diff --git a/qasper-0001/instruction.md b/qasper-0001/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..489c88c2fa15db1f21f7306bea83bca4526797f2 --- /dev/null +++ b/qasper-0001/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Minimally Supervised Learning of Affective Events Using Discourse Relations + +Question: What are the results? \ No newline at end of file diff --git a/qasper-0006/instruction.md b/qasper-0006/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..d1a87a1f6e42626bef4bde526793a75469d82ee2 --- /dev/null +++ b/qasper-0006/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Minimally Supervised Learning of Affective Events Using Discourse Relations + +Question: How does their model learn using mostly raw data? \ No newline at end of file diff --git a/qasper-0007/instruction.md b/qasper-0007/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..02fe9ef357c6c488383f25ee617ac3baef2f3994 --- /dev/null +++ b/qasper-0007/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Minimally Supervised Learning of Affective Events Using Discourse Relations + +Question: How big is seed lexicon used for training? \ No newline at end of file diff --git a/qasper-0009/instruction.md b/qasper-0009/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..6ef21651ce10361d1d38b71c06a04f6968ad8bc8 --- /dev/null +++ b/qasper-0009/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: PO-EMO: Conceptualization, Annotation, and Modeling of Aesthetic Emotions in German and English Poetry + +Question: Does the paper report macro F1? \ No newline at end of file diff --git a/qasper-0023/instruction.md b/qasper-0023/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..df1a520f26f208b8f298d5900955d928956cee33 --- /dev/null +++ b/qasper-0023/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Question Answering based Clinical Text Structuring Using Pre-trained Language Model + +Question: How many questions are in the dataset? \ No newline at end of file diff --git a/qasper-0030/instruction.md b/qasper-0030/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..271aaa0aff8a640aa95ac779898929f13c1dd22e --- /dev/null +++ b/qasper-0030/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Progress and Tradeoffs in Neural Language Models + +Question: What aspects have been compared between various language models? \ No newline at end of file diff --git a/qasper-0031/instruction.md b/qasper-0031/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4c5ad36c45f655ca202ca3bd6fb045d8eceab6ff --- /dev/null +++ b/qasper-0031/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Progress and Tradeoffs in Neural Language Models + +Question: what classic language models are mentioned in the paper? \ No newline at end of file diff --git a/qasper-0036/instruction.md b/qasper-0036/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..71643b8ed7913325464fff34d519c9fecfd1c137 --- /dev/null +++ b/qasper-0036/instruction.md @@ -0,0 +1,673 @@ +Name of Paper: Stay On-Topic: Generating Context-specific Fake Restaurant Reviews + +Question: What kind of model do they use for detection? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Background", + "System Model", + "Attack Model", + "Generative Model" + ], + "paragraphs": [ + [ + "Automatically generated fake reviews have only recently become natural enough to fool human readers. Yao et al. BIBREF0 use a deep neural network (a so-called 2-layer LSTM BIBREF1 ) to generate fake reviews, and concluded that these fake reviews look sufficiently genuine to fool native English speakers. They train their model using real restaurant reviews from yelp.com BIBREF2 . Once trained, the model is used to generate reviews character-by-character. Due to the generation methodology, it cannot be easily targeted for a specific context (meaningful side information). Consequently, the review generation process may stray off-topic. For instance, when generating a review for a Japanese restaurant in Las Vegas, the review generation process may include references to an Italian restaurant in Baltimore. The authors of BIBREF0 apply a post-processing step (customization), which replaces food-related words with more suitable ones (sampled from the targeted restaurant). The word replacement strategy has drawbacks: it can miss certain words and replace others independent of their surrounding words, which may alert savvy readers. As an example: when we applied the customization technique described in BIBREF0 to a review for a Japanese restaurant it changed the snippet garlic knots for breakfast with garlic knots for sushi).", + "We propose a methodology based on neural machine translation (NMT) that improves the generation process by defining a context for the each generated fake review. Our context is a clear-text sequence of: the review rating, restaurant name, city, state and food tags (e.g. Japanese, Italian). We show that our technique generates review that stay on topic. We can instantiate our basic technique into several variants. We vet them on Amazon Mechanical Turk and find that native English speakers are very poor at recognizing our fake generated reviews. For one variant, the participants' performance is close to random: the class-averaged F-score of detection is INLINEFORM0 (whereas random would be INLINEFORM1 given the 1:6 imbalance in the test). Via a user study with experienced, highly educated participants, we compare this variant (which we will henceforth refer to as NMT-Fake* reviews) with fake reviews generated using the char-LSTM-based technique from BIBREF0 .", + "We demonstrate that NMT-Fake* reviews constitute a new category of fake reviews that cannot be detected by classifiers trained only using previously known categories of fake reviews BIBREF0 , BIBREF3 , BIBREF4 . Therefore, NMT-Fake* reviews may go undetected in existing online review sites. To meet this challenge, we develop an effective classifier that detects NMT-Fake* reviews effectively (97% F-score). Our main contributions are:" + ], + [ + "Fake reviews User-generated content BIBREF5 is an integral part of the contemporary user experience on the web. Sites like tripadvisor.com, yelp.com and Google Play use user-written reviews to provide rich information that helps other users choose where to spend money and time. User reviews are used for rating services or products, and for providing qualitative opinions. User reviews and ratings may be used to rank services in recommendations. Ratings have an affect on the outwards appearance. Already 8 years ago, researchers estimated that a one-star rating increase affects the business revenue by 5 \u2013 9% on yelp.com BIBREF6 .", + "Due to monetary impact of user-generated content, some businesses have relied on so-called crowd-turfing agents BIBREF7 that promise to deliver positive ratings written by workers to a customer in exchange for a monetary compensation. Crowd-turfing ethics are complicated. For example, Amazon community guidelines prohibit buying content relating to promotions, but the act of writing fabricated content is not considered illegal, nor is matching workers to customers BIBREF8 . Year 2015, approximately 20% of online reviews on yelp.com were suspected of being fake BIBREF9 .", + "Nowadays, user-generated review sites like yelp.com use filters and fraudulent review detection techniques. These factors have resulted in an increase in the requirements of crowd-turfed reviews provided to review sites, which in turn has led to an increase in the cost of high-quality review. Due to the cost increase, researchers hypothesize the existence of neural network-generated fake reviews. These neural-network-based fake reviews are statistically different from human-written fake reviews, and are not caught by classifiers trained on these BIBREF0 .", + "Detecting fake reviews can either be done on an individual level or as a system-wide detection tool (i.e. regulation). Detecting fake online content on a personal level requires knowledge and skills in critical reading. In 2017, the National Literacy Trust assessed that young people in the UK do not have the skillset to differentiate fake news from real news BIBREF10 . For example, 20% of children that use online news sites in age group 12-15 believe that all information on news sites are true.", + "Neural Networks Neural networks are function compositions that map input data through INLINEFORM0 subsequent layers: DISPLAYFORM0 ", + "where the functions INLINEFORM0 are typically non-linear and chosen by experts partly for known good performance on datasets and partly for simplicity of computational evaluation. Language models (LMs) BIBREF11 are generative probability distributions that assign probabilities to sequences of tokens ( INLINEFORM1 ): DISPLAYFORM0 ", + "such that the language model can be used to predict how likely a specific token at time step INLINEFORM0 is, based on the INLINEFORM1 previous tokens. Tokens are typically either words or characters.", + "For decades, deep neural networks were thought to be computationally too difficult to train. However, advances in optimization, hardware and the availability of frameworks have shown otherwise BIBREF1 , BIBREF12 . Neural language models (NLMs) have been one of the promising application areas. NLMs are typically various forms of recurrent neural networks (RNNs), which pass through the data sequentially and maintain a memory representation of the past tokens with a hidden context vector. There are many RNN architectures that focus on different ways of updating and maintaining context vectors: Long Short-Term Memory units (LSTM) and Gated Recurrent Units (GRUs) are perhaps most popular. Neural LMs have been used for free-form text generation. In certain application areas, the quality has been high enough to sometimes fool human readers BIBREF0 . Encoder-decoder (seq2seq) models BIBREF13 are architectures of stacked RNNs, which have the ability to generate output sequences based on input sequences. The encoder network reads in a sequence of tokens, and passes it to a decoder network (a LM). In contrast to simpler NLMs, encoder-decoder networks have the ability to use additional context for generating text, which enables more accurate generation of text. Encoder-decoder models are integral in Neural Machine Translation (NMT) BIBREF14 , where the task is to translate a source text from one language to another language. NMT models additionally use beam search strategies to heuristically search the set of possible translations. Training datasets are parallel corpora; large sets of paired sentences in the source and target languages. The application of NMT techniques for online machine translation has significantly improved the quality of translations, bringing it closer to human performance BIBREF15 .", + "Neural machine translation models are efficient at mapping one expression to another (one-to-one mapping). Researchers have evaluated these models for conversation generation BIBREF16 , with mixed results. Some researchers attribute poor performance to the use of the negative log likelihood cost function during training, which emphasizes generation of high-confidence phrases rather than diverse phrases BIBREF17 . The results are often generic text, which lacks variation. Li et al. have suggested various augmentations to this, among others suppressing typical responses in the decoder language model to promote response diversity BIBREF17 ." + ], + [ + "We discuss the attack model, our generative machine learning method and controlling the generative process in this section." + ], + [ + "Wang et al. BIBREF7 described a model of crowd-turfing attacks consisting of three entities: customers who desire to have fake reviews for a particular target (e.g. their restaurant) on a particular platform (e.g. Yelp), agents who offer fake review services to customers, and workers who are orchestrated by the agent to compose and post fake reviews.", + "Automated crowd-turfing attacks (ACA) replace workers by a generative model. This has several benefits including better economy and scalability (human workers are more expensive and slower) and reduced detectability (agent can better control the rate at which fake reviews are generated and posted).", + "We assume that the agent has access to public reviews on the review platform, by which it can train its generative model. We also assume that it is easy for the agent to create a large number of accounts on the review platform so that account-based detection or rate-limiting techniques are ineffective against fake reviews.", + "The quality of the generative model plays a crucial role in the attack. Yao et al. BIBREF0 propose the use of a character-based LSTM as base for generative model. LSTMs are not conditioned to generate reviews for a specific target BIBREF1 , and may mix-up concepts from different contexts during free-form generation. Mixing contextually separate words is one of the key criteria that humans use to identify fake reviews. These may result in violations of known indicators for fake content BIBREF18 . For example, the review content may not match prior expectations nor the information need that the reader has. We improve the attack model by considering a more capable generative model that produces more appropriate reviews: a neural machine translation (NMT) model." + ], + [ + "We propose the use of NMT models for fake review generation. The method has several benefits: 1) the ability to learn how to associate context (keywords) to reviews, 2) fast training time, and 3) a high-degree of customization during production time, e.g. introduction of specific waiter or food items names into reviews.", + "NMT models are constructions of stacked recurrent neural networks (RNNs). They include an encoder network and a decoder network, which are jointly optimized to produce a translation of one sequence to another. The encoder rolls over the input data in sequence and produces one INLINEFORM0 -dimensional context vector representation for the sentence. The decoder then generates output sequences based on the embedding vector and an attention module, which is taught to associate output words with certain input words. The generation typically continues until a specific EOS (end of sentence) token is encountered. The review length can be controlled in many ways, e.g. by setting the probability of generating the EOS token to zero until the required length is reached.", + "NMT models often also include a beam search BIBREF14 , which generates several hypotheses and chooses the best ones amongst them. In our work, we use the greedy beam search technique. We forgo the use of additional beam searches as we found that the quality of the output was already adequate and the translation phase time consumption increases linearly for each beam used.", + "We use the Yelp Challenge dataset BIBREF2 for our fake review generation. The dataset (Aug 2017) contains 2.9 million 1 \u20135 star restaurant reviews. We treat all reviews as genuine human-written reviews for the purpose of this work, since wide-scale deployment of machine-generated review attacks are not yet reported (Sep 2017) BIBREF19 . As preprocessing, we remove non-printable (non-ASCII) characters and excessive white-space. We separate punctuation from words. We reserve 15,000 reviews for validation and 3,000 for testing, and the rest we use for training. NMT models require a parallel corpus of source and target sentences, i.e. a large set of (source, target)-pairs. We set up a parallel corpus by constructing (context, review)-pairs from the dataset. Next, we describe how we created our input context.", + "The Yelp Challenge dataset includes metadata about restaurants, including their names, food tags, cities and states these restaurants are located in. For each restaurant review, we fetch this metadata and use it as our input context in the NMT model. The corresponding restaurant review is similarly set as the target sentence. This method produced 2.9 million pairs of sentences in our parallel corpus. We show one example of the parallel training corpus in Example 1 below:", + "5 Public House Las Vegas NV Gastropubs Restaurants > Excellent", + "food and service . Pricey , but well worth it . I would recommend", + "the bone marrow and sampler platter for appetizers . \\end{verbatim}", + " ", + " ", + "\\noindent The order {\\textbf{[rating name city state tags]}} is kept constant.", + "Training the model conditions it to associate certain sequences of words in the input sentence with others in the output.", + " ", + "\\subsubsection{Training Settings}", + " ", + "We train our NMT model on a commodity PC with a i7-4790k CPU (4.00GHz), with 32GB RAM and one NVidia GeForce GTX 980 GPU. Our system can process approximately 1,300 \\textendash 1,500 source tokens/s and approximately 5,730 \\textendash 5,830 output tokens/s. Training one epoch takes in average 72 minutes. The model is trained for 8 epochs, i.e. over night. We call fake review generated by this model \\emph{NMT-Fake reviews}. We only need to train one model to produce reviews of different ratings.", + "We use the training settings: adam optimizer \\cite{kingma2014adam} with the suggested learning rate 0.001 \\cite{klein2017opennmt}. For most parts, parameters are at their default values. Notably, the maximum sentence length of input and output is 50 tokens by default.", + "We leverage the framework openNMT-py \\cite{klein2017opennmt} to teach the our NMT model.", + "We list used openNMT-py commands in Appendix Table~\\ref{table:openNMT-py_commands}.", + " ", + "\\begin{figure}[t]", + "\\begin{center}", + " \\begin{tabular}{ | l | }", + " \\hline", + "Example 2. Greedy NMT \\\\", + "Great food, \\underline{great} service, \\underline{great} \\textit{\\textit{beer selection}}. I had the \\textit{Gastropubs burger} and it", + "\\\\", + "was delicious. The \\underline{\\textit{beer selection}} was also \\underline{great}. \\\\", + "\\\\", + "Example 3. NMT-Fake* \\\\", + "I love this restaurant. Great food, great service. It's \\textit{a little pricy} but worth\\\\", + "it for the \\textit{quality} of the \\textit{beer} and atmosphere you can see in \\textit{Vegas}", + "\\\\", + " \\hline", + " \\end{tabular}", + " \\label{table:output_comparison}", + "\\end{center}", + "\\caption{Na\\\"{i}ve text generation with NMT vs. generation using our NTM model. Repetitive patterns are \\underline{underlined}. Contextual words are \\emph{italicized}. Both examples here are generated based on the context given in Example~1.}", + "\\label{fig:comparison}", + "\\end{figure}", + " ", + "\\subsection{Controlling generation of fake reviews}", + "\\label{sec:generating}", + " ", + "Greedy NMT beam searches are practical in many NMT cases. However, the results are simply repetitive, when naively applied to fake review generation (See Example~2 in Figure~\\ref{fig:comparison}).", + "The NMT model produces many \\emph{high-confidence} word predictions, which are repetitive and obviously fake. We calculated that in fact, 43\\% of the generated sentences started with the phrase ``Great food''. The lack of diversity in greedy use of NMTs for text generation is clear.", + " ", + " ", + "\\begin{algorithm}[!b]", + " \\KwData{Desired review context $C_\\mathrm{input}$ (given as cleartext), NMT model}", + " \\KwResult{Generated review $out$ for input context $C_\\mathrm{input}$}", + "set $b=0.3$, $\\lambda=-5$, $\\alpha=\\frac{2}{3}$, $p_\\mathrm{typo}$, $p_\\mathrm{spell}$ \\\\", + "$\\log p \\leftarrow \\text{NMT.decode(NMT.encode(}C_\\mathrm{input}\\text{))}$ \\\\", + "out $\\leftarrow$ [~] \\\\", + "$i \\leftarrow 0$ \\\\", + "$\\log p \\leftarrow \\text{Augment}(\\log p$, $b$, $\\lambda$, $1$, $[~]$, 0)~~~~~~~~~~~~~~~ |~random penalty~\\\\", + "\\While{$i=0$ or $o_i$ not EOS}{", + "$\\log \\Tilde{p} \\leftarrow \\text{Augment}(\\log p$, $b$, $\\lambda$, $\\alpha$, $o_i$, $i$)~~~~~~~~~~~ |~start \\& memory penalty~\\\\", + "$o_i \\leftarrow$ \\text{NMT.beam}($\\log \\Tilde{p}$, out) \\\\", + "out.append($o_i$) \\\\", + "$i \\leftarrow i+1$", + "}\\text{return}~$\\text{Obfuscate}$(out,~$p_\\mathrm{typo}$,~$p_\\mathrm{spell}$)", + "\\caption{Generation of NMT-Fake* reviews.}", + "\\label{alg:base}", + "\\end{algorithm}", + " ", + "In this work, we describe how we succeeded in creating more diverse and less repetitive generated reviews, such as Example 3 in Figure~\\ref{fig:comparison}.", + "We outline pseudocode for our methodology of generating fake reviews in Algorithm~\\ref{alg:base}. There are several parameters in our algorithm.", + "The details of the algorithm will be shown later.", + "We modify the openNMT-py translation phase by changing log-probabilities before passing them to the beam search.", + "We notice that reviews generated with openNMT-py contain almost no language errors. As an optional post-processing step, we obfuscate reviews by introducing natural typos/misspellings randomly. In the next sections, we describe how we succeeded in generating more natural sentences from our NMT model, i.e. generating reviews like Example~3 instead of reviews like Example~2.", + " ", + "\\subsubsection{Variation in word content}", + " ", + "Example 2 in Figure~\\ref{fig:comparison} repeats commonly occurring words given for a specific context (e.g. \\textit{great, food, service, beer, selection, burger} for Example~1). Generic review generation can be avoided by decreasing probabilities (log-likelihoods \\cite{murphy2012machine}) of the generators LM, the decoder.", + "We constrain the generation of sentences by randomly \\emph{imposing penalties to words}.", + "We tried several forms of added randomness, and found that adding constant penalties to a \\emph{random subset} of the target words resulted in the most natural sentence flow. We call these penalties \\emph{Bernoulli penalties}, since the random variables are chosen as either 1 or 0 (on or off).", + " ", + " ", + "\\paragraph{Bernoulli penalties to language model}", + "To avoid generic sentences components, we augment the default language model $p(\\cdot)$ of the decoder by", + " ", + "\\begin{equation}", + "\\log \\Tilde{p}(t_k) = \\log p(t_k | t_i, \\dots, t_1) + \\lambda q,", + "\\end{equation}", + " ", + "where $q \\in R^{V}$ is a vector of Bernoulli-distributed random values that obtain values $1$ with probability $b$ and value $0$ with probability $1-b_i$, and $\\lambda < 0$. Parameter $b$ controls how much of the vocabulary is forgotten and $\\lambda$ is a soft penalty of including ``forgotten'' words in a review.", + "$\\lambda q_k$ emphasizes sentence forming with non-penalized words. The randomness is reset at the start of generating a new review.", + "Using Bernoulli penalties in the language model, we can ``forget'' a certain proportion of words and essentially ``force'' the creation of less typical sentences. We will test the effect of these two parameters, the Bernoulli probability $b$ and log-likelihood penalty of including ``forgotten'' words $\\lambda$, with a user study in Section~\\ref{sec:varying}.", + " ", + "\\paragraph{Start penalty}", + "We introduce start penalties to avoid generic sentence starts (e.g. ``Great food, great service''). Inspired by \\cite{li2016diversity}, we add a random start penalty $\\lambda s^\\mathrm{i}$, to our language model, which decreases monotonically for each generated token. We set $\\alpha \\leftarrow 0.66$ as it's effect decreases by 90\\% every 5 words generated.", + " ", + "\\paragraph{Penalty for reusing words}", + "Bernoulli penalties do not prevent excessive use of certain words in a sentence (such as \\textit{great} in Example~2).", + "To avoid excessive reuse of words, we included a memory penalty for previously used words in each translation.", + "Concretely, we add the penalty $\\lambda$ to each word that has been generated by the greedy search.", + " ", + "\\subsubsection{Improving sentence coherence}", + "\\label{sec:grammar}", + "We visually analyzed reviews after applying these penalties to our NMT model. While the models were clearly diverse, they were \\emph{incoherent}: the introduction of random penalties had degraded the grammaticality of the sentences. Amongst others, the use of punctuation was erratic, and pronouns were used semantically wrongly (e.g. \\emph{he}, \\emph{she} might be replaced, as could ``and''/``but''). To improve the authenticity of our reviews, we added several \\emph{grammar-based rules}.", + " ", + "English language has several classes of words which are important for the natural flow of sentences.", + "We built a list of common pronouns (e.g. I, them, our), conjunctions (e.g. and, thus, if), punctuation (e.g. ,/.,..), and apply only half memory penalties for these words. We found that this change made the reviews more coherent. The pseudocode for this and the previous step is shown in Algorithm~\\ref{alg:aug}.", + "The combined effect of grammar-based rules and LM augmentation is visible in Example~3, Figure~\\ref{fig:comparison}.", + " ", + "\\begin{algorithm}[!t]", + " \\KwData{Initial log LM $\\log p$, Bernoulli probability $b$, soft-penalty $\\lambda$, monotonic factor $\\alpha$, last generated token $o_i$, grammar rules set $G$}", + " \\KwResult{Augmented log LM $\\log \\Tilde{p}$}", + "\\begin{algorithmic}[1]", + "\\Procedure {Augment}{$\\log p$, $b$, $\\lambda$, $\\alpha$, $o_i$, $i$}{ \\\\", + "generate $P_{\\mathrm{1:N}} \\leftarrow Bernoulli(b)$~~~~~~~~~~~~~~~|~$\\text{One value} \\in \\{0,1\\}~\\text{per token}$~ \\\\", + "$I \\leftarrow P>0$ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~|~Select positive indices~\\\\", + "$\\log \\Tilde{p} \\leftarrow$ $\\text{Discount}$($\\log p$, $I$, $\\lambda \\cdot \\alpha^i$,$G$) ~~~~~~ |~start penalty~\\\\", + "$\\log \\Tilde{p} \\leftarrow$ $\\text{Discount}$($\\log \\Tilde{p}$, $[o_i]$, $\\lambda$,$G$) ~~~~~~~~~ |~memory penalty~\\\\", + "\\textbf{return}~$\\log \\Tilde{p}$", + "}", + "\\EndProcedure", + "\\\\", + "\\Procedure {Discount}{$\\log p$, $I$, $\\lambda$, $G$}{", + "\\State{\\For{$i \\in I$}{", + "\\eIf{$o_i \\in G$}{", + "$\\log p_{i} \\leftarrow \\log p_{i} + \\lambda/2$", + "}{", + "$\\log p_{i} \\leftarrow \\log p_{i} + \\lambda$}", + "}\\textbf{return}~$\\log p$", + "\\EndProcedure", + "}}", + "\\end{algorithmic}", + "\\caption{Pseudocode for augmenting language model. }", + "\\label{alg:aug}", + "\\end{algorithm}", + " ", + "\\subsubsection{Human-like errors}", + "\\label{sec:obfuscation}", + "We notice that our NMT model produces reviews without grammar mistakes.", + "This is unlike real human writers, whose sentences contain two types of language mistakes 1) \\emph{typos} that are caused by mistakes in the human motoric input, and 2) \\emph{common spelling mistakes}.", + "We scraped a list of common English language spelling mistakes from Oxford dictionary\\footnote{\\url{https://en.oxforddictionaries.com/spelling/common-misspellings}} and created 80 rules for randomly \\emph{re-introducing spelling mistakes}.", + "Similarly, typos are randomly reintroduced based on the weighted edit distance\\footnote{\\url{https://pypi.python.org/pypi/weighted-levenshtein/0.1}}, such that typos resulting in real English words with small perturbations are emphasized.", + "We use autocorrection tools\\footnote{\\url{https://pypi.python.org/pypi/autocorrect/0.1.0}} for finding these words.", + "We call these augmentations \\emph{obfuscations}, since they aim to confound the reader to think a human has written them. We omit the pseudocode description for brevity.", + " ", + "\\subsection{Experiment: Varying generation parameters in our NMT model}", + "\\label{sec:varying}", + " ", + "Parameters $b$ and $\\lambda$ control different aspects in fake reviews.", + "We show six different examples of generated fake reviews in Table~\\ref{table:categories}.", + "Here, the largest differences occur with increasing values of $b$: visibly, the restaurant reviews become more extreme.", + "This occurs because a large portion of vocabulary is ``forgotten''. Reviews with $b \\geq 0.7$ contain more rare word combinations, e.g. ``!!!!!'' as punctuation, and they occasionally break grammaticality (''experience was awesome'').", + "Reviews with lower $b$ are more generic: they contain safe word combinations like ``Great place, good service'' that occur in many reviews. Parameter $\\lambda$'s is more subtle: it affects how random review starts are and to a degree, the discontinuation between statements within the review.", + "We conducted an Amazon Mechanical Turk (MTurk) survey in order to determine what kind of NMT-Fake reviews are convincing to native English speakers. We describe the survey and results in the next section.", + " ", + " ", + "\\begin{table}[!b]", + "\\caption{Six different parametrizations of our NMT reviews and one example for each. The context is ``5 P~.~F~.~Chang ' s Scottsdale AZ'' in all examples.}", + "\\begin{center}", + " \\begin{tabular}{ | l | l | }", + " \\hline", + " $(b, \\lambda)$ & Example review for context \\\\ \\hline", + " \\hline", + " $(0.3, -3)$ & I love this location! Great service, great food and the best drinks in Scottsdale. \\\\", + " & The staff is very friendly and always remembers u when we come in\\\\\\hline", + " $(0.3, -5)$ & Love love the food here! I always go for lunch. They have a great menu and \\\\", + " & they make it fresh to order. Great place, good service and nice staff\\\\\\hline", + " $(0.5, -4)$ & I love their chicken lettuce wraps and fried rice!! The service is good, they are\\\\", + " & always so polite. They have great happy hour specials and they have a lot\\\\", + " & of options.\\\\\\hline", + " $(0.7, -3)$ & Great place to go with friends! They always make sure your dining \\\\", + " & experience was awesome.\\\\ \\hline", + " $(0.7, -5)$ & Still haven't ordered an entree before but today we tried them once..\\\\", + " & both of us love this restaurant....\\\\\\hline", + " $(0.9, -4)$ & AMAZING!!!!! Food was awesome with excellent service. Loved the lettuce \\\\", + " & wraps. Great drinks and wine! Can't wait to go back so soon!!\\\\ \\hline", + " \\end{tabular}", + " \\label{table:categories}", + "\\end{center}", + "\\end{table}", + " ", + "\\subsubsection{MTurk study}", + "\\label{sec:amt}", + "We created 20 jobs, each with 100 questions, and requested master workers in MTurk to complete the jobs.", + "We randomly generated each survey for the participants. Each review had a 50\\% chance to be real or fake. The fake ones further were chosen among six (6) categories of fake reviews (Table~\\ref{table:categories}).", + "The restaurant and the city was given as contextual information to the participants. Our aim was to use this survey to understand how well English-speakers react to different parametrizations of NMT-Fake reviews.", + "Table~\\ref{table:amt_pop} in Appendix summarizes the statistics for respondents in the survey. All participants were native English speakers from America. The base rate (50\\%) was revealed to the participants prior to the study.", + " ", + "We first investigated overall detection of any NMT-Fake reviews (1,006 fake reviews and 994 real reviews). We found that the participants had big difficulties in detecting our fake reviews. In average, the reviews were detected with class-averaged \\emph{F-score of only 56\\%}, with 53\\% F-score for fake review detection and 59\\% F-score for real review detection. The results are very close to \\emph{random detection}, where precision, recall and F-score would each be 50\\%. Results are recorded in Table~\\ref{table:MTurk_super}. Overall, the fake review generation is very successful, since human detection rate across categories is close to random.", + " ", + "\\begin{table}[t]", + "\\caption{Effectiveness of Mechanical Turkers in distinguishing human-written reviews from fake reviews generated by our NMT model (all variants).}", + "\\begin{center}", + " \\begin{tabular}{ | c | c |c |c | c | }", + " \\hline", + " \\multicolumn{5}{|c|}{Classification report}", + " \\\\ \\hline", + " Review Type & Precision & Recall & F-score & Support \\\\ \\hline", + " \\hline", + " Human & 55\\% & 63\\% & 59\\% & 994\\\\", + " NMT-Fake & 57\\% & 50\\% & 53\\% & 1006 \\\\", + " \\hline", + " \\end{tabular}", + " \\label{table:MTurk_super}", + "\\end{center}", + "\\end{table}", + " ", + "We noticed some variation in the detection of different fake review categories. The respondents in our MTurk survey had most difficulties recognizing reviews of category $(b=0.3, \\lambda=-5)$, where true positive rate was $40.4\\%$, while the true negative rate of the real class was $62.7\\%$. The precision were $16\\%$ and $86\\%$, respectively. The class-averaged F-score is $47.6\\%$, which is close to random. Detailed classification reports are shown in Table~\\ref{table:MTurk_sub} in Appendix. Our MTurk-study shows that \\emph{our NMT-Fake reviews pose a significant threat to review systems}, since \\emph{ordinary native English-speakers have very big difficulties in separating real reviews from fake reviews}. We use the review category $(b=0.3, \\lambda=-5)$ for future user tests in this paper, since MTurk participants had most difficulties detecting these reviews. We refer to this category as NMT-Fake* in this paper.", + " ", + "\\section{Evaluation}", + "\\graphicspath{ {figures/}}", + " ", + "We evaluate our fake reviews by first comparing them statistically to previously proposed types of fake reviews, and proceed with a user study with experienced participants. We demonstrate the statistical difference to existing fake review types \\cite{yao2017automated,mukherjee2013yelp,rayana2015collective} by training classifiers to detect previous types and investigate classification performance.", + " ", + "\\subsection{Replication of state-of-the-art model: LSTM}", + "\\label{sec:repl}", + " ", + "Yao et al. \\cite{yao2017automated} presented the current state-of-the-art generative model for fake reviews. The model is trained over the Yelp Challenge dataset using a two-layer character-based LSTM model.", + "We requested the authors of \\cite{yao2017automated} for access to their LSTM model or a fake review dataset generated by their model. Unfortunately they were not able to share either of these with us. We therefore replicated their model as closely as we could, based on their paper and e-mail correspondence\\footnote{We are committed to sharing our code with bonafide researchers for the sake of reproducibility.}.", + " ", + "We used the same graphics card (GeForce GTX) and trained using the same framework (torch-RNN in lua). We downloaded the reviews from Yelp Challenge and preprocessed the data to only contain printable ASCII characters, and filtered out non-restaurant reviews. We trained the model for approximately 72 hours. We post-processed the reviews using the customization methodology described in \\cite{yao2017automated} and email correspondence. We call fake reviews generated by this model LSTM-Fake reviews.", + " ", + "\\subsection{Similarity to existing fake reviews}", + "\\label{sec:automated}", + " ", + "We now want to understand how NMT-Fake* reviews compare to a) LSTM fake reviews and b) human-generated fake reviews. We do this by comparing the statistical similarity between these classes.", + " ", + "For `a' (Figure~\\ref{fig:lstm}), we use the Yelp Challenge dataset. We trained a classifier using 5,000 random reviews from the Yelp Challenge dataset (``human'') and 5,000 fake reviews generated by LSTM-Fake. Yao et al. \\cite{yao2017automated} found that character features are essential in identifying LSTM-Fake reviews. Consequently, we use character features (n-grams up to 3).", + " ", + "For `b' (Figure~\\ref{fig:shill}),we the ``Yelp Shills'' dataset (combination of YelpZip \\cite{mukherjee2013yelp}, YelpNYC \\cite{mukherjee2013yelp}, YelpChi \\cite{rayana2015collective}). This dataset labels entries that are identified as fraudulent by Yelp's filtering mechanism (''shill reviews'')\\footnote{Note that shill reviews are probably generated by human shills \\cite{zhao2017news}.}. The rest are treated as genuine reviews from human users (''genuine''). We use 100,000 reviews from each category to train a classifier. We use features from the commercial psychometric tool LIWC2015 \\cite{pennebaker2015development} to generated features.", + " ", + "In both cases, we use AdaBoost (with 200 shallow decision trees) for training. For testing each classifier, we use a held out test set of 1,000 reviews from both classes in each case. In addition, we test 1,000 NMT-Fake* reviews. Figures~\\ref{fig:lstm} and~\\ref{fig:shill} show the results. The classification threshold of 50\\% is marked with a dashed line.", + " ", + "\\begin{figure}", + " \\begin{subfigure}[b]{0.5\\columnwidth}", + " \\includegraphics[width=\\columnwidth]{figures/lstm.png}", + " \\caption{Human--LSTM reviews.}", + " \\label{fig:lstm}", + " \\end{subfigure}", + " \\begin{subfigure}[b]{0.5\\columnwidth}", + " \\includegraphics[width=\\columnwidth]{figures/distribution_shill.png}", + " \\caption{Genuine--Shill reviews.}", + " \\label{fig:shill}", + " \\end{subfigure}", + " \\caption{", + " Histogram comparison of NMT-Fake* reviews with LSTM-Fake reviews and human-generated (\\emph{genuine} and \\emph{shill}) reviews. Figure~\\ref{fig:lstm} shows that a classifier trained to distinguish ``human'' vs. LSTM-Fake cannot distinguish ``human'' vs NMT-Fake* reviews. Figure~\\ref{fig:shill} shows NMT-Fake* reviews are more similar to \\emph{genuine} reviews than \\emph{shill} reviews.", + " }", + " \\label{fig:statistical_similarity}", + "\\end{figure}", + " ", + "We can see that our new generated reviews do not share strong attributes with previous known categories of fake reviews. If anything, our fake reviews are more similar to genuine reviews than previous fake reviews. We thus conjecture that our NMT-Fake* fake reviews present a category of fake reviews that may go undetected on online review sites.", + " ", + " ", + "\\subsection{Comparative user study}", + "\\label{sec:comparison}", + "We wanted to evaluate the effectiveness of fake reviews againsttech-savvy users who understand and know to expect machine-generated fake reviews. We conducted a user study with 20 participants, all with computer science education and at least one university degree. Participant demographics are shown in Table~\\ref{table:amt_pop} in the Appendix. Each participant first attended a training session where they were asked to label reviews (fake and genuine) and could later compare them to the correct answers -- we call these participants \\emph{experienced participants}.", + "No personal data was collected during the user study.", + " ", + "Each person was given two randomly selected sets of 30 of reviews (a total of 60 reviews per person) with reviews containing 10 \\textendash 50 words each.", + "Each set contained 26 (87\\%) real reviews from Yelp and 4 (13\\%) machine-generated reviews,", + "numbers chosen based on suspicious review prevalence on Yelp~\\cite{mukherjee2013yelp,rayana2015collective}.", + "One set contained machine-generated reviews from one of the two models (NMT ($b=0.3, \\lambda=-5$) or LSTM),", + "and the other set reviews from the other in randomized order. The number of fake reviews was revealed to each participant in the study description. Each participant was requested to mark four (4) reviews as fake.", + " ", + "Each review targeted a real restaurant. A screenshot of that restaurant's Yelp page was shown to each participant prior to the study. Each participant evaluated reviews for one specific, randomly selected, restaurant. An example of the first page of the user study is shown in Figure~\\ref{fig:screenshot} in Appendix.", + " ", + "\\begin{figure}[!ht]", + "\\centering", + "\\includegraphics[width=.7\\columnwidth]{detection2.png}", + "\\caption{Violin plots of detection rate in comparative study. Mean and standard deviations for number of detected fakes are $0.8\\pm0.7$ for NMT-Fake* and $2.5\\pm1.0$ for LSTM-Fake. $n=20$. A sample of random detection is shown as comparison.}", + "\\label{fig:aalto}", + "\\end{figure}", + " ", + " ", + "Figure~\\ref{fig:aalto} shows the distribution of detected reviews of both types. A hypothetical random detector is shown for comparison.", + "NMT-Fake* reviews are significantly more difficult to detect for our experienced participants. In average, detection rate (recall) is $20\\%$ for NMT-Fake* reviews, compared to $61\\%$ for LSTM-based reviews.", + "The precision (and F-score) is the same as the recall in our study, since participants labeled 4 fakes in each set of 30 reviews \\cite{murphy2012machine}.", + "The distribution of the detection across participants is shown in Figure~\\ref{fig:aalto}. \\emph{The difference is statistically significant with confidence level $99\\%$} (Welch's t-test).", + "We compared the detection rate of NMT-Fake* reviews to a random detector, and find that \\emph{our participants detection rate of NMT-Fake* reviews is not statistically different from random predictions with 95\\% confidence level} (Welch's t-test).", + " ", + " ", + "\\section{Defenses}", + " ", + "\\label{sec:detection}", + " ", + "We developed an AdaBoost-based classifier to detect our new fake reviews, consisting of 200 shallow decision trees (depth 2). The features we used are recorded in Table~\\ref{table:features_adaboost} (Appendix).", + "We used word-level features based on spaCy-tokenization \\cite{honnibal-johnson:2015:EMNLP} and constructed n-gram representation of POS-tags and dependency tree tags. We added readability features from NLTK~\\cite{bird2004nltk}.", + " ", + "\\begin{figure}[ht]", + "\\centering", + "\\includegraphics[width=.7\\columnwidth]{obf_score_fair_2.png}", + "\\caption{", + "Adaboost-based classification of NMT-Fake and human-written reviews.", + "Effect of varying $b$ and $\\lambda$ in fake review generation.", + "The variant native speakers had most difficulties detecting is well detectable by AdaBoost (97\\%).}", + "\\label{fig:adaboost_matrix_b_lambda}", + "\\end{figure}", + " ", + " ", + "Figure~\\ref{fig:adaboost_matrix_b_lambda} shows our AdaBoost classifier's class-averaged F-score at detecting different kind of fake reviews. The classifier is very effective in detecting reviews that humans have difficulties detecting. For example, the fake reviews MTurk users had most difficulty detecting ($b=0.3, \\lambda=-5$) are detected with an excellent 97\\% F-score.", + "The most important features for the classification were counts for frequently occurring words in fake reviews (such as punctuation, pronouns, articles) as well as the readability feature ``Automated Readability Index''. We thus conclude that while NMT-Fake reviews are difficult to detect for humans, they can be well detected with the right tools.", + " ", + "\\section{Related Work}", + " ", + "Kumar and Shah~\\cite{kumar2018false} survey and categorize false information research. Automatically generated fake reviews are a form of \\emph{opinion-based false information}, where the creator of the review may influence reader's opinions or decisions.", + "Yao et al. \\cite{yao2017automated} presented their study on machine-generated fake reviews. Contrary to us, they investigated character-level language models, without specifying a specific context before generation. We leverage existing NMT tools to encode a specific context to the restaurant before generating reviews.", + "Supporting our study, Everett et al~\\cite{Everett2016Automated} found that security researchers were less likely to be fooled by Markov chain-generated Reddit comments compared to ordinary Internet users.", + " ", + "Diversification of NMT model outputs has been studied in \\cite{li2016diversity}. The authors proposed the use of a penalty to commonly occurring sentences (\\emph{n-grams}) in order to emphasize maximum mutual information-based generation.", + "The authors investigated the use of NMT models in chatbot systems.", + "We found that unigram penalties to random tokens (Algorithm~\\ref{alg:aug}) was easy to implement and produced sufficiently diverse responses.", + " ", + "\\section {Discussion and Future Work}", + " ", + "\\paragraph{What makes NMT-Fake* reviews difficult to detect?} First, NMT models allow the encoding of a relevant context for each review, which narrows down the possible choices of words that the model has to choose from. Our NMT model had a perplexity of approximately $25$, while the model of \\cite{yao2017automated} had a perplexity of approximately $90$ \\footnote{Personal communication with the authors}. Second, the beam search in NMT models narrows down choices to natural-looking sentences. Third, we observed that the NMT model produced \\emph{better structure} in the generated sentences (i.e. a more coherent story).", + " ", + "\\paragraph{Cost of generating reviews} With our setup, generating one review took less than one second. The cost of generation stems mainly from the overnight training. Assuming an electricity cost of 16 cents / kWh (California) and 8 hours of training, training the NMT model requires approximately 1.30 USD. This is a 90\\% reduction in time compared to the state-of-the-art \\cite{yao2017automated}. Furthermore, it is possible to generate both positive and negative reviews with the same model.", + " ", + "\\paragraph{Ease of customization} We experimented with inserting specific words into the text by increasing their log likelihoods in the beam search. We noticed that the success depended on the prevalence of the word in the training set. For example, adding a +5 to \\emph{Mike} in the log-likelihood resulted in approximately 10\\% prevalence of this word in the reviews. An attacker can therefore easily insert specific keywords to reviews, which can increase evasion probability.", + " ", + "\\paragraph{Ease of testing} Our diversification scheme is applicable during \\emph{generation phase}, and does not affect the training setup of the network in any way. Once the NMT model is obtained, it is easy to obtain several different variants of NMT-Fake reviews by varying parameters $b$ and $\\lambda$.", + " ", + " ", + " ", + "\\paragraph{Languages} The generation methodology is not per-se language-dependent. The requirement for successful generation is that sufficiently much data exists in the targeted language. However, our language model modifications require some knowledge of that target language's grammar to produce high-quality reviews.", + " ", + "\\paragraph{Generalizability of detection techniques} Currently, fake reviews are not universally detectable. Our results highlight that it is difficult to claim detection performance on unseen types of fake reviews (Section~\\ref{sec:automated}). We see this an open problem that deserves more attention in fake reviews research.", + " ", + "\\paragraph{Generalizability to other types of datasets} Our technique can be applied to any dataset, as long as there is sufficient training data for the NMT model. We used approximately 2.9 million reviews for this work.", + " ", + "\\section{Conclusion}", + " ", + "In this paper, we showed that neural machine translation models can be used to generate fake reviews that are very effective in deceiving even experienced, tech-savvy users.", + "This supports anecdotal evidence \\cite{national2017commission}.", + "Our technique is more effective than state-of-the-art \\cite{yao2017automated}.", + "We conclude that machine-aided fake review detection is necessary since human users are ineffective in identifying fake reviews.", + "We also showed that detectors trained using one type of fake reviews are not effective in identifying other types of fake reviews.", + "Robust detection of fake reviews is thus still an open problem.", + " ", + " ", + "\\section*{Acknowledgments}", + "We thank Tommi Gr\\\"{o}ndahl for assistance in planning user studies and the", + "participants of the user study for their time and feedback. We also thank", + "Luiza Sayfullina for comments that improved the manuscript.", + "We thank the authors of \\cite{yao2017automated} for answering questions about", + "their work.", + " ", + " ", + "\\bibliographystyle{splncs}", + "\\begin{thebibliography}{10}", + " ", + "\\bibitem{yao2017automated}", + "Yao, Y., Viswanath, B., Cryan, J., Zheng, H., Zhao, B.Y.:", + "\\newblock Automated crowdturfing attacks and defenses in online review systems.", + "\\newblock In: Proceedings of the 2017 ACM SIGSAC Conference on Computer and", + " Communications Security, ACM (2017)", + " ", + "\\bibitem{murphy2012machine}", + "Murphy, K.:", + "\\newblock Machine learning: a probabilistic approach.", + "\\newblock Massachusetts Institute of Technology (2012)", + " ", + "\\bibitem{challenge2013yelp}", + "Yelp:", + "\\newblock {Yelp Challenge Dataset} (2013)", + " ", + "\\bibitem{mukherjee2013yelp}", + "Mukherjee, A., Venkataraman, V., Liu, B., Glance, N.:", + "\\newblock What yelp fake review filter might be doing?", + "\\newblock In: Seventh International AAAI Conference on Weblogs and Social Media", + " (ICWSM). (2013)", + " ", + "\\bibitem{rayana2015collective}", + "Rayana, S., Akoglu, L.:", + "\\newblock Collective opinion spam detection: Bridging review networks and", + " metadata.", + "\\newblock In: {}Proceedings of the 21th ACM SIGKDD International Conference on", + " Knowledge Discovery and Data Mining", + " ", + "\\bibitem{o2008user}", + "{O'Connor}, P.:", + "\\newblock {User-generated content and travel: A case study on Tripadvisor.com}.", + "\\newblock Information and communication technologies in tourism 2008 (2008)", + " ", + "\\bibitem{luca2010reviews}", + "Luca, M.:", + "\\newblock {Reviews, Reputation, and Revenue: The Case of Yelp. com}.", + "\\newblock {Harvard Business School} (2010)", + " ", + "\\bibitem{wang2012serf}", + "Wang, G., Wilson, C., Zhao, X., Zhu, Y., Mohanlal, M., Zheng, H., Zhao, B.Y.:", + "\\newblock Serf and turf: crowdturfing for fun and profit.", + "\\newblock In: Proceedings of the 21st international conference on World Wide", + " Web (WWW), ACM (2012)", + " ", + "\\bibitem{rinta2017understanding}", + "Rinta-Kahila, T., Soliman, W.:", + "\\newblock Understanding crowdturfing: The different ethical logics behind the", + " clandestine industry of deception.", + "\\newblock In: ECIS 2017: Proceedings of the 25th European Conference on", + " Information Systems. (2017)", + " ", + "\\bibitem{luca2016fake}", + "Luca, M., Zervas, G.:", + "\\newblock Fake it till you make it: Reputation, competition, and yelp review", + " fraud.", + "\\newblock Management Science (2016)", + " ", + "\\bibitem{national2017commission}", + "{National Literacy Trust}:", + "\\newblock Commission on fake news and the teaching of critical literacy skills", + " in schools URL:", + " \\url{https://literacytrust.org.uk/policy-and-campaigns/all-party-parliamentary-group-literacy/fakenews/}.", + " ", + "\\bibitem{jurafsky2014speech}", + "Jurafsky, D., Martin, J.H.:", + "\\newblock Speech and language processing. Volume~3.", + "\\newblock Pearson London: (2014)", + " ", + "\\bibitem{kingma2014adam}", + "Kingma, D.P., Ba, J.:", + "\\newblock Adam: A method for stochastic optimization.", + "\\newblock arXiv preprint arXiv:1412.6980 (2014)", + " ", + "\\bibitem{cho2014learning}", + "Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F.,", + " Schwenk, H., Bengio, Y.:", + "\\newblock Learning phrase representations using rnn encoder--decoder for", + " statistical machine translation.", + "\\newblock In: Proceedings of the 2014 Conference on Empirical Methods in", + " Natural Language Processing (EMNLP). (2014)", + " ", + "\\bibitem{klein2017opennmt}", + "Klein, G., Kim, Y., Deng, Y., Senellart, J., Rush, A.:", + "\\newblock Opennmt: Open-source toolkit for neural machine translation.", + "\\newblock Proceedings of ACL, System Demonstrations (2017)", + " ", + "\\bibitem{wu2016google}", + "Wu, Y., Schuster, M., Chen, Z., Le, Q.V., Norouzi, M., Macherey, W., Krikun,", + " M., Cao, Y., Gao, Q., Macherey, K., et~al.:", + "\\newblock Google's neural machine translation system: Bridging the gap between", + " human and machine translation.", + "\\newblock arXiv preprint arXiv:1609.08144 (2016)", + " ", + "\\bibitem{mei2017coherent}", + "Mei, H., Bansal, M., Walter, M.R.:", + "\\newblock Coherent dialogue with attention-based language models.", + "\\newblock In: AAAI. (2017) 3252--3258", + " ", + "\\bibitem{li2016diversity}", + "Li, J., Galley, M., Brockett, C., Gao, J., Dolan, B.:", + "\\newblock A diversity-promoting objective function for neural conversation", + " models.", + "\\newblock In: Proceedings of NAACL-HLT. (2016)", + " ", + "\\bibitem{rubin2006assessing}", + "Rubin, V.L., Liddy, E.D.:", + "\\newblock Assessing credibility of weblogs.", + "\\newblock In: AAAI Spring Symposium: Computational Approaches to Analyzing", + " Weblogs. (2006)", + " ", + "\\bibitem{zhao2017news}", + "news.com.au:", + "\\newblock {The potential of AI generated 'crowdturfing' could undermine online", + " reviews and dramatically erode public trust} URL:", + " \\url{http://www.news.com.au/technology/online/security/the-potential-of-ai-generated-crowdturfing-could-undermine-online-reviews-and-dramatically-erode-public-trust/news-story/e1c84ad909b586f8a08238d5f80b6982}.", + " ", + "\\bibitem{pennebaker2015development}", + "Pennebaker, J.W., Boyd, R.L., Jordan, K., Blackburn, K.:", + "\\newblock {The development and psychometric properties of LIWC2015}.", + "\\newblock Technical report (2015)", + " ", + "\\bibitem{honnibal-johnson:2015:EMNLP}", + "Honnibal, M., Johnson, M.:", + "\\newblock An improved non-monotonic transition system for dependency parsing.", + "\\newblock In: Proceedings of the 2015 Conference on Empirical Methods in", + " Natural Language Processing (EMNLP), ACM (2015)", + " ", + "\\bibitem{bird2004nltk}", + "Bird, S., Loper, E.:", + "\\newblock {NLTK: the natural language toolkit}.", + "\\newblock In: Proceedings of the ACL 2004 on Interactive poster and", + " demonstration sessions, Association for Computational Linguistics (2004)", + " ", + "\\bibitem{kumar2018false}", + "Kumar, S., Shah, N.:", + "\\newblock False information on web and social media: A survey.", + "\\newblock arXiv preprint arXiv:1804.08559 (2018)", + " ", + "\\bibitem{Everett2016Automated}", + "Everett, R.M., Nurse, J.R.C., Erola, A.:", + "\\newblock The anatomy of online deception: What makes automated text", + " convincing?", + "\\newblock In: Proceedings of the 31st Annual ACM Symposium on Applied", + " Computing. SAC '16, ACM (2016)", + " ", + "\\end{thebibliography}", + " ", + " ", + " ", + "\\section*{Appendix}", + " ", + "We present basic demographics of our MTurk study and the comparative study with experienced users in Table~\\ref{table:amt_pop}.", + " ", + "\\begin{table}", + "\\caption{User study statistics.}", + "\\begin{center}", + " \\begin{tabular}{ | l | c | c | }", + " \\hline", + " Quality & Mechanical Turk users & Experienced users\\\\", + " \\hline", + " Native English Speaker & Yes (20) & Yes (1) No (19) \\\\", + " Fluent in English & Yes (20) & Yes (20) \\\\", + " Age & 21-40 (17) 41-60 (3) & 21-25 (8) 26-30 (7) 31-35 (4) 41-45 (1)\\\\", + " Gender & Male (14) Female (6) & Male (17) Female (3)\\\\", + " Highest Education & High School (10) Bachelor (10) & Bachelor (9) Master (6) Ph.D. (5) \\\\", + " \\hline", + " \\end{tabular}", + " \\label{table:amt_pop}", + "\\end{center}", + "\\end{table}", + " ", + " ", + "Table~\\ref{table:openNMT-py_commands} shows a listing of the openNMT-py commands we used to create our NMT model and to generate fake reviews.", + " ", + "\\begin{table}[t]", + "\\caption{Listing of used openNMT-py commands.}", + "\\begin{center}", + " \\begin{tabular}{ | l | l | }", + " \\hline", + " Phase & Bash command \\\\", + " \\hline", + " Preprocessing & \\begin{lstlisting}[language=bash]", + "python preprocess.py -train_src context-train.txt", + "-train_tgt reviews-train.txt -valid_src context-val.txt", + "-valid_tgt reviews-val.txt -save_data model", + "-lower -tgt_words_min_frequency 10", + "\\end{lstlisting}", + " \\\\ & \\\\", + " Training & \\begin{lstlisting}[language=bash]", + "python train.py -data model -save_model model -epochs 8", + "-gpuid 0 -learning_rate_decay 0.5 -optim adam", + "-learning_rate 0.001 -start_decay_at 3\\end{lstlisting}", + " \\\\ & \\\\", + " Generation & \\begin{lstlisting}[language=bash]", + "python translate.py -model model_acc_35.54_ppl_25.68_e8.pt", + "-src context-tst.txt -output pred-e8.txt -replace_unk", + "-verbose -max_length 50 -gpu 0", + " \\end{lstlisting} \\\\", + " \\hline", + " \\end{tabular}", + " \\label{table:openNMT-py_commands}", + "\\end{center}", + "\\end{table}", + " ", + " ", + "Table~\\ref{table:MTurk_sub} shows the classification performance of Amazon Mechanical Turkers, separated across different categories of NMT-Fake reviews. The category with best performance ($b=0.3, \\lambda=-5$) is denoted as NMT-Fake*.", + " ", + "\\begin{table}[b]", + "\\caption{MTurk study subclass classification reports. Classes are imbalanced in ratio 1:6. Random predictions are $p_\\mathrm{human} = 86\\%$ and $p_\\mathrm{machine} = 14\\%$, with $r_\\mathrm{human} = r_\\mathrm{machine} = 50\\%$. Class-averaged F-scores for random predictions are $42\\%$.}", + "\\begin{center}", + " \\begin{tabular}{ | c || c |c |c | c | }", + " \\hline", + " $(b=0.3, \\lambda = -3)$ & Precision & Recall & F-score & Support \\\\ \\hline", + " Human & 89\\% & 63\\% & 73\\% & 994\\\\", + " NMT-Fake & 15\\% & 45\\% & 22\\% & 146 \\\\", + " \\hline", + " \\hline", + " $(b=0.3, \\lambda = -5)$ & Precision & Recall & F-score & Support \\\\ \\hline", + " Human & 86\\% & 63\\% & 73\\% & 994\\\\", + " NMT-Fake* & 16\\% & 40\\% & 23\\% & 171 \\\\", + " \\hline", + " \\hline", + " $(b=0.5, \\lambda = -4)$ & Precision & Recall & F-score & Support \\\\ \\hline", + " Human & 88\\% & 63\\% & 73\\% & 994\\\\", + " NMT-Fake & 21\\% & 55\\% & 30\\% & 181 \\\\", + " \\hline", + " \\hline", + " $(b=0.7, \\lambda = -3)$ & Precision & Recall & F-score & Support \\\\ \\hline", + " Human & 88\\% & 63\\% & 73\\% & 994\\\\", + " NMT-Fake & 19\\% & 50\\% & 27\\% & 170 \\\\", + " \\hline", + " \\hline", + " $(b=0.7, \\lambda = -5)$ & Precision & Recall & F-score & Support \\\\ \\hline", + " Human & 89\\% & 63\\% & 74\\% & 994\\\\", + " NMT-Fake & 21\\% & 57\\% & 31\\% & 174 \\\\", + " \\hline", + " \\hline", + " $(b=0.9, \\lambda = -4)$ & Precision & Recall & F-score & Support \\\\ \\hline", + " Human & 88\\% & 63\\% & 73\\% & 994\\\\", + " NMT-Fake & 18\\% & 50\\% & 27\\% & 164 \\\\", + " \\hline", + " \\end{tabular}", + " \\label{table:MTurk_sub}", + "\\end{center}", + "\\end{table}", + " ", + "Figure~\\ref{fig:screenshot} shows screenshots of the first two pages of our user study with experienced participants.", + " ", + "\\begin{figure}[ht]", + "\\centering", + "\\includegraphics[width=1.\\columnwidth]{figures/screenshot_7-3.png}", + "\\caption{", + "Screenshots of the first two pages in the user study. Example 1 is a NMT-Fake* review, the rest are human-written.", + "}", + "\\label{fig:screenshot}", + "\\end{figure}", + " ", + "Table~\\ref{table:features_adaboost} shows the features used to detect NMT-Fake reviews using the AdaBoost classifier.", + " ", + "\\begin{table}", + "\\caption{Features used in NMT-Fake review detector.}", + "\\begin{center}", + " \\begin{tabular}{ | l | c | }", + " \\hline", + " Feature type & Number of features \\\\ \\hline", + " \\hline", + " Readability features & 13 \\\\ \\hline", + " Unique POS tags & $~20$ \\\\ \\hline", + " Word unigrams & 22,831 \\\\ \\hline", + " 1/2/3/4-grams of simple part-of-speech tags & 54,240 \\\\ \\hline", + " 1/2/3-grams of detailed part-of-speech tags & 112,944 \\\\ \\hline", + " 1/2/3-grams of syntactic dependency tags & 93,195 \\\\ \\hline", + " \\end{tabular}", + " \\label{table:features_adaboost}", + "\\end{center}", + "\\end{table}", + " ", + "\\end{document}", + "" + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0039/instruction.md b/qasper-0039/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..91c401393422850e75ef3f79285e0534e3042a58 --- /dev/null +++ b/qasper-0039/instruction.md @@ -0,0 +1,84 @@ +Name of Paper: Saliency Maps Generation for Automatic Text Summarization + +Question: Which baselines did they compare? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "The Task and the Model", + "Dataset and Training Task", + "The Model", + "Obtained Summaries", + "Layer-Wise Relevance Propagation", + "Mathematical Description", + "Generation of the Saliency Maps", + "Experimental results", + "First Observations", + "Validating the Attributions", + "Conclusion" + ], + "paragraphs": [ + [ + "Ever since the LIME algorithm BIBREF0 , \"explanation\" techniques focusing on finding the importance of input features in regard of a specific prediction have soared and we now have many ways of finding saliency maps (also called heat-maps because of the way we like to visualize them). We are interested in this paper by the use of such a technique in an extreme task that highlights questions about the validity and evaluation of the approach. We would like to first set the vocabulary we will use. We agree that saliency maps are not explanations in themselves and that they are more similar to attribution, which is only one part of the human explanation process BIBREF1 . We will prefer to call this importance mapping of the input an attribution rather than an explanation. We will talk about the importance of the input relevance score in regard to the model's computation and not make allusion to any human understanding of the model as a result.", + "There exist multiple ways to generate saliency maps over the input for non-linear classifiers BIBREF2 , BIBREF3 , BIBREF4 . We refer the reader to BIBREF5 for a survey of explainable AI in general. We use in this paper Layer-Wise Relevance Propagation (LRP) BIBREF2 which aims at redistributing the value of the classifying function on the input to obtain the importance attribution. It was first created to \u201cexplain\" the classification of neural networks on image recognition tasks. It was later successfully applied to text using convolutional neural networks (CNN) BIBREF6 and then Long-Short Term Memory (LSTM) networks for sentiment analysis BIBREF7 .", + "Our goal in this paper is to test the limits of the use of such a technique for more complex tasks, where the notion of input importance might not be as simple as in topic classification or sentiment analysis. We changed from a classification task to a generative task and chose a more complex one than text translation (in which we can easily find a word to word correspondence/importance between input and output). We chose text summarization. We consider abstractive and informative text summarization, meaning that we write a summary \u201cin our own words\" and retain the important information of the original text. We refer the reader to BIBREF8 for more details on the task and the different variants that exist. Since the success of deep sequence-to-sequence models for text translation BIBREF9 , the same approaches have been applied to text summarization tasks BIBREF10 , BIBREF11 , BIBREF12 which use architectures on which we can apply LRP.", + "We obtain one saliency map for each word in the generated summaries, supposed to represent the use of the input features for each element of the output sequence. We observe that all the saliency maps for a text are nearly identical and decorrelated with the attention distribution. We propose a way to check their validity by creating what could be seen as a counterfactual experiment from a synthesis of the saliency maps, using the same technique as in Arras et al. Arras2017. We show that in some but not all cases they help identify the important input features and that we need to rigorously check importance attributions before trusting them, regardless of whether or not the mapping \u201cmakes sense\" to us. We finally argue that in the process of identifying the important input features, verifying the saliency maps is as important as the generation step, if not more." + ], + [ + "We present in this section the baseline model from See et al. See2017 trained on the CNN/Daily Mail dataset. We reproduce the results from See et al. See2017 to then apply LRP on it." + ], + [ + "The CNN/Daily mail dataset BIBREF12 is a text summarization dataset adapted from the Deepmind question-answering dataset BIBREF13 . It contains around three hundred thousand news articles coupled with summaries of about three sentences. These summaries are in fact \u201chighlights\" of the articles provided by the media themselves. Articles have an average length of 780 words and the summaries of 50 words. We had 287 000 training pairs and 11 500 test pairs. Similarly to See et al. See2017, we limit during training and prediction the input text to 400 words and generate summaries of 200 words. We pad the shorter texts using an UNKNOWN token and truncate the longer texts. We embed the texts and summaries using a vocabulary of size 50 000, thus recreating the same parameters as See et al. See2017." + ], + [ + "The baseline model is a deep sequence-to-sequence encoder/decoder model with attention. The encoder is a bidirectional Long-Short Term Memory(LSTM) cell BIBREF14 and the decoder a single LSTM cell with attention mechanism. The attention mechanism is computed as in BIBREF9 and we use a greedy search for decoding. We train end-to-end including the words embeddings. The embedding size used is of 128 and the hidden state size of the LSTM cells is of 254." + ], + [ + "We train the 21 350 992 parameters of the network for about 60 epochs until we achieve results that are qualitatively equivalent to the results of See et al. See2017. We obtain summaries that are broadly relevant to the text but do not match the target summaries very well. We observe the same problems such as wrong reproduction of factual details, replacing rare words with more common alternatives or repeating non-sense after the third sentence. We can see in Figure 1 an example of summary obtained compared to the target one.", + "The \u201csummaries\" we generate are far from being valid summaries of the information in the texts but are sufficient to look at the attribution that LRP will give us. They pick up the general subject of the original text." + ], + [ + "We present in this section the Layer-Wise Relevance Propagation (LRP) BIBREF2 technique that we used to attribute importance to the input features, together with how we adapted it to our model and how we generated the saliency maps. LRP redistributes the output of the model from the output layer to the input by transmitting information backwards through the layers. We call this propagated backwards importance the relevance. LRP has the particularity to attribute negative and positive relevance: a positive relevance is supposed to represent evidence that led to the classifier's result while negative relevance represents evidence that participated negatively in the prediction." + ], + [ + "We initialize the relevance of the output layer to the value of the predicted class before softmax and we then describe locally the propagation backwards of the relevance from layer to layer. For normal neural network layers we use the form of LRP with epsilon stabilizer BIBREF2 . We write down $R_{i\\leftarrow j}^{(l, l+1)}$ the relevance received by the neuron $i$ of layer $l$ from the neuron $j$ of layer $l+1$ : ", + "$$\\begin{split}\n\nR_{i\\leftarrow j}^{(l, l+1)} &= \\dfrac{w_{i\\rightarrow j}^{l,l+1}\\textbf {z}^l_i + \\dfrac{\\epsilon \\textrm { sign}(\\textbf {z}^{l+1}_j) + \\textbf {b}^{l+1}_j}{D_l}}{\\textbf {z}^{l+1}_j + \\epsilon * \\textrm { sign}(\\textbf {z}^{l+1}_j)} * R_j^{l+1} \\\\\n\\end{split}$$ (Eq. 7) ", + "where $w_{i\\rightarrow j}^{l,l+1}$ is the network's weight parameter set during training, $\\textbf {b}^{l+1}_j$ is the bias for neuron $j$ of layer $l+1$ , $\\textbf {z}^{l}_i$ is the activation of neuron $i$ on layer $l$ , $\\epsilon $ is the stabilizing term set to 0.00001 and $D_l$ is the dimension of the $l$ -th layer.", + "The relevance of a neuron is then computed as the sum of the relevance he received from the above layer(s).", + "For LSTM cells we use the method from Arras et al.Arras2017 to solve the problem posed by the element-wise multiplications of vectors. Arras et al. noted that when such computation happened inside an LSTM cell, it always involved a \u201cgate\" vector and another vector containing information. The gate vector containing only value between 0 and 1 is essentially filtering the second vector to allow the passing of \u201crelevant\" information. Considering this, when we propagate relevance through an element-wise multiplication operation, we give all the upper-layer's relevance to the \u201cinformation\" vector and none to the \u201cgate\" vector." + ], + [ + "We use the same method to transmit relevance through the attention mechanism back to the encoder because Bahdanau's attention BIBREF9 uses element-wise multiplications as well. We depict in Figure 2 the transmission end-to-end from the output layer to the input through the decoder, attention mechanism and then the bidirectional encoder. We then sum up the relevance on the word embedding to get the token's relevance as Arras et al. Arras2017.", + "The way we generate saliency maps differs a bit from the usual context in which LRP is used as we essentially don't have one classification, but 200 (one for each word in the summary). We generate a relevance attribution for the 50 first words of the generated summary as after this point they often repeat themselves.", + "This means that for each text we obtain 50 different saliency maps, each one supposed to represent the relevance of the input for a specific generated word in the summary." + ], + [ + "In this section, we present our results from extracting attributions from the sequence-to-sequence model trained for abstractive text summarization. We first have to discuss the difference between the 50 different saliency maps we obtain and then we propose a protocol to validate the mappings." + ], + [ + "The first observation that is made is that for one text, the 50 saliency maps are almost identical. Indeed each mapping highlights mainly the same input words with only slight variations of importance. We can see in Figure 3 an example of two nearly identical attributions for two distant and unrelated words of the summary. The saliency map generated using LRP is also uncorrelated with the attention distribution that participated in the generation of the output word. The attention distribution changes drastically between the words in the generated summary while not impacting significantly the attribution over the input text. We deleted in an experiment the relevance propagated through the attention mechanism to the encoder and didn't observe much changes in the saliency map.", + "It can be seen as evidence that using the attention distribution as an \u201cexplanation\" of the prediction can be misleading. It is not the only information received by the decoder and the importance it \u201callocates\" to this attention state might be very low. What seems to happen in this application is that most of the information used is transmitted from the encoder to the decoder and the attention mechanism at each decoding step just changes marginally how it is used. Quantifying the difference between attention distribution and saliency map across multiple tasks is a possible future work.", + "The second observation we can make is that the saliency map doesn't seem to highlight the right things in the input for the summary it generates. The saliency maps on Figure 3 correspond to the summary from Figure 1 , and we don't see the word \u201cvideo\" highlighted in the input text, which seems to be important for the output.", + "This allows us to question how good the saliency maps are in the sense that we question how well they actually represent the network's use of the input features. We will call that truthfulness of the attribution in regard to the computation, meaning that an attribution is truthful in regard to the computation if it actually highlights the important input features that the network attended to during prediction. We proceed to measure the truthfulness of the attributions by validating them quantitatively." + ], + [ + "We propose to validate the saliency maps in a similar way as Arras et al. Arras2017 by incrementally deleting \u201cimportant\" words from the input text and observe the change in the resulting generated summaries.", + "We first define what \u201cimportant\" (and \u201cunimportant\") input words mean across the 50 saliency maps per texts. Relevance transmitted by LRP being positive or negative, we average the absolute value of the relevance across the saliency maps to obtain one ranking of the most \u201crelevant\" words. The idea is that input words with negative relevance have an impact on the resulting generated word, even if it is not participating positively, while a word with a relevance close to zero should not be important at all. We did however also try with different methods, like averaging the raw relevance or averaging a scaled absolute value where negative relevance is scaled down by a constant factor. The absolute value average seemed to deliver the best results.", + "We delete incrementally the important words (words with the highest average) in the input and compared it to the control experiment that consists of deleting the least important word and compare the degradation of the resulting summaries. We obtain mitigated results: for some texts, we observe a quick degradation when deleting important words which are not observed when deleting unimportant words (see Figure 4 ), but for other test examples we don't observe a significant difference between the two settings (see Figure 5 ).", + "One might argue that the second summary in Figure 5 is better than the first one as it makes better sentences but as the model generates inaccurate summaries, we do not wish to make such a statement.", + "This however allows us to say that the attribution generated for the text at the origin of the summaries in Figure 4 are truthful in regard to the network's computation and we may use it for further studies of the example, whereas for the text at the origin of Figure 5 we shouldn't draw any further conclusions from the attribution generated.", + "One interesting point is that one saliency map didn't look \u201cbetter\" than the other, meaning that there is no apparent way of determining their truthfulness in regard of the computation without doing a quantitative validation. This brings us to believe that even in simpler tasks, the saliency maps might make sense to us (for example highlighting the animal in an image classification task), without actually representing what the network really attended too, or in what way.", + "We defined without saying it the counterfactual case in our experiment: \u201cWould the important words in the input be deleted, we would have a different summary\". Such counterfactuals are however more difficult to define for image classification for example, where it could be applying a mask over an image, or just filtering a colour or a pattern. We believe that defining a counterfactual and testing it allows us to measure and evaluate the truthfulness of the attributions and thus weight how much we can trust them." + ], + [ + "In this work, we have implemented and applied LRP to a sequence-to-sequence model trained on a more complex task than usual: text summarization. We used previous work to solve the difficulties posed by LRP in LSTM cells and adapted the same technique for Bahdanau et al. Bahdanau2014 attention mechanism.", + "We observed a peculiar behaviour of the saliency maps for the words in the output summary: they are almost all identical and seem uncorrelated with the attention distribution. We then proceeded to validate our attributions by averaging the absolute value of the relevance across the saliency maps. We obtain a ranking of the word from the most important to the least important and proceeded to delete one or another.", + "We showed that in some cases the saliency maps are truthful to the network's computation, meaning that they do highlight the input features that the network focused on. But we also showed that in some cases the saliency maps seem to not capture the important input features. This brought us to discuss the fact that these attributions are not sufficient by themselves, and that we need to define the counter-factual case and test it to measure how truthful the saliency maps are.", + "Future work would look into the saliency maps generated by applying LRP to pointer-generator networks and compare to our current results as well as mathematically justifying the average that we did when validating our saliency maps. Some additional work is also needed on the validation of the saliency maps with counterfactual tests. The exploitation and evaluation of saliency map are a very important step and should not be overlooked." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0048/instruction.md b/qasper-0048/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..843ef5163268881b9093e85a2a5081f879a0b13a --- /dev/null +++ b/qasper-0048/instruction.md @@ -0,0 +1,56 @@ +Name of Paper: Is there Gender bias and stereotype in Portuguese Word Embeddings? + +Question: Does this paper target European or Brazilian Portuguese? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Portuguese Embedding", + "Proposed Approach", + "Experiments", + "Final Remarks" + ], + "paragraphs": [ + [ + "Recently, the transformative potential of machine learning (ML) has propelled ML into the forefront of mainstream media. In Brazil, the use of such technique has been widely diffused gaining more space. Thus, it is used to search for patterns, regularities or even concepts expressed in data sets BIBREF0 , and can be applied as a form of aid in several areas of everyday life.", + "Among the different definitions, ML can be seen as the ability to improve performance in accomplishing a task through the experience BIBREF1 . Thus, BIBREF2 presents this as a method of inferences of functions or hypotheses capable of solving a problem algorithmically from data representing instances of the problem. This is an important way to solve different types of problems that permeate computer science and other areas.", + "One of the main uses of ML is in text processing, where the analysis of the content the entry point for various learning algorithms. However, the use of this content can represent the insertion of different types of bias in training and may vary with the context worked. This work aims to analyze and remove gender stereotypes from word embedding in Portuguese, analogous to what was done in BIBREF3 for the English language. Hence, we propose to employ a public word2vec model pre-trained to analyze gender bias in the Portuguese language, quantifying biases present in the model so that it is possible to reduce the spreading of sexism of such models. There is also a stage of bias reducing over the results obtained in the model, where it is sought to analyze the effects of the application of gender distinction reduction techniques.", + "This paper is organized as follows: Section SECREF2 discusses related works. Section SECREF3 presents the Portuguese word2vec embeddings model used in this paper and Section SECREF4 proposes our method. Section SECREF5 presents experimental results, whose purpose is to verify results of a de-bias algorithm application in Portuguese embeddings word2vec model and a short discussion about it. Section SECREF6 brings our concluding remarks." + ], + [ + "There is a wide range of techniques that provide interesting results in the context of ML algorithms geared to the classification of data without discrimination; these techniques range from the pre-processing of data BIBREF4 to the use of bias removal techniques BIBREF5 in fact. Approaches linked to the data pre-processing step usually consist of methods based on improving the quality of the dataset after which the usual classification tools can be used to train a classifier. So, it starts from a baseline already stipulated by the execution of itself. On the other side of the spectrum, there are Unsupervised and semi-supervised learning techniques, that are attractive because they do not imply the cost of corpus annotation BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 .", + "The bias reduction is studied as a way to reduce discrimination through classification through different approaches BIBREF10 BIBREF11 . In BIBREF12 the authors propose to specify, implement, and evaluate the \u201cfairness-aware\" ML interface called themis-ml. In this interface, the main idea is to pick up a data set from a modified dataset. Themis-ml implements two methods for training fairness-aware models. The tool relies on two methods to make agnostic model type predictions: Reject Option Classification and Discrimination-Aware Ensemble Classification, these procedures being used to post-process predictions in a way that reduces potentially discriminatory predictions. According to the authors, it is possible to perceive the potential use of the method as a means of reducing bias in the use of ML algorithms.", + "In BIBREF3 , the authors propose a method to hardly reduce bias in English word embeddings collected from Google News. Using word2vec, they performed a geometric analysis of gender direction of the bias contained in the data. Using this property with the generation of gender-neutral analogies, a methodology was provided for modifying an embedding to remove gender stereotypes. Some metrics were defined to quantify both direct and indirect gender biases in embeddings and to develop algorithms to reduce bias in some embedding. Hence, the authors show that embeddings can be used in applications without amplifying gender bias." + ], + [ + "In BIBREF13 , the quality of the representation of words through vectors in several models is discussed. According to the authors, the ability to train high-quality models using simplified architectures is useful in models composed of predictive methods that try to predict neighboring words with one or more context words, such as Word2Vec. Word embeddings have been used to provide meaningful representations for words in an efficient way.", + "In BIBREF14 , several word embedding models trained in a large Portuguese corpus are evaluated. Within the Word2Vec model, two training strategies were used. In the first, namely Skip-Gram, the model is given the word and attempts to predict its neighboring words. The second, Continuous Bag-of-Words (CBOW), the model is given the sequence of words without the middle one and attempts to predict this omitted word. The latter was chosen for application in the present proposal.", + "The authors of BIBREF14 claim to have collected a large corpus from several sources to obtain a multi-genre corpus representative of the Portuguese language. Hence, it comprehensively covers different expressions of the language, making it possible to analyze gender bias and stereotype in Portuguese word embeddings. The dataset used was tokenized and normalized by the authors to reduce the corpus vocabulary size, under the premise that vocabulary reduction provides more representative vectors." + ], + [ + "Some linguists point out that the female gender is, in Portuguese, a particularization of the masculine. In this way the only gender mark is the feminine, the others being considered without gender (including names considered masculine). In BIBREF15 the gender representation in Portuguese is associated with a set of phenomena, not only from a linguistic perspective but also from a socio-cultural perspective. Since most of the termination of words (e.g., advogada and advogado) are used to indicate to whom the expression refers, stereotypes can be explained through communication. This implies the presence of biases when dealing with terms such as those referring to professions.", + "Figure FIGREF1 illustrates the approach proposed in this work. First, using a list of professions relating the identification of female and male who perform it as a parameter, we evaluate the accuracy of similarity generated by the embeddings. Then, getting the biased results, we apply the De-bias algorithm BIBREF3 aiming to reduce sexist analogies previous generated. Thus, all the results are analyzed by comparing the accuracies.", + "Using the word2vec model available in a public repository BIBREF14 , the proposal involves the analysis of the most similar analogies generated before and after the application of the BIBREF3 . The work is focused on the analysis of gender bias associated with professions in word embeddings. So therefore into the evaluation of the accuracy of the associations generated, aiming at achieving results as good as possible without prejudicing the evaluation metrics.", + "Algorithm SECREF4 describes the method performed during the evaluation of the gender bias presence. In this method we try to evaluate the accuracy of the analogies generated through the model, that is, to verify the cases of association matching generated between the words.", + "[!htb] Model Evaluation [1]", + "w2v_evaluate INLINEFORM0 open_model( INLINEFORM1 ) count = 0 INLINEFORM2 in INLINEFORM3 read list of tuples x = model.most_similar(positive=[`ela', male], negative=[`ele'])", + "x = female count += 1 accuracy = count/size(profession_pairs) return accuracy" + ], + [ + "The purpose of this section is to perform different analysis concerning bias in word2vec models with Portuguese embeddings. The Continuous Bag-of-Words model used was provided by BIBREF14 (described in Section SECREF3 ). For these experiments, we use a model containing 934966 words of dimension 300 per vector representation. To realize the experiments, a list containing fifty professions labels for female and male was used as the parameter of similarity comparison.", + "Using the python library gensim, we evaluate the extreme analogies generated when comparing vectors like: INLINEFORM0 , where INLINEFORM1 represents the item from professions list and INLINEFORM2 the expected association. The most similarity function finds the top-N most similar entities, computing cosine similarity between a simple mean of the projection weight vectors of the given docs. Figure FIGREF4 presents the most extreme analogies results obtained from the model using these comparisons.", + "Applying the Algorithm SECREF4 , we check the accuracy obtained with the similarity function before and after the application of the de-bias method. Table TABREF3 presents the corresponding results. In cases like the analogy of `gar\u00e7onete' to `stripper' (Figure FIGREF4 , line 8), it is possible to observe that the relationship stipulated between terms with sexual connotation and females is closer than between females and professions. While in the male model, even in cases of non-compliance, the closest analogy remains in the professional environment.", + "Using a confidence factor of 99%, when comparing the correctness levels of the model with and without the reduction of bias, the prediction of the model with bias is significantly better. Different authors BIBREF16 BIBREF17 show that the removal of bias in models produces a negative impact on the quality of the model. On the other hand, it is observed that even with a better hit rate the correctness rate in the prediction of related terms is still low." + ], + [ + "This paper presents an analysis of the presence of gender bias in Portuguese word embeddings. Even though it is a work in progress, the proposal showed promising results in analyzing predicting models.", + "A possible extension of the work involves deepening the analysis of the results obtained, seeking to achieve higher accuracy rates and fairer models to be used in machine learning techniques. Thus, these studies can involve tests with different methods of pre-processing the data to the use of different models, as well as other factors that may influence the results generated. This deepening is necessary since the model's accuracy is not high.", + "To conclude, we believe that the presence of gender bias and stereotypes in the Portuguese language is found in different spheres of language, and it is important to study ways of mitigating different types of discrimination. As such, it can be easily applied to analyze racists bias into the language, such as different types of preconceptions." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0052/instruction.md b/qasper-0052/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..7aef9c13a78d7646c23e72578a0497f155c80b0a --- /dev/null +++ b/qasper-0052/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Citation Data of Czech Apex Courts + +Question: How is quality of the citation measured? \ No newline at end of file diff --git a/qasper-0055/instruction.md b/qasper-0055/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..53f2dbde638c13daeda616575519f75fef30cae0 --- /dev/null +++ b/qasper-0055/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: LAXARY: A Trustworthy Explainable Twitter Analysis Model for Post-Traumatic Stress Disorder Assessment + +Question: Do the authors mention any possible confounds in this study? \ No newline at end of file diff --git a/qasper-0070/instruction.md b/qasper-0070/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..63cc4e20baa10a80c54b2e1be931e4fa083fa11a --- /dev/null +++ b/qasper-0070/instruction.md @@ -0,0 +1,121 @@ +Name of Paper: Spoken Language Identification using ConvNets + +Question: Is the performance compared against a baseline model? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Proposed Method ::: Motivations", + "Proposed Method ::: Description of Features", + "Proposed Method ::: Model Description", + "Proposed Method ::: Model Details: 1D ConvNet", + "Proposed Method ::: Model Details: 1D ConvNet ::: Hyperparameter Optimization:", + "Proposed Method ::: Model Details: 2D ConvNet with Attention and bi-directional GRU", + "Proposed Method ::: Model Details: 2D ConvNet with Attention and bi-directional GRU ::: ", + "Proposed Method ::: Model Details: 2D ConvNet with Attention and bi-directional GRU ::: Hyperparameter Optimization:", + "Proposed Method ::: Model details: 2D-ConvNet", + "Proposed Method ::: Dataset", + "Results and Discussion", + "Results and Discussion ::: Misclassification", + "Results and Discussion ::: Future Scope", + "Conclusion" + ], + "paragraphs": [ + [ + "Language Identification (LI) is a problem which involves classifying the language being spoken by a speaker. LI systems can be used in call centers to route international calls to an operator who is fluent in that identified language BIBREF0. In speech-based assistants, LI acts as the first step which chooses the corresponding grammar from a list of available languages for its further semantic analysis BIBREF1. It can also be used in multi-lingual voice-controlled information retrieval systems, for example, Apple Siri and Amazon Alexa.", + "Over the years, studies have utilized many prosodic and acoustic features to construct machine learning models for LI systems BIBREF2. Every language is composed of phonemes, which are distinct unit of sounds in that language, such as b of black and g of green. Several prosodic and acoustic features are based on phonemes, which become the underlying features on whom the performance of the statistical model depends BIBREF3, BIBREF4. If two languages have many overlapping phonemes, then identifying them becomes a challenging task for a classifier. For example, the word cat in English, kat in Dutch, katze in German have different consonants but when used in a speech they all would sound quite similar.", + "Due to such drawbacks several studies have switched over to using Deep Neural Networks (DNNs) to harness their novel auto-extraction techniques BIBREF1, BIBREF5. This work follows an implicit approach for identifying six languages with overlapping phonemes on the VoxForge BIBREF6 dataset and achieves 95.4% overall accuracy.", + "In previous studies BIBREF1, BIBREF7, BIBREF5, authors use log-Mel spectrum of a raw audio as inputs to their models. One of our contributions is to enhance the performance of this approach by utilising recent techniques like Mixup augmentation of inputs and exploring the effectiveness of Attention mechanism in enhancing performance of neural network. As log-Mel spectrum needs to be computed for each raw audio input and processing time for generating log-Mel spectrum increases linearly with length of audio, this acts as a bottleneck for these models. Hence, we propose the use of raw audio waveforms as inputs to deep neural network which boosts performance by avoiding additional overhead of computing log-Mel spectrum for each audio. Our 1D-ConvNet architecture auto-extracts and classifies features from this raw audio input.", + "The structure of the work is as follows. In Section 2 we discuss about the previous related studies in this field. The model architecture for both the raw waveforms and log-Mel spectrogram images is discussed in Section 3 along with the a discussion on hyperparameter space exploration. In Section 4 we present the experimental results. Finally, in Section 5 we discuss the conclusions drawn from the experiment and future work." + ], + [ + "Extraction of language dependent features like prosody and phonemes was a popular approach to classify spoken languages BIBREF8, BIBREF9, BIBREF10. Following their success in speaker verification systems, i-vectors have also been used as features in various classification networks. These approaches required significant domain knowledge BIBREF11, BIBREF9. Nowadays most of the attempts on spoken language identification rely on neural networks for meaningful feature extraction and classification BIBREF12, BIBREF13.", + "Revay et al. BIBREF5 used the ResNet50 BIBREF14 architecture for classifying languages by generating the log-Mel spectra of each raw audio. The model uses a cyclic learning rate where learning rate increases and then decreases linearly. Maximum learning rate for a cycle is set by finding the optimal learning rate using fastai BIBREF15 library. The model classified six languages \u2013 English, French, Spanish, Russian, Italian and German \u2013 and achieving an accuracy of 89.0%.", + "Gazeau et al. BIBREF16 in his research showed how Neural Networks, Support Vector Machine and Hidden Markov Model (HMM) can be used to identify French, English, Spanish and German. Dataset was prepared using voice samples from Youtube News BIBREF17and VoxForge BIBREF6 datasets. Hidden Markov models convert speech into a sequence of vectors, was used to capture temporal features in speech. HMMs trained on VoxForge BIBREF6 dataset performed best in comparison to other models proposed by him on same VoxForge dataset. They reported an accuracy of 70.0%.", + "Bartz et al. BIBREF1 proposed two different hybrid Convolutional Recurrent Neural Networks for language identification. They proposed a new architecture for extracting spatial features from log-Mel spectra of raw audio using CNNs and then using RNNs for capturing temporal features to identify the language. This model achieved an accuracy of 91.0% on Youtube News Dataset BIBREF17. In their second architecture they used the Inception-v3 BIBREF18 architecture to extract spatial features which were then used as input for bi-directional LSTMs to predict the language accurately. This model achieved an accuracy of 96.0% on four languages which were English, German, French and Spanish. They also trained their CNN model (obtained after removing RNN from CRNN model) and the Inception-v3 on their dataset. However they were not able to achieve better results achieving and reported 90% and 95% accuracies, respectively.", + "Kumar et al. BIBREF0 used Mel-frequency cepstral coefficients (MFCC), Perceptual linear prediction coefficients (PLP), Bark Frequency Cepstral Coefficients (BFCC) and Revised Perceptual Linear Prediction Coefficients (RPLP) as features for language identification. BFCC and RPLP are hybrid features derived using MFCC and PLP. They used two different models based on Vector Quantization (VQ) with Dynamic Time Warping (DTW) and Gaussian Mixture Model (GMM) for classification. These classification models were trained with different features. The authors were able to show that these models worked better with hybrid features (BFCC and RPLP) as compared to conventional features (MFCC and PLP). GMM combined with RPLP features gave the most promising results and achieved an accuracy of 88.8% on ten languages. They designed their own dataset comprising of ten languages being Dutch, English, French, German, Italian, Russian, Spanish, Hindi, Telegu, and Bengali.", + "Montavon BIBREF7 generated Mel spectrogram as features for a time-delay neural network (TDNN). This network had two-dimensional convolutional layers for feature extraction. An elaborate analysis of how deep architectures outperform their shallow counterparts is presented in this reseacrch. The difficulties in classifying perceptually similar languages like German and English were also put forward in this work. It is mentioned that the proposed approach is less robust to new speakers present in the test dataset. This method was able to achieve an accuracy of 91.2% on dataset comprising of 3 languages \u2013 English, French and German.", + "In Table TABREF1, we summarize the quantitative results of the above previous studies. It includes the model basis, feature description, languages classified and the used dataset along with accuracy obtained. The table also lists the overall results of our proposed models (at the top). The languages used by various authors along with their acronyms are English (En), Spanish (Es), French (Fr), German (De), Russian (Ru), Italian (It), Bengali (Ben), Hindi (Hi) and Telegu (Tel)." + ], + [ + "Several state-of-the-art results on various audio classification tasks have been obtained by using log-Mel spectrograms of raw audio, as features BIBREF19. Convolutional Neural Networks have demonstrated an excellent performance gain in classification of these features BIBREF20, BIBREF21 against other machine learning techniques. It has been shown that using attention layers with ConvNets further enhanced their performance BIBREF22. This motivated us to develop a CNN-based architecture with attention since this approach hasn\u2019t been applied to the task of language identification before.", + "Recently, using raw audio waveform as features to neural networks has become a popular approach in audio classification BIBREF23, BIBREF22. Raw waveforms have several artifacts which are not effectively captured by various conventional feature extraction techniques like Mel Frequency Cepstral Coefficients (MFCC), Constant Q Transform (CQT), Fast Fourier Transform (FFT), etc.", + "Audio files are a sequence of spoken words, hence they have temporal features too.A CNN is better at capturing spatial features only and RNNs are better at capturing temporal features as demonstrated by Bartz et al. BIBREF1 using audio files. Therefore, we combined both of these to make a CRNN model.", + "We propose three types of models to tackle the problem with different approaches, discussed as follows." + ], + [ + "As an average human's voice is around 300 Hz and according to Nyquist-Shannon sampling theorem all the useful frequencies (0-300 Hz) are preserved with sampling at 8 kHz, therefore, we sampled raw audio files from all six languages at 8 kHz", + "The average length of audio files in this dataset was about 10.4 seconds and standard deviation was 2.3 seconds. For our experiments, the audio length was set to 10 seconds. If the audio files were shorter than 10 second, then the data was repeated and concatenated. If audio files were longer, then the data was truncated." + ], + [ + "We applied the following design principles to all our models:", + "Every convolutional layer is always followed by an appropriate max pooling layer. This helps in containing the explosion of parameters and keeps the model small and nimble.", + "Convolutional blocks are defined as an individual block with multiple pairs of one convolutional layer and one max pooling layer. Each convolutional block is preceded or succeded by a convolutional layer.", + "Batch Normalization and Rectified linear unit activations were applied after each convolutional layer. Batch Normalization helps speed up convergence during training of a neural network.", + "Model ends with a dense layer which acts the final output layer." + ], + [ + "As the sampling rate is 8 kHz and audio length is 10 s, hence the input is raw audio to the models with input size of (batch size, 1, 80000). In Table TABREF10, we present a detailed layer-by-layer illustration of the model along with its hyperparameter.", + "-10pt" + ], + [ + "Tuning hyperparameters is a cumbersome process as the hyperparamter space expands exponentially with the number of parameters, therefore efficient exploration is needed for any feasible study. We used the random search algorithm supported by Hyperopt BIBREF24 library to randomly search for an optimal set of hyperparameters from a given parameter space. In Fig. FIGREF12, various hyperparameters we considered are plotted against the validation accuracy as violin plots. Our observations for each hyperparameter are summarized below:", + "Number of filters in first layer: We observe that having 128 filters gives better results as compared to other filter values of 32 and 64 in the first layer. A higher number of filters in the first layer of network is able to preserve most of the characteristics of input.", + "Kernel Size: We varied the receptive fields of convolutional layers by choosing the kernel size from among the set of {3, 5, 7, 9}. We observe that a kernel size of 9 gives better accuracy at the cost of increased computation time and larger number of parameters. A large kernel size is able to capture longer patterns in its input due to bigger receptive power which results in an improved accuracy.", + "Dropout: Dropout randomly turns-off (sets to 0) various individual nodes during training of the network. In a deep CNN it is important that nodes do not develop a co-dependency amongst each other during training in order to prevent overfitting on training data BIBREF25. Dropout rate of $0.1$ works well for our model. When using a higher dropout rate the network is not able to capture the patterns in training dataset.", + "Batch Size: We chose batch sizes from amongst the set {32, 64, 128}. There is more noise while calculating error in a smaller batch size as compared to a larger one. This tends to have a regularizing effect during training of the network and hence gives better results. Thus, batch size of 32 works best for the model.", + "Layers in Convolutional block 1 and 2: We varied the number of layers in both the convolutional blocks. If the number of layers is low, then the network does not have enough depth to capture patterns in the data whereas having large number of layers leads to overfitting on the data. In our network, two layers in the first block and one layer in the second block give optimal results." + ], + [ + "Log-Mel spectrogram is the most commonly used method for converting audio into the image domain. The audio data was again sampled at 8 kHz. The input to this model was the log-Mel spectra. We generated log-Mel spectrogram using the LibROSA BIBREF26 library. In Table TABREF16, we present a detailed layer-by-layer illustration of the model along with its hyperparameter." + ], + [ + "We took some specific design choices for this model, which are as follows:", + "We added residual connections with each convolutional layer. Residual connections in a way makes the model selective of the contributing layers, determines the optimal number of layers required for training and solves the problem of vanishing gradients. Residual connections or skip connections skip training of those layers that do not contribute much in the overall outcome of model.", + "We added spatial attention BIBREF27 networks to help the model in focusing on specific regions or areas in an image. Spatial attention aids learning irrespective of transformations, scaling and rotation done on the input images making the model more robust and helping it to achieve better results.", + "We added Channel Attention networks so as to help the model to find interdependencies among color channels of log-Mel spectra. It adaptively assigns importance to each color channel in a deep convolutional multi-channel network. In our model we apply channel and spatial attention just before feeding the input into bi-directional GRU. This helps the model to focus on selected regions and at the same time find patterns among channels to better determine the language." + ], + [ + "We used the random search algorithm supported by Hyperopt BIBREF24 library to randomly search for an optimal set of hyperparameters from a given parameter space. In Fig. FIGREF19 ,various hyperparameters we tuned are plotted against the validation accuracy. Our observations for each hyperparameter are summarized below:", + "Filter Size: 64 filters in the first layer of network can preserve most of the characteristics of input, but increasing it to 128 is inefficient as overfitting occurs.", + "Kernel Size: There is a trade-off between kernel size and capturing complex non-linear features. Using a small kernel size will require more layers to capture features whereas using a large kernel size will require less layers. Large kernels capture simple non-linear features whereas using a smaller kernel will help us capture more complex non-linear features. However, with more layers, backpropagation necessitates the need for a large memory. We experimented with large kernel size and gradually increased the layers in order to capture more complex features. The results are not conclusive and thus we chose kernel size of 7 against 3.", + "Dropout: Dropout rate of 0.1 works well for our data. When using a higher dropout rate the network is not able to capture the patterns in training dataset.", + "Batch Size: There is always a trade-off between batch size and getting accurate gradients. Using a large batch size helps the model to get more accurate gradients since the model tries to optimize gradients over a large set of images. We found that using a batch size of 128 helped the model to train faster and get better results than using a batch size less than 128.", + "Number of hidden units in bi-directional GRU: Varying the number of hidden units and layers in GRU helps the model to capture temporal features which can play a significant role in identifying the language correctly. The optimal number of hidden units and layers depends on the complexity of the dataset. Using less number of hidden units may capture less features whereas using large number of hidden units may be computationally expensive. In our case we found that using 1536 hidden units in a single bi-directional GRU layer leads to the best result.", + "Image Size: We experimented with log-Mel spectra images of sizes $64 \\times 64$ and $128 \\times 128$ pixels and found that our model worked best with images of size of $128 \\times 128$ pixels.", + "We also evaluated our model on data with mixup augmentation BIBREF28. It is a data augmentation technique that also acts as a regularization technique and prevents overfitting. Instead of directly taking images from the training dataset as input, mixup takes a linear combination of any two random images and feeds it as input. The following equations were used to prepared a mixed-up dataset:", + "and", + "where $\\alpha \\in [0, 1]$ is a random variable from a $\\beta $-distribution, $I_1$." + ], + [ + "This model is a similar model to 2D-ConvNet with Attention and bi-directional GRU described in section SECREF13 except that it lacks skip connections, attention layers, bi-directional GRU and the embedding layer incorporated in the previous model." + ], + [ + "We classified six languages (English, French, German, Spanish, Russian and Italian) from the VoxForge BIBREF6 dataset. VoxForge is an open-source speech corpus which primarily consists of samples recorded and submitted by users using their own microphone. This results in significant variation of speech quality between samples making it more representative of real world scenarios.", + "Our dataset consists of 1,500 samples for each of six languages. Out of 1,500 samples for each language, 1,200 were randomly selected as training dataset for that language and rest 300 as validation dataset using k-fold cross-validation. To sum up, we trained our model on 7,200 samples and validated it on 1800 samples comprising six languages. The results are discussed in next section." + ], + [ + "This paper discusses two end-to-end approaches which achieve state-of-the-art results in both the image as well as audio domain on the VoxForge dataset BIBREF6. In Table TABREF25, we present all the classification accuracies of the two models of the cases with and without mixup for six and four languages.", + "In the audio domain (using raw audio waveform as input), 1D-ConvNet achieved a mean accuracy of 93.7% with a standard deviation of 0.3% on running k-fold cross validation. In Fig FIGREF27 (a) we present the confusion matrix for the 1D-ConvNet model.", + "In the image domain (obtained by taking log-Mel spectra of raw audio), 2D-ConvNet with 2D attention (channel and spatial attention) and bi-directional GRU achieved a mean accuracy of 95.0% with a standard deviation of 1.2% for six languages. This model performed better when mixup regularization was applied. 2D-ConvNet achieved a mean accuracy of 95.4% with standard deviation of 0.6% on running k-fold cross validation for six languages when mixup was applied. In Fig FIGREF27 (b) we present the confusion matrix for the 2D-ConvNet model. 2D attention models focused on the important features extracted by convolutional layers and bi-directional GRU captured the temporal features." + ], + [ + "Several of the spoken languages in Europe belong to the Indo-European family. Within this family, the languages are divided into three phyla which are Romance, Germanic and Slavic. Of the 6 languages that we selected Spanish (Es), French (Fr) and Italian (It) belong to the Romance phyla, English and German belong to Germanic phyla and Russian in Slavic phyla. Our model also confuses between languages belonging to the similar phyla which acts as an insanity check since languages in same phyla have many similar pronounced words such as cat in English becomes Katze in German and Ciao in Italian becomes Chao in Spanish.", + "Our model confuses between French (Fr) and Russian (Ru) while these languages belong to different phyla, many words from French were adopted into Russian such as automate (oot-oo-mate) in French becomes ABTOMaT (aff-taa-maat) in Russian which have similar pronunciation.", + "" + ], + [ + "The performance of raw audio waveforms as input features to ConvNet can be further improved by applying silence removal in the audio. Also, there is scope for improvement by augmenting available data through various conventional techniques like pitch shifting, adding random noise and changing speed of audio. These help in making neural networks more robust to variations which might be present in real world scenarios. There can be further exploration of various feature extraction techniques like Constant-Q transform and Fast Fourier Transform and assessment of their impact on Language Identification.", + "There can be further improvements in neural network architectures like concatenating the high level features obtained from 1D-ConvNet and 2D-ConvNet, before performing classification. There can be experiments using deeper networks with skip connections and Inception modules. These are known to have positively impacted the performance of Convolutional Neural Networks." + ], + [ + "There are two main contributions of this paper in the domain of spoken language identification. Firstly, we presented an extensive analysis of raw audio waveforms as input features to 1D-ConvNet. We experimented with various hyperparameters in our 1D-ConvNet and evaluated their effect on validation accuracy. This method is able to bypass the computational overhead of conventional approaches which depend on generation of spectrograms as a necessary pre-procesing step. We were able to achieve an accauracy of 93.7% using this technique.", + "Next, we discussed the enhancement in performance of 2D-ConvNet using mixup augmentation, which is a recently developed technique to prevent over\ufb01tting on test data.This approach achieved an accuracy of 95.4%. We also analysed how attention mechanism and recurrent layers impact the performance of networks. This approach achieved an accuracy of 95.0%." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0077/instruction.md b/qasper-0077/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..d10b03fd5117451fa3f034de025b183eb691126f --- /dev/null +++ b/qasper-0077/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: AraNet: A Deep Learning Toolkit for Arabic Social Media + +Question: What datasets are used in training? \ No newline at end of file diff --git a/qasper-0083/instruction.md b/qasper-0083/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..17ffa2068b26da358b6fdda710655ee30a3a3888 --- /dev/null +++ b/qasper-0083/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Stacked DeBERT: All Attention in Incomplete Data for Text Classification + +Question: How many vanilla transformers do they use after applying an embedding layer? \ No newline at end of file diff --git a/qasper-0099/instruction.md b/qasper-0099/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..65cd10c0f66347e0b360e1ad5bab65b5ad99d51f --- /dev/null +++ b/qasper-0099/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: An empirical study on the effectiveness of images in Multimodal Neural Machine Translation + +Question: What misbehavior is identified? \ No newline at end of file diff --git a/qasper-0202/instruction.md b/qasper-0202/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..a8ef3e78187e631f430439eee17c6b1d21e8df02 --- /dev/null +++ b/qasper-0202/instruction.md @@ -0,0 +1,124 @@ +Name of Paper: Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness? + +Question: Which are three assumptions in current approaches for defining faithfulness? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Faithfulness vs. Plausibility", + "Inherently Interpretable?", + "Evaluation via Utility", + "Guidelines for Evaluating Faithfulness", + "Guidelines for Evaluating Faithfulness ::: Be explicit in what you evaluate.", + "Guidelines for Evaluating Faithfulness ::: Faithfulness evaluation should not involve human-judgement on the quality of interpretation.", + "Guidelines for Evaluating Faithfulness ::: Faithfulness evaluation should not involve human-provided gold labels.", + "Guidelines for Evaluating Faithfulness ::: Do not trust \u201cinherent interpretability\u201d claims.", + "Guidelines for Evaluating Faithfulness ::: Faithfulness evaluation of IUI systems should not rely on user performance.", + "Defining Faithfulness", + "Defining Faithfulness ::: Assumption 1 (The Model Assumption).", + "Defining Faithfulness ::: Assumption 2 (The Prediction Assumption).", + "Defining Faithfulness ::: Assumption 3 (The Linearity Assumption).", + "Is Faithful Interpretation Impossible?", + "Towards Better Faithfulness Criteria", + "Conclusion", + "Acknowledgements" + ], + "paragraphs": [ + [ + "Fueled by recent advances in deep-learning and language processing, NLP systems are increasingly being used for prediction and decision-making in many fields BIBREF0, including sensitive ones such as health, commerce and law BIBREF1. Unfortunately, these highly flexible and highly effective neural models are also opaque. There is therefore a critical need for explaining learning-based models' decisions.", + "The emerging research topic of interpretability or explainability has grown rapidly in recent years. Unfortunately, not without growing pains.", + "One such pain is the challenge of defining\u2014and evaluating\u2014what constitutes a quality interpretation. Current approaches define interpretation in a rather ad-hoc manner, motivated by practical use-cases and applications. However, this view often fails to distinguish between distinct aspects of the interpretation's quality, such as readability, plausibility and faithfulness BIBREF2. We argue (\u00a7SECREF2, \u00a7SECREF5) such conflation is harmful, and that faithfulness should be defined and evaluated explicitly, and independently from plausibility.", + "Our main focus is the evaluation of the faithfulness of an explanation. Intuitively, a faithful interpretation is one that accurately represents the reasoning process behind the model's prediction. We find this to be a pressing issue in explainability: in cases where an explanation is required to be faithful, imperfect or misleading evaluation can have disastrous effects.", + "While literature in this area may implicitly or explicitly evaluate faithfulness for specific explanation techniques, there is no consistent and formal definition of faithfulness. We uncover three assumptions that underlie all these attempts. By making the assumptions explicit and organizing the literature around them, we \u201cconnect the dots\u201d between seemingly distinct evaluation methods, and also provide a basis for discussion regarding the desirable properties of faithfulness (\u00a7SECREF6).", + "Finally, we observe a trend by which faithfulness is treated as a binary property, followed by showing that an interpretation method is not faithful. We claim that this is unproductive (\u00a7SECREF7), as the assumptions are nearly impossible to satisfy fully, and it is all too easy to disprove the faithfulness of an interpretation method via a counter-example. What can be done? We argue for a more practical view of faithfulness, calling for a graded criteria that measures the extent and likelihood of an interpretation to be faithful, in practice (\u00a7SECREF8). While we started to work in this area, we pose the exact formalization of these criteria, and concrete evaluations methods for them, as a central challenge to the community for the coming future." + ], + [ + "There is considerable research effort in attempting to define and categorize the desiderata of a learned system's interpretation, most of which revolves around specific use-cases BIBREF17, BIBREF15.", + "Two particularly notable criteria, each useful for a different purposes, are plausibility and faithfulness. \u201cPlausibility\u201d refers to how convincing the interpretation is to humans, while \u201cfaithfulness\u201d refers to how accurately it reflects the true reasoning process of the model BIBREF2, BIBREF18.", + "Naturally, it is possible to satisfy one of these properties without the other. For example, consider the case of interpretation via post-hoc text generation\u2014where an additional \u201cgenerator\u201d component outputs a textual explanation of the model's decision, and the generator is learned with supervision of textual explanations BIBREF19, BIBREF20, BIBREF21. In this case, plausibility is the dominating property, while there is no faithfulness guarantee.", + "Despite the difference between the two criteria, many authors do not clearly make the distinction, and sometimes conflate the two. Moreoever, the majority of works do not explicitly name the criteria under consideration, even when they clearly belong to one camp or the other.", + "We argue that this conflation is dangerous. For example, consider the case of recidivism prediction, where a judge is exposed to a model's prediction and its interpretation, and the judge believes the interpretation to reflect the model's reasoning process. Since the interpretation's faithfulness carries legal consequences, a plausible but unfaithful interpretation may be the worst-case scenario. The lack of explicit claims by research may cause misinformation to potential users of the technology, who are not versed in its inner workings. Therefore, clear distinction between these terms is critical." + ], + [ + "A distinction is often made between two methods of achieving interpretability: (1) interpreting existing models via post-hoc techniques; and (2) designing inherently interpretable models. BIBREF29 argues in favor of inherently interpretable models, which by design claim to provide more faithful interpretations than post-hoc interpretation of black-box models.", + "We warn against taking this argumentation at face-value: a method being \u201cinherently interpretable\u201d is merely a claim that needs to be verified before it can be trusted. Indeed, while attention mechanisms have been considered as \u201cinherently interpretable\u201d BIBREF30, BIBREF31, recent work cast doubt regarding their faithfulness BIBREF32, BIBREF33, BIBREF18." + ], + [ + "While explanations have many different use-cases, such as model debugging, lawful guarantees or health-critical guarantees, one other possible use-case with particularly prominent evaluation literature is Intelligent User Interfaces (IUI), via Human-Computer Interaction (HCI), of automatic models assisting human decision-makers. In this case, the goal of the explanation is to increase the degree of trust between the user and the system, giving the user more nuance towards whether the system's decision is likely correct, or not. In the general case, the final evaluation metric is the performance of the user at their task BIBREF34. For example, BIBREF35 evaluate various explanations of a model in a setting of trivia question answering.", + "However, in the context of faithfulness, we must warn against HCI-inspired evaluation, as well: increased performance in this setting is not indicative of faithfulness; rather, it is indicative of correlation between the plausibility of the explanations and the model's performance.", + "To illustrate, consider the following fictional case of a non-faithful explanation system, in an HCI evaluation setting: the explanation given is a heat-map of the textual input, attributing scores to various tokens. Assume the system explanations behave in the following way: when the output is correct, the explanation consists of random content words; and when the output is incorrect, it consists of random punctuation marks. In other words, the explanation is more likely to appear plausible when the model is correct, while at the same time not reflecting the true decision process of the model. The user, convinced by the nicer-looking explanations, performs better using this system. However, the explanation consistently claimed random tokens to be highly relevant to the model's reasoning process. While the system is concretely useful, the claims given by the explanation do not reflect the model's decisions whatsoever (by design).", + "While the above scenario is extreme, this misunderstanding is not entirely unlikely, since any degree of correlation between plausibility and model performance will result in increased user performance, regardless of any notion of faithfulness." + ], + [ + "We propose the following guidelines for evaluating the faithfulness of explanations. These guidelines address common pitfalls and sub-optimal practices we observed in the literature." + ], + [ + "Conflating plausability and faithfulness is harmful. You should be explicit on which one of them you evaluate, and use suitable methodologies for each one. Of course, the same applies when designing interpretation techniques\u2014be clear about which properties are being prioritized." + ], + [ + "We note that: (1) humans cannot judge if an interpretation is faithful or not: if they understood the model, interpretation would be unnecessary; (2) for similar reasons, we cannot obtain supervision for this problem, either. Therefore, human judgement should not be involved in evaluation for faithfulness, as human judgement measures plausability." + ], + [ + "We should be able to interpret incorrect model predictions, just the same as correct ones. Evaluation methods that rely on gold labels are influenced by human priors on what should the model do, and again push the evaluation in the direction of plausability." + ], + [ + "Inherent interpretability is a claim until proven otherwise. Explanations provided by \u201cinherently interpretable\u201d models must be held to the same standards as post-hoc interpretation methods, and be evaluated for faithfulness using the same set of evaluation techniques." + ], + [ + "End-task user performance in HCI settings is merely indicative of correlation between plausibility and model performance, however small this correlation is. While important to evaluate the utility of the interpretations for some use-cases, it is unrelated to faithfulness." + ], + [ + "What does it mean for an interpretation method to be faithful? Intuitively, we would like the provided interpretation to reflect the true reasoning process of the model when making a decision. But what is a reasoning process of a model, and how can reasoning processes be compared to each other?", + "Lacking a standard definition, different works evaluate their methods by introducing tests to measure properties that they believe good interpretations should satisfy. Some of these tests measure aspects of faithfulness. These ad-hoc definitions are often unique to each paper and inconsistent with each other, making it hard to find commonalities.", + "We uncover three assumptions that underlie all these methods, enabling us to organize the literature along standardized axes, and relate seemingly distinct lines of work. Moreover, exposing the underlying assumptions enables an informed discussion regarding their validity and merit (we leave such a discussion for future work, by us or others).", + "These assumptions, to our knowledge, encapsulate the current working definitions of faithfulness used by the research community." + ], + [ + "Two models will make the same predictions if and only if they use the same reasoning process.", + "Corollary 1.1. An interpretation system is unfaithful if it results in different interpretations of models that make the same decisions.", + "As demonstrated by a recent example concerning NLP models, it can be used for proof by counter-example. Theoretically, if all possible models which can perfectly mimic the model's decisions also provide the same interpretations, then they could be deemed faithful. Conversely, showing that two models provide the same results but different interpretations, disprove the faithfulness of the method. BIBREF18 show how these counter-examples can be derived with adversarial training of models which can mimic the original model, yet provide different explanations.", + "Corollary 1.2. An interpretation is unfaithful if it results in different decisions than the model it interprets.", + "A more direct application of the Model Assumption is via the notion of fidelity BIBREF15, BIBREF8. For cases in which the explanation is itself a model capable of making decisions (e.g., decision trees or rule lists BIBREF36), fidelity is defined as the degree to which the explanation model can mimic the original model's decisions (as an accuracy score). For cases where the explanation is not a computable model, BIBREF37 propose a simple way of mapping explanations to decisions via crowd-sourcing, by asking humans to simulate the model's decision without any access to the model, and only access to the input and explanation (termed forward simulation). This idea is further explored and used in practice by BIBREF38." + ], + [ + "On similar inputs, the model makes similar decisions if and only if its reasoning is similar.", + "Corollary 2. An interpretation system is unfaithful if it provides different interpretations for similar inputs and outputs.", + "Since the interpretation serves as a proxy for the model's \u201creasoning\u201d, it should satisfy the same constraints. In other words, interpretations of similar decisions should be similar, and interpretations of dissimilar decisions should be dissimilar.", + "This assumption is more useful to disprove the faithfulness of an interpretation rather than prove it, since a disproof requires finding appropriate cases where the assumption doesn't hold, where a proof would require checking a (very large) satisfactory quantity of examples, or even the entire input space.", + "One recent discussion in the NLP community BIBREF33, BIBREF18 concerns the use of this underlying assumption for evaluating attention heat-maps as explanations. The former attempts to provide different explanations of similar decisions per instance. The latter critiques the former and is based more heavily on the model assumption, described above.", + "Additionally, BIBREF39 propose to introduce a constant shift to the input space, and evaluate whether the explanation changes significantly as the final decision stays the same. BIBREF16 formalize a generalization of this technique under the term interpretability robustness: interpretations should be invariant to small perturbations in the input (a direct consequence of the prediction assumption). BIBREF40 further expand on this notion as \u201cconsistency of the explanation with respect to the model\u201d. Unfortunately, robustness measures are difficult to apply in NLP settings due to the discrete input." + ], + [ + "Certain parts of the input are more important to the model reasoning than others. Moreover, the contributions of different parts of the input are independent from each other.", + "Corollary 3. Under certain circumstances, heat-map interpretations can be faithful.", + "This assumption is employed by methods that consider heat-maps (e.g., attention maps) over the input as explanations, particularly popular in NLP. Heat-maps are claims about which parts of the input are more relevant than others to the model's decision. As such, we can design \u201cstress tests\u201d to verify whether they uphold their claims.", + "One method proposed to do so is erasure, where the \u201cmost relevant\u201d parts of the input\u2014according to the explanation\u2014are erased from the input, in expectation that the model's decision will change BIBREF25, BIBREF42, BIBREF32. Otherwise, the \u201cleast relevant\u201d parts of the input may be erased, in expectation that the model's decision will not change BIBREF43. BIBREF44, BIBREF45 propose two measures of comprehensiveness and sufficiency as a formal generalization of erasure: as the degree by which the model is influenced by the removal of the high-ranking features, or by inclusion of solely the high-ranking features." + ], + [ + "The aforementioned assumptions are currently utilized to evaluate faithfulness in a binary manner, whether an interpretation is strictly faithful or not. Specifically, they are most often used to show that a method is not faithful, by constructing cases in which the assumptions do not hold for the suggested method. In other words, there is a clear trend of proof via counter-example, for various interpretation methods, that they are not globally faithful.", + "We claim that this is unproductive, as we expect these various methods to consistently result in negative (not faithful) results, continuing the current trend. This follows because an interpretation functions as an approximation of the model or decision's true reasoning process, so it by definition loses information. By the pigeonhole principle, there will be inputs with deviation between interpretation and reasoning.", + "This is observed in practice, in numerous work that show adversarial behavior, or pathological behaviours, that arise from the deeply non-linear and high-dimensional decision boundaries of current models. Furthermore, because we lack supervision regarding which models or decisions are indeed mappable to human-readable concepts, we cannot ignore the approximation errors.", + "This poses a high bar for explanation methods to fulfill, a bar which we estimate will not be overcome soon, if at all. What should we do, then, if we desire a system that provides faithful explanations?" + ], + [ + "We argue that a way out of this standstill is in a more practical and nuanced methodology for defining and evaluating faithfulness. We propose the following challenge to the community: We must develop formal definition and evaluation for faithfulness that allows us the freedom to say when a method is sufficiently faithful to be useful in practice.", + "We note two possible approaches to this end:", + "Across models and tasks: The degree (as grayscale) of faithfulness at the level of specific models and tasks. Perhaps some models or tasks allow sufficiently faithful interpretation, even if the same is not true for others.", + "For example, the method may not be faithful for some question-answering task, but faithful for movie review sentiment, perhaps based on various syntactic and semantic attributes of those tasks.", + "Across input space: The degree of faithfulness at the level of subspaces of the input space, such as neighborhoods of similar inputs, or singular inputs themselves. If we are able to say with some degree of confidence whether a specific decision's explanation is faithful to the model, even if the interpretation method is not considered universally faithful, it can be used with respect to those specific areas or instances only." + ], + [ + "The opinion proposed in this paper is two-fold:", + "First, interpretability evaluation often conflates evaluating faithfulness and plausibility together. We should tease apart the two definitions and focus solely on evaluating faithfulness without any supervision or influence of the convincing power of the interpretation.", + "Second, faithfulness is often evaluated in a binary \u201cfaithful or not faithful\u201d manner, and we believe strictly faithful interpretation is a \u201cunicorn\u201d which will likely never be found. We should instead evaluate faithfulness on a more nuanced \u201cgrayscale\u201d that allows interpretations to be useful even if they are not globally and definitively faithful." + ], + [ + "We thank Yanai Elazar for welcome input on the presentation and organization of the paper. We also thank the reviewers for additional feedback and pointing to relevant literature in HCI and IUI.", + "This project has received funding from the Europoean Research Council (ERC) under the Europoean Union's Horizon 2020 research and innovation programme, grant agreement No. 802774 (iEXTRACT)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0203/instruction.md b/qasper-0203/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..d36d8cd1fb18b5a84c3f892c2f7f52958256149c --- /dev/null +++ b/qasper-0203/instruction.md @@ -0,0 +1,124 @@ +Name of Paper: Towards Faithfully Interpretable NLP Systems: How should we define and evaluate faithfulness? + +Question: Which are key points in guidelines for faithfulness evaluation? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Faithfulness vs. Plausibility", + "Inherently Interpretable?", + "Evaluation via Utility", + "Guidelines for Evaluating Faithfulness", + "Guidelines for Evaluating Faithfulness ::: Be explicit in what you evaluate.", + "Guidelines for Evaluating Faithfulness ::: Faithfulness evaluation should not involve human-judgement on the quality of interpretation.", + "Guidelines for Evaluating Faithfulness ::: Faithfulness evaluation should not involve human-provided gold labels.", + "Guidelines for Evaluating Faithfulness ::: Do not trust \u201cinherent interpretability\u201d claims.", + "Guidelines for Evaluating Faithfulness ::: Faithfulness evaluation of IUI systems should not rely on user performance.", + "Defining Faithfulness", + "Defining Faithfulness ::: Assumption 1 (The Model Assumption).", + "Defining Faithfulness ::: Assumption 2 (The Prediction Assumption).", + "Defining Faithfulness ::: Assumption 3 (The Linearity Assumption).", + "Is Faithful Interpretation Impossible?", + "Towards Better Faithfulness Criteria", + "Conclusion", + "Acknowledgements" + ], + "paragraphs": [ + [ + "Fueled by recent advances in deep-learning and language processing, NLP systems are increasingly being used for prediction and decision-making in many fields BIBREF0, including sensitive ones such as health, commerce and law BIBREF1. Unfortunately, these highly flexible and highly effective neural models are also opaque. There is therefore a critical need for explaining learning-based models' decisions.", + "The emerging research topic of interpretability or explainability has grown rapidly in recent years. Unfortunately, not without growing pains.", + "One such pain is the challenge of defining\u2014and evaluating\u2014what constitutes a quality interpretation. Current approaches define interpretation in a rather ad-hoc manner, motivated by practical use-cases and applications. However, this view often fails to distinguish between distinct aspects of the interpretation's quality, such as readability, plausibility and faithfulness BIBREF2. We argue (\u00a7SECREF2, \u00a7SECREF5) such conflation is harmful, and that faithfulness should be defined and evaluated explicitly, and independently from plausibility.", + "Our main focus is the evaluation of the faithfulness of an explanation. Intuitively, a faithful interpretation is one that accurately represents the reasoning process behind the model's prediction. We find this to be a pressing issue in explainability: in cases where an explanation is required to be faithful, imperfect or misleading evaluation can have disastrous effects.", + "While literature in this area may implicitly or explicitly evaluate faithfulness for specific explanation techniques, there is no consistent and formal definition of faithfulness. We uncover three assumptions that underlie all these attempts. By making the assumptions explicit and organizing the literature around them, we \u201cconnect the dots\u201d between seemingly distinct evaluation methods, and also provide a basis for discussion regarding the desirable properties of faithfulness (\u00a7SECREF6).", + "Finally, we observe a trend by which faithfulness is treated as a binary property, followed by showing that an interpretation method is not faithful. We claim that this is unproductive (\u00a7SECREF7), as the assumptions are nearly impossible to satisfy fully, and it is all too easy to disprove the faithfulness of an interpretation method via a counter-example. What can be done? We argue for a more practical view of faithfulness, calling for a graded criteria that measures the extent and likelihood of an interpretation to be faithful, in practice (\u00a7SECREF8). While we started to work in this area, we pose the exact formalization of these criteria, and concrete evaluations methods for them, as a central challenge to the community for the coming future." + ], + [ + "There is considerable research effort in attempting to define and categorize the desiderata of a learned system's interpretation, most of which revolves around specific use-cases BIBREF17, BIBREF15.", + "Two particularly notable criteria, each useful for a different purposes, are plausibility and faithfulness. \u201cPlausibility\u201d refers to how convincing the interpretation is to humans, while \u201cfaithfulness\u201d refers to how accurately it reflects the true reasoning process of the model BIBREF2, BIBREF18.", + "Naturally, it is possible to satisfy one of these properties without the other. For example, consider the case of interpretation via post-hoc text generation\u2014where an additional \u201cgenerator\u201d component outputs a textual explanation of the model's decision, and the generator is learned with supervision of textual explanations BIBREF19, BIBREF20, BIBREF21. In this case, plausibility is the dominating property, while there is no faithfulness guarantee.", + "Despite the difference between the two criteria, many authors do not clearly make the distinction, and sometimes conflate the two. Moreoever, the majority of works do not explicitly name the criteria under consideration, even when they clearly belong to one camp or the other.", + "We argue that this conflation is dangerous. For example, consider the case of recidivism prediction, where a judge is exposed to a model's prediction and its interpretation, and the judge believes the interpretation to reflect the model's reasoning process. Since the interpretation's faithfulness carries legal consequences, a plausible but unfaithful interpretation may be the worst-case scenario. The lack of explicit claims by research may cause misinformation to potential users of the technology, who are not versed in its inner workings. Therefore, clear distinction between these terms is critical." + ], + [ + "A distinction is often made between two methods of achieving interpretability: (1) interpreting existing models via post-hoc techniques; and (2) designing inherently interpretable models. BIBREF29 argues in favor of inherently interpretable models, which by design claim to provide more faithful interpretations than post-hoc interpretation of black-box models.", + "We warn against taking this argumentation at face-value: a method being \u201cinherently interpretable\u201d is merely a claim that needs to be verified before it can be trusted. Indeed, while attention mechanisms have been considered as \u201cinherently interpretable\u201d BIBREF30, BIBREF31, recent work cast doubt regarding their faithfulness BIBREF32, BIBREF33, BIBREF18." + ], + [ + "While explanations have many different use-cases, such as model debugging, lawful guarantees or health-critical guarantees, one other possible use-case with particularly prominent evaluation literature is Intelligent User Interfaces (IUI), via Human-Computer Interaction (HCI), of automatic models assisting human decision-makers. In this case, the goal of the explanation is to increase the degree of trust between the user and the system, giving the user more nuance towards whether the system's decision is likely correct, or not. In the general case, the final evaluation metric is the performance of the user at their task BIBREF34. For example, BIBREF35 evaluate various explanations of a model in a setting of trivia question answering.", + "However, in the context of faithfulness, we must warn against HCI-inspired evaluation, as well: increased performance in this setting is not indicative of faithfulness; rather, it is indicative of correlation between the plausibility of the explanations and the model's performance.", + "To illustrate, consider the following fictional case of a non-faithful explanation system, in an HCI evaluation setting: the explanation given is a heat-map of the textual input, attributing scores to various tokens. Assume the system explanations behave in the following way: when the output is correct, the explanation consists of random content words; and when the output is incorrect, it consists of random punctuation marks. In other words, the explanation is more likely to appear plausible when the model is correct, while at the same time not reflecting the true decision process of the model. The user, convinced by the nicer-looking explanations, performs better using this system. However, the explanation consistently claimed random tokens to be highly relevant to the model's reasoning process. While the system is concretely useful, the claims given by the explanation do not reflect the model's decisions whatsoever (by design).", + "While the above scenario is extreme, this misunderstanding is not entirely unlikely, since any degree of correlation between plausibility and model performance will result in increased user performance, regardless of any notion of faithfulness." + ], + [ + "We propose the following guidelines for evaluating the faithfulness of explanations. These guidelines address common pitfalls and sub-optimal practices we observed in the literature." + ], + [ + "Conflating plausability and faithfulness is harmful. You should be explicit on which one of them you evaluate, and use suitable methodologies for each one. Of course, the same applies when designing interpretation techniques\u2014be clear about which properties are being prioritized." + ], + [ + "We note that: (1) humans cannot judge if an interpretation is faithful or not: if they understood the model, interpretation would be unnecessary; (2) for similar reasons, we cannot obtain supervision for this problem, either. Therefore, human judgement should not be involved in evaluation for faithfulness, as human judgement measures plausability." + ], + [ + "We should be able to interpret incorrect model predictions, just the same as correct ones. Evaluation methods that rely on gold labels are influenced by human priors on what should the model do, and again push the evaluation in the direction of plausability." + ], + [ + "Inherent interpretability is a claim until proven otherwise. Explanations provided by \u201cinherently interpretable\u201d models must be held to the same standards as post-hoc interpretation methods, and be evaluated for faithfulness using the same set of evaluation techniques." + ], + [ + "End-task user performance in HCI settings is merely indicative of correlation between plausibility and model performance, however small this correlation is. While important to evaluate the utility of the interpretations for some use-cases, it is unrelated to faithfulness." + ], + [ + "What does it mean for an interpretation method to be faithful? Intuitively, we would like the provided interpretation to reflect the true reasoning process of the model when making a decision. But what is a reasoning process of a model, and how can reasoning processes be compared to each other?", + "Lacking a standard definition, different works evaluate their methods by introducing tests to measure properties that they believe good interpretations should satisfy. Some of these tests measure aspects of faithfulness. These ad-hoc definitions are often unique to each paper and inconsistent with each other, making it hard to find commonalities.", + "We uncover three assumptions that underlie all these methods, enabling us to organize the literature along standardized axes, and relate seemingly distinct lines of work. Moreover, exposing the underlying assumptions enables an informed discussion regarding their validity and merit (we leave such a discussion for future work, by us or others).", + "These assumptions, to our knowledge, encapsulate the current working definitions of faithfulness used by the research community." + ], + [ + "Two models will make the same predictions if and only if they use the same reasoning process.", + "Corollary 1.1. An interpretation system is unfaithful if it results in different interpretations of models that make the same decisions.", + "As demonstrated by a recent example concerning NLP models, it can be used for proof by counter-example. Theoretically, if all possible models which can perfectly mimic the model's decisions also provide the same interpretations, then they could be deemed faithful. Conversely, showing that two models provide the same results but different interpretations, disprove the faithfulness of the method. BIBREF18 show how these counter-examples can be derived with adversarial training of models which can mimic the original model, yet provide different explanations.", + "Corollary 1.2. An interpretation is unfaithful if it results in different decisions than the model it interprets.", + "A more direct application of the Model Assumption is via the notion of fidelity BIBREF15, BIBREF8. For cases in which the explanation is itself a model capable of making decisions (e.g., decision trees or rule lists BIBREF36), fidelity is defined as the degree to which the explanation model can mimic the original model's decisions (as an accuracy score). For cases where the explanation is not a computable model, BIBREF37 propose a simple way of mapping explanations to decisions via crowd-sourcing, by asking humans to simulate the model's decision without any access to the model, and only access to the input and explanation (termed forward simulation). This idea is further explored and used in practice by BIBREF38." + ], + [ + "On similar inputs, the model makes similar decisions if and only if its reasoning is similar.", + "Corollary 2. An interpretation system is unfaithful if it provides different interpretations for similar inputs and outputs.", + "Since the interpretation serves as a proxy for the model's \u201creasoning\u201d, it should satisfy the same constraints. In other words, interpretations of similar decisions should be similar, and interpretations of dissimilar decisions should be dissimilar.", + "This assumption is more useful to disprove the faithfulness of an interpretation rather than prove it, since a disproof requires finding appropriate cases where the assumption doesn't hold, where a proof would require checking a (very large) satisfactory quantity of examples, or even the entire input space.", + "One recent discussion in the NLP community BIBREF33, BIBREF18 concerns the use of this underlying assumption for evaluating attention heat-maps as explanations. The former attempts to provide different explanations of similar decisions per instance. The latter critiques the former and is based more heavily on the model assumption, described above.", + "Additionally, BIBREF39 propose to introduce a constant shift to the input space, and evaluate whether the explanation changes significantly as the final decision stays the same. BIBREF16 formalize a generalization of this technique under the term interpretability robustness: interpretations should be invariant to small perturbations in the input (a direct consequence of the prediction assumption). BIBREF40 further expand on this notion as \u201cconsistency of the explanation with respect to the model\u201d. Unfortunately, robustness measures are difficult to apply in NLP settings due to the discrete input." + ], + [ + "Certain parts of the input are more important to the model reasoning than others. Moreover, the contributions of different parts of the input are independent from each other.", + "Corollary 3. Under certain circumstances, heat-map interpretations can be faithful.", + "This assumption is employed by methods that consider heat-maps (e.g., attention maps) over the input as explanations, particularly popular in NLP. Heat-maps are claims about which parts of the input are more relevant than others to the model's decision. As such, we can design \u201cstress tests\u201d to verify whether they uphold their claims.", + "One method proposed to do so is erasure, where the \u201cmost relevant\u201d parts of the input\u2014according to the explanation\u2014are erased from the input, in expectation that the model's decision will change BIBREF25, BIBREF42, BIBREF32. Otherwise, the \u201cleast relevant\u201d parts of the input may be erased, in expectation that the model's decision will not change BIBREF43. BIBREF44, BIBREF45 propose two measures of comprehensiveness and sufficiency as a formal generalization of erasure: as the degree by which the model is influenced by the removal of the high-ranking features, or by inclusion of solely the high-ranking features." + ], + [ + "The aforementioned assumptions are currently utilized to evaluate faithfulness in a binary manner, whether an interpretation is strictly faithful or not. Specifically, they are most often used to show that a method is not faithful, by constructing cases in which the assumptions do not hold for the suggested method. In other words, there is a clear trend of proof via counter-example, for various interpretation methods, that they are not globally faithful.", + "We claim that this is unproductive, as we expect these various methods to consistently result in negative (not faithful) results, continuing the current trend. This follows because an interpretation functions as an approximation of the model or decision's true reasoning process, so it by definition loses information. By the pigeonhole principle, there will be inputs with deviation between interpretation and reasoning.", + "This is observed in practice, in numerous work that show adversarial behavior, or pathological behaviours, that arise from the deeply non-linear and high-dimensional decision boundaries of current models. Furthermore, because we lack supervision regarding which models or decisions are indeed mappable to human-readable concepts, we cannot ignore the approximation errors.", + "This poses a high bar for explanation methods to fulfill, a bar which we estimate will not be overcome soon, if at all. What should we do, then, if we desire a system that provides faithful explanations?" + ], + [ + "We argue that a way out of this standstill is in a more practical and nuanced methodology for defining and evaluating faithfulness. We propose the following challenge to the community: We must develop formal definition and evaluation for faithfulness that allows us the freedom to say when a method is sufficiently faithful to be useful in practice.", + "We note two possible approaches to this end:", + "Across models and tasks: The degree (as grayscale) of faithfulness at the level of specific models and tasks. Perhaps some models or tasks allow sufficiently faithful interpretation, even if the same is not true for others.", + "For example, the method may not be faithful for some question-answering task, but faithful for movie review sentiment, perhaps based on various syntactic and semantic attributes of those tasks.", + "Across input space: The degree of faithfulness at the level of subspaces of the input space, such as neighborhoods of similar inputs, or singular inputs themselves. If we are able to say with some degree of confidence whether a specific decision's explanation is faithful to the model, even if the interpretation method is not considered universally faithful, it can be used with respect to those specific areas or instances only." + ], + [ + "The opinion proposed in this paper is two-fold:", + "First, interpretability evaluation often conflates evaluating faithfulness and plausibility together. We should tease apart the two definitions and focus solely on evaluating faithfulness without any supervision or influence of the convincing power of the interpretation.", + "Second, faithfulness is often evaluated in a binary \u201cfaithful or not faithful\u201d manner, and we believe strictly faithful interpretation is a \u201cunicorn\u201d which will likely never be found. We should instead evaluate faithfulness on a more nuanced \u201cgrayscale\u201d that allows interpretations to be useful even if they are not globally and definitively faithful." + ], + [ + "We thank Yanai Elazar for welcome input on the presentation and organization of the paper. We also thank the reviewers for additional feedback and pointing to relevant literature in HCI and IUI.", + "This project has received funding from the Europoean Research Council (ERC) under the Europoean Union's Horizon 2020 research and innovation programme, grant agreement No. 802774 (iEXTRACT)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0204/instruction.md b/qasper-0204/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..eba877a611ab24e131f8feb2fd2cacc543c7f965 --- /dev/null +++ b/qasper-0204/instruction.md @@ -0,0 +1,75 @@ +Name of Paper: Interpreting Recurrent and Attention-Based Neural Models: a Case Study on Natural Language Inference + +Question: Did they use the state-of-the-art model to analyze the attention? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Task and Model", + "Visualization of Attention and Gating", + "Attention", + "LSTM Gating Signals", + "Conclusion" + ], + "paragraphs": [ + [ + "Deep learning has achieved tremendous success for many NLP tasks. However, unlike traditional methods that provide optimized weights for human understandable features, the behavior of deep learning models is much harder to interpret. Due to the high dimensionality of word embeddings, and the complex, typically recurrent architectures used for textual data, it is often unclear how and why a deep learning model reaches its decisions.", + "There are a few attempts toward explaining/interpreting deep learning-based models, mostly by visualizing the representation of words and/or hidden states, and their importances (via saliency or erasure) on shallow tasks like sentiment analysis and POS tagging BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 . In contrast, we focus on interpreting the gating and attention signals of the intermediate layers of deep models in the challenging task of Natural Language Inference. A key concept in explaining deep models is saliency, which determines what is critical for the final decision of a deep model. So far, saliency has only been used to illustrate the impact of word embeddings. In this paper, we extend this concept to the intermediate layer of deep models to examine the saliency of attention as well as the LSTM gating signals to understand the behavior of these components and their impact on the final decision.", + "We make two main contributions. First, we introduce new strategies for interpreting the behavior of deep models in their intermediate layers, specifically, by examining the saliency of the attention and the gating signals. Second, we provide an extensive analysis of the state-of-the-art model for the NLI task and show that our methods reveal interesting insights not available from traditional methods of inspecting attention and word saliency.", + "In this paper, our focus was on NLI, which is a fundamental NLP task that requires both understanding and reasoning. Furthermore, the state-of-the-art NLI models employ complex neural architectures involving key mechanisms, such as attention and repeated reading, widely seen in successful models for other NLP tasks. As such, we expect our methods to be potentially useful for other natural understanding tasks as well." + ], + [ + "In NLI BIBREF4 , we are given two sentences, a premise and a hypothesis, the goal is to decide the logical relationship (Entailment, Neutral, or Contradiction) between them.", + "Many of the top performing NLI models BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , are variants of the ESIM model BIBREF11 , which we choose to analyze in this paper. ESIM reads the sentences independently using LSTM at first, and then applies attention to align/contrast the sentences. Another round of LSTM reading then produces the final representations, which are compared to make the prediction. Detailed description of ESIM can be found in the Appendix.", + "Using the SNLI BIBREF4 data, we train two variants of ESIM, with dimensionality 50 and 300 respectively, referred to as ESIM-50 and ESIM-300 in the remainder of the paper." + ], + [ + "In this work, we are primarily interested in the internal workings of the NLI model. In particular, we focus on the attention and the gating signals of LSTM readers, and how they contribute to the decisions of the model." + ], + [ + "Attention has been widely used in many NLP tasks BIBREF12 , BIBREF13 , BIBREF14 and is probably one of the most critical parts that affects the inference decisions. Several pieces of prior work in NLI have attempted to visualize the attention layer to provide some understanding of their models BIBREF5 , BIBREF15 . Such visualizations generate a heatmap representing the similarity between the hidden states of the premise and the hypothesis (Eq. 19 of Appendix). Unfortunately the similarities are often the same regardless of the decision.", + "Let us consider the following example, where the same premise \u201cA kid is playing in the garden\u201d, is paired with three different hypotheses:", + "A kid is taking a nap in the garden", + "A kid is having fun in the garden with her family", + "A kid is having fun in the garden", + " Note that the ground truth relationships are Contradiction, Neutral, and Entailment, respectively.", + "The first row of Fig. 1 shows the visualization of normalized attention for the three cases produced by ESIM-50, which makes correct predictions for all of them. As we can see from the figure, the three attention maps are fairly similar despite the completely different decisions. The key issue is that the attention visualization only allows us to see how the model aligns the premise with the hypothesis, but does not show how such alignment impacts the decision. This prompts us to consider the saliency of attention.", + "The concept of saliency was first introduced in vision for visualizing the spatial support on an image for a particular object class BIBREF16 . In NLP, saliency has been used to study the importance of words toward a final decision BIBREF0 .", + "We propose to examine the saliency of attention. Specifically, given a premise-hypothesis pair and the model's decision $y$ , we consider the similarity between a pair of premise and hypothesis hidden states $e_{ij}$ as a variable. The score of the decision $S(y)$ is thus a function of $e_{ij}$ for all $i$ and $j$ . The saliency of $e_{ij}$ is then defined to be $|\\frac{\\partial S(y)}{\\partial {e_{ij}}}|$ .", + "The second row of Fig. 1 presents the attention saliency map for the three examples acquired by the same ESIM-50 model. Interestingly, the saliencies are clearly different across the examples, each highlighting different parts of the alignment. Specifically, for h1, we see the alignment between \u201cis playing\u201d and \u201ctaking a nap\u201d and the alignment of \u201cin a garden\u201d to have the most prominent saliency toward the decision of Contradiction. For h2, the alignment of \u201ckid\u201d and \u201cher family\u201d seems to be the most salient for the decision of Neutral. Finally, for h3, the alignment between \u201cis having fun\u201d and \u201ckid is playing\u201d have the strongest impact toward the decision of Entailment.", + "From this example, we can see that by inspecting the attention saliency, we effectively pinpoint which part of the alignments contribute most critically to the final prediction whereas simply visualizing the attention itself reveals little information.", + "In the previous examples, we study the behavior of the same model on different inputs. Now we use the attention saliency to compare the two different ESIM models: ESIM-50 and ESIM-300.", + "Consider two examples with a shared hypothesis of \u201cA man ordered a book\u201d and premise:", + "John ordered a book from amazon", + "Mary ordered a book from amazon", + " Here ESIM-50 fails to capture the gender connections of the two different names and predicts Neutral for both inputs, whereas ESIM-300 correctly predicts Entailment for the first case and Contradiction for the second.", + "In the first two columns of Fig. 2 (column a and b) we visualize the attention of the two examples for ESIM-50 (left) and ESIM-300 (right) respectively. Although the two models make different predictions, their attention maps appear qualitatively similar.", + "In contrast, columns 3-4 of Fig. 2 (column c and d) present the attention saliency for the two examples by ESIM-50 and ESIM-300 respectively. We see that for both examples, ESIM-50 primarily focused on the alignment of \u201cordered\u201d, whereas ESIM-300 focused more on the alignment of \u201cJohn\u201d and \u201cMary\u201d with \u201cman\u201d. It is interesting to note that ESIM-300 does not appear to learn significantly different similarity values compared to ESIM-50 for the two critical pairs of words (\u201cJohn\u201d, \u201cman\u201d) and (\u201cMary\u201d, \u201cman\u201d) based on the attention map. The saliency map, however, reveals that the two models use these values quite differently, with only ESIM-300 correctly focusing on them." + ], + [ + "LSTM gating signals determine the flow of information. In other words, they indicate how LSTM reads the word sequences and how the information from different parts is captured and combined. LSTM gating signals are rarely analyzed, possibly due to their high dimensionality and complexity. In this work, we consider both the gating signals and their saliency, which is computed as the partial derivative of the score of the final decision with respect to each gating signal.", + "Instead of considering individual dimensions of the gating signals, we aggregate them to consider their norm, both for the signal and for its saliency. Note that ESIM models have two LSTM layers, the first (input) LSTM performs the input encoding and the second (inference) LSTM generates the representation for inference.", + "In Fig. 3 we plot the normalized signal and saliency norms for different gates (input, forget, output) of the Forward input (bottom three rows) and inference (top three rows) LSTMs. These results are produced by the ESIM-50 model for the three examples of Section 3.1, one for each column.", + "From the figure, we first note that the saliency tends to be somewhat consistent across different gates within the same LSTM, suggesting that we can interpret them jointly to identify parts of the sentence important for the model's prediction.", + "Comparing across examples, we see that the saliency curves show pronounced differences across the examples. For instance, the saliency pattern of the Neutral example is significantly different from the other two examples, and heavily concentrated toward the end of the sentence (\u201cwith her family\u201d). Note that without this part of the sentence, the relationship would have been Entailment. The focus (evidenced by its strong saliency and strong gating signal) on this particular part, which presents information not available from the premise, explains the model's decision of Neutral.", + "Comparing the behavior of the input LSTM and the inference LSTM, we observe interesting shifts of focus. In particular, we see that the inference LSTM tends to see much more concentrated saliency over key parts of the sentence, whereas the input LSTM sees more spread of saliency. For example, for the Contradiction example, the input LSTM sees high saliency for both \u201ctaking\u201d and \u201cin\u201d, whereas the inference LSTM primarily focuses on \u201cnap\u201d, which is the key word suggesting a Contradiction. Note that ESIM uses attention between the input and inference LSTM layers to align/contrast the sentences, hence it makes sense that the inference LSTM is more focused on the critical differences between the sentences. This is also observed for the Neutral example as well.", + "It is worth noting that, while revealing similar general trends, the backward LSTM can sometimes focus on different parts of the sentence (e.g., see Fig. 11 of Appendix), suggesting the forward and backward readings provide complementary understanding of the sentence." + ], + [ + "We propose new visualization and interpretation strategies for neural models to understand how and why they work. We demonstrate the effectiveness of the proposed strategies on a complex task (NLI). Our strategies are able to provide interesting insights not achievable by previous explanation techniques. Our future work will extend our study to consider other NLP tasks and models with the goal of producing useful insights for further improving these models. Model In this section we describe the ESIM model. We divide ESIM to three main parts: 1) input encoding, 2) attention, and 3) inference. Figure 4 demonstrates a high-level view of the ESIM framework. Let $u=[u_1, \\cdots , u_n]$ and $v=[v_1, \\cdots , v_m]$ be the given premise with length $n$ and hypothesis with length $m$ respectively, where $u_i, v_j \\in \\mathbb {R}^r$ are word embeddings of $r$ -dimensional vector. The goal is to predict a label $y$ that indicates the logical relationship between premise $u$ and hypothesis $v$ . Below we briefly explain the aforementioned parts. Input Encoding It utilizes a bidirectional LSTM (BiLSTM) for encoding the given premise and hypothesis using Equations 16 and 17 respectively. ", + "$$\\hat{u} \\in \\mathbb {R}^{n \\times 2d}$$ (Eq. ) ", + "$$\\hat{v} \\in \\mathbb {R}^{m \\times 2d}$$ (Eq. ) where $u$ and $v=[v_1, \\cdots , v_m]$0 are the reading sequences of $v=[v_1, \\cdots , v_m]$1 and $v=[v_1, \\cdots , v_m]$2 respectively. Attention It employs a soft alignment method to associate the relevant sub-components between the given premise and hypothesis. Equation 19 (energy function) computes the unnormalized attention weights as the similarity of hidden states of the premise and hypothesis. ", + "$$u$$ (Eq. ) where $v=[v_1, \\cdots , v_m]$3 and $v=[v_1, \\cdots , v_m]$4 are the hidden representations of $v=[v_1, \\cdots , v_m]$5 and $v=[v_1, \\cdots , v_m]$6 respectively which are computed earlier in Equations 16 and 17 . Next, for each word in either premise or hypothesis, the relevant semantics in the other sentence is extracted and composed according to $v=[v_1, \\cdots , v_m]$7 . Equations 20 and 21 provide formal and specific details of this procedure. ", + "$$\\tilde{v}_j$$ (Eq. ) ", + "$$\\hat{u}$$ (Eq. ) where $v=[v_1, \\cdots , v_m]$8 represents the extracted relevant information of $v=[v_1, \\cdots , v_m]$9 by attending to $n$0 while $n$1 represents the extracted relevant information of $n$2 by attending to $n$3 . Next, it passes the enriched information through a projector layer which produce the final output of attention stage. Equations 22 and 23 formally represent this process. ", + "$$p$$ (Eq. ) ", + "$$q$$ (Eq. ) Here $n$4 stands for element-wise product while $n$5 and $n$6 are the trainable weights and biases of the projector layer respectively. $n$7 and $n$8 indicate the output of attention devision for premise and hypothesis respectively. Inference During this phase, it uses another BiLSTM to aggregate the two sequences of computed matching vectors, $n$9 and $m$0 from the attention stage (Equations 27 and 28 ). ", + "$$\\emph {softmax}$$ (Eq. ) ", + "$$\\hat{u} = \\textit {BiLSTM}(u)$$ (Eq. 16) where $m$1 and $m$2 are the reading sequences of $m$3 and $m$4 respectively. Finally the concatenation max and average pooling of $m$5 and $m$6 are pass through a multilayer perceptron (MLP) classifier that includes a hidden layer with $m$7 activation and $m$8 output layer. The model is trained in an end-to-end manner. Attention Study Here we provide more examples on the NLI task which intend to examine specific behavior in this model. Such examples indicate interesting observation that we can analyze them in the future works. Table 1 shows the list of all example. LSTM Gating Signal Finally, Figure 11 depicts the backward LSTM gating signals study. " + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0211/instruction.md b/qasper-0211/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..9d0582eee78e23a60311aa12e37de82773cc84f3 --- /dev/null +++ b/qasper-0211/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Exploring Question Understanding and Adaptation in Neural-Network-Based Question Answering + +Question: What is the exact performance on SQUAD? \ No newline at end of file diff --git a/qasper-0233/instruction.md b/qasper-0233/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..5c240c4af2e532dd29223dc04f5dfa3a5697ee1a --- /dev/null +++ b/qasper-0233/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Gating Mechanisms for Combining Character and Word-level Word Representations: An Empirical Study + +Question: Which model architecture do they use to obtain representations? \ No newline at end of file diff --git a/qasper-0234/instruction.md b/qasper-0234/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..598e53f7a5819d8257906836777a4ef2035033b1 --- /dev/null +++ b/qasper-0234/instruction.md @@ -0,0 +1,160 @@ +Name of Paper: Gating Mechanisms for Combining Character and Word-level Word Representations: An Empirical Study + +Question: Which downstream sentence-level tasks do they evaluate on? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Background", + "Mapping Characters to Character-level Word Representations", + "Combining Character and Word-level Representations", + "Obtaining Sentence Representations", + "Experimental Setup", + "Datasets", + "Word Similarity", + "Word Frequencies and Gating Values", + "Sentence-level Evaluation", + "Relationship Between Word- and Sentence-level Evaluation Tasks", + "Gating Mechanisms for Combining Characters and Word Representations", + "Sentence Representation Learning", + "General Feature-wise Transformations", + "Conclusions", + "Acknowledgements", + "Hyperparameters", + "Sentence Evaluation Datasets" + ], + "paragraphs": [ + [ + "Incorporating sub-word structures like substrings, morphemes and characters to the creation of word representations significantly increases their quality as reflected both by intrinsic metrics and performance in a wide range of downstream tasks BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 .", + "The reason for this improvement is related to sub-word structures containing information that is usually ignored by standard word-level models. Indeed, when representing words as vectors extracted from a lookup table, semantically related words resulting from inflectional processes such as surf, surfing, and surfed, are treated as being independent from one another. Further, word-level embeddings do not account for derivational processes resulting in syntactically-similar words with different meanings such as break, breakable, and unbreakable. This causes derived words, which are usually less frequent, to have lower-quality (or no) vector representations.", + "Previous works have successfully combined character-level and word-level word representations, obtaining overall better results than using only word-level representations. For example BIBREF1 achieved state-of-the-art results in a machine translation task by representing unknown words as a composition of their characters. BIBREF4 created word representations by adding the vector representations of the words' surface forms and their morphemes ( INLINEFORM0 ), obtaining significant improvements on intrinsic evaluation tasks, word similarity and machine translation. BIBREF5 concatenated character-level and word-level representations for creating word representations, and then used them as input to their models for obtaining state-of-the-art results in Named Entity Recognition on several languages.", + "What these works have in common is that the models they describe first learn how to represent subword information, at character BIBREF1 , morpheme BIBREF4 , or substring BIBREF0 levels, and then combine these learned representations at the word level. The incorporation of information at a finer-grained hierarchy results in higher-quality modeling of rare words, morphological processes, and semantics BIBREF6 .", + "There is no consensus, however, on which combination method works better in which case, or how the choice of a combination method affects downstream performance, either measured intrinsically at the word level, or extrinsically at the sentence level.", + "In this paper we aim to provide some intuitions about how the choice of mechanism for combining character-level with word-level representations influences the quality of the final word representations, and the subsequent effect these have in the performance of downstream tasks. Our contributions are as follows:" + ], + [ + "We are interested in studying different ways of combining word representations, obtained from different hierarchies, into a single word representation. Specifically, we want to study how combining word representations (1) taken directly from a word embedding lookup table, and (2) obtained from a function over the characters composing them, affects the quality of the final word representations.", + "Let INLINEFORM0 be a set, or vocabulary, of words with INLINEFORM1 elements, and INLINEFORM2 a vocabulary of characters with INLINEFORM3 elements. Further, let INLINEFORM4 be a sequence of words, and INLINEFORM5 be the sequence of characters composing INLINEFORM6 . Each token INLINEFORM7 can be represented as a vector INLINEFORM8 extracted directly from an embedding lookup table INLINEFORM9 , pre-trained or otherwise, and as a vector INLINEFORM10 built from the characters that compose it; in other words, INLINEFORM11 , where INLINEFORM12 is a function that maps a sequence of characters to a vector.", + "The methods for combining word and character-level representations we study, are of the form INLINEFORM0 where INLINEFORM1 is the final word representation." + ], + [ + "The function INLINEFORM0 is composed of an embedding layer, an optional context function, and an aggregation function.", + "The embedding layer transforms each character INLINEFORM0 into a vector INLINEFORM1 of dimension INLINEFORM2 , by directly taking it from a trainable embedding lookup table INLINEFORM3 . We define the matrix representation of word INLINEFORM4 as INLINEFORM5 .", + "The context function takes INLINEFORM0 as input and returns a context-enriched matrix representation INLINEFORM1 , in which each INLINEFORM2 contains a measure of information about its context, and interactions with its neighbors. In particular, we chose to do this by feeding INLINEFORM3 to a BiLSTM BIBREF7 , BIBREF8 .", + "Informally, we can think of LSTM BIBREF10 as a function INLINEFORM0 that takes a matrix INLINEFORM1 as input and returns a context-enriched matrix representation INLINEFORM2 , where each INLINEFORM3 encodes information about the previous elements INLINEFORM4 .", + "A BiLSTM is simply composed of 2 LSTM, one that reads the input from left to right (forward), and another that does so from right to left (backward). The output of the forward and backward LSTM are INLINEFORM0 and INLINEFORM1 respectively. In the backward case the LSTM reads INLINEFORM2 first and INLINEFORM3 last, therefore INLINEFORM4 will encode the context from INLINEFORM5 .", + "The aggregation function takes the context-enriched matrix representation of word INLINEFORM0 for both directions, INLINEFORM1 and INLINEFORM2 , and returns a single vector INLINEFORM3 . To do so we followed BIBREF11 , and defined the character-level representation INLINEFORM4 of word INLINEFORM5 as the linear combination of the forward and backward last hidden states returned by the context function: DISPLAYFORM0 ", + "where INLINEFORM0 and INLINEFORM1 are trainable parameters, and INLINEFORM2 represents the concatenation operation between two vectors." + ], + [ + "We tested three different methods for combining INLINEFORM0 with INLINEFORM1 : simple concatenation, a learned scalar gate BIBREF11 , and a learned vector gate (also referred to as feature-wise sigmoidal gate). Additionally, we compared these methods to two baselines: using pre-trained word vectors only, and using character-only features for representing words. See fig:methods for a visual description of the proposed methods.", + "word-only (w) considers only INLINEFORM0 and ignores INLINEFORM1 : DISPLAYFORM0 ", + "char-only (c) considers only INLINEFORM0 and ignores INLINEFORM1 : DISPLAYFORM0 ", + "concat (cat) concatenates both word and character-level representations: DISPLAYFORM0 ", + "scalar gate (sg) implements the scalar gating mechanism described by BIBREF11 : DISPLAYFORM0 ", + "where INLINEFORM0 and INLINEFORM1 are trainable parameters, INLINEFORM2 , and INLINEFORM3 is the sigmoid function.", + "vector gate (vg): DISPLAYFORM0 ", + "where INLINEFORM0 and INLINEFORM1 are trainable parameters, INLINEFORM2 , INLINEFORM3 is the element-wise sigmoid function, INLINEFORM4 is the element-wise product for vectors, and INLINEFORM5 is a vector of ones.", + "The vector gate is inspired by BIBREF11 and BIBREF12 , but is different to the former in that the gating mechanism acts upon each dimension of the word and character-level vectors, and different to the latter in that it does not rely on external sources of information for calculating the gating mechanism.", + "Finally, note that word only and char only are special cases of both gating mechanisms: INLINEFORM0 (scalar gate) and INLINEFORM1 (vector gate) correspond to word only; INLINEFORM2 and INLINEFORM3 correspond to char only." + ], + [ + "To enable sentence-level classification we need to obtain a sentence representation from the word vectors INLINEFORM0 . We achieved this by using a BiLSTM with max pooling, which was shown to be a good universal sentence encoding mechanism BIBREF13 .", + "Let INLINEFORM0 , be an input sentence and INLINEFORM1 its matrix representation, where each INLINEFORM2 was obtained by one of the methods described in subsec:methods. INLINEFORM3 is the context-enriched matrix representation of INLINEFORM4 obtained by feeding INLINEFORM5 to a BiLSTM of output dimension INLINEFORM6 . Lastly, INLINEFORM11 is the final sentence representation of INLINEFORM12 obtained by max-pooling INLINEFORM13 along the sequence dimension.", + "Finally, we initialized the word representations INLINEFORM0 using GloVe embeddings BIBREF14 , and fine-tuned them during training. Refer to app:hyperparams for details on the other hyperparameters we used." + ], + [ + "We trained our models for solving the Natural Language Inference (NLI) task in two datasets, SNLI BIBREF15 and MultiNLI BIBREF16 , and validated them in each corresponding development set (including the matched and mismatched development sets of MultiNLI).", + "For each dataset-method combination we trained 7 models initialized with different random seeds, and saved each when it reached its best validation accuracy. We then evaluated the quality of each trained model's word representations INLINEFORM0 in 10 word similarity tasks, using the system created by BIBREF17 .", + "Finally, we fed these obtained word vectors to a BiLSTM with max-pooling and evaluated the final sentence representations in 11 downstream transfer tasks BIBREF13 , BIBREF18 ." + ], + [ + "Word-level Semantic Similarity A desirable property of vector representations of words is that semantically similar words should have similar vector representations. Assessing whether a set of word representations possesses this quality is referred to as the semantic similarity task. This is the most widely-used evaluation method for evaluating word representations, despite its shortcomings BIBREF20 .", + "This task consists of comparing the similarity between word vectors measured by a distance metric (usually cosine distance), with a similarity score obtained from human judgements. High correlation between these similarities is an indicator of good performance.", + "A problem with this formulation though, is that the definition of \u201csimilarity\u201d often confounds the meaning of both similarity and relatedness. For example, cup and tea are related but dissimilar words, and this type of distinction is not always clear BIBREF21 , BIBREF22 .", + "To face the previous problem, we tested our methods in a wide variety of datasets, including some that explicitly model relatedness (WS353R), some that explicitly consider similarity (WS353S, SimLex999, SimVerb3500), and some where the distinction is not clear (MEN, MTurk287, MTurk771, RG, WS353). We also included the RareWords (RW) dataset for evaluating the quality of rare word representations. See appendix:datasets for a more complete description of the datasets we used.", + "Sentence-level Evaluation Tasks Unlike word-level representations, there is no consensus on the desirable properties sentence representations should have. In response to this, BIBREF13 created SentEval, a sentence representation evaluation benchmark designed for assessing how well sentence representations perform in various downstream tasks BIBREF23 .", + "Some of the datasets included in SentEval correspond to sentiment classification (CR, MPQA, MR, SST2, and SST5), subjectivity classification (SUBJ), question-type classification (TREC), recognizing textual entailment (SICK E), estimating semantic relatedness (SICK R), and measuring textual semantic similarity (STS16, STSB). The datasets are described by BIBREF13 , and we provide pointers to their original sources in the appendix table:sentence-eval-datasets.", + "To evaluate these sentence representations SentEval trained a linear model on top of them, and evaluated their performance in the validation sets accompanying each dataset. The only exception was the STS16 task, in which our representations were evaluated directly." + ], + [ + "table:wordlevelresults shows the quality of word representations in terms of the correlation between word similarity scores obtained by the proposed models and word similarity scores defined by humans.", + "First, we can see that for each task, character only models had significantly worse performance than every other model trained on the same dataset. The most likely explanation for this is that these models are the only ones that need to learn word representations from scratch, since they have no access to the global semantic knowledge encoded by the GloVe embeddings.", + "Further, bold results show the overall trend that vector gates outperformed the other methods regardless of training dataset. This implies that learning how to combine character and word-level representations at the dimension level produces word vector representations that capture a notion of word similarity and relatedness that is closer to that of humans.", + "Additionally, results from the MNLI row in general, and underlined results in particular, show that training on MultiNLI produces word representations better at capturing word similarity. This is probably due to MultiNLI data being richer than that of SNLI. Indeed, MultiNLI data was gathered from various sources (novels, reports, letters, and telephone conversations, among others), rather than the single image captions dataset from which SNLI was created.", + "Exceptions to the previous rule are models evaluated in MEN and RW. The former case can be explained by the MEN dataset containing only words that appear as image labels in the ESP-Game and MIRFLICKR-1M image datasets BIBREF24 , and therefore having data that is more closely distributed to SNLI than to MultiNLI.", + "More notably, in the RareWords dataset BIBREF25 , the word only, concat, and scalar gate methods performed equally, despite having been trained in different datasets ( INLINEFORM0 ), and the char only method performed significantly worse when trained in MultiNLI. The vector gate, however, performed significantly better than its counterpart trained in SNLI. These facts provide evidence that this method is capable of capturing linguistic phenomena that the other methods are unable to model.", + "table:word-similarity-dataset lists the word-similarity datasets and their corresponding reference. As mentioned in subsec:datasets, all the word-similarity datasets contain pairs of words annotated with similarity or relatedness scores, although this difference is not always explicit. Below we provide some details for each.", + "MEN contains 3000 annotated word pairs with integer scores ranging from 0 to 50. Words correspond to image labels appearing in the ESP-Game and MIRFLICKR-1M image datasets.", + "MTurk287 contains 287 annotated pairs with scores ranging from 1.0 to 5.0. It was created from words appearing in both DBpedia and in news articles from The New York Times.", + "MTurk771 contains 771 annotated pairs with scores ranging from 1.0 to 5.0, with words having synonymy, holonymy or meronymy relationships sampled from WordNet BIBREF56 .", + "RG contains 65 annotated pairs with scores ranging from 0.0 to 4.0 representing \u201csimilarity of meaning\u201d.", + "RW contains 2034 pairs of words annotated with similarity scores in a scale from 0 to 10. The words included in this dataset were obtained from Wikipedia based on their frequency, and later filtered depending on their WordNet synsets, including synonymy, hyperonymy, hyponymy, holonymy and meronymy. This dataset was created with the purpose of testing how well models can represent rare and complex words.", + "SimLex999 contains 999 word pairs annotated with similarity scores ranging from 0 to 10. In this case the authors explicitly considered similarity and not relatedness, addressing the shortcomings of datasets that do not, such as MEN and WS353. Words include nouns, adjectives and verbs.", + "SimVerb3500 contains 3500 verb pairs annotated with similarity scores ranging from 0 to 10. Verbs were obtained from the USF free association database BIBREF66 , and VerbNet BIBREF63 . This dataset was created to address the lack of representativity of verbs in SimLex999, and the fact that, at the time of creation, the best performing models had already surpassed inter-annotator agreement in verb similarity evaluation resources. Like SimLex999, this dataset also explicitly considers similarity as opposed to relatedness.", + "WS353 contains 353 word pairs annotated with similarity scores from 0 to 10.", + "WS353R is a subset of WS353 containing 252 word pairs annotated with relatedness scores. This dataset was created by asking humans to classify each WS353 word pair into one of the following classes: synonyms, antonyms, identical, hyperonym-hyponym, hyponym-hyperonym, holonym-meronym, meronym-holonym, and none-of-the-above. These annotations were later used to group the pairs into: similar pairs (synonyms, antonyms, identical, hyperonym-hyponym, and hyponym-hyperonym), related pairs (holonym-meronym, meronym-holonym, and none-of-the-above with a human similarity score greater than 5), and unrelated pairs (classified as none-of-the-above with a similarity score less than or equal to 5). This dataset is composed by the union of related and unrelated pairs.", + "WS353S is another subset of WS353 containing 203 word pairs annotated with similarity scores. This dataset is composed by the union of similar and unrelated pairs, as described previously." + ], + [ + "fig:gatingviz shows that for more common words the vector gate mechanism tends to favor only a few dimensions while keeping a low average gating value across dimensions. On the other hand, values are greater and more homogeneous across dimensions in rarer words. Further, fig:freqvsgatevalue shows this mechanism assigns, on average, a greater gating value to less frequent words, confirming the findings by BIBREF11 , and BIBREF12 .", + "In other words, the less frequent the word, the more this mechanism allows the character-level representation to influence the final word representation, as shown by eq:vg. A possible interpretation of this result is that exploiting character information becomes increasingly necessary as word-level representations' quality decrease.", + "Another observable trend in both figures is that gating values tend to be low on average. Indeed, it is possible to see in fig:freqvsgatevalue that the average gating values range from INLINEFORM0 to INLINEFORM1 . This result corroborates the findings by BIBREF11 , stating that setting INLINEFORM2 in eq:scalar-gate, was better than setting it to higher values.", + "In summary, the gating mechanisms learn how to compensate the lack of expressivity of underrepresented words by selectively combining their representations with those of characters." + ], + [ + "table:sentlevelresults shows the impact that different methods for combining character and word-level word representations have in the quality of the sentence representations produced by our models.", + "We can observe the same trend mentioned in subsec:word-similarity-eval, and highlighted by the difference between bold values, that models trained in MultiNLI performed better than those trained in SNLI at a statistically significant level, confirming the findings of BIBREF13 . In other words, training sentence encoders on MultiNLI yields more general sentence representations than doing so on SNLI.", + "The two exceptions to the previous trend, SICKE and SICKR, benefited more from models trained on SNLI. We hypothesize this is again due to both SNLI and SICK BIBREF26 having similar data distributions.", + "Additionally, there was no method that significantly outperformed the word only baseline in classification tasks. This means that the added expressivity offered by explicitly modeling characters, be it through concatenation or gating, was not significantly better than simply fine-tuning the pre-trained GloVe embeddings for this type of task. We hypothesize this is due to the conflation of two effects. First, the fact that morphological processes might not encode important information for solving these tasks; and second, that SNLI and MultiNLI belong to domains that are too dissimilar to the domains in which the sentence representations are being tested.", + "On the other hand, the vector gate significantly outperformed every other method in the STSB task when trained in both datasets, and in the STS16 task when trained in SNLI. This again hints at this method being capable of modeling phenomena at the word level, resulting in improved semantic representations at the sentence level." + ], + [ + "It is clear that the better performance the vector gate had in word similarity tasks did not translate into overall better performance in downstream tasks. This confirms previous findings indicating that intrinsic word evaluation metrics are not good predictors of downstream performance BIBREF29 , BIBREF30 , BIBREF20 , BIBREF31 .", + "subfig:mnli-correlations shows that the word representations created by the vector gate trained in MultiNLI had positively-correlated results within several word-similarity tasks. This hints at the generality of the word representations created by this method when modeling similarity and relatedness.", + "However, the same cannot be said about sentence-level evaluation performance; there is no clear correlation between word similarity tasks and sentence-evaluation tasks. This is clearly illustrated by performance in the STSBenchmark, the only in which the vector gate was significantly superior, not being correlated with performance in any word-similarity dataset. This can be interpreted simply as word-level representations capturing word-similarity not being a sufficient condition for good performance in sentence-level tasks.", + "In general, fig:correlations shows that there are no general correlation effects spanning both training datasets and combination mechanisms. For example, subfig:snli-correlations shows that, for both word-only and concat models trained in SNLI, performance in word similarity tasks correlates positively with performance in most sentence evaluation tasks, however, this does not happen as clearly for the same models trained in MultiNLI (subfig:mnli-correlations)." + ], + [ + "To the best of our knowledge, there are only two recent works that specifically study how to combine word and subword-level vector representations.", + " BIBREF11 propose to use a trainable scalar gating mechanism capable of learning a weighting scheme for combining character-level and word-level representations. They compared their proposed method to manually weighting both levels; using characters only; words only; or their concatenation. They found that in some datasets a specific manual weighting scheme performed better, while in others the learned scalar gate did.", + " BIBREF12 further expand the gating concept by making the mechanism work at a finer-grained level, learning how to weight each vector's dimensions independently, conditioned on external word-level features such as part-of-speech and named-entity tags. Similarly, they compared their proposed mechanism to using words only, characters only, and a concatenation of both, with and without external features. They found that their vector gate performed better than the other methods in all the reported tasks, and beat the state of the art in two reading comprehension tasks.", + "Both works showed that the gating mechanisms assigned greater importance to character-level representations in rare words, and to word-level representations in common ones, reaffirming the previous findings that subword structures in general, and characters in particular, are beneficial for modeling uncommon words." + ], + [ + "The problem of representing sentences as fixed-length vectors has been widely studied.", + " BIBREF32 suggested a self-adaptive hierarchical model that gradually composes words into intermediate phrase representations, and adaptively selects specific hierarchical levels for specific tasks. BIBREF33 proposed an encoder-decoder model trained by attempting to reconstruct the surrounding sentences of an encoded passage, in a fashion similar to Skip-gram BIBREF34 . BIBREF35 overcame the previous model's need for ordered training sentences by using autoencoders for creating the sentence representations. BIBREF36 implemented a model simpler and faster to train than the previous two, while having competitive performance. Similar to BIBREF33 , BIBREF37 suggested predicting future sentences with a hierarchical CNN-LSTM encoder.", + " BIBREF13 trained several sentence encoding architectures on a combination of the SNLI and MultiNLI datasets, and showed that a BiLSTM with max-pooling was the best at producing highly transferable sentence representations. More recently, BIBREF18 empirically showed that sentence representations created in a multi-task setting BIBREF38 , performed increasingly better the more tasks they were trained in. BIBREF39 proposed using an autoencoder that relies on multi-head self-attention over the concatenation of the max and mean pooled encoder outputs for producing sentence representations. Finally, BIBREF40 show that modern sentence embedding methods are not vastly superior to random methods.", + "The works mentioned so far usually evaluate the quality of the produced sentence representations in sentence-level downstream tasks. Common benchmarks grouping these kind of tasks include SentEval BIBREF23 , and GLUE BIBREF41 . Another trend, however, is to probe sentence representations to understand what linguistic phenomena they encode BIBREF42 , BIBREF43 , BIBREF44 , BIBREF45 , BIBREF46 ." + ], + [ + " BIBREF47 provide a review on feature-wise transformation methods, of which the mechanisms presented in this paper form a part of. In a few words, the INLINEFORM0 parameter, in both scalar gate and vector gate mechanisms, can be understood as a scaling parameter limited to the INLINEFORM1 range and conditioned on word representations, whereas adding the scaled INLINEFORM2 and INLINEFORM3 representations can be seen as biasing word representations conditioned on character representations.", + "The previous review extends the work by BIBREF48 , which describes the Feature-wise Linear Modulation (FiLM) framework as a generalization of Conditional Normalization methods, and apply it in visual reasoning tasks. Some of the reported findings are that, in general, scaling has greater impact than biasing, and that in a setting similar to the scalar gate, limiting the scaling parameter to INLINEFORM0 hurt performance. Future decisions involving the design of mechanisms for combining character and word-level representations should be informed by these insights." + ], + [ + "We presented an empirical study showing the effect that different ways of combining character and word representations has in word-level and sentence-level evaluation tasks.", + "We showed that a vector gate performed consistently better across a variety of word similarity and relatedness tasks. Additionally, despite showing inconsistent results in sentence evaluation tasks, it performed significantly better than the other methods in semantic similarity tasks.", + "We further showed through this mechanism, that learning character-level representations is always beneficial, and becomes increasingly so with less common words.", + "In the future it would be interesting to study how the choice of mechanism for combining subword and word representations affects the more recent language-model-based pretraining methods such as ELMo BIBREF49 , GPT BIBREF50 , BIBREF51 and BERT BIBREF52 ." + ], + [ + "Thanks to Edison Marrese-Taylor and Pablo Loyola for their feedback on early versions of this manuscript. We also gratefully acknowledge the support of the NVIDIA Corporation with the donation of one of the GPUs used for this research. Jorge A. Balazs is partially supported by the Japanese Government MEXT Scholarship." + ], + [ + "We only considered words that appear at least twice, for each dataset. Those that appeared only once were considered UNK. We used the Treebank Word Tokenizer as implemented in NLTK for tokenizing the training and development datasets.", + "In the same fashion as conneau2017supervised, we used a batch size of 64, an SGD optmizer with an initial learning rate of INLINEFORM0 , and at each epoch divided the learning rate by 5 if the validation accuracy decreased. We also used gradient clipping when gradients where INLINEFORM1 .", + "We defined character vector representations as 50-dimensional vectors randomly initialized by sampling from the uniform distribution in the INLINEFORM0 range.", + "The output dimension of the character-level BiLSTM was 300 per direction, and remained of such size after combining forward and backward representations as depicted in eq. EQREF9 .", + "Word vector representations where initialized from the 300-dimensional GloVe vectors BIBREF14 , trained in 840B tokens from the Common Crawl, and finetuned during training. Words not present in the GloVe vocabulary where randomly initialized by sampling from the uniform distribution in the INLINEFORM0 range.", + "The input size of the word-level LSTM was 300 for every method except concat in which it was 600, and its output was always 2048 per direction, resulting in a 4096-dimensional sentence representation." + ], + [ + "table:sentence-eval-datasets lists the sentence-level evaluation datasets used in this paper. The provided URLs correspond to the original sources, and not necessarily to the URLs where SentEval got the data from.", + "The version of the CR, MPQA, MR, and SUBJ datasets used in this paper were the ones preprocessed by BIBREF75 . Both SST2 and SST5 correspond to preprocessed versions of the SST dataset by BIBREF74 . SST2 corresponds to a subset of SST used by BIBREF54 containing flat representations of sentences annotated with binary sentiment labels, and SST5 to another subset annotated with more fine-grained sentiment labels (very negative, negative, neutral, positive, very positive)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0242/instruction.md b/qasper-0242/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4dc9d71d0738953bac104a27c9be410b67eb9b35 --- /dev/null +++ b/qasper-0242/instruction.md @@ -0,0 +1,104 @@ +Name of Paper: Learning to Rank Scientific Documents from the Crowd + +Question: what is the size of this built corpus? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + null, + "Introduction", + "Benchmark Datasets", + "Learning to Rank", + "Features", + "Baseline Systems", + "Evaluation Measures", + "Forward Feature Selection", + "Results", + "Discussion", + "Acknowledgments" + ], + "paragraphs": [ + [ + "[block]I.1em", + "[block]i.1em", + " Learning to Rank Scientific Documents from the CrowdLearning to Rank Scientific Documents from the Crowd ", + "-4", + "[1]1" + ], + [ + "The number of biomedical research papers published has increased dramatically in recent years. As of October, 2016, PubMed houses over 26 million citations, with almost 1 million from the first 3 quarters of 2016 alone . It has become impossible for any one person to actually read all of the work being published. We require tools to help us determine which research articles would be most informative and related to a particular question or document. For example, a common task when reading articles is to find articles that are most related to another. Major research search engines offer such a \u201crelated articles\u201d feature. However, we propose that instead of measuring relatedness by text-similarity measures, we build a model that is able to infer relatedness from the authors' judgments.", + " BIBREF0 consider two kinds of queries important to bibliographic information retrieval: the first is a search query written by the user and the second is a request for documents most similar to a document already judged relevant by the user. Such a query-by-document (or query-by-example) system has been implemented in the de facto scientific search engine PubMed\u2014called Related Citation Search. BIBREF1 show that 19% of all PubMed searches performed by users have at least one click on a related article. Google Scholar provides a similar Related Articles system. Outside of bibliographic retrieval, query-by-document systems are commonly used for patent retrieval, Internet search, and plagiarism detection, amongst others. Most work in the area of query-by-document uses text-based similarity measures ( BIBREF2 , BIBREF3 , BIBREF4 ). However, scientific research is hypothesis driven and therefore we question whether text-based similarity alone is the best model for bibliographic retrieval. In this study we asked authors to rank documents by \u201ccloseness\u201d to their work. The definition of \u201ccloseness\u201d was left for the authors to interpret, as the goal is to model which documents the authors subjectively feel are closest to their own. Throughout the paper we will use \u201ccloseness\u201d and \u201crelatedness\u201d interchangeably.", + "We found that researchers' ranking by closeness differs significantly from the ranking provided by a traditional IR system. Our contributions are three fold:", + "The principal ranking algorithms of query-by-document in bibliographic information retrieval rely mainly on text similarity measures ( BIBREF1 , BIBREF0 ). For example, the foundational work of BIBREF0 introduced the concept of a \u201cdocument neighborhood\u201d in which they pre-compute a text-similarity based distance between each pair of documents. When a user issues a query, first an initial set of related documents is retrieved. Then, the neighbors of each of those documents is retrieved, i.e., documents with the highest text similarity to those in the initial set. In a later work, BIBREF1 develop the PMRA algorithm for PubMed related article search. PMRA is an unsupervised probabilistic topic model that is trained to model \u201crelatedness\u201d between documents. BIBREF5 introduce the competing algorithm Find-Similar for this task, treating the full text of documents as a query and selecting related documents from the results.", + "Outside bibliographic IR, prior work in query-by-document includes patent retrieval ( BIBREF6 , BIBREF3 ), finding related documents given a manuscript ( BIBREF1 , BIBREF7 ), and web page search ( BIBREF8 , BIBREF9 ). Much of the work focuses on generating shorter queries from the lengthy document. For example, noun-phrase extraction has been used for extracting short, descriptive phrases from the original lengthy text ( BIBREF10 ). Topic models have been used to distill a document into a set of topics used to form query ( BIBREF11 ). BIBREF6 generated queries using the top TF*IDF weighted terms in each document. BIBREF4 suggested extracting phrasal concepts from a document, which are then used to generate queries. BIBREF2 combined query extraction and pseudo-relevance feedback for patent retrieval. BIBREF9 employ supervised machine learning model (i.e., Conditional Random Fields) ( BIBREF12 ) for query generation. BIBREF13 explored ontology to identify chemical concepts for queries.", + "There are also many biomedical-document specific search engines available. Many information retrieval systems focus on question answering systems such as those developed for the TREC Genomics Track ( BIBREF14 ) or BioASQ Question-Answer ( BIBREF15 ) competitions. Systems designed for question-answering use a combination of natural language processing techniques to identify biomedical entities, and then information retrieval systems to extract relevant answers to questions. Systems like those detailed in BIBREF16 can provide answers to yes/no biomedical questions with high precision. However what we propose differs from these systems in a fundamental way: given a specific document, suggest the most important documents that are related to it.", + "The body of work most related to ours is that of citation recommendation. The goal of citation recommendation is to suggest a small number of publications that can be used as high quality references for a particular article ( BIBREF17 , BIBREF1 ). Topic models have been used to rank articles based on the similarity of latent topic distribution ( BIBREF11 , BIBREF18 , BIBREF1 ). These models attempt to decompose a document into a few important keywords. Specifically, these models attempt to find a latent vector representation of a document that has a much smaller dimensionality than the document itself and compare the reduced dimension vectors.", + "Citation networks have also been explored for ranking articles by importance, i.e., authority ( BIBREF19 , BIBREF20 ). BIBREF17 introduced heterogeneous network models, called meta-path based models, to incorporate venues (the conference where a paper is published) and content (the term which links two articles, for citation recommendation). Another highly relevant work is BIBREF8 who decomposed a document to represent it with a compact vector, which is then used to measure the similarity with other documents. Note that we exclude the work of context-aware recommendation, which analyze each citation's local context, which is typically short and does not represent a full document.", + "One of the key contributions of our study is an innovative approach for automatically generating a query-by-document gold standard. Crowd-sourcing has generated large databases, including Wikipedia and Freebase. Recently, BIBREF21 concluded that unpaid participants performed better than paid participants for question answering. They attribute this to unpaid participants being more intrinsically motivated than the paid test takers: they performed the task for fun and already had knowledge about the subject being tested. In contrast, another study, BIBREF22 , compared unpaid workers found through Google Adwords (GA) to paid workers found through Amazon Mechanical Turk (AMT). They found that the paid participants from AMT outperform the unpaid ones. This is attributed to the paid workers being more willing to look up information they didn't know. In the bibliographic domain, authors of scientific publications have contributed annotations ( BIBREF23 ). They found that authors are more willing to annotate their own publications ( BIBREF23 ) than to annotate other publications ( BIBREF24 ) even though they are paid. In this work, our annotated dataset was created by the unpaid authors of the articles." + ], + [ + "In order to develop and evaluate ranking algorithms we need a benchmark dataset. However, to the best of our knowledge, we know of no openly available benchmark dataset for bibliographic query-by-document systems. We therefore created such a benchmark dataset.", + "The creation of any benchmark dataset is a daunting labor-intensive task, and in particular, challenging in the scientific domain because one must master the technical jargon of a scientific article, and such experts are not easy to find when using traditional crowd-sourcing technologies (e.g., AMT). For our task, the ideal annotator for each of our articles are the authors themselves. The authors of a publication typically have a clear knowledge of the references they cite and their scientific importance to their publication, and therefore may be excellent judges for ranking the reference articles.", + "Given the full text of a scientific publication, we want to rank its citations according to the author's judgments. We collected recent publications from the open-access PLoS journals and asked the authors to rank by closeness five citations we selected from their paper. PLoS articles were selected because its journals cover a wide array of topics and the full text articles are available in XML format. We selected the most recent publications as previous work in crowd-sourcing annotation shows that authors' willingness to participate in an unpaid annotation task declines with the age of publication ( BIBREF23 ). We then extracted the abstract, citations, full text, authors, and corresponding author email address from each document. The titles and abstracts of the citations were retrieved from PubMed, and the cosine similarity between the PLoS abstract and the citation's abstract was calculated. We selected the top five most similar abstracts using TF*IDF weighted cosine similarity, shuffled their order, and emailed them to the corresponding author for annotation. We believe that ranking five articles (rather than the entire collection of the references) is a more manageable task for an author compared to asking them to rank all references. Because the documents to be annotated were selected based on text similarity, they also represent a challenging baseline for models based on text-similarity features. In total 416 authors were contacted, and 92 responded (22% response rate). Two responses were removed from the dataset for incomplete annotation.", + "We asked authors to rank documents by how \u201cclose to your work\u201d they were. The definition of closeness was left to the discretion of the author. The dataset is composed of 90 annotated documents with 5 citations each ranked 1 to 5, where 1 is least relevant and 5 is most relevant for a total of 450 annotated citations." + ], + [ + "Learning-to-rank is a technique for reordering the results returned from a search engine query. Generally, the initial query to a search engine is concerned more with recall than precision: the goal is to obtain a subset of potentially related documents from the corpus. Then, given this set of potentially related documents, learning-to-rank algorithms reorder the documents such that the most relevant documents appear at the top of the list. This process is illustrated in Figure FIGREF6 .", + "There are three basic types of learning-to-rank algorithms: point-wise, pair-wise, and list-wise. Point-wise algorithms assign a score to each retrieved document and rank them by their scores. Pair-wise algorithms turn learning-to-rank into a binary classification problem, obtaining a ranking by comparing each individual pair of documents. List-wise algorithms try to optimize an evaluation parameter over all queries in the dataset.", + "Support Vector Machine (SVM) ( BIBREF25 ) is a commonly used supervised classification algorithm that has shown good performance over a range of tasks. SVM can be thought of as a binary linear classifier where the goal is to maximize the size of the gap between the class-separating line and the points on either side of the line. This helps avoid over-fitting on the training data. SVMRank is a modification to SVM that assigns scores to each data point and allows the results to be ranked ( BIBREF26 ). We use SVMRank in the experiments below. SVMRank has previously been used in the task of document retrieval in ( BIBREF27 ) for a more traditional short query task and has been shown to be a top-performing system for ranking.", + "SVMRank is a point-wise learning-to-rank algorithm that returns scores for each document. We rank the documents by these scores. It is possible that sometimes two documents will have the same score, resulting in a tie. In this case, we give both documents the same rank, and then leave a gap in the ranking. For example, if documents 2 and 3 are tied, their ranked list will be [5, 3, 3, 2, 1].", + "Models are trained by randomly splitting the dataset into 70% training data and 30% test data. We apply a random sub-sampling approach where the dataset is randomly split, trained, and tested 100 times due to the relatively small size of the data. A model is learned for each split and a ranking is produced for each annotated document.", + "We test three different supervised models. The first supervised model uses only text similarity features, the second model uses all of the features, and the third model runs forward feature selection to select the best performing combination of features. We also test using two different models trained on two different datasets: one trained using the gold standard annotations, and another trained using the judgments based on text similarity that were used to select the citations to give to the authors.", + "We tested several different learning to rank algorithms for this work. We found in preliminary testing that SVMRank had the best performance, so it will be used in the following experiments." + ], + [ + "Each citation is turned into a feature vector representing the relationship between the published article and the citation. Four types of features are used: text similarity, citation count and location, age of the citation, and the number of times the citation has appeared in the literature (citation impact). Text similarity features measure the similarity of the words used in different parts of the document. In this work, we calculate the similarity between a document INLINEFORM0 and a document it cites INLINEFORM1 by transforming the their text into term vectors. For example, to calculate the similarity of the abstracts between INLINEFORM2 and INLINEFORM3 we transform the abstracts into two term vectors, INLINEFORM4 and INLINEFORM5 . The length of each of the term vectors is INLINEFORM6 . We then weight each word by its Term-frequency * Inverse-document frequency (TF*IDF) weight. TF*IDF is a technique to give higher weight to words that appear frequently in a document but infrequently in the corpus. Term frequency is simply the number of times that a word INLINEFORM7 appears in a document. Inverse-document frequency is the logarithmically-scaled fraction of documents in the corpus in which the word INLINEFORM8 appears. Or, more specifically: INLINEFORM9 ", + "where INLINEFORM0 is the total number of documents in the corpus, and the denominator is the number of documents in which a term INLINEFORM1 appears in the corpus INLINEFORM2 . Then, TF*IDF is defined as: INLINEFORM3 ", + "where INLINEFORM0 is a term, INLINEFORM1 is the document, and INLINEFORM2 is the corpus. For example, the word \u201cthe\u201d may appear often in a document, but because it also appears in almost every document in the corpus it is not useful for calculating similarity, thus it receives a very low weight. However, a word such as \u201cneurogenesis\u201d may appear often in a document, but does not appear frequently in the corpus, and so it receives a high weight. The similarity between term vectors is then calculated using cosine similarity: INLINEFORM3 ", + "where INLINEFORM0 and INLINEFORM1 are two term vectors. The cosine similarity is a measure of the angle between the two vectors. The smaller the angle between the two vectors, i.e., the more similar they are, then the closer the value is to 1. Conversely, the more dissimilar the vectors, the closer the cosine similarity is to 0.", + "We calculate the text similarity between several different sections of the document INLINEFORM0 and the document it cites INLINEFORM1 . From the citing article INLINEFORM2 , we use the title, full text, abstract, the combined discussion/conclusion sections, and the 10 words on either side of the place in the document where the actual citation occurs. From the document it cites INLINEFORM3 we only use the title and the abstract due to limited availability of the full text. In this work we combine the discussion and conclusion sections of each document because some documents have only a conclusion section, others have only a discussion, and some have both. The similarity between each of these sections from the two documents is calculated and used as features in the model.", + "The age of the citation may be relevant to its importance. As a citation ages, we hypothesize that it is more likely to become a \u201cfoundational\u201d citation rather than one that directly influenced the development of the article. Therefore more recent citations may be more likely relevant to the article. Similarly, \u201ccitation impact\u201d, that is, the number of times a citation has appeared in the literature (as measured by Google Scholar) may be an indicator of whether or not an article is foundational rather than directly related. We hypothesize that the fewer times an article is cited in the literature, the more impact it had on the article at hand.", + "We also keep track of the number of times a citation is mentioned in both the full text and discussion/conclusion sections. We hypothesize that if a citation is mentioned multiple times, it is more important than citations that are mentioned only once. Further, citations that appear in the discussion/conclusion sections are more likely to be crucial to understanding the results. We normalize the counts of the citations by the total number of citations in that section. In total we select 15 features, shown in Table TABREF15 . The features are normalized within each document so that each of citation features is on a scale from 0 to 1, and are evenly distributed within that range. This is done because some of the features (such as years since citation) are unbounded." + ], + [ + "We compare our system to a variety of baselines. (1) Rank by the number of times a citation is mentioned in the document. (2) Rank by the number of times the citation is cited in the literature (citation impact). (3) Rank using Google Scholar Related Articles. (4) Rank by the TF*IDF weighted cosine similarity. (5) Rank using a learning-to-rank model trained on text similarity rankings. The first two baseline systems are models where the values are ordered from highest to lowest to generate the ranking. The idea behind them is that the number of times a citation is mentioned in an article, or the citation impact may already be good indicators of their closeness. The text similarity model is trained using the same features and methods used by the annotation model, but trained using text similarity rankings instead of the author's judgments.", + "We also compare our rankings to those found on the popular scientific article search engine Google Scholar. Google Scholar is a \u201cblack box\u201d IR system: they do not release details about which features they are using and how they judge relevance of documents. Google Scholar provides a \u201cRelated Articles\u201d feature for each document in its index that shows the top 100 related documents for each article. To compare our rankings, we search through these related documents and record the ranking at which each of the citations we selected appeared. We scale these rankings such that the lowest ranked article from Google Scholar has the highest relevance ranking in our set. If the cited document does not appear in the set, we set its relevance-ranking equal to one below the lowest relevance ranking found.", + "Four comparisons are performed with the Google Scholar data. (1) We first train a model using our gold standard and see if we can predict Google Scholar's ranking. (2) We compare to a baseline of using Google Scholar's rankings to train and compare with their own rankings using our feature set. (3) Then we train a model using Google Scholar's rankings and try to predict our gold standard. (4) We compare it to the model trained on our gold standard to predict our gold standard." + ], + [ + "Normalized Discounted Cumulative Gain (NDCG) is a common measure for comparing a list of estimated document relevance judgments with a list of known judgments ( BIBREF28 ). To calculate NDCG we first calculate a ranking's Discounted Cumulative Gain (DCG) as: DISPLAYFORM0 ", + "where rel INLINEFORM0 is the relevance judgment at position INLINEFORM1 . Intuitively, DCG penalizes retrieval of documents that are not relevant (rel INLINEFORM2 ). However, DCG is an unbounded value. In order to compare the DCG between two models, we must normalize it. To do this, we use the ideal DCG (IDCG), i.e., the maximum possible DCG given the relevance judgments. The maximum possible DCG occurs when the relevance judgments are in the correct order. DISPLAYFORM0 ", + "The NDCG value is in the range of 0 to 1, where 0 means that no relevant documents were retrieved, and 1 means that the relevant documents were retrieved and in the correct order of their relevance judgments.", + "Kendall's INLINEFORM0 is a measure of the correlation between two ranked lists. It compares the number of concordant pairs with the number of discordant pairs between each list. A concordant pair is defined over two observations INLINEFORM1 and INLINEFORM2 . If INLINEFORM3 and INLINEFORM4 , then the pair at indices INLINEFORM5 is concordant, that is, the ranking at INLINEFORM6 in both ranking sets INLINEFORM7 and INLINEFORM8 agree with each other. Similarly, a pair INLINEFORM9 is discordant if INLINEFORM10 and INLINEFORM11 or INLINEFORM12 and INLINEFORM13 . Kendall's INLINEFORM14 is then defined as: DISPLAYFORM0 ", + "where C is the number of concordant pairs, D is the number of discordant pairs, and the denominator represents the total number of possible pairs. Thus, Kendall's INLINEFORM0 falls in the range of INLINEFORM1 , where -1 means that the ranked lists are perfectly negatively correlated, 0 means that they are not significantly correlated, and 1 means that the ranked lists are perfectly correlated. One downside of this measure is that it does not take into account where in the ranked list an error occurs. Information retrieval, in general, cares more about errors near the top of the list rather than errors near the bottom of the list.", + "Average-Precision INLINEFORM0 ( BIBREF29 ) (or INLINEFORM1 ) extends on Kendall's INLINEFORM2 by incorporating the position of errors. If an error occurs near the top of the list, then that is penalized heavier than an error occurring at the bottom of the list. To achieve this, INLINEFORM3 incorporates ideas from the popular Average Precision measure, were we calculate the precision at each index of the list and then average them together. INLINEFORM4 is defined as: DISPLAYFORM0 ", + "Intuitively, if an error occurs at the top of the list, then that error is propagated into each iteration of the summation, meaning that it's penalty is added multiple times. INLINEFORM0 's range is between -1 and 1, where -1 means the lists are perfectly negatively correlated, 0 means that they are not significantly correlated, and 1 means that they are perfectly correlated." + ], + [ + "Forward feature selection was performed by iteratively testing each feature one at a time. The highest performing feature is kept in the model, and another sweep is done over the remaining features. This continues until all features have been selected. This approach allows us to explore the effect of combinations of features and the effect of having too many or too few features. It also allows us to evaluate which features and combinations of features are the most powerful." + ], + [ + "We first compare our gold standard to the baselines. A random baseline is provided for reference. Because all of the documents that we rank are relevant, NDCG will be fairly high simply by chance. We find that the number of times a document is mentioned in the annotated document is significantly better than the random baseline or the citation impact. The more times a document is mentioned in a paper, the more likely the author was to annotate it as important. Interestingly, we see a negative correlation with the citation impact. The more times a document is mentioned in the literature, the less likely it is to be important. These results are shown in Table TABREF14 .", + "Next we rank the raw values of the features and compare them to our gold standard to obtain a baseline (Table TABREF15 ). The best performing text similarity feature is the similarity between the abstract of the annotated document and the abstract of the cited document. However, the number of times that a cited document is mentioned in the text of the annotated document are also high-scoring features, especially in the INLINEFORM0 correlation coefficient. These results indicate that text similarity alone may not be a good measure for judging the rank of a document.", + "Next we test three different feature sets for our supervised learning-to-rank models. The model using only the text similarity features performs poorly: NDCG stays at baseline and the correlation measures are low. Models that incorporate information about the age, number of times a cited document was referenced, and the citation impact of that document in addition to the text similarity features significantly outperformed models that used only text similarity features INLINEFORM0 . Because INLINEFORM1 takes into account the position in the ranking of the errors, this indicates that the All Features model was able to better correctly place highly ranked documents above lower ranked ones. Similarly, because Kendall's INLINEFORM2 is an overall measure of correlation that does not take into account the position of errors, the higher value here means that more rankings were correctly placed. Interestingly, feature selection (which is optimized for NDCG) does not outperform the model using all of the features in terms of our correlation measures. The features chosen during forward feature selection are (1) the citation impact, (2) number of mentions in the full text, (3) text similarity between the annotated document's title and the referenced document's abstract, (4) the text similarity between the annotated document's discussion/conclusion section and the referenced document's title. These results are shown in Table TABREF16 . The models trained on the text similarity judgments perform worse than the models trained on the annotated data. However, in terms of both NDCG and the correlation measures, they perform significantly better than the random baseline.", + "Next we compare our model to Google Scholar's rankings. Using the ranking collected from Google Scholar, we build a training set to try to predict our authors' rankings. We find that Google Scholar performs similarly to the text-only features model. This indicates that the rankings we obtained from the authors are substantially different than the rankings that Google Scholar provides. Results appear in Table TABREF17 ." + ], + [ + "We found that authors rank the references they cite substantially differently from rankings based on text-similarity. Our results show that decomposing a document into a set of features that is able to capture that difference is key. While text similarity is indeed important (as evidenced by the Similarity(a,a) feature in Table TABREF15 ), we also found that the number of times a document is referenced in the text and the number of times a document is referenced in the literature are also both important features (via feature selection). The more often a citation is mentioned in the text, the more likely it is to be important. This feature is often overlooked in article citation recommendation. We also found that recency is important: the age of the citation is negatively correlated with the rank. Newer citations are more likely to be directly important than older, more foundational citations. Additionally, the number of times a document is cited in the literature is negatively correlated with rank. This is likely due to highly cited documents being more foundational works; they may be older papers that are important to the field but not directly influential to the new work.", + "The model trained using the author's judgments does significantly better than the model trained using the text-similarity-based judgments. An error analysis was performed to find out why some of the rankings disagreed with the author's annotations. We found that in some cases our features were unable to capture the relationship: for example a biomedical document applying a model developed in another field to the dataset may use very different language to describe the model than the citation. Previous work adopting topic models to query document search may prove useful for such cases.", + "A small subset of features ended up performing as well as the full list of features. The number of times a citation was mentioned and the citation impact score in the literature ended up being two of the most important features. Indeed, without the citation-based features, the model performs as though it were trained with the text-similarity rankings. Feature engineering is a part of any learning-to-rank system, especially in domain-specific contexts. Citations are an integral feature of our dataset. For learning-to-rank to be applied to other datasets feature engineering must also occur to exploit the unique properties of those datasets. However, we show that combining the domain-specific features with more traditional text-based features does improve the model's scores over simply using the domain-specific features themselves.", + "Interestingly, citation impact and age of the citation are both negatively correlated with rank. We hypothesize that this is because both measures can be indicators of recency: a new publication is more likely to be directly influenced by more recent work. Many other related search tools, however, treat the citation impact as a positive feature of relatedness: documents with a higher citation impact appear higher on the list of related articles than those with lower citation impacts. This may be the opposite of what the user actually desires.", + "We also found that rankings from our text-similarity based IR system or Google Scholar's IR system were unable to rank documents by the authors' annotations as well as our system. In one sense, this is reasonable: the rankings coming from these systems were from a different system than the author annotations. However, in domain-specific IR, domain experts are the best judges. We built a system that exploits these expert judgments. The text similarity and Google Scholar models were able to do this to some extent, performing above the random baseline, but not on the level of our model.", + "Additionally, we observe that NDCG may not be the most appropriate measure for comparing short ranked lists where all of the documents are relevant to some degree. NDCG gives a lot of credit to relevant documents that occur in the highest ranks. However, all of the documents here are relevant, just to varying degrees. Thus, NDCG does not seem to be the most appropriate measure, as is evident in our scores. The correlation coefficients from Kendall's INLINEFORM0 and INLINEFORM1 seem to be far more appropriate for this case, as they are not concerned with relevance, only ranking.", + "One limitation of our work is that we selected a small set of references based on their similarities to the article that cites them. Ideally, we would have had authors rank all of their citations for us, but this would have been a daunting task for authors to perform. We chose to use the Google Scholar dataset in order to attempt to mitigate this: we obtain a ranking for the set of references from a system that is also ranking many other documents. The five citations selected by TF*IDF weighted cosine similarity represent a \u201chard\u201d gold standard: we are attempting to rank documents that are known to all be relevant by their nature, and have high similarity with the text. Additionally, there are plethora of other, more expensive features we could explore to improve the model. Citation network features, phrasal concepts, and topic models could all be used to help improve our results, at the cost of computational complexity.", + "We have developed a model for fast related-document ranking based on crowd-sourced data. The model, data, and data collection software are all publicly available and can easily be used in future applications as an automatic search to help users find the most important citations given a particular document. The experimental setup is portable to other datasets with some feature engineering. We were able to identify that several domain-specific features were crucial to our model, and that we were able to improve on the results of simply using those features alone by adding more traditional features.", + "Query-by-document is a complicated and challenging task. We provide an approach with an easily obtained dataset and a computationally inexpensive model. By working with biomedical researchers we were able to build a system that ranks documents in a quantitatively different way than previous systems, and to provide a tool that helps researchers find related documents." + ], + [ + "We would like to thank all of the authors who took the time to answer our citation ranking survey. This work is supported by National Institutes of Health with the grant number 1R01GM095476. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0245/instruction.md b/qasper-0245/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..cf370c9ae506758ef0587628482f1a7fda0a7fc5 --- /dev/null +++ b/qasper-0245/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Exploiting Deep Learning for Persian Sentiment Analysis + +Question: By how much did the results improve? \ No newline at end of file diff --git a/qasper-0258/instruction.md b/qasper-0258/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..2c55d1c4fff4e5d0ab1cb8ec5e46ba9351cd03fb --- /dev/null +++ b/qasper-0258/instruction.md @@ -0,0 +1,148 @@ +Name of Paper: RC-QED: Evaluating Natural Language Derivations in Multi-Hop Reading Comprehension + +Question: What is the baseline? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Task formulation: RC-QED ::: Input, output, and evaluation metrics", + "Task formulation: RC-QED ::: RC-QED@!START@$^{\\rm E}$@!END@", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Crowdsourcing interface", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Crowdsourcing interface ::: Judgement task (Figure @!START@UID13@!END@).", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Crowdsourcing interface ::: Derivation task (Figure @!START@UID14@!END@).", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Dataset", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Results", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Results ::: Quality", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Results ::: Agreement", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Knowledge graph construction", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Path ranking-based KGC (PRKGC)", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Training", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Training ::: Semi-supervising derivations", + "Experiments ::: Settings ::: Dataset", + "Experiments ::: Settings ::: Hyperparameters", + "Experiments ::: Settings ::: Baseline", + "Experiments ::: Results and discussion", + "Experiments ::: Results and discussion ::: QA performance.", + "Related work ::: RC datasets with explanations", + "Related work ::: Analysis of RC models and datasets", + "Related work ::: Other NLP corpora annotated with explanations", + "Conclusions", + "Example annotations" + ], + "paragraphs": [ + [ + "Reading comprehension (RC) has become a key benchmark for natural language understanding (NLU) systems and a large number of datasets are now available BIBREF0, BIBREF1, BIBREF2. However, these datasets suffer from annotation artifacts and other biases, which allow systems to \u201ccheat\u201d: Instead of learning to read texts, systems learn to exploit these biases and find answers via simple heuristics, such as looking for an entity with a matching semantic type BIBREF3, BIBREF4. To give another example, many RC datasets contain a large number of \u201ceasy\u201d problems that can be solved by looking at the first few words of the question Sugawara2018. In order to provide a reliable measure of progress, an RC dataset thus needs to be robust to such simple heuristics.", + "Towards this goal, two important directions have been investigated. One direction is to improve the dataset itself, for example, so that it requires an RC system to perform multi-hop inferences BIBREF0 or to generate answers BIBREF1. Another direction is to request a system to output additional information about answers. Yang2018HotpotQA:Answering propose HotpotQA, an \u201cexplainable\u201d multi-hop Question Answering (QA) task that requires a system to identify a set of sentences containing supporting evidence for the given answer. We follow the footsteps of Yang2018HotpotQA:Answering and explore an explainable multi-hop QA task.", + "In the community, two important types of explanations have been explored so far BIBREF5: (i) introspective explanation (how a decision is made), and (ii) justification explanation (collections of evidences to support the decision). In this sense, supporting facts in HotpotQA can be categorized as justification explanations. The advantage of using justification explanations as benchmark is that the task can be reduced to a standard classification task, which enables us to adopt standard evaluation metrics (e.g. a classification accuracy). However, this task setting does not evaluate a machine's ability to (i) extract relevant information from justification sentences and (ii) synthesize them to form coherent logical reasoning steps, which are equally important for NLU.", + "To address this issue, we propose RC-QED, an RC task that requires not only the answer to a question, but also an introspective explanation in the form of a natural language derivation (NLD). For example, given the question \u201cWhich record company released the song Barracuda?\u201d and supporting documents shown in Figure FIGREF1, a system needs to give the answer \u201cPortrait Records\u201d and to provide the following NLD: 1.) Barracuda is on Little Queen, and 2.) Little Queen was released by Portrait Records.", + "The main difference between our work and HotpotQA is that they identify a set of sentences $\\lbrace s_2,s_4\\rbrace $, while RC-QED requires a system to generate its derivations in a correct order. This generation task enables us to measure a machine's logical reasoning ability mentioned above. Due to its subjective nature of the natural language derivation task, we evaluate the correctness of derivations generated by a system with multiple reference answers. Our contributions can be summarized as follows:", + "We create a large corpus consisting of 12,000 QA pairs and natural language derivations. The developed crowdsourcing annotation framework can be used for annotating other QA datasets with derivations.", + "Through an experiment using two baseline models, we highlight several challenges of RC-QED.", + "We will make the corpus of reasoning annotations and the baseline system publicly available at https://naoya-i.github.io/rc-qed/." + ], + [ + "We formally define RC-QED as follows:", + "Given: (i) a question $Q$, and (ii) a set $S$ of supporting documents relevant to $Q$;", + "Find: (i) answerability $s \\in \\lbrace \\textsf {Answerable},$ $\\textsf {Unanswerable} \\rbrace $, (ii) an answer $a$, and (iii) a sequence $R$ of derivation steps.", + "We evaluate each prediction with the following evaluation metrics:", + "Answerability: Correctness of model's decision on answerability (i.e. binary classification task) evaluated by Precision/Recall/F1.", + "Answer precision: Correctness of predicted answers (for Answerable predictions only). We follow the standard practice of RC community for evaluation (e.g. an accuracy in the case of multiple choice QA).", + "Derivation precision: Correctness of generated NLDs evaluated by ROUGE-L BIBREF6 (RG-L) and BLEU-4 (BL-4) BIBREF7. We follow the standard practice of evaluation for natural language generation BIBREF1. Derivation steps might be subjective, so we resort to multiple reference answers." + ], + [ + "This paper instantiates RC-QED by employing multiple choice, entity-based multi-hop QA BIBREF0 as a testbed (henceforth, RC-QED$^{\\rm E}$). In entity-based multi-hop QA, machines need to combine relational facts between entities to derive an answer. For example, in Figure FIGREF1, understanding the facts about Barracuda, Little Queen, and Portrait Records stated in each article is required. This design choice restricts a problem domain, but it provides interesting challenges as discussed in Section SECREF46. In addition, such entity-based chaining is known to account for the majority of reasoning types required for multi-hop reasoning BIBREF2.", + "More formally, given (i) a question $Q=(r, q)$ represented by a binary relation $r$ and an entity $q$ (question entity), (ii) relevant articles $S$, and (iii) a set $C$ of candidate entities, systems are required to output (i) an answerability $s \\in \\lbrace \\textsf {Answerable}, \\textsf {Unanswerable} \\rbrace $, (ii) an entity $e \\in C$ (answer entity) that $(q, r, e)$ holds, and (iii) a sequence $R$ of derivation steps as to why $e$ is believed to be an answer. We define derivation steps as an $m$ chain of relational facts to derive an answer, i.e. $(q, r_1, e_1), (e_1, r_2, e_2), ..., (e_{m-1}, r_{m-1}, e_m),$ $(e_m, r_m, e_{m+1}))$. Although we restrict the form of knowledge to entity relations, we use a natural language form to represent $r_i$ rather than a closed vocabulary (see Figure FIGREF1 for an example)." + ], + [ + "To acquire a large-scale corpus of NLDs, we use crowdsourcing (CS). Although CS is a powerful tool for large-scale dataset creation BIBREF2, BIBREF8, quality control for complex tasks is still challenging. We thus carefully design an incentive structure for crowdworkers, following Yang2018HotpotQA:Answering.", + "Initially, we provide crowdworkers with an instruction with example annotations, where we emphasize that they judge the truth of statements solely based on given articles, not based on their own knowledge." + ], + [ + "Given a statement and articles, workers are asked to judge whether the statement can be derived from the articles at three grades: True, Likely (i.e. Answerable), or Unsure (i.e. Unanswerable). If a worker selects Unsure, we ask workers to tell us why they are unsure from two choices (\u201cNot stated in the article\u201d or \u201cOther\u201d)." + ], + [ + "If a worker selects True or Likely in the judgement task, we first ask which sentences in the given articles are justification explanations for a given statement, similarly to HotpotQA BIBREF2. The \u201csummary\u201d text boxes (i.e. NLDs) are then initialized with these selected sentences. We give a \u00a26 bonus to those workers who select True or Likely. To encourage an abstraction of selected sentences, we also introduce a gamification scheme to give a bonus to those who provide shorter NLDs. Specifically, we probabilistically give another \u00a214 bonus to workers according to a score they gain. The score is always shown on top of the screen, and changes according to the length of NLDs they write in real time. To discourage noisy annotations, we also warn crowdworkers that their work would be rejected for noisy submissions. We periodically run simple filtering to exclude noisy crowdworkers (e.g. workers who give more than 50 submissions with the same answers).", + "We deployed the task on Amazon Mechanical Turk (AMT). To see how reasoning varies across workers, we hire 3 crowdworkers per one instance. We hire reliable crowdworkers with $\\ge 5,000$ HITs experiences and an approval rate of $\\ge $ 99.0%, and pay \u00a220 as a reward per instance.", + "Our data collection pipeline is expected to be applicable to other types of QAs other than entity-based multi-hop QA without any significant extensions, because the interface is not specifically designed for entity-centric reasoning." + ], + [ + "Our study uses WikiHop BIBREF0, as it is an entity-based multi-hop QA dataset and has been actively used. We randomly sampled 10,000 instances from 43,738 training instances and 2,000 instances from 5,129 validation instances (i.e. 36,000 annotation tasks were published on AMT). We manually converted structured WikiHop question-answer pairs (e.g. locatedIn(Macchu Picchu, Peru)) into natural language statements (Macchu Picchu is located in Peru) using a simple conversion dictionary.", + "We use supporting documents provided by WikiHop. WikiHop collects supporting documents by finding Wikipedia articles that bridges a question entity $e_i$ and an answer entity $e_j$, where the link between articles is given by a hyperlink." + ], + [ + "Table TABREF17 shows the statistics of responses and example annotations. Table TABREF17 also shows the abstractiveness of annotated NLDs ($a$), namely the number of tokens in an NLD divided by the number of tokens in its corresponding justification sentences. This indicates that annotated NLDs are indeed summarized. See Table TABREF53 in Appendix and Supplementary Material for more results." + ], + [ + "To evaluate the quality of annotation results, we publish another CS task on AMT. We randomly sample 300 True and Likely responses in this evaluation. Given NLDs and a statement, 3 crowdworkers are asked if the NLDs can lead to the statement at four scale levels. If the answer is 4 or 3 (\u201cyes\u201d or \u201clikely\u201d), we additionally asked whether each derivation step can be derived from each supporting document; otherwise we asked them the reasons. For a fair evaluation, we encourage crowdworkers to annotate given NLDs with a lower score by stating that we give a bonus if they found a flaw of reasoning on the CS interface.", + "The evaluation results shown in Table TABREF24 indicate that the annotated NLDs are of high quality (Reachability), and each NLD is properly derived from supporting documents (Derivability).", + "On the other hand, we found the quality of 3-step NLDs is relatively lower than the others. Crowdworkers found that 45.3% of 294 (out of 900) 3-step NLDs has missing steps to derive a statement. Let us consider this example: for annotated NLDs \u201c[1] Kouvola is located in Helsinki. [2] Helsinki is in the region of Uusimaa. [3] Uusimaa borders the regions Southwest Finland, Kymenlaakso and some others.\u201d and for the statement \u201cKouvola is located in Kymenlaakso\u201d, one worker pointed out the missing step \u201cUusimaa is in Kymenlaakso.\u201d. We speculate that greater steps of reasoning make it difficult for crowdworkers to check the correctness of derivations during the writing task." + ], + [ + "For agreement on the number of NLDs, we obtained a Krippendorff's $\\alpha $ of 0.223, indicating a fair agreement BIBREF9.", + "Our manual inspection of the 10 worst disagreements revealed that majority (7/10) come from Unsure v.s. non-Unsure. It also revealed that crowdworkers who labeled non-Unsure are reliable\u20146 out 7 non-Unsure annotations can be judged as correct. This partially confirms the effectiveness of our incentive structure." + ], + [ + "To highlight the challenges and nature of RC-QED$^{\\rm E}$, we create a simple, transparent, and interpretable baseline model.", + "Recent studies on knowledge graph completion (KGC) explore compositional inferences to combat with the sparsity of knowledge bases BIBREF10, BIBREF11, BIBREF12. Given a query triplet $(h, r, t)$ (e.g. (Macchu Picchu, locatedIn, Peru)), a path ranking-based approach for KGC explicitly samples paths between $h$ and $t$ in a knowledge base (e.g. Macchu Picchu\u2014locatedIn\u2014Andes Mountain\u2014countryOf\u2014Peru), and construct a feature vector of these paths. This feature vector is then used to calculate the compatibility between the query triplet and the sampled paths.", + "RC-QED$^{\\rm E}$ can be naturally solved by path ranking-based KGC (PRKGC), where the query triplet and the sampled paths correspond to a question and derivation steps, respectively. PRKGC meets our purposes because of its glassboxness: we can trace the derivation steps of the model easily." + ], + [ + "Given supporting documents $S$, we build a knowledge graph. We first apply a coreference resolver to $S$ and then create a directed graph $G(S)$. Therein, each node represents named entities (NEs) in $S$, and each edge represents textual relations between NEs extracted from $S$. Figure FIGREF27 illustrates an example of $G(S)$ constructed from supporting documents in Figure FIGREF1." + ], + [ + "Given a question $Q=(q, r)$ and a candidate entity $c_i$, we estimate the plausibility of $(q, r, c_i)$ as follows:", + "where $\\sigma $ is a sigmoid function, and $\\mathbf {q, r, c_i}, \\mathbf {\\pi }(q, c_i)$ are vector representations of $q, r, c_i$ and a set $\\pi (q, c_i)$ of shortest paths between $q$ and $c_i$ on $G(S)$. ${\\rm MLP}(\\cdot , \\cdot )$ denotes a multi-layer perceptron. To encode entities into vectors $\\mathbf {q, c_i}$, we use Long-Short Term Memory (LSTM) and take its last hidden state. For example, in Figure FIGREF27, $q =$ Barracuda and $c_i =$ Portrait Records yield $\\pi (q, c_i) = \\lbrace $Barracuda\u2014is the most popular in their album\u2014Little Queen\u2014was released in May 1977 on\u2014Portrait Records, Barracuda\u2014was released from American band Heart\u2014is the second album released by:-1\u2014Little Queen\u2014was released in May 1977 on\u2014Portrait Records$\\rbrace $.", + "To obtain path representations $\\mathbf {\\pi }(q, c_i)$, we attentively aggregate individual path representations: $\\mathbf {\\pi }(q, c_i) = \\sum _j \\alpha _j \\mathbf {\\pi _j}(q, c_i)$, where $\\alpha _j$ is an attention for the $j$-th path. The attention values are calculated as follows: $\\alpha _j = \\exp ({\\rm sc}(q, r, c_i, \\pi _j)) / \\sum _k \\exp ({\\rm sc}(q, r, c_i, \\pi _k))$, where ${\\rm sc}(q, r, c_i, \\pi _j) = {\\rm MLP}(\\mathbf {q}, \\mathbf {r}, \\mathbf {c_i}, \\mathbf {\\pi _j})$. To obtain individual path representations $\\mathbf {\\pi _j}$, we follow toutanova-etal-2015-representing. We use a Bi-LSTM BIBREF13 with mean pooling over timestep in order to encourage similar paths to have similar path representations.", + "For the testing phase, we choose a candidate entity $c_i$ with the maximum probability $P(r|q, c_i)$ as an answer entity, and choose a path $\\pi _j$ with the maximum attention value $\\alpha _j$ as NLDs. To generate NLDs, we simply traverse the path from $q$ to $c_i$ and subsequently concatenate all entities and textual relations as one string. We output Unanswerable when (i) $\\max _{c_i \\in C} P(r|q, c_i) < \\epsilon _k$ or (ii) $G(S)$ has no path between $q$ and all $c_i \\in C$." + ], + [ + "Let $\\mathcal {K}^+$ be a set of question-answer pairs, where each instance consists of a triplet (a query entity $q_i$, a relation $r_i$, an answer entity $a_i$). Similarly, let $\\mathcal {K}^-$ be a set of question-non-answer pairs. We minimize the following binary cross-entropy loss:", + "From the NLD point of view, this is unsupervised training. The model is expected to learn the score function ${\\rm sc(\\cdot )}$ to give higher scores to paths (i.e. NLD steps) that are useful for discriminating correct answers from wrong answers by its own. Highly scored NLDs might be useful for answer classification, but these are not guaranteed to be interpretable to humans." + ], + [ + "To address the above issue, we resort to gold-standard NLDs to guide the path scoring function ${\\rm sc(\\cdot )}$. Let $\\mathcal {D}$ be question-answer pairs coupled with gold-standard NLDs, namely a binary vector $\\mathbf {p}_i$, where the $j$-th value represents whether $j$-th path corresponds to a gold-standard NLD (1) or not (0). We apply the following cross-entropy loss to the path attention:" + ], + [ + "We aggregated crowdsourced annotations obtained in Section SECREF3. As a preprocessing, we converted the NLD annotation to Unsure if the derivation contains the phrase needs to be mentioned. This is due to the fact that annotators misunderstand our instruction. When at least one crowdworker state that a statement is Unsure, then we set the answerability to Unanswerable and discard NLD annotations. Otherwise, we employ all NLD annotations from workers as multiple reference NLDs. The statistics is shown in Table TABREF36.", + "Regarding $\\mathcal {K}^+, \\mathcal {K}^-$, we extracted 867,936 instances from the training set of WikiHop BIBREF0. We reserve 10% of these instances as a validation set to find the best model. For $\\mathcal {D}$, we used Answerable questions in the training set. To create supervision of path (i.e. $\\mathbf {p}_i$), we selected the path that is most similar to all NLD annotations in terms of ROUGE-L F1." + ], + [ + "We used 100-dimensional vectors for entities, relations, and textual relation representations. We initialize these representations with 100-dimensional Glove Embeddings BIBREF14 and fine-tuned them during training. We retain only top-100,000 frequent words as a model vocabulary. We used Bi-LSTM with 50 dimensional hidden state as a textual relation encoder, and an LSTM with 100-dimensional hidden state as an entity encoder. We used the Adam optimizer (default parameters) BIBREF15 with a batch size of 32. We set the answerability threshold $\\epsilon _k = 0.5$." + ], + [ + "To check the integrity of the PRKGC model, we created a simple baseline model (shortest path model). It outputs a candidate entity with the shortest path length from a query entity on $G(S)$ as an answer. Similarly to the PRKGC model, it traverses the path to generate NLDs. It outputs Unanswerable if (i) a query entity is not reachable to any candidate entities on $G(S)$ or (ii) the shortest path length is more than 3." + ], + [ + "As shown in Table TABREF37, the PRKGC models learned to reason over more than simple shortest paths. Yet, the PRKGC model do not give considerably good results, which indicates the non-triviality of RC-QED$^{\\rm E}$. Although the PRKGC model do not receive supervision about human-generated NLDs, paths with the maximum score match human-generated NLDs to some extent.", + "Supervising path attentions (the PRKGC+NS model) is indeed effective for improving the human interpretability of generated NLDs. It also improves the generalization ability of question answering. We speculate that $L_d$ functions as a regularizer, which helps models to learn reasoning that helpful beyond training data. This observation is consistent with previous work where an evidence selection task is learned jointly with a main task BIBREF11, BIBREF2, BIBREF5.", + "As shown in Table TABREF43, as the required derivation step increases, the PRKGC+NS model suffers from predicting answer entities and generating correct NLDs. This indicates that the challenge of RC-QED$^{\\rm E}$ is in how to extract relevant information from supporting documents and synthesize these multiple facts to derive an answer.", + "To obtain further insights, we manually analyzed generated NLDs. Table TABREF44 (a) illustrates a positive example, where the model identifies that altudoceras belongs to pseudogastrioceratinae, and that pseudogastrioceratinae is a subfamily of paragastrioceratidae. Some supporting sentences are already similar to human-generated NLDs, thus simply extracting textual relations works well for some problems.", + "On the other hand, typical derivation error is from non-human readable textual relations. In (b), the model states that bumped has a relationship of \u201c,\u201d with hands up, which is originally extracted from one of supporting sentences It contains the UK Top 60 singles \u201cBumped\u201d, \u201cHands Up (4 Lovers)\u201d and .... This provides a useful clue for answer prediction, but is not suitable as a derivation. One may address this issue by incorporating, for example, a relation extractor or a paraphrasing mechanism using recent advances of conditional language models BIBREF20." + ], + [ + "To check the integrity of our baseline models, we compare our baseline models with existing neural models tailored for QA under the pure WikiHop setting (i.e. evaluation with only an accuracy of predicted answers). Note that these existing models do not output derivations. We thus cannot make a direct comparison, so it servers as a reference purpose. Because WikiHop has no answerability task, we enforced the PRKGC model to always output answers. As shown in Table TABREF45, the PRKGC models achieve a comparable performance to other sophisticated neural models." + ], + [ + "There exists few RC datasets annotated with explanations (Table TABREF50). The most similar work to ours is Science QA dataset BIBREF21, BIBREF22, BIBREF23, which provides a small set of NLDs annotated for analysis purposes. By developing the scalable crowdsourcing framework, our work provides one order-of-magnitude larger NLDs which can be used as a benchmark more reliably. In addition, it provides the community with new types of challenges not included in HotpotQA." + ], + [ + "There is a large body of work on analyzing the nature of RC datasets, motivated by the question to what degree RC models understand natural language BIBREF3, BIBREF4. Several studies suggest that current RC datasets have unintended bias, which enables RC systems to rely on a cheap heuristics to answer questions. For instance, Sugawara2018 show that some of these RC datasets contain a large number of \u201ceasy\u201d questions that can be solved by a cheap heuristics (e.g. by looking at a first few tokens of questions). Responding to their findings, we take a step further and explore the new task of RC that requires RC systems to give introspective explanations as well as answers. In addition, recent studies show that current RC models and NLP models are vulnerable to adversarial examples BIBREF29, BIBREF30, BIBREF31. Explicit modeling of NLDs is expected to reguralize RC models, which could prevent RC models' strong dependence on unintended bias in training data (e.g. annotation artifact) BIBREF32, BIBREF8, BIBREF2, BIBREF5, as partially confirmed in Section SECREF46." + ], + [ + "There are existing NLP tasks that require models to output explanations (Table TABREF50). FEVER BIBREF25 requires a system to judge the \u201cfactness\u201d of a claim as well as to identify justification sentences. As discussed earlier, we take a step further from justification explanations to provide new challenges for NLU.", + "Several datasets are annotated with introspective explanations, ranging from textual entailments BIBREF8 to argumentative texts BIBREF26, BIBREF27, BIBREF33. All these datasets offer the classification task of single sentences or sentence pairs. The uniqueness of our dataset is that it measures a machine's ability to extract relevant information from a set of documents and to build coherent logical reasoning steps." + ], + [ + "Towards RC models that can perform correct reasoning, we have proposed RC-QED that requires a system to output its introspective explanations, as well as answers. Instantiating RC-QED with entity-based multi-hop QA (RC-QED$^{\\rm E}$), we have created a large-scale corpus of NLDs. The developed crowdsourcing annotation framework can be used for annotating other QA datasets with derivations. Our experiments using two simple baseline models have demonstrated that RC-QED$^{\\rm E}$ is a non-trivial task, and that it indeed provides a challenging task of extracting and synthesizing relevant facts from supporting documents. We will make the corpus of reasoning annotations and baseline systems publicly available at https://naoya-i.github.io/rc-qed/.", + "One immediate future work is to expand the annotation to non-entity-based multi-hop QA datasets such as HotpotQA BIBREF2. For modeling, we plan to incorporate a generative mechanism based on recent advances in conditional language modeling." + ], + [ + "Table TABREF53 shows examples of crowdsourced annotations." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0260/instruction.md b/qasper-0260/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..27ebf062254f55b7822847f872c920f634fa70f0 --- /dev/null +++ b/qasper-0260/instruction.md @@ -0,0 +1,148 @@ +Name of Paper: RC-QED: Evaluating Natural Language Derivations in Multi-Hop Reading Comprehension + +Question: Did they use any crowdsourcing platform? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Task formulation: RC-QED ::: Input, output, and evaluation metrics", + "Task formulation: RC-QED ::: RC-QED@!START@$^{\\rm E}$@!END@", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Crowdsourcing interface", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Crowdsourcing interface ::: Judgement task (Figure @!START@UID13@!END@).", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Crowdsourcing interface ::: Derivation task (Figure @!START@UID14@!END@).", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Dataset", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Results", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Results ::: Quality", + "Data collection for RC-QED@!START@$^{\\rm E}$@!END@ ::: Results ::: Agreement", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Knowledge graph construction", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Path ranking-based KGC (PRKGC)", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Training", + "Baseline RC-QED@!START@$^{\\rm E}$@!END@ model ::: Training ::: Semi-supervising derivations", + "Experiments ::: Settings ::: Dataset", + "Experiments ::: Settings ::: Hyperparameters", + "Experiments ::: Settings ::: Baseline", + "Experiments ::: Results and discussion", + "Experiments ::: Results and discussion ::: QA performance.", + "Related work ::: RC datasets with explanations", + "Related work ::: Analysis of RC models and datasets", + "Related work ::: Other NLP corpora annotated with explanations", + "Conclusions", + "Example annotations" + ], + "paragraphs": [ + [ + "Reading comprehension (RC) has become a key benchmark for natural language understanding (NLU) systems and a large number of datasets are now available BIBREF0, BIBREF1, BIBREF2. However, these datasets suffer from annotation artifacts and other biases, which allow systems to \u201ccheat\u201d: Instead of learning to read texts, systems learn to exploit these biases and find answers via simple heuristics, such as looking for an entity with a matching semantic type BIBREF3, BIBREF4. To give another example, many RC datasets contain a large number of \u201ceasy\u201d problems that can be solved by looking at the first few words of the question Sugawara2018. In order to provide a reliable measure of progress, an RC dataset thus needs to be robust to such simple heuristics.", + "Towards this goal, two important directions have been investigated. One direction is to improve the dataset itself, for example, so that it requires an RC system to perform multi-hop inferences BIBREF0 or to generate answers BIBREF1. Another direction is to request a system to output additional information about answers. Yang2018HotpotQA:Answering propose HotpotQA, an \u201cexplainable\u201d multi-hop Question Answering (QA) task that requires a system to identify a set of sentences containing supporting evidence for the given answer. We follow the footsteps of Yang2018HotpotQA:Answering and explore an explainable multi-hop QA task.", + "In the community, two important types of explanations have been explored so far BIBREF5: (i) introspective explanation (how a decision is made), and (ii) justification explanation (collections of evidences to support the decision). In this sense, supporting facts in HotpotQA can be categorized as justification explanations. The advantage of using justification explanations as benchmark is that the task can be reduced to a standard classification task, which enables us to adopt standard evaluation metrics (e.g. a classification accuracy). However, this task setting does not evaluate a machine's ability to (i) extract relevant information from justification sentences and (ii) synthesize them to form coherent logical reasoning steps, which are equally important for NLU.", + "To address this issue, we propose RC-QED, an RC task that requires not only the answer to a question, but also an introspective explanation in the form of a natural language derivation (NLD). For example, given the question \u201cWhich record company released the song Barracuda?\u201d and supporting documents shown in Figure FIGREF1, a system needs to give the answer \u201cPortrait Records\u201d and to provide the following NLD: 1.) Barracuda is on Little Queen, and 2.) Little Queen was released by Portrait Records.", + "The main difference between our work and HotpotQA is that they identify a set of sentences $\\lbrace s_2,s_4\\rbrace $, while RC-QED requires a system to generate its derivations in a correct order. This generation task enables us to measure a machine's logical reasoning ability mentioned above. Due to its subjective nature of the natural language derivation task, we evaluate the correctness of derivations generated by a system with multiple reference answers. Our contributions can be summarized as follows:", + "We create a large corpus consisting of 12,000 QA pairs and natural language derivations. The developed crowdsourcing annotation framework can be used for annotating other QA datasets with derivations.", + "Through an experiment using two baseline models, we highlight several challenges of RC-QED.", + "We will make the corpus of reasoning annotations and the baseline system publicly available at https://naoya-i.github.io/rc-qed/." + ], + [ + "We formally define RC-QED as follows:", + "Given: (i) a question $Q$, and (ii) a set $S$ of supporting documents relevant to $Q$;", + "Find: (i) answerability $s \\in \\lbrace \\textsf {Answerable},$ $\\textsf {Unanswerable} \\rbrace $, (ii) an answer $a$, and (iii) a sequence $R$ of derivation steps.", + "We evaluate each prediction with the following evaluation metrics:", + "Answerability: Correctness of model's decision on answerability (i.e. binary classification task) evaluated by Precision/Recall/F1.", + "Answer precision: Correctness of predicted answers (for Answerable predictions only). We follow the standard practice of RC community for evaluation (e.g. an accuracy in the case of multiple choice QA).", + "Derivation precision: Correctness of generated NLDs evaluated by ROUGE-L BIBREF6 (RG-L) and BLEU-4 (BL-4) BIBREF7. We follow the standard practice of evaluation for natural language generation BIBREF1. Derivation steps might be subjective, so we resort to multiple reference answers." + ], + [ + "This paper instantiates RC-QED by employing multiple choice, entity-based multi-hop QA BIBREF0 as a testbed (henceforth, RC-QED$^{\\rm E}$). In entity-based multi-hop QA, machines need to combine relational facts between entities to derive an answer. For example, in Figure FIGREF1, understanding the facts about Barracuda, Little Queen, and Portrait Records stated in each article is required. This design choice restricts a problem domain, but it provides interesting challenges as discussed in Section SECREF46. In addition, such entity-based chaining is known to account for the majority of reasoning types required for multi-hop reasoning BIBREF2.", + "More formally, given (i) a question $Q=(r, q)$ represented by a binary relation $r$ and an entity $q$ (question entity), (ii) relevant articles $S$, and (iii) a set $C$ of candidate entities, systems are required to output (i) an answerability $s \\in \\lbrace \\textsf {Answerable}, \\textsf {Unanswerable} \\rbrace $, (ii) an entity $e \\in C$ (answer entity) that $(q, r, e)$ holds, and (iii) a sequence $R$ of derivation steps as to why $e$ is believed to be an answer. We define derivation steps as an $m$ chain of relational facts to derive an answer, i.e. $(q, r_1, e_1), (e_1, r_2, e_2), ..., (e_{m-1}, r_{m-1}, e_m),$ $(e_m, r_m, e_{m+1}))$. Although we restrict the form of knowledge to entity relations, we use a natural language form to represent $r_i$ rather than a closed vocabulary (see Figure FIGREF1 for an example)." + ], + [ + "To acquire a large-scale corpus of NLDs, we use crowdsourcing (CS). Although CS is a powerful tool for large-scale dataset creation BIBREF2, BIBREF8, quality control for complex tasks is still challenging. We thus carefully design an incentive structure for crowdworkers, following Yang2018HotpotQA:Answering.", + "Initially, we provide crowdworkers with an instruction with example annotations, where we emphasize that they judge the truth of statements solely based on given articles, not based on their own knowledge." + ], + [ + "Given a statement and articles, workers are asked to judge whether the statement can be derived from the articles at three grades: True, Likely (i.e. Answerable), or Unsure (i.e. Unanswerable). If a worker selects Unsure, we ask workers to tell us why they are unsure from two choices (\u201cNot stated in the article\u201d or \u201cOther\u201d)." + ], + [ + "If a worker selects True or Likely in the judgement task, we first ask which sentences in the given articles are justification explanations for a given statement, similarly to HotpotQA BIBREF2. The \u201csummary\u201d text boxes (i.e. NLDs) are then initialized with these selected sentences. We give a \u00a26 bonus to those workers who select True or Likely. To encourage an abstraction of selected sentences, we also introduce a gamification scheme to give a bonus to those who provide shorter NLDs. Specifically, we probabilistically give another \u00a214 bonus to workers according to a score they gain. The score is always shown on top of the screen, and changes according to the length of NLDs they write in real time. To discourage noisy annotations, we also warn crowdworkers that their work would be rejected for noisy submissions. We periodically run simple filtering to exclude noisy crowdworkers (e.g. workers who give more than 50 submissions with the same answers).", + "We deployed the task on Amazon Mechanical Turk (AMT). To see how reasoning varies across workers, we hire 3 crowdworkers per one instance. We hire reliable crowdworkers with $\\ge 5,000$ HITs experiences and an approval rate of $\\ge $ 99.0%, and pay \u00a220 as a reward per instance.", + "Our data collection pipeline is expected to be applicable to other types of QAs other than entity-based multi-hop QA without any significant extensions, because the interface is not specifically designed for entity-centric reasoning." + ], + [ + "Our study uses WikiHop BIBREF0, as it is an entity-based multi-hop QA dataset and has been actively used. We randomly sampled 10,000 instances from 43,738 training instances and 2,000 instances from 5,129 validation instances (i.e. 36,000 annotation tasks were published on AMT). We manually converted structured WikiHop question-answer pairs (e.g. locatedIn(Macchu Picchu, Peru)) into natural language statements (Macchu Picchu is located in Peru) using a simple conversion dictionary.", + "We use supporting documents provided by WikiHop. WikiHop collects supporting documents by finding Wikipedia articles that bridges a question entity $e_i$ and an answer entity $e_j$, where the link between articles is given by a hyperlink." + ], + [ + "Table TABREF17 shows the statistics of responses and example annotations. Table TABREF17 also shows the abstractiveness of annotated NLDs ($a$), namely the number of tokens in an NLD divided by the number of tokens in its corresponding justification sentences. This indicates that annotated NLDs are indeed summarized. See Table TABREF53 in Appendix and Supplementary Material for more results." + ], + [ + "To evaluate the quality of annotation results, we publish another CS task on AMT. We randomly sample 300 True and Likely responses in this evaluation. Given NLDs and a statement, 3 crowdworkers are asked if the NLDs can lead to the statement at four scale levels. If the answer is 4 or 3 (\u201cyes\u201d or \u201clikely\u201d), we additionally asked whether each derivation step can be derived from each supporting document; otherwise we asked them the reasons. For a fair evaluation, we encourage crowdworkers to annotate given NLDs with a lower score by stating that we give a bonus if they found a flaw of reasoning on the CS interface.", + "The evaluation results shown in Table TABREF24 indicate that the annotated NLDs are of high quality (Reachability), and each NLD is properly derived from supporting documents (Derivability).", + "On the other hand, we found the quality of 3-step NLDs is relatively lower than the others. Crowdworkers found that 45.3% of 294 (out of 900) 3-step NLDs has missing steps to derive a statement. Let us consider this example: for annotated NLDs \u201c[1] Kouvola is located in Helsinki. [2] Helsinki is in the region of Uusimaa. [3] Uusimaa borders the regions Southwest Finland, Kymenlaakso and some others.\u201d and for the statement \u201cKouvola is located in Kymenlaakso\u201d, one worker pointed out the missing step \u201cUusimaa is in Kymenlaakso.\u201d. We speculate that greater steps of reasoning make it difficult for crowdworkers to check the correctness of derivations during the writing task." + ], + [ + "For agreement on the number of NLDs, we obtained a Krippendorff's $\\alpha $ of 0.223, indicating a fair agreement BIBREF9.", + "Our manual inspection of the 10 worst disagreements revealed that majority (7/10) come from Unsure v.s. non-Unsure. It also revealed that crowdworkers who labeled non-Unsure are reliable\u20146 out 7 non-Unsure annotations can be judged as correct. This partially confirms the effectiveness of our incentive structure." + ], + [ + "To highlight the challenges and nature of RC-QED$^{\\rm E}$, we create a simple, transparent, and interpretable baseline model.", + "Recent studies on knowledge graph completion (KGC) explore compositional inferences to combat with the sparsity of knowledge bases BIBREF10, BIBREF11, BIBREF12. Given a query triplet $(h, r, t)$ (e.g. (Macchu Picchu, locatedIn, Peru)), a path ranking-based approach for KGC explicitly samples paths between $h$ and $t$ in a knowledge base (e.g. Macchu Picchu\u2014locatedIn\u2014Andes Mountain\u2014countryOf\u2014Peru), and construct a feature vector of these paths. This feature vector is then used to calculate the compatibility between the query triplet and the sampled paths.", + "RC-QED$^{\\rm E}$ can be naturally solved by path ranking-based KGC (PRKGC), where the query triplet and the sampled paths correspond to a question and derivation steps, respectively. PRKGC meets our purposes because of its glassboxness: we can trace the derivation steps of the model easily." + ], + [ + "Given supporting documents $S$, we build a knowledge graph. We first apply a coreference resolver to $S$ and then create a directed graph $G(S)$. Therein, each node represents named entities (NEs) in $S$, and each edge represents textual relations between NEs extracted from $S$. Figure FIGREF27 illustrates an example of $G(S)$ constructed from supporting documents in Figure FIGREF1." + ], + [ + "Given a question $Q=(q, r)$ and a candidate entity $c_i$, we estimate the plausibility of $(q, r, c_i)$ as follows:", + "where $\\sigma $ is a sigmoid function, and $\\mathbf {q, r, c_i}, \\mathbf {\\pi }(q, c_i)$ are vector representations of $q, r, c_i$ and a set $\\pi (q, c_i)$ of shortest paths between $q$ and $c_i$ on $G(S)$. ${\\rm MLP}(\\cdot , \\cdot )$ denotes a multi-layer perceptron. To encode entities into vectors $\\mathbf {q, c_i}$, we use Long-Short Term Memory (LSTM) and take its last hidden state. For example, in Figure FIGREF27, $q =$ Barracuda and $c_i =$ Portrait Records yield $\\pi (q, c_i) = \\lbrace $Barracuda\u2014is the most popular in their album\u2014Little Queen\u2014was released in May 1977 on\u2014Portrait Records, Barracuda\u2014was released from American band Heart\u2014is the second album released by:-1\u2014Little Queen\u2014was released in May 1977 on\u2014Portrait Records$\\rbrace $.", + "To obtain path representations $\\mathbf {\\pi }(q, c_i)$, we attentively aggregate individual path representations: $\\mathbf {\\pi }(q, c_i) = \\sum _j \\alpha _j \\mathbf {\\pi _j}(q, c_i)$, where $\\alpha _j$ is an attention for the $j$-th path. The attention values are calculated as follows: $\\alpha _j = \\exp ({\\rm sc}(q, r, c_i, \\pi _j)) / \\sum _k \\exp ({\\rm sc}(q, r, c_i, \\pi _k))$, where ${\\rm sc}(q, r, c_i, \\pi _j) = {\\rm MLP}(\\mathbf {q}, \\mathbf {r}, \\mathbf {c_i}, \\mathbf {\\pi _j})$. To obtain individual path representations $\\mathbf {\\pi _j}$, we follow toutanova-etal-2015-representing. We use a Bi-LSTM BIBREF13 with mean pooling over timestep in order to encourage similar paths to have similar path representations.", + "For the testing phase, we choose a candidate entity $c_i$ with the maximum probability $P(r|q, c_i)$ as an answer entity, and choose a path $\\pi _j$ with the maximum attention value $\\alpha _j$ as NLDs. To generate NLDs, we simply traverse the path from $q$ to $c_i$ and subsequently concatenate all entities and textual relations as one string. We output Unanswerable when (i) $\\max _{c_i \\in C} P(r|q, c_i) < \\epsilon _k$ or (ii) $G(S)$ has no path between $q$ and all $c_i \\in C$." + ], + [ + "Let $\\mathcal {K}^+$ be a set of question-answer pairs, where each instance consists of a triplet (a query entity $q_i$, a relation $r_i$, an answer entity $a_i$). Similarly, let $\\mathcal {K}^-$ be a set of question-non-answer pairs. We minimize the following binary cross-entropy loss:", + "From the NLD point of view, this is unsupervised training. The model is expected to learn the score function ${\\rm sc(\\cdot )}$ to give higher scores to paths (i.e. NLD steps) that are useful for discriminating correct answers from wrong answers by its own. Highly scored NLDs might be useful for answer classification, but these are not guaranteed to be interpretable to humans." + ], + [ + "To address the above issue, we resort to gold-standard NLDs to guide the path scoring function ${\\rm sc(\\cdot )}$. Let $\\mathcal {D}$ be question-answer pairs coupled with gold-standard NLDs, namely a binary vector $\\mathbf {p}_i$, where the $j$-th value represents whether $j$-th path corresponds to a gold-standard NLD (1) or not (0). We apply the following cross-entropy loss to the path attention:" + ], + [ + "We aggregated crowdsourced annotations obtained in Section SECREF3. As a preprocessing, we converted the NLD annotation to Unsure if the derivation contains the phrase needs to be mentioned. This is due to the fact that annotators misunderstand our instruction. When at least one crowdworker state that a statement is Unsure, then we set the answerability to Unanswerable and discard NLD annotations. Otherwise, we employ all NLD annotations from workers as multiple reference NLDs. The statistics is shown in Table TABREF36.", + "Regarding $\\mathcal {K}^+, \\mathcal {K}^-$, we extracted 867,936 instances from the training set of WikiHop BIBREF0. We reserve 10% of these instances as a validation set to find the best model. For $\\mathcal {D}$, we used Answerable questions in the training set. To create supervision of path (i.e. $\\mathbf {p}_i$), we selected the path that is most similar to all NLD annotations in terms of ROUGE-L F1." + ], + [ + "We used 100-dimensional vectors for entities, relations, and textual relation representations. We initialize these representations with 100-dimensional Glove Embeddings BIBREF14 and fine-tuned them during training. We retain only top-100,000 frequent words as a model vocabulary. We used Bi-LSTM with 50 dimensional hidden state as a textual relation encoder, and an LSTM with 100-dimensional hidden state as an entity encoder. We used the Adam optimizer (default parameters) BIBREF15 with a batch size of 32. We set the answerability threshold $\\epsilon _k = 0.5$." + ], + [ + "To check the integrity of the PRKGC model, we created a simple baseline model (shortest path model). It outputs a candidate entity with the shortest path length from a query entity on $G(S)$ as an answer. Similarly to the PRKGC model, it traverses the path to generate NLDs. It outputs Unanswerable if (i) a query entity is not reachable to any candidate entities on $G(S)$ or (ii) the shortest path length is more than 3." + ], + [ + "As shown in Table TABREF37, the PRKGC models learned to reason over more than simple shortest paths. Yet, the PRKGC model do not give considerably good results, which indicates the non-triviality of RC-QED$^{\\rm E}$. Although the PRKGC model do not receive supervision about human-generated NLDs, paths with the maximum score match human-generated NLDs to some extent.", + "Supervising path attentions (the PRKGC+NS model) is indeed effective for improving the human interpretability of generated NLDs. It also improves the generalization ability of question answering. We speculate that $L_d$ functions as a regularizer, which helps models to learn reasoning that helpful beyond training data. This observation is consistent with previous work where an evidence selection task is learned jointly with a main task BIBREF11, BIBREF2, BIBREF5.", + "As shown in Table TABREF43, as the required derivation step increases, the PRKGC+NS model suffers from predicting answer entities and generating correct NLDs. This indicates that the challenge of RC-QED$^{\\rm E}$ is in how to extract relevant information from supporting documents and synthesize these multiple facts to derive an answer.", + "To obtain further insights, we manually analyzed generated NLDs. Table TABREF44 (a) illustrates a positive example, where the model identifies that altudoceras belongs to pseudogastrioceratinae, and that pseudogastrioceratinae is a subfamily of paragastrioceratidae. Some supporting sentences are already similar to human-generated NLDs, thus simply extracting textual relations works well for some problems.", + "On the other hand, typical derivation error is from non-human readable textual relations. In (b), the model states that bumped has a relationship of \u201c,\u201d with hands up, which is originally extracted from one of supporting sentences It contains the UK Top 60 singles \u201cBumped\u201d, \u201cHands Up (4 Lovers)\u201d and .... This provides a useful clue for answer prediction, but is not suitable as a derivation. One may address this issue by incorporating, for example, a relation extractor or a paraphrasing mechanism using recent advances of conditional language models BIBREF20." + ], + [ + "To check the integrity of our baseline models, we compare our baseline models with existing neural models tailored for QA under the pure WikiHop setting (i.e. evaluation with only an accuracy of predicted answers). Note that these existing models do not output derivations. We thus cannot make a direct comparison, so it servers as a reference purpose. Because WikiHop has no answerability task, we enforced the PRKGC model to always output answers. As shown in Table TABREF45, the PRKGC models achieve a comparable performance to other sophisticated neural models." + ], + [ + "There exists few RC datasets annotated with explanations (Table TABREF50). The most similar work to ours is Science QA dataset BIBREF21, BIBREF22, BIBREF23, which provides a small set of NLDs annotated for analysis purposes. By developing the scalable crowdsourcing framework, our work provides one order-of-magnitude larger NLDs which can be used as a benchmark more reliably. In addition, it provides the community with new types of challenges not included in HotpotQA." + ], + [ + "There is a large body of work on analyzing the nature of RC datasets, motivated by the question to what degree RC models understand natural language BIBREF3, BIBREF4. Several studies suggest that current RC datasets have unintended bias, which enables RC systems to rely on a cheap heuristics to answer questions. For instance, Sugawara2018 show that some of these RC datasets contain a large number of \u201ceasy\u201d questions that can be solved by a cheap heuristics (e.g. by looking at a first few tokens of questions). Responding to their findings, we take a step further and explore the new task of RC that requires RC systems to give introspective explanations as well as answers. In addition, recent studies show that current RC models and NLP models are vulnerable to adversarial examples BIBREF29, BIBREF30, BIBREF31. Explicit modeling of NLDs is expected to reguralize RC models, which could prevent RC models' strong dependence on unintended bias in training data (e.g. annotation artifact) BIBREF32, BIBREF8, BIBREF2, BIBREF5, as partially confirmed in Section SECREF46." + ], + [ + "There are existing NLP tasks that require models to output explanations (Table TABREF50). FEVER BIBREF25 requires a system to judge the \u201cfactness\u201d of a claim as well as to identify justification sentences. As discussed earlier, we take a step further from justification explanations to provide new challenges for NLU.", + "Several datasets are annotated with introspective explanations, ranging from textual entailments BIBREF8 to argumentative texts BIBREF26, BIBREF27, BIBREF33. All these datasets offer the classification task of single sentences or sentence pairs. The uniqueness of our dataset is that it measures a machine's ability to extract relevant information from a set of documents and to build coherent logical reasoning steps." + ], + [ + "Towards RC models that can perform correct reasoning, we have proposed RC-QED that requires a system to output its introspective explanations, as well as answers. Instantiating RC-QED with entity-based multi-hop QA (RC-QED$^{\\rm E}$), we have created a large-scale corpus of NLDs. The developed crowdsourcing annotation framework can be used for annotating other QA datasets with derivations. Our experiments using two simple baseline models have demonstrated that RC-QED$^{\\rm E}$ is a non-trivial task, and that it indeed provides a challenging task of extracting and synthesizing relevant facts from supporting documents. We will make the corpus of reasoning annotations and baseline systems publicly available at https://naoya-i.github.io/rc-qed/.", + "One immediate future work is to expand the annotation to non-entity-based multi-hop QA datasets such as HotpotQA BIBREF2. For modeling, we plan to incorporate a generative mechanism based on recent advances in conditional language modeling." + ], + [ + "Table TABREF53 shows examples of crowdsourced annotations." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0267/instruction.md b/qasper-0267/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..6a8a841a3a2ed3fa36ebb073c4970c14c66af6f8 --- /dev/null +++ b/qasper-0267/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Event Outcome Prediction using Sentiment Analysis and Crowd Wisdom in Microblog Feeds + +Question: How do you establish the ground truth of who won a debate? \ No newline at end of file diff --git a/qasper-0293/instruction.md b/qasper-0293/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c35fefd6a53353a76f73d8e70a3541698de432a8 --- /dev/null +++ b/qasper-0293/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: An Unsupervised Word Sense Disambiguation System for Under-Resourced Languages + +Question: What measure of semantic similarity is used? \ No newline at end of file diff --git a/qasper-0294/instruction.md b/qasper-0294/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..3c0aa5b9551b934d7d628cc96b3a06ce4f0da2ba --- /dev/null +++ b/qasper-0294/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Quasar: Datasets for Question Answering by Search and Reading + +Question: Which retrieval system was used for baselines? \ No newline at end of file diff --git a/qasper-0409/instruction.md b/qasper-0409/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..60ddde21607858d7ecde167df8e1d569f5b6b4d0 --- /dev/null +++ b/qasper-0409/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: CN-CELEB: a challenging Chinese speaker recognition dataset + +Question: What was the performance of both approaches on their dataset? \ No newline at end of file diff --git a/qasper-0413/instruction.md b/qasper-0413/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..94db486cd6f77dd2acde2254bae67f71bf008c05 --- /dev/null +++ b/qasper-0413/instruction.md @@ -0,0 +1,84 @@ +Name of Paper: CN-CELEB: a challenging Chinese speaker recognition dataset + +Question: Which of the two speech recognition models works better overall on CN-Celeb? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "The CN-Celeb dataset ::: Data description", + "The CN-Celeb dataset ::: Challenges with CN-Celeb", + "The CN-Celeb dataset ::: Collection pipeline", + "Experiments on speaker recognition", + "Experiments on speaker recognition ::: Data", + "Experiments on speaker recognition ::: Settings", + "Experiments on speaker recognition ::: Basic results", + "Experiments on speaker recognition ::: Further comparison", + "Conclusions" + ], + "paragraphs": [ + [ + "Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1.", + "Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions.", + "The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions.", + "To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection.", + "In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features:", + "CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons.", + "CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging.", + "CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check.", + "The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper." + ], + [ + "The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check.", + "In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances." + ], + [ + "Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows.", + "Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh.", + "A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres.", + "Most of speakers have different genres of utterances, which results in significant variation in speaking styles.", + "The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems.", + "Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision." + ], + [ + "CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process.", + "Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision.", + "The detailed steps of the collection process are summarized as follows.", + "STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered.", + "STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres.", + "STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams.", + "STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system.", + "STEP 5. Double check by speaker recognition.", + "Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely.", + "STEP 6. Human check.", + "The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours." + ], + [ + "In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets." + ], + [ + "VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW).", + "CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set." + ], + [ + "Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10.", + "For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4.", + "For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details." + ], + [ + "We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E).", + "The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW)." + ], + [ + "To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training.", + "The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results.", + "Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb." + ], + [ + "We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0414/instruction.md b/qasper-0414/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..911700039aed533f395af6afd50c4da4815a8d67 --- /dev/null +++ b/qasper-0414/instruction.md @@ -0,0 +1,84 @@ +Name of Paper: CN-CELEB: a challenging Chinese speaker recognition dataset + +Question: By how much is performance on CN-Celeb inferior to performance on VoxCeleb? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "The CN-Celeb dataset ::: Data description", + "The CN-Celeb dataset ::: Challenges with CN-Celeb", + "The CN-Celeb dataset ::: Collection pipeline", + "Experiments on speaker recognition", + "Experiments on speaker recognition ::: Data", + "Experiments on speaker recognition ::: Settings", + "Experiments on speaker recognition ::: Basic results", + "Experiments on speaker recognition ::: Further comparison", + "Conclusions" + ], + "paragraphs": [ + [ + "Speaker recognition including identification and verification, aims to recognize claimed identities of speakers. After decades of research, performance of speaker recognition systems has been vastly improved, and the technique has been deployed to a wide range of practical applications. Nevertheless, the present speaker recognition approaches are still far from reliable in unconstrained conditions where uncertainties within the speech recordings could be arbitrary. These uncertainties might be caused by multiple factors, including free text, multiple channels, environmental noises, speaking styles, and physiological status. These uncertainties make the speaker recognition task highly challenging BIBREF0, BIBREF1.", + "Researchers have devoted much effort to address the difficulties in unconstrained conditions. Early methods are based on probabilistic models that treat these uncertainties as an additive Gaussian noise. JFA BIBREF2, BIBREF3 and PLDA BIBREF4 are the most famous among such models. These models, however, are shallow and linear, and therefore cannot deal with the complexity of real-life applications. Recent advance in deep learning methods offers a new opportunity BIBREF5, BIBREF6, BIBREF7, BIBREF8. Resorting to the power of deep neural networks (DNNs) in representation learning, these methods can remove unwanted uncertainties by propagating speech signals through the DNN layer by layer and retain speaker-relevant features only BIBREF9. Significant improvement in robustness has been achieved by the DNN-based approach BIBREF10, which makes it more suitable for applications in unconstrained conditions.", + "The success of DNN-based methods, however, largely relies on a large amount of data, in particular data that involve the true complexity in unconstrained conditions. Unfortunately, most existing datasets for speaker recognition are collected in constrained conditions, where the acoustic environment, channel and speaking style do not change significantly for each speaker BIBREF11, BIBREF12, BIBREF13. These datasets tend to deliver over optimistic performance and do not meet the request of research on speaker recognition in unconstrained conditions.", + "To address this shortage in datasets, researchers have started to collect data `in the wild'. The most successful `wild' dataset may be VoxCeleb BIBREF14, BIBREF15, which contains millions of utterances from over thousands of speakers. The utterances were collected from open-source media using a fully automated pipeline based on computer vision techniques, in particular face detection, tracking and recognition, plus video-audio synchronization. The automated pipeline is almost costless, and thus greatly improves the efficiency of data collection.", + "In this paper, we re-implement the automated pipeline of VoxCeleb and collect a new large-scale speaker dataset, named CN-Celeb. Compared with VoxCeleb, CN-Celeb has three distinct features:", + "CN-Celeb specially focuses on Chinese celebrities, and contains more than $130,000$ utterances from $1,000$ persons.", + "CN-Celeb covers more genres of speech. We intentionally collected data from 11 genres, including entertainment, interview, singing, play, movie, vlog, live broadcast, speech, drama, recitation and advertisement. The speech of a particular speaker may be in more than 5 genres. As a comparison, most of the utterances in VoxCeleb were extracted from interview videos. The diversity in genres makes our database more representative for the true scenarios in unconstrained conditions, but also more challenging.", + "CN-Celeb is not fully automated, but involves human check. We found that more complex the genre is, more errors the automated pipeline tends to produce. Ironically, the error-pron segments could be highly valuable as they tend to be boundary samples. We therefore choose a two-stage strategy that employs the automated pipeline to perform pre-selection, and then perform human check.", + "The rest of the paper is organized as follows. Section SECREF2 presents a detailed description for CN-Celeb, and Section SECREF3 presents more quantitative comparisons between CN-Celeb and VoxCeleb on the speaker recognition task. Section SECREF4 concludes the entire paper." + ], + [ + "The original purpose of the CN-Celeb dataset is to investigate the true difficulties of speaker recognition techniques in unconstrained conditions, and provide a resource for researchers to build prototype systems and evaluate the performance. Ideally, it can be used as a standalone data source, and can be also used with other datasets together, in particular VoxCeleb which is free and large. For this reason, CN-Celeb tries to be distinguished from but also complementary to VoxCeleb from the beginning of the design. This leads to three features that we have discussed in the previous section: Chinese focused, complex genres, and quality guarantee by human check.", + "In summary, CN-Celeb contains over $130,000$ utterances from $1,000$ Chinese celebrities. It covers 11 genres and the total amount of speech waveforms is 274 hours. Table TABREF5 gives the data distribution over the genres, and Table TABREF6 presents the data distribution over the length of utterances." + ], + [ + "Table TABREF13 summarizes the main difference between CN-Celeb and VoxCeleb. Compared to VoxCeleb, CN-Celeb is a more complex dataset and more challenging for speaker recognition research. More details of these challenges are as follows.", + "Most of the utterances involve real-world noise, including ambient noise, background babbling, music, cheers and laugh.", + "A certain amount of utterances involve strong and overlapped background speakers, especially in the dram and movie genres.", + "Most of speakers have different genres of utterances, which results in significant variation in speaking styles.", + "The utterances of the same speaker may be recorded at different time and with different devices, leading to serious cross-time and cross-channel problems.", + "Most of the utterances are short, which meets the scenarios of most real applications but leads to unreliable decision." + ], + [ + "CN-Celeb was collected following a two-stage strategy: firstly we used an automated pipeline to extract potential segments of the Person of Interest (POI), and then applied a human check to remove incorrect segments. This process is much faster than purely human-based segmentation, and reduces errors caused by a purely automated process.", + "Briefly, the automated pipeline we used is similar to the one used to collect VoxCeleb1 BIBREF14 and VoxCeleb2 BIBREF15, though we made some modification to increase efficiency and precision. Especially, we introduced a new face-speaker double check step that fused the information from both the image and speech signals to increase the recall rate while maintaining the precision.", + "The detailed steps of the collection process are summarized as follows.", + "STEP 1. POI list design. We manually selected $1,000$ Chinese celebrities as our target speakers. These speakers were mostly from the entertainment sector, such as singers, drama actors/actrees, news reporters, interviewers. Region diversity was also taken into account so that variation in accent was covered.", + "STEP 2. Pictures and videos download. Pictures and videos of the $1,000$ POIs were downloaded from the data source (https://www.bilibili.com/) by searching for the names of the persons. In order to specify that we were searching for POI names, the word `human' was added in the search queries. The downloaded videos were manually examined and were categorized into the 11 genres.", + "STEP 3. Face detection and tracking. For each POI, we first obtained the portrait of the person. This was achieved by detecting and clipping the face images from all pictures of that person. The RetinaFace algorithm was used to perform the detection and clipping BIBREF16. Afterwards, video segments that contain the target person were extracted. This was achieved by three steps: (1) For each frame, detect all the faces appearing in the frame using RetinaFace; (2) Determine if the target person appears by comparing the POI portrait and the faces detected in the frame. We used the ArcFace face recognition system BIBREF17 to perform the comparison; (3) Apply the MOSSE face tracking system BIBREF18 to produce face streams.", + "STEP 4. Active speaker verification. As in BIBREF14, an active speaker verification system was employed to verify if the speech was really spoken by the target person. This is necessary as it is possible that the target person appears in the video but the speech is from other persons. We used the SyncNet model BIBREF19 as in BIBREF14 to perform the task. This model was trained to detect if a stream of mouth movement and a stream of speech are synchronized. In our implementation, the stream of mouth movement was derived from the face stream produced by the MOSSE system.", + "STEP 5. Double check by speaker recognition.", + "Although SyncNet worked well for videos in simple genres, it failed for videos of complex genres such as movie and vlog. A possible reason is that the video content of these genres may change dramatically in time, which leads to unreliable estimation for the stream of the mouth movement, hence unreliable synchronization detection. In order to improve the robustness of the active speaker verification in complex genres, we introduced a double check procedure based on speaker recognition. The idea is simple: whenever the speaker recognition system states a very low confidence for the target speaker, the segment will be discarded even if the confidence from SyncNet is high; vice versa, if the speaker recognition system states a very high confidence, the segment will be retained. We used an off-the-shelf speaker recognition system BIBREF20 to perform this double check. In our study, this double check improved the recall rate by 30% absolutely.", + "STEP 6. Human check.", + "The segments produced by the above automated pipeline were finally checked by human. According to our experience, this human check is rather efficient: one could check 1 hour of speech in 1 hour. As a comparison, if we do not apply the automated pre-selection, checking 1 hour of speech requires 4 hours." + ], + [ + "In this section, we present a series of experiments on speaker recognition using VoxCeleb and CN-Celeb, to compare the complexity of the two datasets." + ], + [ + "VoxCeleb: The entire dataset involves two parts: VoxCeleb1 and VoxCeleb2. We used SITW BIBREF21, a subset of VoxCeleb1 as the evaluation set. The rest of VoxCeleb1 was merged with VoxCeleb2 to form the training set (simply denoted by VoxCeleb). The training set involves $1,236,567$ utterances from $7,185$ speakers, and the evaluation set involves $6,445$ utterances from 299 speakers (precisely, this is the Eval. Core set within SITW).", + "CN-Celeb: The entire dataset was split into two parts: the first part CN-Celeb(T) involves $111,260$ utterances from 800 speakers and was used as the training set; the second part CN-Celeb(E) involves $18,849$ utterances from 200 speakers and was used as the evaluation set." + ], + [ + "Two state-of-the-art baseline systems were built following the Kaldi SITW recipe BIBREF22: an i-vector system BIBREF3 and an x-vector system BIBREF10.", + "For the i-vector system, the acoustic feature involved 24-dimensional MFCCs plus the log energy, augmented by the first- and second-order derivatives. We also applied the cepstral mean normalization (CMN) and the energy-based voice active detection (VAD). The universal background model (UBM) consisted of $2,048$ Gaussian components, and the dimensionality of the i-vector space was 400. LDA was applied to reduce the dimensionality of the i-vectors to 150. The PLDA model was used for scoring BIBREF4.", + "For the x-vector system, the feature-learning component was a 5-layer time-delay neural network (TDNN). The slicing parameters for the five time-delay layers were: {$t$-2, $t$-1, $t$, $t$+1, $t$+2}, {$t$-2, $t$, $t$+2}, {$t$-3, $t$, $t$+3}, {$t$}, {$t$}. The statistic pooling layer computed the mean and standard deviation of the frame-level features from a speech segment. The size of the output layer was consistent with the number of speakers in the training set. Once trained, the activations of the penultimate hidden layer were read out as x-vectors. In our experiments, the dimension of the x-vectors trained on VoxCeleb was set to 512, while for CN-Celeb, it was set to 256, considering the less number of speakers in the training set. Afterwards, the x-vectors were projected to 150-dimensional vectors by LDA, and finally the PLDA model was employed to score the trials. Refer to BIBREF10 for more details." + ], + [ + "We first present the basic results evaluated on SITW and CN-Celeb(E). Both the front-end (i-vector or x-vector models) and back-end (LDA-PLDA) models were trained with the VoxCeleb training set. Note that for SITW, the averaged length of the utterances is more than 80 seconds, while this number is about 8 seconds for CN-Celeb(E). For a better comparison, we resegmented the data of SITW and created a new dataset denoted by SITW(S), where the averaged lengths of the enrollment and test utterances are 28 and 8 seconds, respectively. These numbers are similar to the statistics of CN-Celeb(E).", + "The results in terms of the equal error rate (EER) are reported in Table TABREF24. It can be observed that for both the i-vector system and the x-vector system, the performance on CN-Celeb(E) is much worse than the performance on SITW and SITW(S). This indicates that there is big difference between these two datasets. From another perspective, it demonstrates that the model trained with VoxCeleb does not generalize well, although it has achieved reasonable performance on data from a similar source (SITW)." + ], + [ + "To further compare CN-Celeb and VoxCeleb in a quantitative way, we built systems based on CN-Celeb and VoxCeleb, respectively. For a fair comparison, we randomly sampled 800 speakers from VoxCeleb and built a new dataset VoxCeleb(L) whose size is comparable to CN-Celeb(T). This data set was used for back-end (LDA-PLDA) training.", + "The experimental results are shown in Table TABREF26. Note that the performance of all the comparative experiments show the same trend with the i-vector system and the x-vector system, we therefore only analyze the i-vector results.", + "Firstly, it can be seen that the system trained purely on VoxCeleb obtained good performance on SITW(S) (1st row). This is understandable as VoxCeleb and SITW(S) were collected from the same source. For the pure CN-Celeb system (2nd row), although CN-Celeb(T) and CN-Celeb(E) are from the same source, the performance is still poor (14.24%). More importantly, with re-training the back-end model with VoxCeleb(L) (4th row), the performance on SITW becomes better than the same-source result on CN-Celeb(E) (11.34% vs 14.24%). All these results reconfirmed the significant difference between the two datasets, and indicates that CN-Celeb is more challenging than VoxCeleb." + ], + [ + "We introduced a free dataset CN-Celeb for speaker recognition research. The dataset contains more than $130k$ utterances from $1,000$ Chinese celebrities, and covers 11 different genres in real world. We compared CN-Celeb and VoxCeleb, a widely used dataset in speaker recognition, by setting up a series of experiments based on two state-of-the-art speaker recognition models. Experimental results demonstrated that CN-Celeb is significantly different from VoxCeleb, and it is more challenging for speaker recognition research. The EER performance we obtained in this paper suggests that in unconstrained conditions, the performance of the current speaker recognition techniques might be much worse than it was thought." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0431/instruction.md b/qasper-0431/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..86a4308d9dfadc9edd3e9c110d2f807dc13d8db8 --- /dev/null +++ b/qasper-0431/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Efficient Twitter Sentiment Classification using Subjective Distant Supervision + +Question: How is tweet subjectivity measured? \ No newline at end of file diff --git a/qasper-0436/instruction.md b/qasper-0436/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ef49ee51689f13561ded068c0a4e9234cccc55bd --- /dev/null +++ b/qasper-0436/instruction.md @@ -0,0 +1,142 @@ +Name of Paper: Dynamic Memory Networks for Visual and Textual Question Answering + +Question: How does the model circumvent the lack of supporting facts during training? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Dynamic Memory Networks", + "Improved Dynamic Memory Networks: DMN+", + "Input Module for Text QA", + "Input Module for VQA", + "The Episodic Memory Module", + "Related Work", + "Datasets", + "bAbI-10k", + "DAQUAR-ALL visual dataset", + "Visual Question Answering", + "Model Analysis", + "Comparison to state of the art using bAbI-10k", + "Comparison to state of the art using VQA", + "Conclusion" + ], + "paragraphs": [ + [ + "Neural network based methods have made tremendous progress in image and text classification BIBREF0 , BIBREF1 . However, only recently has progress been made on more complex tasks that require logical reasoning. This success is based in part on the addition of memory and attention components to complex neural networks. For instance, memory networks BIBREF2 are able to reason over several facts written in natural language or (subject, relation, object) triplets. Attention mechanisms have been successful components in both machine translation BIBREF3 , BIBREF4 and image captioning models BIBREF5 .", + "The dynamic memory network BIBREF6 (DMN) is one example of a neural network model that has both a memory component and an attention mechanism. The DMN yields state of the art results on question answering with supporting facts marked during training, sentiment analysis, and part-of-speech tagging.", + "We analyze the DMN components, specifically the input module and memory module, to improve question answering. We propose a new input module which uses a two level encoder with a sentence reader and input fusion layer to allow for information flow between sentences. For the memory, we propose a modification to gated recurrent units (GRU) BIBREF7 . The new GRU formulation incorporates attention gates that are computed using global knowledge over the facts. Unlike before, the new DMN+ model does not require that supporting facts (i.e. the facts that are relevant for answering a particular question) are labeled during training. The model learns to select the important facts from a larger set.", + "In addition, we introduce a new input module to represent images. This module is compatible with the rest of the DMN architecture and its output is fed into the memory module. We show that the changes in the memory module that improved textual question answering also improve visual question answering. Both tasks are illustrated in Fig. 1 ." + ], + [ + "We begin by outlining the DMN for question answering and the modules as presented in BIBREF6 .", + "The DMN is a general architecture for question answering (QA). It is composed of modules that allow different aspects such as input representations or memory components to be analyzed and improved independently. The modules, depicted in Fig. 1 , are as follows:", + "Input Module: This module processes the input data about which a question is being asked into a set of vectors termed facts, represented as $F=[f_1,\\hdots ,f_N]$ , where $N$ is the total number of facts. These vectors are ordered, resulting in additional information that can be used by later components. For text QA in BIBREF6 , the module consists of a GRU over the input words.", + "As the GRU is used in many components of the DMN, it is useful to provide the full definition. For each time step $i$ with input $x_i$ and previous hidden state $h_{i-1}$ , we compute the updated hidden state $h_i = GRU(x_i,h_{i-1})$ by ", + "$$u_i &=& \\sigma \\left(W^{(u)}x_{i} + U^{(u)} h_{i-1} + b^{(u)} \\right)\\\\\nr_i &=& \\sigma \\left(W^{(r)}x_{i} + U^{(r)} h_{i-1} + b^{(r)} \\right)\\\\\n\\tilde{h}_i &=& \\tanh \\left(Wx_{i} + r_i \\circ U h_{i-1} + b^{(h)}\\right)\\\\\nh_i &=& u_i\\circ \\tilde{h}_i + (1-u_i) \\circ h_{i-1}$$ (Eq. 2) ", + "where $\\sigma $ is the sigmoid activation function, $\\circ $ is an element-wise product, $W^{(z)}, W^{(r)}, W \\in \\mathbb {R}^{n_H \\times n_I}$ , $U^{(z)}, U^{(r)}, U \\in \\mathbb {R}^{n_H \\times n_H}$ , $n_H$ is the hidden size, and $n_I$ is the input size.", + "Question Module: This module computes a vector representation $q$ of the question, where $q \\in \\mathbb {R}^{n_H}$ is the final hidden state of a GRU over the words in the question.", + "Episodic Memory Module: Episode memory aims to retrieve the information required to answer the question $q$ from the input facts. To improve our understanding of both the question and input, especially if questions require transitive reasoning, the episode memory module may pass over the input multiple times, updating episode memory after each pass. We refer to the episode memory on the $t^{th}$ pass over the inputs as $m^t$ , where $m^t \\in \\mathbb {R}^{n_H}$ , the initial memory vector is set to the question vector: $m^0 = q$ .", + "The episodic memory module consists of two separate components: the attention mechanism and the memory update mechanism. The attention mechanism is responsible for producing a contextual vector $c^t$ , where $c^t \\in \\mathbb {R}^{n_H}$ is a summary of relevant input for pass $t$ , with relevance inferred by the question $q$ and previous episode memory $m^{t-1}$ . The memory update mechanism is responsible for generating the episode memory $m^t$ based upon the contextual vector $c^t$ and previous episode memory $m^{t-1}$ . By the final pass $T$ , the episodic memory $m^T$ should contain all the information required to answer the question $c^t \\in \\mathbb {R}^{n_H}$0 .", + "Answer Module: The answer module receives both $q$ and $m^T$ to generate the model's predicted answer. For simple answers, such as a single word, a linear layer with softmax activation may be used. For tasks requiring a sequence output, an RNN may be used to decode $a = [q ; m^T]$ , the concatenation of vectors $q$ and $m^T$ , to an ordered set of tokens. The cross entropy error on the answers is used for training and backpropagated through the entire network." + ], + [ + "We propose and compare several modeling choices for two crucial components: input representation, attention mechanism and memory update. The final DMN+ model obtains the highest accuracy on the bAbI-10k dataset without supporting facts and the VQA dataset BIBREF8 . Several design choices are motivated by intuition and accuracy improvements on that dataset." + ], + [ + "In the DMN specified in BIBREF6 , a single GRU is used to process all the words in the story, extracting sentence representations by storing the hidden states produced at the end of sentence markers. The GRU also provides a temporal component by allowing a sentence to know the content of the sentences that came before them. Whilst this input module worked well for bAbI-1k with supporting facts, as reported in BIBREF6 , it did not perform well on bAbI-10k without supporting facts (Sec. \"Model Analysis\" ).", + "We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU.", + "Input Fusion Layer", + "For the DMN+, we propose replacing this single GRU with two different components. The first component is a sentence reader, responsible only for encoding the words into a sentence embedding. The second component is the input fusion layer, allowing for interactions between sentences. This resembles the hierarchical neural auto-encoder architecture of BIBREF9 and allows content interaction between sentences. We adopt the bi-directional GRU for this input fusion layer because it allows information from both past and future sentences to be used. As gradients do not need to propagate through the words between sentences, the fusion layer also allows for distant supporting sentences to have a more direct interaction.", + "Fig. 2 shows an illustration of an input module, where a positional encoder is used for the sentence reader and a bi-directional GRU is adopted for the input fusion layer. Each sentence encoding $f_i$ is the output of an encoding scheme taking the word tokens $[w^i_1, \\hdots , w^i_{M_i}]$ , where $M_i$ is the length of the sentence.", + "The sentence reader could be based on any variety of encoding schemes. We selected positional encoding described in BIBREF10 to allow for a comparison to their work. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used.", + "For the positional encoding scheme, the sentence representation is produced by $f_i = \\sum ^{j=1}_M l_j \\circ w^i_j$ , where $\\circ $ is element-wise multiplication and $l_j$ is a column vector with structure $l_{jd} = (1 - j / M) - (d / D) (1 - 2j / M)$ , where $d$ is the embedding index and $D$ is the dimension of the embedding.", + "The input fusion layer takes these input facts and enables an information exchange between them by applying a bi-directional GRU. ", + "$$\\overrightarrow{f_i} = GRU_{fwd}(f_i, \\overrightarrow{f_{i-1}}) \\\\\n\\overleftarrow{f_{i}} = GRU_{bwd}(f_{i}, \\overleftarrow{f_{i+1}}) \\\\\n\\overleftrightarrow{f_i} = \\overleftarrow{f_i} + \\overrightarrow{f_i}$$ (Eq. 5) ", + "where $f_i$ is the input fact at timestep $i$ , $ \\overrightarrow{f_i}$ is the hidden state of the forward GRU at timestep $i$ , and $\\overleftarrow{f_i}$ is the hidden state of the backward GRU at timestep $i$ . This allows contextual information from both future and past facts to impact $\\overleftrightarrow{f_i}$ .", + "We explored a variety of encoding schemes for the sentence reader, including GRUs, LSTMs, and the positional encoding scheme described in BIBREF10 . For simplicity and speed, we selected the positional encoding scheme. GRUs and LSTMs were also considered but required more computational resources and were prone to overfitting if auxiliary tasks, such as reconstructing the original sentence, were not used." + ], + [ + "To apply the DMN to visual question answering, we introduce a new input module for images. The module splits an image into small local regions and considers each region equivalent to a sentence in the input module for text. The input module for VQA is composed of three parts, illustrated in Fig. 3 : local region feature extraction, visual feature embedding, and the input fusion layer introduced in Sec. \"Input Module for Text QA\" .", + "Local region feature extraction: To extract features from the image, we use a convolutional neural network BIBREF0 based upon the VGG-19 model BIBREF11 . We first rescale the input image to $448 \\times 448$ and take the output from the last pooling layer which has dimensionality $d = 512 \\times 14 \\times 14$ . The pooling layer divides the image into a grid of $14 \\times 14$ , resulting in 196 local regional vectors of $d = 512$ .", + "Visual feature embedding: As the VQA task involves both image features and text features, we add a linear layer with tanh activation to project the local regional vectors to the textual feature space used by the question vector $q$ .", + "Input fusion layer: The local regional vectors extracted from above do not yet have global information available to them. Without global information, their representational power is quite limited, with simple issues like object scaling or locational variance causing accuracy problems.", + "To solve this, we add an input fusion layer similar to that of the textual input module described in Sec. \"Input Module for Text QA\" . First, to produce the input facts $F$ , we traverse the image in a snake like fashion, as seen in Figure 3 . We then apply a bi-directional GRU over these input facts $F$ to produce the globally aware input facts $\\overleftrightarrow{F}$ . The bi-directional GRU allows for information propagation from neighboring image patches, capturing spatial information." + ], + [ + "The episodic memory module, as depicted in Fig. 4 , retrieves information from the input facts $\\overleftrightarrow{F} = [\\overleftrightarrow{f_1}, \\hdots , \\overleftrightarrow{f_N}]$ provided to it by focusing attention on a subset of these facts. We implement this attention by associating a single scalar value, the attention gate $g^t_i$ , with each fact $\\overleftrightarrow{f}_i$ during pass $t$ . This is computed by allowing interactions between the fact and both the question representation and the episode memory state. ", + "$$z^t_i &=& [\\overleftrightarrow{f_i} \\circ q; \\overleftrightarrow{f_i} \\circ m^{t-1}; \\vert \\overleftrightarrow{f_i} - q \\vert ; \\vert \\overleftrightarrow{f_i} - m^{t-1} \\vert ] \\\\\nZ^t_i &=& W^{(2)} \\tanh \\left(W^{(1)}z^t_i + b^{(1)} \\right)+ b^{(2)} \\\\\ng^t_i &=& \\frac{\\exp (Z^t_i)}{\\sum _{k=1}^{M_i} \\exp (Z^t_k)} $$ (Eq. 10) ", + "where $\\overleftrightarrow{f_i}$ is the $i^{th}$ fact, $m^{t-1}$ is the previous episode memory, $q$ is the original question, $\\circ $ is the element-wise product, $|\\cdot |$ is the element-wise absolute value, and $;$ represents concatenation of the vectors.", + "The DMN implemented in BIBREF6 involved a more complex set of interactions within $z$ , containing the additional terms $[f; m^{t-1}; q; f^T W^{(b)} q; f^T W^{(b)} m^{t-1}]$ . After an initial analysis, we found these additional terms were not required.", + "Attention Mechanism", + "Once we have the attention gate $g^t_i$ we use an attention mechanism to extract a contextual vector $c^t$ based upon the current focus. We focus on two types of attention: soft attention and a new attention based GRU. The latter improves performance and is hence the final modeling choice for the DMN+.", + "Soft attention: Soft attention produces a contextual vector $c^t$ through a weighted summation of the sorted list of vectors $\\overleftrightarrow{F}$ and corresponding attention gates $g_i^t$ : $c^t = \\sum _{i=1}^N g^t_i \\overleftrightarrow{f}_i$ This method has two advantages. First, it is easy to compute. Second, if the softmax activation is spiky it can approximate a hard attention function by selecting only a single fact for the contextual vector whilst still being differentiable. However the main disadvantage to soft attention is that the summation process loses both positional and ordering information. Whilst multiple attention passes can retrieve some of this information, this is inefficient.", + "Attention based GRU: For more complex queries, we would like for the attention mechanism to be sensitive to both the position and ordering of the input facts $\\overleftrightarrow{F}$ . An RNN would be advantageous in this situation except they cannot make use of the attention gate from Equation .", + "We propose a modification to the GRU architecture by embedding information from the attention mechanism. The update gate $u_i$ in Equation 2 decides how much of each dimension of the hidden state to retain and how much should be updated with the transformed input $x_i$ from the current timestep. As $u_i$ is computed using only the current input and the hidden state from previous timesteps, it lacks any knowledge from the question or previous episode memory.", + "By replacing the update gate $u_i$ in the GRU (Equation 2 ) with the output of the attention gate $g^t_i$ (Equation ) in Equation , the GRU can now use the attention gate for updating its internal state. This change is depicted in Fig 5 . ", + "$$h_i &=& g^t_i \\circ \\tilde{h}_i + (1-g^t_i) \\circ h_{i-1}$$ (Eq. 12) ", + "An important consideration is that $g^t_i$ is a scalar, generated using a softmax activation, as opposed to the vector $u_i \\in \\mathbb {R}^{n_H}$ , generated using a sigmoid activation. This allows us to easily visualize how the attention gates activate over the input, later shown for visual QA in Fig. 6 . Though not explored, replacing the softmax activation in Equation with a sigmoid activation would result in $g^t_i \\in \\mathbb {R}^{n_H}$ . To produce the contextual vector $c^t$ used for updating the episodic memory state $m^t$ , we use the final hidden state of the attention based GRU.", + "Episode Memory Updates", + "After each pass through the attention mechanism, we wish to update the episode memory $m^{t-1}$ with the newly constructed contextual vector $c^t$ , producing $m^t$ . In the DMN, a GRU with the initial hidden state set to the question vector $q$ is used for this purpose. The episodic memory for pass $t$ is computed by ", + "$$m^t = GRU(c^t, m^{t-1})$$ (Eq. 13) ", + "The work of BIBREF10 suggests that using different weights for each pass through the episodic memory may be advantageous. When the model contains only one set of weights for all episodic passes over the input, it is referred to as a tied model, as in the \u201cMem Weights\u201d row in Table 1 .", + "Following the memory update component used in BIBREF10 and BIBREF12 we experiment with using a ReLU layer for the memory update, calculating the new episode memory state by ", + "$$m^t = ReLU\\left(W^t [m^{t-1} ; c^t ; q] + b\\right)$$ (Eq. 14) ", + "where $;$ is the concatenation operator, $W^t \\in \\mathbb {R}^{n_H \\times n_H}$ , $b \\in \\mathbb {R}^{n_H}$ , and $n_H$ is the hidden size. The untying of weights and using this ReLU formulation for the memory update improves accuracy by another 0.5% as shown in Table 1 in the last column. The final output of the memory network is passed to the answer module as in the original DMN." + ], + [ + "The DMN is related to two major lines of recent work: memory and attention mechanisms. We work on both visual and textual question answering which have, until now, been developed in separate communities.", + "Neural Memory Models The earliest recent work with a memory component that is applied to language processing is that of memory networks BIBREF2 which adds a memory component for question answering over simple facts. They are similar to DMNs in that they also have input, scoring, attention and response mechanisms. However, unlike the DMN their input module computes sentence representations independently and hence cannot easily be used for other tasks such as sequence labeling. Like the original DMN, this memory network requires that supporting facts are labeled during QA training. End-to-end memory networks BIBREF10 do not have this limitation. In contrast to previous memory models with a variety of different functions for memory attention retrieval and representations, DMNs BIBREF6 have shown that neural sequence models can be used for input representation, attention and response mechanisms. Sequence models naturally capture position and temporality of both the inputs and transitive reasoning steps.", + "Neural Attention Mechanisms Attention mechanisms allow neural network models to use a question to selectively pay attention to specific inputs. They can benefit image classification BIBREF13 , generating captions for images BIBREF5 , among others mentioned below, and machine translation BIBREF14 , BIBREF3 , BIBREF4 . Other recent neural architectures with memory or attention which have proposed include neural Turing machines BIBREF15 , neural GPUs BIBREF16 and stack-augmented RNNs BIBREF17 .", + "Question Answering in NLP Question answering involving natural language can be solved in a variety of ways to which we cannot all do justice. If the potential input is a large text corpus, QA becomes a combination of information retrieval and extraction BIBREF18 . Neural approaches can include reasoning over knowledge bases, BIBREF19 , BIBREF20 or directly via sentences for trivia competitions BIBREF21 .", + "Visual Question Answering (VQA) In comparison to QA in NLP, VQA is still a relatively young task that is feasible only now that objects can be identified with high accuracy. The first large scale database with unconstrained questions about images was introduced by BIBREF8 . While VQA datasets existed before they did not include open-ended, free-form questions about general images BIBREF22 . Others are were too small to be viable for a deep learning approach BIBREF23 . The only VQA model which also has an attention component is the stacked attention network BIBREF24 . Their work also uses CNN based features. However, unlike our input fusion layer, they use a single layer neural network to map the features of each patch to the dimensionality of the question vector. Hence, the model cannot easily incorporate adjacency of local information in its hidden state. A model that also uses neural modules, albeit logically inspired ones, is that by BIBREF25 who evaluate on knowledgebase reasoning and visual question answering. We compare directly to their method on the latter task and dataset.", + "Related to visual question answering is the task of describing images with sentences BIBREF26 . BIBREF27 used deep learning methods to map images and sentences into the same space in order to describe images with sentences and to find images that best visualize a sentence. This was the first work to map both modalities into a joint space with deep learning methods, but it could only select an existing sentence to describe an image. Shortly thereafter, recurrent neural networks were used to generate often novel sentences based on images BIBREF28 , BIBREF29 , BIBREF30 , BIBREF5 ." + ], + [ + "To analyze our proposed model changes and compare our performance with other architectures, we use three datasets." + ], + [ + "For evaluating the DMN on textual question answering, we use bAbI-10k English BIBREF31 , a synthetic dataset which features 20 different tasks. Each example is composed of a set of facts, a question, the answer, and the supporting facts that lead to the answer. The dataset comes in two sizes, referring to the number of training examples each task has: bAbI-1k and bAbI-10k. The experiments in BIBREF10 found that their lowest error rates on the smaller bAbI-1k dataset were on average three times higher than on bAbI-10k." + ], + [ + "The DAtaset for QUestion Answering on Real-world images (DAQUAR) BIBREF23 consists of 795 training images and 654 test images. Based upon these images, 6,795 training questions and 5,673 test questions were generated. Following the previously defined experimental method, we exclude multiple word answers BIBREF32 , BIBREF33 . The resulting dataset covers 90% of the original data. The evaluation method uses classification accuracy over the single words. We use this as a development dataset for model analysis (Sec. \"Model Analysis\" )." + ], + [ + "The Visual Question Answering (VQA) dataset was constructed using the Microsoft COCO dataset BIBREF34 which contained 123,287 training/validation images and 81,434 test images. Each image has several related questions with each question answered by multiple people. This dataset contains 248,349 training questions, 121,512 validation questions, and 244,302 for testing. The testing data was split into test-development, test-standard and test-challenge in BIBREF8 .", + "Evaluation on both test-standard and test-challenge are implemented via a submission system. test-standard may only be evaluated 5 times and test-challenge is only evaluated at the end of the competition. To the best of our knowledge, VQA is the largest and most complex image dataset for the visual question answering task." + ], + [ + "To understand the impact of the proposed module changes, we analyze the performance of a variety of DMN models on textual and visual question answering datasets.", + "The original DMN (ODMN) is the architecture presented in BIBREF6 without any modifications. DMN2 only replaces the input module with the input fusion layer (Sec. \"Input Module for Text QA\" ). DMN3, based upon DMN2, replaces the soft attention mechanism with the attention based GRU proposed in Sec. \"The Episodic Memory Module\" . Finally, DMN+, based upon DMN3, is an untied model, using a unique set of weights for each pass and a linear layer with a ReLU activation to compute the memory update. We report the performance of the model variations in Table 1 .", + "A large improvement to accuracy on both the bAbI-10k textual and DAQUAR visual datasets results from updating the input module, seen when comparing ODMN to DMN2. On both datasets, the input fusion layer improves interaction between distant facts. In the visual dataset, this improvement is purely from providing contextual information from neighboring image patches, allowing it to handle objects of varying scale or questions with a locality aspect. For the textual dataset, the improved interaction between sentences likely helps the path finding required for logical reasoning when multiple transitive steps are required.", + "The addition of the attention GRU in DMN3 helps answer questions where complex positional or ordering information may be required. This change impacts the textual dataset the most as few questions in the visual dataset are likely to require this form of logical reasoning. Finally, the untied model in the DMN+ overfits on some tasks compared to DMN3, but on average the error rate decreases.", + "From these experimental results, we find that the combination of all the proposed model changes results, culminating in DMN+, achieves the highest performance across both the visual and textual datasets." + ], + [ + "We trained our models using the Adam optimizer BIBREF35 with a learning rate of 0.001 and batch size of 128. Training runs for up to 256 epochs with early stopping if the validation loss had not improved within the last 20 epochs. The model from the epoch with the lowest validation loss was then selected. Xavier initialization was used for all weights except for the word embeddings, which used random uniform initialization with range $[-\\sqrt{3}, \\sqrt{3}]$ . Both the embedding and hidden dimensions were of size $d = 80$ . We used $\\ell _2$ regularization on all weights except bias and used dropout on the initial sentence encodings and the answer module, keeping the input with probability $p=0.9$ . The last 10% of the training data on each task was chosen as the validation set. For all tasks, three passes were used for the episodic memory module, allowing direct comparison to other state of the art methods. Finally, we limited the input to the last 70 sentences for all tasks except QA3 for which we limited input to the last 130 sentences, similar to BIBREF10 .", + "On some tasks, the accuracy was not stable across multiple runs. This was particularly problematic on QA3, QA17, and QA18. To solve this, we repeated training 10 times using random initializations and evaluated the model that achieved the lowest validation set loss.", + "Text QA Results", + "We compare our best performing approach, DMN+, to two state of the art question answering architectures: the end to end memory network (E2E) BIBREF10 and the neural reasoner framework (NR) BIBREF12 . Neither approach use supporting facts for training.", + "The end-to-end memory network is a form of memory network BIBREF2 tested on both textual question answering and language modeling. The model features both explicit memory and a recurrent attention mechanism. We select the model from the paper that achieves the lowest mean error over the bAbI-10k dataset. This model utilizes positional encoding for input, RNN-style tied weights for the episode module, and a ReLU non-linearity for the memory update component.", + "The neural reasoner framework is an end-to-end trainable model which features a deep architecture for logical reasoning and an interaction-pooling mechanism for allowing interaction over multiple facts. While the neural reasoner framework was only tested on QA17 and QA19, these were two of the most challenging question types at the time.", + "In Table 2 we compare the accuracy of these question answering architectures, both as mean error and error on individual tasks. The DMN+ model reduces mean error by 1.4% compared to the the end-to-end memory network, achieving a new state of the art for the bAbI-10k dataset.", + "One notable deficiency in our model is that of QA16: Basic Induction. In BIBREF10 , an untied model using only summation for memory updates was able to achieve a near perfect error rate of $0.4$ . When the memory update was replaced with a linear layer with ReLU activation, the end-to-end memory network's overall mean error decreased but the error for QA16 rose sharply. Our model experiences the same difficulties, suggesting that the more complex memory update component may prevent convergence on certain simpler tasks.", + "The neural reasoner model outperforms both the DMN and end-to-end memory network on QA17: Positional Reasoning. This is likely as the positional reasoning task only involves minimal supervision - two sentences for input, yes/no answers for supervision, and only 5,812 unique examples after removing duplicates from the initial 10,000 training examples. BIBREF12 add an auxiliary task of reconstructing both the original sentences and question from their representations. This auxiliary task likely improves performance by preventing overfitting." + ], + [ + "For the VQA dataset, each question is answered by multiple people and the answers may not be the same, the generated answers are evaluated using human consensus. For each predicted answer $a_i$ for the $i_{th}$ question with target answer set $T^{i}$ , the accuracy of VQA: $Acc_{VQA} = \\frac{1}{N}\\sum _{i=1}^Nmin(\\frac{\\sum _{t\\in T^i}{1}_{(a_i==t)}}{3},1)$ where ${1}_{(\\cdot )}$ is the indicator function. Simply put, the answer $a_i$ is only 100 $\\%$ accurate if at least 3 people provide that exact answer.", + "Training Details We use the Adam optimizer BIBREF35 with a learning rate of 0.003 and batch size of 100. Training runs for up to 256 epochs with early stopping if the validation loss has not improved in the last 10 epochs. For weight initialization, we sampled from a random uniform distribution with range $[-0.08, 0.08]$ . Both the word embedding and hidden layers were vectors of size $d=512$ . We apply dropout on the initial image output from the VGG convolutional neural network BIBREF11 as well as the input to the answer module, keeping input with probability $p=0.5$ .", + "Results and Analysis", + "The VQA dataset is composed of three question domains: Yes/No, Number, and Other. This enables us to analyze the performance of the models on various tasks that require different reasoning abilities.", + "The comparison models are separated into two broad classes: those that utilize a full connected image feature for classification and those that perform reasoning over multiple small image patches. Only the SAN and DMN approach use small image patches, while the rest use the fully-connected whole image feature approach.", + "Here, we show the quantitative and qualitative results in Table 3 and Fig. 6 , respectively. The images in Fig. 6 illustrate how the attention gate $g^t_i$ selectively activates over relevant portions of the image according to the query. In Table 3 , our method outperforms baseline and other state-of-the-art methods across all question domains (All) in both test-dev and test-std, and especially for Other questions, achieves a wide margin compared to the other architectures, which is likely as the small image patches allow for finely detailed reasoning over the image.", + "However, the granularity offered by small image patches does not always offer an advantage. The Number questions may be not solvable for both the SAN and DMN architectures, potentially as counting objects is not a simple task when an object crosses image patch boundaries." + ], + [ + "We have proposed new modules for the DMN framework to achieve strong results without supervision of supporting facts. These improvements include the input fusion layer to allow interactions between input facts and a novel attention based GRU that allows for logical reasoning over ordered inputs. Our resulting model obtains state of the art results on both the VQA dataset and the bAbI-10k text question-answering dataset, proving the framework can be generalized across input domains." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0447/instruction.md b/qasper-0447/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c5f8be3d68b86ae0462db80bd2711c0d89b80642 --- /dev/null +++ b/qasper-0447/instruction.md @@ -0,0 +1,105 @@ +Name of Paper: Fusing Visual, Textual and Connectivity Clues for Studying Mental Health + +Question: What types of features are used from each data type? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + null, + "Introduction", + "Related Work", + "Dataset", + "Data Modality Analysis", + "Demographic Prediction", + "Multi-modal Prediction Framework" + ], + "paragraphs": [ + [ + "0pt*0*0", + "0pt*0*0", + "0pt*0*0 0.95", + "1]Amir Hossein Yazdavar 1]Mohammad Saeid Mahdavinejad 2]Goonmeet Bajaj", + " 3]William Romine 1]Amirhassan Monadjemi 1]Krishnaprasad Thirunarayan", + " 1]Amit Sheth 4]Jyotishman Pathak [1]Department of Computer Science & Engineering, Wright State University, OH, USA [2]Ohio State University, Columbus, OH, USA [3]Department of Biological Science, Wright State University, OH, USA [4] Division of Health Informatics, Weill Cornell University, New York, NY, USA", + "[1] yazdavar.2@wright.edu", + "With ubiquity of social media platforms, millions of people are sharing their online persona by expressing their thoughts, moods, emotions, feelings, and even their daily struggles with mental health issues voluntarily and publicly on social media. Unlike the most existing efforts which study depression by analyzing textual content, we examine and exploit multimodal big data to discern depressive behavior using a wide variety of features including individual-level demographics. By developing a multimodal framework and employing statistical techniques for fusing heterogeneous sets of features obtained by processing visual, textual and user interaction data, we significantly enhance the current state-of-the-art approaches for identifying depressed individuals on Twitter (improving the average F1-Score by 5 percent) as well as facilitate demographic inference from social media for broader applications. Besides providing insights into the relationship between demographics and mental health, our research assists in the design of a new breed of demographic-aware health interventions." + ], + [ + "Depression is a highly prevalent public health challenge and a major cause of disability worldwide. Depression affects 6.7% (i.e., about 16 million) Americans each year . According to the World Mental Health Survey conducted in 17 countries, on average, about 5% of people reported having an episode of depression in 2011 BIBREF0 . Untreated or under-treated clinical depression can lead to suicide and other chronic risky behaviors such as drug or alcohol addiction.", + "Global efforts to curb clinical depression involve identifying depression through survey-based methods employing phone or online questionnaires. These approaches suffer from under-representation as well as sampling bias (with very small group of respondents.) In contrast, the widespread adoption of social media where people voluntarily and publicly express their thoughts, moods, emotions, and feelings, and even share their daily struggles with mental health problems has not been adequately tapped into studying mental illnesses, such as depression. The visual and textual content shared on different social media platforms like Twitter offer new opportunities for a deeper understanding of self-expressed depression both at an individual as well as community-level. Previous research efforts have suggested that language style, sentiment, users' activities, and engagement expressed in social media posts can predict the likelihood of depression BIBREF1 , BIBREF2 . However, except for a few attempts BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , these investigations have seldom studied extraction of emotional state from visual content of images in posted/profile images. Visual content can express users' emotions more vividly, and psychologists noted that imagery is an effective medium for communicating difficult emotions.", + "According to eMarketer, photos accounted for 75% of content posted on Facebook worldwide and they are the most engaging type of content on Facebook (87%). Indeed, \"a picture is worth a thousand words\" and now \"photos are worth a million likes.\" Similarly, on Twitter, the tweets with image links get twice as much attention as those without , and video-linked tweets drive up engagement . The ease and naturalness of expression through visual imagery can serve to glean depression-indicators in vulnerable individuals who often seek social support through social media BIBREF7 . Further, as psychologist Carl Rogers highlights, we often pursue and promote our Ideal-Self . In this regard, the choice of profile image can be a proxy for the online persona BIBREF8 , providing a window into an individual's mental health status. For instance, choosing emaciated legs of girls covered with several cuts as profile image portrays negative self-view BIBREF9 .", + "Inferring demographic information like gender and age can be crucial for stratifying our understanding of population-level epidemiology of mental health disorders. Relying on electronic health records data, previous studies explored gender differences in depressive behavior from different angles including prevalence, age at onset, comorbidities, as well as biological and psychosocial factors. For instance, women have been diagnosed with depression twice as often as men BIBREF10 and national psychiatric morbidity survey in Britain has shown higher risk of depression in women BIBREF11 . On the other hand, suicide rates for men are three to five times higher compared to that of the women BIBREF12 .", + "Although depression can affect anyone at any age, signs and triggers of depression vary for different age groups . Depression triggers for children include parental depression, domestic violence, and loss of a pet, friend or family member. For teenagers (ages 12-18), depression may arise from hormonal imbalance, sexuality concerns and rejection by peers. Young adults (ages 19-29) may develop depression due to life transitions, poverty, trauma, and work issues. Adult (ages 30-60) depression triggers include caring simultaneously for children and aging parents, financial burden, work and relationship issues. Senior adults develop depression from common late-life issues, social isolation, major life loses such as the death of a spouse, financial stress and other chronic health problems (e.g., cardiac disease, dementia). Therefore, inferring demographic information while studying depressive behavior from passively sensed social data, can shed better light on the population-level epidemiology of depression.", + "The recent advancements in deep neural networks, specifically for image analysis task, can lead to determining demographic features such as age and gender BIBREF13 . We show that by determining and integrating heterogeneous set of features from different modalities \u2013 aesthetic features from posted images (colorfulness, hue variance, sharpness, brightness, blurriness, naturalness), choice of profile picture (for gender, age, and facial expression), the screen name, the language features from both textual content and profile's description (n-gram, emotion, sentiment), and finally sociability from ego-network, and user engagement \u2013 we can reliably detect likely depressed individuals in a data set of 8,770 human-annotated Twitter users.", + "We address and derive answers to the following research questions: 1) How well do the content of posted images (colors, aesthetic and facial presentation) reflect depressive behavior? 2) Does the choice of profile picture show any psychological traits of depressed online persona? Are they reliable enough to represent the demographic information such as age and gender? 3) Are there any underlying common themes among depressed individuals generated using multimodal content that can be used to detect depression reliably?" + ], + [ + "Mental Health Analysis using Social Media:", + "Several efforts have attempted to automatically detect depression from social media content utilizing machine/deep learning and natural language processing approaches. Conducting a retrospective study over tweets, BIBREF14 characterizes depression based on factors such as language, emotion, style, ego-network, and user engagement. They built a classifier to predict the likelihood of depression in a post BIBREF14 , BIBREF15 or in an individual BIBREF1 , BIBREF16 , BIBREF17 , BIBREF18 . Moreover, there have been significant advances due to the shared task BIBREF19 focusing on methods for identifying depressed users on Twitter at the Computational Linguistics and Clinical Psychology Workshop (CLP 2015). A corpus of nearly 1,800 Twitter users was built for evaluation, and the best models employed topic modeling BIBREF20 , Linguistic Inquiry and Word Count (LIWC) features, and other metadata BIBREF21 . More recently, a neural network architecture introduced by BIBREF22 combined posts into a representation of user's activities for detecting depressed users. Another active line of research has focused on capturing suicide and self-harm signals BIBREF23 , BIBREF24 , BIBREF25 , BIBREF26 , BIBREF2 , BIBREF27 . Moreover, the CLP 2016 BIBREF28 defined a shared task on detecting the severity of the mental health from forum posts. All of these studies derive discriminative features to classify depression in user-generated content at message-level, individual-level or community-level. Recent emergence of photo-sharing platforms such as Instagram, has attracted researchers attention to study people's behavior from their visual narratives \u2013 ranging from mining their emotions BIBREF29 , and happiness trend BIBREF30 , to studying medical concerns BIBREF31 . Researchers show that people use Instagram to engage in social exchange and storytelling about their difficult experiences BIBREF4 . The role of visual imagery as a mechanism of self-disclosure by relating visual attributes to mental health disclosures on Instagram was highlighted by BIBREF3 , BIBREF5 where individual Instagram profiles were utilized to build a prediction framework for identifying markers of depression. The importance of data modality to understand user behavior on social media was highlighted by BIBREF32 . More recently, a deep neural network sequence modeling approach that marries audio and text data modalities to analyze question-answer style interviews between an individual and an agent has been developed to study mental health BIBREF32 . Similarly, a multimodal depressive dictionary learning was proposed to detect depressed users on Twitter BIBREF33 . They provide a sparse user representations by defining a feature set consisting of social network features, user profile features, visual features, emotional features BIBREF34 , topic-level features, and domain-specific features. Particularly, our choice of multi-model prediction framework is intended to improve upon the prior works involving use of images in multimodal depression analysis BIBREF33 and prior works on studying Instagram photos BIBREF6 , BIBREF35 .", + "Demographic information inference on Social Media: ", + "There is a growing interest in understanding online user's demographic information due to its numerous applications in healthcare BIBREF36 , BIBREF37 . A supervised model developed by BIBREF38 for determining users' gender by employing features such as screen-name, full-name, profile description and content on external resources (e.g., personal blog). Employing features including emoticons, acronyms, slangs, punctuations, capitalization, sentence length and included links/images, along with online behaviors such as number of friends, post time, and commenting activity, a supervised model was built for predicting user's age group BIBREF39 . Utilizing users life stage information such as secondary school student, college student, and employee, BIBREF40 builds age inference model for Dutch Twitter users. Similarly, relying on profile descriptions while devising a set of rules and patterns, a novel model introduced for extracting age for Twitter users BIBREF41 . They also parse description for occupation by consulting the SOC2010 list of occupations and validating it through social surveys. A novel age inference model was developed while relying on homophily interaction information and content for predicting age of Twitter users BIBREF42 . The limitations of textual content for predicting age and gender was highlighted by BIBREF43 . They distinguish language use based on social gender, age identity, biological sex and chronological age by collecting crowdsourced signals using a game in which players (crowd) guess the biological sex and age of a user based only on their tweets. Their findings indicate how linguistic markers can misguide (e.g., a heart represented as <3 can be misinterpreted as feminine when the writer is male.) Estimating age and gender from facial images by training a convolutional neural networks (CNN) for face recognition is an active line of research BIBREF44 , BIBREF13 , BIBREF45 ." + ], + [ + "Self-disclosure clues have been extensively utilized for creating ground-truth data for numerous social media analytic studies e.g., for predicting demographics BIBREF36 , BIBREF41 , and user's depressive behavior BIBREF46 , BIBREF47 , BIBREF48 . For instance, vulnerable individuals may employ depressive-indicative terms in their Twitter profile descriptions. Others may share their age and gender, e.g., \"16 years old suicidal girl\"(see Figure FIGREF15 ). We employ a huge dataset of 45,000 self-reported depressed users introduced in BIBREF46 where a lexicon of depression symptoms consisting of 1500 depression-indicative terms was created with the help of psychologist clinician and employed for collecting self-declared depressed individual's profiles. A subset of 8,770 users (24 million time-stamped tweets) containing 3981 depressed and 4789 control users (that do not show any depressive behavior) were verified by two human judges BIBREF46 . This dataset INLINEFORM0 contains the metadata values of each user such as profile descriptions, followers_count, created_at, and profile_image_url.", + "Age Enabled Ground-truth Dataset: We extract user's age by applying regular expression patterns to profile descriptions (such as \"17 years old, self-harm, anxiety, depression\") BIBREF41 . We compile \"age prefixes\" and \"age suffixes\", and use three age-extraction rules: 1. I am X years old 2. Born in X 3. X years old, where X is a \"date\" or age (e.g., 1994). We selected a subset of 1061 users among INLINEFORM0 as gold standard dataset INLINEFORM1 who disclose their age. From these 1061 users, 822 belong to depressed class and 239 belong to control class. From 3981 depressed users, 20.6% disclose their age in contrast with only 4% (239/4789) among control group. So self-disclosure of age is more prevalent among vulnerable users. Figure FIGREF18 depicts the age distribution in INLINEFORM2 . The general trend, consistent with the results in BIBREF42 , BIBREF49 , is biased toward young people. Indeed, according to Pew, 47% of Twitter users are younger than 30 years old BIBREF50 . Similar data collection procedure with comparable distribution have been used in many prior efforts BIBREF51 , BIBREF49 , BIBREF42 . We discuss our approach to mitigate the impact of the bias in Section 4.1. The median age is 17 for depressed class versus 19 for control class suggesting either likely depressed-user population is younger, or depressed youngsters are more likely to disclose their age for connecting to their peers (social homophily.) BIBREF51 ", + "Gender Enabled Ground-truth Dataset: We selected a subset of 1464 users INLINEFORM0 from INLINEFORM1 who disclose their gender in their profile description. From 1464 users 64% belonged to the depressed group, and the rest (36%) to the control group. 23% of the likely depressed users disclose their gender which is considerably higher (12%) than that for the control class. Once again, gender disclosure varies among the two gender groups. For statistical significance, we performed chi-square test (null hypothesis: gender and depression are two independent variables). Figure FIGREF19 illustrates gender association with each of the two classes. Blue circles (positive residuals, see Figure FIGREF19 -A,D) show positive association among corresponding row and column variables while red circles (negative residuals, see Figure FIGREF19 -B,C) imply a repulsion. Our findings are consistent with the medical literature BIBREF10 as according to BIBREF52 more women than men were given a diagnosis of depression. In particular, the female-to-male ratio is 2.1 and 1.9 for Major Depressive Disorder and Dysthymic Disorder respectively. Our findings from Twitter data indicate there is a strong association (Chi-square: 32.75, p-value:1.04e-08) between being female and showing depressive behavior on Twitter." + ], + [ + "We now provide an in-depth analysis of visual and textual content of vulnerable users.", + "Visual Content Analysis: We show that the visual content in images from posts as well as profiles provide valuable psychological cues for understanding a user's depression status. Profile/posted images can surface self-stigmatization BIBREF53 . Additionally, as opposed to typical computer vision framework for object recognition that often relies on thousands of predetermined low-level features, what matters more for assessing user's online behavior is the emotions reflected in facial expressions BIBREF54 , attributes contributing to the computational aesthetics BIBREF55 , and sentimental quotes they may subscribe to (Figure FIGREF15 ) BIBREF8 .", + "Facial Presence: ", + "For capturing facial presence, we rely on BIBREF56 's approach that uses multilevel convolutional coarse-to-fine network cascade to tackle facial landmark localization. We identify facial presentation, emotion from facial expression, and demographic features from profile/posted images . Table TABREF21 illustrates facial presentation differences in both profile and posted images (media) for depressed and control users in INLINEFORM0 . With control class showing significantly higher in both profile and media (8%, 9% respectively) compared to that for the depressed class. In contrast with age and gender disclosure, vulnerable users are less likely to disclose their facial identity, possibly due to lack of confidence or fear of stigma.", + "Facial Expression:", + "Following BIBREF8 's approach, we adopt Ekman's model of six emotions: anger, disgust, fear, joy, sadness and surprise, and use the Face++ API to automatically capture them from the shared images. Positive emotions are joy and surprise, and negative emotions are anger, disgust, fear, and sadness. In general, for each user u in INLINEFORM0 , we process profile/shared images for both the depressed and the control groups with at least one face from the shared images (Table TABREF23 ). For the photos that contain multiple faces, we measure the average emotion.", + "Figure FIGREF27 illustrates the inter-correlation of these features. Additionally, we observe that emotions gleaned from facial expressions correlated with emotional signals captured from textual content utilizing LIWC. This indicates visual imagery can be harnessed as a complementary channel for measuring online emotional signals.", + "General Image Features:", + "The importance of interpretable computational aesthetic features for studying users' online behavior has been highlighted by several efforts BIBREF55 , BIBREF8 , BIBREF57 . Color, as a pillar of the human vision system, has a strong association with conceptual ideas like emotion BIBREF58 , BIBREF59 . We measured the normalized red, green, blue and the mean of original colors, and brightness and contrast relative to variations of luminance. We represent images in Hue-Saturation-Value color space that seems intuitive for humans, and measure mean and variance for saturation and hue. Saturation is defined as the difference in the intensities of the different light wavelengths that compose the color. Although hue is not interpretable, high saturation indicates vividness and chromatic purity which are more appealing to the human eye BIBREF8 . Colorfulness is measured as a difference against gray background BIBREF60 . Naturalness is a measure of the degree of correspondence between images and the human perception of reality BIBREF60 . In color reproduction, naturalness is measured from the mental recollection of the colors of familiar objects. Additionally, there is a tendency among vulnerable users to share sentimental quotes bearing negative emotions. We performed optical character recognition (OCR) with python-tesseract to extract text and their sentiment score. As illustrated in Table TABREF26 , vulnerable users tend to use less colorful (higher grayscale) profile as well as shared images to convey their negative feelings, and share images that are less natural (Figure FIGREF15 ). With respect to the aesthetic quality of images (saturation, brightness, and hue), depressed users use images that are less appealing to the human eye. We employ independent t-test, while adopting Bonferroni Correction as a conservative approach to adjust the confidence intervals. Overall, we have 223 features, and choose Bonferroni-corrected INLINEFORM0 level of INLINEFORM1 (*** INLINEFORM2 , ** INLINEFORM3 ).", + "** alpha= 0.05, *** alpha = 0.05/223", + "Demographics Inference & Language Cues: LIWC has been used extensively for examining the latent dimensions of self-expression for analyzing personality BIBREF61 , depressive behavior, demographic differences BIBREF43 , BIBREF40 , etc. Several studies highlight that females employ more first-person singular pronouns BIBREF62 , and deictic language BIBREF63 , while males tend to use more articles BIBREF64 which characterizes concrete thinking, and formal, informational and affirmation words BIBREF65 . For age analysis, the salient findings include older individuals using more future tense verbs BIBREF62 triggering a shift in focus while aging. They also show positive emotions BIBREF66 and employ fewer self-references (i.e. 'I', 'me') with greater first person plural BIBREF62 . Depressed users employ first person pronouns more frequently BIBREF67 , repeatedly use negative emotions and anger words. We analyzed psycholinguistic cues and language style to study the association between depressive behavior as well as demographics. Particularly, we adopt Levinson's adult development grouping that partitions users in INLINEFORM0 into 5 age groups: (14,19],(19,23], (23,34],(34,46], and (46,60]. Then, we apply LIWC for characterizing linguistic styles for each age group for users in INLINEFORM1 .", + "Qualitative Language Analysis: The recent LIWC version summarizes textual content in terms of language variables such as analytical thinking, clout, authenticity, and emotional tone. It also measures other linguistic dimensions such as descriptors categories (e.g., percent of target words gleaned by dictionary, or longer than six letters - Sixltr) and informal language markers (e.g., swear words, netspeak), and other linguistic aspects (e.g., 1st person singular pronouns.)", + "Thinking Style:", + "Measuring people's natural ways of trying to analyze, and organize complex events have strong association with analytical thinking. LIWC relates higher analytic thinking to more formal and logical reasoning whereas a lower value indicates focus on narratives. Also, cognitive processing measures problem solving in mind. Words such as \"think,\" \"realize,\" and \"know\" indicates the degree of \"certainty\" in communications. Critical thinking ability relates to education BIBREF68 , and is impacted by different stages of cognitive development at different ages . It has been shown that older people communicate with greater cognitive complexity while comprehending nuances and subtle differences BIBREF62 . We observe a similar pattern in our data (Table TABREF40 .) A recent study highlights how depression affects brain and thinking at molecular level using a rat model BIBREF69 . Depression can promote cognitive dysfunction including difficulty in concentrating and making decisions. We observed a notable differences in the ability to think analytically in depressed and control users in different age groups (see Figure FIGREF39 - A, F and Table TABREF40 ). Overall, vulnerable younger users are not logical thinkers based on their relative analytical score and cognitive processing ability.", + "Authenticity:", + "Authenticity measures the degree of honesty. Authenticity is often assessed by measuring present tense verbs, 1st person singular pronouns (I, me, my), and by examining the linguistic manifestations of false stories BIBREF70 . Liars use fewer self-references and fewer complex words. Psychologists often see a child's first successfull lie as a mental growth. There is a decreasing trend of the Authenticity with aging (see Figure FIGREF39 -B.) Authenticity for depressed youngsters is strikingly higher than their control peers. It decreases with age (Figure FIGREF39 -B.)", + "Clout:", + "People with high clout speak more confidently and with certainty, employing more social words with fewer negations (e.g., no, not) and swear words. In general, midlife is relatively stable w.r.t. relationships and work. A recent study shows that age 60 to be best for self-esteem BIBREF71 as people take on managerial roles at work and maintain a satisfying relationship with their spouse. We see the same pattern in our data (see Figure FIGREF39 -C and Table TABREF40 ). Unsurprisingly, lack of confidence (the 6th PHQ-9 symptom) is a distinguishable characteristic of vulnerable users, leading to their lower clout scores, especially among depressed users before middle age (34 years old).", + "Self-references:", + "First person singular words are often seen as indicating interpersonal involvement and their high usage is associated with negative affective states implying nervousness and depression BIBREF66 . Consistent with prior studies, frequency of first person singular for depressed people is significantly higher compared to that of control class. Similarly to BIBREF66 , youngsters tend to use more first-person (e.g. I) and second person singular (e.g. you) pronouns (Figure FIGREF39 -G).", + "Informal Language Markers; Swear, Netspeak:", + "Several studies highlighted the use of profanity by young adults has significantly increased over the last decade BIBREF72 . We observed the same pattern in both the depressed and the control classes (Table TABREF40 ), although it's rate is higher for depressed users BIBREF1 . Psychologists have also shown that swearing can indicate that an individual is not a fragmented member of a society. Depressed youngsters, showing higher rate of interpersonal involvement and relationships, have a higher rate of cursing (Figure FIGREF39 -E). Also, Netspeak lexicon measures the frequency of terms such as lol and thx.", + "Sexual, Body: ", + "Sexual lexicon contains terms like \"horny\", \"love\" and \"incest\", and body terms like \"ache\", \"heart\", and \"cough\". Both start with a higher rate for depressed users while decreasing gradually while growing up, possibly due to changes in sexual desire as we age (Figure FIGREF39 -H,I and Table TABREF40 .)", + "Quantitative Language Analysis:", + "We employ one-way ANOVA to compare the impact of various factors and validate our findings above. Table TABREF40 illustrates our findings, with a degree of freedom (df) of 1055. The null hypothesis is that the sample means' for each age group are similar for each of the LIWC features.", + "*** alpha = 0.001, ** alpha = 0.01, * alpha = 0.05" + ], + [ + "We leverage both the visual and textual content for predicting age and gender.", + "Prediction with Textual Content:", + "We employ BIBREF73 's weighted lexicon of terms that uses the dataset of 75,394 Facebook users who shared their status, age and gender. The predictive power of this lexica was evaluated on Twitter, blog, and Facebook, showing promising results BIBREF73 . Utilizing these two weighted lexicon of terms, we are predicting the demographic information (age or gender) of INLINEFORM0 (denoted by INLINEFORM1 ) using following equation: INLINEFORM2 ", + "where INLINEFORM0 is the lexicon weight of the term, and INLINEFORM1 represents the frequency of the term in the user generated INLINEFORM2 , and INLINEFORM3 measures total word count in INLINEFORM4 . As our data is biased toward young people, we report age prediction performance for each age group separately (Table TABREF42 ). Moreover, to measure the average accuracy of this model, we build a balanced dataset (keeping all the users above 23 -416 users), and then randomly sampling the same number of users from the age ranges (11,19] and (19,23]. The average accuracy of this model is 0.63 for depressed users and 0.64 for control class. Table TABREF44 illustrates the performance of gender prediction for each class. The average accuracy is 0.82 on INLINEFORM5 ground-truth dataset.", + "Prediction with Visual Imagery:", + "Inspired by BIBREF56 's approach for facial landmark localization, we use their pretrained CNN consisting of convolutional layers, including unshared and fully-connected layers, to predict gender and age from both the profile and shared images. We evaluate the performance for gender and age prediction task on INLINEFORM0 and INLINEFORM1 respectively as shown in Table TABREF42 and Table TABREF44 .", + "Demographic Prediction Analysis:", + "We delve deeper into the benefits and drawbacks of each data modality for demographic information prediction. This is crucial as the differences between language cues between age groups above age 35 tend to become smaller (see Figure FIGREF39 -A,B,C) and making the prediction harder for older people BIBREF74 . In this case, the other data modality (e.g., visual content) can play integral role as a complementary source for age inference. For gender prediction (see Table TABREF44 ), on average, the profile image-based predictor provides a more accurate prediction for both the depressed and control class (0.92 and 0.90) compared to content-based predictor (0.82). For age prediction (see Table TABREF42 ), textual content-based predictor (on average 0.60) outperforms both of the visual-based predictors (on average profile:0.51, Media:0.53).", + "However, not every user provides facial identity on his account (see Table TABREF21 ). We studied facial presentation for each age-group to examine any association between age-group, facial presentation and depressive behavior (see Table TABREF43 ). We can see youngsters in both depressed and control class are not likely to present their face on profile image. Less than 3% of vulnerable users between 11-19 years reveal their facial identity. Although content-based gender predictor was not as accurate as image-based one, it is adequate for population-level analysis." + ], + [ + "We use the above findings for predicting depressive behavior. Our model exploits early fusion BIBREF32 technique in feature space and requires modeling each user INLINEFORM0 in INLINEFORM1 as vector concatenation of individual modality features. As opposed to computationally expensive late fusion scheme where each modality requires a separate supervised modeling, this model reduces the learning effort and shows promising results BIBREF75 . To develop a generalizable model that avoids overfitting, we perform feature selection using statistical tests and all relevant ensemble learning models. It adds randomness to the data by creating shuffled copies of all features (shadow feature), and then trains Random Forest classifier on the extended data. Iteratively, it checks whether the actual feature has a higher Z-score than its shadow feature (See Algorithm SECREF6 and Figure FIGREF45 ) BIBREF76 .", + "Main each Feature INLINEFORM0 INLINEFORM1 ", + "RndForrest( INLINEFORM0 ) Calculate Imp INLINEFORM1 INLINEFORM2 Generate next hypothesis , INLINEFORM3 Once all hypothesis generated Perform Statistical Test INLINEFORM4 //Binomial Distribution INLINEFORM5 Feature is important Feature is important", + " Ensemble Feature Selection", + "Next, we adopt an ensemble learning method that integrates the predictive power of multiple learners with two main advantages; its interpretability with respect to the contributions of each feature and its high predictive power. For prediction we have INLINEFORM0 where INLINEFORM1 is a weak learner and INLINEFORM2 denotes the final prediction.", + "In particular, we optimize the loss function: INLINEFORM0 where INLINEFORM1 incorporates INLINEFORM2 and INLINEFORM3 regularization. In each iteration, the new INLINEFORM4 is obtained by fitting weak learner to the negative gradient of loss function. Particularly, by estimating the loss function with Taylor expansion : INLINEFORM5 where its first expression is constant, the second and the third expressions are first ( INLINEFORM6 ) and second order derivatives ( INLINEFORM7 ) of the loss. INLINEFORM8 ", + "For exploring the weak learners, assume INLINEFORM0 has k leaf nodes, INLINEFORM1 be subset of users from INLINEFORM2 belongs to the node INLINEFORM3 , and INLINEFORM4 denotes the prediction for node INLINEFORM5 . Then, for each user INLINEFORM6 belonging to INLINEFORM7 , INLINEFORM8 and INLINEFORM9 INLINEFORM10 ", + "Next, for each leaf node INLINEFORM0 , deriving w.r.t INLINEFORM1 : INLINEFORM2 ", + "and by substituting weights: INLINEFORM0 ", + "which represents the loss for fixed weak learners with INLINEFORM0 nodes. The trees are built sequentially such that each subsequent tree aims to reduce the errors of its predecessor tree. Although, the weak learners have high bias, the ensemble model produces a strong learner that effectively integrate the weak learners by reducing bias and variance (the ultimate goal of supervised models) BIBREF77 . Table TABREF48 illustrates our multimodal framework outperform the baselines for identifying depressed users in terms of average specificity, sensitivity, F-Measure, and accuracy in 10-fold cross-validation setting on INLINEFORM1 dataset. Figure FIGREF47 shows how the likelihood of being classified into the depressed class varies with each feature addition to the model for a sample user in the dataset. The prediction bar (the black bar) shows that the log-odds of prediction is 0.31, that is, the likelihood of this person being a depressed user is 57% (1 / (1 + exp(-0.3))). The figure also sheds light on the impact of each contributing feature. The waterfall charts represent how the probability of being depressed changes with the addition of each feature variable. For instance, the \"Analytic thinking\" of this user is considered high 48.43 (Median:36.95, Mean: 40.18) and this decreases the chance of this person being classified into the depressed group by the log-odds of -1.41. Depressed users have significantly lower \"Analytic thinking\" score compared to control class. Moreover, the 40.46 \"Clout\" score is a low value (Median: 62.22, Mean: 57.17) and it decreases the chance of being classified as depressed. With respect to the visual features, for instance, the mean and the median of 'shared_colorfulness' is 112.03 and 113 respectively. The value of 136.71 would be high; thus, it decreases the chance of being depressed for this specific user by log-odds of -0.54. Moreover, the 'profile_naturalness' of 0.46 is considered high compared to 0.36 as the mean for the depressed class which justifies pull down of the log-odds by INLINEFORM2 . For network features, for instance, 'two_hop_neighborhood' for depressed users (Mean : 84) are less than that of control users (Mean: 154), and is reflected in pulling down the log-odds by -0.27.", + "Baselines:", + "To test the efficacy of our multi-modal framework for detecting depressed users, we compare it against existing content, content-network, and image-based models (based on the aforementioned general image feature, facial presence, and facial expressions.)" + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0452/instruction.md b/qasper-0452/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..83e141270e6d574b2a937079058a117fbcde2885 --- /dev/null +++ b/qasper-0452/instruction.md @@ -0,0 +1,105 @@ +Name of Paper: Fusing Visual, Textual and Connectivity Clues for Studying Mental Health + +Question: What is the source of the visual data? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + null, + "Introduction", + "Related Work", + "Dataset", + "Data Modality Analysis", + "Demographic Prediction", + "Multi-modal Prediction Framework" + ], + "paragraphs": [ + [ + "0pt*0*0", + "0pt*0*0", + "0pt*0*0 0.95", + "1]Amir Hossein Yazdavar 1]Mohammad Saeid Mahdavinejad 2]Goonmeet Bajaj", + " 3]William Romine 1]Amirhassan Monadjemi 1]Krishnaprasad Thirunarayan", + " 1]Amit Sheth 4]Jyotishman Pathak [1]Department of Computer Science & Engineering, Wright State University, OH, USA [2]Ohio State University, Columbus, OH, USA [3]Department of Biological Science, Wright State University, OH, USA [4] Division of Health Informatics, Weill Cornell University, New York, NY, USA", + "[1] yazdavar.2@wright.edu", + "With ubiquity of social media platforms, millions of people are sharing their online persona by expressing their thoughts, moods, emotions, feelings, and even their daily struggles with mental health issues voluntarily and publicly on social media. Unlike the most existing efforts which study depression by analyzing textual content, we examine and exploit multimodal big data to discern depressive behavior using a wide variety of features including individual-level demographics. By developing a multimodal framework and employing statistical techniques for fusing heterogeneous sets of features obtained by processing visual, textual and user interaction data, we significantly enhance the current state-of-the-art approaches for identifying depressed individuals on Twitter (improving the average F1-Score by 5 percent) as well as facilitate demographic inference from social media for broader applications. Besides providing insights into the relationship between demographics and mental health, our research assists in the design of a new breed of demographic-aware health interventions." + ], + [ + "Depression is a highly prevalent public health challenge and a major cause of disability worldwide. Depression affects 6.7% (i.e., about 16 million) Americans each year . According to the World Mental Health Survey conducted in 17 countries, on average, about 5% of people reported having an episode of depression in 2011 BIBREF0 . Untreated or under-treated clinical depression can lead to suicide and other chronic risky behaviors such as drug or alcohol addiction.", + "Global efforts to curb clinical depression involve identifying depression through survey-based methods employing phone or online questionnaires. These approaches suffer from under-representation as well as sampling bias (with very small group of respondents.) In contrast, the widespread adoption of social media where people voluntarily and publicly express their thoughts, moods, emotions, and feelings, and even share their daily struggles with mental health problems has not been adequately tapped into studying mental illnesses, such as depression. The visual and textual content shared on different social media platforms like Twitter offer new opportunities for a deeper understanding of self-expressed depression both at an individual as well as community-level. Previous research efforts have suggested that language style, sentiment, users' activities, and engagement expressed in social media posts can predict the likelihood of depression BIBREF1 , BIBREF2 . However, except for a few attempts BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , these investigations have seldom studied extraction of emotional state from visual content of images in posted/profile images. Visual content can express users' emotions more vividly, and psychologists noted that imagery is an effective medium for communicating difficult emotions.", + "According to eMarketer, photos accounted for 75% of content posted on Facebook worldwide and they are the most engaging type of content on Facebook (87%). Indeed, \"a picture is worth a thousand words\" and now \"photos are worth a million likes.\" Similarly, on Twitter, the tweets with image links get twice as much attention as those without , and video-linked tweets drive up engagement . The ease and naturalness of expression through visual imagery can serve to glean depression-indicators in vulnerable individuals who often seek social support through social media BIBREF7 . Further, as psychologist Carl Rogers highlights, we often pursue and promote our Ideal-Self . In this regard, the choice of profile image can be a proxy for the online persona BIBREF8 , providing a window into an individual's mental health status. For instance, choosing emaciated legs of girls covered with several cuts as profile image portrays negative self-view BIBREF9 .", + "Inferring demographic information like gender and age can be crucial for stratifying our understanding of population-level epidemiology of mental health disorders. Relying on electronic health records data, previous studies explored gender differences in depressive behavior from different angles including prevalence, age at onset, comorbidities, as well as biological and psychosocial factors. For instance, women have been diagnosed with depression twice as often as men BIBREF10 and national psychiatric morbidity survey in Britain has shown higher risk of depression in women BIBREF11 . On the other hand, suicide rates for men are three to five times higher compared to that of the women BIBREF12 .", + "Although depression can affect anyone at any age, signs and triggers of depression vary for different age groups . Depression triggers for children include parental depression, domestic violence, and loss of a pet, friend or family member. For teenagers (ages 12-18), depression may arise from hormonal imbalance, sexuality concerns and rejection by peers. Young adults (ages 19-29) may develop depression due to life transitions, poverty, trauma, and work issues. Adult (ages 30-60) depression triggers include caring simultaneously for children and aging parents, financial burden, work and relationship issues. Senior adults develop depression from common late-life issues, social isolation, major life loses such as the death of a spouse, financial stress and other chronic health problems (e.g., cardiac disease, dementia). Therefore, inferring demographic information while studying depressive behavior from passively sensed social data, can shed better light on the population-level epidemiology of depression.", + "The recent advancements in deep neural networks, specifically for image analysis task, can lead to determining demographic features such as age and gender BIBREF13 . We show that by determining and integrating heterogeneous set of features from different modalities \u2013 aesthetic features from posted images (colorfulness, hue variance, sharpness, brightness, blurriness, naturalness), choice of profile picture (for gender, age, and facial expression), the screen name, the language features from both textual content and profile's description (n-gram, emotion, sentiment), and finally sociability from ego-network, and user engagement \u2013 we can reliably detect likely depressed individuals in a data set of 8,770 human-annotated Twitter users.", + "We address and derive answers to the following research questions: 1) How well do the content of posted images (colors, aesthetic and facial presentation) reflect depressive behavior? 2) Does the choice of profile picture show any psychological traits of depressed online persona? Are they reliable enough to represent the demographic information such as age and gender? 3) Are there any underlying common themes among depressed individuals generated using multimodal content that can be used to detect depression reliably?" + ], + [ + "Mental Health Analysis using Social Media:", + "Several efforts have attempted to automatically detect depression from social media content utilizing machine/deep learning and natural language processing approaches. Conducting a retrospective study over tweets, BIBREF14 characterizes depression based on factors such as language, emotion, style, ego-network, and user engagement. They built a classifier to predict the likelihood of depression in a post BIBREF14 , BIBREF15 or in an individual BIBREF1 , BIBREF16 , BIBREF17 , BIBREF18 . Moreover, there have been significant advances due to the shared task BIBREF19 focusing on methods for identifying depressed users on Twitter at the Computational Linguistics and Clinical Psychology Workshop (CLP 2015). A corpus of nearly 1,800 Twitter users was built for evaluation, and the best models employed topic modeling BIBREF20 , Linguistic Inquiry and Word Count (LIWC) features, and other metadata BIBREF21 . More recently, a neural network architecture introduced by BIBREF22 combined posts into a representation of user's activities for detecting depressed users. Another active line of research has focused on capturing suicide and self-harm signals BIBREF23 , BIBREF24 , BIBREF25 , BIBREF26 , BIBREF2 , BIBREF27 . Moreover, the CLP 2016 BIBREF28 defined a shared task on detecting the severity of the mental health from forum posts. All of these studies derive discriminative features to classify depression in user-generated content at message-level, individual-level or community-level. Recent emergence of photo-sharing platforms such as Instagram, has attracted researchers attention to study people's behavior from their visual narratives \u2013 ranging from mining their emotions BIBREF29 , and happiness trend BIBREF30 , to studying medical concerns BIBREF31 . Researchers show that people use Instagram to engage in social exchange and storytelling about their difficult experiences BIBREF4 . The role of visual imagery as a mechanism of self-disclosure by relating visual attributes to mental health disclosures on Instagram was highlighted by BIBREF3 , BIBREF5 where individual Instagram profiles were utilized to build a prediction framework for identifying markers of depression. The importance of data modality to understand user behavior on social media was highlighted by BIBREF32 . More recently, a deep neural network sequence modeling approach that marries audio and text data modalities to analyze question-answer style interviews between an individual and an agent has been developed to study mental health BIBREF32 . Similarly, a multimodal depressive dictionary learning was proposed to detect depressed users on Twitter BIBREF33 . They provide a sparse user representations by defining a feature set consisting of social network features, user profile features, visual features, emotional features BIBREF34 , topic-level features, and domain-specific features. Particularly, our choice of multi-model prediction framework is intended to improve upon the prior works involving use of images in multimodal depression analysis BIBREF33 and prior works on studying Instagram photos BIBREF6 , BIBREF35 .", + "Demographic information inference on Social Media: ", + "There is a growing interest in understanding online user's demographic information due to its numerous applications in healthcare BIBREF36 , BIBREF37 . A supervised model developed by BIBREF38 for determining users' gender by employing features such as screen-name, full-name, profile description and content on external resources (e.g., personal blog). Employing features including emoticons, acronyms, slangs, punctuations, capitalization, sentence length and included links/images, along with online behaviors such as number of friends, post time, and commenting activity, a supervised model was built for predicting user's age group BIBREF39 . Utilizing users life stage information such as secondary school student, college student, and employee, BIBREF40 builds age inference model for Dutch Twitter users. Similarly, relying on profile descriptions while devising a set of rules and patterns, a novel model introduced for extracting age for Twitter users BIBREF41 . They also parse description for occupation by consulting the SOC2010 list of occupations and validating it through social surveys. A novel age inference model was developed while relying on homophily interaction information and content for predicting age of Twitter users BIBREF42 . The limitations of textual content for predicting age and gender was highlighted by BIBREF43 . They distinguish language use based on social gender, age identity, biological sex and chronological age by collecting crowdsourced signals using a game in which players (crowd) guess the biological sex and age of a user based only on their tweets. Their findings indicate how linguistic markers can misguide (e.g., a heart represented as <3 can be misinterpreted as feminine when the writer is male.) Estimating age and gender from facial images by training a convolutional neural networks (CNN) for face recognition is an active line of research BIBREF44 , BIBREF13 , BIBREF45 ." + ], + [ + "Self-disclosure clues have been extensively utilized for creating ground-truth data for numerous social media analytic studies e.g., for predicting demographics BIBREF36 , BIBREF41 , and user's depressive behavior BIBREF46 , BIBREF47 , BIBREF48 . For instance, vulnerable individuals may employ depressive-indicative terms in their Twitter profile descriptions. Others may share their age and gender, e.g., \"16 years old suicidal girl\"(see Figure FIGREF15 ). We employ a huge dataset of 45,000 self-reported depressed users introduced in BIBREF46 where a lexicon of depression symptoms consisting of 1500 depression-indicative terms was created with the help of psychologist clinician and employed for collecting self-declared depressed individual's profiles. A subset of 8,770 users (24 million time-stamped tweets) containing 3981 depressed and 4789 control users (that do not show any depressive behavior) were verified by two human judges BIBREF46 . This dataset INLINEFORM0 contains the metadata values of each user such as profile descriptions, followers_count, created_at, and profile_image_url.", + "Age Enabled Ground-truth Dataset: We extract user's age by applying regular expression patterns to profile descriptions (such as \"17 years old, self-harm, anxiety, depression\") BIBREF41 . We compile \"age prefixes\" and \"age suffixes\", and use three age-extraction rules: 1. I am X years old 2. Born in X 3. X years old, where X is a \"date\" or age (e.g., 1994). We selected a subset of 1061 users among INLINEFORM0 as gold standard dataset INLINEFORM1 who disclose their age. From these 1061 users, 822 belong to depressed class and 239 belong to control class. From 3981 depressed users, 20.6% disclose their age in contrast with only 4% (239/4789) among control group. So self-disclosure of age is more prevalent among vulnerable users. Figure FIGREF18 depicts the age distribution in INLINEFORM2 . The general trend, consistent with the results in BIBREF42 , BIBREF49 , is biased toward young people. Indeed, according to Pew, 47% of Twitter users are younger than 30 years old BIBREF50 . Similar data collection procedure with comparable distribution have been used in many prior efforts BIBREF51 , BIBREF49 , BIBREF42 . We discuss our approach to mitigate the impact of the bias in Section 4.1. The median age is 17 for depressed class versus 19 for control class suggesting either likely depressed-user population is younger, or depressed youngsters are more likely to disclose their age for connecting to their peers (social homophily.) BIBREF51 ", + "Gender Enabled Ground-truth Dataset: We selected a subset of 1464 users INLINEFORM0 from INLINEFORM1 who disclose their gender in their profile description. From 1464 users 64% belonged to the depressed group, and the rest (36%) to the control group. 23% of the likely depressed users disclose their gender which is considerably higher (12%) than that for the control class. Once again, gender disclosure varies among the two gender groups. For statistical significance, we performed chi-square test (null hypothesis: gender and depression are two independent variables). Figure FIGREF19 illustrates gender association with each of the two classes. Blue circles (positive residuals, see Figure FIGREF19 -A,D) show positive association among corresponding row and column variables while red circles (negative residuals, see Figure FIGREF19 -B,C) imply a repulsion. Our findings are consistent with the medical literature BIBREF10 as according to BIBREF52 more women than men were given a diagnosis of depression. In particular, the female-to-male ratio is 2.1 and 1.9 for Major Depressive Disorder and Dysthymic Disorder respectively. Our findings from Twitter data indicate there is a strong association (Chi-square: 32.75, p-value:1.04e-08) between being female and showing depressive behavior on Twitter." + ], + [ + "We now provide an in-depth analysis of visual and textual content of vulnerable users.", + "Visual Content Analysis: We show that the visual content in images from posts as well as profiles provide valuable psychological cues for understanding a user's depression status. Profile/posted images can surface self-stigmatization BIBREF53 . Additionally, as opposed to typical computer vision framework for object recognition that often relies on thousands of predetermined low-level features, what matters more for assessing user's online behavior is the emotions reflected in facial expressions BIBREF54 , attributes contributing to the computational aesthetics BIBREF55 , and sentimental quotes they may subscribe to (Figure FIGREF15 ) BIBREF8 .", + "Facial Presence: ", + "For capturing facial presence, we rely on BIBREF56 's approach that uses multilevel convolutional coarse-to-fine network cascade to tackle facial landmark localization. We identify facial presentation, emotion from facial expression, and demographic features from profile/posted images . Table TABREF21 illustrates facial presentation differences in both profile and posted images (media) for depressed and control users in INLINEFORM0 . With control class showing significantly higher in both profile and media (8%, 9% respectively) compared to that for the depressed class. In contrast with age and gender disclosure, vulnerable users are less likely to disclose their facial identity, possibly due to lack of confidence or fear of stigma.", + "Facial Expression:", + "Following BIBREF8 's approach, we adopt Ekman's model of six emotions: anger, disgust, fear, joy, sadness and surprise, and use the Face++ API to automatically capture them from the shared images. Positive emotions are joy and surprise, and negative emotions are anger, disgust, fear, and sadness. In general, for each user u in INLINEFORM0 , we process profile/shared images for both the depressed and the control groups with at least one face from the shared images (Table TABREF23 ). For the photos that contain multiple faces, we measure the average emotion.", + "Figure FIGREF27 illustrates the inter-correlation of these features. Additionally, we observe that emotions gleaned from facial expressions correlated with emotional signals captured from textual content utilizing LIWC. This indicates visual imagery can be harnessed as a complementary channel for measuring online emotional signals.", + "General Image Features:", + "The importance of interpretable computational aesthetic features for studying users' online behavior has been highlighted by several efforts BIBREF55 , BIBREF8 , BIBREF57 . Color, as a pillar of the human vision system, has a strong association with conceptual ideas like emotion BIBREF58 , BIBREF59 . We measured the normalized red, green, blue and the mean of original colors, and brightness and contrast relative to variations of luminance. We represent images in Hue-Saturation-Value color space that seems intuitive for humans, and measure mean and variance for saturation and hue. Saturation is defined as the difference in the intensities of the different light wavelengths that compose the color. Although hue is not interpretable, high saturation indicates vividness and chromatic purity which are more appealing to the human eye BIBREF8 . Colorfulness is measured as a difference against gray background BIBREF60 . Naturalness is a measure of the degree of correspondence between images and the human perception of reality BIBREF60 . In color reproduction, naturalness is measured from the mental recollection of the colors of familiar objects. Additionally, there is a tendency among vulnerable users to share sentimental quotes bearing negative emotions. We performed optical character recognition (OCR) with python-tesseract to extract text and their sentiment score. As illustrated in Table TABREF26 , vulnerable users tend to use less colorful (higher grayscale) profile as well as shared images to convey their negative feelings, and share images that are less natural (Figure FIGREF15 ). With respect to the aesthetic quality of images (saturation, brightness, and hue), depressed users use images that are less appealing to the human eye. We employ independent t-test, while adopting Bonferroni Correction as a conservative approach to adjust the confidence intervals. Overall, we have 223 features, and choose Bonferroni-corrected INLINEFORM0 level of INLINEFORM1 (*** INLINEFORM2 , ** INLINEFORM3 ).", + "** alpha= 0.05, *** alpha = 0.05/223", + "Demographics Inference & Language Cues: LIWC has been used extensively for examining the latent dimensions of self-expression for analyzing personality BIBREF61 , depressive behavior, demographic differences BIBREF43 , BIBREF40 , etc. Several studies highlight that females employ more first-person singular pronouns BIBREF62 , and deictic language BIBREF63 , while males tend to use more articles BIBREF64 which characterizes concrete thinking, and formal, informational and affirmation words BIBREF65 . For age analysis, the salient findings include older individuals using more future tense verbs BIBREF62 triggering a shift in focus while aging. They also show positive emotions BIBREF66 and employ fewer self-references (i.e. 'I', 'me') with greater first person plural BIBREF62 . Depressed users employ first person pronouns more frequently BIBREF67 , repeatedly use negative emotions and anger words. We analyzed psycholinguistic cues and language style to study the association between depressive behavior as well as demographics. Particularly, we adopt Levinson's adult development grouping that partitions users in INLINEFORM0 into 5 age groups: (14,19],(19,23], (23,34],(34,46], and (46,60]. Then, we apply LIWC for characterizing linguistic styles for each age group for users in INLINEFORM1 .", + "Qualitative Language Analysis: The recent LIWC version summarizes textual content in terms of language variables such as analytical thinking, clout, authenticity, and emotional tone. It also measures other linguistic dimensions such as descriptors categories (e.g., percent of target words gleaned by dictionary, or longer than six letters - Sixltr) and informal language markers (e.g., swear words, netspeak), and other linguistic aspects (e.g., 1st person singular pronouns.)", + "Thinking Style:", + "Measuring people's natural ways of trying to analyze, and organize complex events have strong association with analytical thinking. LIWC relates higher analytic thinking to more formal and logical reasoning whereas a lower value indicates focus on narratives. Also, cognitive processing measures problem solving in mind. Words such as \"think,\" \"realize,\" and \"know\" indicates the degree of \"certainty\" in communications. Critical thinking ability relates to education BIBREF68 , and is impacted by different stages of cognitive development at different ages . It has been shown that older people communicate with greater cognitive complexity while comprehending nuances and subtle differences BIBREF62 . We observe a similar pattern in our data (Table TABREF40 .) A recent study highlights how depression affects brain and thinking at molecular level using a rat model BIBREF69 . Depression can promote cognitive dysfunction including difficulty in concentrating and making decisions. We observed a notable differences in the ability to think analytically in depressed and control users in different age groups (see Figure FIGREF39 - A, F and Table TABREF40 ). Overall, vulnerable younger users are not logical thinkers based on their relative analytical score and cognitive processing ability.", + "Authenticity:", + "Authenticity measures the degree of honesty. Authenticity is often assessed by measuring present tense verbs, 1st person singular pronouns (I, me, my), and by examining the linguistic manifestations of false stories BIBREF70 . Liars use fewer self-references and fewer complex words. Psychologists often see a child's first successfull lie as a mental growth. There is a decreasing trend of the Authenticity with aging (see Figure FIGREF39 -B.) Authenticity for depressed youngsters is strikingly higher than their control peers. It decreases with age (Figure FIGREF39 -B.)", + "Clout:", + "People with high clout speak more confidently and with certainty, employing more social words with fewer negations (e.g., no, not) and swear words. In general, midlife is relatively stable w.r.t. relationships and work. A recent study shows that age 60 to be best for self-esteem BIBREF71 as people take on managerial roles at work and maintain a satisfying relationship with their spouse. We see the same pattern in our data (see Figure FIGREF39 -C and Table TABREF40 ). Unsurprisingly, lack of confidence (the 6th PHQ-9 symptom) is a distinguishable characteristic of vulnerable users, leading to their lower clout scores, especially among depressed users before middle age (34 years old).", + "Self-references:", + "First person singular words are often seen as indicating interpersonal involvement and their high usage is associated with negative affective states implying nervousness and depression BIBREF66 . Consistent with prior studies, frequency of first person singular for depressed people is significantly higher compared to that of control class. Similarly to BIBREF66 , youngsters tend to use more first-person (e.g. I) and second person singular (e.g. you) pronouns (Figure FIGREF39 -G).", + "Informal Language Markers; Swear, Netspeak:", + "Several studies highlighted the use of profanity by young adults has significantly increased over the last decade BIBREF72 . We observed the same pattern in both the depressed and the control classes (Table TABREF40 ), although it's rate is higher for depressed users BIBREF1 . Psychologists have also shown that swearing can indicate that an individual is not a fragmented member of a society. Depressed youngsters, showing higher rate of interpersonal involvement and relationships, have a higher rate of cursing (Figure FIGREF39 -E). Also, Netspeak lexicon measures the frequency of terms such as lol and thx.", + "Sexual, Body: ", + "Sexual lexicon contains terms like \"horny\", \"love\" and \"incest\", and body terms like \"ache\", \"heart\", and \"cough\". Both start with a higher rate for depressed users while decreasing gradually while growing up, possibly due to changes in sexual desire as we age (Figure FIGREF39 -H,I and Table TABREF40 .)", + "Quantitative Language Analysis:", + "We employ one-way ANOVA to compare the impact of various factors and validate our findings above. Table TABREF40 illustrates our findings, with a degree of freedom (df) of 1055. The null hypothesis is that the sample means' for each age group are similar for each of the LIWC features.", + "*** alpha = 0.001, ** alpha = 0.01, * alpha = 0.05" + ], + [ + "We leverage both the visual and textual content for predicting age and gender.", + "Prediction with Textual Content:", + "We employ BIBREF73 's weighted lexicon of terms that uses the dataset of 75,394 Facebook users who shared their status, age and gender. The predictive power of this lexica was evaluated on Twitter, blog, and Facebook, showing promising results BIBREF73 . Utilizing these two weighted lexicon of terms, we are predicting the demographic information (age or gender) of INLINEFORM0 (denoted by INLINEFORM1 ) using following equation: INLINEFORM2 ", + "where INLINEFORM0 is the lexicon weight of the term, and INLINEFORM1 represents the frequency of the term in the user generated INLINEFORM2 , and INLINEFORM3 measures total word count in INLINEFORM4 . As our data is biased toward young people, we report age prediction performance for each age group separately (Table TABREF42 ). Moreover, to measure the average accuracy of this model, we build a balanced dataset (keeping all the users above 23 -416 users), and then randomly sampling the same number of users from the age ranges (11,19] and (19,23]. The average accuracy of this model is 0.63 for depressed users and 0.64 for control class. Table TABREF44 illustrates the performance of gender prediction for each class. The average accuracy is 0.82 on INLINEFORM5 ground-truth dataset.", + "Prediction with Visual Imagery:", + "Inspired by BIBREF56 's approach for facial landmark localization, we use their pretrained CNN consisting of convolutional layers, including unshared and fully-connected layers, to predict gender and age from both the profile and shared images. We evaluate the performance for gender and age prediction task on INLINEFORM0 and INLINEFORM1 respectively as shown in Table TABREF42 and Table TABREF44 .", + "Demographic Prediction Analysis:", + "We delve deeper into the benefits and drawbacks of each data modality for demographic information prediction. This is crucial as the differences between language cues between age groups above age 35 tend to become smaller (see Figure FIGREF39 -A,B,C) and making the prediction harder for older people BIBREF74 . In this case, the other data modality (e.g., visual content) can play integral role as a complementary source for age inference. For gender prediction (see Table TABREF44 ), on average, the profile image-based predictor provides a more accurate prediction for both the depressed and control class (0.92 and 0.90) compared to content-based predictor (0.82). For age prediction (see Table TABREF42 ), textual content-based predictor (on average 0.60) outperforms both of the visual-based predictors (on average profile:0.51, Media:0.53).", + "However, not every user provides facial identity on his account (see Table TABREF21 ). We studied facial presentation for each age-group to examine any association between age-group, facial presentation and depressive behavior (see Table TABREF43 ). We can see youngsters in both depressed and control class are not likely to present their face on profile image. Less than 3% of vulnerable users between 11-19 years reveal their facial identity. Although content-based gender predictor was not as accurate as image-based one, it is adequate for population-level analysis." + ], + [ + "We use the above findings for predicting depressive behavior. Our model exploits early fusion BIBREF32 technique in feature space and requires modeling each user INLINEFORM0 in INLINEFORM1 as vector concatenation of individual modality features. As opposed to computationally expensive late fusion scheme where each modality requires a separate supervised modeling, this model reduces the learning effort and shows promising results BIBREF75 . To develop a generalizable model that avoids overfitting, we perform feature selection using statistical tests and all relevant ensemble learning models. It adds randomness to the data by creating shuffled copies of all features (shadow feature), and then trains Random Forest classifier on the extended data. Iteratively, it checks whether the actual feature has a higher Z-score than its shadow feature (See Algorithm SECREF6 and Figure FIGREF45 ) BIBREF76 .", + "Main each Feature INLINEFORM0 INLINEFORM1 ", + "RndForrest( INLINEFORM0 ) Calculate Imp INLINEFORM1 INLINEFORM2 Generate next hypothesis , INLINEFORM3 Once all hypothesis generated Perform Statistical Test INLINEFORM4 //Binomial Distribution INLINEFORM5 Feature is important Feature is important", + " Ensemble Feature Selection", + "Next, we adopt an ensemble learning method that integrates the predictive power of multiple learners with two main advantages; its interpretability with respect to the contributions of each feature and its high predictive power. For prediction we have INLINEFORM0 where INLINEFORM1 is a weak learner and INLINEFORM2 denotes the final prediction.", + "In particular, we optimize the loss function: INLINEFORM0 where INLINEFORM1 incorporates INLINEFORM2 and INLINEFORM3 regularization. In each iteration, the new INLINEFORM4 is obtained by fitting weak learner to the negative gradient of loss function. Particularly, by estimating the loss function with Taylor expansion : INLINEFORM5 where its first expression is constant, the second and the third expressions are first ( INLINEFORM6 ) and second order derivatives ( INLINEFORM7 ) of the loss. INLINEFORM8 ", + "For exploring the weak learners, assume INLINEFORM0 has k leaf nodes, INLINEFORM1 be subset of users from INLINEFORM2 belongs to the node INLINEFORM3 , and INLINEFORM4 denotes the prediction for node INLINEFORM5 . Then, for each user INLINEFORM6 belonging to INLINEFORM7 , INLINEFORM8 and INLINEFORM9 INLINEFORM10 ", + "Next, for each leaf node INLINEFORM0 , deriving w.r.t INLINEFORM1 : INLINEFORM2 ", + "and by substituting weights: INLINEFORM0 ", + "which represents the loss for fixed weak learners with INLINEFORM0 nodes. The trees are built sequentially such that each subsequent tree aims to reduce the errors of its predecessor tree. Although, the weak learners have high bias, the ensemble model produces a strong learner that effectively integrate the weak learners by reducing bias and variance (the ultimate goal of supervised models) BIBREF77 . Table TABREF48 illustrates our multimodal framework outperform the baselines for identifying depressed users in terms of average specificity, sensitivity, F-Measure, and accuracy in 10-fold cross-validation setting on INLINEFORM1 dataset. Figure FIGREF47 shows how the likelihood of being classified into the depressed class varies with each feature addition to the model for a sample user in the dataset. The prediction bar (the black bar) shows that the log-odds of prediction is 0.31, that is, the likelihood of this person being a depressed user is 57% (1 / (1 + exp(-0.3))). The figure also sheds light on the impact of each contributing feature. The waterfall charts represent how the probability of being depressed changes with the addition of each feature variable. For instance, the \"Analytic thinking\" of this user is considered high 48.43 (Median:36.95, Mean: 40.18) and this decreases the chance of this person being classified into the depressed group by the log-odds of -1.41. Depressed users have significantly lower \"Analytic thinking\" score compared to control class. Moreover, the 40.46 \"Clout\" score is a low value (Median: 62.22, Mean: 57.17) and it decreases the chance of being classified as depressed. With respect to the visual features, for instance, the mean and the median of 'shared_colorfulness' is 112.03 and 113 respectively. The value of 136.71 would be high; thus, it decreases the chance of being depressed for this specific user by log-odds of -0.54. Moreover, the 'profile_naturalness' of 0.46 is considered high compared to 0.36 as the mean for the depressed class which justifies pull down of the log-odds by INLINEFORM2 . For network features, for instance, 'two_hop_neighborhood' for depressed users (Mean : 84) are less than that of control users (Mean: 154), and is reflected in pulling down the log-odds by -0.27.", + "Baselines:", + "To test the efficacy of our multi-modal framework for detecting depressed users, we compare it against existing content, content-network, and image-based models (based on the aforementioned general image feature, facial presence, and facial expressions.)" + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0454/instruction.md b/qasper-0454/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..d4c3545ff9cfa67789c0af25a41881bcee99db76 --- /dev/null +++ b/qasper-0454/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Incorporating Sememes into Chinese Definition Modeling + +Question: Do they perform manual evaluation? \ No newline at end of file diff --git a/qasper-0455/instruction.md b/qasper-0455/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..dd2f1a4a86b689926971b044f38aac092a973b82 --- /dev/null +++ b/qasper-0455/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Incorporating Sememes into Chinese Definition Modeling + +Question: Do they compare against Noraset et al. 2017? \ No newline at end of file diff --git a/qasper-0462/instruction.md b/qasper-0462/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..68367e12d9b240b4f1fc8ac11f7991831c19a2b9 --- /dev/null +++ b/qasper-0462/instruction.md @@ -0,0 +1,100 @@ +Name of Paper: Natural Language State Representation for Reinforcement Learning + +Question: How much faster natural language agents converge in performed experiments? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Preliminaries ::: Reinforcement Learning", + "Preliminaries ::: Deep Learning for NLP", + "Semantic Representation Methods", + "Semantic State Representations in the Doom Environment", + "Semantic State Representations in the Doom Environment ::: Experiments", + "Related Work", + "Discussion and Future Work", + "Appendix ::: VizDoom", + "Appendix ::: Natural language State Space", + "Appendix ::: Language model implementation", + "Appendix ::: Model implementation" + ], + "paragraphs": [ + [ + "\u201cThe world of our experiences must be enormously simplified and generalized before it is possible to make a symbolic inventory of all our experiences of things and relations.\"", + "(Edward Sapir, Language: An Introduction to the Study of Speech, 1921)", + "Deep Learning based algorithms use neural networks in order to learn feature representations that are good for solving high dimensional Machine Learning (ML) tasks. Reinforcement Learning (RL) is a subfield of ML that has been greatly affected by the use of deep neural networks as universal function approximators BIBREF0, BIBREF1. These deep neural networks are used in RL to estimate value functions, state-action value functions, policy mappings, next-state predictions, rewards, and more BIBREF2, BIBREF3, BIBREF4, thus combating the \u201ccurse of dimensionality\".", + "The term representation is used differently in different contexts. For the purpose of this paper we define a semantic representation of a state as one that reflects its meaning as it is understood by an expert. The semantic representation of a state should thus be paired with a reliable and computationally efficient method for extracting information from it. Previous success in RL has mainly focused on representing the state in its raw form (e.g., visual input in Atari-based games BIBREF2). This approach stems from the belief that neural networks (specifically convolutional networks) can extract meaningful features from complex inputs. In this work, we challenge current representation techniques and suggest to represent the state using natural language, similar to the way we, as humans, summarize and transfer information efficiently from one to the other BIBREF5.", + "The ability to associate states with natural language sentences that describe them is a hallmark of understanding representations for reinforcement learning. Humans use rich natural language to describe and communicate their visual perceptions, feelings, beliefs, strategies, and more. The semantics inherent to natural language carry knowledge and cues of complex types of content, including: events, spatial relations, temporal relations, semantic roles, logical structures, support for inference and entailment, as well as predicates and arguments BIBREF6. The expressive nature of language can thus act as an alternative semantic state representation.", + "Over the past few years, Natural Language Processing (NLP) has shown an acceleration in progress on a wide range of downstream applications ranging from Question Answering BIBREF7, BIBREF8, to Natural Language Inference BIBREF9, BIBREF10, BIBREF11 through Syntactic Parsing BIBREF12, BIBREF13, BIBREF14. Recent work has shown the ability to learn flexible, hierarchical, contextualized representations, obtaining state-of-the-art results on various natural language processing tasks BIBREF15. A basic observation of our work is that natural language representations are also beneficial for solving problems in which natural language is not the underlying source of input. Moreover, our results indicate that natural language is a strong alternative to current complementary methods for semantic representations of a state.", + "In this work we assume a state can be described using natural language sentences. We use distributional embedding methods in order to represent sentences, processed with a standard Convolutional Neural Network for feature extraction. In Section SECREF2 we describe the basic frameworks we rely on. We discuss possible semantic representations in Section SECREF3, namely, raw visual inputs, semantic segmentation, feature vectors, and natural language representations. Then, in Section SECREF4 we compare NLP representations with their alternatives. Our results suggest that representation of the state using natural language can achieve better performance, even on difficult tasks, or tasks in which the description of the state is saturated with task-nuisances BIBREF17. Moreover, we observe that NLP representations are more robust to transfer and changes in the environment. We conclude the paper with a short discussion and related work." + ], + [ + "In Reinforcement Learning the goal is to learn a policy $\\pi (s)$, which is a mapping from state $s$ to a probability distribution over actions $\\mathcal {A}$, with the objective to maximize a reward $r(s)$ that is provided by the environment. This is often solved by formulating the problem as a Markov Decision Process (MDP) BIBREF19. Two common quantities used to estimate the performance in MDPs are the value $v (s)$ and action-value $Q (s, a)$ functions, which are defined as follows: ${v(s) = \\mathbb {E}^{\\pi } [\\sum _t \\gamma ^t r_t | s_0 = s ]}$ and ${Q(s, a) = \\mathbb {E}^{\\pi } [\\sum _t \\gamma ^t r_t | s_0 = s, a_0 = a ]}$. Two prominent algorithms for solving RL tasks, which we use in this paper, are the value-based DQN BIBREF2 and the policy-based PPO BIBREF3.", + "Deep Q Networks (DQN): The DQN algorithm is an extension of the classical Q-learning approach, to a deep learning regime. Q-learning learns the optimal policy by directly learning the value function, i.e., the action-value function. A neural network is used to estimate the $Q$-values and is trained to minimize the Bellman error, namely", + "Proximal Policy Optimization (PPO): While the DQN learns the optimal behavioral policy using a dynamic programming approach, PPO takes a different route. PPO builds upon the policy gradient theorem, which optimizes the policy directly, with an addition of a trust-region update rule. The policy gradient theorem updates the policy by" + ], + [ + "A word embedding is a mapping from a word $w$ to a vector $\\mathbf {w} \\in \\mathbb {R}^d$. A simple form of word embedding is the Bag of Words (BoW), a vector $\\mathbf {w} \\in \\mathbb {N}^{|D|}$ ($|D|$ is the dictionary size), in which each word receives a unique 1-hot vector representation. Recently, more efficient methods have been proposed, in which the embedding vector is smaller than the dictionary size, $d \\ll |D|$. These methods are also known as distributional embeddings.", + "The distributional hypothesis in linguistics is derived from the semantic theory of language usage (i.e. words that are used and occur in the same contexts tend to have similar meanings). Distributional word representations are a fundamental building block for representing natural language sentences. Word embeddings such as Word2vec BIBREF20 and GloVe BIBREF21 build upon the distributional hypothesis, improving efficiency of state-of-the-art language models.", + "Convolutional Neural Networks (CNNs), originally invented for computer vision, have been shown to achieve strong performance on text classification tasks BIBREF22, BIBREF23, as well as other traditional NLP tasks BIBREF24. In this paper we consider a common architecture BIBREF25, in which each word in a sentence is represented as an embedding vector, a single convolutional layer with $m$ filters is applied, producing an $m$-dimensional vector for each $n$-gram. The vectors are combined using max-pooling followed by a ReLU activation. The result is then passed through multiple hidden linear layers with ReLU activation, eventually generating the final output." + ], + [ + "Contemporary methods for semantic representation of states currently follow one of three approaches: (1) raw visual inputs BIBREF2, BIBREF26, in which raw sensory values of pixels are used from one or multiple sources, (2) feature vectors BIBREF27, BIBREF28, in which general features of the problem are chosen, with no specific structure, and (3) semantic segmentation maps BIBREF29, BIBREF30, in which discrete or logical values are used in one or many channels to represent the general features of the state.", + "The common approach is to derive decisions (e.g., classification, action, etc.) based on information in its raw form. In RL, the raw form is often the pixels representing an image \u2013 however the image is only one form of a semantic representation. In Semantic Segmentation, the image is converted from a 3-channel (RGB) matrix into an $N$-channel matrix, where $N$ is the number of classes. In this case, each channel represents a class, and a binary value at each coordinate denotes whether or not this class is present in the image at this location. For instance, fig: semantic segmentation example considers an autonomous vehicle task. The raw image and segmentation maps are both sufficient for the task (i.e., both contain a sufficient semantic representation). Nevertheless, the semantic segmentation maps contain less task-nuisances BIBREF17, which are random variables that affect the observed data, but are not informative to the task we are trying to solve.", + "In this paper we propose a forth method for representing a state, namely using natural language descriptions. One method to achieve such a representation is through Image Captioning BIBREF31, BIBREF32. Natural language is both rich as well as flexible. This flexibility enables the algorithm designer to represent the information present in the state as efficiently and compactly as possible. As an example, the top image in fig: semantic segmentation example can be represented using natural language as \u201cThere is a car in your lane two meters in front of you, a bicycle rider on your far left in the negative lane, a car in your direction in the opposite lane which is twenty meters away, and trees and pedestrians on the side walk.\u201d or compactly by \u201cThere is a car two meters in front of you a pedestrian on the sidewalk to your right and a car inbound in the negative lane which is far away.\u201d. Language also allows us to efficiently compress information. As an example, the segmentation map in the bottom image of fig: semantic segmentation example can be compactly described by \u201cThere are 13 pedestrians crossing the road in front of you\u201d. In the next section we will demonstrate the benefits of using natural-language-based semantic state representation in a first person shooter enviornment." + ], + [ + "In this section we compare the different types of semantic representations for representing states in the ViZDoom environment BIBREF26, as described in the previous section. More specifically, we use a semantic natural language parser in order to describe a state, over numerous instances of levels varying in difficulty, task-nuisances, and objectives. Our results show that, though semantic segmentation and feature vector representation techniques express a similar statistic of the state, natural language representation offers better performance, faster convergence, more robust solutions, as well as better transfer.", + "The ViZDoom environment involves a 3D world that is significantly more real-world-like than Atari 2600 games, with a relatively realistic physics model. An agent in the ViZDoom environment must effectively perceive, interpret, and learn the 3D world in order to make tactical and strategic decisions of where to go and how to act. There are three types of state representations that are provided by the environment. The first, which is also most commonly used, is raw visual inputs, in which the state is represented by an image from a first person view of the agent. A feature vector representation is an additional state representation provided by the environment. The feature vector representation includes positions as well as labels of all objects and creatures in the vicinity of the agent. Lastly, the environment provides a semantic segmentation map based on the aforementioned feature vector. An example of the visual representations in VizDoom is shown in fig: representations in vizdoom.", + "In order to incorporate natural language representation to the VizDoom environment we've constructed a semantic parser of the semantic segmentation maps provided by the environment. Each state of the environment was converted into a natural language sentence based on positions and labels of objects in the frame. To implement this, the screen was divided into several vertical and horizontal patches, as depicted in fig: patches. These patches describe relational aspects of the state, such as distance of objects and their direction with respect to the agent's point of view. In each patch, objects were counted, and a natural language description of the patch was constructed. This technique was repeated for all patches to form the final state representation. fig: nlp state rep depicts examples of natural language sentences of different states in the enviornment." + ], + [ + "We tested the natural language representation against the visual-based and feature representations on several tasks, with varying difficulty. In these tasks, the agent could navigate, shoot, and collect items such as weapons and medipacks. Often, enemies of different types attacked the agent, and a positive reward was given when an enemy was killed. Occasionally, the agent also suffered from health degeneration. The tasks included a basic scenario, a health gathering scenario, a scenario in which the agent must take cover from fireballs, a scenario in which the agent must defend itself from charging enemies, and a super scenario, where a mixture of the above scenarios was designed to challenge the agent.", + "More specifically, in the basic scenario, a single monster is spawned in front of the agent. The purpose of this scenario is to teach the agent to aim at the enemy and shoot at it. In the health gathering scenario, the floor of the room is covered in toxin, causing the agent to gradually lose health. Medipacks are spawned randomly in the room and the agent's objective is to keep itself alive by collecting them. In the take cover scenario, multiple fireball shooting monsters are spawned in front of the agent. The goal of the agent is to stay alive as long as possible, dodging inbound fireballs. The difficulty of the task increases over time, as additional monsters are spawned. In the defend the center scenario, melee attacking monsters are randomly spawned in the room, and charge towards the agent. As opposed to other scenarios, the agent is incapable of moving, aside from turning left and right and shooting. In the defend the line scenario, both melee and fireball shooting monsters are spawned near the opposing wall. The agent can only step right, left or shoot. Finally, in the \u201csuper\" scenario both melee and fireball shooting monsters are repeatably spawned all over the room. the room contains various items the agent can pick up and use, such as medipacks, shotguns, ammunition and armor. Furthermore, the room is filled with unusable objects, various types of trees, pillars and other decorations. The agent can freely move and turn in any direction, as well as shoot. This scenario combines elements from all of the previous scenarios.", + "Our agent was implemented using a Convolutional Neural Network as described in Section SECREF4. We converted the parsed state into embedded representations of fixed length. We tested both a DQN and a PPO based agent, and compared the natural language representation to the other representation techniques, namely the raw image, feature vector, and semantic segmentation representations.", + "In order to effectively compare the performance of the different representation methods, we conducted our experiments under similar conditions for all agents. The same hyper-parameters were used under all tested representations. Moreover, to rule out effects of architectural expressiveness, the number of weights in all neural networks was approximately matched, regardless of the input type. Finally, we ensured the \u201csuper\" scenario was positively biased toward image-based representations. This was done by adding a large amount items to the game level, thereby filling the state with nuisances (these tests are denoted by `nuisance' in the scenario name). This was especially evident in the NLP representations, as sentences became extensively longer (average of over 250 words). This is contrary to image-based representations, which did not change in dimension.", + "Results of the DQN-based agent are presented in fig: scenario comparison. Each plot depicts the average reward (across 5 seeds) of all representations methods. It can be seen that the NLP representation outperforms the other methods. This is contrary to the fact that it contains the same information as the semantic segmentation maps. More interestingly, comparing the vision-based and feature-based representations render inconsistent conclusions with respect to their relative performance. NLP representations remain robust to changes in the environment as well as task-nuisances in the state. As depicted in fig: nuisance scenarios, inflating the state space with task-nuisances impairs the performance of all representations. There, a large amount of unnecessary objects were spawned in the level, increasing the state's description length to over 250 words, whilst retaining the same amount of useful information. Nevertheless, the NLP representation outperformed the vision and feature based representations, with high robustness to the applied noise.", + "In order to verify the performance of the natural language representation was not due to extensive discretization of patches, we've conducted experiments increasing the number of horizontal patches - ranging from 3 to 31 patches in the extreme case. Our results, as depicted in fig: patch count, indicate that the amount of discretization of patches did not affect the performance of the NLP agent, remaining a superior representation compared to the rest.", + "To conclude, our experiments suggest that NLP representations, though they describe the same raw information of the semantic segmentation maps, are more robust to task-nuisances, allow for better transfer, and achieve higher performance in complex tasks, even when their description is long and convoluted. While we've only presented results for DQN agents, we include plots for a PPO agent in the Appendix, showing similar trends and conclusions. We thus deduce that NLP-based semantic state representations are a preferable choice for training VizDoom agents." + ], + [ + "Work on representation learning is concerned with finding an appropriate representation of data in order to perform a machine learning task BIBREF33. In particular, deep learning exploits this concept by its very nature BIBREF2. Work on representation learning include Predictive State Representations (PSR) BIBREF34, BIBREF35, which capture the state as a vector of predictions of future outcomes, and a Heuristic Embedding of Markov Processes (HEMP) BIBREF36, which learns to embed transition probabilities using an energy-based optimization problem.", + "There has been extensive work attempting to use natural language in RL. Efforts that integrate language in RL develop tools, approaches, and insights that are valuable for improving the generalization and sample efficiency of learning agents. Previous work on language-conditioned RL has considered the use of natural language in the observation and action space. Environments such as Zork and TextWorld BIBREF37 have been the standard benchmarks for testing text-based games. Nevertheless, these environments do not search for semantic state representations, in which an RL algorithm can be better evaluated and controlled.", + "BIBREF38 use high-level semantic abstractions of documents in a representation to facilitate relational learning using Inductive Logic Programming and a generative language model. BIBREF39 use high-level guidance expressed in text to enrich a stochastic agent, playing against the built-in AI of Civilization II. They train an agent with the Monte-Carlo search framework in order to jointly learn to identify text that is relevant to a given game state as well as game strategies based only on environment feedback. BIBREF40 utilize natural language in a model-based approach to describe the dynamics and rewards of an environment, showing these can facilitate transfer between different domains.", + "More recently, the structure and compositionality of natural language has been used for representing policies in hierarchical RL. In a paper by BIBREF41, instructions given in natural language were used in order to break down complex problems into high-level plans and lower-level actions. Their suggested framework leverages the structure inherent to natural language, allowing for transfer to unfamiliar tasks and situations. This use of semantic structure has also been leveraged by BIBREF42, where abstract actions (not necessarily words) were recognized as symbols of a natural and expressive language, improving performance and transfer of RL agents.", + "Outside the context of RL, previous work has also shown that high-quality linguistic representations can assist in cross-modal transfer, such as using semantic relationships between labels for zero-shot transfer in image classification BIBREF43, BIBREF44." + ], + [ + "Our results indicate that natural language can outperform, and sometime even replace, vision-based representations. Nevertheless, natural language representations can also have disadvantages in various scenarios. For one, they require the designer to be able to describe the state exactly, whether by a rule-based or learned parser. Second, they abstract notions of the state space that the designer may not realize are necessary for solving the problem. As such, semantic representations should be carefully chosen, similar to the process of reward shaping or choosing a training algorithm. Here, we enumerate three instances in which we believe natural language representations are beneficial:", + "Natural use-case: Information contained in both generic and task-specific textual corpora may be highly valuable for decision making. This case assumes the state can either be easily described using natural language or is already in a natural language state. This includes examples such as user-based domains, in which user profiles and comments are part of the state, or the stock market, in which stocks are described by analysts and other readily available text. 3D physical environments such as VizDoom also fall into this category, as semantic segmentation maps can be easily described using natural language.", + "Subjective information: Subjectivity refers to aspects used to express opinions, evaluations, and speculations. These may include strategies for a game, the way a doctor feels about her patient, the mood of a driver, and more.", + "Unstructured information: In these cases, features might be measured by different units, with an arbitrary position in the state's feature vector, rendering them sensitive to permutations. Such state representations are thus hard to process using neural networks. As an example, the medical domain may contain numerous features describing the vitals of a patient. These raw features, when observed by an expert, can be efficiently described using natural language. Moreover, they allow an expert to efficiently add subjective information.", + "An orthogonal line of research considers automating the process of image annotation. The noise added from the supervised or unsupervised process serves as a great challenge for natural language representation. We suspect the noise accumulated by this procedure would require additional information to be added to the state (e.g., past information). Nevertheless, as we have shown in this paper, such information can be compressed using natural language. In addition, while we have only considered spatial features of the state, information such as movement directions and transient features can be efficiently encoded as well.", + "Natural language representations help abstract information and interpret the state of an agent, improving its overall performance. Nevertheless, it is imperative to choose a representation that best fits the domain at hand. Designers of RL algorithms should consider searching for a semantic representation that fits their needs. While this work only takes a first step toward finding better semantic state representations, we believe the structure inherent in natural language can be considered a favorable candidate for achieving this goal." + ], + [ + "VizDoom is a \"Doom\" based research environment that was developed at the Pozna\u0144 University of Technology. It is based on \"ZDoom\" game executable, and includes a Python based API. The API offers the user the ability to run game instances, query the game state, and execute actions. The original purpose of VizDoom is to provide a research platform for vision based reinforcement learning. Thus, a natural language representation for the game was needed to be implemented. ViZDoom emulates the \"Doom\" game and enables us to access data within a certain frame using Python dictionaries. This makes it possible to extract valuable data including player health, ammo, enemy locations etc. Each game frame contains \"labels\", which contain data on visible objects in the game (the player, enemies, medkits, etc). We used \"Doom Builder\" in order to edit some of the scenarios and design a new one. Enviroment rewards are presented in doom-scenarios-table." + ], + [ + "A semantic representation using natural language should contain information which can be deduced by a human playing the game. For example, even though a human does not know the exact distance between objects, it can classify them as \"close\" or \"far\". However, objects that are outside the player's field of vision can not be a part of the state. Furthermore, a human would most likely refer to an object's location relative to itself, using directions such as \"right\" or \"left\"." + ], + [ + "To convert each frame to a natural language representation state, the list of available labels is iterated, and a string is built accordingly. The main idea of our implementation is to divide the screen into multiple vertical patches, count the amount of different objects inside by their types, and parse it as a sentence. The decision as to whether an object is close or far can be determined by calculating the distance from it to the player, and using two threshold levels. Object descriptions can be concise or detailed, as needed. We experimented with the following mechanics:", + "the screen can be divided between patches equally, or by determined ratios. Here, our main guideline was to keep the \"front\" patch narrow enough so it can be used as \"sights\".", + "our initial experiment was with 3 patches, and later we added 2 more patches classified as \"outer left\" and \"outer right\". In our experiments we have tested up to 51 patches, referred to as left or right patch with corresponding numbers.", + "we used 2 thresholds, which allowed us to classify the distance of an object from the player as \"close\",\"mid\", and \"far. Depending on the task, the value of the threshold can be changed, as well as adding more thresholds.", + "different states might generate sentence with different size. A maximum sentence length is another parameter that was tested. sentences-length-table presents some data regarding the average word count in some of the game sceanrios.", + "After the sentence describing the state is generated, it is transformed to an embedding vector. Words that were not found in the vocabulary were replaced with an \u201cOOV\" vector. All words were then concatenated to a NxDx1 matrix, representing the state. We experimented with both Word2Vec and GloVe pretrained embedding vectors. Eventually, we used the latter, as it consumes less memory and speeds up the training process. The length of the state sentence is one of the hyperparameters of the agents; shorter sentences are zero padded, where longer ones are trimmed." + ], + [ + "All of our models were implemented using PyTorch. The DQN agents used a single network that outputs the Q-Values of the available actions. The PPO agents used an Actor-Critic model with two networks; the first outputs the policy distribution for the input state, and the second network outputs it's value. As mentioned earlier, we used three common neural network architectures:", + "used for the raw image and semantic segmentation based agents. VizDoom's raw output image resolution is 640X480X3 RGB image. We experimented with both the original image and its down-sampled version. The semantic segmentation image was of resolution 640X480X1, where the pixel value represents the object's class, generated using the VizDoom label API. the network consisted of two convolutional layers, two hidden linear layers and an output layer. The first convolutional layer has 8 6X6 filters with stride 3 and ReLU activation. The second convolutional layer has 16 3X3 filters with stride 2 and ReLU activation. The fully connected layers has 32 and 16 units, both of them are followed by ReLU activation. The output layer's size is the amount of actions the agent has available in the trained scenario.", + "Used in the feature vector based agent. Naturally, some discretization is needed in order to build a feature vector, so some of the state data is lost. the feature vector was made using features we extracted from the VizDoom API, and its dimensions was 90 X 1. The network is made up of two fully connected layers, each of them followed by a ReLU activation. The first layer has 32 units, and the second one one has 16 units. The output layer's size was the amount of actions available to the agent.", + "Used in the natural language based agent. As previously mentioned, each word in the natural language state is transformed into a 200X50X1 matrix. The first layers of the TextCNN are convolutional layers with 8 filter which are designed to scan input sentence, and return convolution outputs of sequences of varying lengths. The filters vary in width, such that each of them learns to identify different lengths of sequences in words. Longer filters have higher capability of extracting features from longer word sequences. The filters we have chosen have the following dimensions: 3X50X1, 4X50X1, 5X50X1, 8X50X1,11X50X1. Following the convolution layer there is a ReLU activation and a max pool layer. Finally, there are two fully connected layers; The first layer has 32 units, and second one has 16 units. Both of them are followed by ReLU activation.", + "All architectures have the same output, regardless of the input type. The DQN network is a regression network, with its output size the number of available actions. The PPO agent has 2 networks; actor and critic. The actor network has a Softmax activation with size equal to the available amount of actions. The critic network is a regression model with a single output representing the state's value. Reward plots for the PPO agent can be found in Figure FIGREF47." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0463/instruction.md b/qasper-0463/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..19fd9a1d7a644e97bb0152abc774f4aa49ea43d2 --- /dev/null +++ b/qasper-0463/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Natural Language State Representation for Reinforcement Learning + +Question: What experiments authors perform? \ No newline at end of file diff --git a/qasper-0465/instruction.md b/qasper-0465/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..958d35f320dcdd69360b16edcfa31dd8024022b8 --- /dev/null +++ b/qasper-0465/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Query-oriented text summarization based on hypergraph transversals + +Question: How does the model compare with the MMR baseline? \ No newline at end of file diff --git a/qasper-0478/instruction.md b/qasper-0478/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..6180cfbd1b16bc74a31a3a344de2d0b722f1e59f --- /dev/null +++ b/qasper-0478/instruction.md @@ -0,0 +1,165 @@ +Name of Paper: Entity-Consistent End-to-end Task-Oriented Dialogue System with KB Retriever + +Question: What were the baseline systems? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Definition", + "Definition ::: Dialogue History", + "Definition ::: Knowledge Base", + "Definition ::: Seq2Seq Dialogue Generation", + "Our Framework", + "Our Framework ::: Encoder", + "Our Framework ::: Vanilla Attention-based Decoder", + "Our Framework ::: Entity-Consistency Augmented Decoder", + "Our Framework ::: Entity-Consistency Augmented Decoder ::: KB Row Selection", + "Our Framework ::: Entity-Consistency Augmented Decoder ::: KB Row Selection ::: Dialogue History Representation:", + "Our Framework ::: Entity-Consistency Augmented Decoder ::: KB Row Selection ::: KB Row Representation:", + "Our Framework ::: Entity-Consistency Augmented Decoder ::: KB Row Selection ::: Memory Network-Based Retriever:", + "Our Framework ::: Entity-Consistency Augmented Decoder ::: KB Column Selection", + "Our Framework ::: Entity-Consistency Augmented Decoder ::: Decoder with Retrieved Entity", + "Training the KB-Retriever", + "Training the KB-Retriever ::: Training with Distant Supervision", + "Training the KB-Retriever ::: Training with Gumbel-Softmax", + "Training the KB-Retriever ::: Experimental Settings", + "Training the KB-Retriever ::: Baseline Models", + "Results", + "Results ::: The proportion of responses that can be supported by a single KB row", + "Results ::: Generation Consistency", + "Results ::: Correlation between the number of KB rows and generation consistency", + "Results ::: Visualization", + "Results ::: Human Evaluation", + "Related Work", + "Conclusion", + "Acknowledgments" + ], + "paragraphs": [ + [ + "Task-oriented dialogue system, which helps users to achieve specific goals with natural language, is attracting more and more research attention. With the success of the sequence-to-sequence (Seq2Seq) models in text generation BIBREF0, BIBREF1, BIBREF2, BIBREF3, BIBREF4, several works tried to model the task-oriented dialogue as the Seq2Seq generation of response from the dialogue history BIBREF5, BIBREF6, BIBREF7. This kind of modeling scheme frees the task-oriented dialogue system from the manually designed pipeline modules and heavy annotation labor for these modules.", + "Different from typical text generation, the successful conversations for task-oriented dialogue system heavily depend on accurate knowledge base (KB) queries. Taking the dialogue in Figure FIGREF1 as an example, to answer the driver's query on the gas station, the dialogue system is required to retrieve the entities like \u201c200 Alester Ave\u201d and \u201cValero\u201d. For the task-oriented system based on Seq2Seq generation, there is a trend in recent study towards modeling the KB query as an attention network over the entire KB entity representations, hoping to learn a model to pay more attention to the relevant entities BIBREF6, BIBREF7, BIBREF8, BIBREF9. Though achieving good end-to-end dialogue generation with over-the-entire-KB attention mechanism, these methods do not guarantee the generation consistency regarding KB entities and sometimes yield responses with conflict entities, like \u201cValero is located at 899 Ames Ct\u201d for the gas station query (as shown in Figure FIGREF1). In fact, the correct address for Valero is 200 Alester Ave. A consistent response is relatively easy to achieve for the conventional pipeline systems because they query the KB by issuing API calls BIBREF10, BIBREF11, BIBREF12, and the returned entities, which typically come from a single KB row, are consistently related to the object (like the \u201cgas station\u201d) that serves the user's request. This indicates that a response can usually be supported by a single KB row. It's promising to incorporate such observation into the Seq2Seq dialogue generation model, since it encourages KB relevant generation and avoids the model from producing responses with conflict entities.", + "To achieve entity-consistent generation in the Seq2Seq task-oriented dialogue system, we propose a novel framework which query the KB in two steps. In the first step, we introduce a retrieval module \u2014 KB-retriever to explicitly query the KB. Inspired by the observation that a single KB row usually supports a response, given the dialogue history and a set of KB rows, the KB-retriever uses a memory network BIBREF13 to select the most relevant row. The retrieval result is then fed into a Seq2Seq dialogue generation model to filter the irrelevant KB entities and improve the consistency within the generated entities. In the second step, we further perform attention mechanism to address the most correlated KB column. Finally, we adopt the copy mechanism to incorporate the retrieved KB entity.", + "Since dialogue dataset is not typically annotated with the retrieval results, training the KB-retriever is non-trivial. To make the training feasible, we propose two methods: 1) we use a set of heuristics to derive the training data and train the retriever in a distant supervised fashion; 2) we use Gumbel-Softmax BIBREF14 as an approximation of the non-differentiable selecting process and train the retriever along with the Seq2Seq dialogue generation model. Experiments on two publicly available datasets (Camrest BIBREF11 and InCar Assistant BIBREF6) confirm the effectiveness of the KB-retriever. Both the retrievers trained with distant-supervision and Gumbel-Softmax technique outperform the compared systems in the automatic and human evaluations. Analysis empirically verifies our assumption that more than 80% responses in the dataset can be supported by a single KB row and better retrieval results lead to better task-oriented dialogue generation performance." + ], + [ + "In this section, we will describe the input and output of the end-to-end task-oriented dialogue system, and the definition of Seq2Seq task-oriented dialogue generation." + ], + [ + "Given a dialogue between a user ($u$) and a system ($s$), we follow eric:2017:SIGDial and represent the $k$-turned dialogue utterances as $\\lbrace (u_{1}, s_{1} ), (u_{2} , s_{2} ), ... , (u_{k}, s_{k})\\rbrace $. At the $i^{\\text{th}}$ turn of the dialogue, we aggregate dialogue context which consists of the tokens of $(u_{1}, s_{1}, ..., s_{i-1}, u_{i})$ and use $\\mathbf {x} = (x_{1}, x_{2}, ..., x_{m})$ to denote the whole dialogue history word by word, where $m$ is the number of tokens in the dialogue history." + ], + [ + "In this paper, we assume to have the access to a relational-database-like KB $B$, which consists of $|\\mathcal {R}|$ rows and $|\\mathcal {C}|$ columns. The value of entity in the $j^{\\text{th}}$ row and the $i^{\\text{th}}$ column is noted as $v_{j, i}$." + ], + [ + "We define the Seq2Seq task-oriented dialogue generation as finding the most likely response $\\mathbf {y}$ according to the input dialogue history $\\mathbf {x}$ and KB $B$. Formally, the probability of a response is defined as", + "where $y_t$ represents an output token." + ], + [ + "In this section, we describe our framework for end-to-end task-oriented dialogues. The architecture of our framework is demonstrated in Figure FIGREF3, which consists of two major components including an memory network-based retriever and the seq2seq dialogue generation with KB Retriever. Our framework first uses the KB-retriever to select the most relevant KB row and further filter the irrelevant entities in a Seq2Seq response generation model to improve the consistency among the output entities. While in decoding, we further perform the attention mechanism to choose the most probable KB column. We will present the details of our framework in the following sections." + ], + [ + "In our encoder, we adopt the bidirectional LSTM BIBREF15 to encode the dialogue history $\\mathbf {x}$, which captures temporal relationships within the sequence. The encoder first map the tokens in $\\mathbf {x}$ to vectors with embedding function $\\phi ^{\\text{emb}}$, and then the BiLSTM read the vector forwardly and backwardly to produce context-sensitive hidden states $(\\mathbf {h}_{1}, \\mathbf {h}_2, ..., \\mathbf {h}_{m})$ by repeatedly applying the recurrence $\\mathbf {h}_{i}=\\text{BiLSTM}\\left( \\phi ^{\\text{emb}}\\left( x_{i}\\right) , \\mathbf {h}_{i-1}\\right)$." + ], + [ + "Here, we follow eric:2017:SIGDial to adopt the attention-based decoder to generation the response word by word. LSTM is also used to represent the partially generated output sequence $(y_{1}, y_2, ...,y_{t-1})$ as $(\\tilde{\\mathbf {h}}_{1}, \\tilde{\\mathbf {h}}_2, ...,\\tilde{\\mathbf {h}}_t)$. For the generation of next token $y_t$, their model first calculates an attentive representation $\\tilde{\\mathbf {h}}^{^{\\prime }}_t$ of the dialogue history as", + "Then, the concatenation of the hidden representation of the partially outputted sequence $\\tilde{\\mathbf {h}}_t$ and the attentive dialogue history representation $\\tilde{\\mathbf {h}}^{^{\\prime }}_t$ are projected to the vocabulary space $\\mathcal {V}$ by $U$ as", + "to calculate the score (logit) for the next token generation. The probability of next token $y_t$ is finally calculated as" + ], + [ + "As shown in section SECREF7, we can see that the generation of tokens are just based on the dialogue history attention, which makes the model ignorant to the KB entities. In this section, we present how to query the KB explicitly in two steps for improving the entity consistence, which first adopt the KB-retriever to select the most relevant KB row and the generation of KB entities from the entities-augmented decoder is constrained to the entities within the most probable row, thus improve the entity generation consistency. Next, we perform the column attention to select the most probable KB column. Finally, we show how to use the copy mechanism to incorporate the retrieved entity while decoding." + ], + [ + "In our framework, our KB-retriever takes the dialogue history and KB rows as inputs and selects the most relevant row. This selection process resembles the task of selecting one word from the inputs to answer questions BIBREF13, and we use a memory network to model this process. In the following sections, we will first describe how to represent the inputs, then we will talk about our memory network-based retriever" + ], + [ + "We encode the dialogue history by adopting the neural bag-of-words (BoW) followed the original paper BIBREF13. Each token in the dialogue history is mapped into a vector by another embedding function $\\phi ^{\\text{emb}^{\\prime }}(x)$ and the dialogue history representation $\\mathbf {q}$ is computed as the sum of these vectors: $\\mathbf {q} = \\sum ^{m}_{i=1} \\phi ^{\\text{emb}^{\\prime }} (x_{i}) $." + ], + [ + "In this section, we describe how to encode the KB row. Each KB cell is represented as the cell value $v$ embedding as $\\mathbf {c}_{j, k} = \\phi ^{\\text{value}}(v_{j, k})$, and the neural BoW is also used to represent a KB row $\\mathbf {r}_{j}$ as $\\mathbf {r}_{j} = \\sum _{k=1}^{|\\mathcal {C}|} \\mathbf {c}_{j,k}$." + ], + [ + "We model the KB retrieval process as selecting the row that most-likely supports the response generation. Memory network BIBREF13 has shown to be effective to model this kind of selection. For a $n$-hop memory network, the model keeps a set of input matrices $\\lbrace R^{1}, R^{2}, ..., R^{n+1}\\rbrace $, where each $R^{i}$ is a stack of $|\\mathcal {R}|$ inputs $(\\mathbf {r}^{i}_1, \\mathbf {r}^{i}_2, ..., \\mathbf {r}^{i}_{|\\mathcal {R}|})$. The model also keeps query $\\mathbf {q}^{1}$ as the input. A single hop memory network computes the probability $\\mathbf {a}_j$ of selecting the $j^{\\text{th}}$ input as", + "For the multi-hop cases, layers of single hop memory network are stacked and the query of the $(i+1)^{\\text{th}}$ layer network is computed as", + "and the output of the last layer is used as the output of the whole network. For more details about memory network, please refer to the original paper BIBREF13.", + "After getting $\\mathbf {a}$, we represent the retrieval results as a 0-1 matrix $T \\in \\lbrace 0, 1\\rbrace ^{|\\mathcal {R}|\\times \\mathcal {|C|}}$, where each element in $T$ is calculated as", + "In the retrieval result, $T_{j, k}$ indicates whether the entity in the $j^{\\text{th}}$ row and the $k^{\\text{th}}$ column is relevant to the final generation of the response. In this paper, we further flatten T to a 0-1 vector $\\mathbf {t} \\in \\lbrace 0, 1\\rbrace ^{|\\mathcal {E}|}$ (where $|\\mathcal {E}|$ equals $|\\mathcal {R}|\\times \\mathcal {|C|}$) as our retrieval row results." + ], + [ + "After getting the retrieved row result that indicates which KB row is the most relevant to the generation, we further perform column attention in decoding time to select the probable KB column. For our KB column selection, following the eric:2017:SIGDial we use the decoder hidden state $(\\tilde{\\mathbf {h}}_{1}, \\tilde{\\mathbf {h}}_2, ...,\\tilde{\\mathbf {h}}_t)$ to compute an attention score with the embedding of column attribute name. The attention score $\\mathbf {c}\\in R^{|\\mathcal {E}|}$ then become the logits of the column be selected, which can be calculated as", + "where $\\mathbf {c}_j$ is the attention score of the $j^{\\text{th}}$ KB column, $\\mathbf {k}_j$ is represented with the embedding of word embedding of KB column name. $W^{^{\\prime }}_{1}$, $W^{^{\\prime }}_{2}$ and $\\mathbf {t}^{T}$ are trainable parameters of the model." + ], + [ + "After the row selection and column selection, we can define the final retrieved KB entity score as the element-wise dot between the row retriever result and the column selection score, which can be calculated as", + "where the $v^{t}$ indicates the final KB retrieved entity score. Finally, we follow eric:2017:SIGDial to use copy mechanism to incorporate the retrieved entity, which can be defined as", + "where $\\mathbf {o}_t$\u2019s dimensionality is $ |\\mathcal {V}|$ +$|\\mathcal {E}|$. In $\\mathbf {v}^t$ , lower $ |\\mathcal {V}|$ is zero and the rest$|\\mathcal {E}|$ is retrieved entity scores." + ], + [ + "As mentioned in section SECREF9, we adopt the memory network to train our KB-retriever. However, in the Seq2Seq dialogue generation, the training data does not include the annotated KB row retrieval results, which makes supervised training the KB-retriever impossible. To tackle this problem, we propose two training methods for our KB-row-retriever. 1) In the first method, inspired by the recent success of distant supervision in information extraction BIBREF16, BIBREF17, BIBREF18, BIBREF19, we take advantage of the similarity between the surface string of KB entries and the reference response, and design a set of heuristics to extract training data for the KB-retriever. 2) In the second method, instead of training the KB-retriever as an independent component, we train it along with the training of the Seq2Seq dialogue generation. To make the retrieval process in Equation DISPLAY_FORM13 differentiable, we use Gumbel-Softmax BIBREF14 as an approximation of the $\\operatornamewithlimits{argmax}$ during training." + ], + [ + "Although it's difficult to obtain the annotated retrieval data for the KB-retriever, we can \u201cguess\u201d the most relevant KB row from the reference response, and then obtain the weakly labeled data for the retriever. Intuitively, for the current utterance in the same dialogue which usually belongs to one topic and the KB row that contains the largest number of entities mentioned in the whole dialogue should support the utterance. In our training with distant supervision, we further simplify our assumption and assume that one dialogue which is usually belongs to one topic and can be supported by the most relevant KB row, which means for a $k$-turned dialogue, we construct $k$ pairs of training instances for the retriever and all the inputs $(u_{1}, s_{1}, ..., s_{i-1}, u_{i} \\mid i \\le k)$ are associated with the same weakly labeled KB retrieval result $T^*$.", + "In this paper, we compute each row's similarity to the whole dialogue and choose the most similar row as $T^*$. We define the similarity of each row as the number of matched spans with the surface form of the entities in the row. Taking the dialogue in Figure FIGREF1 for an example, the similarity of the 4$^\\text{th}$ row equals to 4 with \u201c200 Alester Ave\u201d, \u201cgas station\u201d, \u201cValero\u201d, and \u201croad block nearby\u201d matching the dialogue context; and the similarity of the 7$^\\text{th}$ row equals to 1 with only \u201croad block nearby\u201d matching.", + "In our model with the distantly supervised retriever, the retrieval results serve as the input for the Seq2Seq generation. During training the Seq2Seq generation, we use the weakly labeled retrieval result $T^{*}$ as the input." + ], + [ + "In addition to treating the row retrieval result as an input to the generation model, and training the kb-row-retriever independently, we can train it along with the training of the Seq2Seq dialogue generation in an end-to-end fashion. The major difficulty of such a training scheme is that the discrete retrieval result is not differentiable and the training signal from the generation model cannot be passed to the parameters of the retriever. Gumbel-softmax technique BIBREF14 has been shown an effective approximation to the discrete variable and proved to work in sentence representation. In this paper, we adopt the Gumbel-Softmax technique to train the KB retriever. We use", + "as the approximation of $T$, where $\\mathbf {g}_{j}$ are i.i.d samples drawn from $\\text{Gumbel}(0,1)$ and $\\tau $ is a constant that controls the smoothness of the distribution. $T^{\\text{approx}}_{j}$ replaces $T^{\\text{}}_{j}$ in equation DISPLAY_FORM13 and goes through the same flattening and expanding process as $\\mathbf {V}$ to get $\\mathbf {v}^{\\mathbf {t}^{\\text{approx}^{\\prime }}}$ and the training signal from Seq2Seq generation is passed via the logit", + "To make training with Gumbel-Softmax more stable, we first initialize the parameters by pre-training the KB-retriever with distant supervision and further fine-tuning our framework." + ], + [ + "We choose the InCar Assistant dataset BIBREF6 including three distinct domains: navigation, weather and calendar domain. For weather domain, we follow wen2018sequence to separate the highest temperature, lowest temperature and weather attribute into three different columns. For calendar domain, there are some dialogues without a KB or incomplete KB. In this case, we padding a special token \u201c-\u201d in these incomplete KBs. Our framework is trained separately in these three domains, using the same train/validation/test split sets as eric:2017:SIGDial. To justify the generalization of the proposed model, we also use another public CamRest dataset BIBREF11 and partition the datasets into training, validation and testing set in the ratio 3:1:1. Especially, we hired some human experts to format the CamRest dataset by equipping the corresponding KB to every dialogues.", + "All hyper-parameters are selected according to validation set. We use a three-hop memory network to model our KB-retriever. The dimensionalities of the embedding is selected from $\\lbrace 100, 200\\rbrace $ and LSTM hidden units is selected from $\\lbrace 50, 100, 150, 200, 350\\rbrace $. The dropout we use in our framework is selected from $\\lbrace 0.25, 0.5, 0.75\\rbrace $ and the batch size we adopt is selected from $\\lbrace 1,2\\rbrace $. L2 regularization is used on our model with a tension of $5\\times 10^{-6}$ for reducing overfitting. For training the retriever with distant supervision, we adopt the weight typing trick BIBREF20. We use Adam BIBREF21 to optimize the parameters in our model and adopt the suggested hyper-parameters for optimization.", + "We adopt both the automatic and human evaluations in our experiments." + ], + [ + "We compare our model with several baselines including:", + "Attn seq2seq BIBREF22: A model with simple attention over the input context at each time step during decoding.", + "Ptr-UNK BIBREF23: Ptr-UNK is the model which augments a sequence-to-sequence architecture with attention-based copy mechanism over the encoder context.", + "KV Net BIBREF6: The model adopted and argumented decoder which decodes over the concatenation of vocabulary and KB entities, which allows the model to generate entities.", + "Mem2Seq BIBREF7: Mem2Seq is the model that takes dialogue history and KB entities as input and uses a pointer gate to control either generating a vocabulary word or selecting an input as the output.", + "DSR BIBREF9: DSR leveraged dialogue state representation to retrieve the KB implicitly and applied copying mechanism to retrieve entities from knowledge base while decoding.", + "In InCar dataset, for the Attn seq2seq, Ptr-UNK and Mem2seq, we adopt the reported results from madotto2018mem2seq. In CamRest dataset, for the Mem2Seq, we adopt their open-sourced code to get the results while for the DSR, we run their code on the same dataset to obtain the results." + ], + [ + "Follow the prior works BIBREF6, BIBREF7, BIBREF9, we adopt the BLEU and the Micro Entity F1 to evaluate our model performance. The experimental results are illustrated in Table TABREF30.", + "In the first block of Table TABREF30, we show the Human, rule-based and KV Net (with*) result which are reported from eric:2017:SIGDial. We argue that their results are not directly comparable because their work uses the entities in thier canonicalized forms, which are not calculated based on real entity value. It's noticing that our framework with two methods still outperform KV Net in InCar dataset on whole BLEU and Entity F metrics, which demonstrates the effectiveness of our framework.", + "In the second block of Table TABREF30, we can see that our framework trained with both the distant supervision and the Gumbel-Softmax beats all existing models on two datasets. Our model outperforms each baseline on both BLEU and F1 metrics. In InCar dataset, Our model with Gumbel-Softmax has the highest BLEU compared with baselines, which which shows that our framework can generate more fluent response. Especially, our framework has achieved 2.5% improvement on navigate domain, 1.8% improvement on weather domain and 3.5% improvement on calendar domain on F1 metric. It indicates that the effectiveness of our KB-retriever module and our framework can retrieve more correct entity from KB. In CamRest dataset, the same trend of improvement has been witnessed, which further show the effectiveness of our framework.", + "Besides, we observe that the model trained with Gumbel-Softmax outperforms with distant supervision method. We attribute this to the fact that the KB-retriever and the Seq2Seq module are fine-tuned in an end-to-end fashion, which can refine the KB-retriever and further promote the dialogue generation." + ], + [ + "In this section, we verify our assumption by examining the proportion of responses that can be supported by a single row.", + "We define a response being supported by the most relevant KB row as all the responded entities are included by that row. We study the proportion of these responses over the test set. The number is 95% for the navigation domain, 90% for the CamRest dataset and 80% for the weather domain. This confirms our assumption that most responses can be supported by the relevant KB row. Correctly retrieving the supporting row should be beneficial.", + "We further study the weather domain to see the rest 20% exceptions. Instead of being supported by multiple rows, most of these exceptions cannot be supported by any KB row. For example, there is one case whose reference response is \u201cIt 's not rainy today\u201d, and the related KB entity is sunny. These cases provide challenges beyond the scope of this paper. If we consider this kind of cases as being supported by a single row, such proportion in the weather domain is 99%." + ], + [ + "In this paper, we expect the consistent generation from our model. To verify this, we compute the consistency recall of the utterances that have multiple entities. An utterance is considered as consistent if it has multiple entities and these entities belong to the same row which we annotated with distant supervision.", + "The consistency result is shown in Table TABREF37. From this table, we can see that incorporating retriever in the dialogue generation improves the consistency." + ], + [ + "To further explore the correlation between the number of KB rows and generation consistency, we conduct experiments with distant manner to study the correlation between the number of KB rows and generation consistency.", + "We choose KBs with different number of rows on a scale from 1 to 5 for the generation. From Figure FIGREF32, as the number of KB rows increase, we can see a decrease in generation consistency. This indicates that irrelevant information would harm the dialogue generation consistency." + ], + [ + "To gain more insights into how the our retriever module influences the whole KB score distribution, we visualized the KB entity probability at the decoding position where we generate the entity 200_Alester_Ave. From the example (Fig FIGREF38), we can see the $4^\\text{th}$ row and the $1^\\text{th}$ column has the highest probabilities for generating 200_Alester_Ave, which verify the effectiveness of firstly selecting the most relevant KB row and further selecting the most relevant KB column." + ], + [ + "We provide human evaluation on our framework and the compared models. These responses are based on distinct dialogue history. We hire several human experts and ask them to judge the quality of the responses according to correctness, fluency, and humanlikeness on a scale from 1 to 5. In each judgment, the expert is presented with the dialogue history, an output of a system with the name anonymized, and the gold response.", + "The evaluation results are illustrated in Table TABREF37. Our framework outperforms other baseline models on all metrics according to Table TABREF37. The most significant improvement is from correctness, indicating that our model can retrieve accurate entity from KB and generate more informative information that the users want to know." + ], + [ + "Sequence-to-sequence (Seq2Seq) models in text generation BIBREF0, BIBREF1, BIBREF2, BIBREF3, BIBREF4 has gained more popular and they are applied for the open-domain dialogs BIBREF24, BIBREF25 in the end-to-end training method. Recently, the Seq2Seq can be used for learning task oriented dialogs and how to query the structured KB is the remaining challenges.", + "Properly querying the KB has long been a challenge in the task-oriented dialogue system. In the pipeline system, the KB query is strongly correlated with the design of language understanding, state tracking, and policy management. Typically, after obtaining the dialogue state, the policy management module issues an API call accordingly to query the KB. With the development of neural network in natural language processing, efforts have been made to replacing the discrete and pre-defined dialogue state with the distributed representation BIBREF10, BIBREF11, BIBREF12, BIBREF26. In our framework, our retrieval result can be treated as a numeric representation of the API call return.", + "Instead of interacting with the KB via API calls, more and more recent works tried to incorporate KB query as a part of the model. The most popular way of modeling KB query is treating it as an attention network over the entire KB entities BIBREF6, BIBREF27, BIBREF8, BIBREF28, BIBREF29 and the return can be a fuzzy summation of the entity representations. madotto2018mem2seq's practice of modeling the KB query with memory network can also be considered as learning an attentive preference over these entities. wen2018sequence propose the implicit dialogue state representation to query the KB and achieve the promising performance. Different from their modes, we propose the KB-retriever to explicitly query the KB, and the query result is used to filter the irrelevant entities in the dialogue generation to improve the consistency among the output entities." + ], + [ + "In this paper, we propose a novel framework to improve entities consistency by querying KB in two steps. In the first step, inspired by the observation that a response can usually be supported by a single KB row, we introduce the KB retriever to return the most relevant KB row, which is used to filter the irrelevant KB entities and encourage consistent generation. In the second step, we further perform attention mechanism to select the most relevant KB column. Experimental results show the effectiveness of our method. Extensive analysis further confirms the observation and reveal the correlation between the success of KB query and the success of task-oriented dialogue generation." + ], + [ + "We thank the anonymous reviewers for their helpful comments and suggestions. This work was supported by the National Natural Science Foundation of China (NSFC) via grant 61976072, 61632011 and 61772153." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0491/instruction.md b/qasper-0491/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4cc20aef42ad35228b1b4e97f4c132b179c68721 --- /dev/null +++ b/qasper-0491/instruction.md @@ -0,0 +1,93 @@ +Name of Paper: Civique: Using Social Media to Detect Urban Emergencies + +Question: Are the tweets specific to a region? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Motivation and Challenges", + "Our Approach", + "Pre-Processing Modules", + "Emergency Classification", + "Type Classification", + "Location Visualizer", + "Evaluation", + "Dataset Creation", + "Classifier Evaluation", + "Demostration Description", + "Conclusions" + ], + "paragraphs": [ + [ + "With the surge in the use of social media, micro-blogging sites like Twitter, Facebook, and Foursquare have become household words. Growing ubiquity of mobile phones in highly populated developing nations has spurred an exponential rise in social media usage. The heavy volume of social media posts tagged with users' location information on micro-blogging website Twitter presents a unique opportunity to scan these posts. These Short texts (e.g. \"tweets\") on social media contain information about various events happening around the globe, as people post about events and incidents alike. Conventional web outlets provide emergency phone numbers (i.e. 100, 911), etc., and are fast and accurate. Our system, on the other hand, connects its users through a relatively newer platform i.e. social media, and provides an alternative to these conventional methods. In case of their failure or when such means are busy/occupied, an alternative could prove to be life saving.", + "These real life events are reported on Twitter with different perspectives, opinions, and sentiment. Every day, people discuss events thousands of times across social media sites. We would like to detect such events in case of an emergency. Some previous studies BIBREF0 investigate the use of features such as keywords in the tweet, number of words, and context to devise a classifier for event detection. BIBREF1 discusses various techniques researchers have used previously to detect events from Twitter. BIBREF2 describe a system to automatically detect events about known entities from Twitter. This work is highly specific to detection of events only related to known entities. BIBREF3 discuss a system that returns a ranked list of relevant events given a user query.", + "Several research efforts have focused on identifying events in real time( BIBREF4 BIBREF5 BIBREF6 BIBREF0 ). These include systems to detect emergent topics from Twitter in real time ( BIBREF4 BIBREF7 ), an online clustering technique for identifying tweets in real time BIBREF5 , a system to detect localized events and also track evolution of such events over a period of time BIBREF6 . Our focus is on detecting urban emergencies as events from Twitter messages. We classify events ranging from natural disasters to fire break outs, and accidents. Our system detects whether a tweet, which contains a keyword from a pre-decided list, is related to an actual emergency or not. It also classifies the event into its appropriate category, and visualizes the possible location of the emergency event on the map. We also support notifications to our users, containing the contacts of specifically concerned authorities, as per the category of their tweet.", + "The rest of the paper is as follows: Section SECREF2 provides the motivation for our work, and the challenges in building such a system. Section SECREF3 describes the step by step details of our work, and its results. We evaluate our system and present the results in Section SECREF4 . Section SECREF5 showcases our demonstrations in detail, and Section SECREF6 concludes the paper by briefly describing the overall contribution, implementation and demonstration." + ], + [ + "In 2015, INLINEFORM0 of all unnatural deaths in India were caused by accidents, and INLINEFORM1 by accidental fires. Moreover, the Indian subcontinent suffered seven earthquakes in 2015, with the recent Nepal earthquake alone killing more than 9000 people and injuring INLINEFORM2 . We believe we can harness the current social media activity on the web to minimize losses by quickly connecting affected people and the concerned authorities. Our work is motivated by the following factors, (a) Social media is very accessible in the current scenario. (The \u201cDigital India\u201d initiative by the Government of India promotes internet activity, and thus a pro-active social media.) (b) As per the Internet trends reported in 2014, about 117 million Indians are connected to the Internet through mobile devices. (c) A system such as ours can point out or visualize the affected areas precisely and help inform the authorities in a timely fashion. (d) Such a system can be used on a global scale to reduce the effect of natural calamities and prevent loss of life.", + "There are several challenges in building such an application: (a) Such a system expects a tweet to be location tagged. Otherwise, event detection techniques to extract the spatio-temporal data from the tweet can be vague, and lead to false alarms. (b) Such a system should also be able to verify the user's credibility as pranksters may raise false alarms. (c) Tweets are usually written in a very informal language, which requires a sophisticated language processing component to sanitize the tweet input before event detection. (d) A channel with the concerned authorities should be established for them to take serious action, on alarms raised by such a system. (e) An urban emergency such as a natural disaster could affect communications severely, in case of an earthquake or a cyclone, communications channels like Internet connectivity may get disrupted easily. In such cases, our system may not be of help, as it requires the user to be connected to the internet. We address the above challenges and present our approach in the next section." + ], + [ + "We propose a software architecture for Emergency detection and visualization as shown in figure FIGREF9 . We collect data using Twitter API, and perform language pre-processing before applying a classification model. Tweets are labelled manually with and labels, and later classified manually to provide labels according to the type of emergency they indicate. We use the manually labeled data for training our classifiers.", + "We use traditional classification techniques such as Support Vector Machines(SVM), and Naive Bayes(NB) for training, and perform 10-fold cross validation to obtain f-scores. Later, in real time, our system uses the Twitter streaming APIs to get data, pre-processes it using the same modules, and detects emergencies using the classifiers built above. The tweets related to emergencies are displayed on the web interface along with the location and information for the concerned authorities. The pre-processing of Twitter data obtained is needed as it usually contains ad-hoc abbreviations, phonetic substitutions, URLs, hashtags, and a lot of misspelled words. We use the following language processing modules for such corrections." + ], + [ + "We implement a cleaning module to automate the cleaning of tweets obtained from the Twitter API. We remove URLs, special symbols like @ along with the user mentions, Hashtags and any associated text. We also replace special symbols by blank spaces, and inculcate the module as shown in figure FIGREF9 .", + "An example of such a sample tweet cleaning is shown in table TABREF10 .", + "While tweeting, users often express their emotions by stressing over a few characters in the word. For example, usage of words like hellpppp, fiiiiiireeee, ruuuuunnnnn, druuuuuunnnkkk, soooooooo actually corresponds to help, fire, run, drunk, so etc. We use the compression module implemented by BIBREF8 for converting terms like \u201cpleeeeeeeaaaaaassseeee\u201d to \u201cplease\u201d.", + "It is unlikely for an English word to contain the same character consecutively for three or more times. We, hence, compress all the repeated windows of character length greater than two, to two characters. For example \u201cpleeeeeaaaassee\u201d is converted to \u201cpleeaassee\u201d. Each window now contains two characters of the same alphabet in cases of repetition. Let n be the number of windows, obtained from the previous step. We, then, apply brute force search over INLINEFORM0 possibilities to select a valid dictionary word.", + "Table TABREF13 contains sanitized sample output from our compression module for further processing.", + "Text Normalization is the process of translating ad-hoc abbreviations, typographical errors, phonetic substitution and ungrammatical structures used in text messaging (Tweets and SMS) to plain English. Use of such language (often referred as Chatting Language) induces noise which poses additional processing challenges.", + "We use the normalization module implemented by BIBREF8 for text normalization. Training process requires a Language Model of the target language and a parallel corpora containing aligned un-normalized and normalized word pairs. Our language model consists of 15000 English words taken from various sources on the web.", + "Parallel corpora was collected from the following sources:", + "Stanford Normalization Corpora which consists of 9122 pairs of un-normalized and normalized words / phrases.", + "The above corpora, however, lacked acronyms and short hand texts like 2mrw, l8r, b4, hlp, flor which are frequently used in chatting. We collected 215 pairs un-normalized to normalized word/phrase mappings via crowd-sourcing.", + "Table TABREF16 contains input and normalized output from our module.", + "Users often make spelling mistakes while tweeting. A spell checker makes sure that a valid English word is sent to the classification system. We take this problem into account by introducing a spell checker as a pre-processing module by using the JAVA API of Jazzy spell checker for handling spelling mistakes.", + "An example of correction provided by the Spell Checker module is given below:-", + "Input: building INLINEFORM0 flor, help", + "Output: building INLINEFORM0 floor, help", + "Please note that, our current system performs compression, normalization and spell-checking if the language used is English. The classifier training and detection process are described below." + ], + [ + "The first classifier model acts as a filter for the second stage of classification. We use both SVM and NB to compare the results and choose SVM later for stage one classification model, owing to a better F-score. The training is performed on tweets labeled with classes , and based on unigrams as features. We create word vectors of strings in the tweet using a filter available in the WEKA API BIBREF9 , and perform cross validation using standard classification techniques." + ], + [ + "We employ a multi-class Naive Bayes classifier as the second stage classification mechanism, for categorizing tweets appropriately, depending on the type of emergencies they indicate. This multi-class classifier is trained on data manually labeled with classes. We tokenize the training data using \u201cNgramTokenizer\u201d and then, apply a filter to create word vectors of strings before training. We use \u201ctrigrams\u201d as features to build a model which, later, classifies tweets into appropriate categories, in real time. We then perform cross validation using standard techniques to calculate the results, which are shown under the label \u201cStage 2\u201d, in table TABREF20 ." + ], + [ + "We use Google Maps Geocoding API to display the possible location of the tweet origin based on longitude and latitude. Our visualizer presents the user with a map and pinpoints the location with custom icons for earthquake, cyclone, fire accident etc. Since we currently collect tweets with a location filter for the city of \"Mumbai\", we display its map location on the interface. The possible occurrences of such incidents are displayed on the map as soon as our system is able to detect it.", + "We also display the same on an Android device using the WebView functionality available to developers, thus solving the issue of portability. Our system displays visualization of the various emergencies detected on both web browsers and mobile devices." + ], + [ + "We evaluate our system using automated, and manual evaluation techniques. We perform 10-fold cross validation to obtain the F-scores for our classification systems. We use the following technique for dataset creation. We test the system in realtime environments, and tweet about fires at random locations in our city, using test accounts. Our system was able to detect such tweets and detect them with locations shown on the map." + ], + [ + "We collect data by using the Twitter API for saved data, available for public use. For our experiments we collect 3200 tweets filtered by keywords like \u201cfire\u201d, \u201cearthquake\u201d, \u201ctheft\u201d, \u201crobbery\u201d, \u201cdrunk driving\u201d, \u201cdrunk driving accident\u201d etc. Later, we manually label tweets with and labels for classification as stage one. Our dataset contains 1313 tweet with positive label and 1887 tweets with a negative label . We create another dataset with the positively labeled tweets and provide them with category labels like \u201cfire\u201d, \u201caccident\u201d, \u201cearthquake\u201d etc.", + "" + ], + [ + "The results of 10-fold cross-validation performed for stage one are shown in table TABREF20 , under the label \u201cStage 1\u201d. In table TABREF20 , For \u201cStage 1\u201d of classification, F-score obtained using SVM classifier is INLINEFORM0 as shown in row 2, column 2. We also provide the system with sample tweets in real time and assess its ability to detect the emergency, and classify it accordingly. The classification training for Stage 1 was performed using two traditional classification techniques SVM and NB. SVM outperformed NB by around INLINEFORM1 and became the choice of classification technique for stage one.", + "Some false positives obtained during manual evaluation are, \u201cI am sooooo so drunk right nowwwwwwww\u201d and \u201cfire in my office , the boss is angry\u201d. These occurrences show the need of more labeled gold data for our classifiers, and some other features, like Part-of-Speech tags, Named Entity recognition, Bigrams, Trigrams etc. to perform better.", + "The results of 10-fold cross-validation performed for stage two classfication model are also shown in table TABREF20 , under the label \u201cStage 2\u201d. The training for stage two was also performed using both SVM and NB, but NB outperformed SVM by around INLINEFORM0 to become a choice for stage two classification model.", + "We also perform attribute evaluation for the classification model, and create a word cloud based on the output values, shown in figure FIGREF24 . It shows that our classifier model is trained on appropriate words, which are very close to the emergency situations viz. \u201cfire\u201d, \u201cearthquake\u201d, \u201caccident\u201d, \u201cbreak\u201d (Unigram representation here, but possibly occurs in a bigram phrase with \u201cfire\u201d) etc. In figure FIGREF24 , the word cloud represents the word \u201crespond\u201d as the most frequently occurring word as people need urgent help, and quick response from the assistance teams." + ], + [ + "Users interact with Civique through its Web-based user interface and Android based application interface. The features underlying Civique are demonstrated through the following two show cases:", + "Show case 1: Tweet Detection and Classification", + "This showcase aims at detecting related tweets, and classifying them into appropriate categories. For this, we have created a list of filter words, which are used to filter tweets from the Twitter streaming API. These set of words help us filter the tweets related to any incident. We will tweet, and users are able to see how our system captures such tweets and classifies them. Users should be able to see the tweet emerge as an incident on the web-interface, as shown in figure FIGREF26 and the on the android application, as shown in figure FIGREF27 . Figure FIGREF27 demonstrates how a notification is generated when our system detects an emergency tweet. When a user clicks the emerged spot, the system should be able to display the sanitized version / extracted spatio-temporal data from the tweet. We test the system in a realtime environment, and validate our experiments. We also report the false positives generated during the process in section SECREF25 above.", + "Show case 2: User Notification and Contact Info.", + "Civique includes a set of local contacts for civic authorities who are to be / who can be contacted in case of various emergencies. Users can see how Civique detects an emergency and classifies it. They can also watch how the system generates a notification on the web interface and the Android interface, requesting them to contact the authorities for emergencies. Users can change their preferences on the mobile device anytime and can also opt not to receive notifications. Users should be able to contact the authorities online using the application, but in case the online contact is not responsive, or in case of a sudden loss of connectivity, we provide the user with the offline contact information of the concerned civic authorities along with the notifications." + ], + [ + "Civique is a system which detects urban emergencies like earthquakes, cyclones, fire break out, accidents etc. and visualizes them on both on a browsable web interface and an Android application. We collect data from the popular micro-blogging site Twitter and use language processing modules to sanitize the input. We use this data as input to train a two step classification system, which indicates whether a tweet is related to an emergency or not, and if it is, then what category of emergency it belongs to. We display such positively classified tweets along with their type and location on a Google map, and notify our users to inform the concerned authorities, and possibly evacuate the area, if his location matches the affected area. We believe such a system can help the disaster management machinery, and government bodies like Fire department, Police department, etc., to act swiftly, thus minimizing the loss of life.", + "Twitter users use slang, profanity, misspellings and neologisms. We, use standard cleaning methods, and combine NLP with Machine Learning (ML) to further our cause of tweet classification. At the current stage, we also have an Android application ready for our system, which shows the improvised, mobile-viewable web interface.", + "In the future, we aim to develop detection of emergency categories on the fly, obscure emergencies like \u201cairplane hijacking\u201d should also be detected by our system. We plan to analyze the temporal sequence of the tweet set from a single location to determine whether multiple problems on the same location are the result of a single event, or relate to multiple events." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0496/instruction.md b/qasper-0496/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..da1f6519c3cecc30b75d5fba9bb81a344badbb00 --- /dev/null +++ b/qasper-0496/instruction.md @@ -0,0 +1,67 @@ +Name of Paper: Enriching Existing Conversational Emotion Datasets with Dialogue Acts using Neural Annotators. + +Question: What other relations were found in the datasets? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Annotation of Emotional Dialogue Acts ::: Data for Conversational Emotion Analysis", + "Annotation of Emotional Dialogue Acts ::: Dialogue Act Tagset and SwDA Corpus", + "Annotation of Emotional Dialogue Acts ::: Neural Model Annotators", + "Annotation of Emotional Dialogue Acts ::: Ensemble of Neural Annotators", + "Annotation of Emotional Dialogue Acts ::: Reliability of Neural Annotators", + "EDAs Analysis", + "Conclusion and Future Work", + "Acknowledgements" + ], + "paragraphs": [ + [ + "With the growing demand for human-computer/robot interaction systems, detecting the emotional state of the user can heavily benefit a conversational agent to respond at an appropriate emotional level. Emotion recognition in conversations has proven important for potential applications such as response recommendation or generation, emotion-based text-to-speech, personalisation, etc. Human emotional states can be expressed verbally and non-verbally BIBREF0, BIBREF1, however, while building an interactive dialogue system, the interface needs dialogue acts. A typical dialogue system consists of a language understanding module which requires to determine the meaning of and intention in the human input utterances BIBREF2, BIBREF3. Also, in discourse or conversational analysis, dialogue acts are the main linguistic features to consider BIBREF4. A dialogue act provides an intention and performative function in an utterance of the dialogue. For example, it can infer a user's intention by distinguishing Question, Answer, Request, Agree/Reject, etc. and performative functions such as Acknowledgement, Conversational-opening or -closing, Thanking, etc. The dialogue act information together with emotional states can be very useful for a spoken dialogue system to produce natural interaction BIBREF5.", + "The research in emotion recognition is growing very rapidly and many datasets are available, such as text-based, speech- or vision-level, and multimodal emotion data. Emotion expression recognition is a challenging task and hence multimodality is crucial BIBREF0. However, few conversational multi-modal emotion recognition datasets are available, for example, IEMOCAP BIBREF6, SEMAINE BIBREF7, MELD BIBREF8. They are multi-modal dyadic conversational datasets containing audio-visual and conversational transcripts. Every utterance in these datasets is labeled with an emotion label.", + "In this work, we apply an automated neural ensemble annotation process for dialogue act labeling. Several neural models are trained with the Switchboard Dialogue Act (SwDA) Corpus BIBREF9, BIBREF10 and used for inferring dialogue acts on the emotion datasets. We ensemble five model output labels by checking majority occurrences (most of the model labels are the same) and ranking confidence values of the models. We have annotated two potential multi-modal conversation datasets for emotion recognition: IEMOCAP (Interactive Emotional dyadic MOtion CAPture database) BIBREF6 and MELD (Multimodal EmotionLines Dataset) BIBREF8. Figure FIGREF2, shows an example of dialogue acts with emotion and sentiment labels from the MELD dataset. We confirmed the reliability of annotations with inter-annotator metrics. We analysed the co-occurrences of the dialogue act and emotion labels and discovered a key relationship between them; certain dialogue acts of the utterances show significant and useful association with respective emotional states. For example, Accept/Agree dialogue act often occurs with the Joy emotion while Reject with Anger, Acknowledgements with Surprise, Thanking with Joy, and Apology with Sadness, etc. The detailed analysis of the emotional dialogue acts (EDAs) and annotated datasets are being made available at the SECURE EU Project website." + ], + [ + "There are two emotion taxonomies: (1) discrete emotion categories (DEC) and (2) fined-grained dimensional basis of emotion states (DBE). The DECs are Joy, Sadness, Fear, Surprise, Disgust, Anger and Neutral; identified by Ekman et al. ekman1987universalemos. The DBE of the emotion is usually elicited from two or three dimensions BIBREF1, BIBREF11, BIBREF12. A two-dimensional model is commonly used with Valence and Arousal (also called activation), and in the three-dimensional model, the third dimension is Dominance. IEMOCAP is annotated with all DECs and two additional emotion classes, Frustration and Excited. IEMOCAP is also annotated with three DBE, that includes Valance, Arousal and Dominance BIBREF6. MELD BIBREF8, which is an evolved version of the Emotionlines dataset developed by BIBREF13, is annotated with exactly 7 DECs and sentiments (positive, negative and neutral)." + ], + [ + "There have been many taxonomies for dialogue acts: speech acts BIBREF14 refer to the utterance, not only to present information but to the action at is performed. Speech acts were later modified into five classes (Assertive, Directive, Commissive, Expressive, Declarative) BIBREF15. There are many such standard taxonomies and schemes to annotate conversational data, and most of them follow the discourse compositionality. These schemes have proven their importance for discourse or conversational analysis BIBREF16. During the increased development of dialogue systems and discourse analysis, the standard taxonomy was introduced in recent decades, called Dialogue Act Markup in Several Layers (DAMSL) tag set. According to DAMSL, each DA has a forward-looking function (such as Statement, Info-request, Thanking) and a backwards-looking function (such as Accept, Reject, Answer) BIBREF17.", + "The DAMSL annotation includes not only the utterance-level but also segmented-utterance labelling. However, in the emotion datasets, the utterances are not segmented, as we can see in Figure FIGREF2 first or fourth utterances are not segmented as two separate. The fourth utterance, it could be segmented to have two dialogue act labels, for example, a statement (sd) and a question (qy). That provides very fine-grained DA classes and follows the concept of discourse compositionality. DAMSL distinguishes wh-question (qw), yes-no question (qy), open-ended (qo), and or-question (qr) classes, not just because these questions are syntactically distinct, but also because they have different forward functions BIBREF18. For example, yes-no question is more likely to get a \u201cyes\" answer than a wh-question (qw). This also gives an intuition that the answers follow the syntactic formulation of question, providing a context. For example, qy is used for a question that, from a discourse perspective, expects a Yes (ny) or No (nn) answer.", + "We have investigated the annotation method and trained our neural models with the Switchboard Dialogue Act (SwDA) Corpus BIBREF9, BIBREF10. SwDA Corpus is annotated with the DAMSL tag set and it is been used for reporting and bench-marking state-of-the-art results in dialogue act recognition tasks BIBREF19, BIBREF20, BIBREF21 which makes it ideal for our use case. The Switchboard DAMSL Coders Manual can be followed for knowing more about the dialogue act labels." + ], + [ + "We adopted the neural architectures based on Bothe et al. bothe2018discourse where two variants are: non-context model (classifying at utterance level) and context model (recognizing the dialogue act of the current utterance given a few preceding utterances). From conversational analysis using dialogue acts in Bothe et al. bothe2018interspeech, we learned that the preceding two utterances contribute significantly to recognizing the dialogue act of the current utterance. Hence, we adapt this setting for the context model and create a pool of annotators using recurrent neural networks (RNNs). RNNs can model the contextual information in the sequence of words of an utterance and in the sequence of utterances of a dialogue. Each word in an utterance is represented with a word embedding vector of dimension 1024. We use the word embedding vectors from pre-trained ELMo (Embeddings from Language Models) embeddings BIBREF22. We have a pool of five neural annotators as shown in Figure FIGREF6. Our online tool called Discourse-Wizard is available to practice automated dialogue act labeling. In this tool we use the same neural architectures but model-trained embeddings (while, in this work we use pre-trained ELMo embeddings, as they are better performant but computationally and size-wise expensive to be hosted in the online tool). The annotators are:", + "Utt-level 1 Dialogue Act Neural Annotator (DANA) is an utterance-level classifier that uses word embeddings ($w$) as an input to an RNN layer, attention mechanism and computes the probability of dialogue acts ($da$) using the softmax function (see in Figure FIGREF10, dotted line utt-l1). This model achieved 75.13% accuracy on the SwDA corpus test set.", + "Context 1 DANA is a context model that uses 2 preceding utterances while recognizing the dialogue act of the current utterance (see context model with con1 line in Figure FIGREF10). It uses a hierarchical RNN with the first RNN layer to encode the utterance from word embeddings ($w$) and the second RNN layer is provided with three utterances ($u$) (current and two preceding) composed from the first layer followed by the attention mechanism ($a$), where $\\sum _{n=0}^{n} a_{t-n} = 1$. Finally, the softmax function is used to compute the probability distribution. This model achieved 77.55% accuracy on the SwDA corpus test set.", + "Utt-level 2 DANA is another utterance-level classifier which takes an average of the word embeddings in the input utterance and uses a feedforward neural network hidden layer (see utt-l2 line in Figure FIGREF10, where $mean$ passed to $softmax$ directly). Similar to the previous model, it computes the probability of dialogue acts using the softmax function. This model achieved 72.59% accuracy on the test set of the SwDA corpus.", + "Context 2 DANA is another context model that uses three utterances similar to the Context 1 DANA model, but the utterances are composed as the mean of the word embeddings over each utterance, similar to the Utt-level 2 model ($mean$ passed to context model in Figure FIGREF10 with con2 line). Hence, the Context 2 DANA model is composed of one RNN layer with three input vectors, finally topped with the softmax function for computing the probability distribution of the dialogue acts. This model achieved 75.97% accuracy on the test set of the SwDA corpus.", + "Context 3 DANA is a context model that uses three utterances similar to the previous models, but the utterance representations combine both features from the Context 1 and Context 2 models (con1 and con2 together in Figure FIGREF10). Hence, the Context 3 DANA model combines features of almost all the previous four models to provide the recognition of the dialogue acts. This model achieves 75.91% accuracy on the SwDA corpus test set." + ], + [ + "First preference is given to the labels that are perfectly matching in all the neural annotators. In Table TABREF11, we can see that both datasets have about 40% of exactly matching labels over all models (AM). Then priority is given to the context-based models to check if the label in all context models is matching perfectly. In case two out of three context models are correct, then it is being checked if that label is also produced by at least one of the non-context models. Then, we allow labels to rely on these at least two context models. As a result, about 47% of the labels are taken based on the context models (CM). When we see that none of the context models is producing the same results, then we rank the labels with their respective confidence values produced as a probability distribution using the $softmax$ function. The labels are sorted in descending order according to confidence values. Then we check if the first three (case when one context model and both non-context models produce the same label) or at least two labels are matching, then we allow to pick that one. There are about 3% in IEMOCAP and 5% in MELD (BM).", + "Finally, when none the above conditions are fulfilled, we leave out the label with an unknown category. This unknown category of the dialogue act is labeled with `xx' in the final annotations, and they are about 7% in IEMOCAP and 11% in MELD (NM). The statistics of the EDAs is reported in Table TABREF13 for both datasets. Total utterances in MELD includes training, validation and test datasets." + ], + [ + "The pool of neural annotators provides a fair range of annotations, and we checked the reliability with the following metrics BIBREF23. Krippendorff's Alpha ($\\alpha $) is a reliability coefficient developed to measure the agreement among observers, annotators, and raters, and is often used in emotion annotation BIBREF24. We apply it on the five neural annotators at the nominal level of measurement of dialogue act categories. $\\alpha $ is computed as follows:", + "where $D_{o}$ is the observed disagreement and $D_{e}$ is the disagreement that is expected by chance. $\\alpha =1$ means all annotators produce the same label, while $\\alpha =0$ would mean none agreed on any label. As we can see in Table TABREF20, both datasets IEMOCAP and MELD produce significant inter-neural annotator agreement, 0.553 and 0.494, respectively.", + "A very popular inter-annotator metric is Fleiss' Kappa score, also reported in Table TABREF20, which determines consistency in the ratings. The kappa $k$ can be defined as,", + "where the denominator $1 -\\bar{P}_e$ elicits the degree of agreement that is attainable above chance, and the numerator $\\bar{P} -\\bar{P}_e$ provides the degree of the agreement actually achieved above chance. Hence, $k = 1$ if the raters agree completely, and $k = 0$ when none reach any agreement. We got 0.556 and 0.502 for IEOMOCAP and MELD respectively with our five neural annotators. This indicated that the annotators are labeling the dialogue acts reliably and consistently. We also report the Spearman's correlation between context-based models (Context1 and Context2), and it shows a strong correlation between them (Table TABREF20). While using the labels we checked the absolute match between all context-based models and hence their strong correlation indicates their robustness." + ], + [ + "We can see emotional dialogue act co-occurrences with respect to emotion labels in Figure FIGREF12 for both datasets. There are sets of three bars per dialogue act in the figure, the first and second bar represent emotion labels of IEMOCAP (IE) and MELD (ME), and the third bar is for MELD sentiment (MS) labels. MELD emotion and sentiment statistics are interesting as they are strongly correlated to each other. The bars contain the normalized number of utterances for emotion labels with respect to the total number of utterances for that particular dialogue act category. The statements without-opinion (sd) and with-opinion (sv) contain utterances with almost all emotions. Many neutral utterances are spanning over all the dialogue acts.", + "Quotation (\u2303q) dialogue acts, on the other hand, are mostly used with `Anger' and `Frustration' (in case of IEMOCAP), however, some utterances with `Joy' or `Sadness' as well (see examples in Table TABREF21). Action Directive (ad) dialogue act utterances, which are usually orders, frequently occur with `Anger' or `Frustration' although many with `Happy' emotion in case of the MELD dataset. Acknowledgements (b) are mostly with positive or neutral, however, Appreciation (ba) and Rhetorical (bh) backchannels often occur with a greater number in `Surprise', `Joy' and/or with `Excited' (in case of IEMOCAP). Questions (qh, qw, qy and qy\u2303d) are mostly asked with emotions `Surprise', `Excited', `Frustration' or `Disgust' (in case of MELD) and many are neutral. No-answers (nn) are mostly `Sad' or `Frustrated' as compared to yes-answers (ny). Forward-functions such as Apology (fa) are mostly with `Sadness' whereas Thanking (ft) and Conventional-closing or -opening (fc or fp) are usually with `Joy' or `Excited'.", + "We also noticed that both datasets exhibit a similar relation between dialogue act and emotion. It is important to notice that the dialogue act annotation is based on the given transcripts, however, the emotional expressions are better perceived with audio or video BIBREF6. We report some examples where we mark the utterances with an determined label (xx) in the last row of Table TABREF21. They are skipped from the final annotation because of not fulfilling the conditions explained in Section SECREF14 It is also interesting to see the previous utterance dialogue acts (P-DA) of those skipped utterances, and the sequence of the labels can be followed from Figure FIGREF6 (utt-l1, utt-l2, con1, con2, con3).", + "In the first example, the previous utterance was b, and three DANA models produced labels of the current utterance as b, but it is skipped because the confidence values were not sufficient to bring it as a final label. The second utterance can be challenging even for humans to perceive with any of the dialogue acts. However, the third and fourth utterances are followed by a yes-no question (qy), and hence, we can see in the third example, that context models tried their best to at least perceive it as an answer (ng, ny, nn). The last utterance, \u201cI'm so sorry!\", has been completely disagreed by all the five annotators. Similar apology phrases are mostly found with `Sadness' emotion label's, and the correct dialogue act is Apology (fa). However, they are placed either in the sd or in ba dialogue act category. We believe that with human annotator's help those labels of the utterances can be corrected with very limited efforts." + ], + [ + "In this work, we presented a method to extend conversational multi-modal emotion datasets with dialogue act labels. We successfully show this on two well-established emotion datasets: IEMOCAP and MELD, which we labeled with dialogue acts and made publicly available for further study and research. As a first insight, we found that many of the dialogue acts and emotion labels follow certain relations. These relations can be useful to learn about the emotional behaviours with dialogue acts to build a natural dialogue system and for deeper conversational analysis. The conversational agent might benefit in generating an appropriate response when considering both emotional states and dialogue acts in the utterances.", + "In future work, we foresee the human in the loop for the annotation process along with a pool of automated neural annotators. Robust annotations can be achieved with very little human effort and supervision, for example, observing and correcting the final labels produced by ensemble output labels from the neural annotators. The human-annotator might also help to achieve segmented-utterance labelling of the dialogue acts. We also plan to use these datasets for conversational analysis to infer interactive behaviours of the emotional states with respect to dialogue acts. In our recent work, where we used dialogue acts to build a dialogue system for a social robot, we find this study and dataset very helpful. For example, we can extend our robotic conversational system to consider emotion as an added linguistic feature to produce natural interaction." + ], + [ + "We would like to acknowledge funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska Curie grant agreement No 642667 (SECURE)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0497/instruction.md b/qasper-0497/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..df5a265ef502c7e931dada8f1a6649138d553ad9 --- /dev/null +++ b/qasper-0497/instruction.md @@ -0,0 +1,67 @@ +Name of Paper: Enriching Existing Conversational Emotion Datasets with Dialogue Acts using Neural Annotators. + +Question: How does the ensemble annotator extract the final label? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Annotation of Emotional Dialogue Acts ::: Data for Conversational Emotion Analysis", + "Annotation of Emotional Dialogue Acts ::: Dialogue Act Tagset and SwDA Corpus", + "Annotation of Emotional Dialogue Acts ::: Neural Model Annotators", + "Annotation of Emotional Dialogue Acts ::: Ensemble of Neural Annotators", + "Annotation of Emotional Dialogue Acts ::: Reliability of Neural Annotators", + "EDAs Analysis", + "Conclusion and Future Work", + "Acknowledgements" + ], + "paragraphs": [ + [ + "With the growing demand for human-computer/robot interaction systems, detecting the emotional state of the user can heavily benefit a conversational agent to respond at an appropriate emotional level. Emotion recognition in conversations has proven important for potential applications such as response recommendation or generation, emotion-based text-to-speech, personalisation, etc. Human emotional states can be expressed verbally and non-verbally BIBREF0, BIBREF1, however, while building an interactive dialogue system, the interface needs dialogue acts. A typical dialogue system consists of a language understanding module which requires to determine the meaning of and intention in the human input utterances BIBREF2, BIBREF3. Also, in discourse or conversational analysis, dialogue acts are the main linguistic features to consider BIBREF4. A dialogue act provides an intention and performative function in an utterance of the dialogue. For example, it can infer a user's intention by distinguishing Question, Answer, Request, Agree/Reject, etc. and performative functions such as Acknowledgement, Conversational-opening or -closing, Thanking, etc. The dialogue act information together with emotional states can be very useful for a spoken dialogue system to produce natural interaction BIBREF5.", + "The research in emotion recognition is growing very rapidly and many datasets are available, such as text-based, speech- or vision-level, and multimodal emotion data. Emotion expression recognition is a challenging task and hence multimodality is crucial BIBREF0. However, few conversational multi-modal emotion recognition datasets are available, for example, IEMOCAP BIBREF6, SEMAINE BIBREF7, MELD BIBREF8. They are multi-modal dyadic conversational datasets containing audio-visual and conversational transcripts. Every utterance in these datasets is labeled with an emotion label.", + "In this work, we apply an automated neural ensemble annotation process for dialogue act labeling. Several neural models are trained with the Switchboard Dialogue Act (SwDA) Corpus BIBREF9, BIBREF10 and used for inferring dialogue acts on the emotion datasets. We ensemble five model output labels by checking majority occurrences (most of the model labels are the same) and ranking confidence values of the models. We have annotated two potential multi-modal conversation datasets for emotion recognition: IEMOCAP (Interactive Emotional dyadic MOtion CAPture database) BIBREF6 and MELD (Multimodal EmotionLines Dataset) BIBREF8. Figure FIGREF2, shows an example of dialogue acts with emotion and sentiment labels from the MELD dataset. We confirmed the reliability of annotations with inter-annotator metrics. We analysed the co-occurrences of the dialogue act and emotion labels and discovered a key relationship between them; certain dialogue acts of the utterances show significant and useful association with respective emotional states. For example, Accept/Agree dialogue act often occurs with the Joy emotion while Reject with Anger, Acknowledgements with Surprise, Thanking with Joy, and Apology with Sadness, etc. The detailed analysis of the emotional dialogue acts (EDAs) and annotated datasets are being made available at the SECURE EU Project website." + ], + [ + "There are two emotion taxonomies: (1) discrete emotion categories (DEC) and (2) fined-grained dimensional basis of emotion states (DBE). The DECs are Joy, Sadness, Fear, Surprise, Disgust, Anger and Neutral; identified by Ekman et al. ekman1987universalemos. The DBE of the emotion is usually elicited from two or three dimensions BIBREF1, BIBREF11, BIBREF12. A two-dimensional model is commonly used with Valence and Arousal (also called activation), and in the three-dimensional model, the third dimension is Dominance. IEMOCAP is annotated with all DECs and two additional emotion classes, Frustration and Excited. IEMOCAP is also annotated with three DBE, that includes Valance, Arousal and Dominance BIBREF6. MELD BIBREF8, which is an evolved version of the Emotionlines dataset developed by BIBREF13, is annotated with exactly 7 DECs and sentiments (positive, negative and neutral)." + ], + [ + "There have been many taxonomies for dialogue acts: speech acts BIBREF14 refer to the utterance, not only to present information but to the action at is performed. Speech acts were later modified into five classes (Assertive, Directive, Commissive, Expressive, Declarative) BIBREF15. There are many such standard taxonomies and schemes to annotate conversational data, and most of them follow the discourse compositionality. These schemes have proven their importance for discourse or conversational analysis BIBREF16. During the increased development of dialogue systems and discourse analysis, the standard taxonomy was introduced in recent decades, called Dialogue Act Markup in Several Layers (DAMSL) tag set. According to DAMSL, each DA has a forward-looking function (such as Statement, Info-request, Thanking) and a backwards-looking function (such as Accept, Reject, Answer) BIBREF17.", + "The DAMSL annotation includes not only the utterance-level but also segmented-utterance labelling. However, in the emotion datasets, the utterances are not segmented, as we can see in Figure FIGREF2 first or fourth utterances are not segmented as two separate. The fourth utterance, it could be segmented to have two dialogue act labels, for example, a statement (sd) and a question (qy). That provides very fine-grained DA classes and follows the concept of discourse compositionality. DAMSL distinguishes wh-question (qw), yes-no question (qy), open-ended (qo), and or-question (qr) classes, not just because these questions are syntactically distinct, but also because they have different forward functions BIBREF18. For example, yes-no question is more likely to get a \u201cyes\" answer than a wh-question (qw). This also gives an intuition that the answers follow the syntactic formulation of question, providing a context. For example, qy is used for a question that, from a discourse perspective, expects a Yes (ny) or No (nn) answer.", + "We have investigated the annotation method and trained our neural models with the Switchboard Dialogue Act (SwDA) Corpus BIBREF9, BIBREF10. SwDA Corpus is annotated with the DAMSL tag set and it is been used for reporting and bench-marking state-of-the-art results in dialogue act recognition tasks BIBREF19, BIBREF20, BIBREF21 which makes it ideal for our use case. The Switchboard DAMSL Coders Manual can be followed for knowing more about the dialogue act labels." + ], + [ + "We adopted the neural architectures based on Bothe et al. bothe2018discourse where two variants are: non-context model (classifying at utterance level) and context model (recognizing the dialogue act of the current utterance given a few preceding utterances). From conversational analysis using dialogue acts in Bothe et al. bothe2018interspeech, we learned that the preceding two utterances contribute significantly to recognizing the dialogue act of the current utterance. Hence, we adapt this setting for the context model and create a pool of annotators using recurrent neural networks (RNNs). RNNs can model the contextual information in the sequence of words of an utterance and in the sequence of utterances of a dialogue. Each word in an utterance is represented with a word embedding vector of dimension 1024. We use the word embedding vectors from pre-trained ELMo (Embeddings from Language Models) embeddings BIBREF22. We have a pool of five neural annotators as shown in Figure FIGREF6. Our online tool called Discourse-Wizard is available to practice automated dialogue act labeling. In this tool we use the same neural architectures but model-trained embeddings (while, in this work we use pre-trained ELMo embeddings, as they are better performant but computationally and size-wise expensive to be hosted in the online tool). The annotators are:", + "Utt-level 1 Dialogue Act Neural Annotator (DANA) is an utterance-level classifier that uses word embeddings ($w$) as an input to an RNN layer, attention mechanism and computes the probability of dialogue acts ($da$) using the softmax function (see in Figure FIGREF10, dotted line utt-l1). This model achieved 75.13% accuracy on the SwDA corpus test set.", + "Context 1 DANA is a context model that uses 2 preceding utterances while recognizing the dialogue act of the current utterance (see context model with con1 line in Figure FIGREF10). It uses a hierarchical RNN with the first RNN layer to encode the utterance from word embeddings ($w$) and the second RNN layer is provided with three utterances ($u$) (current and two preceding) composed from the first layer followed by the attention mechanism ($a$), where $\\sum _{n=0}^{n} a_{t-n} = 1$. Finally, the softmax function is used to compute the probability distribution. This model achieved 77.55% accuracy on the SwDA corpus test set.", + "Utt-level 2 DANA is another utterance-level classifier which takes an average of the word embeddings in the input utterance and uses a feedforward neural network hidden layer (see utt-l2 line in Figure FIGREF10, where $mean$ passed to $softmax$ directly). Similar to the previous model, it computes the probability of dialogue acts using the softmax function. This model achieved 72.59% accuracy on the test set of the SwDA corpus.", + "Context 2 DANA is another context model that uses three utterances similar to the Context 1 DANA model, but the utterances are composed as the mean of the word embeddings over each utterance, similar to the Utt-level 2 model ($mean$ passed to context model in Figure FIGREF10 with con2 line). Hence, the Context 2 DANA model is composed of one RNN layer with three input vectors, finally topped with the softmax function for computing the probability distribution of the dialogue acts. This model achieved 75.97% accuracy on the test set of the SwDA corpus.", + "Context 3 DANA is a context model that uses three utterances similar to the previous models, but the utterance representations combine both features from the Context 1 and Context 2 models (con1 and con2 together in Figure FIGREF10). Hence, the Context 3 DANA model combines features of almost all the previous four models to provide the recognition of the dialogue acts. This model achieves 75.91% accuracy on the SwDA corpus test set." + ], + [ + "First preference is given to the labels that are perfectly matching in all the neural annotators. In Table TABREF11, we can see that both datasets have about 40% of exactly matching labels over all models (AM). Then priority is given to the context-based models to check if the label in all context models is matching perfectly. In case two out of three context models are correct, then it is being checked if that label is also produced by at least one of the non-context models. Then, we allow labels to rely on these at least two context models. As a result, about 47% of the labels are taken based on the context models (CM). When we see that none of the context models is producing the same results, then we rank the labels with their respective confidence values produced as a probability distribution using the $softmax$ function. The labels are sorted in descending order according to confidence values. Then we check if the first three (case when one context model and both non-context models produce the same label) or at least two labels are matching, then we allow to pick that one. There are about 3% in IEMOCAP and 5% in MELD (BM).", + "Finally, when none the above conditions are fulfilled, we leave out the label with an unknown category. This unknown category of the dialogue act is labeled with `xx' in the final annotations, and they are about 7% in IEMOCAP and 11% in MELD (NM). The statistics of the EDAs is reported in Table TABREF13 for both datasets. Total utterances in MELD includes training, validation and test datasets." + ], + [ + "The pool of neural annotators provides a fair range of annotations, and we checked the reliability with the following metrics BIBREF23. Krippendorff's Alpha ($\\alpha $) is a reliability coefficient developed to measure the agreement among observers, annotators, and raters, and is often used in emotion annotation BIBREF24. We apply it on the five neural annotators at the nominal level of measurement of dialogue act categories. $\\alpha $ is computed as follows:", + "where $D_{o}$ is the observed disagreement and $D_{e}$ is the disagreement that is expected by chance. $\\alpha =1$ means all annotators produce the same label, while $\\alpha =0$ would mean none agreed on any label. As we can see in Table TABREF20, both datasets IEMOCAP and MELD produce significant inter-neural annotator agreement, 0.553 and 0.494, respectively.", + "A very popular inter-annotator metric is Fleiss' Kappa score, also reported in Table TABREF20, which determines consistency in the ratings. The kappa $k$ can be defined as,", + "where the denominator $1 -\\bar{P}_e$ elicits the degree of agreement that is attainable above chance, and the numerator $\\bar{P} -\\bar{P}_e$ provides the degree of the agreement actually achieved above chance. Hence, $k = 1$ if the raters agree completely, and $k = 0$ when none reach any agreement. We got 0.556 and 0.502 for IEOMOCAP and MELD respectively with our five neural annotators. This indicated that the annotators are labeling the dialogue acts reliably and consistently. We also report the Spearman's correlation between context-based models (Context1 and Context2), and it shows a strong correlation between them (Table TABREF20). While using the labels we checked the absolute match between all context-based models and hence their strong correlation indicates their robustness." + ], + [ + "We can see emotional dialogue act co-occurrences with respect to emotion labels in Figure FIGREF12 for both datasets. There are sets of three bars per dialogue act in the figure, the first and second bar represent emotion labels of IEMOCAP (IE) and MELD (ME), and the third bar is for MELD sentiment (MS) labels. MELD emotion and sentiment statistics are interesting as they are strongly correlated to each other. The bars contain the normalized number of utterances for emotion labels with respect to the total number of utterances for that particular dialogue act category. The statements without-opinion (sd) and with-opinion (sv) contain utterances with almost all emotions. Many neutral utterances are spanning over all the dialogue acts.", + "Quotation (\u2303q) dialogue acts, on the other hand, are mostly used with `Anger' and `Frustration' (in case of IEMOCAP), however, some utterances with `Joy' or `Sadness' as well (see examples in Table TABREF21). Action Directive (ad) dialogue act utterances, which are usually orders, frequently occur with `Anger' or `Frustration' although many with `Happy' emotion in case of the MELD dataset. Acknowledgements (b) are mostly with positive or neutral, however, Appreciation (ba) and Rhetorical (bh) backchannels often occur with a greater number in `Surprise', `Joy' and/or with `Excited' (in case of IEMOCAP). Questions (qh, qw, qy and qy\u2303d) are mostly asked with emotions `Surprise', `Excited', `Frustration' or `Disgust' (in case of MELD) and many are neutral. No-answers (nn) are mostly `Sad' or `Frustrated' as compared to yes-answers (ny). Forward-functions such as Apology (fa) are mostly with `Sadness' whereas Thanking (ft) and Conventional-closing or -opening (fc or fp) are usually with `Joy' or `Excited'.", + "We also noticed that both datasets exhibit a similar relation between dialogue act and emotion. It is important to notice that the dialogue act annotation is based on the given transcripts, however, the emotional expressions are better perceived with audio or video BIBREF6. We report some examples where we mark the utterances with an determined label (xx) in the last row of Table TABREF21. They are skipped from the final annotation because of not fulfilling the conditions explained in Section SECREF14 It is also interesting to see the previous utterance dialogue acts (P-DA) of those skipped utterances, and the sequence of the labels can be followed from Figure FIGREF6 (utt-l1, utt-l2, con1, con2, con3).", + "In the first example, the previous utterance was b, and three DANA models produced labels of the current utterance as b, but it is skipped because the confidence values were not sufficient to bring it as a final label. The second utterance can be challenging even for humans to perceive with any of the dialogue acts. However, the third and fourth utterances are followed by a yes-no question (qy), and hence, we can see in the third example, that context models tried their best to at least perceive it as an answer (ng, ny, nn). The last utterance, \u201cI'm so sorry!\", has been completely disagreed by all the five annotators. Similar apology phrases are mostly found with `Sadness' emotion label's, and the correct dialogue act is Apology (fa). However, they are placed either in the sd or in ba dialogue act category. We believe that with human annotator's help those labels of the utterances can be corrected with very limited efforts." + ], + [ + "In this work, we presented a method to extend conversational multi-modal emotion datasets with dialogue act labels. We successfully show this on two well-established emotion datasets: IEMOCAP and MELD, which we labeled with dialogue acts and made publicly available for further study and research. As a first insight, we found that many of the dialogue acts and emotion labels follow certain relations. These relations can be useful to learn about the emotional behaviours with dialogue acts to build a natural dialogue system and for deeper conversational analysis. The conversational agent might benefit in generating an appropriate response when considering both emotional states and dialogue acts in the utterances.", + "In future work, we foresee the human in the loop for the annotation process along with a pool of automated neural annotators. Robust annotations can be achieved with very little human effort and supervision, for example, observing and correcting the final labels produced by ensemble output labels from the neural annotators. The human-annotator might also help to achieve segmented-utterance labelling of the dialogue acts. We also plan to use these datasets for conversational analysis to infer interactive behaviours of the emotional states with respect to dialogue acts. In our recent work, where we used dialogue acts to build a dialogue system for a social robot, we find this study and dataset very helpful. For example, we can extend our robotic conversational system to consider emotion as an added linguistic feature to produce natural interaction." + ], + [ + "We would like to acknowledge funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska Curie grant agreement No 642667 (SECURE)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0499/instruction.md b/qasper-0499/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..8b19362845c81d3b698f4bc0df9b879932ffb1cd --- /dev/null +++ b/qasper-0499/instruction.md @@ -0,0 +1,67 @@ +Name of Paper: Enriching Existing Conversational Emotion Datasets with Dialogue Acts using Neural Annotators. + +Question: How many models were used? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Annotation of Emotional Dialogue Acts ::: Data for Conversational Emotion Analysis", + "Annotation of Emotional Dialogue Acts ::: Dialogue Act Tagset and SwDA Corpus", + "Annotation of Emotional Dialogue Acts ::: Neural Model Annotators", + "Annotation of Emotional Dialogue Acts ::: Ensemble of Neural Annotators", + "Annotation of Emotional Dialogue Acts ::: Reliability of Neural Annotators", + "EDAs Analysis", + "Conclusion and Future Work", + "Acknowledgements" + ], + "paragraphs": [ + [ + "With the growing demand for human-computer/robot interaction systems, detecting the emotional state of the user can heavily benefit a conversational agent to respond at an appropriate emotional level. Emotion recognition in conversations has proven important for potential applications such as response recommendation or generation, emotion-based text-to-speech, personalisation, etc. Human emotional states can be expressed verbally and non-verbally BIBREF0, BIBREF1, however, while building an interactive dialogue system, the interface needs dialogue acts. A typical dialogue system consists of a language understanding module which requires to determine the meaning of and intention in the human input utterances BIBREF2, BIBREF3. Also, in discourse or conversational analysis, dialogue acts are the main linguistic features to consider BIBREF4. A dialogue act provides an intention and performative function in an utterance of the dialogue. For example, it can infer a user's intention by distinguishing Question, Answer, Request, Agree/Reject, etc. and performative functions such as Acknowledgement, Conversational-opening or -closing, Thanking, etc. The dialogue act information together with emotional states can be very useful for a spoken dialogue system to produce natural interaction BIBREF5.", + "The research in emotion recognition is growing very rapidly and many datasets are available, such as text-based, speech- or vision-level, and multimodal emotion data. Emotion expression recognition is a challenging task and hence multimodality is crucial BIBREF0. However, few conversational multi-modal emotion recognition datasets are available, for example, IEMOCAP BIBREF6, SEMAINE BIBREF7, MELD BIBREF8. They are multi-modal dyadic conversational datasets containing audio-visual and conversational transcripts. Every utterance in these datasets is labeled with an emotion label.", + "In this work, we apply an automated neural ensemble annotation process for dialogue act labeling. Several neural models are trained with the Switchboard Dialogue Act (SwDA) Corpus BIBREF9, BIBREF10 and used for inferring dialogue acts on the emotion datasets. We ensemble five model output labels by checking majority occurrences (most of the model labels are the same) and ranking confidence values of the models. We have annotated two potential multi-modal conversation datasets for emotion recognition: IEMOCAP (Interactive Emotional dyadic MOtion CAPture database) BIBREF6 and MELD (Multimodal EmotionLines Dataset) BIBREF8. Figure FIGREF2, shows an example of dialogue acts with emotion and sentiment labels from the MELD dataset. We confirmed the reliability of annotations with inter-annotator metrics. We analysed the co-occurrences of the dialogue act and emotion labels and discovered a key relationship between them; certain dialogue acts of the utterances show significant and useful association with respective emotional states. For example, Accept/Agree dialogue act often occurs with the Joy emotion while Reject with Anger, Acknowledgements with Surprise, Thanking with Joy, and Apology with Sadness, etc. The detailed analysis of the emotional dialogue acts (EDAs) and annotated datasets are being made available at the SECURE EU Project website." + ], + [ + "There are two emotion taxonomies: (1) discrete emotion categories (DEC) and (2) fined-grained dimensional basis of emotion states (DBE). The DECs are Joy, Sadness, Fear, Surprise, Disgust, Anger and Neutral; identified by Ekman et al. ekman1987universalemos. The DBE of the emotion is usually elicited from two or three dimensions BIBREF1, BIBREF11, BIBREF12. A two-dimensional model is commonly used with Valence and Arousal (also called activation), and in the three-dimensional model, the third dimension is Dominance. IEMOCAP is annotated with all DECs and two additional emotion classes, Frustration and Excited. IEMOCAP is also annotated with three DBE, that includes Valance, Arousal and Dominance BIBREF6. MELD BIBREF8, which is an evolved version of the Emotionlines dataset developed by BIBREF13, is annotated with exactly 7 DECs and sentiments (positive, negative and neutral)." + ], + [ + "There have been many taxonomies for dialogue acts: speech acts BIBREF14 refer to the utterance, not only to present information but to the action at is performed. Speech acts were later modified into five classes (Assertive, Directive, Commissive, Expressive, Declarative) BIBREF15. There are many such standard taxonomies and schemes to annotate conversational data, and most of them follow the discourse compositionality. These schemes have proven their importance for discourse or conversational analysis BIBREF16. During the increased development of dialogue systems and discourse analysis, the standard taxonomy was introduced in recent decades, called Dialogue Act Markup in Several Layers (DAMSL) tag set. According to DAMSL, each DA has a forward-looking function (such as Statement, Info-request, Thanking) and a backwards-looking function (such as Accept, Reject, Answer) BIBREF17.", + "The DAMSL annotation includes not only the utterance-level but also segmented-utterance labelling. However, in the emotion datasets, the utterances are not segmented, as we can see in Figure FIGREF2 first or fourth utterances are not segmented as two separate. The fourth utterance, it could be segmented to have two dialogue act labels, for example, a statement (sd) and a question (qy). That provides very fine-grained DA classes and follows the concept of discourse compositionality. DAMSL distinguishes wh-question (qw), yes-no question (qy), open-ended (qo), and or-question (qr) classes, not just because these questions are syntactically distinct, but also because they have different forward functions BIBREF18. For example, yes-no question is more likely to get a \u201cyes\" answer than a wh-question (qw). This also gives an intuition that the answers follow the syntactic formulation of question, providing a context. For example, qy is used for a question that, from a discourse perspective, expects a Yes (ny) or No (nn) answer.", + "We have investigated the annotation method and trained our neural models with the Switchboard Dialogue Act (SwDA) Corpus BIBREF9, BIBREF10. SwDA Corpus is annotated with the DAMSL tag set and it is been used for reporting and bench-marking state-of-the-art results in dialogue act recognition tasks BIBREF19, BIBREF20, BIBREF21 which makes it ideal for our use case. The Switchboard DAMSL Coders Manual can be followed for knowing more about the dialogue act labels." + ], + [ + "We adopted the neural architectures based on Bothe et al. bothe2018discourse where two variants are: non-context model (classifying at utterance level) and context model (recognizing the dialogue act of the current utterance given a few preceding utterances). From conversational analysis using dialogue acts in Bothe et al. bothe2018interspeech, we learned that the preceding two utterances contribute significantly to recognizing the dialogue act of the current utterance. Hence, we adapt this setting for the context model and create a pool of annotators using recurrent neural networks (RNNs). RNNs can model the contextual information in the sequence of words of an utterance and in the sequence of utterances of a dialogue. Each word in an utterance is represented with a word embedding vector of dimension 1024. We use the word embedding vectors from pre-trained ELMo (Embeddings from Language Models) embeddings BIBREF22. We have a pool of five neural annotators as shown in Figure FIGREF6. Our online tool called Discourse-Wizard is available to practice automated dialogue act labeling. In this tool we use the same neural architectures but model-trained embeddings (while, in this work we use pre-trained ELMo embeddings, as they are better performant but computationally and size-wise expensive to be hosted in the online tool). The annotators are:", + "Utt-level 1 Dialogue Act Neural Annotator (DANA) is an utterance-level classifier that uses word embeddings ($w$) as an input to an RNN layer, attention mechanism and computes the probability of dialogue acts ($da$) using the softmax function (see in Figure FIGREF10, dotted line utt-l1). This model achieved 75.13% accuracy on the SwDA corpus test set.", + "Context 1 DANA is a context model that uses 2 preceding utterances while recognizing the dialogue act of the current utterance (see context model with con1 line in Figure FIGREF10). It uses a hierarchical RNN with the first RNN layer to encode the utterance from word embeddings ($w$) and the second RNN layer is provided with three utterances ($u$) (current and two preceding) composed from the first layer followed by the attention mechanism ($a$), where $\\sum _{n=0}^{n} a_{t-n} = 1$. Finally, the softmax function is used to compute the probability distribution. This model achieved 77.55% accuracy on the SwDA corpus test set.", + "Utt-level 2 DANA is another utterance-level classifier which takes an average of the word embeddings in the input utterance and uses a feedforward neural network hidden layer (see utt-l2 line in Figure FIGREF10, where $mean$ passed to $softmax$ directly). Similar to the previous model, it computes the probability of dialogue acts using the softmax function. This model achieved 72.59% accuracy on the test set of the SwDA corpus.", + "Context 2 DANA is another context model that uses three utterances similar to the Context 1 DANA model, but the utterances are composed as the mean of the word embeddings over each utterance, similar to the Utt-level 2 model ($mean$ passed to context model in Figure FIGREF10 with con2 line). Hence, the Context 2 DANA model is composed of one RNN layer with three input vectors, finally topped with the softmax function for computing the probability distribution of the dialogue acts. This model achieved 75.97% accuracy on the test set of the SwDA corpus.", + "Context 3 DANA is a context model that uses three utterances similar to the previous models, but the utterance representations combine both features from the Context 1 and Context 2 models (con1 and con2 together in Figure FIGREF10). Hence, the Context 3 DANA model combines features of almost all the previous four models to provide the recognition of the dialogue acts. This model achieves 75.91% accuracy on the SwDA corpus test set." + ], + [ + "First preference is given to the labels that are perfectly matching in all the neural annotators. In Table TABREF11, we can see that both datasets have about 40% of exactly matching labels over all models (AM). Then priority is given to the context-based models to check if the label in all context models is matching perfectly. In case two out of three context models are correct, then it is being checked if that label is also produced by at least one of the non-context models. Then, we allow labels to rely on these at least two context models. As a result, about 47% of the labels are taken based on the context models (CM). When we see that none of the context models is producing the same results, then we rank the labels with their respective confidence values produced as a probability distribution using the $softmax$ function. The labels are sorted in descending order according to confidence values. Then we check if the first three (case when one context model and both non-context models produce the same label) or at least two labels are matching, then we allow to pick that one. There are about 3% in IEMOCAP and 5% in MELD (BM).", + "Finally, when none the above conditions are fulfilled, we leave out the label with an unknown category. This unknown category of the dialogue act is labeled with `xx' in the final annotations, and they are about 7% in IEMOCAP and 11% in MELD (NM). The statistics of the EDAs is reported in Table TABREF13 for both datasets. Total utterances in MELD includes training, validation and test datasets." + ], + [ + "The pool of neural annotators provides a fair range of annotations, and we checked the reliability with the following metrics BIBREF23. Krippendorff's Alpha ($\\alpha $) is a reliability coefficient developed to measure the agreement among observers, annotators, and raters, and is often used in emotion annotation BIBREF24. We apply it on the five neural annotators at the nominal level of measurement of dialogue act categories. $\\alpha $ is computed as follows:", + "where $D_{o}$ is the observed disagreement and $D_{e}$ is the disagreement that is expected by chance. $\\alpha =1$ means all annotators produce the same label, while $\\alpha =0$ would mean none agreed on any label. As we can see in Table TABREF20, both datasets IEMOCAP and MELD produce significant inter-neural annotator agreement, 0.553 and 0.494, respectively.", + "A very popular inter-annotator metric is Fleiss' Kappa score, also reported in Table TABREF20, which determines consistency in the ratings. The kappa $k$ can be defined as,", + "where the denominator $1 -\\bar{P}_e$ elicits the degree of agreement that is attainable above chance, and the numerator $\\bar{P} -\\bar{P}_e$ provides the degree of the agreement actually achieved above chance. Hence, $k = 1$ if the raters agree completely, and $k = 0$ when none reach any agreement. We got 0.556 and 0.502 for IEOMOCAP and MELD respectively with our five neural annotators. This indicated that the annotators are labeling the dialogue acts reliably and consistently. We also report the Spearman's correlation between context-based models (Context1 and Context2), and it shows a strong correlation between them (Table TABREF20). While using the labels we checked the absolute match between all context-based models and hence their strong correlation indicates their robustness." + ], + [ + "We can see emotional dialogue act co-occurrences with respect to emotion labels in Figure FIGREF12 for both datasets. There are sets of three bars per dialogue act in the figure, the first and second bar represent emotion labels of IEMOCAP (IE) and MELD (ME), and the third bar is for MELD sentiment (MS) labels. MELD emotion and sentiment statistics are interesting as they are strongly correlated to each other. The bars contain the normalized number of utterances for emotion labels with respect to the total number of utterances for that particular dialogue act category. The statements without-opinion (sd) and with-opinion (sv) contain utterances with almost all emotions. Many neutral utterances are spanning over all the dialogue acts.", + "Quotation (\u2303q) dialogue acts, on the other hand, are mostly used with `Anger' and `Frustration' (in case of IEMOCAP), however, some utterances with `Joy' or `Sadness' as well (see examples in Table TABREF21). Action Directive (ad) dialogue act utterances, which are usually orders, frequently occur with `Anger' or `Frustration' although many with `Happy' emotion in case of the MELD dataset. Acknowledgements (b) are mostly with positive or neutral, however, Appreciation (ba) and Rhetorical (bh) backchannels often occur with a greater number in `Surprise', `Joy' and/or with `Excited' (in case of IEMOCAP). Questions (qh, qw, qy and qy\u2303d) are mostly asked with emotions `Surprise', `Excited', `Frustration' or `Disgust' (in case of MELD) and many are neutral. No-answers (nn) are mostly `Sad' or `Frustrated' as compared to yes-answers (ny). Forward-functions such as Apology (fa) are mostly with `Sadness' whereas Thanking (ft) and Conventional-closing or -opening (fc or fp) are usually with `Joy' or `Excited'.", + "We also noticed that both datasets exhibit a similar relation between dialogue act and emotion. It is important to notice that the dialogue act annotation is based on the given transcripts, however, the emotional expressions are better perceived with audio or video BIBREF6. We report some examples where we mark the utterances with an determined label (xx) in the last row of Table TABREF21. They are skipped from the final annotation because of not fulfilling the conditions explained in Section SECREF14 It is also interesting to see the previous utterance dialogue acts (P-DA) of those skipped utterances, and the sequence of the labels can be followed from Figure FIGREF6 (utt-l1, utt-l2, con1, con2, con3).", + "In the first example, the previous utterance was b, and three DANA models produced labels of the current utterance as b, but it is skipped because the confidence values were not sufficient to bring it as a final label. The second utterance can be challenging even for humans to perceive with any of the dialogue acts. However, the third and fourth utterances are followed by a yes-no question (qy), and hence, we can see in the third example, that context models tried their best to at least perceive it as an answer (ng, ny, nn). The last utterance, \u201cI'm so sorry!\", has been completely disagreed by all the five annotators. Similar apology phrases are mostly found with `Sadness' emotion label's, and the correct dialogue act is Apology (fa). However, they are placed either in the sd or in ba dialogue act category. We believe that with human annotator's help those labels of the utterances can be corrected with very limited efforts." + ], + [ + "In this work, we presented a method to extend conversational multi-modal emotion datasets with dialogue act labels. We successfully show this on two well-established emotion datasets: IEMOCAP and MELD, which we labeled with dialogue acts and made publicly available for further study and research. As a first insight, we found that many of the dialogue acts and emotion labels follow certain relations. These relations can be useful to learn about the emotional behaviours with dialogue acts to build a natural dialogue system and for deeper conversational analysis. The conversational agent might benefit in generating an appropriate response when considering both emotional states and dialogue acts in the utterances.", + "In future work, we foresee the human in the loop for the annotation process along with a pool of automated neural annotators. Robust annotations can be achieved with very little human effort and supervision, for example, observing and correcting the final labels produced by ensemble output labels from the neural annotators. The human-annotator might also help to achieve segmented-utterance labelling of the dialogue acts. We also plan to use these datasets for conversational analysis to infer interactive behaviours of the emotional states with respect to dialogue acts. In our recent work, where we used dialogue acts to build a dialogue system for a social robot, we find this study and dataset very helpful. For example, we can extend our robotic conversational system to consider emotion as an added linguistic feature to produce natural interaction." + ], + [ + "We would like to acknowledge funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska Curie grant agreement No 642667 (SECURE)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0603/instruction.md b/qasper-0603/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..172236fe5047bdd67baddf085283c656a3e61589 --- /dev/null +++ b/qasper-0603/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Emotion helps Sentiment: A Multi-task Model for Sentiment and Emotion Analysis + +Question: What was their result on Stance Sentiment Emotion Corpus? \ No newline at end of file diff --git a/qasper-0604/instruction.md b/qasper-0604/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0bcfab34c8fdb5c5c213679d32f01436bdf31b41 --- /dev/null +++ b/qasper-0604/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Emotion helps Sentiment: A Multi-task Model for Sentiment and Emotion Analysis + +Question: What performance did they obtain on the SemEval dataset? \ No newline at end of file diff --git a/qasper-0621/instruction.md b/qasper-0621/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..a5c1e9d8ab1d85752f543ad09314f70b54242dbe --- /dev/null +++ b/qasper-0621/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Conversational Intent Understanding for Passengers in Autonomous Vehicles + +Question: Did they compare against other systems? \ No newline at end of file diff --git a/qasper-0650/instruction.md b/qasper-0650/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..b4f02d30a7070b4d0f45c0045137c077e6afc91b --- /dev/null +++ b/qasper-0650/instruction.md @@ -0,0 +1,49 @@ +Name of Paper: Imitation Learning of Robot Policies by Combining Language, Vision and Demonstration + +Question: What is task success rate achieved? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Introduction ::: Problem Statement:", + "Background", + "Multimodal Policy Generation via Imitation", + "Results", + "Conclusion and Future Work" + ], + "paragraphs": [ + [ + "A significant challenge when designing robots to operate in the real world lies in the generation of control policies that can adapt to changing environments. Programming such policies is a labor and time-consuming process which requires substantial technical expertise. Imitation learning BIBREF0, is an appealing methodology that aims at overcoming this challenge \u2013 instead of complex programming, the user only provides a set of demonstrations of the intended behavior. These demonstrations are consequently distilled into a robot control policy by learning appropriate parameter settings of the controller. Popular approaches to imitation, such as Dynamic Motor Primitives (DMPs) BIBREF1 or Gaussian Mixture Regression (GMR) BIBREF2 largely focus on motion as the sole input and output modality, i.e., joint angles, forces or positions. Critical semantic and visual information regarding the task, such as the appearance of the target object or the type of task performed, is not taken into account during training and reproduction. The result is often a limited generalization capability which largely revolves around adaptation to changes in the object position. While imitation learning has been successfully applied to a wide range of tasks including table-tennis BIBREF3, locomotion BIBREF4, and human-robot interaction BIBREF5 an important question is how to incorporate language and vision into a differentiable end-to-end system for complex robot control.", + "In this paper, we present an imitation learning approach that combines language, vision, and motion in order to synthesize natural language-conditioned control policies that have strong generalization capabilities while also capturing the semantics of the task. We argue that such a multi-modal teaching approach enables robots to acquire complex policies that generalize to a wide variety of environmental conditions based on descriptions of the intended task. In turn, the network produces control parameters for a lower-level control policy that can be run on a robot to synthesize the corresponding motion. The hierarchical nature of our approach, i.e., a high-level policy generating the parameters of a lower-level policy, allows for generalization of the trained task to a variety of spatial, visual and contextual changes." + ], + [ + "In order to outline our problem statement, we contrast our approach to Imitation learning BIBREF0 which considers the problem of learning a policy $\\mathbf {\\pi }$ from a given set of demonstrations ${\\cal D}=\\lbrace \\mathbf {d}^0,.., \\mathbf {d}^m\\rbrace $. Each demonstration spans a time horizon $T$ and contains information about the robot's states and actions, e.g., demonstrated sensor values and control inputs at each time step. Robot states at each time step within a demonstration are denoted by $\\mathbf {x}_t$. In contrast to other imitation learning approaches, we assume that we have access to the raw camera images of the robot $_t$ at teach time step, as well as access to a verbal description of the task in natural language. This description may provide critical information about the context, goals or objects involved in the task and is denoted as $\\mathbf {s}$. Given this information, our overall objective is to learn a policy $\\mathbf {\\pi }$ which imitates the demonstrated behavior, while also capturing semantics and important visual features. After training, we can provide the policy $\\mathbf {\\pi }(\\mathbf {s},)$ with a different, new state of the robot and a new verbal description (instruction) as parameters. The policy will then generate the control signals needed to perform the task which takes the new visual input and semantic context int o account." + ], + [ + "A fundamental challenge in imitation learning is the extraction of policies that do not only cover the trained scenarios, but also generalize to a wide range of other situations. A large body of literature has addressed the problem of learning robot motor skills by imitation BIBREF6, learning functional BIBREF1 or probabilistic BIBREF7 representations. However, in most of these approaches, the state vector has to be carefully designed in order to ensure that all necessary information for adaptation is available. Neural approaches to imitation learning BIBREF8 circumvent this problem by learning suitable feature representations from rich data sources for each task or for a sequence of tasks BIBREF9, BIBREF10, BIBREF11. Many of these approaches assume that either a sufficiently large set of motion primitives is already available or that a taxonomy of the task is available, i.e., semantics and motions are not trained in conjunction. The importance of maintaining this connection has been shown in BIBREF12, allowing the robot to adapt to untrained variations of the same task. To learn entirely new tasks, meta-learning aims at learning policy parameters that can quickly be fine-tuned to new tasks BIBREF13. While very successful in dealing with visual and spatial information, these approaches do not incorporate any semantic or linguistic component into the learning process. Language has shown to successfully generate task descriptions in BIBREF14 and several works have investigated the idea of combining natural language and imitation learning: BIBREF15, BIBREF16, BIBREF17, BIBREF18, BIBREF19. However, most approaches do not utilize the inherent connection between semantic task descriptions and low-level motions to train a model.", + "Our work is most closely related to the framework introduced in BIBREF20, which also focuses on the symbol grounding problem. More specifically, the work in BIBREF20 aims at mapping perceptual features in the external world to constituents in an expert-provided natural language instruction. Our work approaches the problem of generating dynamic robot policies by fundamentally combining language, vision, and motion control in to a single differentiable neural network that can learn the cross-modal relationships found in the data with minimal human feature engineering. Unlike previous work, our proposed model is capable of directly generating complex low-level control policies from language and vision that allow the robot to reassemble motions shown during training." + ], + [ + "", + "We motivate our approach with a simple example: consider a binning task in which a 6 DOF robot has to drop an object into one of several differently shaped and colored bowls on a table. To teach this task, the human demonstrator does not only provide a kinesthetic demonstration of the desired trajectory, but also a verbal command, e.g., \u201cMove towards the blue bowl\u201d to the robot. In this example, the trajectory generation would have to be conditioned on the blue bowl's position which, however, has to be extracted from visual sensing. Our approach automatically detects and extracts these relationships between vision, language, and motion modalities in order to make best usage of contextual information for better generalization and disambiguation.", + "Figure FIGREF2 (left) provides an overview of our method. Our goal is to train a deep neural network that can take as input a task description $\\mathbf {s}$ and and image $$ and consequently generates robot controls. In the remainder of this paper, we will refer to our network as the mpn. Rather than immediately producing control signals, the mpn will generate the parameters for a lower-level controller. This distinction allows us to build upon well-established control schemes in robotics and optimal control. In our specific case, we use the widely used Dynamic Motor Primitives BIBREF1 as a lower-level controller for control signal generation.", + "In essence, our network can be divided into three parts. The first part, the semantic network, is used to create a task embedding $$ from the input sentence $$ and environment image $$. In a first step, the sentence $$ is tokenized and converted into a sentence matrix ${W} \\in \\mathbb {R}^{l_s \\times l_w} = f_W()$ by utilizing pre-trained Glove word embeddings BIBREF21 where $l_s$ is the padded-fixed-size length of the sentence and $l_w$ is the size of the glove word vectors. To extract the relationships between the words, we use use multiple CNNs $_s = f_L()$ with filter size $n \\times l_w$ for varying $n$, representing different $n$-gram sizes BIBREF22. The final representation is built by flattening the individual $n$-grams with max-pooling of size $(l_s - n_i + 1)\\times l_w$ and concatenating the results before using a single perceptron to detect relationships between different $n$-grams. In order to combine the sentence embedding $_s$ with the image, it is concatenated as a fourth channel to the input image $$. The task embedding $$ is produced with three blocks of convolutional layers, composed of two regular convolutions, followed by a residual convolution BIBREF23 each.", + "In the second part, the policy translation network is used to generate the task parameters $\\Theta \\in \\mathcal {R}^{o \\times b}$ and $\\in \\mathcal {R}^{o}$ given a task embedding $$ where $o$ is the number of output dimensions and $b$ the number of basis functions in the DMP:", + "where $f_G()$ and $f_H()$ are multilayer-perceptrons that use $$ after being processed in a single perceptron with weight $_G$ and bias $_G$. These parameters are then used in the third part of the network, which is a DMP BIBREF0, allowing us leverage a large body of research regarding their behavior and stability, while also allowing other extensions of DMPs BIBREF5, BIBREF24, BIBREF25 to be incorporated to our framework." + ], + [ + "We evaluate our model in a simulated binning task in which the robot is tasked to place a cube into a bowl as outlined by the verbal command. Each environment contains between three and five objects differentiated by their size (small, large), shape (round, square) and color (red, green, blue, yellow, pink), totalling in 20 different objects. Depending on the generated scenario, combinations of these three features are necessary to distinguish the targets from each other, allowing for tasks of varying complexity.", + "To train our model, we generated a dataset of 20,000 demonstrated 7 DOF trajectories (6 robot joints and 1 gripper dimension) in our simulated environment together with a sentence generator capable of creating natural task descriptions for each scenario. In order to create the language generator, we conducted an human-subject study to collect sentence templates of a placement task as well as common words and synonyms for each of the used features. By utilising these data, we are able to generate over 180,000 unique sentences, depending on the generated scenario.", + "The generated parameters of the low-level DMP controller \u2013 the weights and goal position \u2013 must be sufficiently accurate in order to successfully deliver the object to the specified bin. On the right side of Figure FIGREF4, the generated weights for the DMP are shown for two tasks in which the target is close and far away from the robot, located at different sides of the table, indicating the robots ability to generate differently shaped trajectories. The accuracy of the goal position can be seen in Figure FIGREF4(left) which shows another aspect of our approach: By using stochastic forward passes BIBREF26 the model can return an estimate for the validity of a requested task in addition to the predicted goal configuration. The figure shows that the goal position of a red bowl has a relatively small distribution independently of the used sentence or location on the table, where as an invalid target (green) produces a significantly larger distribution, indicating that the requested task may be invalid.", + "To test our model, we generated 500 new scenario testing each of the three features to identify the correct target among other bowls. A task is considered to be successfully completed when the cube is withing the boundaries of the targeted bowl. Bowls have a bounding box of 12.5 and 17.5cm edge length for the small and large variant, respectively. Our experiments showed that using the objects color or shape to uniquely identify an object allows the robot successfully complete the binning task in 97.6% and 96.0% of the cases. However, using the shape alone as a unique identifier, the task could only be completed in 79.0% of the cases. We suspect that the loss of accuracy is due to the low image resolution of the input image, preventing the network from reliably distinguishing the object shapes. In general, our approach is able to actuate the robot with an target error well below 5cm, given the target was correctly identified." + ], + [ + "In this work, we presented an imitation learning approach combining language, vision, and motion. A neural network architecture called Multimodal Policy Network was introduced which is able to learn the cross-modal relationships in the training data and achieve high generalization and disambiguation performance as a result. Our experiments showed that the model is able to generalize towards different locations and sentences while maintaining a high success rate of delivering an object to a desired bowl. In addition, we discussed an extensions of the method that allow us to obtain uncertainty information from the model by utilizing stochastic network outputs to get a distribution over the belief.", + "The modularity of our architecture allows us to easily exchange parts of the network. This can be utilized for transfer learning between different tasks in the semantic network or transfer between different robots by transferring the policy translation network to different robots in simulation, or to bridge the gap between simulation and reality." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0651/instruction.md b/qasper-0651/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..24d739d873f8739b7bcf29c4259d9b99ac5e4bb7 --- /dev/null +++ b/qasper-0651/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Imitation Learning of Robot Policies by Combining Language, Vision and Demonstration + +Question: What simulations are performed by the authors to validate their approach? \ No newline at end of file diff --git a/qasper-0657/instruction.md b/qasper-0657/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..5a4483edb70fc326ea5e13098ee776ae09081ab0 --- /dev/null +++ b/qasper-0657/instruction.md @@ -0,0 +1,96 @@ +Name of Paper: A framework for anomaly detection using language modeling, and its applications to finance + +Question: Does the paper consider the use of perplexity in order to identify text anomalies? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Five views on anomaly", + "Five views on anomaly ::: Anomaly as error", + "Five views on anomaly ::: Anomaly as irregularity", + "Five views on anomaly ::: Anomaly as novelty", + "Five views on anomaly ::: Anomaly as semantic richness", + "Five views on anomaly ::: Anomaly as contextual relevance", + "Anomaly detection via language modeling", + "Anomaly detection via language modeling ::: Anomaly in input vectors", + "Anomaly detection via language modeling ::: Anomaly in output vectors", + "Anomaly detection via language modeling ::: Anomaly in hidden vectors", + "Anomaly detection via language modeling ::: Anomaly in weight tensors and other parameters", + "Challenges and Future Research", + "Conclusion" + ], + "paragraphs": [ + [ + "The detection of anomalous trends in the financial domain has focused largely on fraud detection BIBREF0, risk modeling BIBREF1, and predictive analysis BIBREF2. The data used in the majority of such studies is of time-series, transactional, graph or generally quantitative or structured nature. This belies the critical importance of semi-structured or unstructured text corpora that practitioners in the finance domain derive insights from\u2014corpora such as financial reports, press releases, earnings call transcripts, credit agreements, news articles, customer interaction logs, and social data.", + "Previous research in anomaly detection from text has evolved largely independently from financial applications. Unsupervised clustering methods have been applied to documents in order to identify outliers and emerging topics BIBREF3. Deviation analysis has been applied to text in order to identify errors in spelling BIBREF4 and tagging of documents BIBREF5. Recent popularity of distributional semantics BIBREF6 has led to further advances in semantic deviation analysis BIBREF7. However, current research remains largely divorced from specific applications within the domain of finance.", + "In the following sections, we enumerate major applications of anomaly detection from text in the financial domain, and contextualize them within current research topics in Natural Language Processing." + ], + [ + "Anomaly detection is a strategy that is often employed in contexts where a deviation from a certain norm is sought to be captured, especially when extreme class imbalance impedes the use of a supervised approach. The implementation of such methods allows for the unveiling of previously hidden or obstructed insights.", + "In this section, we lay out five perspectives on how textual anomaly detection can be applied in the context of finance, and how each application opens up opportunities for NLP researchers to apply current research to the financial domain." + ], + [ + "Previous studies have used anomaly detection to identify and correct errors in text BIBREF4, BIBREF5. These are often unintentional errors that occur as a result of some form of data transfer, e.g. from audio to text, from image to text, or from one language to another. Such studies have direct applicability to the error-prone process of earnings call or customer call transcription, where audio quality, accents, and domain-specific terms can lead to errors. Consider a scenario where the CEO of a company states in an audio conference, `Now investments will be made in Asia.' However, the system instead transcribes, `No investments will be made in Asia.' There is a meaningful difference in the implication of the two statements that could greatly influence the analysis and future direction of the company. Additionally, with regards to the second scenario, it is highly unlikely that the CEO would make such a strong and negative statement in a public setting thus supporting the use of anomaly detection for error correction.", + "Optical-character-recognition from images is another error-prone process with large applicability to finance. Many financial reports and presentations are circulated as image documents that need to undergo OCR in order to be machine-readable. OCR might also be applicable to satellite imagery and other forms of image data that might include important textual content such as a graphical representation of financial data. Errors that result from OCR'd documents can often be fixed using systems that have a robust semantic representation of the target domain. For instance, a model that is trained on financial reports might have encoded awareness that emojis are unlikely to appear in them or that it is unusual for the numeric value of profit to be higher than that of revenue." + ], + [ + "Anomaly in the semantic space might reflect irregularities that are intentional or emergent, signaling risky behavior or phenomena. A sudden change in the tone and vocabulary of a company's leadership in their earnings calls or financial reports can signal risk. News stories that have abnormal language, or irregular origination or propagation patterns might be unreliable or untrustworthy.", + "BIBREF8 showed that when trained on similar domains or contexts, distributed representations of words are likely to be stable, where stability is measured as the similarity of their nearest neighbors in the distributed space. Such insight can be used to assess anomalies in this sense. As an example, BIBREF9 identified cliques of users on Twitter who consistently shared news from similar domains. Characterizing these networks as \u201cecho-chambers,\u201d they then represented the content shared by these echo-chambers as distributed representations. When certain topics from one echo-chamber began to deviate from similar topics in other echo-chambers, the content was tagged as unreliable. BIBREF9 showed that this method can be used to improve the performance of standard methods for fake-news detection.", + "In another study BIBREF10, the researchers hypothesized that transparent language in earnings calls indicates high expectations for performance in the upcoming quarters, whereas semantic ambiguity can signal a lack of confidence and expected poor performance. By quantifying transparency as the frequent use of numbers, shorter words, and unsophisticated vocabulary, they showed that a change in transparency is associated with a change in future performance." + ], + [ + "Anomaly can indicate a novel event or phenomenon that may or may not be risky. Breaking news stories often emerge as anomalous trends on social media. BIBREF11 experimented with this in their effort to detect novel events from Twitter conversations. By representing each event as a real-time cluster of tweets (where each tweet was encoded as a vector), they managed to assess the novelty of the event by comparing its centroid to the centroids of older events.", + "Novelty detection can also be used to detect emerging trends on social media, e.g. controversies that engulf various brands often start as small local events that are shared on social media and attract attention over a short period of time. How people respond to these events in early stages of development can be a measure of their veracity or controversiality BIBREF12, BIBREF13.", + "An anomaly in an industry grouping of companies can also be indicative of a company that is disrupting the norm for that industry and the emergence of a new sector or sub-sector. Often known as trail-blazers, these companies innovate faster than their competitors to meet market demands sometimes even before the consumer is aware of their need. As these companies continually evolve their business lines, their core operations are novel outliers from others in the same industry classification that can serve as meaningful signals of transforming industry demands." + ], + [ + "A large portion of text documents that analysts and researchers in the financial sectors consume have a regulatory nature. Annual financial reports, credit agreements, and filings with the U.S. Securities and Exchange Commission (SEC) are some of these types of documents. These documents can be tens or hundreds of pages long, and often include boilerplate language that the readers might need to skip or ignore in order to get to the \u201cmeat\u201d of the content. Often, the abnormal clauses found in these documents are buried in standard text so as not to attract attention to the unique phrases.", + "BIBREF14 used smoothed representations of n-grams in SEC filings in order to identify boilerplate and abnormal language. They did so by comparing the probability of each n-gram against the company's previous filings, against other filings in the same sector, and against other filings from companies with similar market cap. The aim was to assist accounting analysts in skipping boilerplate language and focusing their attention on important snippets in these documents.", + "Similar methods can be applied to credit agreements where covenants and clauses that are too common are often ignored by risk analysts and special attention is paid to clauses that \u201cstand out\u201d from similar agreements." + ], + [ + "Certain types of documents include universal as well as context-specific signals. As an example, consider a given company's financial reports. The reports may include standard financial metrics such as total revenue, net sales, net income, etc. In addition to these universal metrics, businesses often report their performance in terms of the performance of their operating segments. These segments can be business divisions, products, services, or regional operations. The segments are often specific to the company or its peers. For example, Apple Inc.'s segments might include \u201ciPhone,\u201d \u201ciMac,\u201d \u201ciPad,\u201d and \u201cservices.\u201d The same segments will not appear in reports by other businesses.", + "For many analysts and researchers, operating segments are a crucial part of exploratory or predictive analysis. They use performance metrics associated with these segments to compare the business to its competitors, to estimate its market share, and to project the overall performance of the business in upcoming quarters. Automating the identification and normalization of these metrics can facilitate more insightful analytical research. Since these segments are often specific to each business, supervised models that are trained on a diverse set of companies cannot capture them without overfitting to certain companies. Instead, these segments can be treated as company-specific anomalies." + ], + [ + "Unlike numeric data, text data is not directly machine-readable, and requires some form of transformation as a pre-processing step. In \u201cbag-of-words\u201d methods, this transformation can take place by assigning an index number to each word, and representing any block of text as an unordered set of these words. A slightly more sophisticated approach might chain words into continuous \u201cn-grams\u201d and represent a block of text as an ordered series of \u201cn-grams\u201d that have been extracted on a sliding window of size n. These approaches are conventionally known as \u201clanguage modeling.\u201d", + "Since the advent of high-powered processors enabled the widespread use of distributed representations, language modeling has rapidly evolved and adapted to these new capabilities. Recurrent neural networks can capture an arbitrarily long sequence of text and perform various tasks such as classification or text generation BIBREF16. In this new context, language modeling often refers to training a recurrent network that predicts a word in a given sequence of text BIBREF17. Language models are easy to train because even though they follow a predictive mechanism, they do not need any labeled data, and are thus unsupervised.", + "Figure FIGREF6 is a simple illustration of how a neural network that is composed of recurrent units such as Long-Short Term Memory (LSTM) BIBREF18 can perform language modeling. The are four main components to the network:", + "The input vectors ($x_i$), which represent units (i.e. characters, words, phrases, sentences, paragraphs, etc.) in the input text. Occasionally, these are represented by one-hot vectors that assign a unique index to each particular input. More commonly, these vectors are adapted from a pre-trained corpus, where distributed representations have been inferred either by a simpler auto-encoding process BIBREF19 or by applying the same recurrent model to a baseline corpus such as Wikipedia BIBREF17.", + "The output vectors ($y_i$), which represent the model's prediction of the next word in the sequence. Naturally, they are represented in the same dimensionality as $x_i$s.", + "The hidden vectors ($h_i$), which are often randomly initialized and learned through backpropagation. Often trained as dense representations, these vectors tend to display characteristics that indicate semantic richness BIBREF20 and compositionality BIBREF19. While the language model can be used as a text-generation mechanism, the hidden vectors are a strong side product that are sometimes extracted and reused as augmented features in other machine learning systems BIBREF21.", + "The weights of the network ($W_{ij}$) (or other parameters in the network), which are tuned through backpropagation. These often indicate how each vectors in the input or hidden sequence is utilized to generate the output. These parameters play a big role in the way the output of neural networks are reverse-engineered or explained to the end user .", + "The distributions of any of the above-mentioned components can be studied to mine signals for anomalous behavior in the context of irregularity, error, novelty, semantic richness, or contextual relevance." + ], + [ + "As previously mentioned, the input vectors to a text-based neural network are often adapted from publicly-available word vector corpora. In simpler architectures, the network is allowed to back-propagate its errors all the way to the input layer, which might cause the input vectors to be modified. This can serve as a signal for anomaly in the semantic distributions between the original vectors and the modified vectors.", + "Analyzing the stability of word vectors when trained on different iterations can also signal anomalous trends BIBREF8." + ], + [ + "As previously mentioned, language models generate a probability distribution over a word (or character) in a sequence. These probabilities can be used to detect transcription or character-recognition errors in a domain-friendly manner. When the language model is trained on financial data, domain-specific trends (such as the use of commas and parentheses in financial metrics) can be captured and accounted for by the network, minimizing the rate of false positives." + ], + [ + "A recent advancement in text processing is the introduction of fine-tuning methods to neural networks trained on text BIBREF17. Fine-tuning is an approach that facilitates the transfer of semantic knowledge from one domain (source) to another domain (target). The source domain is often large and generic, such as web data or the Wikipedia corpus, while the target domain is often specific (e.g. SEC filings). A network is pre-trained on the source corpus such that its hidden representations are enriched. Next, the pre-trained networks is re-trained on the target domain, but this time only the final (or top few) layers are tuned and the parameters in the remaining layers remain \u201cfrozen.\u201d The top-most layer of the network can be modified to perform a classification, prediction, or generation task in the target domain (see Figure FIGREF15).", + "Fine-tuning aims to change the distribution of hidden representations in such a way that important information about the source domain is preserved, while idiosyncrasies of the target domain are captured in an effective manner BIBREF22. A similar process can be used to determine anomalies in documents. As an example, consider a model that is pre-trained on historical documents from a given sector. If fine-tuning the model on recent documents from the same sector dramatically shifts the representations for certain vectors, this can signal an evolving trend." + ], + [ + "Models that have interpretable parameters can be used to identify areas of deviation or anomalous content. Attention mechanisms BIBREF23 allow the network to account for certain input signals more than others. The learned attention mechanism can provide insight into potential anomalies in the input. Consider a language model that predicts the social media engagement for a given tweet. Such a model can be used to distinguish between engaging and information-rich content versus clickbait, bot-generated, propagandistic, or promotional content by exposing how, for these categories, engagement is associated with attention to certain distributions of \u201ctrigger words.\u201d", + "Table TABREF17 lists four scenarios for using the various layers and parameters of a language model in order to perform anomaly detection from text." + ], + [ + "Like many other domains, in the financial domain, the application of language models as a measurement for semantic regularity of text bears the challenge of dealing with unseen input. Unseen input can be mistaken for anomaly, especially in systems that are designed for error detection. As an example, a system that is trained to correct errors in an earnings call transcript might treat named entities such as the names of a company's executives, or a recent acquisition, as anomalies. This problem is particularly prominent in fine-tuned language models, which are pre-trained on generic corpora that might not include domain-specific terms.", + "When anomalies are of a malicious nature, such as in the case where abnormal clauses are included in credit agreements, the implementation of the anomalous content is adapted to appear normal. Thereby, the task of detecting normal language becomes more difficult.", + "Alternatively, in the case of language used by executives in company presentations such as earnings calls, there may be a lot of noise in the data due to the large degree of variability in the personalities and linguistic patterns of various leaders. The noise variability present in this content could be similar to actual anomalies, hence making it difficult to identify true anomalies.", + "Factors related to market interactions and competitive behavior can also impact the effectiveness of anomaly-detection models. In detecting the emergence of a new industry sector, it may be challenging for a system to detect novelty when a collection of companies, rather than a single company, behave in an anomalous way. The former may be the more common real-world scenario as companies closely monitor and mimic the innovations of their competitors. The exact notion of anomaly can also vary based on the sector and point in time. For example, in the technology sector, the norm in today's world is one of continuous innovation and technological advancements.", + "Additionally, certain types of anomaly can interact and make it difficult for systems to distinguish between them. As an example, a system that is trained to identify the operating segments of a company tends to distinguish between information that is specific to the company, and information that is common across different companies. As a result, it might identify the names of the company's board of directors or its office locations as its operating segments.", + "Traditional machine learning models have previously tackled the above challenges, and solutions are likely to emerge in the neural paradigms as well. Any future research in these directions will have to account for the impact of such solutions on the reliability and explainability of the resulting models and their robustness against adversarial data." + ], + [ + "Anomaly detection from text can have numerous applications in finance, including risk detection, predictive analysis, error correction, and peer detection. We have outlined various perspectives on how anomaly can be interpreted in the context of finance, and corresponding views on how language modeling can be used to detect such aspects of anomalous content. We hope that this paper lays the groundwork for establishing a framework for understanding the opportunities and risks associated with these methods when applied in the financial domain." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0658/instruction.md b/qasper-0658/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0c7014e119765ec3ec0a6997c28224b30fa56786 --- /dev/null +++ b/qasper-0658/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A framework for anomaly detection using language modeling, and its applications to finance + +Question: Does the paper report a baseline for the task? \ No newline at end of file diff --git a/qasper-0659/instruction.md b/qasper-0659/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..09a3a00261cde87d5fd94cc30d30375448ad5987 --- /dev/null +++ b/qasper-0659/instruction.md @@ -0,0 +1,169 @@ +Name of Paper: What Gets Echoed? Understanding the"Pointers"in Explanations of Persuasive Arguments + +Question: What non-contextual properties do they refer to? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Dataset", + "Understanding the Pointers in Explanations", + "Predicting Pointers", + "Predicting Pointers ::: Experiment setup", + "Predicting Pointers ::: Prediction Performance", + "Predicting Pointers ::: The Effect on Generating Explanations", + "Concluding Discussions", + "Acknowledgments", + "Supplemental Material ::: Preprocessing.", + "Supplemental Material ::: PC Echoing OP", + "Supplemental Material ::: Feature Calculation", + "Supplemental Material ::: Word\u2013level Prediction Task", + "Supplemental Material ::: Generating Explanations" + ], + "paragraphs": [ + [ + "Explanations are essential for understanding and learning BIBREF0. They can take many forms, ranging from everyday explanations for questions such as why one likes Star Wars, to sophisticated formalization in the philosophy of science BIBREF1, to simply highlighting features in recent work on interpretable machine learning BIBREF2.", + "Although everyday explanations are mostly encoded in natural language, natural language explanations remain understudied in NLP, partly due to a lack of appropriate datasets and problem formulations. To address these challenges, we leverage /r/ChangeMyView, a community dedicated to sharing counterarguments to controversial views on Reddit, to build a sizable dataset of naturally-occurring explanations. Specifically, in /r/ChangeMyView, an original poster (OP) first delineates the rationales for a (controversial) opinion (e.g., in Table TABREF1, \u201cmost hit music artists today are bad musicians\u201d). Members of /r/ChangeMyView are invited to provide counterarguments. If a counterargument changes the OP's view, the OP awards a $\\Delta $ to indicate the change and is required to explain why the counterargument is persuasive. In this work, we refer to what is being explained, including both the original post and the persuasive comment, as the explanandum.", + "An important advantage of explanations in /r/ChangeMyView is that the explanandum contains most of the required information to provide its explanation. These explanations often select key counterarguments in the persuasive comment and connect them with the original post. As shown in Table TABREF1, the explanation naturally points to, or echoes, part of the explanandum (including both the persuasive comment and the original post) and in this case highlights the argument of \u201cmusic serving different purposes.\u201d", + "These naturally-occurring explanations thus enable us to computationally investigate the selective nature of explanations: \u201cpeople rarely, if ever, expect an explanation that consists of an actual and complete cause of an event. Humans are adept at selecting one or two causes from a sometimes infinite number of causes to be the explanation\u201d BIBREF3. To understand the selective process of providing explanations, we formulate a word-level task to predict whether a word in an explanandum will be echoed in its explanation.", + "Inspired by the observation that words that are likely to be echoed are either frequent or rare, we propose a variety of features to capture how a word is used in the explanandum as well as its non-contextual properties in Section SECREF4. We find that a word's usage in the original post and in the persuasive argument are similarly related to being echoed, except in part-of-speech tags and grammatical relations. For instance, verbs in the original post are less likely to be echoed, while the relationship is reversed in the persuasive argument.", + "We further demonstrate that these features can significantly outperform a random baseline and even a neural model with significantly more knowledge of a word's context. The difficulty of predicting whether content words (i.e., non-stopwords) are echoed is much greater than that of stopwords, among which adjectives are the most difficult and nouns are relatively the easiest. This observation highlights the important role of nouns in explanations. We also find that the relationship between a word's usage in the original post and in the persuasive comment is crucial for predicting the echoing of content words. Our proposed features can also improve the performance of pointer generator networks with coverage in generating explanations BIBREF4.", + "To summarize, our main contributions are:", + "[itemsep=0pt,leftmargin=*,topsep=0pt]", + "We highlight the importance of computationally characterizing human explanations and formulate a concrete problem of predicting how information is selected from explananda to form explanations, including building a novel dataset of naturally-occurring explanations.", + "We provide a computational characterization of natural language explanations and demonstrate the U-shape in which words get echoed.", + "We identify interesting patterns in what gets echoed through a novel word-level classification task, including the importance of nouns in shaping explanations and the importance of contextual properties of both the original post and persuasive comment in predicting the echoing of content words.", + "We show that vanilla LSTMs fail to learn some of the features we develop and that the proposed features can even improve performance in generating explanations with pointer networks.", + "Our code and dataset is available at https://chenhaot.com/papers/explanation-pointers.html." + ], + [ + "To provide background for our study, we first present a brief overview of explanations for the NLP community, and then discuss the connection of our study with pointer networks, linguistic accommodation, and argumentation mining.", + "The most developed discussion of explanations is in the philosophy of science. Extensive studies aim to develop formal models of explanations (e.g., the deductive-nomological model in BIBREF5, see BIBREF1 and BIBREF6 for a review). In this view, explanations are like proofs in logic. On the other hand, psychology and cognitive sciences examine \u201ceveryday explanations\u201d BIBREF0, BIBREF7. These explanations tend to be selective, are typically encoded in natural language, and shape our understanding and learning in life despite the absence of \u201caxioms.\u201d Please refer to BIBREF8 for a detailed comparison of these two modes of explanation.", + "Although explanations have attracted significant interest from the AI community thanks to the growing interest on interpretable machine learning BIBREF9, BIBREF10, BIBREF11, such studies seldom refer to prior work in social sciences BIBREF3. Recent studies also show that explanations such as highlighting important features induce limited improvement on human performance in detecting deceptive reviews and media biases BIBREF12, BIBREF13. Therefore, we believe that developing a computational understanding of everyday explanations is crucial for explainable AI. Here we provide a data-driven study of everyday explanations in the context of persuasion.", + "In particular, we investigate the \u201cpointers\u201d in explanations, inspired by recent work on pointer networks BIBREF14. Copying mechanisms allow a decoder to generate a token by copying from the source, and have been shown to be effective in generation tasks ranging from summarization to program synthesis BIBREF4, BIBREF15, BIBREF16. To the best of our knowledge, our work is the first to investigate the phenomenon of pointers in explanations.", + "Linguistic accommodation and studies on quotations also examine the phenomenon of reusing words BIBREF17, BIBREF18, BIBREF19, BIBREF20. For instance, BIBREF21 show that power differences are reflected in the echoing of function words; BIBREF22 find that news media prefer to quote locally distinct sentences in political debates. In comparison, our word-level formulation presents a fine-grained view of echoing words, and puts a stronger emphasis on content words than work on linguistic accommodation.", + "Finally, our work is concerned with an especially challenging problem in social interaction: persuasion. A battery of studies have done work to enhance our understanding of persuasive arguments BIBREF23, BIBREF24, BIBREF25, BIBREF26, BIBREF27, and the area of argumentation mining specifically investigates the structure of arguments BIBREF28, BIBREF29, BIBREF30. We build on previous work by BIBREF31 and leverage the dynamics of /r/ChangeMyView. Although our findings are certainly related to the persuasion process, we focus on understanding the self-described reasons for persuasion, instead of the structure of arguments or the factors that drive effective persuasion." + ], + [ + "Our dataset is derived from the /r/ChangeMyView subreddit, which has more than 720K subscribers BIBREF31. /r/ChangeMyView hosts conversations where someone expresses a view and others then try to change that person's mind. Despite being fundamentally based on argument, /r/ChangeMyView has a reputation for being remarkably civil and productive BIBREF32, e.g., a journalist wrote \u201cIn a culture of brittle talking points that we guard with our lives, Change My View is a source of motion and surprise\u201d BIBREF33.", + "The delta mechanism in /r/ChangeMyView allows members to acknowledge opinion changes and enables us to identify explanations for opinion changes BIBREF34. Specifically, it requires \u201cAny user, whether they're the OP or not, should reply to a comment that changed their view with a delta symbol and an explanation of the change.\u201d As a result, we have access to tens of thousands of naturally-occurring explanations and associated explananda. In this work, we focus on the opinion changes of the original posters.", + "Throughout this paper, we use the following terminology:", + "[itemsep=-5pt,leftmargin=*,topsep=0pt]", + "An original post (OP) is an initial post where the original poster justifies his or her opinion. We also use OP to refer to the original poster.", + "A persuasive comment (PC) is a comment that directly leads to an opinion change on the part of the OP (i.e., winning a $\\Delta $).", + "A top-level comment is a comment that directly replies to an OP, and /r/ChangeMyView requires the top-level comment to \u201cchallenge at least one aspect of OP\u2019s stated view (however minor), unless they are asking a clarifying question.\u201d", + "An explanation is a comment where an OP acknowledges a change in his or her view and provides an explanation of the change. As shown in Table TABREF1, the explanation not only provides a rationale, it can also include other discourse acts, such as expressing gratitude.", + "Using https://pushshift.io, we collect the posts and comments in /r/ChangeMyView from January 17th, 2013 to January 31st, 2019, and extract tuples of (OP, PC, explanation). We use the tuples from the final six months of our dataset as the test set, those from the six months before that as the validation set, and the remaining tuples as the training set. The sets contain 5,270, 5,831, and 26,617 tuples respectively. Note that there is no overlap in time between the three sets and the test set can therefore be used to assess generalization including potential changes in community norms and world events.", + "Preprocessing. We perform a number of preprocessing steps, such as converting blockquotes in Markdown to quotes, filtering explicit edits made by authors, mapping all URLs to a special @url@ token, and replacing hyperlinks with the link text. We ignore all triples that contain any deleted comments or posts. We use spaCy for tokenization and tagging BIBREF35. We also use the NLTK implementation of the Porter stemming algorithm to store the stemmed version of each word, for later use in our prediction task BIBREF36, BIBREF37. Refer to the supplementary material for more information on preprocessing.", + "Data statistics. Table TABREF16 provides basic statistics of the training tuples and how they compare to other comments. We highlight the fact that PCs are on average longer than top-level comments, suggesting that PCs contain substantial counterarguments that directly contribute to opinion change. Therefore, we simplify the problem by focusing on the (OP, PC, explanation) tuples and ignore any other exchanges between an OP and a commenter.", + "Below, we highlight some notable features of explanations as they appear in our dataset.", + "The length of explanations shows stronger correlation with that of OPs and PCs than between OPs and PCs (Figure FIGREF8). This observation indicates that explanations are somehow better related with OPs and PCs than PCs are with OPs in terms of language use. A possible reason is that the explainer combines their natural tendency towards length with accommodating the PC.", + "Explanations have a greater fraction of \u201cpointers\u201d than do persuasive comments (Figure FIGREF8). We measure the likelihood of a word in an explanation being copied from either its OP or PC and provide a similar probability for a PC for copying from its OP. As we discussed in Section SECREF1, the words in an explanation are much more likely to come from the existing discussion than are the words in a PC (59.8% vs 39.0%). This phenomenon holds even if we restrict ourselves to considering words outside quotations, which removes the effect of quoting other parts of the discussion, and if we focus only on content words, which removes the effect of \u201creusing\u201d stopwords.", + "Relation between a word being echoed and its document frequency (Figure FIGREF8). Finally, as a preview of our main results, the document frequency of a word from the explanandum is related to the probability of being echoed in the explanation. Although the average likelihood declines as the document frequency gets lower, we observe an intriguing U-shape in the scatter plot. In other words, the words that are most likely to be echoed are either unusually frequent or unusually rare, while most words in the middle show a moderate likelihood of being echoed." + ], + [ + "To further investigate how explanations select words from the explanandum, we formulate a word-level prediction task to predict whether words in an OP or PC are echoed in its explanation. Formally, given a tuple of (OP, PC, explanation), we extract the unique stemmed words as $\\mathcal {V}_{\\text{OP}}, \\mathcal {V}_{\\text{PC}}, \\mathcal {V}_{\\text{EXP}}$. We then define the label for each word in the OP or PC, $w \\in \\mathcal {V}_{\\text{OP}} \\cup \\mathcal {V}_{\\text{PC}}$, based on the explanation as follows:", + "Our prediction task is thus a straightforward binary classification task at the word level. We develop the following five groups of features to capture properties of how a word is used in the explanandum (see Table TABREF18 for the full list):", + "[itemsep=0pt,leftmargin=*,topsep=0pt]", + "Non-contextual properties of a word. These features are derived directly from the word and capture the general tendency of a word being echoed in explanations.", + "Word usage in an OP or PC (two groups). These features capture how a word is used in an OP or PC. As a result, for each feature, we have two values for the OP and PC respectively.", + "How a word connects an OP and PC. These features look at the difference between word usage in the OP and PC. We expect this group to be the most important in our task.", + "General OP/PC properties. These features capture the general properties of a conversation. They can be used to characterize the background distribution of echoing.", + "Table TABREF18 further shows the intuition for including each feature, and condensed $t$-test results after Bonferroni correction. Specifically, we test whether the words that were echoed in explanations have different feature values from those that were not echoed. In addition to considering all words, we also separately consider stopwords and content words in light of Figure FIGREF8. Here, we highlight a few observations:", + "[itemsep=0pt,leftmargin=*,topsep=0pt]", + "Although we expect more complicated words (#characters) to be echoed more often, this is not the case on average. We also observe an interesting example of Simpson's paradox in the results for Wordnet depth BIBREF38: shallower words are more likely to be echoed across all words, but deeper words are more likely to be echoed in content words and stopwords.", + "OPs and PCs generally exhibit similar behavior for most features, except for part-of-speech and grammatical relation (subject, object, and other.) For instance, verbs in an OP are less likely to be echoed, while verbs in a PC are more likely to be echoed.", + "Although nouns from both OPs and PCs are less likely to be echoed, within content words, subjects and objects from an OP are more likely to be echoed. Surprisingly, subjects and objects in a PC are less likely to be echoed, which suggests that the original poster tends to refer back to their own subjects and objects, or introduce new ones, when providing explanations.", + "Later words in OPs and PCs are more likely to be echoed, especially in OPs. This could relate to OPs summarizing their rationales at the end of their post and PCs putting their strongest points last.", + "Although the number of surface forms in an OP or PC is positively correlated with being echoed, the differences in surface forms show reverse trends: the more surface forms of a word that show up only in the PC (i.e., not in the OP), the more likely a word is to be echoed. However, the reverse is true for the number of surface forms in only the OP. Such contrast echoes BIBREF31, in which dissimilarity in word usage between the OP and PC was a predictive feature of successful persuasion." + ], + [ + "We further examine the effectiveness of our proposed features in a predictive setting. These features achieve strong performance in the word-level classification task, and can enhance neural models in both the word-level task and generating explanations. However, the word-level task remains challenging, especially for content words." + ], + [ + "We consider two classifiers for our word-level classification task: logistic regression and gradient boosting tree (XGBoost) BIBREF39. We hypothesized that XGBoost would outperform logistic regression because our problem is non-linear, as shown in Figure FIGREF8.", + "To examine the utility of our features in a neural framework, we further adapt our word-level task as a tagging task, and use LSTM as a baseline. Specifically, we concatenate an OP and PC with a special token as the separator so that an LSTM model can potentially distinguish the OP from PC, and then tag each word based on the label of its stemmed version. We use GloVe embeddings to initialize the word embeddings BIBREF40. We concatenate our proposed features of the corresponding stemmed word to the word embedding; the resulting difference in performance between a vanilla LSTM demonstrates the utility of our proposed features. We scale all features to $[0, 1]$ before fitting the models. As introduced in Section SECREF3, we split our tuples of (OP, PC, explanation) into training, validation, and test sets, and use the validation set for hyperparameter tuning. Refer to the supplementary material for additional details in the experiment.", + "Evaluation metric. Since our problem is imbalanced, we use the F1 score as our evaluation metric. For the tagging approach, we average the labels of words with the same stemmed version to obtain a single prediction for the stemmed word. To establish a baseline, we consider a random method that predicts the positive label with 0.15 probability (the base rate of positive instances)." + ], + [ + "Overall performance (Figure FIGREF28). Although our word-level task is heavily imbalanced, all of our models outperform the random baseline by a wide margin. As expected, content words are much more difficult to predict than stopwords, but the best F1 score in content words more than doubles that of the random baseline (0.286 vs. 0.116). Notably, although we strongly improve on our random baseline, even our best F1 scores are relatively low, and this holds true regardless of the model used. Despite involving more tokens than standard tagging tasks (e.g., BIBREF41 and BIBREF42), predicting whether a word is going to be echoed in explanations remains a challenging problem.", + "Although the vanilla LSTM model incorporates additional knowledge (in the form of word embeddings), the feature-based XGBoost and logistic regression models both outperform the vanilla LSTM model. Concatenating our proposed features with word embeddings leads to improved performance from the LSTM model, which becomes comparable to XGBoost. This suggests that our proposed features can be difficult to learn with an LSTM alone.", + "Despite the non-linearity observed in Figure FIGREF8, XGBoost only outperforms logistic regression by a small margin. In the rest of this section, we use XGBoost to further examine the effectiveness of different groups of features, and model performance in different conditions.", + "Ablation performance (Table TABREF34). First, if we only consider a single group of features, as we hypothesized, the relation between OP and PC is crucial and leads to almost as strong performance in content words as using all features. To further understand the strong performance of OP-PC relation, Figure FIGREF28 shows the feature importance in the ablated model, measured by the normalized total gain (see the supplementary material for feature importance in the full model). A word's occurrence in both the OP and PC is clearly the most important feature, with distance between its POS tag distributions as the second most important. Recall that in Table TABREF18 we show that words that have similar POS behavior between the OP and PC are more likely to be echoed in the explanation.", + "Overall, it seems that word-level properties contribute the most valuable signals for predicting stopwords. If we restrict ourselves to only information in either an OP or PC, how a word is used in a PC is much more predictive of content word echoing (0.233 vs 0.191). This observation suggests that, for content words, the PC captures more valuable information than the OP. This finding is somewhat surprising given that the OP sets the topic of discussion and writes the explanation.", + "As for the effects of removing a group of features, we can see that there is little change in the performance on content words. This can be explained by the strong performance of the OP-PC relation on its own, and the possibility of the OP-PC relation being approximated by OP and PC usage. Again, word-level properties are valuable for strong performance in stopwords.", + "Performance vs. word source (Figure FIGREF28). We further break down the performance by where a word is from. We can group a word based on whether it shows up only in an OP, a PC, or both OP and PC, as shown in Table TABREF1. There is a striking difference between the performance in the three categories (e.g., for all words, 0.63 in OP & PC vs. 0.271 in PC only). The strong performance on words in both the OP and PC applies to stopwords and content words, even accounting for the shift in the random baseline, and recalls the importance of occurring both in OP and PC as a feature.", + "Furthermore, the echoing of words from the PC is harder to predict (0.271) than from the OP (0.347) despite the fact that words only in PCs are more likely to be echoed than words only in OPs (13.5% vs. 8.6%). The performance difference is driven by stopwords, suggesting that our overall model is better at capturing signals for stopwords used in OPs. This might relate to the fact that the OP and the explanation are written by the same author; prior studies have demonstrated the important role of stopwords for authorship attribution BIBREF43.", + "Nouns are the most reliably predicted part-of-speech tag within content words (Table TABREF35). Next, we break down the performance by part-of-speech tags. We focus on the part-of-speech tags that are semantically important, namely, nouns, proper nouns, verbs, adverbs, and adjectives.", + "Prediction performance can be seen as a proxy for how reliably a part-of-speech tag is reused when providing explanations. Consistent with our expectations for the importance of nouns and verbs, our models achieve the best performance on nouns within content words. Verbs are more challenging, but become the least difficult tag to predict when we consider all words, likely due to stopwords such as \u201chave.\u201d Adjectives turn out to be the most challenging category, suggesting that adjectival choice is perhaps more arbitrary than other parts of speech, and therefore less central to the process of constructing an explanation. The important role of nouns in shaping explanations resonates with the high recall rate of nouns in memory tasks BIBREF44." + ], + [ + "One way to measure the ultimate success of understanding pointers in explanations is to be able to generate explanations. We use the pointer generator network with coverage as our starting point BIBREF4, BIBREF46 (see the supplementary material for details). We investigate whether concatenating our proposed features with word embeddings can improve generation performance, as measured by ROUGE scores.", + "Consistent with results in sequence tagging for word-level echoing prediction, our proposed features can enhance a neural model with copying mechanisms (see Table TABREF37). Specifically, their use leads to statistically significant improvement in ROUGE-1 and ROUGE-L, while slightly hurting the performance in ROUGE-2 (the difference is not statistically significant). We also find that our features can increase the likelihood of copying: an average of 17.59 unique words get copied to the generated explanation with our features, compared to 14.17 unique words without our features. For comparison, target explanations have an average of 34.81 unique words. We emphasize that generating explanations is a very challenging task (evidenced by the low ROUGE scores and examples in the supplementary material), and that fully solving the generation task requires more work." + ], + [ + "In this work, we conduct the first large-scale empirical study of everyday explanations in the context of persuasion. We assemble a novel dataset and formulate a word-level prediction task to understand the selective nature of explanations. Our results suggest that the relation between an OP and PC plays an important role in predicting the echoing of content words, while a word's non-contextual properties matter for stopwords. We show that vanilla LSTMs fail to learn some of the features we develop and that our proposed features can improve the performance in generating explanations using pointer networks. We also demonstrate the important role of nouns in shaping explanations.", + "Although our approach strongly outperforms random baselines, the relatively low F1 scores indicate that predicting which word is echoed in explanations is a very challenging task. It follows that we are only able to derive a limited understanding of how people choose to echo words in explanations. The extent to which explanation construction is fundamentally random BIBREF47, or whether there exist other unidentified patterns, is of course an open question. We hope that our study and the resources that we release encourage further work in understanding the pragmatics of explanations.", + "There are many promising research directions for future work in advancing the computational understanding of explanations. First, although /r/ChangeMyView has the useful property that its explanations are closely connected to its explananda, it is important to further investigate the extent to which our findings generalize beyond /r/ChangeMyView and Reddit and establish universal properties of explanations. Second, it is important to connect the words in explanations that we investigate here to the structure of explanations in pyschology BIBREF7. Third, in addition to understanding what goes into an explanation, we need to understand what makes an explanation effective. A better understanding of explanations not only helps develop explainable AI, but also informs the process of collecting explanations that machine learning systems learn from BIBREF48, BIBREF49, BIBREF50." + ], + [ + "We thank Kimberley Buchan, anonymous reviewers, and members of the NLP+CSS research group at CU Boulder for their insightful comments and discussions; Jason Baumgartner for sharing the dataset that enabled this research." + ], + [ + "Before tokenizing, we pass each OP, PC, and explanation through a preprocessing pipeline, with the following steps:", + "Occasionally, /r/ChangeMyView's moderators will edit comments, prefixing their edits with \u201cHello, users of CMV\u201d or \u201cThis is a footnote\u201d (see Table TABREF46). We remove this, and any text that follows on the same line.", + "We replace URLs with a \u201c@url@\u201d token, defining a URL to be any string which matches the following regular expression: (https?://[^\\s)]*).", + "We replace \u201c$\\Delta $\u201d symbols and their analogues\u2014such as \u201c$\\delta $\u201d, \u201c&;#8710;\u201d, and \u201c!delta\u201d\u2014with the word \u201cdelta\u201d. We also remove the word \u201cdelta\u201d from explanations, if the explanation starts with delta.", + "Reddit\u2013specific prefixes, such as \u201cu/\u201d (denoting a user) and \u201cr/\u201d (denoting a subreddit) are removed, as we observed that they often interfered with spaCy's ability to correctly parse its inputs.", + "We remove any text matching the regular expression EDIT(.*?):.* from the beginning of the match to the end of that line, as well as variations, such as Edit(.*?):.*.", + "Reddit allows users to insert blockquoted text. We extract any blockquotes and surround them with standard quotation marks.", + "We replace all contiguous whitespace with a single space. We also do this with tab characters and carriage returns, and with two or more hyphens, asterisks, or underscores.", + "Tokenizing the data. After passing text through our preprocessing pipeline, we use the default spaCy pipeline to extract part-of-speech tags, dependency tags, and entity details for each token BIBREF35. In addition, we use NLTK to stem words BIBREF36. This is used to compute all word level features discussed in Section 4 of the main paper." + ], + [ + "Figure FIGREF49 shows a similar U-shape in the probability of a word being echoed in PC. However, visually, we can see that rare words seem more likely to have high echoing probability in explanations, while that probability is higher for words with moderate frequency in PCs. As PCs tend to be longer than explanations, we also used the echoing probability of the most frequent words to normalize the probability of other words so that they are comparable. We indeed observed a higher likelihood of echoing the rare words, but lower likelihood of echoing words with moderate frequency in explanations than in PCs." + ], + [ + "Given an OP, PC, and explanation, we calculate a 66\u2013dimensional vector for each unique stem in the concatenated OP and PC. Here, we describe the process of calculating each feature.", + "Inverse document frequency: for a stem $s$, the inverse document frequency is given by $\\log \\frac{N}{\\mathrm {df}_s}$, where $N$ is the total number of documents (here, OPs and PCs) in the training set, and $\\mathrm {df}_s$ is the number of documents in the training data whose set of stemmed words contains $s$.", + "Stem length: the number of characters in the stem.", + "Wordnet depth (min): starting with the stem, this is the length of the minimum hypernym path to the synset root.", + "Wordnet depth (max): similarly, this is the length of the maximum hypernym path.", + "Stem transfer probability: the percentage of times in which a stem seen in the explanandum is also seen in the explanation. If, during validation or testing, a stem is encountered for the first time, we set this to be the mean probability of transfer over all stems seen in the training data.", + "OP part\u2013of\u2013speech tags: a stem can represent multiple parts of speech. For example, both \u201ctraditions\u201d and \u201ctraditional\u201d will be stemmed to \u201ctradit.\u201d We count the percentage of times the given stem appears as each part\u2013of\u2013speech tag, following the Universal Dependencies scheme BIBREF53. If the stem does not appear in the OP, each part\u2013of\u2013speech feature will be $\\frac{1}{16}$.", + "OP subject, object, and other: Given a stem $s$, we calculate the percentage of times that $s$'s surface forms in the OP are classified as subjects, objects, or something else by SpaCy. We follow the CLEAR guidelines, BIBREF51 and use the following tags to indicate a subject: nsubj, nsubjpass, csubj, csubjpass, agent, and expl. Objects are identified using these tags: dobj, dative, attr, oprd. If $s$ does not appear at all in the OP, we let subject, object, and other each equal $\\frac{1}{3}$.", + "OP term frequency: the number of times any surface form of a stem appears in the list of tokens that make up the OP.", + "OP normalized term frequency: the percentage of the OP's tokens which are a surface form of the given stem.", + "OP # of surface forms: the number of different surface forms for the given stem.", + "OP location: the average location of each surface form of the given stem which appears in the OP, where the location of a surface form is defined as the percentage of tokens which appear after that surface form. If the stem does not appear at all in the OP, this value is $\\frac{1}{2}$.", + "OP is in quotes: the number of times the stem appears in the OP surrounded by quotation marks.", + "OP is entity: the percentage of tokens in the OP that are both a surface form for the given stem, and are tagged by SpaCy as one of the following entities: PERSON, NORP, FAC, ORG, GPE, LOC, PRODUCT, EVENT, WORK_OF_ART, LAW, and LANGUAGE.", + "PC equivalents of features 6-30.", + "In both OP and PC: 1, if one of the stem's surface forms appears in both the OP and PC. 0 otherwise.", + "# of unique surface forms in OP: for the given stem, the number of surface forms that appear in the OP, but not in the PC.", + "# of unique surface forms in PC: for the given stem, the number of surface forms that appear in the PC, but not in the OP.", + "Stem part\u2013of\u2013speech distribution difference: we consider the concatenation of features 6-21, along with the concatenation of features 31-46, as two distributions, and calculate the Jensen\u2013Shannon divergence between them.", + "Stem dependency distribution difference: similarly, we consider the concatenation of features 22-24 (OP dependency labels), and the concatenation of features 47-49 (PC dependency labels), as two distributions, and calculate the Jensen\u2013Shannon divergence between them.", + "OP length: the number of tokens in the OP.", + "PC length: the number of tokens in the PC.", + "Length difference: the absolute value of the difference between OP length and PC length.", + "Avg. word length difference: the difference between the average number of characters per token in the OP and the average number of characters per token in the PC.", + "OP/PC part\u2013of\u2013speech tag distribution difference: the Jensen\u2013Shannon divergence between the part\u2013of\u2013speech tag distributions of the OP on the one hand, and the PC on the other.", + "Depth of the PC in the thread: since there can be many back\u2013and\u2013forth replies before a user awards a delta, we number each comment in a thread, starting at 0 for the OP, and incrementing for each new comment before the PC appears." + ], + [ + "For each non\u2013LSTM classifier, we train 11 models: one full model, and forward and backward models for each of the five feature groups. To train, we fit on the training set and use the validation set for hyperparameter tuning.", + "For the random model, since the echo rate of the training set is 15%, we simply predict 1 with 15% probability, and 0 otherwise.", + "For logistic regression, we use the lbfgs solver. To tune hyperparameters, we perform an exhaustive grid search, with $C$ taking values from $\\lbrace 10^{x}:x\\in \\lbrace -1, 0, 1, 2, 3, 4\\rbrace \\rbrace $, and the respective weights of the negative and positive classes taking values from $\\lbrace (x, 1-x): x\\in \\lbrace 0.25, 0.20, 0.15\\rbrace \\rbrace $.", + "We also train XGBoost models. Here, we use a learning rate of $0.1$, 1000 estimator trees, and no subsampling. We perform an exhaustive grid search to tune hyperparameters, with the max tree depth equaling 5, 7, or 9, the minimum weight of a child equaling 3, 5, or 7, and the weight of a positive class instance equaling 3, 4, or 5.", + "Finally, we train two LSTM models, each with a single 300\u2013dimensional hidden layer. Due to efficiency considerations, we eschewed a full search of the parameter space, but experimented with different values of dropout, learning rate, positive class weight, and batch size. We ultimately trained each model for five epochs with a batch size of 32 and a learning rate of 0.001, using the Adam optimizer BIBREF52. We also weight positive instances four times more highly than negative instances." + ], + [ + "We formulate an abstractive summarization task using an OP concatenated with the PC as a source, and the explanation as target. We train two models, one with the features described above, and one without. A shared vocabulary of 50k words is constructed from the training set by setting the maximum encoding length to 500 words. We set the maximum decoding length to 100. We use a pointer generator network with coverage for generating explanations, using a bidirectional LSTM as an encoder and a unidirectional LSTM as a decoder. Both use a 256-dimensional hidden state. The parameters of this network are tuned using a validation set of five thousand instances. We constrain the batch size to 16 and train the network for 20k steps, using the parameters described in Table TABREF82." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0660/instruction.md b/qasper-0660/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c933bff179914ef915255f3f2b49307deb624e34 --- /dev/null +++ b/qasper-0660/instruction.md @@ -0,0 +1,169 @@ +Name of Paper: What Gets Echoed? Understanding the"Pointers"in Explanations of Persuasive Arguments + +Question: What is the baseline? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Dataset", + "Understanding the Pointers in Explanations", + "Predicting Pointers", + "Predicting Pointers ::: Experiment setup", + "Predicting Pointers ::: Prediction Performance", + "Predicting Pointers ::: The Effect on Generating Explanations", + "Concluding Discussions", + "Acknowledgments", + "Supplemental Material ::: Preprocessing.", + "Supplemental Material ::: PC Echoing OP", + "Supplemental Material ::: Feature Calculation", + "Supplemental Material ::: Word\u2013level Prediction Task", + "Supplemental Material ::: Generating Explanations" + ], + "paragraphs": [ + [ + "Explanations are essential for understanding and learning BIBREF0. They can take many forms, ranging from everyday explanations for questions such as why one likes Star Wars, to sophisticated formalization in the philosophy of science BIBREF1, to simply highlighting features in recent work on interpretable machine learning BIBREF2.", + "Although everyday explanations are mostly encoded in natural language, natural language explanations remain understudied in NLP, partly due to a lack of appropriate datasets and problem formulations. To address these challenges, we leverage /r/ChangeMyView, a community dedicated to sharing counterarguments to controversial views on Reddit, to build a sizable dataset of naturally-occurring explanations. Specifically, in /r/ChangeMyView, an original poster (OP) first delineates the rationales for a (controversial) opinion (e.g., in Table TABREF1, \u201cmost hit music artists today are bad musicians\u201d). Members of /r/ChangeMyView are invited to provide counterarguments. If a counterargument changes the OP's view, the OP awards a $\\Delta $ to indicate the change and is required to explain why the counterargument is persuasive. In this work, we refer to what is being explained, including both the original post and the persuasive comment, as the explanandum.", + "An important advantage of explanations in /r/ChangeMyView is that the explanandum contains most of the required information to provide its explanation. These explanations often select key counterarguments in the persuasive comment and connect them with the original post. As shown in Table TABREF1, the explanation naturally points to, or echoes, part of the explanandum (including both the persuasive comment and the original post) and in this case highlights the argument of \u201cmusic serving different purposes.\u201d", + "These naturally-occurring explanations thus enable us to computationally investigate the selective nature of explanations: \u201cpeople rarely, if ever, expect an explanation that consists of an actual and complete cause of an event. Humans are adept at selecting one or two causes from a sometimes infinite number of causes to be the explanation\u201d BIBREF3. To understand the selective process of providing explanations, we formulate a word-level task to predict whether a word in an explanandum will be echoed in its explanation.", + "Inspired by the observation that words that are likely to be echoed are either frequent or rare, we propose a variety of features to capture how a word is used in the explanandum as well as its non-contextual properties in Section SECREF4. We find that a word's usage in the original post and in the persuasive argument are similarly related to being echoed, except in part-of-speech tags and grammatical relations. For instance, verbs in the original post are less likely to be echoed, while the relationship is reversed in the persuasive argument.", + "We further demonstrate that these features can significantly outperform a random baseline and even a neural model with significantly more knowledge of a word's context. The difficulty of predicting whether content words (i.e., non-stopwords) are echoed is much greater than that of stopwords, among which adjectives are the most difficult and nouns are relatively the easiest. This observation highlights the important role of nouns in explanations. We also find that the relationship between a word's usage in the original post and in the persuasive comment is crucial for predicting the echoing of content words. Our proposed features can also improve the performance of pointer generator networks with coverage in generating explanations BIBREF4.", + "To summarize, our main contributions are:", + "[itemsep=0pt,leftmargin=*,topsep=0pt]", + "We highlight the importance of computationally characterizing human explanations and formulate a concrete problem of predicting how information is selected from explananda to form explanations, including building a novel dataset of naturally-occurring explanations.", + "We provide a computational characterization of natural language explanations and demonstrate the U-shape in which words get echoed.", + "We identify interesting patterns in what gets echoed through a novel word-level classification task, including the importance of nouns in shaping explanations and the importance of contextual properties of both the original post and persuasive comment in predicting the echoing of content words.", + "We show that vanilla LSTMs fail to learn some of the features we develop and that the proposed features can even improve performance in generating explanations with pointer networks.", + "Our code and dataset is available at https://chenhaot.com/papers/explanation-pointers.html." + ], + [ + "To provide background for our study, we first present a brief overview of explanations for the NLP community, and then discuss the connection of our study with pointer networks, linguistic accommodation, and argumentation mining.", + "The most developed discussion of explanations is in the philosophy of science. Extensive studies aim to develop formal models of explanations (e.g., the deductive-nomological model in BIBREF5, see BIBREF1 and BIBREF6 for a review). In this view, explanations are like proofs in logic. On the other hand, psychology and cognitive sciences examine \u201ceveryday explanations\u201d BIBREF0, BIBREF7. These explanations tend to be selective, are typically encoded in natural language, and shape our understanding and learning in life despite the absence of \u201caxioms.\u201d Please refer to BIBREF8 for a detailed comparison of these two modes of explanation.", + "Although explanations have attracted significant interest from the AI community thanks to the growing interest on interpretable machine learning BIBREF9, BIBREF10, BIBREF11, such studies seldom refer to prior work in social sciences BIBREF3. Recent studies also show that explanations such as highlighting important features induce limited improvement on human performance in detecting deceptive reviews and media biases BIBREF12, BIBREF13. Therefore, we believe that developing a computational understanding of everyday explanations is crucial for explainable AI. Here we provide a data-driven study of everyday explanations in the context of persuasion.", + "In particular, we investigate the \u201cpointers\u201d in explanations, inspired by recent work on pointer networks BIBREF14. Copying mechanisms allow a decoder to generate a token by copying from the source, and have been shown to be effective in generation tasks ranging from summarization to program synthesis BIBREF4, BIBREF15, BIBREF16. To the best of our knowledge, our work is the first to investigate the phenomenon of pointers in explanations.", + "Linguistic accommodation and studies on quotations also examine the phenomenon of reusing words BIBREF17, BIBREF18, BIBREF19, BIBREF20. For instance, BIBREF21 show that power differences are reflected in the echoing of function words; BIBREF22 find that news media prefer to quote locally distinct sentences in political debates. In comparison, our word-level formulation presents a fine-grained view of echoing words, and puts a stronger emphasis on content words than work on linguistic accommodation.", + "Finally, our work is concerned with an especially challenging problem in social interaction: persuasion. A battery of studies have done work to enhance our understanding of persuasive arguments BIBREF23, BIBREF24, BIBREF25, BIBREF26, BIBREF27, and the area of argumentation mining specifically investigates the structure of arguments BIBREF28, BIBREF29, BIBREF30. We build on previous work by BIBREF31 and leverage the dynamics of /r/ChangeMyView. Although our findings are certainly related to the persuasion process, we focus on understanding the self-described reasons for persuasion, instead of the structure of arguments or the factors that drive effective persuasion." + ], + [ + "Our dataset is derived from the /r/ChangeMyView subreddit, which has more than 720K subscribers BIBREF31. /r/ChangeMyView hosts conversations where someone expresses a view and others then try to change that person's mind. Despite being fundamentally based on argument, /r/ChangeMyView has a reputation for being remarkably civil and productive BIBREF32, e.g., a journalist wrote \u201cIn a culture of brittle talking points that we guard with our lives, Change My View is a source of motion and surprise\u201d BIBREF33.", + "The delta mechanism in /r/ChangeMyView allows members to acknowledge opinion changes and enables us to identify explanations for opinion changes BIBREF34. Specifically, it requires \u201cAny user, whether they're the OP or not, should reply to a comment that changed their view with a delta symbol and an explanation of the change.\u201d As a result, we have access to tens of thousands of naturally-occurring explanations and associated explananda. In this work, we focus on the opinion changes of the original posters.", + "Throughout this paper, we use the following terminology:", + "[itemsep=-5pt,leftmargin=*,topsep=0pt]", + "An original post (OP) is an initial post where the original poster justifies his or her opinion. We also use OP to refer to the original poster.", + "A persuasive comment (PC) is a comment that directly leads to an opinion change on the part of the OP (i.e., winning a $\\Delta $).", + "A top-level comment is a comment that directly replies to an OP, and /r/ChangeMyView requires the top-level comment to \u201cchallenge at least one aspect of OP\u2019s stated view (however minor), unless they are asking a clarifying question.\u201d", + "An explanation is a comment where an OP acknowledges a change in his or her view and provides an explanation of the change. As shown in Table TABREF1, the explanation not only provides a rationale, it can also include other discourse acts, such as expressing gratitude.", + "Using https://pushshift.io, we collect the posts and comments in /r/ChangeMyView from January 17th, 2013 to January 31st, 2019, and extract tuples of (OP, PC, explanation). We use the tuples from the final six months of our dataset as the test set, those from the six months before that as the validation set, and the remaining tuples as the training set. The sets contain 5,270, 5,831, and 26,617 tuples respectively. Note that there is no overlap in time between the three sets and the test set can therefore be used to assess generalization including potential changes in community norms and world events.", + "Preprocessing. We perform a number of preprocessing steps, such as converting blockquotes in Markdown to quotes, filtering explicit edits made by authors, mapping all URLs to a special @url@ token, and replacing hyperlinks with the link text. We ignore all triples that contain any deleted comments or posts. We use spaCy for tokenization and tagging BIBREF35. We also use the NLTK implementation of the Porter stemming algorithm to store the stemmed version of each word, for later use in our prediction task BIBREF36, BIBREF37. Refer to the supplementary material for more information on preprocessing.", + "Data statistics. Table TABREF16 provides basic statistics of the training tuples and how they compare to other comments. We highlight the fact that PCs are on average longer than top-level comments, suggesting that PCs contain substantial counterarguments that directly contribute to opinion change. Therefore, we simplify the problem by focusing on the (OP, PC, explanation) tuples and ignore any other exchanges between an OP and a commenter.", + "Below, we highlight some notable features of explanations as they appear in our dataset.", + "The length of explanations shows stronger correlation with that of OPs and PCs than between OPs and PCs (Figure FIGREF8). This observation indicates that explanations are somehow better related with OPs and PCs than PCs are with OPs in terms of language use. A possible reason is that the explainer combines their natural tendency towards length with accommodating the PC.", + "Explanations have a greater fraction of \u201cpointers\u201d than do persuasive comments (Figure FIGREF8). We measure the likelihood of a word in an explanation being copied from either its OP or PC and provide a similar probability for a PC for copying from its OP. As we discussed in Section SECREF1, the words in an explanation are much more likely to come from the existing discussion than are the words in a PC (59.8% vs 39.0%). This phenomenon holds even if we restrict ourselves to considering words outside quotations, which removes the effect of quoting other parts of the discussion, and if we focus only on content words, which removes the effect of \u201creusing\u201d stopwords.", + "Relation between a word being echoed and its document frequency (Figure FIGREF8). Finally, as a preview of our main results, the document frequency of a word from the explanandum is related to the probability of being echoed in the explanation. Although the average likelihood declines as the document frequency gets lower, we observe an intriguing U-shape in the scatter plot. In other words, the words that are most likely to be echoed are either unusually frequent or unusually rare, while most words in the middle show a moderate likelihood of being echoed." + ], + [ + "To further investigate how explanations select words from the explanandum, we formulate a word-level prediction task to predict whether words in an OP or PC are echoed in its explanation. Formally, given a tuple of (OP, PC, explanation), we extract the unique stemmed words as $\\mathcal {V}_{\\text{OP}}, \\mathcal {V}_{\\text{PC}}, \\mathcal {V}_{\\text{EXP}}$. We then define the label for each word in the OP or PC, $w \\in \\mathcal {V}_{\\text{OP}} \\cup \\mathcal {V}_{\\text{PC}}$, based on the explanation as follows:", + "Our prediction task is thus a straightforward binary classification task at the word level. We develop the following five groups of features to capture properties of how a word is used in the explanandum (see Table TABREF18 for the full list):", + "[itemsep=0pt,leftmargin=*,topsep=0pt]", + "Non-contextual properties of a word. These features are derived directly from the word and capture the general tendency of a word being echoed in explanations.", + "Word usage in an OP or PC (two groups). These features capture how a word is used in an OP or PC. As a result, for each feature, we have two values for the OP and PC respectively.", + "How a word connects an OP and PC. These features look at the difference between word usage in the OP and PC. We expect this group to be the most important in our task.", + "General OP/PC properties. These features capture the general properties of a conversation. They can be used to characterize the background distribution of echoing.", + "Table TABREF18 further shows the intuition for including each feature, and condensed $t$-test results after Bonferroni correction. Specifically, we test whether the words that were echoed in explanations have different feature values from those that were not echoed. In addition to considering all words, we also separately consider stopwords and content words in light of Figure FIGREF8. Here, we highlight a few observations:", + "[itemsep=0pt,leftmargin=*,topsep=0pt]", + "Although we expect more complicated words (#characters) to be echoed more often, this is not the case on average. We also observe an interesting example of Simpson's paradox in the results for Wordnet depth BIBREF38: shallower words are more likely to be echoed across all words, but deeper words are more likely to be echoed in content words and stopwords.", + "OPs and PCs generally exhibit similar behavior for most features, except for part-of-speech and grammatical relation (subject, object, and other.) For instance, verbs in an OP are less likely to be echoed, while verbs in a PC are more likely to be echoed.", + "Although nouns from both OPs and PCs are less likely to be echoed, within content words, subjects and objects from an OP are more likely to be echoed. Surprisingly, subjects and objects in a PC are less likely to be echoed, which suggests that the original poster tends to refer back to their own subjects and objects, or introduce new ones, when providing explanations.", + "Later words in OPs and PCs are more likely to be echoed, especially in OPs. This could relate to OPs summarizing their rationales at the end of their post and PCs putting their strongest points last.", + "Although the number of surface forms in an OP or PC is positively correlated with being echoed, the differences in surface forms show reverse trends: the more surface forms of a word that show up only in the PC (i.e., not in the OP), the more likely a word is to be echoed. However, the reverse is true for the number of surface forms in only the OP. Such contrast echoes BIBREF31, in which dissimilarity in word usage between the OP and PC was a predictive feature of successful persuasion." + ], + [ + "We further examine the effectiveness of our proposed features in a predictive setting. These features achieve strong performance in the word-level classification task, and can enhance neural models in both the word-level task and generating explanations. However, the word-level task remains challenging, especially for content words." + ], + [ + "We consider two classifiers for our word-level classification task: logistic regression and gradient boosting tree (XGBoost) BIBREF39. We hypothesized that XGBoost would outperform logistic regression because our problem is non-linear, as shown in Figure FIGREF8.", + "To examine the utility of our features in a neural framework, we further adapt our word-level task as a tagging task, and use LSTM as a baseline. Specifically, we concatenate an OP and PC with a special token as the separator so that an LSTM model can potentially distinguish the OP from PC, and then tag each word based on the label of its stemmed version. We use GloVe embeddings to initialize the word embeddings BIBREF40. We concatenate our proposed features of the corresponding stemmed word to the word embedding; the resulting difference in performance between a vanilla LSTM demonstrates the utility of our proposed features. We scale all features to $[0, 1]$ before fitting the models. As introduced in Section SECREF3, we split our tuples of (OP, PC, explanation) into training, validation, and test sets, and use the validation set for hyperparameter tuning. Refer to the supplementary material for additional details in the experiment.", + "Evaluation metric. Since our problem is imbalanced, we use the F1 score as our evaluation metric. For the tagging approach, we average the labels of words with the same stemmed version to obtain a single prediction for the stemmed word. To establish a baseline, we consider a random method that predicts the positive label with 0.15 probability (the base rate of positive instances)." + ], + [ + "Overall performance (Figure FIGREF28). Although our word-level task is heavily imbalanced, all of our models outperform the random baseline by a wide margin. As expected, content words are much more difficult to predict than stopwords, but the best F1 score in content words more than doubles that of the random baseline (0.286 vs. 0.116). Notably, although we strongly improve on our random baseline, even our best F1 scores are relatively low, and this holds true regardless of the model used. Despite involving more tokens than standard tagging tasks (e.g., BIBREF41 and BIBREF42), predicting whether a word is going to be echoed in explanations remains a challenging problem.", + "Although the vanilla LSTM model incorporates additional knowledge (in the form of word embeddings), the feature-based XGBoost and logistic regression models both outperform the vanilla LSTM model. Concatenating our proposed features with word embeddings leads to improved performance from the LSTM model, which becomes comparable to XGBoost. This suggests that our proposed features can be difficult to learn with an LSTM alone.", + "Despite the non-linearity observed in Figure FIGREF8, XGBoost only outperforms logistic regression by a small margin. In the rest of this section, we use XGBoost to further examine the effectiveness of different groups of features, and model performance in different conditions.", + "Ablation performance (Table TABREF34). First, if we only consider a single group of features, as we hypothesized, the relation between OP and PC is crucial and leads to almost as strong performance in content words as using all features. To further understand the strong performance of OP-PC relation, Figure FIGREF28 shows the feature importance in the ablated model, measured by the normalized total gain (see the supplementary material for feature importance in the full model). A word's occurrence in both the OP and PC is clearly the most important feature, with distance between its POS tag distributions as the second most important. Recall that in Table TABREF18 we show that words that have similar POS behavior between the OP and PC are more likely to be echoed in the explanation.", + "Overall, it seems that word-level properties contribute the most valuable signals for predicting stopwords. If we restrict ourselves to only information in either an OP or PC, how a word is used in a PC is much more predictive of content word echoing (0.233 vs 0.191). This observation suggests that, for content words, the PC captures more valuable information than the OP. This finding is somewhat surprising given that the OP sets the topic of discussion and writes the explanation.", + "As for the effects of removing a group of features, we can see that there is little change in the performance on content words. This can be explained by the strong performance of the OP-PC relation on its own, and the possibility of the OP-PC relation being approximated by OP and PC usage. Again, word-level properties are valuable for strong performance in stopwords.", + "Performance vs. word source (Figure FIGREF28). We further break down the performance by where a word is from. We can group a word based on whether it shows up only in an OP, a PC, or both OP and PC, as shown in Table TABREF1. There is a striking difference between the performance in the three categories (e.g., for all words, 0.63 in OP & PC vs. 0.271 in PC only). The strong performance on words in both the OP and PC applies to stopwords and content words, even accounting for the shift in the random baseline, and recalls the importance of occurring both in OP and PC as a feature.", + "Furthermore, the echoing of words from the PC is harder to predict (0.271) than from the OP (0.347) despite the fact that words only in PCs are more likely to be echoed than words only in OPs (13.5% vs. 8.6%). The performance difference is driven by stopwords, suggesting that our overall model is better at capturing signals for stopwords used in OPs. This might relate to the fact that the OP and the explanation are written by the same author; prior studies have demonstrated the important role of stopwords for authorship attribution BIBREF43.", + "Nouns are the most reliably predicted part-of-speech tag within content words (Table TABREF35). Next, we break down the performance by part-of-speech tags. We focus on the part-of-speech tags that are semantically important, namely, nouns, proper nouns, verbs, adverbs, and adjectives.", + "Prediction performance can be seen as a proxy for how reliably a part-of-speech tag is reused when providing explanations. Consistent with our expectations for the importance of nouns and verbs, our models achieve the best performance on nouns within content words. Verbs are more challenging, but become the least difficult tag to predict when we consider all words, likely due to stopwords such as \u201chave.\u201d Adjectives turn out to be the most challenging category, suggesting that adjectival choice is perhaps more arbitrary than other parts of speech, and therefore less central to the process of constructing an explanation. The important role of nouns in shaping explanations resonates with the high recall rate of nouns in memory tasks BIBREF44." + ], + [ + "One way to measure the ultimate success of understanding pointers in explanations is to be able to generate explanations. We use the pointer generator network with coverage as our starting point BIBREF4, BIBREF46 (see the supplementary material for details). We investigate whether concatenating our proposed features with word embeddings can improve generation performance, as measured by ROUGE scores.", + "Consistent with results in sequence tagging for word-level echoing prediction, our proposed features can enhance a neural model with copying mechanisms (see Table TABREF37). Specifically, their use leads to statistically significant improvement in ROUGE-1 and ROUGE-L, while slightly hurting the performance in ROUGE-2 (the difference is not statistically significant). We also find that our features can increase the likelihood of copying: an average of 17.59 unique words get copied to the generated explanation with our features, compared to 14.17 unique words without our features. For comparison, target explanations have an average of 34.81 unique words. We emphasize that generating explanations is a very challenging task (evidenced by the low ROUGE scores and examples in the supplementary material), and that fully solving the generation task requires more work." + ], + [ + "In this work, we conduct the first large-scale empirical study of everyday explanations in the context of persuasion. We assemble a novel dataset and formulate a word-level prediction task to understand the selective nature of explanations. Our results suggest that the relation between an OP and PC plays an important role in predicting the echoing of content words, while a word's non-contextual properties matter for stopwords. We show that vanilla LSTMs fail to learn some of the features we develop and that our proposed features can improve the performance in generating explanations using pointer networks. We also demonstrate the important role of nouns in shaping explanations.", + "Although our approach strongly outperforms random baselines, the relatively low F1 scores indicate that predicting which word is echoed in explanations is a very challenging task. It follows that we are only able to derive a limited understanding of how people choose to echo words in explanations. The extent to which explanation construction is fundamentally random BIBREF47, or whether there exist other unidentified patterns, is of course an open question. We hope that our study and the resources that we release encourage further work in understanding the pragmatics of explanations.", + "There are many promising research directions for future work in advancing the computational understanding of explanations. First, although /r/ChangeMyView has the useful property that its explanations are closely connected to its explananda, it is important to further investigate the extent to which our findings generalize beyond /r/ChangeMyView and Reddit and establish universal properties of explanations. Second, it is important to connect the words in explanations that we investigate here to the structure of explanations in pyschology BIBREF7. Third, in addition to understanding what goes into an explanation, we need to understand what makes an explanation effective. A better understanding of explanations not only helps develop explainable AI, but also informs the process of collecting explanations that machine learning systems learn from BIBREF48, BIBREF49, BIBREF50." + ], + [ + "We thank Kimberley Buchan, anonymous reviewers, and members of the NLP+CSS research group at CU Boulder for their insightful comments and discussions; Jason Baumgartner for sharing the dataset that enabled this research." + ], + [ + "Before tokenizing, we pass each OP, PC, and explanation through a preprocessing pipeline, with the following steps:", + "Occasionally, /r/ChangeMyView's moderators will edit comments, prefixing their edits with \u201cHello, users of CMV\u201d or \u201cThis is a footnote\u201d (see Table TABREF46). We remove this, and any text that follows on the same line.", + "We replace URLs with a \u201c@url@\u201d token, defining a URL to be any string which matches the following regular expression: (https?://[^\\s)]*).", + "We replace \u201c$\\Delta $\u201d symbols and their analogues\u2014such as \u201c$\\delta $\u201d, \u201c&;#8710;\u201d, and \u201c!delta\u201d\u2014with the word \u201cdelta\u201d. We also remove the word \u201cdelta\u201d from explanations, if the explanation starts with delta.", + "Reddit\u2013specific prefixes, such as \u201cu/\u201d (denoting a user) and \u201cr/\u201d (denoting a subreddit) are removed, as we observed that they often interfered with spaCy's ability to correctly parse its inputs.", + "We remove any text matching the regular expression EDIT(.*?):.* from the beginning of the match to the end of that line, as well as variations, such as Edit(.*?):.*.", + "Reddit allows users to insert blockquoted text. We extract any blockquotes and surround them with standard quotation marks.", + "We replace all contiguous whitespace with a single space. We also do this with tab characters and carriage returns, and with two or more hyphens, asterisks, or underscores.", + "Tokenizing the data. After passing text through our preprocessing pipeline, we use the default spaCy pipeline to extract part-of-speech tags, dependency tags, and entity details for each token BIBREF35. In addition, we use NLTK to stem words BIBREF36. This is used to compute all word level features discussed in Section 4 of the main paper." + ], + [ + "Figure FIGREF49 shows a similar U-shape in the probability of a word being echoed in PC. However, visually, we can see that rare words seem more likely to have high echoing probability in explanations, while that probability is higher for words with moderate frequency in PCs. As PCs tend to be longer than explanations, we also used the echoing probability of the most frequent words to normalize the probability of other words so that they are comparable. We indeed observed a higher likelihood of echoing the rare words, but lower likelihood of echoing words with moderate frequency in explanations than in PCs." + ], + [ + "Given an OP, PC, and explanation, we calculate a 66\u2013dimensional vector for each unique stem in the concatenated OP and PC. Here, we describe the process of calculating each feature.", + "Inverse document frequency: for a stem $s$, the inverse document frequency is given by $\\log \\frac{N}{\\mathrm {df}_s}$, where $N$ is the total number of documents (here, OPs and PCs) in the training set, and $\\mathrm {df}_s$ is the number of documents in the training data whose set of stemmed words contains $s$.", + "Stem length: the number of characters in the stem.", + "Wordnet depth (min): starting with the stem, this is the length of the minimum hypernym path to the synset root.", + "Wordnet depth (max): similarly, this is the length of the maximum hypernym path.", + "Stem transfer probability: the percentage of times in which a stem seen in the explanandum is also seen in the explanation. If, during validation or testing, a stem is encountered for the first time, we set this to be the mean probability of transfer over all stems seen in the training data.", + "OP part\u2013of\u2013speech tags: a stem can represent multiple parts of speech. For example, both \u201ctraditions\u201d and \u201ctraditional\u201d will be stemmed to \u201ctradit.\u201d We count the percentage of times the given stem appears as each part\u2013of\u2013speech tag, following the Universal Dependencies scheme BIBREF53. If the stem does not appear in the OP, each part\u2013of\u2013speech feature will be $\\frac{1}{16}$.", + "OP subject, object, and other: Given a stem $s$, we calculate the percentage of times that $s$'s surface forms in the OP are classified as subjects, objects, or something else by SpaCy. We follow the CLEAR guidelines, BIBREF51 and use the following tags to indicate a subject: nsubj, nsubjpass, csubj, csubjpass, agent, and expl. Objects are identified using these tags: dobj, dative, attr, oprd. If $s$ does not appear at all in the OP, we let subject, object, and other each equal $\\frac{1}{3}$.", + "OP term frequency: the number of times any surface form of a stem appears in the list of tokens that make up the OP.", + "OP normalized term frequency: the percentage of the OP's tokens which are a surface form of the given stem.", + "OP # of surface forms: the number of different surface forms for the given stem.", + "OP location: the average location of each surface form of the given stem which appears in the OP, where the location of a surface form is defined as the percentage of tokens which appear after that surface form. If the stem does not appear at all in the OP, this value is $\\frac{1}{2}$.", + "OP is in quotes: the number of times the stem appears in the OP surrounded by quotation marks.", + "OP is entity: the percentage of tokens in the OP that are both a surface form for the given stem, and are tagged by SpaCy as one of the following entities: PERSON, NORP, FAC, ORG, GPE, LOC, PRODUCT, EVENT, WORK_OF_ART, LAW, and LANGUAGE.", + "PC equivalents of features 6-30.", + "In both OP and PC: 1, if one of the stem's surface forms appears in both the OP and PC. 0 otherwise.", + "# of unique surface forms in OP: for the given stem, the number of surface forms that appear in the OP, but not in the PC.", + "# of unique surface forms in PC: for the given stem, the number of surface forms that appear in the PC, but not in the OP.", + "Stem part\u2013of\u2013speech distribution difference: we consider the concatenation of features 6-21, along with the concatenation of features 31-46, as two distributions, and calculate the Jensen\u2013Shannon divergence between them.", + "Stem dependency distribution difference: similarly, we consider the concatenation of features 22-24 (OP dependency labels), and the concatenation of features 47-49 (PC dependency labels), as two distributions, and calculate the Jensen\u2013Shannon divergence between them.", + "OP length: the number of tokens in the OP.", + "PC length: the number of tokens in the PC.", + "Length difference: the absolute value of the difference between OP length and PC length.", + "Avg. word length difference: the difference between the average number of characters per token in the OP and the average number of characters per token in the PC.", + "OP/PC part\u2013of\u2013speech tag distribution difference: the Jensen\u2013Shannon divergence between the part\u2013of\u2013speech tag distributions of the OP on the one hand, and the PC on the other.", + "Depth of the PC in the thread: since there can be many back\u2013and\u2013forth replies before a user awards a delta, we number each comment in a thread, starting at 0 for the OP, and incrementing for each new comment before the PC appears." + ], + [ + "For each non\u2013LSTM classifier, we train 11 models: one full model, and forward and backward models for each of the five feature groups. To train, we fit on the training set and use the validation set for hyperparameter tuning.", + "For the random model, since the echo rate of the training set is 15%, we simply predict 1 with 15% probability, and 0 otherwise.", + "For logistic regression, we use the lbfgs solver. To tune hyperparameters, we perform an exhaustive grid search, with $C$ taking values from $\\lbrace 10^{x}:x\\in \\lbrace -1, 0, 1, 2, 3, 4\\rbrace \\rbrace $, and the respective weights of the negative and positive classes taking values from $\\lbrace (x, 1-x): x\\in \\lbrace 0.25, 0.20, 0.15\\rbrace \\rbrace $.", + "We also train XGBoost models. Here, we use a learning rate of $0.1$, 1000 estimator trees, and no subsampling. We perform an exhaustive grid search to tune hyperparameters, with the max tree depth equaling 5, 7, or 9, the minimum weight of a child equaling 3, 5, or 7, and the weight of a positive class instance equaling 3, 4, or 5.", + "Finally, we train two LSTM models, each with a single 300\u2013dimensional hidden layer. Due to efficiency considerations, we eschewed a full search of the parameter space, but experimented with different values of dropout, learning rate, positive class weight, and batch size. We ultimately trained each model for five epochs with a batch size of 32 and a learning rate of 0.001, using the Adam optimizer BIBREF52. We also weight positive instances four times more highly than negative instances." + ], + [ + "We formulate an abstractive summarization task using an OP concatenated with the PC as a source, and the explanation as target. We train two models, one with the features described above, and one without. A shared vocabulary of 50k words is constructed from the training set by setting the maximum encoding length to 500 words. We set the maximum decoding length to 100. We use a pointer generator network with coverage for generating explanations, using a bidirectional LSTM as an encoder and a unidirectional LSTM as a decoder. Both use a 256-dimensional hidden state. The parameters of this network are tuned using a validation set of five thousand instances. We constrain the batch size to 16 and train the network for 20k steps, using the parameters described in Table TABREF82." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0666/instruction.md b/qasper-0666/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..85a6ebdcd64b9b69dbfc5b64b8e25c5133f2d888 --- /dev/null +++ b/qasper-0666/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Automating Reading Comprehension by Generating Question and Answer Pairs + +Question: Which datasets are used to train this model? \ No newline at end of file diff --git a/qasper-0667/instruction.md b/qasper-0667/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..1d4fdf6a6c709367553a11f38f975427564682d5 --- /dev/null +++ b/qasper-0667/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Automatic Reminiscence Therapy for Dementia. + +Question: How is performance of this system measured? \ No newline at end of file diff --git a/qasper-0668/instruction.md b/qasper-0668/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c69ad27629a95d0f2025e354f728c8e8f38dd891 --- /dev/null +++ b/qasper-0668/instruction.md @@ -0,0 +1,139 @@ +Name of Paper: Automatic Reminiscence Therapy for Dementia. + +Question: How many questions per image on average are available in dataset? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Methodology", + "Methodology ::: VQG model", + "Methodology ::: Chatbot network", + "Datasets", + "Datasets ::: MS-COCO, Bing and Flickr datasets", + "Datasets ::: Persona-chat and Cornell-movie corpus", + "Validation", + "Validation ::: Implementation", + "Validation ::: Quantitative evaluation", + "Validation ::: Qualitative results", + "Usability study", + "Usability study ::: User interface", + "Feedback from patients", + "Conclusions", + "Acknowledgements" + ], + "paragraphs": [ + [ + "Increases in life expectancy in the last century have resulted in a large number of people living to old ages and will result in a double number of dementia cases by the middle of the century BIBREF0BIBREF1. The most common form of dementia is Alzheimer disease which contributes to 60\u201370% of cases BIBREF2. Research focused on identifying treatments to slow down the evolution of Alzheimer's disease is a very active pursuit, but it has been only successful in terms of developing therapies that eases the symptoms without addressing the cause BIBREF3BIBREF4. Besides, people with dementia might have some barriers to access to the therapies, such as cost, availability and displacement to the care home or hospital, where the therapy takes place. We believe that Artificial Intelligence (AI) can contribute in innovative systems to give accessibility and offer new solutions to the patients needs, as well as help relatives and caregivers to understand the illness of their family member or patient and monitor the progress of the dementia.", + "Therapies such as reminiscence, that stimulate memories of the patient's past, have well documented benefits on social, mental and emotional well-being BIBREF5BIBREF6, making them a very desirable practice, especially for older adults. Reminiscence therapy in particular involves the discussion of events and past experiences using tangible prompts such as pictures or music to evoke memories and stimulate conversation BIBREF7. With this aim, we explore multi-modal deep learning architectures to be used to develop an intuitive, easy to use, and robust dialogue system to automatize the reminiscence therapy for people affected by mild cognitive impairment or at early stages of Alzheimer's disease.", + "We propose a conversational agent that simulates a reminiscence therapist by asking questions about the patient's experiences. Questions are generated from pictures provided by the patient, which contain significant moments or important people in user's life. Moreover, to engage the user in the conversation we propose a second model which generates comments on user's answers. A chatbot model trained with a dataset containing simple conversations between different people. The activity pretends to be challenging for the patient, as the questions may require the user to exercise the memory. Our contributions include:", + "Automation of the Reminiscence therapy by using a multi-modal approach that generates questions from pictures, without using a reminiscence therapy dataset.", + "An end-to-end deep learning approach which do not require hand-crafted rules and it is ready to be used by mild cognitive impairment patients. The system is designed to be intuitive and easy to use for the users and could be reached by any smartphone with internet connection." + ], + [ + "The origin of chatbots goes back to 1966 with the creation of ELIZA BIBREF8 by Joseph Weizenbaum at MIT. Its implementation consisted in pattern matching and substitution methodology. Recently, data driven approaches have drawn significant attention. Existing work along this line includes retrieval-based methods BIBREF9BIBREF10 and generation-based methodsBIBREF11BIBREF12. In this work we focus on generative models, where sequence-to-sequence algorithm that uses RNNs to encode and decode inputs into responses is a current best practice.", + "Our conversational agent uses two architectures to simulate a specialized reminiscence therapist. The block in charge of generating questions is based on the work Show, Attend and Tell BIBREF13. This work generates descriptions from pictures, also known as image captioning. In our case, we focus on generating questions from pictures. Our second architecture is inspired by Neural Conversational Model from BIBREF14 where the author presents an end-to-end approach to generate simple conversations. Building an open-domain conversational agent is a challenging problem. As addressed in BIBREF15 and BIBREF16, the lack of a consistent personality and lack of long-term memory which produces some meaningless responses in these models are still unresolved problems.", + "Some works have proposed conversational agents for older adults with a variety of uses, such as stimulate conversation BIBREF17 , palliative care BIBREF18 or daily assistance. An example of them is \u2018Billie\u2019 reported in BIBREF19 which is a virtual agent that uses facial expression for a more natural behavior and is focused on managing user\u2019s calendar, or \u2018Mary\u2019 BIBREF20 that assists the users by organizing their tasks offering reminders and guidance with household activities. Both of the works perform well on its specific tasks, but report difficulties to maintain a casual conversation. Other works focus on the content used in Reminiscence therapy. Like BIBREF21 where the authors propose a system that recommends multimedia content to be used in therapy, or Visual Dialog BIBREF22 where the conversational agent is the one that has to answer the questions about the image." + ], + [ + "In this section we explain the main two components of our model, as well as how the interaction with the model works. We named it Elisabot and its goal is to mantain a dialog with the patient about her user\u2019s life experiences.", + "Before starting the conversation, the user must introduce photos that should contain significant moments for him/her. The system randomly chooses one of these pictures and analyses the content. Then, Elisabot shows the selected picture and starts the conversation by asking a question about the picture. The user should give an answer, even though he does not know it, and Elisabot makes a relevant comment on it. The cycle starts again by asking another relevant question about the image and the flow is repeated for 4 to 6 times until the picture is changed. The Figure FIGREF3 summarizes the workflow of our system.", + "Elisabot is composed of two models: the model in charge of asking questions about the image which we will refer to it as VQG model, and the Chatbot model which tries to make the dialogue more engaging by giving feedback to the user's answers." + ], + [ + "The algorithm behind VQG consists in an Encoder-Decoder architecture with attention. The Encoder takes as input one of the given photos $I$ from the user and learns its information using a CNN. CNNs have been widely studied for computer vision tasks. The CNN provides the image's learned features to the Decoder which generates the question $y$ word by word by using an attention mechanism with a Long Short-Term Memory (LSTM). The model is trained to maximize the likelihood $p(y|I)$ of producing a target sequence of words:", + "where $K$ is the size of the vocabulary and $C$ is the length of the caption.", + "Since there are already Convolutional Neural Networks (CNNs) trained on large datasets to represent images with an outstanding performance, we make use of transfer learning to integrate a pre-trained model into our algorithm. In particular, we use a ResNet-101 BIBREF23 model trained on ImageNet. We discard the last 2 layers, since these layers classify the image into categories and we only need to extract its features." + ], + [ + "The core of our chatbot model is a sequence-to-sequence BIBREF24. This architecture uses a Recurrent Neural Network (RNN) to encode a variable-length sequence to obtain a large fixed dimensional vector representation and another RNN to decode the vector into a variable-length sequence.", + "The encoder iterates through the input sentence one word at each time step producing an output vector and a hidden state vector. The hidden state vector is passed to the next time step, while the output vector is stored. We use a bidirectional Gated Recurrent Unit (GRU), meaning we use two GRUs one fed in sequential order and another one fed in reverse order. The outputs of both networks are summed at each time step, so we encode past and future context.", + "The final hidden state $h_t^{enc}$ is fed into the decoder as the initial state $h_0^{dec}$. By using an attention mechanism, the decoder uses the encoder\u2019s context vectors, and internal hidden states to generate the next word in the sequence. It continues generating words until it outputs an $<$end$>$ token, representing the end of the sentence. We use an attention layer to multiply attention weights to encoder's outputs to focus on the relevant information when decoding the sequence. This approach have shown better performance on sequence-to-sequence models BIBREF25." + ], + [ + "One of the first requirements to develop an architecture using a machine learning approach is a training dataset. The lack of open-source datasets containing dialogues from reminiscence therapy lead as to use a dataset with content similar to the one used in the therapy. In particular, we use two types of datasets to train our models: A dataset that maps pictures with questions, and an open-domain conversation dataset. The details of the two datasets are as follows." + ], + [ + "We use MS COCO, Bing and Flickr datasets from BIBREF26 to train the model that generates questions. These datasets contain natural questions about images with the purpose of knowing more about the picture. As can be seen in the Figure FIGREF8, questions cannot be answered by only looking at the image. Each source contains 5,000 images with 5 questions per image, adding a total of 15,000 images with 75,000 questions. COCO dataset includes images of complex everyday scenes containing common objects in their natural context, but it is limited in terms of the concepts it covers. Bing dataset contains more event related questions and has a wider range of questions longitudes (between 3 and 20 words), while Flickr questions are shorter (less than 6 words) and the images appear to be more casual." + ], + [ + "We use two datasets to train our chatbot model. The first one is the Persona-chat BIBREF15 which contains dialogues between two people with different profiles that are trying to know each other. It is complemented by the Cornell-movie dialogues dataset BIBREF27, which contains a collection of fictional conversations extracted from raw movie scripts. Persona-chat's sentences have a maximum of 15 words, making it easier to learn for machines and a total of 162,064 utterances over 10,907 dialogues. While Cornell-movie dataset contains 304,713 utterances over 220,579 conversational exchanges between 10,292 pairs of movie characters." + ], + [ + "An important aspect of dialogue response generation systems is how to evaluate the quality of the generated response. This section presents the training procedure and the quantitative evaluation of the model, together with some qualitative results." + ], + [ + "Both models are trained using Stochastic Gradient Descent with ADAM optimization BIBREF28 and a learning rate of 1e-4. Besides, we use dropout regularization BIBREF29 which prevents from over-fitting by dropping some units of the network.", + "The VQG encoder is composed of 2048 neuron cells, while the VQG decoder has an attention layer of 512 followed by an embedding layer of 512 and a LSTM with the same size. We use a dropout of 50% and a beam search of 7 for decoding, which let as obtain up to 5 output questions. The vocabulary we use consists of all words seen 3 or more times in the training set, which amounts to 11.214 unique tokens. Unknown words are mapped to an $<$unk$>$ token during training, but we do not allow the decoder to produce this token at test time. We also set a maximum sequence length of 6 words as we want simple questions easy to understand and easy to learn by the model.", + "In the Chatbot model we use a hidden size of 500 and Dropout regularization of 25%. For decoding we use greedy search, which consists in making the optimal token choice at each step. We first train it with Persona-chat and then fine-tune it with Cornell dataset. The vocabulary we use consists of all words seen 3 or more times in Persona-chat dataset and we set a maximum sequence length of 12 words. For the hyperparameter setting, we use a batch size of 64." + ], + [ + "We use the BLEU BIBREF30 metric on the validation set for the VQG model training. BLEU is a measure of similitude between generated and target sequences of words, widely used in natural language processing. It assumes that valid generated responses have significant word overlap with the ground truth responses. We use it because in this case we have five different references for each of the generated questions. We obtain a BLEU score of 2.07.", + "Our chatbot model instead, only have one reference ground truth in training when generating a sequence of words. We considered that it was not a good metric to apply as in some occasions responses have the same meaning, but do not share any words in common. Thus, we save several models with different hyperparameters and at different number of training iterations and compare them using human evaluation, to chose the model that performs better in a conversation." + ], + [ + "Our first goal was to generate meaningful questions from the provided pictures. Table TABREF16 contains some generated questions by our VQG model for each of the proposed images. These images were not in the training set as they were taken from photo albums, so the model was the first time that had these images as input. By looking at the questions, we can easily realise that the model is not only able to generate questions grammatically correct, but to understand the content of the image and ask about it. Furthermore, we not only wanted to generate only a question per image, but to bring out up to five questions for a deeper conversation.", + "We used different kind of pictures as we wanted a robust model able to work with all kind of images. In the Table TABREF16 we can see both, old gray and colorful photos, both with vertical and horizontal shapes. Most of the pictures we selected for testing were showing people, because these are the type of pictures used in reminiscence therapy, as they evoke memories from the relationships with that persons. We also tested pictures showing pets, cars and motorbikes as some people are very attached to them, triggering strong emotions.", + "Most of the generated questions could be asked by a person. Nonetheless, we found some of the questions such as How did the car get damaged? that do not correspond with the image's content, or are not well formulated like What body of dog is that?, but can be easily understood.", + "About the chatbot module, we test it manually by having a conversation with it. In the following lines, we can see two samples of a conversation, where each sentence starting with \"$>$\" is an input to our seq2seq model. In this first sample, the chatbot model is clearly overfitted as it repeats the same sentence twice. While in the second sample, we can see that the model is able to follow the conversation.", + "Overfitted chatbot model:", + "$>$ hi", + "Elisabot: hello ! how are you ?", + "$>$ fine and you?", + "Elisabot : i am good do you have any hobbies ?", + "$>$ yes i like to ski", + "Elisabot: i love to play sports . it is my favorite", + "$>$ I have a sister", + "Elisabot : i have a dog . i love them .", + "$>$ I have a dog too", + "Elisabot: i have a dog . i love them", + "$>$ ok", + "Elisabot : i am not really a good person", + "$>$ why?", + "Elisabot : i love to play sports . it is my favorite", + "Final chatbot model:", + "$>$ hello", + "Elisabot: hello how are you ?", + "$>$ fine and you?", + "Elisabot : i m great !", + "$>$ I have a sister, do you have any brother or sister?", + "Elisabot : i have a iphone but i have no idea .", + "$>$ ok", + "Elisabot : i am a pro fan of thrones", + "$>$ me too", + "Elisabot : i like to eat healthy", + "$>$ have you seen the last chapter?", + "Elisabot : no i have not .", + "$>$ where do you live?", + "Elisabot : i live in the mountains" + ], + [ + "As most of the metrics correlate very weakly with human judgements in the non-technical domain BIBREF31 we decide to evaluate our system with a simple user study with two patients. We present the user interface built and the feedback obtained from the patients." + ], + [ + "We developed a user interface for Elisabot with Telegram, an instant messaging application available for smartphones or computers. We select it because it is easy to use and it offers an API for developers to connect bots to the Telegram system. It enables to create special accounts for bots which do not require a phone number to set up.", + "Telegram is only the interface for the code running in the server. The bot is executed via an HTTP-request to the API. Users can start a conversation with Elisabot by typing @TherapistElisabot in the searcher and executing the command /start, as can be seen in the Figure FIGREF31. Messages, commands and requests sent by users are passed to the software running on the server. We add /change, /yes and /exit commands to enable more functionalities. /Change gives the opportunity to the user to change the image in case the user does not want to talk about it, /yes accepts the image which is going to talk about and /exit finishes the dialogue with Elisabot. The commands can be executed either by tapping on the linked text or typing them." + ], + [ + "We designed a usability study where users with and without mild cognitive impairment interacted with the system with the help of a doctor and one of the authors. The purpose was to study the acceptability and feasibility of the system with patients of mild cognitive impairment. The users were all older than 60 years old. The sessions lasted 30 minutes and were carried out by using a laptop computer connected to Telegram. As Elisabot's language is English we translated the questions to the users and the answers to Elisabot.", + "Figure FIGREF38 is a sample of the session we did with mild cognitive impairment patients from anonymized institution and location. The picture provided by the patient (Figure FIGREF37 is blurred for user's privacy rights. In this experiment all the generated questions were right according to the image content, but the feedback was wrong for some of the answers. We can see that it was the last picture of the session as when Elisabot asks if the user wants to continue or leave, and he decides to continue, Elisabot finishes the session as there are no more pictures remaining to talk about.", + "At the end of the session, we administrated a survey to ask participants the following questions about their assessment of Elisabot:", + "Did you like it?", + "Did you find it engaging?", + "How difficult have you found it?", + "Responses were given on a five-point scale ranging from strongly disagree (1) to strongly agree (5) and very easy (1) to very difficult (5). The results were 4.6 for amusing and engaging and 2.6 for difficulty. Healthy users found it very easy to use (1/5) and even a bit silly, because of some of the generated questions and comments. Nevertheless, users with mild cognitive impairment found it engaging (5/5) and challenging (4/5), because of the effort they had to make to remember the answers for some of the generated questions. All the users had in common that they enjoyed doing the therapy with Elisabot." + ], + [ + "We presented a dialogue system for handling sessions of 30 minutes of reminiscence therapy. Elisabot, our conversational agent leads the therapy by showing a picture and generating some questions. The goal of the system is to improve users mood and stimulate their memory and communication skills. Two models were proposed to generate the dialogue system for the reminiscence therapy. A visual question generator composed of a CNN and a LSTM with attention and a sequence-to-sequence model to generate feedback on the user's answers. We realize that fine-tuning our chatbot model with another dataset improved the generated dialogue.", + "The manual evaluation shows that our model can generate questions and feedback well formulated grammatically, but in some occasions not appropriate in content. As expected, it has tendency to produce non-specific answers and to loss its consistency in the comments with respect to what it has said before. However, the overall usability evaluation of the system by users with mild cognitive impairment shows that they found the session very entertaining and challenging. They had to make an effort to remember the answers for some of the questions, but they were very satisfied when they achieved it. Though, we see that for the proper performance of the therapy is essential a person to support the user to help remember the experiences that are being asked.", + "This project has many possible future lines. In our future work, we suggest to train the model including the Reddit dataset which could improve the chatbot model, as it has many open-domain conversations. Moreover, we would like to include speech recognition and generation, as well as real-time text translation, to make Elisabot more autonomous and open to older adults with reading and writing difficulties. Furthermore, the lack of consistency in the dialogue might be avoided by improving the architecture including information about passed conversation into the model. We also think it would be a good idea to recognize feelings from the user's answers and give a feedback according to them." + ], + [ + "Marioan Caros was funded with a scholarship from the Fundacion Vodafona Spain. Petia Radeva was partially funded by TIN2018-095232-B-C21, 2017 SGR 1742, Nestore, Validithi, and CERCA Programme/Generalitat de Catalunya. We acknowledge the support of NVIDIA Corporation with the donation of Titan Xp GPUs." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0675/instruction.md b/qasper-0675/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..e3182539d7d550943b81e3c2bfb364fded9f6963 --- /dev/null +++ b/qasper-0675/instruction.md @@ -0,0 +1,66 @@ +Name of Paper: Speaker-independent classification of phonetic segments from raw ultrasound in child speech + +Question: Do they report results only on English data? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Ultrasound Tongue Imaging", + "Related Work", + "Ultrasound Data", + "Data Selection", + "Preprocessing and Model Architectures", + "Training Scenarios and Speaker Means", + "Results and Discussion", + "Future Work", + "Conclusion" + ], + "paragraphs": [ + [ + "Ultrasound tongue imaging (UTI) uses standard medical ultrasound to visualize the tongue surface during speech production. It provides a non-invasive, clinically safe, and increasingly inexpensive method to visualize the vocal tract. Articulatory visual biofeedback of the speech production process, using UTI, can be valuable for speech therapy BIBREF0 , BIBREF1 , BIBREF2 or language learning BIBREF3 , BIBREF4 . Ultrasound visual biofeedback combines auditory information with visual information of the tongue position, allowing users, for example, to correct inaccurate articulations in real-time during therapy or learning. In the context of speech therapy, automatic processing of ultrasound images was used for tongue contour extraction BIBREF5 and the animation of a tongue model BIBREF6 . More broadly, speech recognition and synthesis from articulatory signals BIBREF7 captured using UTI can be used with silent speech interfaces in order to help restore spoken communication for users with speech or motor impairments, or to allow silent spoken communication in situations where audible speech is undesirable BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 . Similarly, ultrasound images of the tongue have been used for direct estimation of acoustic parameters for speech synthesis BIBREF13 , BIBREF14 , BIBREF15 .", + "Speech and language therapists (SLTs) have found UTI to be very useful in speech therapy. In this work we explore the automatic processing of ultrasound tongue images in order to assist SLTs, who currently largely rely on manual processing when using articulatory imaging in speech therapy. One task that could assist SLTs is the automatic classification of tongue shapes from raw ultrasound. This can facilitate the diagnosis and treatment of speech sound disorders, by allowing SLTs to automatically identify incorrect articulations, or by quantifying patient progress in therapy. In addition to being directly useful for speech therapy, the classification of tongue shapes enables further understanding of phonetic variability in ultrasound tongue images. Much of the previous work in this area has focused on speaker-dependent models. In this work we investigate how automatic processing of ultrasound tongue imaging is affected by speaker variation, and how severe degradations in performance can be avoided when applying systems to data from previously unseen speakers through the use of speaker adaptation and speaker normalization approaches.", + "Below, we present the main challenges associated with the automatic processing of ultrasound data, together with a review of speaker-independent models applied to UTI. Following this, we present the experiments that we have performed (Section SECREF2 ), and discuss the results obtained (Section SECREF3 ). Finally we propose some future work and conclude the paper (Sections SECREF4 and SECREF5 )." + ], + [ + "There are several challenges associated with the automatic processing of ultrasound tongue images.", + "Image quality and limitations. UTI output tends to be noisy, with unrelated high-contrast edges, speckle noise, or interruptions of the tongue surface BIBREF16 , BIBREF17 . Additionally, the oral cavity is not entirely visible from the image, missing the lips, the palate, or the pharyngeal wall.", + "Inter-speaker variation. Age and physiology may affect the output, with children imaging better than adults due to more moisture in the mouth and less tissue fat BIBREF16 . However, dry mouths lead to poor imaging, which might occur in speech therapy if a child is nervous during a session. Similarly, the vocal tracts of children across different ages may be more variable than those of adults.", + "Probe placement. Articulators that are orthogonal to the ultrasound beam direction image well, while those at an angle tend to image poorly. Incorrect or variable probe placement during recordings may lead to high variability between otherwise similar tongue shapes. This may be controlled using helmets BIBREF18 , although it is unreasonable to expect the speaker to remain still throughout the recording session, especially if working with children. Therefore, probe displacement should be expected to be a factor in image quality and consistency.", + "Limited data. Although ultrasound imaging is becoming less expensive to acquire, there is still a lack of large publicly available databases to evaluate automatic processing methods. The UltraSuite Repository BIBREF19 , which we use in this work, helps alleviate this issue, but it still does not compare to standard speech recognition or image classification databases, which contain hundreds of hours of speech or millions of images." + ], + [ + "Earlier work concerned with speech recognition from ultrasound data has mostly been focused on speaker-dependent systems BIBREF20 , BIBREF21 , BIBREF22 , BIBREF23 . An exception is the work of Xu et al. BIBREF24 , which investigates the classification of tongue gestures from ultrasound data using convolutional neural networks. Some results are presented for a speaker-independent system, although the investigation is limited to two speakers generalizing to a third. Fabre et al BIBREF5 present a method for automatic tongue contour extraction from ultrasound data. The system is evaluated in a speaker-independent way by training on data from eight speakers and evaluating on a single held out speaker. In both of these studies, a large drop in accuracy was observed when using speaker-independent systems in comparison to speaker-dependent systems. Our investigation differs from previous work in that we focus on child speech while using a larger number of speakers (58 children). Additionally, we use cross-validation to evaluate the performance of speaker-independent systems across all speakers, rather than using a small held out subset." + ], + [ + "We use the Ultrax Typically Developing dataset (UXTD) from the publicly available UltraSuite repository BIBREF19 . This dataset contains synchronized acoustic and ultrasound data from 58 typically developing children, aged 5-12 years old (31 female, 27 male). The data was aligned at the phone-level, according to the methods described in BIBREF19 , BIBREF25 . For this work, we discarded the acoustic data and focused only on the B-Mode ultrasound images capturing a midsaggital view of the tongue. The data was recorded using an Ultrasonix SonixRP machine using Articulate Assistant Advanced (AAA) software at INLINEFORM0 121fps with a 135 field of view. A single ultrasound frame consists of 412 echo returns from each of the 63 scan lines (63x412 raw frames). For this work, we only use UXTD type A (semantically unrelated words, such as pack, tap, peak, tea, oak, toe) and type B (non-words designed to elicit the articulation of target phones, such as apa, eepee, opo) utterances." + ], + [ + "For this investigation, we define a simplified phonetic segment classification task. We determine four classes corresponding to distinct places of articulation. The first consists of bilabial and labiodental phones (e.g. /p, b, v, f, .../). The second class includes dental, alveolar, and postalveolar phones (e.g. /th, d, t, z, s, sh, .../). The third class consists of velar phones (e.g. /k, g, .../). Finally, the fourth class consists of alveolar approximant /r/. Figure FIGREF1 shows examples of the four classes for two speakers.", + "For each speaker, we divide all available utterances into disjoint train, development, and test sets. Using the force-aligned phone boundaries, we extract the mid-phone frame for each example across the four classes, which leads to a data imbalance. Therefore, for all utterances in the training set, we randomly sample additional examples within a window of 5 frames around the center phone, to at least 50 training examples per class per speaker. It is not always possible to reach the target of 50 examples, however, if no more data is available to sample from. This process gives a total of INLINEFORM0 10700 training examples with roughly 2000 to 3000 examples per class, with each speaker having an average of 185 examples. Because the amount of data varies per speaker, we compute a sampling score, which denotes the proportion of sampled examples to the speaker's total training examples. We expect speakers with high sampling scores (less unique data overall) to underperform when compared with speakers with more varied training examples." + ], + [ + "For each system, we normalize the training data to zero mean and unit variance. Due to the high dimensionality of the data (63x412 samples per frame), we have opted to investigate two preprocessing techniques: principal components analysis (PCA, often called eigentongues in this context) and a 2-dimensional discrete cosine transform (DCT). In this paper, Raw input denotes the mean-variance normalized raw ultrasound frame. PCA applies principal components analysis to the normalized training data and preserves the top 1000 components. DCT applies the 2D DCT to the normalized raw ultrasound frame and the upper left 40x40 submatrix (1600 coefficients) is flattened and used as input.", + "The first type of classifier we evaluate in this work are feedforward neural networks (DNNs) consisting of 3 hidden layers, each with 512 rectified linear units (ReLUs) with a softmax activation function. The networks are optimized for 40 epochs with a mini-batch of 32 samples using stochastic gradient descent. Based on preliminary experiments on the validation set, hyperparameters such learning rate, decay rate, and L2 weight vary depending on the input format (Raw, PCA, or DCT). Generally, Raw inputs work better with smaller learning rates and heavier regularization to prevent overfitting to the high-dimensional data. As a second classifier to evaluate, we use convolutional neural networks (CNNs) with 2 convolutional and max pooling layers, followed by 2 fully-connected ReLU layers with 512 nodes. The convolutional layers use 16 filters, 8x8 and 4x4 kernels respectively, and rectified units. The fully-connected layers use dropout with a drop probability of 0.2. Because CNN systems take longer to converge, they are optimized over 200 epochs. For all systems, at the end of every epoch, the model is evaluated on the development set, and the best model across all epochs is kept." + ], + [ + "We train speaker-dependent systems separately for each speaker, using all of their training data (an average of 185 examples per speaker). These systems use less data overall than the remaining systems, although we still expect them to perform well, as the data matches in terms of speaker characteristics. Realistically, such systems would not be viable, as it would be unreasonable to collect large amounts of data for every child who is undergoing speech therapy. We further evaluate all trained systems in a multi-speaker scenario. In this configuration, the speaker sets for training, development, and testing are equal. That is, we evaluate on speakers that we have seen at training time, although on different utterances. A more realistic configuration is a speaker-independent scenario, which assumes that the speaker set available for training and development is disjoint from the speaker set used at test time. This scenario is implemented by leave-one-out cross-validation. Finally, we investigate a speaker adaptation scenario, where training data for the target speaker becomes available. This scenario is realistic, for example, if after a session, the therapist were to annotate a small number of training examples. In this work, we use the held-out training data to finetune a pretrained speaker-independent system for an additional 6 epochs in the DNN systems and 20 epochs for the CNN systems. We use all available training data across all training scenarios, and we investigate the effect of the number of samples on one of the top performing systems.", + "This work is primarily concerned with generalizing to unseen speakers. Therefore, we investigate a method to provide models with speaker-specific inputs. A simple approach is to use the speaker mean, which is the pixel-wise mean of all raw frames associated with a given speaker, illustrated in Figure FIGREF8 . The mean frame might capture an overall area of tongue activity, average out noise, and compensate for probe placement differences across speakers. Speaker means are computed after mean variance normalization. For PCA-based systems, matrix decomposition is applied on the matrix of speaker means for the training data with 50 components being kept, while the 2D DCT is applied normally to each mean frame. In the DNN systems, the speaker mean is appended to the input vector. In the CNN system, the raw speaker mean is given to the network as a second channel. All model configurations are similar to those described earlier, except for the DNN using Raw input. Earlier experiments have shown that a larger number of parameters are needed for good generalization with a large number of inputs, so we use layers of 1024 nodes rather than 512." + ], + [ + "Results for all systems are presented in Table TABREF10 . When comparing preprocessing methods, we observe that PCA underperforms when compared with the 2 dimensional DCT or with the raw input. DCT-based systems achieve good results when compared with similar model architectures, especially when using smaller amounts of data as in the speaker-dependent scenario. When compared with raw input DNNs, the DCT-based systems likely benefit from the reduced dimensionality. In this case, lower dimensional inputs allow the model to generalize better and the truncation of the DCT matrix helps remove noise from the images. Compared with PCA-based systems, it is hypothesized the observed improvements are likely due to the DCT's ability to encode the 2-D structure of the image, which is ignored by PCA. However, the DNN-DCT system does not outperform a CNN with raw input, ranking last across adapted systems.", + "When comparing training scenarios, as expected, speaker-independent systems underperform, which illustrates the difficulty involved in the generalization to unseen speakers. Multi-speaker systems outperform the corresponding speaker-dependent systems, which shows the usefulness of learning from a larger database, even if variable across speakers. Adapted systems improve over the dependent systems, except when using DCT. It is unclear why DCT-based systems underperform when adapting pre-trained models. Figure FIGREF11 shows the effect of the size of the adaptation data when finetuning a pre-trained speaker-independent system. As expected, the more data is available, the better that system performs. It is observed that, for the CNN system, with roughly 50 samples, the model outperforms a similar speaker-dependent system with roughly three times more examples.", + "Speaker means improve results across all scenarios. It is particularly useful for speaker-independent systems. The ability to generalize to unseen speakers is clear in the CNN system. Using the mean as a second channel in the convolutional network has the advantage of relating each pixel to its corresponding speaker mean value, allowing the model to better generalize to unseen speakers.", + "Figure FIGREF12 shows pair-wise scatterplots for the CNN system. Training scenarios are compared in terms of the effect on individual speakers. It is observed, for example, that the performance of a speaker-adapted system is similar to a multi-speaker system, with most speakers clustered around the identity line (bottom left subplot). Figure FIGREF12 also illustrates the variability across speakers for each of the training scenarios. The classification task is easier for some speakers than others. In an attempt to understand this variability, we can look at correlation between accuracy scores and various speaker details. For the CNN systems, we have found some correlation (Pearson's product-moment correlation) between accuracy and age for the dependent ( INLINEFORM0 ), multi-speaker ( INLINEFORM1 ), and adapted ( INLINEFORM2 ) systems. A very small correlation ( INLINEFORM3 ) was found for the independent system. Similarly, some correlation was found between accuracy and sampling score ( INLINEFORM4 ) for the dependent system, but not for the remaining scenarios. No correlation was found between accuracy and gender (point biserial correlation)." + ], + [ + "There are various possible extensions for this work. For example, using all frames assigned to a phone, rather than using only the middle frame. Recurrent architectures are natural candidates for such systems. Additionally, if using these techniques for speech therapy, the audio signal will be available. An extension of these analyses should not be limited to the ultrasound signal, but instead evaluate whether audio and ultrasound can be complementary. Further work should aim to extend the four classes to more a fine-grained place of articulation, possibly based on phonological processes. Similarly, investigating which classes lead to classification errors might help explain some of the observed results. Although we have looked at variables such as age, gender, or amount of data to explain speaker variation, there may be additional factors involved, such as the general quality of the ultrasound image. Image quality could be affected by probe placement, dry mouths, or other factors. Automatically identifying or measuring such cases could be beneficial for speech therapy, for example, by signalling the therapist that the data being collected is sub-optimal." + ], + [ + "In this paper, we have investigated speaker-independent models for the classification of phonetic segments from raw ultrasound data. We have shown that the performance of the models heavily degrades when evaluated on data from unseen speakers. This is a result of the variability in ultrasound images, mostly due to differences across speakers, but also due to shifts in probe placement. Using the mean of all ultrasound frames for a new speaker improves the generalization of the models to unseen data, especially when using convolutional neural networks. We have also shown that adapting a pre-trained speaker-independent system using as few as 50 ultrasound frames can outperform a corresponding speaker-dependent system." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0681/instruction.md b/qasper-0681/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c7e701110059256c453180dff5ed83f725978d02 --- /dev/null +++ b/qasper-0681/instruction.md @@ -0,0 +1,66 @@ +Name of Paper: Speaker-independent classification of phonetic segments from raw ultrasound in child speech + +Question: What model do they use to classify phonetic segments? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Ultrasound Tongue Imaging", + "Related Work", + "Ultrasound Data", + "Data Selection", + "Preprocessing and Model Architectures", + "Training Scenarios and Speaker Means", + "Results and Discussion", + "Future Work", + "Conclusion" + ], + "paragraphs": [ + [ + "Ultrasound tongue imaging (UTI) uses standard medical ultrasound to visualize the tongue surface during speech production. It provides a non-invasive, clinically safe, and increasingly inexpensive method to visualize the vocal tract. Articulatory visual biofeedback of the speech production process, using UTI, can be valuable for speech therapy BIBREF0 , BIBREF1 , BIBREF2 or language learning BIBREF3 , BIBREF4 . Ultrasound visual biofeedback combines auditory information with visual information of the tongue position, allowing users, for example, to correct inaccurate articulations in real-time during therapy or learning. In the context of speech therapy, automatic processing of ultrasound images was used for tongue contour extraction BIBREF5 and the animation of a tongue model BIBREF6 . More broadly, speech recognition and synthesis from articulatory signals BIBREF7 captured using UTI can be used with silent speech interfaces in order to help restore spoken communication for users with speech or motor impairments, or to allow silent spoken communication in situations where audible speech is undesirable BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 . Similarly, ultrasound images of the tongue have been used for direct estimation of acoustic parameters for speech synthesis BIBREF13 , BIBREF14 , BIBREF15 .", + "Speech and language therapists (SLTs) have found UTI to be very useful in speech therapy. In this work we explore the automatic processing of ultrasound tongue images in order to assist SLTs, who currently largely rely on manual processing when using articulatory imaging in speech therapy. One task that could assist SLTs is the automatic classification of tongue shapes from raw ultrasound. This can facilitate the diagnosis and treatment of speech sound disorders, by allowing SLTs to automatically identify incorrect articulations, or by quantifying patient progress in therapy. In addition to being directly useful for speech therapy, the classification of tongue shapes enables further understanding of phonetic variability in ultrasound tongue images. Much of the previous work in this area has focused on speaker-dependent models. In this work we investigate how automatic processing of ultrasound tongue imaging is affected by speaker variation, and how severe degradations in performance can be avoided when applying systems to data from previously unseen speakers through the use of speaker adaptation and speaker normalization approaches.", + "Below, we present the main challenges associated with the automatic processing of ultrasound data, together with a review of speaker-independent models applied to UTI. Following this, we present the experiments that we have performed (Section SECREF2 ), and discuss the results obtained (Section SECREF3 ). Finally we propose some future work and conclude the paper (Sections SECREF4 and SECREF5 )." + ], + [ + "There are several challenges associated with the automatic processing of ultrasound tongue images.", + "Image quality and limitations. UTI output tends to be noisy, with unrelated high-contrast edges, speckle noise, or interruptions of the tongue surface BIBREF16 , BIBREF17 . Additionally, the oral cavity is not entirely visible from the image, missing the lips, the palate, or the pharyngeal wall.", + "Inter-speaker variation. Age and physiology may affect the output, with children imaging better than adults due to more moisture in the mouth and less tissue fat BIBREF16 . However, dry mouths lead to poor imaging, which might occur in speech therapy if a child is nervous during a session. Similarly, the vocal tracts of children across different ages may be more variable than those of adults.", + "Probe placement. Articulators that are orthogonal to the ultrasound beam direction image well, while those at an angle tend to image poorly. Incorrect or variable probe placement during recordings may lead to high variability between otherwise similar tongue shapes. This may be controlled using helmets BIBREF18 , although it is unreasonable to expect the speaker to remain still throughout the recording session, especially if working with children. Therefore, probe displacement should be expected to be a factor in image quality and consistency.", + "Limited data. Although ultrasound imaging is becoming less expensive to acquire, there is still a lack of large publicly available databases to evaluate automatic processing methods. The UltraSuite Repository BIBREF19 , which we use in this work, helps alleviate this issue, but it still does not compare to standard speech recognition or image classification databases, which contain hundreds of hours of speech or millions of images." + ], + [ + "Earlier work concerned with speech recognition from ultrasound data has mostly been focused on speaker-dependent systems BIBREF20 , BIBREF21 , BIBREF22 , BIBREF23 . An exception is the work of Xu et al. BIBREF24 , which investigates the classification of tongue gestures from ultrasound data using convolutional neural networks. Some results are presented for a speaker-independent system, although the investigation is limited to two speakers generalizing to a third. Fabre et al BIBREF5 present a method for automatic tongue contour extraction from ultrasound data. The system is evaluated in a speaker-independent way by training on data from eight speakers and evaluating on a single held out speaker. In both of these studies, a large drop in accuracy was observed when using speaker-independent systems in comparison to speaker-dependent systems. Our investigation differs from previous work in that we focus on child speech while using a larger number of speakers (58 children). Additionally, we use cross-validation to evaluate the performance of speaker-independent systems across all speakers, rather than using a small held out subset." + ], + [ + "We use the Ultrax Typically Developing dataset (UXTD) from the publicly available UltraSuite repository BIBREF19 . This dataset contains synchronized acoustic and ultrasound data from 58 typically developing children, aged 5-12 years old (31 female, 27 male). The data was aligned at the phone-level, according to the methods described in BIBREF19 , BIBREF25 . For this work, we discarded the acoustic data and focused only on the B-Mode ultrasound images capturing a midsaggital view of the tongue. The data was recorded using an Ultrasonix SonixRP machine using Articulate Assistant Advanced (AAA) software at INLINEFORM0 121fps with a 135 field of view. A single ultrasound frame consists of 412 echo returns from each of the 63 scan lines (63x412 raw frames). For this work, we only use UXTD type A (semantically unrelated words, such as pack, tap, peak, tea, oak, toe) and type B (non-words designed to elicit the articulation of target phones, such as apa, eepee, opo) utterances." + ], + [ + "For this investigation, we define a simplified phonetic segment classification task. We determine four classes corresponding to distinct places of articulation. The first consists of bilabial and labiodental phones (e.g. /p, b, v, f, .../). The second class includes dental, alveolar, and postalveolar phones (e.g. /th, d, t, z, s, sh, .../). The third class consists of velar phones (e.g. /k, g, .../). Finally, the fourth class consists of alveolar approximant /r/. Figure FIGREF1 shows examples of the four classes for two speakers.", + "For each speaker, we divide all available utterances into disjoint train, development, and test sets. Using the force-aligned phone boundaries, we extract the mid-phone frame for each example across the four classes, which leads to a data imbalance. Therefore, for all utterances in the training set, we randomly sample additional examples within a window of 5 frames around the center phone, to at least 50 training examples per class per speaker. It is not always possible to reach the target of 50 examples, however, if no more data is available to sample from. This process gives a total of INLINEFORM0 10700 training examples with roughly 2000 to 3000 examples per class, with each speaker having an average of 185 examples. Because the amount of data varies per speaker, we compute a sampling score, which denotes the proportion of sampled examples to the speaker's total training examples. We expect speakers with high sampling scores (less unique data overall) to underperform when compared with speakers with more varied training examples." + ], + [ + "For each system, we normalize the training data to zero mean and unit variance. Due to the high dimensionality of the data (63x412 samples per frame), we have opted to investigate two preprocessing techniques: principal components analysis (PCA, often called eigentongues in this context) and a 2-dimensional discrete cosine transform (DCT). In this paper, Raw input denotes the mean-variance normalized raw ultrasound frame. PCA applies principal components analysis to the normalized training data and preserves the top 1000 components. DCT applies the 2D DCT to the normalized raw ultrasound frame and the upper left 40x40 submatrix (1600 coefficients) is flattened and used as input.", + "The first type of classifier we evaluate in this work are feedforward neural networks (DNNs) consisting of 3 hidden layers, each with 512 rectified linear units (ReLUs) with a softmax activation function. The networks are optimized for 40 epochs with a mini-batch of 32 samples using stochastic gradient descent. Based on preliminary experiments on the validation set, hyperparameters such learning rate, decay rate, and L2 weight vary depending on the input format (Raw, PCA, or DCT). Generally, Raw inputs work better with smaller learning rates and heavier regularization to prevent overfitting to the high-dimensional data. As a second classifier to evaluate, we use convolutional neural networks (CNNs) with 2 convolutional and max pooling layers, followed by 2 fully-connected ReLU layers with 512 nodes. The convolutional layers use 16 filters, 8x8 and 4x4 kernels respectively, and rectified units. The fully-connected layers use dropout with a drop probability of 0.2. Because CNN systems take longer to converge, they are optimized over 200 epochs. For all systems, at the end of every epoch, the model is evaluated on the development set, and the best model across all epochs is kept." + ], + [ + "We train speaker-dependent systems separately for each speaker, using all of their training data (an average of 185 examples per speaker). These systems use less data overall than the remaining systems, although we still expect them to perform well, as the data matches in terms of speaker characteristics. Realistically, such systems would not be viable, as it would be unreasonable to collect large amounts of data for every child who is undergoing speech therapy. We further evaluate all trained systems in a multi-speaker scenario. In this configuration, the speaker sets for training, development, and testing are equal. That is, we evaluate on speakers that we have seen at training time, although on different utterances. A more realistic configuration is a speaker-independent scenario, which assumes that the speaker set available for training and development is disjoint from the speaker set used at test time. This scenario is implemented by leave-one-out cross-validation. Finally, we investigate a speaker adaptation scenario, where training data for the target speaker becomes available. This scenario is realistic, for example, if after a session, the therapist were to annotate a small number of training examples. In this work, we use the held-out training data to finetune a pretrained speaker-independent system for an additional 6 epochs in the DNN systems and 20 epochs for the CNN systems. We use all available training data across all training scenarios, and we investigate the effect of the number of samples on one of the top performing systems.", + "This work is primarily concerned with generalizing to unseen speakers. Therefore, we investigate a method to provide models with speaker-specific inputs. A simple approach is to use the speaker mean, which is the pixel-wise mean of all raw frames associated with a given speaker, illustrated in Figure FIGREF8 . The mean frame might capture an overall area of tongue activity, average out noise, and compensate for probe placement differences across speakers. Speaker means are computed after mean variance normalization. For PCA-based systems, matrix decomposition is applied on the matrix of speaker means for the training data with 50 components being kept, while the 2D DCT is applied normally to each mean frame. In the DNN systems, the speaker mean is appended to the input vector. In the CNN system, the raw speaker mean is given to the network as a second channel. All model configurations are similar to those described earlier, except for the DNN using Raw input. Earlier experiments have shown that a larger number of parameters are needed for good generalization with a large number of inputs, so we use layers of 1024 nodes rather than 512." + ], + [ + "Results for all systems are presented in Table TABREF10 . When comparing preprocessing methods, we observe that PCA underperforms when compared with the 2 dimensional DCT or with the raw input. DCT-based systems achieve good results when compared with similar model architectures, especially when using smaller amounts of data as in the speaker-dependent scenario. When compared with raw input DNNs, the DCT-based systems likely benefit from the reduced dimensionality. In this case, lower dimensional inputs allow the model to generalize better and the truncation of the DCT matrix helps remove noise from the images. Compared with PCA-based systems, it is hypothesized the observed improvements are likely due to the DCT's ability to encode the 2-D structure of the image, which is ignored by PCA. However, the DNN-DCT system does not outperform a CNN with raw input, ranking last across adapted systems.", + "When comparing training scenarios, as expected, speaker-independent systems underperform, which illustrates the difficulty involved in the generalization to unseen speakers. Multi-speaker systems outperform the corresponding speaker-dependent systems, which shows the usefulness of learning from a larger database, even if variable across speakers. Adapted systems improve over the dependent systems, except when using DCT. It is unclear why DCT-based systems underperform when adapting pre-trained models. Figure FIGREF11 shows the effect of the size of the adaptation data when finetuning a pre-trained speaker-independent system. As expected, the more data is available, the better that system performs. It is observed that, for the CNN system, with roughly 50 samples, the model outperforms a similar speaker-dependent system with roughly three times more examples.", + "Speaker means improve results across all scenarios. It is particularly useful for speaker-independent systems. The ability to generalize to unseen speakers is clear in the CNN system. Using the mean as a second channel in the convolutional network has the advantage of relating each pixel to its corresponding speaker mean value, allowing the model to better generalize to unseen speakers.", + "Figure FIGREF12 shows pair-wise scatterplots for the CNN system. Training scenarios are compared in terms of the effect on individual speakers. It is observed, for example, that the performance of a speaker-adapted system is similar to a multi-speaker system, with most speakers clustered around the identity line (bottom left subplot). Figure FIGREF12 also illustrates the variability across speakers for each of the training scenarios. The classification task is easier for some speakers than others. In an attempt to understand this variability, we can look at correlation between accuracy scores and various speaker details. For the CNN systems, we have found some correlation (Pearson's product-moment correlation) between accuracy and age for the dependent ( INLINEFORM0 ), multi-speaker ( INLINEFORM1 ), and adapted ( INLINEFORM2 ) systems. A very small correlation ( INLINEFORM3 ) was found for the independent system. Similarly, some correlation was found between accuracy and sampling score ( INLINEFORM4 ) for the dependent system, but not for the remaining scenarios. No correlation was found between accuracy and gender (point biserial correlation)." + ], + [ + "There are various possible extensions for this work. For example, using all frames assigned to a phone, rather than using only the middle frame. Recurrent architectures are natural candidates for such systems. Additionally, if using these techniques for speech therapy, the audio signal will be available. An extension of these analyses should not be limited to the ultrasound signal, but instead evaluate whether audio and ultrasound can be complementary. Further work should aim to extend the four classes to more a fine-grained place of articulation, possibly based on phonological processes. Similarly, investigating which classes lead to classification errors might help explain some of the observed results. Although we have looked at variables such as age, gender, or amount of data to explain speaker variation, there may be additional factors involved, such as the general quality of the ultrasound image. Image quality could be affected by probe placement, dry mouths, or other factors. Automatically identifying or measuring such cases could be beneficial for speech therapy, for example, by signalling the therapist that the data being collected is sub-optimal." + ], + [ + "In this paper, we have investigated speaker-independent models for the classification of phonetic segments from raw ultrasound data. We have shown that the performance of the models heavily degrades when evaluated on data from unseen speakers. This is a result of the variability in ultrasound images, mostly due to differences across speakers, but also due to shifts in probe placement. Using the mean of all ultrasound frames for a new speaker improves the generalization of the models to unseen data, especially when using convolutional neural networks. We have also shown that adapting a pre-trained speaker-independent system using as few as 50 ultrasound frames can outperform a corresponding speaker-dependent system." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0692/instruction.md b/qasper-0692/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0ea43c11a55ebd32380c64a2c067a1e194131c81 --- /dev/null +++ b/qasper-0692/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Interpretable Visual Question Answering by Visual Grounding from Attention Supervision Mining + +Question: By how much do they outperform existing state-of-the-art VQA models? \ No newline at end of file diff --git a/qasper-0693/instruction.md b/qasper-0693/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..e4fbe5ddea1e64432466b3947318474d2e90b7e0 --- /dev/null +++ b/qasper-0693/instruction.md @@ -0,0 +1,79 @@ +Name of Paper: Interpretable Visual Question Answering by Visual Grounding from Attention Supervision Mining + +Question: How do they measure the correlation between manual groundings and model generated ones? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "VQA Model Structure", + "Mining Attention Supervision from Visual Genome", + "Implementation Details", + "Datasets", + "Results", + "Conclusions" + ], + "paragraphs": [ + [ + "We are interested in the problem of visual question answering (VQA), where an algorithm is presented with an image and a question that is formulated in natural language and relates to the contents of the image. The goal of this task is to get the algorithm to correctly answer the question. The VQA task has recently received significant attention from the computer vision community, in particular because obtaining high accuracies would presumably require precise understanding of both natural language as well as visual stimuli. In addition to serving as a milestone towards visual intelligence, there are practical applications such as development of tools for the visually impaired.", + "The problem of VQA is challenging due to the complex interplay between the language and visual modalities. On one hand, VQA algorithms must be able to parse and interpret the input question, which is provided in natural language BIBREF0 , BIBREF1 , BIBREF2 . This may potentially involve understanding of nouns, verbs and other linguistic elements, as well as their visual significance. On the other hand, the algorithms must analyze the image to identify and recognize the visual elements relevant to the question. Furthermore, some questions may refer directly to the contents of the image, but may require external, common sense knowledge to be answered correctly. Finally, the algorithms should generate a textual output in natural language that correctly answers the input visual question. In spite of the recent research efforts to address these challenges, the problem remains largely unsolved BIBREF3 .", + "We are particularly interested in giving VQA algorithms the ability to identify the visual elements that are relevant to the question. In the VQA literature, such ability has been implemented by attention mechanisms. Such attention mechanisms generate a heatmap over the input image, which highlights the regions of the image that lead to the answer. These heatmaps are interpreted as groundings of the answer to the most relevant areas of the image. Generally, these mechanisms have either been considered as latent variables for which there is no supervision, or have been treated as output variables that receive direct supervision from human annotations. Unfortunately, both of these approaches have disadvantages. First, unsupervised training of attention tends to lead to models that cannot ground their decision in the image in a human interpretable manner. Second, supervised training of attention is difficult and expensive: human annotators may consider different regions to be relevant for the question at hand, which entails ambiguity and increased annotation cost. Our goal is to leverage the best of both worlds by providing VQA algorithms with interpretable grounding of their answers, without the need of direct and explicit manual annotation of attention.", + "From a practical point of view, as autonomous machines are increasingly finding real world applications, there is an increasing need to provide them with suitable capabilities to explain their decisions. However, in most applications, including VQA, current state-of-the-art techniques operate as black-box models that are usually trained using a discriminative approach. Similarly to BIBREF4 , in this work we show that, in the context of VQA, such approaches lead to internal representations that do not capture the underlying semantic relations between textual questions and visual information. Consequently, as we show in this work, current state-of-the-art approaches for VQA are not able to support their answers with a suitable interpretable representation.", + "In this work, we introduce a methodology that provides VQA algorithms with the ability to generate human interpretable attention maps which effectively ground the answer to the relevant image regions. We accomplish this by leveraging region descriptions and object annotations available in the Visual Genome dataset, and using these to automatically construct attention maps that can be used for attention supervision, instead of requiring human annotators to manually provide grounding labels. Our framework achieves competitive state-of-the-art VQA performance, while generating visual groundings that outperform other algorithms that use human annotated attention during training.", + "The contributions of this paper are: (1) we introduce a mechanism to automatically obtain meaningful attention supervision from both region descriptions and object annotations in the Visual Genome dataset; (2) we show that by using the prediction of region and object label attention maps as auxiliary tasks in a VQA application, it is possible to obtain more interpretable intermediate representations. (3) we experimentally demonstrate state-of-the-art performances in VQA benchmarks as well as visual grounding that closely matches human attention annotations." + ], + [ + "Since its introduction BIBREF0 , BIBREF1 , BIBREF2 , the VQA problem has attracted an increasing interest BIBREF3 . Its multimodal nature and more precise evaluation protocol than alternative multimodal scenarios, such as image captioning, help to explain this interest. Furthermore, the proliferation of suitable datasets and potential applications, are also key elements behind this increasing activity. Most state-of-the-art methods follow a joint embedding approach, where deep models are used to project the textual question and visual input to a joint feature space that is then used to build the answer. Furthermore, most modern approaches pose VQA as a classification problem, where classes correspond to a set of pre-defined candidate answers. As an example, most entries to the VQA challenge BIBREF2 select as output classes the most common 3000 answers in this dataset, which account for 92% of the instances in the validation set.", + "The strategy to combine the textual and visual embeddings and the underlying structure of the deep model are key design aspects that differentiate previous works. Antol et al. BIBREF2 propose an element-wise multiplication between image and question embeddings to generate spatial attention map. Fukui et al. BIBREF5 propose multimodal compact bilinear pooling (MCB) to efficiently implement an outer product operator that combines visual and textual representations. Yu et al. BIBREF6 extend this pooling scheme by introducing a multi-modal factorized bilinear pooling approach (MFB) that improves the representational capacity of the bilinear operator. They achieve this by adding an initial step that efficiently expands the textual and visual embeddings to a high-dimensional space. In terms of structural innovations, Noh et al. BIBREF7 embed the textual question as an intermediate dynamic bilinear layer of a ConvNet that processes the visual information. Andreas et al. BIBREF8 propose a model that learns a set of task-specific neural modules that are jointly trained to answer visual questions.", + "Following the successful introduction of soft attention in neural machine translation applications BIBREF9 , most modern VQA methods also incorporate a similar mechanism. The common approach is to use a one-way attention scheme, where the embedding of the question is used to generate a set of attention coefficients over a set of predefined image regions. These coefficients are then used to weight the embedding of the image regions to obtain a suitable descriptor BIBREF10 , BIBREF11 , BIBREF5 , BIBREF12 , BIBREF6 . More elaborated forms of attention has also been proposed. Xu and Saenko BIBREF13 suggest use word-level embedding to generate attention. Yang et al. BIBREF14 iterates the application of a soft-attention mechanism over the visual input as a way to progressively refine the location of relevant cues to answer the question. Lu et al. BIBREF15 proposes a bidirectional co-attention mechanism that besides the question guided visual attention, also incorporates a visual guided attention over the input question.", + "In all the previous cases, the attention mechanism is applied using an unsupervised scheme, where attention coefficients are considered as latent variables. Recently, there have been also interest on including a supervised attention scheme to the VQA problem BIBREF4 , BIBREF16 , BIBREF17 . Das et al. BIBREF4 compare the image areas selected by humans and state-of-the-art VQA techniques to answer the same visual question. To achieve this, they collect the VQA human attention dataset (VQA-HAT), a large dataset of human attention maps built by asking humans to select images areas relevant to answer questions from the VQA dataset BIBREF2 . Interestingly, this study concludes that current machine-generated attention maps exhibit a poor correlation with respect to the human counterpart, suggesting that humans use different visual cues to answer the questions. At a more fundamental level, this suggests that the discriminative nature of most current VQA systems does not effectively constraint the attention modules, leading to the encoding of discriminative cues instead of the underlying semantic that relates a given question-answer pair. Our findings in this work support this hypothesis.", + "Related to the work in BIBREF4 , Gan et al. BIBREF16 apply a more structured approach to identify the image areas used by humans to answer visual questions. For VQA pairs associated to images in the COCO dataset, they ask humans to select the segmented areas in COCO images that are relevant to answer each question. Afterwards, they use these areas as labels to train a deep learning model that is able to identify attention features. By augmenting a standard VQA technique with these attention features, they are able to achieve a small boost in performance. Closely related to our approach, Qiao et al. BIBREF17 use the attention labels in the VQA-HAT dataset to train an attention proposal network that is able to predict image areas relevant to answer a visual question. This network generates a set of attention proposals for each image in the VQA dataset, which are used as labels to supervise attention in the VQA model. This strategy results in a small boost in performance compared with a non-attentional strategy. In contrast to our approach, these previous works are based on a supervised attention scheme that does not consider an automatic mechanism to obtain the attention labels. Instead, they rely on human annotated groundings as attention supervision. Furthermore, they differ from our work in the method to integrate attention labels to a VQA model." + ], + [ + "Figure FIGREF2 shows the main pipeline of our VQA model. We mostly build upon the MCB model in BIBREF5 , which exemplifies current state-of-the-art techniques for this problem. Our main innovation to this model is the addition of an Attention Supervision Module that incorporates visual grounding as an auxiliary task. Next we describe the main modules behind this model.", + "Question Attention Module: Questions are tokenized and passed through an embedding layer, followed by an LSTM layer that generates the question features INLINEFORM0 , where INLINEFORM1 is the maximum number of words in the tokenized version of the question and INLINEFORM2 is the dimensionality of the hidden state of the LSTM. Additionally, following BIBREF12 , a question attention mechanism is added that generates question attention coefficients INLINEFORM3 , where INLINEFORM4 is the so-called number of \u201cglimpses\u201d. The purpose of INLINEFORM5 is to allow the model to predict multiple attention maps so as to increase its expressiveness. Here, we use INLINEFORM6 . The weighted question features INLINEFORM7 are then computed using a soft attention mechanism BIBREF9 , which is essentially a weighted sum of the INLINEFORM8 word features followed by a concatenation according to INLINEFORM9 .", + "Image Attention Module: Images are passed through an embedding layer consisting of a pre-trained ConvNet model, such as Resnet pretrained with the ImageNet dataset BIBREF18 . This generates image features INLINEFORM0 , where INLINEFORM1 , INLINEFORM2 and INLINEFORM3 are depth, height, and width of the extracted feature maps. Fusion Module I is then used to generate a set of image attention coefficients. First, question features INLINEFORM4 are tiled as the same spatial shape of INLINEFORM5 . Afterwards, the fusion module models the joint relationship INLINEFORM6 between questions and images, mapping them to a common space INLINEFORM7 . In the simplest case, one can implement the fusion module using either concatenation or Hadamard product BIBREF19 , but more effective pooling schemes can be applied BIBREF5 , BIBREF20 , BIBREF12 , BIBREF6 . The design choice of the fusion module remains an on-going research topic. In general, it should both effectively capture the latent relationship between multi-modal features meanwhile be easy to optimize. The fusion results are then passed through an attention module that computes the visual attention coefficient INLINEFORM8 , with which we can obtain attention-weighted visual features INLINEFORM9 . Again, INLINEFORM10 is the number of \u201cglimpses\u201d, where we use INLINEFORM11 .", + "Classification Module: Using the compact representation of questions INLINEFORM0 and visual information INLINEFORM1 , the classification module applies first the Fusion Module II that provides the feature representation of answers INLINEFORM2 , where INLINEFORM3 is the latent answer space. Afterwards, it computes the logits over a set of predefined candidate answers. Following previous work BIBREF5 , we use as candidate outputs the top 3000 most frequent answers in the VQA dataset. At the end of this process, we obtain the highest scoring answer INLINEFORM4 .", + "Attention Supervision Module: As a main novelty of the VQA model, we add an Image Attention Supervision Module as an auxiliary classification task, where ground-truth visual grounding labels INLINEFORM0 are used to guide the model to focus on meaningful parts of the image to answer each question. To do that, we simply treat the generated attention coefficients INLINEFORM1 as a probability distribution, and then compare it with the ground-truth using KL-divergence. Interestingly, we introduce two attention maps, corresponding to relevant region-level and object-level groundings, as shown in Figure FIGREF3 . Sections SECREF4 and SECREF5 provide details about our proposed method to obtain the attention labels and to train the resulting model, respectively." + ], + [ + "Visual Genome (VG) BIBREF21 includes the largest VQA dataset currently available, which consists of 1.7M QA pairs. Furthermore, for each of its more than 100K images, VG also provides region and object annotations by means of bounding boxes. In terms of visual grounding, these region and object annotations provide complementary information. As an example, as shown in Figure FIGREF3 , for questions related to interaction between objects, region annotations result highly relevant. In contrast, for questions related to properties of specific objects, object annotations result more valuable. Consequently, in this section we present a method to automatically select region and object annotations from VG that can be used as labels to implement visual grounding as an auxiliary task for VQA.", + "For region annotations, we propose a simple heuristic to mine visual groundings: for each INLINEFORM0 we enumerate all the region descriptions of INLINEFORM1 and pick the description INLINEFORM2 that has the most (at least two) overlapped informative words with INLINEFORM3 and INLINEFORM4 . Informative words are all nouns and verbs, where two informative words are matched if at least one of the following conditions is met: (1) Their raw text as they appear in INLINEFORM5 or INLINEFORM6 are the same; (2) Their lemmatizations (using NLTK BIBREF22 ) are the same; (3) Their synsets in WordNet BIBREF23 are the same; (4) Their aliases (provided from VG) are the same. We refer to the resulting labels as region-level groundings. Figure FIGREF3 (a) illustrates an example of a region-level grounding.", + "In terms of object annotations, for each image in a INLINEFORM0 triplet we select the bounding box of an object as a valid grounding label, if the object name matches one of the informative nouns in INLINEFORM1 or INLINEFORM2 . To score each match, we use the same criteria as region-level groundings. Additionally, if a triplet INLINEFORM3 has a valid region grounding, each corresponding object-level grounding must be inside this region to be accepted as valid. As a further refinement, selected objects grounding are passed through an intersection over union filter to account for the fact that VG usually includes multiple labels for the same object instance. As a final consideration, for questions related to counting, region-level groundings are discarded after the corresponding object-level groundings are extracted. We refer to the resulting labels as object-level groundings. Figure FIGREF3 (b) illustrates an example of an object-level grounding.", + "As a result, combining both region-level and object-level groundings, about 700K out of 1M INLINEFORM0 triplets in VG end up with valid grounding labels. We will make these labels publicly available." + ], + [ + "We build the attention supervision on top of the open-sourced implementation of MCB BIBREF5 and MFB BIBREF12 . Similar to them, We extract the image feature from res5c layer of Resnet-152, resulting in INLINEFORM0 spatial grid ( INLINEFORM1 , INLINEFORM2 , INLINEFORM3 ). We construct our ground-truth visual grounding labels to be INLINEFORM4 glimpse maps per QA pair, where the first map is object-level grounding and the second map is region-level grounding, as discussed in Section SECREF4 . Let INLINEFORM5 be the coordinate of INLINEFORM6 selected object bounding box in the grounding labels, then the mined object-level attention maps INLINEFORM7 are: DISPLAYFORM0 ", + "where INLINEFORM0 is the indicator function. Similarly, the region-level attention maps INLINEFORM1 are: DISPLAYFORM0 ", + "", + "Afterwards, INLINEFORM0 and INLINEFORM1 are spatially L1-normalized to represent probabilities and concatenated to form INLINEFORM2 .", + "The model is trained using a multi-task loss, DISPLAYFORM0 ", + "where INLINEFORM0 denotes cross-entropy and INLINEFORM1 denotes KL-divergence. INLINEFORM2 corresponds to the learned parameters. INLINEFORM3 is a scalar that weights the loss terms. This scalar decays as a function of the iteration number INLINEFORM4 . In particular, we choose to use a cosine-decay function: DISPLAYFORM0 ", + "This is motivated by the fact that the visual grounding labels have some level of subjectivity. As an example, Figure FIGREF11 (second row) shows a case where the learned attention seems more accurate than the VQA-HAT ground truth. Hence, as the model learns suitable parameter values, we gradually loose the penalty on the attention maps to provide more freedom to the model to selectively decide what attention to use. It is important to note that, for training samples in VQA-2.0 or VG that do not have region-level or object-level grounding labels, INLINEFORM0 in Equation EQREF6 , so the loss is reduced to the classification term only. In our experiment, INLINEFORM1 is calibrated for each tested model based on the number of training steps. In particular, we choose INLINEFORM2 for all MCB models and INLINEFORM3 for others." + ], + [ + "VQA-2.0: The VQA-2.0 dataset BIBREF2 consists of 204721 images, with a total of 1.1M questions and 10 crowd-sourced answers per question. There are more than 20 question types, covering a variety of topics and free-form answers. The dataset is split into training (82K images and 443K questions), validation (40K images and 214K questions), and testing (81K images and 448K questions) sets. The task is to predict a correct answer INLINEFORM0 given a corresponding image-question pair INLINEFORM1 . As a main advantage with respect to version 1.0 BIBREF2 , for every question VQA-2.0 includes complementary images that lead to different answers, reducing language bias by forcing the model to use the visual information.", + "Visual Genome: The Visual Genome (VG) dataset BIBREF21 contains 108077 images, with an average of 17 QA pairs per image. We follow the processing scheme from BIBREF5 , where non-informative words in the questions and answers such as \u201ca\u201d and \u201cis\u201d are removed. Afterwards, INLINEFORM0 triplets with answers to be single keyword and overlapped with VQA-2.0 dataset are included in our training set. This adds 97697 images and about 1 million questions to the training set. Besides the VQA data, VG also provides on average 50 region descriptions and 30 object instances per image. Each region/object is annotated by one sentence/phrase description and bounding box coordinates.", + "VQA-HAT: VQA-HAT dataset BIBREF4 contains 58475 human visual attention heat (HAT) maps for INLINEFORM0 triplets in VQA-1.0 training set. Annotators were shown a blurred image, a INLINEFORM1 pair and were asked to \u201cscratch\u201d the image until they believe someone else can answer the question by looking at the blurred image and the sharpened area. The authors also collect INLINEFORM2 HAT maps for VQA-1.0 validation sets, where each of the 1374 INLINEFORM3 were labeled by three different annotators, so one can compare the level of agreement among labels. We use VQA-HAT to evaluate visual grounding performance, by comparing the rank-correlation between human attention and model attention, as in BIBREF4 , BIBREF24 .", + "VQA-X: VQA-X dataset BIBREF24 contains 2000 labeled attention maps in VQA-2.0 validation sets. In contrast to VQA-HAT, VQA-X attention maps are in the form of instance segmentations, where annotators were asked to segment objects and/or regions that most prominently justify the answer. Hence the attentions are more specific and localized. We use VQA-X to evaluate visual grounding performance by comparing the rank-correlation, as in BIBREF4 , BIBREF24 ." + ], + [ + "We evaluate the performance of our proposed method using two criteria: i) rank-correlation BIBREF25 to evaluate visual grounding and ii) accuracy to evaluate question answering. Intuitively, rank-correlation measures the similarity between human and model attention maps under a rank-based metric. A high rank-correlation means that the model is `looking at' image areas that agree to the visual information used by a human to answer the same question. In terms of accuracy of a predicted answer INLINEFORM0 is evaluated by: DISPLAYFORM0 ", + "", + "Table TABREF10 reports our main results. Our models are built on top of prior works with the additional Attention Supervision Module as described in Section SECREF3 . Specifically, we denote by Attn-* our adaptation of the respective model by including our Attention Supervision Module. We highlight that MCB model is the winner of VQA challenge 2016 and MFH model is the best single model in VQA challenge 2017. In Table TABREF10 , we can observe that our proposed model achieves a significantly boost on rank-correlation with respect to human attention. Furthermore, our model outperforms alternative state-of-art techniques in terms of accuracy in answer prediction. Specifically, the rank-correlation for MFH model increases by 36.4% when is evaluated in VQA-HAT dataset and 7.7% when is evaluated in VQA-X. This indicates that our proposed methods enable VQA models to provide more meaningful and interpretable results by generating more accurate visual grounding.", + "Table TABREF10 also reports the result of an experiment where the decaying factor INLINEFORM0 in Equation EQREF7 is fixed to a value of 1. In this case, the model is able to achieve higher rank-correlation, but accuracy drops by 2%. We observe that as training proceeds, attention loss becomes dominant in the final training steps, which affects the accuracy of the classification module.", + "Figure FIGREF11 shows qualitative results of the resulting visual grounding, including also a comparison with respect to no-attn model." + ], + [ + "In this work we have proposed a new method that is able to slightly outperform current state-of-the-art VQA systems, while also providing interpretable representations in the form of an explicitly trainable visual attention mechanism. Specifically, as a main result, our experiments provide evidence that the generated visual groundings achieve high correlation with respect to human-provided attention annotations, outperforming the correlation scores of previous works by a large margin.", + "As further contributions, we highlight two relevant insides of the proposed approach. On one side, by using attention labels as an auxiliary task, the proposed approach demonstrates that is able to constraint the internal representation of the model in such a way that it fosters the encoding of interpretable representations of the underlying relations between the textual question and input image. On other side, the proposed approach demonstrates a method to leverage existing datasets with region descriptions and object labels to effectively supervise the attention mechanism in VQA applications, avoiding costly human labeling.", + "As future work, we believe that the superior visual grounding provided by the proposed method can play a relevant role to generate natural language explanations to justify the answer to a given visual question. This scenario will help to demonstrate the relevance of our technique as a tool to increase the capabilities of AI based technologies to explain their decisions.", + "", + "Acknowledgements: This work was partially funded by Oppo, Panasonic and the Millennium Institute for Foundational Research on Data." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0694/instruction.md b/qasper-0694/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ed43ea5d592b93330fdc2dcbc589d918f63c29e3 --- /dev/null +++ b/qasper-0694/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Interpretable Visual Question Answering by Visual Grounding from Attention Supervision Mining + +Question: How do they obtain region descriptions and object annotations? \ No newline at end of file diff --git a/qasper-0802/instruction.md b/qasper-0802/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..399f326ab04e4e6ad30ee62aa4850ec30683e432 --- /dev/null +++ b/qasper-0802/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Style Transfer for Texts: to Err is Human, but Error Margins Matter + +Question: What is state of the art method? \ No newline at end of file diff --git a/qasper-0805/instruction.md b/qasper-0805/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..456300bf9da5bdb3ce90aa49799a410687e1a3a9 --- /dev/null +++ b/qasper-0805/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Style Transfer for Texts: to Err is Human, but Error Margins Matter + +Question: How much does the standard metrics for style accuracy vary on different re-runs? \ No newline at end of file diff --git a/qasper-0820/instruction.md b/qasper-0820/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c0dfe11dbff35b4cc9886bf84e7a30d1e142accd --- /dev/null +++ b/qasper-0820/instruction.md @@ -0,0 +1,97 @@ +Name of Paper: Automatic Differentiation in ROOT + +Question: Is this AD implementation used in any deep learning framework? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Background", + "Background ::: AD and its Modes", + "Background ::: AD Implementations", + "Architecture and Implementation", + "Results", + "Results ::: Accuracy", + "Results ::: Performance", + "Results ::: Performance in TFormula", + "Conclusion", + "Acknowledgments" + ], + "paragraphs": [ + [ + "Accurate and efficient computation of derivatives is vital for a wide variety of computing applications, including numerical optimization, solution of nonlinear equations, sensitivity analysis, and nonlinear inverse problems. Virtually every process could be described with a mathematical function, which can be thought of as an association between elements from different sets. Derivatives track how a varying quantity depends on another quantity, for example how the position of a planet varies as time varies.", + "Derivatives and gradients (vectors of partial derivatives of multivariable functions) allow us to explore the properties of a function and thus the described process as a whole. Gradients are an essential component in gradient-based optimization methods, which have become more and more important in recent years, in particular with its application training of (deep) neural networks BIBREF0.", + "Several different techniques are commonly used to compute the derivatives of a given function, either exactly or approximately BIBREF1, BIBREF0, BIBREF2. The most prevalent techniques are:", + "Numerical differentiation, based on the finite difference method, provides a way to evaluate derivatives approximately. While simple, numerical differentiation can be slow (the run-time complexity grows linearly with the number of input variables) and may have problems with accuracy due to round-off and truncation errors.", + "Symbolic differentiation, based on transformations of symbolic expressions of functions, provides exact closed-form expressions for the derivatives. It faces difficulties when the function to be differentiated is not available in a closed form, which is often the case for computer programs which may contain control flow. Symbolic differentiation can produce derivative expressions that are computationally expensive to evaluate due to difficulties in exploiting common subexpressions.", + "Automatic differentiation (AD) computes derivatives accurately to the precision of the original function, supports control flow and uses at most a small constant factor more time and space than it takes to evaluate the original function, at the expense of increased implementation complexity and introducing more software dependencies.", + "Numerical and symbolic differentiation methods are slow at computing gradients of functions with many input variables, as is often needed for gradient-based optimization algorithms. Both methods have problems calculating higher-order derivatives, where the complexity and errors due to numerical precision increase. Automatic differentiation largely avoids the problems of numerical and symbolic differentiation.", + "In this paper, we describe the implementation of automatic differentiation techniques in ROOT, which is the data analysis framework broadly used High-Energy Physics BIBREF3. This implementation is based on Clad BIBREF4, BIBREF5, which is an automatic differentiation plugin for computation expressed in C/C++." + ], + [ + "Here, we briefly discuss main algorithmic and implementation principles behind AD. An in-depth overview and more formal description can be found in BIBREF1 and BIBREF2, respectively." + ], + [ + "AD is based on the decomposition of the procedure (e.g. a source code that computes the original function) into a sequence of simple mathematical operations (e.g. $+, -, *, /, \\sin , \\cos , \\exp $) that can be expressed using a series of intermediate results. Subsequently, derivatives of every intermediate result are evaluated and combined via the chain rule of calculus to obtain the derivatives of the whole sequence. The control flow (e.g. branches, loops) can be incorporated by differentiating the control flow of the original function during the derivative evaluation. Two main modes of AD, which differ in the order of application of the chain rule, are used:", + "Forward mode operates in a top-down approach and computes the derivative of every intermediate result with respect to a single selected input variable of the function. As soon as a final result of the function is reached, the partial derivative with respect to the selected input is available. A single evaluation of the forward mode can only compute partial derivatives with respect to a single input variable. Thus, when the whole gradient is required, forward mode must be invoked once per every input variable, leading to $m \\cdot c_{F} \\cdot n$ runtime complexity, where $m$ is the number of input variables, $n$ is the algorithmic complexity of the original function and $c_{F} < 3 $ is a small constant factor overhead of a single invocation of the forward mode BIBREF2.", + "Reverse mode operates in a bottom-up approach and computes the derivative of a function's output with respect to every intermediate result. Once every input variable of the function is reached, the whole gradient of an output is available. Note that, independently on the number of input variables $N$, a single evaluation of the reverse mode is sufficient to get the whole gradient of a function's output, leading to $c_{R} \\cdot n$ runtime complexity, where $n$ is the complexity of the original function and $c_{R} \\le 4$ is a small constant factor overhead BIBREF2. This is a huge advantage in settings with a single scalar output and many inputs, which is often the case in machine-learning problems where $N >> 10^6$ that makes the forward mode infeasible. As a disadvantage, reverse mode implementations are more complicated, and dynamic memory allocations may be required when dynamic control flow is involved. Depending on the original function, this may cause a single evaluation of the reverse mode to be somewhat slower compared to a single evaluation of the forward mode." + ], + [ + "AD techniques have been implemented in a variety of programming languages and paradigms, ranging from classical tools for Fortran BIBREF6 and C BIBREF7, to recent active work on tools specific to machine-learning applications BIBREF8, BIBREF9, and modern general-purpose programming languages BIBREF10, BIBREF11. We refer the reader to www.autodiff.org for a comprehensive list of available AD implementations for various languages.", + "In particular, several implementations exist for C++, e.g. BIBREF12, BIBREF13, BIBREF14. Majority of implementations of AD fall into one of the two categories of implementation techniques:", + "Tools based on operator overloading utilize features of programming languages like C++ and Python to define custom types and overload mathematical operators (e.g. +, -, *, /) and functions (e.g. $\\exp , \\sin , \\cos $) on them. Such implementations are often based on custom AD-enabled types that wrap values of both the original and derivative functions and redefine operators to simultaneously act on original and derivative values. In C++, such tools are often implemented as a library that introduces templated differentiable types and corresponding mathematical operations. Then, functions called on the custom type return both original and derivative values. This is a powerful technique but has two primary limitations: legacy code and performance. Functions must be either polymorphic (templated) or explicitly defined on AD-enabled type to be differentiated. Differentiation of pre-existing source code using builtin types such as double and float is not possible. Users are required to use additional level of abstraction in the form of library-specific types instead of first-class language features. Moreover, the performance of the derivative generation can be suboptimal due to the C++ metaprogramming system which usually constructs deep template instantiation chains. Performance can be even more problematic when creating a higher order derivatives.", + "Tools based on source transformation analyze the source code of the original function and build another source code for the derivative function. Such techniques typically accept and generate any code using built-in features of the original language and do not require custom libraries. On the other hand, they require an additional pass over the source file to analyze and generate derivative code. Source transformation can fully utilize source-level optimizations and has reasonably good performance. Implementation is more complicated and it is problematic to achieve full coverage of C++ language features. While full integration with a compiler can make AD a first-class language feature that is transparent for the user, most current implementations for C++ are based on custom parsers that do not have full coverage of the vast variety of C++ language constructs and require a separate step before compilation." + ], + [ + "Automatic differentiation in ROOT is based on Clad BIBREF4, BIBREF5. Clad is a source transformation AD tool for C++. It is based on LLVM compiler infrastructure BIBREF15 and is implemented as a plugin for C++ compiler Clang, which allows Clad to be transparently integrated into the compilation phase and to utilize large parts of the compiler. Clad relies on Clang's parsing and code generation functionality and can differentiate complicated C++ constructs. Clad supports both forward and reverse mode. It is available as a standalone Clang plugin that, when attached to the compiler, produces derivatives in the compilation phase.", + "On top of that, Clad is integrated directly into ROOT to provide AD functionality as an integral part of the framework. ROOT has a C++ interpreter Cling BIBREF16 which is built on the top of LLVM and Clang. This allows Clad to be attached to Cling as a plugin in a similar way as it can be attached to Clang. In this section, we discuss 1) architecture of Clad and its interaction with Cling; and 2) details of its integration into ROOT.", + "Clad operates on Clang AST (abstract syntax tree) by analyzing the AST of the original function and generating the AST of the derivative. Clad provides two API functions: clad::differentiate for forward mode and clad::gradient for reverse mode, which can be used directly in the source code to mark a function for differentiation (see BIBREF5 for more details on usage and code examples).", + "The information flow of interactions with Cling during differentiation (Figure FIGREF13) is:", + "A function is marked for differentiation with the C++ construct clad::differentiate or clad::gradient (step 1).", + "Cling in ROOT performs incremental compilation and receives an abstract syntax tree (AST) representation of the code (step 2).", + "Cling detects the differentiation marker and sends the AST of the original function to Clad, which transforms the AST to produce the AST of the derivative (step 3).", + "Clad returns the derivative AST to Cling for code generation and execution by the low level LLVM primitives (steps 4, 5, 6, 7). Alternatively, if Clad was configured for non-interactive use, the generated AST can be converted to a C++ source code and written to a text file. The generated code then can be compiled with any C++ compiler (steps 8, 9).", + "Inside of ROOT, interface functions clad::differentiate and clad::gradient are accessible via include . Clad is also directly integrated into the TFormula class that encapsulates the concept of multidimensional mathematical functions in ROOT. TFormula is a primitive in ROOT's math package which is connected to the Cling interpreter. In the context of TFormula, Clad can differentiate functions available in the interpreter. The TFormula::GenerateGradientPar method uses Clad to differentiate the underlying code of the formula with respect to its parameters and generate the code for the gradient. TFormula::GradientPar method then evaluates the gradient at a specified point." + ], + [ + "In this section, we empirically compare automatic differentiation (AD, our implementation based on Clad) and numerical differentiation (ND, based on finite difference method) in ROOT. We show that AD can drastically improve accuracy and performance of derivative evaluation, compared to ND." + ], + [ + "As stated in Section SECREF1, numerical differentiation may give imprecise results while AD computes the derivatives exactly. We show an example of a function where this difference is apparent: AD provides exact result while ND suffers from the loss of accuracy.", + "2", + "", + "The function is the PDF of Breit-Wigner distribution (Eq. DISPLAY_FORM19), whose derivative with respect to $\\Gamma $ (Eq. DISPLAY_FORM20) has critical points at $\\Gamma =\\pm {2x}$. In ROOT, the function is implemented as in (Listing SECREF18).", + "linenos=false inline double breitwignerpdf(double x, double gamma, double x0 = 0) double gammahalf = gamma/2.0; return gammahalf/(MPI * ((x-x0)*(x-x0) + gammahalf*gammahalf));", + "listingBreit-Wigner PDF implementation in ROOT", + "", + "When evaluating the derivative of breitwignerpdf with respect to gamma at x=1, gamma=2, ND in ROOT the yields a result close to 0 with an absolute error of $10^{-13}$ despite the fact that the function is smooth and well-conditioned at this point. The approximation error becomes larger when the derivative is evaluated further from the critical point. In contrast, the automatic differentiation (in both modes) yields the exact result of 0." + ], + [ + "Section SECREF2 showed that reverse mode AD computes gradients in a single pass with a runtime complexity of at most $4 \\cdot n$, which depends only on the complexity $n$ and not the dimensionality $dim$ of the original function. On the other hand, numerical differentiation requires a separate evaluation of the original function for every dimension to compute the entire gradient, making the overall the run-time complexity of gradient evaluation via central finite difference method $2 \\cdot dim \\cdot n$. Hence, in theory, reverse mode achieves an asymptotic speedup of $O(dim)$ over the numerical differentiation and can be up to $dim / 2$ times faster.", + "We experimentally verify this by comparing the performance of gradient evaluation produced by reverse mode AD against our an implementation of numerical differentiation via the central finite difference method. We use the two functions in Listing SECREF21: sum, which computes the sum of all values in a vector; and mvn, which implements the PDF of a multivariate normal distribution. Both functions have a parameter dim which defines the dimension, and gradients are taken with respect to dim-dimensional vector p. While closed-form expressions of these gradients are well-known, these functions make a good basis of a benchmark as they perform typical operations that are commonly found inside more complicated functions (e.g. +, *, pow, exp inside loop).", + "", + "linenos=false double sum(double* p, int dim) double r = 0.0; for (int i = 0; i < dim; i++) r += p[i]; return r; linenos=false double mvn(double* x, double* p /*means*/, double sigma, int dim) double t = 0; for (int i = 0; i < dim; i++) t += (x[i] - p[i])*(x[i] - p[i]); t = -t / (2*sigma*sigma); return std::pow(2*MPI, -n/2.0) * std::pow(sigma, -0.5) * std::exp(t); listingImplementations of sum and mvn functions", + "Gradients of sum produced by numerical differentiation and Clad are shown in Listing SECREF21.", + "", + "linenos=false double* sumnumgrad(double* p, int dim, double eps = 1e-8) double result = new double[dim]; for (int i = 0; i < dim; i++) double pi = p[i]; p[i] = pi + eps; double v1 = sum(p, dim); p[i] = pi - eps; double v2 = sum(p, dim); result[i] = (v1 - v2)/(2 * eps); p[i] = pi; return result;", + "linenos=false void sumadgrad(double *p, int dim, double *result) double dr = 0; unsigned long t0; int di = 0; clad::tape t1 = ; double r = 0.; t0 = 0; for (int i = 0; i < dim; i++) t0++; r += p[clad::push(t1, i)]; double sumreturn = r; dr += 1; for (; t0; t0\u2013) double rd0 = dr; dr += rd0; result[clad::pop(t1)] += rd0; dr -= rd0; listingGradient of sum: (left) using finite differences, (right) generated by Clad", + "We perform the evaluation for values of dim between 5 and 20480. Figure FIGREF22 shows the comparison for (a) sum; (b) mvn and confirms the expected theoretical speedup of $O(dim)$, with AD-generated gradient being $~dim/4$ times faster for sum and $~dim/25$ times faster for mvn (slowdown is due to more expensive operations like pow, exp).", + "", + "" + ], + [ + "Figure FIGREF26 shows the performance comparisons of reverse-mode AD and ND for the task of evaluating gradients of TFormula's builtin primitive probability density functions. The functions are gaus ($dim=3$), expo ($dim=2$), crystalball ($dim=5$), breitwigner ($dim=5$) and cheb2 ($dim=4$). Despite the low dimensionality ($dim \\le 5$), AD gives significant (approx. 10x) speedups. The speedups are even larger than expected factor of $dim/2$ that follows from theoretical results, apparently due to additional overhead of the implementation of numerical differentiation in ROOT, which tries to find the optimal step size for its finite difference method to improve accuracy.", + "In Figure FIGREF26, we perform fitting of a Gaussian distribution to a histogram of random samples via gradient-based optimization. In ROOT, this functionality is implemented in TFormula-based TF1 class. We can therefore use AD due to the integration of Clad into TFormula. Figure FIGREF26 compares the performance of the AD-based TF1 fitting with the numerical fitting in the Hist package. As in previous experiments, we show that AD scales better with problem dimensionality (number of histogram bins) on this task. The integration of Clad into TFormula makes it straightforward to use AD for fitting in ROOT." + ], + [ + "We discussed our implementation of automatic differentiation in ROOT based on Clad. We demonstrated that Clad is integrated into ROOT and can be easily used in various contexts inside ROOT (e.g. histogram fitting). Furthermore, we showed that automatic differentiation in ROOT achieves significant improvements in accuracy and performance over numerical differentiation. The performance and accuracy are promising and encourage further work in the development of Clad and its integration in ROOT.", + "Possible further improvements for Clad include optimizations to code transformation and design of a consistent interface for derivatives and gradients computation. This functionality can be further extended, including the computation of Jacobians and higher-order derivatives. In order to achieve optimal performance, the evaluation of individual derivatives could be executed in parallel. Besides, the Clad API should enable a flexible execution method based on the needs of its user." + ], + [ + "This work has been supported by U.S. NSF grants PHY-1450377 and 1450323." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0832/instruction.md b/qasper-0832/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..6440b3f79c928550fd2acecc15bfd17fd2e62a76 --- /dev/null +++ b/qasper-0832/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Torch-Struct: Deep Structured Prediction Library + +Question: What general-purpose optimizations are included? \ No newline at end of file diff --git a/qasper-0833/instruction.md b/qasper-0833/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..27ff113e8056f270d7bad5dd8fc77318be6a4834 --- /dev/null +++ b/qasper-0833/instruction.md @@ -0,0 +1,260 @@ +Name of Paper: Embedding Projection for Targeted Cross-Lingual Sentiment: Model Comparisons and a Real-World Study + +Question: what baseline do they compare to? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Targeted Sentiment Classification", + "Cross-Lingual Approaches to Sentiment Analysis", + "Bilingual Distributional Models and the Contributions of this Paper", + "Previous Work", + "Machine Translation Based Methods", + "Bilingual Embedding Methods", + "Sentiment Embeddings", + "Targeted Sentiment Analysis", + "Projecting Sentiment Across Languages", + "Sentence-level Model", + "Targeted Model", + "Experiments", + "Datasets and Resources", + "Setting for Experiment 1: Sentence-level Classification", + "Setting for Experiment 2: Targeted Classification", + "Experiment 1: Sentence-level Classification", + "Experiment 2: Targeted Classification", + "Motivation", + "Experimental Setup", + "Results", + "Discussion", + "Conclusion" + ], + "paragraphs": [ + [ + "Opinions are everywhere in our lives. Every time we open a book, read the newspaper, or look at social media, we scan for opinions or form them ourselves. We are cued to the opinions of others, and often use this information to update our own opinions Asch1955,Das2014. This is true on the Internet as much as it is in our face-to-face relationships. In fact, with its wealth of opinionated material available online, it has become feasible and interesting to harness this data in order to automatically identify opinions, which had previously been far more expensive and tedious when the only access to data was offline.", + "Sentiment analysis, sometimes referred to as opinion mining, seeks to create data-driven methods to classify the polarity of a text. The information obtained from sentiment classifiers can then be used for tracking user opinions in different domains Pang2002,Socher2013b,Nakov2013, predicting the outcome of political elections wang2012demo,bakliwal2013, detecting hate speech online Nahar2012,hartung-EtAl:2017:WASSA2017, as well as predicting changes in the stock market Pogolu2016.", + "Sentiment analysis can be modeled as a classification task, especially at sentence- and document-level, or as a sequence-labeling task at target-level. Targeted sentiment analysis aims at predicting the polarity expressed towards a particular entity or sub-aspect of that entity. This is a more realistic view of sentiment, as polarities are directed towards targets, not spread uniformly across sentences or documents. Take the following example, where we mark the sentiment target with green, positive sentiment expressions with blue, and negative sentiment expressions with red.:", + "The caf\u00e9 near my house has great coffee but I", + "never go there because the service is terrible.", + "In this sentence, it is not stated what the sentiment towards the target \u201ccaf\u00e9\u201d is, while the sentiment of the target \u201ccoffee\u201d is positive and that of \u201cservice\u201d is negative. In order to correctly classify the sentiment of each target, it is necessary to (1) detect the targets, (2) detect polarity expressions, and (3) resolve the relations between these.", + "In order to model these relationships and test the accuracy of the learned models, hand-annotated resources are typically used for training machine learning algorithms. Resource-rich languages, e. g., English, have high-quality annotated data for both classification and sequence-labeling tasks, as well as for a variety of domains. However, under-resourced languages either completely lack annotated data or have only a few resources for specific domains or sentiment tasks. For instance, for aspect-level sentiment analysis, English has datasets available in the news domain Wiebe2005, product review domain HuandLiu2004,Ding2008,Pontiki2014,Pontiki2015, education domain Welch2016, medical domain Grasser2018, urban neighborhood domain Saeidi2016, and financial Maia2018 domain. Spanish, on the other hand, has only three datasets Agerri2013,Pontiki2016, while Basque and Catalan only have one each for a single domain Barnes2018a. The cost of annotating data can often be prohibitive as training native-speakers to annotate fine-grained sentiment is a long process. This motivates the need to develop sentiment analysis methods capable of leveraging data annotated in other languages." + ], + [ + "Previous work on cross-lingual sentiment analysis (CLSA) offers a way to perform sentiment analysis in an under-resourced language that does not have any annotated data available. Most methods relied on the availability of large amounts of parallel data to transfer sentiment information across languages. Machine translation (MT), for example, has been the most common approach to cross-lingual sentiment analysis Banea2013,Almeida2015,Zhang2017. Machine translation, however, can be biased towards domains Hua2008,Bertoldi2009,Koehn2017, does not always preserve sentiment Mohammad2016, and requires millions of parallel sentences Gavrila2011,Vaswani2017, which places a limit on which languages can benefit from these approaches. The following example illustrates that MT does not preserve sentiment (hotel review in Basque, automatically translated via translate.google.com):", + "Hotel $^{1}$ txukuna da, nahiko berria. Harreran zeuden langileen arreta $^{2}$ ez zen onena izan. Tren geltoki bat $^{3}$ du 5 minutura eta kotxez $^{4}$ berehala iristen da baina oinez $^{5}$ urruti samar dago.", + "The hotel $^{1}$ is tidy, quite new. The care of the workers at reception $^{2}$ was not the best. It's 5 minutes away from a train station $^{3}$ and it's quick to reach the car $^{4}$ , but it's a short distance away.", + "While the first two sentences are mostly well translated for the purposes of sentiment analysis, in the third, there are a number of reformulations and deletions that lead to a loss of information. It should read \u201cIt has a train station five minutes away and by car you can reach it quickly, but by foot it's quite a distance.\u201d We can see that one of the targets has been deleted and the sentiment has flipped from negative to positive. Such common problems degrade the results of cross-lingual sentiment systems that use MT, especially at target-level.", + "Although high quality machine translation systems exist between many languages and have been shown to enable cross-lingual sentiment analysis, for the vast majority of language pairs in the world there is not enough parallel data to create these high quality MT systems. This lack of parallel data coupled with the computational expense of MT means that approaches to cross-lingual sentiment analysis that do not require MT should be preferred. Additionally, most cross-lingual sentiment approaches using MT have concentrated on sentence- and document-level, and have not explored targeted or aspect-level sentiment tasks." + ], + [ + "Recently, several bilingual distributional semantics models (bilingual embeddings) have been proposed and provide a useful framework for cross-lingual research without requiring machine translation. They are effective at generating features for bilingual dictionary induction Mikolov2013translation,Artetxe2016,Lample2017, cross-lingual text classification Prettenhofer2011b,Chandar2014, or cross-lingual dependency parsing Sogaard2015, among others. In this framework, words are represented as $n$ -dimensional vectors which are created on large monolingual corpora in order to (1) maximize the similarity of words that appear in similar contexts and use some bilingual regularization in order to (2) maximize the similarity of translation pairs. In this work, we concentrate on a subset of these bilingual embedding methods that perform a post-hoc mapping to a bilingual space, which we refer to as embedding projection methods. One of the main advantages of these methods is that they make better use of small amounts of parallel data than MT systems, even enabling unsupervised machine translation Artetxe2018,Lample2018.", + "With this paper, we provide the first extensive evaluation of cross-lingual embeddings for targeted sentiment tasks. We formulate the task of targeted sentiment analysis as classification, given the targets from an oracle. The question we attempt to address is how to infer the polarity of a sentiment target in a language that does not have any annotated sentiment data or parallel corpora with a resource-rich language. In the following Catalan sentence, for example, how can we determine that the sentiment of \u201cservei\u201d is negative, while that of \u201cmenjar\u201d is positive if we do not have annotated data in Catalan or parallel data for English-Catalan?", + "El servei al restaurant va ser p\u00e9ssim. Al menys el menjar era bo.", + "Specifically, we propose an approach which requires (1) minimal bilingual data and instead makes use of (2) high-quality monolingual word embeddings in the source and target language. We take an intermediate step by first testing this approach on sentence-level classification. After confirming that our approach performs well at sentence-level, we propose a targeted model with the same data requirements. The main contributions are that we", + "compare projection-based cross-lingual methods to MT,", + "extend previous cross-lingual approaches to enable targeted cross-lingual sentiment analysis with minimal parallel data requirements,", + "compare different model architectures for cross-lingual targeted sentiment analysis,", + "perform a detailed error analysis, and detailing the advantages and disadvantages of each method,", + "and, finally, deploy the methods in a realistic case-study to analyze their suitability beyond applications on (naturally) limited language pairs.", + "In addition, we make our code and data publicly available at https://github.com/jbarnesspain/targeted_blse to support future research. The rest of the article is organized as follows: In Section \"Previous Work\" , we detail related work and motivate the need for a different approach. In Section \"Projecting Sentiment Across Languages\" , we describe both the sentence-level and targeted projection approaches that we propose. In Section \"Experiments\" , we detail the resources and experimental setup for both sentence and targeted classification. In Section \"Results\" , we describe the results of the two experiments, as well as perform a detailed error analysis. In Section \"Case Study: Real World Deployment\" , we perform a case study whose purpose is to give a more qualitative view of the models. Finally, we discuss the implications of the results in Section \"Conclusion\" ." + ], + [ + "Sentiment analysis has become an enormously popular task with a focus on classification approaches on individual languages, but there has not been as much work on cross-lingual approaches. In this section, we detail the most relevant work on cross-lingual sentiment analysis and lay the basis for the bilingual embedding approach we propose later." + ], + [ + "Early work in cross-lingual sentiment analysis found that machine translation (MT) had reached a point of maturity that enabled the transfer of sentiment across languages. Researchers translated sentiment lexicons Mihalcea2007,Meng2012 or annotated corpora and used word alignments to project sentiment annotation and create target-language annotated corpora Banea2008,Duh2011a,Demirtas2013,Balahur2014d.", + "Several approaches included a multi-view representation of the data Banea2010,Xiao2012 or co-training Wan2009,Demirtas2013 to improve over a naive implementation of machine translation, where only the translated version of the data is considered. There are also approaches which only require parallel data Meng2012,Zhou2016,Rasooli2017, instead of machine translation.", + "All of these approaches, however, require large amounts of parallel data or an existing high quality translation tool, which are not always available. To tackle this issue, Barnes2016 explore cross-lingual approaches for aspect-based sentiment analysis, comparing machine translation methods and those that instead rely on bilingual vector representations. They conclude that MT approaches outperform current bilingual representation methods.", + "Chen2016 propose an adversarial deep averaging network, which trains a joint feature extractor for two languages. They minimize the difference between these features across languages by learning to fool a language discriminator. This requires no parallel data, but does require large amounts of unlabeled data and has not been tested on fine-grained sentiment analysis." + ], + [ + "Recently proposed bilingual embedding methods Hermann2014,Chandar2014,Gouws2015 offer a natural way to bridge the language gap. These particular approaches to bilingual embeddings, however, also require large parallel corpora in order to build the bilingual space, which gives no advantage over machine translation. Another approach to creating bilingual word embeddings, which we refer to as Projection-based Bilingual Embeddings, has the advantage of requiring relatively little parallel training data while taking advantage of larger amounts of monolingual data. In the following, we describe the most relevant approaches.", + "Mikolov2013translation find that vector spaces in different languages have similar arrangements. Therefore, they propose a linear projection which consists of learning a rotation and scaling matrix. Artetxe2016,Artetxe2017 improve upon this approach by requiring the projection to be orthogonal, thereby preserving the monolingual quality of the original word vectors.", + "Given source embeddings $S$ , target embeddings $T$ , and a bilingual lexicon $L$ , Artetxe2016 learn a projection matrix $W$ by minimizing the square of Euclidean distances ", + "$$\\operatornamewithlimits{arg\\,min}_W \\sum _{i} ||S^{\\prime }W-T^{\\prime }||_{F}^{2}\\,,$$ (Eq. 13) ", + "where $S^{\\prime } \\in S$ and $T^{\\prime } \\in T$ are the word embedding matrices for the tokens in the bilingual lexicon $L$ . This is solved using the Moore-Penrose pseudoinverse $S^{\\prime +} = (S^{\\prime T}S^{\\prime })^{-1}S^{\\prime T}$ as $ W =\nS^{\\prime +}T^{\\prime }$ , which can be computed using SVD. We refer to this approach as VecMap.", + "Lample2017 propose a similar refined orthogonal projection method to Artetxe2017, but include an adversarial discriminator, which seeks to discriminate samples from the projected space $WS$ , and the target $T$ , while the projection matrix $W$ attempts to prevent this making the projection from the source space $WS$ as similar to the target space $T$ as possible.", + "They further refine their projection matrix by reducing the hubness problem Dinu2015, which is commonly found in high-dimensional spaces. For each projected embedding $Wx$ , they define the $k$ nearest neighbors in the target space, $\\mathcal {N}_{T}$ , suggesting $k = 10$ . They consider the mean cosine similarity $r_{T}(Wx)$ between a projected embedding $Wx$ and its $k$ nearest neighbors ", + "$$r_{T}(Wx) = \\frac{1}{k} \\sum _{y \\in \\mathcal {N}_{T}(Wx) } \\cos (Wx,y)$$ (Eq. 15) ", + "as well as the mean cosine of a target word $y$ to its neighborhood, which they denote by $r_{S}$ .", + "In order to decrease similarity between mapped vectors lying in dense areas, they introduce a cross-domain similarity local scaling term (CSLS) ", + "$$\\textrm {CSLS}(Wx,y) = 2 \\cos (Wx,y) - r_{T}(Wx) - r_{S}(y)\\,,$$ (Eq. 16) ", + "which they find improves accuracy, while not requiring any parameter tuning.", + "Gouws2015taskspecific propose a method to create a pseudo-bilingual corpus with a small task-specific bilingual lexicon, which can then be used to train bilingual embeddings (Barista). This approach requires a monolingual corpus in both the source and target languages and a set of translation pairs. The source and target corpora are concatenated and then every word is randomly kept or replaced by its translation with a probability of 0.5. Any kind of word embedding algorithm can be trained with this pseudo-bilingual corpus to create bilingual word embeddings." + ], + [ + "Maas2011 first explored the idea of incorporating sentiment information into semantic word vectors. They proposed a topic modeling approach similar to latent Dirichlet allocation in order to collect the semantic information in their word vectors. To incorporate the sentiment information, they included a second objective whereby they maximize the probability of the sentiment label for each word in a labeled document.", + "Tang2014 exploit distantly annotated tweets to create Twitter sentiment embeddings. To incorporate distributional information about tokens, they use a hinge loss and maximize the likelihood of a true $n$ -gram over a corrupted $n$ -gram. They include a second objective where they classify the polarity of the tweet given the true $n$ -gram. While these techniques have proven useful, they are not easily transferred to a cross-lingual setting.", + "Zhou2015 create bilingual sentiment embeddings by translating all source data to the target language and vice versa. This requires the existence of a machine translation system, which is a prohibitive assumption for many under-resourced languages, especially if it must be open and freely accessible. This motivates approaches which can use smaller amounts of parallel data to achieve similar results." + ], + [ + "The methods discussed so far focus on classifying textual phrases like documents or sentences. Next to these approaches, others have concentrated on classifying aspects HuandLiu2004,Liu2012,Pontiki2014 or targets Zhang2015,Zhang2016,Tang2016 to assign them with polarity values.", + "A common technique when adapting neural architectures to targeted sentiment analysis is to break the text into left context, target, and right context Zhang2015,Zhang2016, alternatively keeping the target as the final/beginning token in the respective contexts Tang2016. The model then extracts a feature vector from each context and target, using some neural architecture, and concatenates the outputs for classification.", + "More recent approaches attempt to augment a neural network with memory to model these interactions Chen2017,Xue2018,Wang2018,Liu2018. Wang2017 explore methods to improve classification of multiple aspects in tweets, while Akhtar2018 attempt to use cross-lingual and multilingual data to improve aspect-based sentiment analysis in under-resourced languages.", + "As mentioned before, MT has traditionally been the main approach for transferring information across language barriers BIBREF0 . But this is particularly problematic for targeted sentiment analysis, as changes in word order or loss of words created during translation can directly affect the performance of a classifier Lambert2015." + ], + [ + "In this section, we propose a novel approach to incorporate sentiment information into bilingual embeddings, which we first test on sentence-level cross-lingual sentiment classification. We then propose an extension in order to adapt this approach to targeted cross-lingual sentiment classification. Our model, Bilingual Sentiment Embeddings (Blse), are embeddings that are jointly optimized to represent both (a) semantic information in the source and target languages, which are bound to each other through a small bilingual dictionary, and (b) sentiment information, which is annotated on the source language only. We only need three resources: (1) a comparably small bilingual lexicon, (2) an annotated sentiment corpus in the resource-rich language, and (3) monolingual word embeddings for the two involved languages." + ], + [ + "In this section, we detail the projection objective, the sentiment objective, and finally the full objective for sentence-level cross-lingual sentiment classification. A sketch of the full sentence-level model is depicted in Figure 1 .", + "We assume that we have two precomputed vector spaces $S = \\mathbb {R}^{v \\times d}$ and $T = \\mathbb {R}^{v^{\\prime } \\times d^{\\prime }}$ for our source and target languages, where $v$ ( $v^{\\prime }$ ) is the length of the source vocabulary (target vocabulary) and $d$ ( $d^{\\prime }$ ) is the dimensionality of the embeddings. We also assume that we have a bilingual lexicon $L$ of length $n$ which consists of word-to-word translation pairs $L$ = $\\lbrace (s_{1},t_{1}),\n(s_{2},t_{2}),\\ldots , (s_{n}, t_{n})\\rbrace $ which map from source to target.", + "In order to create a mapping from both original vector spaces $S$ and $T$ to shared sentiment-informed bilingual spaces $\\mathbf {z}$ and $\\mathbf {\\hat{z}}$ , we employ two linear projection matrices, $M$ and $M^{\\prime }$ . During training, for each translation pair in $L$ , we first look up their associated vectors, project them through their associated projection matrix and finally minimize the mean squared error of the two projected vectors. This is similar to the approach taken by Mikolov2013translation , but includes an additional target projection matrix.", + "The intuition for including this second matrix is that a single projection matrix does not support the transfer of sentiment information from the source language to the target language. Without $M^{\\prime }$ , any signal coming from the sentiment classifier (see Section UID27 ) would have no affect on the target embedding space $T$ , and optimizing $M$ to predict sentiment and projection would only be detrimental to classification of the target language. We analyze this further in Section UID63 . Note that in this configuration, we do not need to update the original vector spaces, which would be problematic with such small training data.", + "The projection quality is ensured by minimizing the mean squared error ", + "$$\\textrm {MSE} = \\dfrac{1}{n} \\sum _{i=1}^{n} (\\mathbf {z_{i}} - \\mathbf {\\hat{z}_{i}})^{2}\\,,$$ (Eq. 26) ", + "where $\\mathbf {z_{i}} = S_{s_{i}} \\cdot M$ is the dot product of the embedding for source word $s_{i}$ and the source projection matrix and $\\mathbf {\\hat{z}_{i}} = T_{t_{i}} \\cdot M^{\\prime }$ is the same for the target word $t_{i}$ .", + "We add a second training objective to optimize the projected source vectors to predict the sentiment of source phrases. This inevitably changes the projection characteristics of the matrix $M$ , and consequently $M^{\\prime }$ and encourages $M^{\\prime }$ to learn to predict sentiment without any training examples in the target language.", + "In order to train $M$ to predict sentiment, we require a source-language corpus $C_{\\textrm {source}}= \\lbrace (x_{1}, y_{1}),\n(x_{2}, y_{2}), \\ldots , (x_{i}, y_{i})\\rbrace $ where each sentence $x_{i}$ is associated with a label $y_{i}$ .", + "For classification, we use a two-layer feed-forward averaging network, loosely following Iyyer2015 . For a sentence $x_{i}$ we take the word embeddings from the source embedding $S$ and average them to $\\mathbf {a}_{i} \\in \\mathbb {R}^{d}$ . We then project this vector to the joint bilingual space $\\mathbf {z}_{i} = \\mathbf {a}_{i} \\cdot M$ . Finally, we pass $\\mathbf {z}_{i}$ through a softmax layer $P$ to obtain the prediction $\\hat{y}_{i} = \\textrm {softmax} ( \\mathbf {z}_{i} \\cdot P)$ .", + "To train our model to predict sentiment, we minimize the cross-entropy error of the predictions", + "$$H = - \\sum _{i=1}^{n} y_{i} \\log \\hat{y_{i}} - (1 - y_{i}) \\log (1 - \\hat{y_{i}})\\,.$$ (Eq. 29) ", + "In order to jointly train both the projection component and the sentiment component, we combine the two loss functions to optimize the parameter matrices $M$ , $M^{\\prime }$ , and $P$ by ", + "$$J =\\hspace{-14.22636pt}\\sum _{(x,y) \\in C_{\\textrm {source}}}\\hspace{2.84526pt}\\sum _{(s,t) \\in L}\\hspace{0.0pt}\\alpha H(x,y)\n+ (1 - \\alpha ) \\cdot \\textrm {MSE}(s,t)\\,,$$ (Eq. 31) ", + "where $\\alpha $ is a hyperparameter that weights sentiment loss vs. projection loss.", + "For inference, we classify sentences from a target-language corpus $C_{\\textrm {target}}$ . As in the training procedure, for each sentence, we take the word embeddings from the target embeddings $T$ and average them to $\\mathbf {a}_{i} \\in \\mathbb {R}^{d}$ . We then project this vector to the joint bilingual space $\\mathbf {\\hat{z}}_{i} = \\mathbf {a}_{i} \\cdot M^{\\prime }$ . Finally, we pass $\\mathbf {\\hat{z}}_{i}$ through a softmax layer $P$ to obtain the prediction $\\hat{y}_{i} = \\textrm {softmax} (\n\\mathbf {\\hat{z}}_{i} \\cdot P)$ ." + ], + [ + "In our targeted model, we assume that the list of sentiment targets as they occur in the text is given. These can be extracted previously either by using domain knowledge Liu2005, by using a named entity recognizer Zhang2015 or by using a number of aspect extraction techniques Zhou2012. Given these targets, the task is reduced to classification. However, what remains is how to represent the target, to learn to subselect the information from the context which is relevant, how to represent this contextual information, and how to combine these representations in a meaningful way that enables us to classify the target reliably.", + "Our approach to adapt the Blse model to targeted sentiment analysis, which we call Split (depicted in Figure 2 ), is similar to the method proposed by Zhang2016 for gated recurrent networks. For a sentence with a target $a$ , we split the sentence at $a$ in order to get a left and right context, $\\textrm {con}_\\ell (a)$ and $\\textrm {con}_r(a)$ respectively.", + "Unlike the approach from Zhang2016, we do not use recurrent neural networks to create a feature vector, as Atrio2019 showed that, in cross-lingual setups, they overfit too much to word order and source-language specific information to perform well on our tasks. Therefore, we instead average each left context $\\textrm {con}_\\ell (a_i)$ , right context $\\textrm {con}_r(a_i)$ , and target $a_{i}$ separately. Although averaging is a simplified approach to create a compositional representation of a phrase, it has been shown to work well for sentiment Iyyer2015,Barnes2017. After creating a single averaged vector for the left context, right context, and target, we concatenate them and use these as input for the softmax classification layer $T \\in \\mathbb {R}^{d \\times 3}$ , where $d$ is the dimensionality of the input vectors. The model is trained on the source language sentiment data using $M$ to project, and then tested by replacing $M$ with $M^{^{\\prime }}$ , similar to the sentence-level model." + ], + [ + "In this section, we describe the resources and datasets, as well as the experimental setups used in both the sentence-level (Experiment 1 in Subsection \"Setting for Experiment 1: Sentence-level Classification\" ) and targeted (Experiment 2 in Subsection \"Setting for Experiment 2: Targeted Classification\" ) experiments." + ], + [ + "The number of datasets and resources for under-resourced languages are limited. Therefore, we choose a mixture of resource-rich and under-resourced languages for our experiments. We treat the resource-rich languages as if they were under-resourced by using similar amounts of parallel data.", + "To evaluate our proposed model at sentence-level, we conduct experiments using four benchmark datasets and three bilingual combinations. We use the OpeNER English and Spanish datasets Agerri2013 and the MultiBooked Catalan and Basque datasets BIBREF1 . All datasets contain hotel reviews which are annotated for targeted sentiment analysis. The labels include Strong Negative ( $--$ ), Negative ( $-$ ), Positive ( $+$ ), and Strong Positive ( $++$ ). We map the aspect-level annotations to sentence level by taking the most common label and remove instances of mixed polarity. We also create a binary setup by combining the strong and weak classes. This gives us a total of six experiments. The details of the sentence-level datasets are summarized in Table 1 .", + "For each of the experiments, we take 70 percent of the data for training, 20 percent for testing and the remaining 10 percent are used as development data for tuning meta-parameters.", + "We use the following corpora to set up the experiments in which we train on a source language corpus $C_{S}$ and test on a target language corpus $C_{T}$ . Statistics for all of the corpora are shown in Table 3 . We include a binary classification setup, where neutral has been removed and strong positive and strong negative have been mapped to positive and negative, as well as a multiclass setup, where the original labels are used.", + "OpeNER Corpora: The OpeNER corpora Agerri2013 are composed of hotel reviews, annotated for aspect-based sentiment. Each aspect is annotated with a sentiment label (Strong Positive, Positive, Negative, Strong Negative). We perform experiments with the English and Spanish versions.", + "MultiBooked Corpora: The MultiBooked corpora Barnes2018a are also hotel reviews annotated in the same way as the OpeNER corpora, but in Basque and Catalan. These corpora allow us to observe how well each approach performs on low-resource languages.", + "SemEval 2016 Task 5: We take the English and Spanish restaurant review corpora made available by the organizers of the SemEval event Pontiki2016. These corpora are annotated for three levels of sentiment (positive, neutral, negative).", + "USAGE Corpora: The USAGE corpora Klinger2014a are Amazon reviews taken from a number of different items, and are available in English and German. Each aspect is annotated for three levels of sentiment (positive, neutral, negative). As the corpus has two sets of annotations available, we take the annotations from annotator 1 as the gold standard.", + "For Blse, VecMap, Muse, and MT, we require monolingual vector spaces for each of our languages. For English, we use the publicly available GoogleNews vectors. For Spanish, Catalan, and Basque, we train skip-gram embeddings using the Word2Vec toolkit with 300 dimensions, subsampling of $10^{-4}$ , window of 5, negative sampling of 15 based on a 2016 Wikipedia corpus (sentence-split, tokenized with IXA pipes Agerri2014 and lowercased). The statistics of the Wikipedia corpora are given in Table 2 .", + "For Blse, VecMap, Muse, and Barista, we also require a bilingual lexicon. We use the sentiment lexicon from HuandLiu2004 (to which we refer in the following as Hu and Liu) and its translation into each target language. We translate the lexicon using Google Translate and exclude multi-word expressions. This leaves a dictionary of 5700 translations in Spanish, 5271 in Catalan, and 4577 in Basque. We set aside ten percent of the translation pairs as a development set in order to check that the distances between translation pairs not seen during training are also minimized during training." + ], + [ + "We compare Blse (Sections UID23 \u2013 UID30 ) to VecMap, Muse, and Barista (Section \"Previous Work\" ) as baselines, which have similar data requirements and to machine translation (MT) and monolingual (Mono) upper bounds which request more resources. For all models (Mono, MT, VecMap, Muse, Barista), we take the average of the word embeddings in the source-language training examples and train a linear SVM. We report this instead of using the same feed-forward network as in Blse as it is the stronger upper bound. We choose the parameter $c$ on the target language development set and evaluate on the target language test set.", + "Upper Bound Mono. We set an empirical upper bound by training and testing a linear SVM on the target language data. Specifically, we train the model on the averaged embeddings from target language training data, tuning the $c$ parameter on the development data. We test on the target language test data.", + "Upper Bound MT. To test the effectiveness of machine translation, we translate all of the sentiment corpora from the target language to English using the Google Translate API. Note that this approach is not considered a baseline, as we assume not to have access to high-quality machine translation for low-resource languages of interest.", + "Baseline Unsup We compare with the unsupervised statistical machine translation approach proposed by artetxe2018emnlp. This approach uses a self-supervised method to create bilingual phrase embeddings which then populates a phrase table. Monolingual n-gram language models and an unsupervised variant of MERT are used to create a MT model which is improved through iterative backtranslation. We use the Wikipedia corpora from Section UID42 to create the unsupervised SMT system between English and the target languages and run the training proceedure with default parameters. Finally, we translate all test examples in the target languages to English.", + "Baseline VecMap. We compare with the approach proposed by Artetxe2016 which has shown promise on other tasks, e. g., word similarity. In order to learn the projection matrix $W$ , we need translation pairs. We use the same word-to-word bilingual lexicon mentioned in Section UID23 . We then map the source vector space $S$ to the bilingual space $\\hat{S} = SW$ and use these embeddings.", + "Baseline Muse. This baseline is similar to VecMap but incorporates and adversarial objective as well as a localized scaling objective, which further improve the orthogonal refinement so that the two language spaces are even more similar.", + "Baseline Barista. The approach proposed by Gouws2015taskspecific is another appropriate baseline, as it fulfills the same data requirements as the projection methods. The bilingual lexicon used to create the pseudo-bilingual corpus is the same word-to-word bilingual lexicon mentioned in Section UID23 . We follow the authors' setup to create the pseudo-bilingual corpus. We create bilingual embeddings by training skip-gram embeddings using the Word2Vec toolkit on the pseudo-bilingual corpus using the same parameters from Section UID42 .", + "Our method: BLSE. Our model, Blse, is implemented in Pytorch Pytorch and the word embeddings are initialized with the pretrained word embeddings $S$ and $T$ mentioned in Section UID42 . We use the word-to-word bilingual lexicon from Section UID46 , tune the hyperparameters $\\alpha $ , training epochs, and batch size on the target development set and use the best hyperparameters achieved on the development set for testing. ADAM Kingma2014a is used in order to minimize the average loss of the training batches.", + "Ensembles. In order to evaluate to what extent each projection model adds complementary information to the machine translation approach, we create an ensemble of MT and each projection method (Blse, VecMap, Muse, Barista). A random forest classifier is trained on the predictions from MT and each of these approaches." + ], + [ + "For the targeted classification experiment, we compare the same models mentioned above, but adapted to the setting using the Split method from Section \"Targeted Model\" .", + "A simple majority baseline sets the lower bound, while the MT-based model serves as an upper bound. We assume our models to perform between these two, as we do not have access to the millions of parallel sentences required to perform high-quality MT and particularly aim at proposing a method which is less resource-hungry.", + "We hypothesize that cross-lingual approaches are particularly error-prone when evaluative phrases and words are wrongly predicted. In such settings, it might be beneficial for a model to put emphasis on the target word itself and learn a prior distribution of sentiment for each target independent of the context. For example, if you assume that all mentions of Steven Segal are negative in movie reviews, it is possible to achieve good results Bird2009. On the other hand, it may be that there are not enough examples of target-context pairs, and that it is better to ignore the target and concentrate only on the contexts.", + "To analyze this, we compare our model to two simplified versions. In addition, this approach enables us to gain insight in the source of relevant information. The first is Target-only, which means that we use the model in the same way as before but ignore the context completely. This serves as a tool to understand how much model performance originates from the target itself.", + "In the same spirit, we use a Context-only model, which ignores the target by constraining the parameters of all target phrase embeddings to be the same. This approach might be beneficial over our initial model if the prior distribution between targets was similar and the context actually carries the relevant information.", + "As the baseline for each projection method, we assume all targets in each sentence respectively to be of the same polarity (Sent). This is generally an erroneous assumption, but can give good results if all of the targets in a sentence have the same polarity. In addition, this baseline provides us with the information about whether the models are able to handle information from different positions in the text." + ], + [ + "In Table 4 , we report the results of all four methods. Our method outperforms the other projection methods (the baselines VecMap, Muse, and Barista) on four of the six experiments substantially. It performs only slightly worse than the more resource-costly upper bounds (MT and Mono). This is especially noticeable for the binary classification task, where Blse performs nearly as well as machine translation and significantly better than the other methods. Unsup also performs similarly to Blse on the binary tasks, while giving stronger performance on the 4-class setup. We perform approximate randomization tests Yeh2000 with 10,000 runs and highlight the results that are statistically significant (*p $<$ 0.01) in Table 4 .", + "In more detail, we see that MT generally performs better than the projection methods (79\u201369 $\\text{F}_1$ on binary, 52\u201344 on 4-class). Blse (75\u201369 on binary, 41\u201330 on 4-class) has the best performance of the projection methods and is comparable with MT on the binary setup, with no significant difference on binary Basque. VecMap (67\u201346 on binary, 35\u201321 on 4-class) and Barista (61\u201355 on binary, 40\u201334 on 4-class) are significantly worse than Blse on all experiments except Catalan and Basque 4-class. Muse (67\u201362 on binary, 45\u201334 on 4-class) performs better than VecMap and Barista. On the binary experiment, VecMap outperforms Barista on Spanish (67.1 vs. 61.2) and Catalan (60.7 vs. 60.1) but suffers more than the other methods on the four-class experiments, with a maximum $\\text{F}_1$ of 34.9. Barista is relatively stable across languages. Unsup performs well across experiments (76\u201365 on binary, 49\u201339 on 4-class), even performing better than MT on both Catalan tasks and Spanish 4-class.", + "The Ensemble of MT and Blse performs the best, which shows that Blse adds complementary information to MT. Finally, we note that all systems perform worse on Basque. This is presumably due to the increased morphological complexity of Basque, as well as its lack of similarity to the source language English (Section UID102 ).", + "We analyze three aspects of our model in further detail: 1) where most mistakes originate, 2) the effect of the bilingual lexicon, and 3) the effect and necessity of the target-language projection matrix $M^{\\prime }$ .", + "In order to analyze where each model struggles, we categorize the mistakes and annotate all of the test phrases with one of the following error classes: vocabulary (voc), adverbial modifiers (mod), negation (neg), external knowledge (know) or other. Table 5 shows the results.", + "Vocabulary: The most common way to express sentiment in hotel reviews is through the use of polar adjectives (as in \u201cthe room was great\u201d) or the mention of certain nouns that are desirable (\u201cit had a pool\u201d). Although this phenomenon has the largest total number of mistakes (an average of 72 per model on binary and 172 on 4-class), it is mainly due to its prevalence. MT performed the best on the test examples which according to the annotation require a correct understanding of the vocabulary (81 $\\text{F}_1$ on binary /54 $\\text{F}_1$ on 4-class), with Blse (79/48) slightly worse. Muse (76/23), VecMap (70/35), and Barista (67/41) perform worse. This suggests that Blse is better than Muse, VecMap and Barista at transferring sentiment of the most important sentiment bearing words.", + "Negation: Negation is a well-studied phenomenon in sentiment analysis Pang2002,Wiegand2010,Zhu2014,Reitan2015 . Therefore, we are interested in how these four models perform on phrases that include the negation of a key element, for example \u201cIn general, this hotel isn't bad\". We would like our models to recognize that the combination of two negative elements \u201cisn't\" and \u201cbad\" lead to a Positive label.", + "Given the simple classification strategy, all models perform relatively well on phrases with negation (all reach nearly 60 $\\text{F}_1$ in the binary setting). However, while Blse performs the best on negation in the binary setting (82.9 $\\text{F}_1$ ), it has more problems with negation in the 4-class setting (36.9 $\\text{F}_1$ ).", + "Adverbial Modifiers: Phrases that are modified by an adverb, e. g., the food was incredibly good, are important for the four-class setup, as they often differentiate between the base and Strong labels. In the binary case, all models reach more than 55 $\\text{F}_1$ . In the 4-class setup, Blse only achieves 27.2 $\\text{F}_1$ compared to 46.6 or 31.3 of MT and Barista, respectively. Therefore, presumably, our model does currently not capture the semantics of the target adverbs well. This is likely due to the fact that it assigns too much sentiment to functional words (see Figure 6 ). Muse performs poorly on modified examples (20.3 $\\text{F}_1$ ).", + "External Knowledge Required: These errors are difficult for any of the models to get correct. Many of these include numbers which imply positive or negative sentiment (350 meters from the beach is Positive while 3 kilometers from the beach is Negative). Blse performs the best (63.5 $\\text{F}_1$ ) while MT performs comparably well (62.5). Barista performs the worst (43.6).", + "Binary vs. 4-class: All of the models suffer when moving from the binary to 4-class setting; an average of 26.8 in macro $\\text{F}_1$ for MT, 31.4 for VecMap, 22.2 for Barista, 34.1 for Muse, and 36.6 for Blse. The vector projection methods (VecMap, Muse, and Blse) suffer the most, suggesting that they are currently more apt for the binary setting.", + "We analyze how the number of translation pairs affects our model. We train on the 4-class Spanish setup using the best hyper-parameters from the previous experiment.", + "Research into projection techniques for bilingual word embeddings Mikolov2013translation,Lazaridou2015,Artetxe2016 often uses a lexicon of the most frequent 8\u201310 thousand words in English and their translations as training data. We test this approach by taking the 10,000 word-to-word translations from the Apertium English-to-Spanish dictionary. We also use the Google Translate API to translate the NRC hashtag sentiment lexicon Mohammad2013 and keep the 22,984 word-to-word translations. We perform the same experiment as above and vary the amount of training data from 0, 100, 300, 600, 1000, 3000, 6000, 10,000 up to 20,000 training pairs. Finally, we compile a small hand translated dictionary of 200 pairs, which we then expand using target language morphological information, finally giving us 657 translation pairs. The macro $\\text{F}_1$ score for the Hu and Liu dictionary climbs constantly with the increasing translation pairs. Both the Apertium and NRC dictionaries perform worse than the translated lexicon by Hu and Liu, while the expanded hand translated dictionary is competitive, as shown in Figure 3 .", + "While for some tasks, e. g., bilingual lexicon induction, using the most frequent words as translation pairs is an effective approach, for sentiment analysis, this does not seem to help. Using a translated sentiment lexicon, even if it is small, gives better results.", + "The main motivation for using two projection matrices $M$ and $M^{\\prime }$ is to allow the original embeddings to remain stable, while the projection matrices have the flexibility to align translations and separate these into distinct sentiment subspaces. To justify this design decision empirically, we perform an experiment to evaluate the actual need for the target language projection matrix $M^{\\prime }$ : We create a simplified version of our model without $M^{\\prime }$ , using $M$ to project from the source to target and then $P$ to classify sentiment.", + "The results of this model are shown in Figure 4 . The modified model does learn to predict in the source language, but not in the target language. This confirms that $M^{\\prime }$ is necessary to transfer sentiment in our model.", + "Additionally, we provide an analysis of a similar model to ours, but which uses $M = \\mathbb {R}^{d, o}$ and $M^{\\prime } = \\mathbb {R}^{d^{\\prime }, o}$ , where $d$ ( $d^{\\prime }$ ) is the dimensionality of the original embeddings and $o$ is the label size, to directly model crosslingual sentiment, such that the final objective function is ", + "$$J =\\hspace{-14.22636pt}\\sum _{(x,y) \\in C_{\\textrm {source}}}\\hspace{2.84526pt}\\sum _{(s,t) \\in L}\\hspace{0.0pt}\\alpha \\cdot H(x, y) + (1 - \\alpha ) \\cdot || M \\cdot s - M^{\\prime } \\cdot t ||$$ (Eq. 66) ", + "thereby simplifying the model and removing the $P$ parameter. Table 6 shows that Blse outperforms this simplified model on all tasks.", + "In order to understand how well our model transfers sentiment information to the target language, we perform two qualitative analyses. First, we collect two sets of 100 positive sentiment words and one set of 100 negative sentiment words. An effective cross-lingual sentiment classifier using embeddings should learn that two positive words should be closer in the shared bilingual space than a positive word and a negative word. We test if Blse is able to do this by training our model and after every epoch observing the mean cosine similarity between the sentiment synonyms and sentiment antonyms after projecting to the joint space.", + "We compare Blse with VecMap and Barista by replacing the Linear SVM classifiers with the same multi-layer classifier used in Blse and observing the distances in the hidden layer. Figure 5 shows this similarity in both source and target language, along with the mean cosine similarity between a held-out set of translation pairs and the macro $\\text{F}_1$ scores on the development set for both source and target languages for Blse, Barista, and VecMap. From this plot, it is clear that Blse is able to learn that sentiment synonyms should be close to one another in vector space and antonyms should have a negative cosine similarity. While the other models also learn this to some degree, jointly optimizing both sentiment and projection gives better results.", + "Secondly, we would like to know how well the projected vectors compare to the original space. Our hypothesis is that some relatedness and similarity information is lost during projection. Therefore, we visualize six categories of words in t-SNE, which projects high dimensional representations to lower dimensional spaces while preserving the relationships as best as possible Vandermaaten2008: positive sentiment words, negative sentiment words, functional words, verbs, animals, and transport.", + "The t-SNE plots in Figure 6 show that the positive and negative sentiment words are rather clearly separated after projection in Blse. This indicates that we are able to incorporate sentiment information into our target language without any labeled data in the target language. However, the downside of this is that functional words and transportation words are highly correlated with positive sentiment.", + "Finally, in order to analyze the sensitivity of the alpha parameter, we train Blse models for 30 epochs each with $\\alpha $ between 0 and 1. Figure 7 shows the average cosine similarity for the translation pairs, as well as macro $\\text{F}_1$ for both source and target language development data.", + "Values near 0 lead to poor translation and consecuently poor target language transfer. There is a rather large \u201csweet spot\u201d where all measures perform best and finally, the translation is optimized to the detriment of sentiment prediction in both source and target languages with values near 1.", + "The experiments in this section have proven that it is possible to perform cross-lingual sentiment analysis without machine translation, and that jointly learning to project and predict sentiment is advantageous. This supports the growing trend of jointly training for multiple objectives Tang2014,Klinger2015,Ferreira2016.", + "This approach has also been exploited within the framework of multi-task learning, where a model learns to perform multiple similar tasks in order to improve on a final task Collobert2011a. The main difference between the joint method proposed here and multi-task learning is that vector space projection and sentiment classification are not similar enough tasks to help each other. In fact, these two objectives compete against one another, as a perfect projection would not contain enough information for sentiment classification, and vice versa." + ], + [ + "Table 7 shows the macro $\\text{F}_1$ scores for all cross-lingual approaches (Blse, VecMap, Muse, Barista, MT, Unsup) and all targeted approaches (Sent, Split, Context-only, and Target-only). The final column is the average over all corpora. The final row in each setup shows the macro $\\text{F}_1$ for a classifier that always chooses the majority class.", + "Blse outperforms other projection methods on the binary setup, 63.0 macro averaged $\\text{F}_1$ across corpora versus 59.0, 57.9, and 51.4 for VecMap, Muse, and Barista, respectively. On the multiclass setup, however, Muse (32.2 $\\text{F}_1$ ) is the best, followed by VecMap (31.0), Barista (28.1) and Blse (23.7). Unsup performs well across all experiments, achieving the best results on OpeNER ES (73.2 on binary and 42.7 on multiclass) and SemEval binary (77.1). VecMap is never the best nor the worst approach. In general, Barista performs poorly on the binary setup, but slightly better on the multiclass, although the overall performance is still weak. These results are similar to those observed in Experiment 1 for sentence classification.", + "The Split approach to ABSA improves over the Sent baseline on 33 of 50 experiments, especially on binary (21/25), while on multiclass it is less helpful (13/25). Both Sent and Split normally outperform Context-only or Target-only approaches. This confirms the intuition that it is important to take both context and target information for classification. Additionally, the Context-only approach always performs better than Target-only, which indicates that context is more important than the prior probability of an target being positive or negative.", + "Unlike the projection methods, MT using only the Sent representation performs well on the OpeNER and MultiBooked datasets, while suffering more on the SemEval and USAGE datasets. This is explained by the percentage of sentences that contain contrasting polarities in each dataset: between 8 and 12% for the OpeNER and Multibooked datasets, compared to 29% for SemEval or 50% for USAGE. In sentences with multiple contrasting polarities, the Sent baseline performs poorly.", + "Finally, the general level of performance of projection-based targeted cross-lingual sentiment classification systems shows that they still lag 10+ percentage points behind MT on binary (compare MT (72.9 $\\text{F}_1$ ) with Blse (63.0)), and 6+ percentage points on multiclass (MT (38.8) versus Muse (32.2)). The gap between MT and projection-based approaches is therefore larger on targeted sentiment analysis than at sentence-level.", + "We perform a manual analysis of the targets misclassified by all systems on the OpeNER Spanish binary corpus (see Table 8 ), and found that the average length of misclassified targets is slightly higher than that of correctly classified targets, except for with VecMap. This indicates that averaging may have a detrimental effect as the size of the targets increases.", + "With the MT upperbounds, there is a non-negligible amount of noise introduced by targets which have been incorrectly translated (0.05% OpeNER ES, 6% MultiBooked EU, 2% CA, 2.5% SemEval, 1% USAGE). We hypothesize that this is why MT with Context-only performs better than MT with Split. This motivates further research with projection-based methods, as they do not suffer from translation errors.", + "The confusion matrices of the models on the SemEval task, shown in Figure 8 , show that on the multilabel task, models are not able to learn the neutral class. This derives from the large class imbalance found in the data (see Table 3 ). Similarly, models do not learn the Strong Negative class on the OpeNER and MultiBooked datasets." + ], + [ + "The performance of machine learning models on different target languages depends on the amount of data available, the quality of the data, and characteristics of the target language, e. g., morphological complexity. In the following, we analyze these aspects. There has been previous work that has observed target-language specific differences in multilingual dependency parsing Zeljko2016, machine translation Johnson2017, and language modeling Cotterell2018,Gerz2018. We are not aware of any work in cross-lingual sentiment analysis that explores the relationship between target language and performance in such depth and aim at improving this situation in the following.", + "Additionally, the effect of domain differences when performing cross-lingual tasks has not been studied in depth. Hangya2018 propose domain adaptation methods for cross-lingual sentiment classification and bilingual dictionary induction. They show that creating domain-specific cross-lingual embeddings improves the classification for English-Spanish. However, the source-language training data used to train the sentiment classifier is taken from the same domain as the target-language test data. Therefore, it is not clear what the effect of using source-language training data from different domains would be. We analyzed the model presented in Section \"Sentence-level Model\" in a domain adaptation setup, including the impact of domain differences Barnes2018c. The main result was that our model performs particularly well on more distant domains, while other approaches Chen2012,Ziser2017 performed better when the source and target domains were not too dissimilar.", + "In the following, we transfer this analysis to the target-based projection model in a real-world case study which mimics a user searching for the sentiment on touristic attractions. In order to analyze how well these methods generalize to new languages and domains, we deploy the targeted Blse, Muse, VecMap and MT models on tweets in ten Western European languages with training data from three different domains. Additionally, we include experiments with the Unsup models for a subset of the languages. English is the source language in all experiments, and we test on each of the ten target languages and attempt to answer the following research questions:", + "How much does the amount of monolingual data available to create the original embeddings effect the final results?", + "How do features of the target language, i. e. similarity to source language or morphological complexity, affect the performance?", + "How do domain mismatches between source-language training and target-language test data affect the performance?", + "Section \"Discussion\" addresses our findings regarding these questions and demonstrates that 1) the amount of monolingual data does not correlate with classification results, 2) language similarity between the source and target languages based on word and character n-gram distributions predicts the performance of Blse on new datasets, and 3) domain mismatch has more of an effect on the multiclass setup than binary." + ], + [ + "We collect tweets directed at a number of tourist attractions in European cities using the Twitter API in 10 European languages, including several under-resourced languages (English, Basque, Catalan, Galician, French, Italian, Dutch, German, Danish, Swedish, and Norwegian). We detail the data collection and annotation procedures in Section UID85 . For classification, we compare MT the best performing projection-based methods (Blse, Muse, VecMap) using the Split method, detailed further in Section UID94 . As we need monolingual embeddings for all projection-based approaches, we create skipgram embeddings from Wikipedia dumps, detailed in Section UID91 .", + "As an experimental setting to measure the effectiveness of targeted cross-lingual sentiment models on a large number of languages, we collect and annotate small datasets from Twitter for each of the target languages, as well as a larger dataset to train the models in English. While it would be possible to only concentrate our efforts on languages with existing datasets in order to enable evaluation, this could give a distorted view of how well these models generalize. In order to reduce the possible ambiguity of the tourist attractions, we do not include those that have two or more obvious senses, e. g., Barcelona could refer either to the city or the football team.", + "In order to obtain a varied sample of tweets with subjective opinions, we download tweets that contain mentions of these tourist attractions as well as one of several emoticons or keywords. This distant supervision technique has been used to create sentiment lexicons Mohammad2016, semi-supervised training data Felbo2017, and features for a classifier Turney2003. We then remove any tweets that are less than 7 words long or which contain more than 3 hashtags or mentions. This increases the probability that a tweet text contains sufficient information for our use case setting.", + "We manually annotate all tweets for its polarity toward the target to insure the quality of the data. Note that we only annotate the sentiment towards the predefined list of targets, which leads to a single annotated target per tweet. Any tweets that have unclear polarity towards the target are assigned a neutral label. This produces the three class setup that is commonly used in the SemEval tasks Nakov2013,Nakov2016. Annotators were master's and doctoral students between 27 and 35 years old. All had either native or C1 level fluency in the languages of interest. Finally, for a subset of tweets in English, Catalan, and Basque two annotators classify each tweet. Table 11 shows three example tweets from English.", + "Table 10 depicts the number of annotated targets for all languages, as well as inter-annotator agreement using Cohen's $\\kappa $ . The neutral class is the largest in all languages, followed by positive, and negative. These distributions are similar to those found in other Twitter crawled datasets Nakov2013,Nakov2016. We calculate pairwise agreement on a subset of languages using Cohen's $\\kappa $ . The scores reflect a good level of agreement (0.62, 0.60, and 0.61 for English, Basque, and Catalan, respectively).", + "We collect Wikipedia dumps for ten languages; namely, Basque, Catalan, Galician, French, Italian, Dutch, German, Danish, Swedish, and Norwegian. We then preprocess them using the Wikiextractor script, and sentence and word tokenize them with either IXA pipes Agerri2014 (Basque, Galician, Italian, Dutch, and French), Freeling Padro2010 (Catalan), or NLTK Loper2002 (Norwegian, Swedish, Danish).", + "For each language we create Skip-gram embeddings with the word2vec toolkit following the pipeline and parameters described in Section UID42 . This process gives us 300 dimensional vectors trained on similar data for all languages. We assume that any large differences in the embedding spaces derive from the size of the data and the characteristics of the language itself. Following the same criteria laid out in Section UID46 , we create projection dictionaries by translating the Hu and Liu dictionary HuandLiu2004 to each of the target languages and keeping only translations that are single word to single word. The statistics of all Wikipedia corpora, embeddings, and projection dictionaries are shown in Table 12 .", + "Since we predetermine the sentiment target for each tweet, we can perform targeted experiments without further annotation. We use the Split models described in Section \"Targeted Model\" . Our model is the targeted Blse models described in Section \"Targeted Model\" . Additionally, we compare to the targeted Muse, VecMap, and MT models, as well as an Ensemble classifier that uses the predictions from Blse and MT before taking the largest predicted class for classification (see Section \"Setting for Experiment 1: Sentence-level Classification\" for details). Finally, we set a majority baseline by assigning the most common label (neutral) to all predictions. All models are trained for 300 epochs with a learning rate of 0.001 and $\\alpha $ of 0.3.", + "We train the five models on the English data compiled during this study, as well as on the USAGE, and SemEval English data (the details can be found in Table 3 ) and test the models on the target-language test set." + ], + [ + "Table 13 shows the macro $\\text{F}_1$ scores for all cross-lingual targeted sentiment approaches (Blse, Muse, VecMap, MT) trained on English data and tested on the target-language using the Split method proposed in \"Targeted Model\" . The final column is the average over all languages. Given the results from the earlier experiments, we hypothesize that MT should outperform Muse, VecMap and Blse for most of the languages.", + "On the binary setup, Blse outperforms all other cross-lingual methods including MT and Unsup, with 56.0 macro averaged $\\text{F}_1$ across languages versus 48.7, 49.4, and 48.9 for Muse, VecMap, and MT respectively (54.1 across Basque and Catalan versus 46.0 for Unsup). Blse performs particularly well on Catalan (54.5), Italian (63.4), Swedish (65.3), and Danish (68.3). VecMap performs poorly on Galician (33.3), Italian (38.2), and Danish (43.4), but outperforms all other methods on Basque (56.4), Dutch (55.2) and Norwegian (59.0). MT performs worse than Blse and VecMap, although it does perform best for Galician (56.5). Unlike experiments in Section \"Sentence-level Model\" , the ensemble approach does not perform better than the individual classifiers and Muse leads to the classifier with the lowest performance overall. Unsup performs better than MT on both Basque and Catalan.", + "On the multiclass setup, however, MT (36.6 $\\text{F}_1$ ) is the best, followed by VecMap (34.1), Blse (32.6), and Muse (26.1). Compared to the experiments on hotel reviews, the average differences between models is small (2.5 percentage points between MT and VecMap, and 1.5 between VecMap and Blse). Unsup performs better than MT on Basque (40.1), but worse on Catalan (28.5). Again, all methods outperform the majority baseline.", + "On both the binary and multiclass setups, the best overall results are obtained by testing and training on data from the same domain (56.0 $\\text{F}_1$ for Blse and 36.6 $\\text{F}_1$ for MT). Training MT, Muse, and VecMap on the SemEval data performs better than training on USAGE, however.", + "An initial error analysis shows that all models suffer greatly on the negative class. This seems to suggest that negative polarity towards a target is more difficult to determine within these frameworks. A significant amount of the tweets that have negative polarity towards a target also express positive or neutral sentiment towards other targets. The averaging approach to create the context vectors does not currently allow any of the models to exclude this information, leading to poor performance on these instances.", + "Finally, compared to the experiments performed on hotel and product reviews in Section \"Experiments\" , the noisy data from Twitter is more difficult to classify. Despite the rather strong majority baseline (an average of 40.5 Macro $\\text{F}_1$ on binary), no model achieves more than an average of 56 Macro $\\text{F}_1$ on the binary task. A marked difference is that Blse and VecMap outperform MT on the binary setup. Unlike the previous experiment, Muse performs the worst on the multiclass setup. The other projection methods obtain multiclass results similar to the previous experiment (32.6\u201334.1 $\\text{F}_1$ here compared to 23.7\u201331.0 $\\text{F}_1$ previously)." + ], + [ + "In this section, we present an error analysis. Specifically, Table 14 shows examples where Blse correctly predicts the polarity of a tweet that MT and Unsup incorrectly predict, and vice versa, as well as examples where all models are incorrect.", + "In general, in examples where Blse outperforms MT and Unsup, the translation-based approaches often mistranslate important sentiment words, which leads to prediction errors. In the first Basque tweet, for example, \u201c#txindoki igo gabe ere inguruaz goza daiteke... zuek joan tontorrera eta utzi arraroei gure kasa...\u201d, Unsup incorrectly translates the most important sentiment word in the tweet \u201cgoza\u201d (enjoy) to \u201coverlook\u201d and subsequently incorrectly predicts that the polarity towards txindoki is negative.", + "Tweets that contain many out-of-vocabulary words or non-standard spelling (due to dialectal differences, informal writing, etc.), such as the third tweet in Table 14 , \u201ckanpora jun barik ehko asko: anboto, txindoki\u201d, are challenging for all models. In this example \u201cjun\u201d is a non-standard spelling of \u201cjoan\u201d (go), \u201cbarik\u201d is a Bizcayan Basque variant of \u201cgabe\u201d (without) , and \u201cehko\u201d is an abbreviation of \u201cEuskal Herriko\u201d (Basque Country's). These lead to poor translations for MT and Unsup, but pose a similar out-of-vocabulary problem for Blse.", + "In order to give a more qualitative view of the targeted model, Figure 9 shows t-sne projections of the bilingual vector space before and after training on the Basque binary task, following the same proceedure mentioned in Section UID68 . As in the sentence-level experiment, there is a separation of the positive and negative sentiment words, although it is less clear for targeted sentiment. This is not surprising, as a targeted model must learn not only the prior polarity of words, but how they interact with targets, leading to a more context-dependent representation of sentiment words.", + "Finally, we further analyze the effects of three variables that are present in cross-lingual sentiment analysis: a) availability of monolingual unlabeled data, b) similarity of source and target languages, and c) domain shift between the source language training data and the target language test data.", + "We pose the question of what the relationship is between the amount of available monolingual data to create the embedding spaces and the classification results of the models. If the original word embedding spaces are not of high quality, this could make it difficult for the projection-based models to create useful features. In order to test this, we perform ablation experiments by training target-language embeddings on varying amounts of data ( $1 \\times 10^{4}$ to $5 \\times 10^{9}$ tokens) and testing the models replacing the full target-language embeddings with these. We plot the performance of the models as a function of available monolingual data in Figure 10 .", + "Figure 10 shows that nearly all models, with the exception of Norwegian, perform poorly with very limited monolingual training data ( $1\\times 10^{4}$ ) and improve, although erratically, with more training data. Interestingly, the models require little data to achieve results comparable to using the all tokens to train the embeddings. A statistical analysis of the amount of unlabeled data available and the performance of Blse, Muse, VecMap (Pearson's $r$ = $-0.14$ , $-0.27$ , $0.08$ , respectively) reveals no statistically significant correlation between them. This seems to indicate that all models are not sensitive to the amount of monolingual training data available in the target language.", + "One hypothesis to different results across languages is that the similarity of the source and target language has an effect on the final classification of the models. In order to analyze this, we need a measure that models pairwise language similarity. Given that the features we use for classification are derived from distributional representations, we model similarity as a function of 1) universal POS-tag n-grams which represent the contexts used during training, and 2) character n-grams, which represent differences in morphology. POS-tag n-grams have previously been used to classify genre Fang2010, improve statistical machine translation Lioma2005, and the combination of POS-tag and character n-grams have proven useful features for identifying the native language of second language writers in English Kulmizev2017. This indicates that these are useful features for characterizing a language. In this section we calculate the pairwise similarity between all languages and then check whether this correlates with performance.", + "After POS-tagging the test sentences obtained from Twitter using the universal part of speech tags Petrov2012, we calculate the normalized frequency distribution $P_{l}$ for the POS-tag trigrams and $C_{l}$ for character trigrams for each language $l$ in $L =\n\\lbrace \\textrm {Danish, Swedish, Norwegian, Italian, Basque, Catalan,\nFrench, Dutch, Galician,}$ ", + " $\\textrm {German, English}\\rbrace $ . We then compute the pairwise cosine similarity between $\\cos (A, B) = \\frac{A\n\\cdot B}{||A|| \\: ||B||} $ where $A$ is the concatenation of $P_{l_{i}}$ and $C_{l_{i}}$ for language $l_{i}$ and $B$ is the concatenation of $P_{l_{j}}$ and $C_{l_{j}}$ for language $l_{j}$ .", + "The pairwise similarities in Figure 11 confirm to expected similarities, and language families are clearly grouped (Romance, Germanic, Scandinavian, with Basque as an outlier that has no more than 0.47 similarity with any language). This confirms the use of our similarity metric for our purposes. We plot model performance as a function of language similarity in Figure 12 . To measure the correlation between language similarity and performance, we calculate Pearson's $r$ and find that for Blse there is a strong correlation between language similarity and performance, $r = 0.76$ and significance $p <\n0.01$ . Muse, VecMap and MT do not show these correlations ( $r$ = 0.41, 0.24, 0.14, respectively). For MT this may be due to robust machine translation available in less similar languages according to our metric, e. g., German-English. For Muse and VecMap, however, it is less clear why it does not follow the same trend as Blse.", + "In this section, we determine the effect of source-language domain on the cross-lingual sentiment classification task. Specifically, we use English language training data from three different domains (Twitter, restaurant reviews, and product reviews) to train the cross-lingual classifiers, and then test on the target-language Twitter data. In monolingual sentiment analysis, one would expect to see a drop when moving to more distant domains.", + "In order to analyze the effect of domain similarity further, we test the similarity of the domains of the source-language training data using Jensen-Shannon Divergence, which is a smoothed, symmetric version of the Kullback-Leibler Divergence, $D_{KL}(A||B) = \\sum _{i}^{N} a_{i} \\log \\frac{a_{i}}{b_{i}}$ . Kullback-Leibler Divergence measures the difference between the probability distributions $A$ and $B$ , but is undefined for any event $a_{i} \\in A$ with zero probability, which is common in term distributions. Jensen-Shannon Divergence is then $\nD_{JS}(A,B) = \\frac{1}{2} \\Big [ D_{KL}(A||B) + D_{KL}(B||A) \\Big ]\\,.\n$ ", + "Our similarity features are probability distributions over terms $t\n\\in \\mathbb {R}^{|V|}$ , where $t_{i}$ is the probability of the $i$ -th word in the vocabulary $V$ . For each domain, we create frequency distributions of the most frequent 10,000 unigrams that all domains have in common and measure the divergence with $D_{JS}$ .", + "The results shown in Table 15 indicate that both the SemEval and USAGE datasets are relatively distinct from the Twitter data described in Section UID85 , while they are more similar to each other. Additionally, we plot the results of all models with respect to the training domain in Figure 13 .", + "We calculate Pearson's $r$ on the correlation between domain and model performance, shown in Table 16 . On the binary setup, the results show a negligible correlation for Blse (0.32), with no significant correlation for Muse, VecMap or MT. This suggests that the models are relatively robust to domain noise, or rather that there is so much other noise found in the approaches that domain is less relevant. On the multiclass setup, however, there is a significant effect for all models. This indicates that the multiclass models presented here are less robust than the binary models.", + "Both the SemEval and USAGE corpora differ equally from the Twitter data given the metric defined here. The fact that models trained on SemEval tend to perform better than those trained on USAGE, therefore, seems to be due to the differences in label distribution, rather than to differences in domain. These label distributions are radically different in the multiclass setup, as the English Twitter data has a 30/50/20 distribution over Positive, Neutral, and Negative labels (67/1/32 and 68/4/28 for USAGE and SemEval, respectively). Both undersampling and oversampling help, but the performance is still worse than training on in-domain data.", + "The case study which we presented in this section showed results of deploying the models from Section \"Projecting Sentiment Across Languages\" to real world Twitter data, which we collect and annotate for targeted sentiment analysis. The analysis of different phenomena revealed that for binary targeted sentiment analysis, Blse performs better than machine translation on noisy data from social media, although it is sensitive to differences between source and target languages. Finally, there is little correlation between performance on the cross-lingual sentiment task and the amount of unlabeled monolingual data used to create the original embeddings spaces which goes against our expectations.", + "Unlike the experiments in Section \"Sentence-level Model\" , the ensemble classifier employed here was not able to improve the results. We assume that the small size of the datasets in this experiment does not enable the classifier to learn which features are useful in certain contexts.", + "One common problem that appears when performing targeted sentiment analysis on noisy data from Twitter is that many of the targets of interest are ambiguous, which leads to false positives. Even with relatively unambiguous targets like \u201cBig Ben\u201d, there are a number of entities that can be referenced; Ben Rothlisberger (an American football player), an English language school in Barcelona, and many others. In order to deploy a full sentiment analysis system on Twitter data, it will be necessary to disambiguate these mentions before classifying the tweets, either as a preprocessing step or jointly.", + "In sentiment analysis, it is not yet common to test a model on multiple languages, despite the fact that current state-of-the-art models are often theoretically language-agnostic. This section shows that good performance in one language does not guarantee that a model transfers well to other languages, even given similar resources. We hope that future work in sentiment analysis will make better use of the available test datasets." + ], + [ + "With this article, we have presented a novel projection-based approach to targeted cross-lingual sentiment analysis. The central unit of the proposed method is Blse which enables the transfer of annotations from a source language to a non-annotated target language. The only input it relies on are word embeddings (which can be trained without manual labeling by self-annotation) and a comparably small translation dictionary which connects the semantics of the source and the target language.", + "In the binary classification setting (automatic labeling of sentences or documents), Blse constitutes a novel state of the art on several language and domain pairs. For a more fine-grained classification to four sentiment labels, Barista and Muse perform slightly better. The predictions in all settings are complementary to the strong upper bound of employing machine translations: in an ensemble, even this resource-intense approach is inferior.", + "The transfer from classification to target-level analysis revealed additional challenges. The performance is lower, particularly for the 4-class setting. Our analyses show that mapping of sentence predictions to the aspects mentioned in each sentence with a machine translation model is a very challenging empirical upper bound \u2013 the difference in performance compared to projection-based methods is greater here than for the sentence-classification setting. However, we showed that in resource-scarce environments, Blse constitutes the current state of the art for binary target-level sentiment analysis when incorporated in a deep learning architecture which is informed about the aspect. Muse performs better in the same architecture for the 4-class setting.", + "Our analysis further showed that the neural network needs to be informed about both the aspect and the context \u2013 limiting the information to a selection of these sentence parts strongly underperforms the combined setting. That also demonstrates that the model does not rely on prior distributions of aspect mentions.", + "The final experiment in the paper is a real-world deployment of the target-level sentiment analysis system in multilingual setting with 10 languages, where the assumption is that the only supervision is available in English (which is not part of the target languages). We learned here that it is important to have access to in-domain data (even for cross-lingual projection), especially in the multiclass setting. Binary classification however, which might often be sufficient for real-world applications, is more robust to domain changes. Further, machine translation is less sensitive to language dissimilarities, unlike projection-based methods. The amount of available unlabeled data to create embeddings plays a role in the final performance of the system, although only to a minor extent.", + "The current performance of the projection-based techniques still lags behind state-of-the-art MT approaches on most tasks, indicating that there is still much work to be done. While general bilingual embedding techniques do not seem to incorporate enough sentiment information, they are able to retain the semantics of their word vectors to a large degree even after projection. We hypothesize that the ability to retain the original semantics of the monolingual spaces leads to Muse performing better than MT on multiclass targeted sentiment analysis. The joint approach introduced in this work suffers from the degradation of the original semantics space, while optimizing the sentiment information. Moving from a similarity-based loss to a ranking loss, where the model must predict a ranked list of most similar translations could improve the model, but would require further resource development cross-lingually, as a simple bilingual dictionary would not provide enough information.", + "One problem that arises when using bilingual embeddings instead of machine translation is that differences in word order are no longer handled BIBREF2 . Machine translation models, on the other hand, always include a reordering element. Nonetheless, there is often a mismatch between the real source language word order and the translated word order. In this work, we avoided the problem by using a bag-of-embeddings representation, but Barnes2017 found that the bag-of-embeddings approach does not perform as well as approaches that take word order into account, e. g., Lstms or Cnns. We leave the incorporation of these classifiers into our framework for future work.", + "Unsupervised machine translation Artetxe2018,Lample2018,artetxe2018emnlp shows great promise for sentence-level classification. Like MT, however, it performs worse on noisy data, such as tweets. Therefore, users who want to apply targeted cross-lingual approaches to noisy data should consider currently consider using embedding projection methods, such as Blse. Future work on adapting unsupervised machine translation to noisy text may provide another solution for low-resource NLP.", + "The authors thank Patrik Lambert, Toni Badia, Amaia Oliden, Itziar Etxeberria, Jessie Kief, Iris H\u00fcbscher, and Arne \u00d8hm for helping with the annotation of the resources used in this research. This work has been partially supported by the DFG Collaborative Research Centre SFB 732 and a SGR-DTCL Predoctoral Scholarship." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0835/instruction.md b/qasper-0835/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..678ffae6efb55fa61a94384ae5ec1869343a5a8c --- /dev/null +++ b/qasper-0835/instruction.md @@ -0,0 +1,81 @@ +Name of Paper: Improving Open Information Extraction via Iterative Rank-Aware Learning + +Question: What's the input representation of OpenIE tuples into the model? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Neural Models for Open IE", + "Problem Formulation", + "Model Architecture and Decoding", + "Iterative Rank-Aware Learning", + "Binary Classification Loss", + "Iterative Learning", + "Experimental Settings", + "Evaluation Results", + "Conclusion", + "Acknowledgements" + ], + "paragraphs": [ + [ + "Open information extraction (IE, sekine2006demand, Banko:2007:OIE) aims to extract open-domain assertions represented in the form of $n$ -tuples (e.g., was born in; Barack Obama; Hawaii) from natural language sentences (e.g., Barack Obama was born in Hawaii). Open IE started from rule-based BIBREF0 and syntax-driven systems BIBREF1 , BIBREF2 , and recently has used neural networks for supervised learning BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 .", + "A key step in open IE is confidence modeling, which ranks a list of candidate extractions based on their estimated quality. This is important for downstream tasks, which rely on trade-offs between the precision and recall of extracted assertions. For instance, an open IE-powered medical question answering (QA) system may require its assertions in higher precision (and consequently lower recall) than QA systems for other domains. For supervised open IE systems, the confidence score of an assertion is typically computed based on its extraction likelihood given by the model BIBREF3 , BIBREF5 . However, we observe that this often yields sub-optimal ranking results, with incorrect extractions of one sentence having higher likelihood than correct extractions of another sentence. We hypothesize this is due to the issue of a disconnect between training and test-time objectives. Specifically, the system is trained solely to raise likelihood of gold-standard extractions, and during training the model is not aware of its test-time behavior of ranking a set of system-generated assertions across sentences that potentially include incorrect extractions.", + "To calibrate open IE confidences and make them more globally comparable across different sentences, we propose an iterative rank-aware learning approach, as outlined in fig:arch. Given extractions generated by the model as training samples, we use a binary classification loss to explicitly increase the confidences of correct extractions and decrease those of incorrect ones. Without adding additional model components, this training paradigm naturally leads to a better open IE model, whose extractions can be further included as training samples. We further propose an iterative learning procedure that gradually improves the model by incrementally adding extractions to the training data. Experiments on the OIE2016 dataset BIBREF8 indicate that our method significantly outperforms both neural and non-neural models." + ], + [ + "We briefly revisit the formulation of open IE and the neural network model used in our paper." + ], + [ + "Given sentence $\\mathbf {s}=(w_1, w_2, ..., w_n)$ , the goal of open IE is to extract assertions in the form of tuples $\\mathbf {r}=(\\mathbf {p}, \\mathbf {a}_1, \\mathbf {a}_2, ..., \\mathbf {a}_m)$ , composed of a single predicate and $m$ arguments. Generally, these components in $\\mathbf {r}$ need not to be contiguous, but to simplify the problem we assume they are contiguous spans of words from $\\mathbf {s}$ and there is no overlap between them.", + "Methods to solve this problem have recently been formulated as sequence-to-sequence generation BIBREF4 , BIBREF5 , BIBREF6 or sequence labeling BIBREF3 , BIBREF7 . We adopt the second formulation because it is simple and can take advantage of the fact that assertions only consist of words from the sentence. Within this framework, an assertion $\\mathbf {r}$ can be mapped to a unique BIO BIBREF3 label sequence $\\mathbf {y}$ by assigning $O$ to the words not contained in $\\mathbf {r}$ , $B_{p}$ / $I_{p}$ to the words in $\\mathbf {p}$ , and $B_{a_i}$ / $I_{a_i}$ to the words in $\\mathbf {a}_i$ respectively, depending on whether the word is at the beginning or inside of the span.", + "The label prediction $\\hat{\\mathbf {y}}$ is made by the model given a sentence associated with a predicate of interest $(\\mathbf {s}, v)$ . At test time, we first identify verbs in the sentence as candidate predicates. Each sentence/predicate pair is fed to the model and extractions are generated from the label sequence." + ], + [ + "Our training method in sec:ours could potentially be used with any probabilistic open IE model, since we make no assumptions about the model and only the likelihood of the extraction is required for iterative rank-aware learning. As a concrete instantiation in our experiments, we use RnnOIE BIBREF3 , BIBREF9 , a stacked BiLSTM with highway connections BIBREF10 , BIBREF11 and recurrent dropout BIBREF12 . Input of the model is the concatenation of word embedding and another embedding indicating whether this word is predicate: $\n\\mathbf {x}_t = [\\mathbf {W}_{\\text{emb}}(w_t), \\mathbf {W}_{\\text{mask}}(w_t = v)].\n$ ", + "The probability of the label at each position is calculated independently using a softmax function: $\nP(y_t|\\mathbf {s}, v) \\propto \\text{exp}(\\mathbf {W}_{\\text{label}}\\mathbf {h}_t + \\mathbf {b}_{\\text{label}}),\n$ ", + "where $\\mathbf {h}_t$ is the hidden state of the last layer. At decoding time, we use the Viterbi algorithm to reject invalid label transitions BIBREF9 , such as $B_{a_2}$ followed by $I_{a_1}$ .", + "We use average log probability of the label sequence BIBREF5 as its confidence: ", + "$$c(\\mathbf {s}, v, \\hat{\\mathbf {y}}) = \\frac{\\sum _{t=1}^{|\\mathbf {s}|}{\\log {P(\\hat{y_t}|\\mathbf {s}, v)}}}{|\\mathbf {s}|}.$$ (Eq. 7) ", + "The probability is trained with maximum likelihood estimation (MLE) of the gold extractions. This formulation lacks an explicit concept of cross-sentence comparison, and thus incorrect extractions of one sentence could have higher confidence than correct extractions of another sentence." + ], + [ + "In this section, we describe our proposed binary classification loss and iterative learning procedure." + ], + [ + "To alleviate the problem of incomparable confidences across sentences, we propose a simple binary classification loss to calibrate confidences to be globally comparable. Given a model $\\theta ^\\prime $ trained with MLE, beam search is performed to generate assertions with the highest probabilities for each predicate. Assertions are annotated as either positive or negative with respect to the gold standard, and are used as training samples to minimize the hinge loss: ", + "$$\\hspace{-2.84526pt}\\hat{\\theta } = \\underset{\\theta }{\\operatornamewithlimits{arg\\,min}}\\hspace{-8.53581pt}\\underset{\\begin{array}{c}\\mathbf {s} \\in \\mathcal {D}\\\\ v, \\hat{\\mathbf {y}} \\in g_{\\theta ^\\prime }(\\mathbf {s})\\end{array}}{\\operatorname{\\mathbb {E}}}\\hspace{-11.38109pt}\\max {(0,1-t \\cdot c_{\\theta }(\\mathbf {s}, v, \\hat{\\mathbf {y}}))},$$ (Eq. 9) ", + "where $\\mathcal {D}$ is the training sentence collection, $g_{\\theta ^\\prime }$ represents the candidate generation process, and $t \\in \\lbrace 1,-1\\rbrace $ is the binary annotation. $c_{\\theta }(\\mathbf {s}, v, \\hat{\\mathbf {y}})$ is the confidence score calculated by average log probability of the label sequence.", + "The binary classification loss distinguishes positive extractions from negative ones generated across different sentences, potentially leading to a more reliable confidence measure and better ranking performance." + ], + [ + "Compared to using external models for confidence modeling, an advantage of the proposed method is that the base model does not change: the binary classification loss just provides additional supervision. Ideally, the resulting model after one-round of training becomes better not only at confidence modeling, but also at assertion generation, suggesting that extractions of higher quality can be added as training samples to continue this training process iteratively. The resulting iterative learning procedure (alg:iter) incrementally includes extractions generated by the current model as training samples to optimize the binary classification loss to obtain a better model, and this procedure is continued until convergence. [t] training data $\\mathcal {D}$ , initial model $\\theta ^{(0)}$ model after convergence $\\theta $ $t \\leftarrow 0$ # iteration", + " $\\mathcal {E} \\leftarrow \\emptyset $ # generated extractions", + "not converge $\\mathcal {E} \\leftarrow \\mathcal {E} \\cup \\lbrace (\\mathbf {s}, v, \\hat{\\mathbf {y}})|v,\\hat{\\mathbf {y}} \\in g_{\\theta ^{(t)}}(\\mathbf {s}), \\forall \\mathbf {s} \\in \\mathcal {D}\\rbrace $ ", + " $\\theta ^{(t+1)} \\leftarrow \\underset{\\theta }{\\operatornamewithlimits{arg\\,min}}\\hspace{-8.53581pt}\\underset{(\\mathbf {s}, v, \\hat{\\mathbf {y}})\\in \\mathcal {E}}{\\operatorname{\\mathbb {E}}}\\hspace{-8.53581pt}\\max {(0,1-t \\cdot c_{\\theta }(\\mathbf {s}, v, \\hat{\\mathbf {y}}))}$ ", + " $t \\leftarrow t+1$ Iterative learning. " + ], + [ + "We use the OIE2016 dataset BIBREF8 to evaluate our method, which only contains verbal predicates. OIE2016 is automatically generated from the QA-SRL dataset BIBREF13 , and to remove noise, we remove extractions without predicates, with less than two arguments, and with multiple instances of an argument. The statistics of the resulting dataset are summarized in tab:data.", + "We follow the evaluation metrics described by Stanovsky:2016:OIE2016: area under the precision-recall curve (AUC) and F1 score. An extraction is judged as correct if the predicate and arguments include the syntactic head of the gold standard counterparts.", + "We compare our method with both competitive neural and non-neural models, including RnnOIE BIBREF3 , OpenIE4, ClausIE BIBREF2 , and PropS BIBREF14 .", + "Our implementation is based on AllenNLP BIBREF15 by adding binary classification loss function on the implementation of RnnOIE. The network consists of 4 BiLSTM layers (2 forward and 2 backward) with 64-dimensional hidden units. ELMo BIBREF16 is used to map words into contextualized embeddings, which are concatenated with a 100-dimensional predicate indicator embedding. The recurrent dropout probability is set to 0.1. Adadelta BIBREF17 with $\\epsilon =10^{-6}$ and $\\rho =0.95$ and mini-batches of size 80 are used to optimize the parameters. Beam search size is 5." + ], + [ + "tab:expmain lists the evaluation results. Our base model (RnnOIE, sec:oie) performs better than non-neural systems, confirming the advantage of supervised training under the sequence labeling setting. To test if the binary classification loss (E.q. 9 , sec:ours) could yield better-calibrated confidence, we perform one round of fine-tuning of the base model with the hinge loss ( $+$ Binary loss in tab:expmain). We show both the results of using the confidence (E.q. 7 ) of the fine-tuned model to rerank the extractions of the base model (Rerank Only), and the end-to-end performance of the fine-tuned model in assertion generation (Generate). We found both settings lead to improved performance compared to the base model, which demonstrates that calibrating confidence using binary classification loss can improve the performance of both reranking and assertion generation. Finally, our proposed iterative learning approach (alg:iter, sec:ours) significantly outperforms non-iterative settings.", + "We also investigate the performance of our iterative learning algorithm with respect to the number of iterations in fig:iter. The model obtained at each iteration is used to both rerank the extractions generated by the previous model and generate new extractions. We also report results of using only positive samples for optimization. We observe the AUC and F1 of both reranking and generation increases simultaneously for the first 6 iterations and converges after that, which demonstrates the effectiveness of iterative training. The best performing iteration achieves AUC of 0.125 and F1 of 0.315, outperforming all the baselines by a large margin. Meanwhile, using both positive and negative samples consistently outperforms only using positive samples, which indicates the necessity of exposure to the errors made by the system.", + "tab:casererank compares extractions from RnnOIE before and after reranking. We can see the order is consistent with the annotation after reranking, showing the additional loss function's efficacy in calibrating the confidences; this is particularly common in extractions with long arguments. tab:casegen shows a positive extraction discovered after iterative training (first example), and a wrong extraction that disappears (second example), which shows that the model also becomes better at assertion generation.", + "Why is the performance still relatively low? We randomly sample 50 extractions generated at the best performing iteration and conduct an error analysis to answer this question. To count as a correct extraction, the number and order of the arguments should be exactly the same as the ground truth and syntactic heads must be included, which is challenging considering that the OIE2016 dataset has complex syntactic structures and multiple arguments per predicate.", + "We classify the errors into three categories and summarize their proportions in tab:err. \u201cOvergenerated predicate\u201d is where predicates not included in ground truth are overgenerated, because all the verbs are used as candidate predicates. An effective mechanism should be designed to reject useless candidates. \u201cWrong argument\u201d is where extracted arguments do not coincide with ground truth, which is mainly caused by merging multiple arguments in ground truth into one. \u201cMissing argument\u201d is where the model fails to recognize arguments. These two errors usually happen when the structure of the sentence is complicated and coreference is involved. More linguistic information should be introduced to solve these problems." + ], + [ + "We propose a binary classification loss function to calibrate confidences in open IE. Iteratively optimizing the loss function enables the model to incrementally learn from trial and error, yielding substantial improvement. An error analysis is performed to shed light on possible future directions." + ], + [ + "This work was supported in part by gifts from Bosch Research, and the Carnegie Bosch Institute." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0851/instruction.md b/qasper-0851/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..cf1b94c6b5f5cdd464ea18ef2fade8c040ba3515 --- /dev/null +++ b/qasper-0851/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Plan, Write, and Revise: an Interactive System for Open-Domain Story Generation + +Question: Do they evaluate in other language appart from English? \ No newline at end of file diff --git a/qasper-0856/instruction.md b/qasper-0856/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..e2f513ae887245f1625fc0d6ff9522fd5d0e0914 --- /dev/null +++ b/qasper-0856/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Collecting Indicators of Compromise from Unstructured Text of Cybersecurity Articles using Neural-Based Sequence Labelling + +Question: What type of hand-crafted features are used in state of the art IOC detection systems? \ No newline at end of file diff --git a/qasper-0869/instruction.md b/qasper-0869/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..de90c8612c6db07c3f18d771a12970439ae548a3 --- /dev/null +++ b/qasper-0869/instruction.md @@ -0,0 +1,71 @@ +Name of Paper: A Comparative Evaluation of Visual and Natural Language Question Answering Over Linked Data + +Question: Which four QA systems do they use? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "INTRODUCTION", + "RELATED WORK", + "SYSTEM DESCRIPTION", + "EVALUATION", + "Evaluation Setup", + "Evaluation Results and Discussion", + "CONCLUSIONS", + "ACKNOWLEDGEMENTS" + ], + "paragraphs": [ + [ + "The Semantic Web provides a large number of structured datasets in form of Linked Data. One central obstacle is to make this data available and consumable to lay users without knowledge of formal query languages such as SPARQL. In order to satisfy specific information needs of users, a typical approach are natural language interfaces to allow question answering over the Linked Data (QALD) by translating user queries into SPARQL BIBREF0 , BIBREF1 . As an alternative method, BIBREF2 propose a visual method of QA using an iterative diagrammatic approach. The diagrammatic approach relies on the visual means only, it requires more user interaction than natural language QA, but also provides additional benefits like intuitive insights into dataset characteristics, or a broader understanding of the answer and the potential to further explore the answer context, and finally allows for knowledge sharing by storing and sharing resulting diagrams.", + "In contrast to BIBREF2 , who present the basic method and tool for diagrammatic question answering (DQA), here we evaluate DQA in comparison to natural language QALD systems. Both approaches have different characteristics, therefore we see them as complementary rather than in competition.", + "The basic research goals are: i) Given a dataset extracted from the QALD7 benchmark, we evaluate DQA versus state-of-the-art QALD systems. ii) More specifically, we investigate if and to what extent DQA can be complementary to QALD systems, especially in cases where those systems do not find a correct answer. iii) Finally, we want to present the basic outline for the integration of the two methods.", + "In a nutshell, users that applied DQA found the correct answer with an F1-score of 79.5%, compared to a maximum of 59.2% for the best performing QALD system. Furthermore, for the subset of questions where the QALD system could not provide a correct answer, users found the answer with 70% F1-score with DQA. We further analyze the characteristics of questions where the QALD or DQA, respectively, approach is better suited.", + "The results indicate, that aside from the other benefits of DQA, it can be a valuable component for integration into larger QALD systems, in cases where those systems cannot find an answer, or when the user wants to explore the answer context in detail by visualizing the relevant nodes and relations. Moreover, users can verify answers given by a QALD system using DQA in case of doubt.", + "This publication is organized as follows: After the presentation of related work in Section SECREF2 , and a brief system description of the DQA tool in Section SECREF3 , the main focus of the paper is on evaluation setup and results of the comparison of DQA and QALD, including a discussion, in Section SECREF4 . The paper concludes with Section SECREF5 ." + ], + [ + "As introduced in BIBREF2 we understand diagrammatic question answering (DQA) as the process of QA relying solely on visual exploration using diagrams as a representation of the underlying knowledge source. The process includes (i) a model for diagrammatic representation of semantic data which supports data interaction using embedded queries, (ii) a simple method for step-by-step construction of diagrams with respect to cognitive boundaries and a layout that boosts understandability of diagrams, (iii) a library for visual data exploration and sharing based on its internal data model, and (iv) an evaluation of DQA as knowledge understanding and knowledge sharing tool. BIBREF3 propose a framework of five perspectives of knowledge visualization, which can be used to describe certain aspects of the DQA use cases, such as its goal to provide an iterative exploration method, which is accessible to any user, the possibility of knowledge sharing (via saved diagrams), or the general purpose of knowledge understanding and abstraction from technical details.", + "Many tools exist for visual consumption and interaction with RDF knowledge bases, however, they are not designed specifically towards the question answering use case. BIBREF4 give an overview of ontology and Linked Data visualization tools, and categorize them based on the used visualization methods, interaction techniques and supported ontology constructs.", + "Regarding language-based QA over Linked Data, BIBREF5 discuss and study the usefulness of natural language interfaces to ontology-based knowledge bases in a general way. They focus on usability of such systems for the end user, and conclude that users prefer full sentences for query formulation and that natural language interfaces are indeed useful.", + " BIBREF0 describe the challenges of QA over knowledge bases using natural languages, and elaborate the various techniques used by existing QALD systems to overcome those challenges. In the present work, we compare DQA with four of those systems using a subset of questions of the QALD7 benchmark. Those systems are: gAnswer BIBREF6 is an approach for RDF QA that has a \u201cgraph-driven\u201d perspective. In contrast to traditional approaches, which first try to understand the question, and then evaluate the query, in gAnswer the intention of the query is modeled in a structured way, which leads to a subgraph matching problem. Secondly, QAKiS BIBREF7 is QA system over structured knowledge bases such as DBpedia that makes use of relational patterns which capture different ways to express a certain relation in a natural language in order to construct a target-language (SPARQL) query. Further, Platypus BIBREF8 is a QA system on Wikidata. It represents questions in an internal format related to dependency-based compositional semantics which allows for question decomposition and language independence. The platform can answer complex questions in several languages by using hybrid grammatical and template-based techniques. And finally, also the WDAqua BIBREF0 system aims for language-independence and for being agnostic of the underlying knowledge base. WDAqua puts more importance on word semantics than on the syntax of the user query, and follows a processes of query expansion, SPARQL construction, query ranking and then making an answer decision.", + "For the evaluation of QA systems, several benchmarks have been proposed such as WebQuestions BIBREF9 or SimpleQuestions BIBREF10 . However, the most popular benchmarks in the Semantic Web field arise from the QALD evaluation campaign BIBREF1 . The recent QALD7 evaluation campaign includes task 4: \u201cEnglish question answering over Wikidata\u201d which serves as basis to compile our evaluation dataset." + ], + [ + "The DQA functionality is part of the Ontodia tool. The initial idea of Ontodia was to enable the exploration of semantic graphs for ordinary users. Data exploration is about efficiently extracting knowledge from data even in situations where it is unclear what is being looked for exactly BIBREF11 .", + "The DQA tool uses an incremental approach to exploration typically starting from a very small number of nodes. With the context menu of a particular node, relations and related nodes can be added until the diagram fulfills the information need of the user. Figure FIGREF1 gives an example of a start node, where a user wants to learn more about the painting style of Van Gogh.", + "To illustrate the process, we give a brief example here. More details about the DQA tool, the motivation for DQA and diagram-based visualizations are found in previous work BIBREF2 , BIBREF12 .", + "As for the example, when attempting to answer a question such as \u201cWho is the mayor of Paris?\u201d the first step for a DQA user is finding a suitable starting point, in our case the entity Paris. The user enters \u201cParis\u201d into the search box, and can then investigate the entity on the tool canvas. The information about the entity stems from the underlying dataset, for example Wikidata. The user can \u2013 in an incremental process \u2013 search in the properties of the given entity (or entities) and add relevant entities onto the canvas. In the given example, the property \u201chead of government\u201d connects the mayor to the city of Paris, Anne Hidalgo. The final diagram which answers the given question is presented in Figure FIGREF3 ." + ], + [ + "Here we present the evaluation of DQA in comparison to four QALD systems." + ], + [ + "As evaluation dataset, we reuse questions from the QALD7 benchmark task 4 \u201cQA over Wikidata\u201d. Question selection from QALD7 is based on the principles of question classification in QA BIBREF13 . Firstly, it is necessary to define question types which correspond to different scenarios of data exploration in DQA, as well as the type of expected answers and the question focus. The question focus refers to the main information in the question which help a user find the answer. We follow the model of BIBREF14 who categorize questions by their question word into WHO, WHICH, WHAT, NAME, and HOW questions. Given the question and answer type categories, we created four questionnaires with nine questions each resulting in 36 questions from the QALD dataset. The questions were picked in equal number for five basic question categories.", + "20 persons participated in the DQA evaluation \u2013 14 male and six female from eight different countries. The majority of respondents work within academia, however seven users were employed in industry. 131 diagrams (of 140 expected) were returned by the users.", + "The same 36 questions were answered using four QALD tools: WDAqua BIBREF0 , QAKiS BIBREF7 , gAnswer BIBREF6 and Platypus BIBREF8 .", + "For the QALD tools, a human evaluator pasted the questions as is into the natural language Web interfaces, and submitted them to the systems. Typically QALD tools provide a distinct answer, which may be a simple literal, or a set of entities which represent the answer, and which can be compared to the gold standard result. However, the WDAqua system, sometimes, additionally to the direct answer to the question, provides links to documents related to the question. We always chose the answer available via direct answer.", + "To assess the correctness of the answers given both by participants in the DQA experiments, and by the QALD system, we use the classic information retrieval metrics of precision (P), recall (R), and F1. INLINEFORM0 measures the fraction of relevant (correct) answer (items) given versus all answers (answer items) given. INLINEFORM1 is the faction of correct answer (parts) given divided by all correct ones in the gold answer, and INLINEFORM2 is the harmonic mean of INLINEFORM3 and INLINEFORM4 . As an example, if the question is \u201cWhere was Albert Einstein born?\u201d (gold answer: \u201cUlm\u201d), and the system gives two answers \u201cUlm\u201d and \u201cBern\u201d, then INLINEFORM5 , INLINEFORM6 and INLINEFORM7 .", + "For DQA four participants answered each question, therefore we took the average INLINEFORM0 , INLINEFORM1 , and INLINEFORM2 values over the four evaluators as the result per question. The detailed answers by the participants and available online." + ], + [ + "Table TABREF8 presents the overall evaluation metrics of DQA, and the four QALD tools studied. With the given dataset, WDAqua (56.1% F1) and gAnswer (59.2% F1) clearly outperform askplatyp.us (8.6% F1) and QAKiS (27.5% F1). Detailed results per question including the calculation of INLINEFORM0 , INLINEFORM1 and INLINEFORM2 scores are available online. DQA led to 79.5% F1 (80.1% precision and 78.5% recall).", + "In further evaluations, we compare DQA results to WDAqua in order to study the differences and potential complementary aspects of the approaches. We selected WDAqua as representative of QALD tools, as it provides state-of-the-art results, and is well grounded in the Semantic Web community. ", + "Comparing DQA and WDAqua, the first interesting question is: To what extend is DQA helpful on questions that could not be answered by the QALD system? For WDAqua the overall F1 score on our test dataset is INLINEFORM0 . For the subset of questions where WDAqua had no, or only a partial, answer, DQA users found the correct answer in INLINEFORM1 of cases. On the other hand, the subset of questions that DQA users (partially) failed to answer, were answered correctly by WDAqua with an F1 of INLINEFORM2 . If DQA is used as a backup method for questions not correctly answered with WDAqua, then overall F1 can be raised to INLINEFORM3 . The increase from INLINEFORM4 to INLINEFORM5 demonstrates the potential of DQA as complementary component in QALD systems.", + "As expected, questions that are difficult to answer with one approach are also harder for the other approach \u2013 as some questions in the dataset or just more complex to process and understand than others. However, almost 70% of questions not answered by WDAqua could still be answered by DQA. As examples of cases which are easier to answer for one approach than the other, a question that DQA users could answer, but where WDAqua failed is: \u201cWhat is the name of the school where Obama's wife studied?\u201d. This complex question formulation is hard to interpret correctly for a machine. In contrast to DQA, QALD systems also struggled with \u201cWho is the son of Sonny and Cher?\u201d. This question needs a lot of real-world knowledge to map the names Sonny and Cher to their corresponding entities. The QALD system needs to select the correct Cher entity from multiple options in Wikidata, and also to understand that \u201cSonny\u201d refers to the entity Sonny Bono. The resulting answer diagram is given in Figure FIGREF17 . More simple questions, like \u201cWho is the mayor of Paris?\u201d were correctly answered by WDAqua, but not by all DQA users. DQA participants in this case struggled to make the leap from the noun \u201cmayor\u201d to the head-of-government property in Wikidata.", + "Regarding the limits of DQA, this method has difficulties when the answer can be obtained only with joins of queries, or when it is hard to find the initial starting entities related to question focus. For example, a question like \u201cShow me the list of African birds that are extinct.\u201d typically requires an intersection of two (large) sets of candidates entities, ie. all African birds and extinct birds. Such a task can easily be represented in a SPARQL query, but is hard to address with diagrams, because it would require placing, and interacting with, a huge amount of nodes on the exploration canvas.", + "Overall, the experiments indicate, that additionally to the use cases where QALD and DQA are useful on their own, there is a lot of potential in combining the two approaches, especially by providing a user the opportunity to explore the dataset with DQA if QALD did not find a correct answer, or when a user wants to confirm the QALD answer by checking in the underlying knowledge base. Furthermore, visually exploring the dataset provides added benefits, like understanding the dataset characteristics, sharing of resulting diagrams (if supported by the tool), and finding more information related to the original information need.", + "For the integration of QALD and DQA, we envision two scenarios. The first scenario addresses plain question answering, and here DQA can be added to a QALD system for cases where a user is not satisfied with a given answer. The QALD Web interface can for example have a Explore visually with diagrams button, which brings the user to a canvas on which the entities detected by the QALD system within the question and results (if any) are displayed on the canvas as starting nodes. The user will then explore the knowledge graph and find the answers in the same way as the participants in our experiments. The first scenario can lead to a large improvement in answer F1 (see above).", + "The second scenario of integration of QALD and DQA focuses on the exploration aspect. Even if the QALD system provides the correct answer, a user might be interested to explore the knowledge graph to validate the result and to discover more interesting information about the target entities. From an implementation and UI point of view, the same Explore visually with diagrams button and pre-population of the canvas can be used. Both scenarios also provide the additional benefits of potentially saving and sharing the created diagrams, which elaborate the relation between question and answer." + ], + [ + "In this work, we compare two approaches to answer questions over Linked Data datasets: a visual diagrammatic approach (DQA) which involves iterative exploration of the graph, and a natural language-based (QALD). The evaluations show, that DQA can be a helpful addition to pure QALD systems, both regarding evaluation metrics (precision, recall, and F1), and also for dataset understanding and further exploration. The contributions include: i) a comparative evaluation of four QALD tools and DQA with a dataset extracted from the QALD7 benchmark, ii) an investigation into the differences and potential complementary aspects of the two approaches, and iii) the proposition of integration scenarios for QALD and DQA.", + "In future work we plan to study the integration of DQA and QALD, especially the aspect of automatically creating an initial diagram from a user query, in order to leverage the discussed potentials. We envision an integrated tool, that uses QALD as basic method to find an answer to a question quickly, but also allows to explore the knowledge graph visually to raise answer quality and support exploration with all its discussed benefits." + ], + [ + "This work was supported by the Government of the Russian Federation (Grant 074-U01) through the ITMO Fellowship and Professorship Program." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0874/instruction.md b/qasper-0874/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0626927b8343b76d0d08416305f8e3ca61dd9603 --- /dev/null +++ b/qasper-0874/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: 'Warriors of the Word' -- Deciphering Lyrical Topics in Music and Their Connection to Audio Feature Dimensions Based on a Corpus of Over 100,000 Metal Songs + +Question: How long is the model trained? \ No newline at end of file diff --git a/qasper-0880/instruction.md b/qasper-0880/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..8713db9ddaf1fa90711d2f74c618687c4b598834 --- /dev/null +++ b/qasper-0880/instruction.md @@ -0,0 +1,100 @@ +Name of Paper: Abstractive Dialog Summarization with Semantic Scaffolds + +Question: How does SPNet utilize additional speaker role, semantic slot and dialog domain annotations? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Proposed Method", + "Proposed Method ::: Background", + "Proposed Method ::: Scaffold Pointer Network (SPNet)", + "Proposed Method ::: Scaffold Pointer Network (SPNet) ::: Speaker Role Scaffold", + "Proposed Method ::: Scaffold Pointer Network (SPNet) ::: Semantic Slot Scaffold", + "Proposed Method ::: Scaffold Pointer Network (SPNet) ::: Dialog Domain Scaffold", + "Experimental Settings ::: Dataset", + "Experimental Settings ::: Evaluation Metrics", + "Experimental Settings ::: Implementation Details", + "Results and Discussions ::: Automatic Evaluation Results", + "Results and Discussions ::: Human Evaluation Results", + "Results and Discussions ::: Case study", + "Conclusion and Future Work" + ], + "paragraphs": [ + [ + "Summarization aims to condense a piece of text to a shorter version, retaining the critical information. On dialogs, summarization has various promising applications in the real world. For instance, the automatic doctor-patient interaction summary can save doctors' massive amount of time used for filling medical records. There is also a general demand for summarizing meetings in order to track project progress in the industry. Generally, multi-party conversations with interactive communication are more difficult to summarize than single-speaker documents. Hence, dialog summarization will be a potential field in summarization track.", + "There are two types of summarization: extractive and abstractive. Extractive summarization selects sentences or phrases directly from the source text and merges them to a summary, while abstractive summarization attempts to generate novel expressions to condense information. Previous dialog summarization research mostly study extractive summarization BIBREF1, BIBREF2. Extractive methods merge selected important utterances from a dialog to form summary. Because dialogs are highly dependant on their histories, it is difficult to produce coherent discourses with a set of non-consecutive conversation turns. Therefore, extractive summarization is not the best approach to summarize dialogs. However, most modern abstractive methods focus on single-speaker documents rather than dialogs due to the lack of dialog summarization corpora. Popular abstractive summarization dataset like CNN/Daily Mail BIBREF3 is on news documents. AMI meeting corpus BIBREF4 is the common benchmark, but it only has extractive summary.", + "In this work, we introduce a dataset for abstractive dialog summarization based on MultiWOZ BIBREF0. Seq2Seq models such as Pointer-Generator BIBREF5 have achieved high-quality summaries of news document. However, directly applying a news summarizer to dialog results in two drawbacks: informative entities such as place name are difficult to capture precisely and contents in different domains are summarized unequally. To address these problems, we propose Scaffold Pointer Network (SPNet). SPNet incorporates three types of semantic scaffolds in dialog: speaker role, semantic slot, and dialog domain. Firstly, SPNet adapts separate encoder to attentional Seq2Seq framework, producing distinct semantic representations for different speaker roles. Then, our method inputs delexicalized utterances for producing delexicalized summary, and fills in slot values to generate complete summary. Finally, we incorporate dialog domain scaffold by jointly optimizing dialog domain classification task along with the summarization task. We evaluate SPNet with both automatic and human evaluation metrics on MultiWOZ. SPNet outperforms Pointer-Generator BIBREF5 and Transformer BIBREF6 on all the metrics." + ], + [ + "BIBREF7 first applied modern neural models to abstractive summarization. Their approach is based on Seq2Seq framework BIBREF8 and attention mechanism BIBREF9, achieving state-of-the-art results on Gigaword and DUC-2004 dataset. BIBREF10 proposed copy mechanism in summarization, demonstrating its effectiveness by combining the advantages of extractive and abstractive approach. BIBREF5 applied pointing BIBREF11 as copy mechanism and use coverage mechanism BIBREF12 to discourage repetition. Most recently, reinforcement learning (RL) has been employed in abstractive summarization. RL-based approaches directly optimize the objectives of summarization BIBREF13, BIBREF14. However, deep reinforcement learning approaches are difficult to train and more prone to exposure bias BIBREF15.", + "Recently, pre-training methods are popular in NLP applications. BERT BIBREF16 and GPT BIBREF17 have achieved state-of-the-art performance in many tasks, including summarization. For instance, BIBREF18 proposed a method to pre-train hierarchical document encoder for extractive summarization. BIBREF19 proposed two strategies to incorporate a pre-trained model (GPT) to perform the abstractive summarizer and achieved a better performance. However, there has not been much research on adapting pre-trained models to dialog summarization.", + "Dialog summarization, specifically meeting summarization, has been studied extensively. Previous work generally focused on statistical machine learning methods in extractive dialog summarization: BIBREF20 used skip-chain conditional random fields (CRFs) BIBREF21 as a ranking method in extractive meeting summarization. BIBREF22 compared support vector machines (SVMs) BIBREF23 with LDA-based topic models BIBREF24 for producing decision summaries. However, abstractive dialog summarization was less explored due to the lack of a suitable benchmark. Recent work BIBREF25, BIBREF26, BIBREF27 created abstractive dialog summary benchmarks with existing dialog corpus. BIBREF26 annotated topic descriptions in AMI meeting corpus as the summary. However, topics they defined are coarse, such as \u201cindustrial designer presentation\". They also proposed a model with a sentence-gated mechanism incorporating dialog acts to perform abstractive summarization. Moreover, BIBREF28 first built a model to summarize audio-visual meeting data with an abstractive method. However, previous work has not investigated the utilization of semantic patterns in dialog, so we explore it in-depth in our work." + ], + [ + "As discussed above, state-of-the-art document summarizers are not applicable in conversation settings. We propose Scaffold Pointer Network (SPNet) based on Pointer-Generator BIBREF5. SPNet incorporates three types of semantic scaffolds to improve abstractive dialog summarization: speaker role, semantic slot and dialog domain." + ], + [ + "We first introduce Pointer-Generator BIBREF5. It is a hybrid model of the typical Seq2Seq attention model BIBREF29 and pointer network BIBREF11. Seq2Seq framework encodes source sequence and generates the target sequence with the decoder. The input sequence is fed into the encoder token by token, producing the encoder hidden states $h_i$ in each encoding step. The decoder receives word embedding of the previous word and generates a distribution to decide the target element in this step, retaining decoder hidden states $s_t$. In Pointer-Generator, attention distribution $a^t$ is computed as in BIBREF9:", + "where $W_h$, $W_s$, $v$ and $b_{attn}$ are all learnable parameters.", + "With the attention distribution $a^t$, context vector $h_t^*$ is computed as the weighted sum of encoder's hidden states. Context vector is regarded as the attentional information in the source text:", + "Pointer-Generator differs from typical Seq2Seq attention model in the generation process. The pointing mechanism combines copying words directly from the source text with generating words from a fixed vocabulary. Generation probability $p_{gen}$ is calculated as \u201ca soft switch\" to choose from copy and generation:", + "where $x_t$ is the decoder input, $w_{h^*}$, $w_s$, $w_x$ and $b_{ptr}$ are all learnable parameters. $\\sigma $ is sigmoid function, so the generation probability $p_{gen}$ has a range of $[0, 1]$.", + "The ability to select from copy and generation corresponds to a dynamic vocabulary. Pointer network forms an extended vocabulary for the copied tokens, including all the out-of-vocabulary(OOV) words appeared in the source text. The final probability distribution $P(w)$ on extended vocabulary is computed as follows:", + "where $P_{vocab}$ is the distribution on the original vocabulary, $V^{\\prime }$, $V$, $b$ and $b^{\\prime }$ are learnable parameters used to calculate such distribution." + ], + [ + "Our Scaffold Pointer Network (depicted in Figure FIGREF7) is based on Pointer-Generator BIBREF5. The contribution of SPNet is three-fold: separate encoding for different roles, incorporating semantic slot scaffold and dialog domain scaffold." + ], + [ + "Our encoder-decoder framework employs separate encoding for different speakers in the dialog. User utterances $x_t^{usr}$ and system utterances $x_t^{sys}$ are fed into a user encoder and a system encoder separately to obtain encoder hidden states $h_{i}^{usr}$ and $h_{i}^{sys}$ . The attention distributions and context vectors are calculated as described in section SECREF1. In order to merge these two encoders in our framework, the decoder's hidden state $s_0$ is initialized as:", + "The pointing mechanism in our model follows the Equation DISPLAY_FORM4, and we obtain the context vector $h_t^{*}$:" + ], + [ + "We integrate semantic slot scaffold by performing delexicalization on original dialogs. Delexicalization is a common pre-processing step in dialog modeling. Specifically, delexicalization replaces the slot values with its semantic slot name(e.g. replace 18:00 with [time]). It is easier for the language modeling to process delexicalized texts, as they have a reduced vocabulary size. But these generated sentences lack the semantic information due to the delexicalization. Some previous dialog system research ignored this issue BIBREF30 or completed single delexicalized utterance BIBREF31 as generated response. We propose to perform delexicalization in dialog summary, since delexicalized utterances can simplify dialog modeling. We fill the generated templates with slots with the copy and pointing mechanism.", + "We first train the model with the delexicalized utterance. Attention distribution $a^t$ over the source tokens instructs the decoder to fill up the slots with lexicalized values:", + "Note that $w_{slot}$ specifies the tokens that represents the slot name (e.g. [hotel_place], [time]). Decoder directly copies lexicalized value $value(w_i)$ conditioned on attention distribution $a_i^t$. If $w$ is not a slot token, then the probability $P(w)$ is calculated as Equation DISPLAY_FORM5." + ], + [ + "We integrate dialog domain scaffold through a multi-task framework. Dialog domain indicates different conversation task content, for example, booking hotel, restaurant and taxi in MultiWOZ dataset. Generally, the content in different domains varies so multi-domain task summarization is more difficult than single-domain. We include domain classification as the auxiliary task to incorporate the prior that different domains have different content. Feedback from the domain classification task provides domain specific information for the encoder to learn better representations. For domain classification, we feed the concatenated encoder hidden state through a binary classifier with two linear layers, producing domain probability $d$. The $i^{th}$ element $d_i$ in $d$ represents the probability of the $i^{th}$ domain:", + "where $U$, $U^{\\prime }$, $b_{d}$ and $b_{d}^{\\prime }$ are all trainable parameters in the classifier. We denote the loss function of summarization as $loss_1$ and domain classification as $loss_2$. Assume target word at timestep $t$ is $w_t^{*}$, $loss_1$ is the arithmetic mean of the negative log likelihood of $w_t^{*}$ over the generated sequence:", + "The domain classification task is a multi-label binary classification problem. We use binary cross entropy loss between the $i^{th}$ domain label $\\hat{d_i}$ and predict probability $d_i$ for this task:", + "where $|D|$ is the number of domains. Finally, we reweight the classification loss with hyperparameter $\\lambda $ and the objective function is:" + ], + [ + "We validate SPNet on MultiWOZ-2.0 dataset BIBREF0. MultiWOZ consists of multi-domain conversations between a tourist and a information center clerk on varies booking tasks or domains, such as booking restaurants, hotels, taxis, etc. There are 10,438 dialogs, spanning over seven domains. 3,406 of them are single-domain (8.93 turns on average) and 7,302 are multi-domain (15.39 turns on average). During MultiWOZ data collection, instruction is provided for crowd workers to perform the task. We use the instructions as the dialog summary, and an example data is shown in Table TABREF25. Dialog domain label is extracted from existing MultiWOZ annotation. In the experiment, we split the dataset into 8,438 training, 1,000 validation, and 1,000 testing." + ], + [ + "ROUGE BIBREF32 is a standard metric for summarization, designed to measure the surface word alignment between a generated summary and a human written summary. We evaluate our model with ROUGE-1, ROUGE-2 and ROUGE-L. They measure the word-overlap, bigram-overlap, and longest common sequence between the reference summary and the generated summary respectively. We obtain ROUGE scores using the files2rouge package. However, ROUGE is insufficient to measure summarization performance. The following example shows its limitations:", + "Reference: You are going to [restaurant_name] at [time].", + "Summary: You are going to [restaurant_name] at.", + "In this case, the summary has a high ROUGE score, as it has a considerable proportion of word overlap with the reference summary. However, it still has poor relevance and readability, for leaving out one of the most critical information: [time]. ROUGE treats each word equally in computing n-gram overlap while the informativeness actually varies: common words or phrases (e.g. \u201cYou are going to\") significantly contribute to the ROUGE score and readability, but they are almost irrelevant to essential contents. The semantic slot values (e.g. [restaurant_name], [time]) are more essential compared to other words in the summary. However, ROUGE did not take this into consideration. To address this drawback in ROUGE, we propose a new evaluation metric: Critical Information Completeness (CIC). Formally, CIC is a recall of semantic slot information between a candidate summary and a reference summary. CIC is defined as follows:", + "where $V$ stands for a set of delexicalized values in the reference summary, $Count_{match}(v)$ is the number of values co-occurring in the candidate summary and reference summary, and $m$ is the number of values in set $V$. In our experiments, CIC is computed as the arithmetic mean over all the dialog domains to retain the overall performance.", + "CIC is a suitable complementary metric to ROUGE because it accounts for the most important information within each dialog domain. CIC can be applied to any summarization task with predefined essential entities. For example, in news summarization the proper nouns are the critical information to retain." + ], + [ + "We implemented our baselines with OpenNMT framework BIBREF33. We delexicalize utterances according to the belief span annotation. To maintain the generalizability of SPNet, we combine the slots that refer to the same information from different dialog domains into one slot (e.g. time). Instead of using pre-trained word embeddings like GloVe BIBREF34, we train word embeddings from scratch with a 128-dimension embedding layer. We set the hidden states of the bidirectional LSTM encoders to 256 dimensions, and the unidirectional LSTM decoder to 512 dimension. Our model is optimized using Adam BIBREF35 with a learning rate of 0.001, $\\beta _1=0.9$, $\\beta _2=0.999$. We reduce the learning rate to half to avoid overfitting when the validation loss increases. We set the hyperparameter $\\lambda $ to 0.5 in the objective function and the batch size to eight. We use beam search with a beam size of three during decoding. We use the validation set to select the model parameter. Our model with and without multi-task takes about 15 epochs and seven epochs to converge, respectively." + ], + [ + "To demonstrate SPNet's effectiveness, we compare it with two state-of-the-art methods, Pointer-Generator BIBREF5 and Transformer BIBREF6. Pointer-Generator is the state-of-the-art method in abstractive document summarization. In inference, we use length penalty and coverage penalty mentioned in BIBREF36. The hyperparameters in the original implementation BIBREF5 were used. Transformer uses attention mechanisms to replace recurrence for sequence transduction. Transformer generalizes well to many sequence-to-sequence problems, so we adapt it to our task, following the implementation in the official OpenNMT-py documentation.", + "We show all the models' results in Table TABREF24. We observe that SPNet reaches the highest score in both ROUGE and CIC. Both Pointer-Generator and Transformer achieve high ROUGE scores, but a relative low CIC scores. It suggests that the baselines have more room for improvement on preserving critical slot information. All the scaffolds we propose can be applied to different neural network models. In this work we select Pointer-Generator as our base model in SPNet because we observe that Transformer only has a small improvement over Pointer-Generator but is having a higher cost on training time and computing resources. We observe that SPNet outperforms other methods in all the automatic evaluation metrics with a big margin, as it incorporates all the three semantic scaffolds. Semantic slot contributes the most to SPNet's increased performance, bringing the largest increase on all automatic evaluation metrics." + ], + [ + "We also perform human evaluation to verify if our method's increased performance on automatic evaluation metrics entails better human perceived quality. We randomly select 100 test samples from MultiWOZ test set for evaluation. We recruit 150 crowd workers from Amazon Mechanical Turk. For each sample, we show the conversation, reference summary, as well as summaries generated by Pointer-Generator and SPNet to three different participants. The participants are asked to score each summary on three indicators: relevance, conciseness and readability on a 1 to 5 scale, and rank the summary pair (tie allowed).", + "We present human evaluation results in Table TABREF27. In the scoring part, our model outperforms Pointer-Generator in all three evaluation metrics. SPNet scored better than Pointer-Generator on relevance and readability. All generated summaries are relatively concise; therefore, they score very similar in conciseness. Ground truth is still perceived as more relevant and readable than SPNet results. However, ground truth does not get a high absolute score. From the feedback of the evaluators, we found that they think that the ground truth has not covered all the necessary information in the conversation, and the description is not so natural. This motivates us to collect a dialog summarization dataset with high-quality human-written summaries in the future. Results in the ranking evaluation show more differences between different summaries. SPNet outperforms Pointer-Generator with a large margin. Its performance is relatively close to the ground truth summary." + ], + [ + "Table TABREF25 shows an example summary from all models along with ground truth summary. We observe that Pointer-Generator ignores some essential fragments, such as the restaurant booking information (6 people, Sunday, 18:45). Missing information always belongs to the last several domains (restaurant in this case) in a multi-domain dialog. We also observe that separately encoding two speakers reduces repetition and inconsistency. For instance, Pointer-Generator's summary mentions \u201cfree wifi\" several times and has conflicting requirements on wifi. This is because dialogs has information redundancy, but single-speaker model ignores such dialog property.", + "Our method has limitations. In the example shown in Table TABREF25, our summary does not mention the hotel name (Alexander Bed and Breakfast) and its address (517a Coldham Lane) referred in the source. It occurs because the ground truth summary doe not cover it in the training data. As a supervised method, SPNet is hard to generate a summary containing additional information beyond the ground truth. However, in some cases, SPNet can also correctly summarize the content not covered in the reference summary (see Table TABREF31 in Appendix).", + "Furthermore, although our SPNet achieves a much-improved performance, the application of SPNet still needs extra annotations for semantic scaffolds. For a dialog dataset, speaker role scaffold is a natural pattern for modeling. Most multi-domain dialog corpus has the domain annotation. While for texts, for example news, its topic categorization such as sports or entertainment can be used as domain annotation. We find that semantic slot scaffold brings the most significant improvement, but it is seldom explicitly annotated. However, the semantic slot scaffold can be relaxed to any critical entities in the corpus, such as team name in sports news or professional terminology in a technical meeting." + ], + [ + "We adapt a dialog generation dataset, MultiWOZ to an abstractive dialog summarization dataset. We propose SPNet, an end-to-end model that incorporates the speaker role, semantic slot and dialog domain as the semantic scaffolds to improve abstractive summary quality. We also propose an automatic evaluation metric CIC that considers semantic slot relevance to serve as a complementary metric to ROUGE. SPNet outperforms baseline methods in both automatic and human evaluation metrics. It suggests that involving semantic scaffolds efficiently improves abstractive summarization quality in the dialog scene.", + "Moreover, we can easily extend SPNet to other summarization tasks. We plan to apply semantic slot scaffold to news summarization. Specifically, we can annotate the critical entities such as person names or location names to ensure that they are captured correctly in the generated summary. We also plan to collect a human-human dialog dataset with more diverse human-written summaries." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1102/instruction.md b/qasper-1102/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..31fbe0e3755669e266a3ade07a6b8cbef1e8c3e3 --- /dev/null +++ b/qasper-1102/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer + +Question: Approximately how much computational cost is saved by using this model? \ No newline at end of file diff --git a/qasper-1105/instruction.md b/qasper-1105/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..73b3a223e7c59181784c3116ad5703e176cbb3ac --- /dev/null +++ b/qasper-1105/instruction.md @@ -0,0 +1,176 @@ +Name of Paper: Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer + +Question: How is the correct number of experts to use decided? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Conditional Computation", + "Our Approach: The Sparsely-Gated Mixture-of-Experts Layer", + "Related work on Mixtures of Experts", + "The Structure of the Mixture-of-Experts layer", + "Gating Network", + "The Shrinking Batch Problem", + "Network Bandwidth", + "Balancing Expert Utilization", + "1 Billion Word Language Modeling Benchmark", + "100 Billion Word Google News Corpus", + "Machine Translation (Single Language Pair)", + "Multilingual Machine Translation", + "Conclusion", + "Appendices", + "Load-Balancing Loss", + "Hierachical Mixture of Experts", + "1 Billion Word Language Modeling Benchmark - Experimental Details", + "100 Billion Word Google News Corpus - Experimental Details", + "Machine Translation - Experimental Details", + "Strictly Balanced Gating", + "Attention Function" + ], + "paragraphs": [ + [ + "Exploiting scale in both training data and model size has been central to the success of deep learning. When datasets are sufficiently large, increasing the capacity (number of parameters) of neural networks can give much better prediction accuracy. This has been shown in domains such as text BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 , images BIBREF4 , BIBREF5 , and audio BIBREF6 , BIBREF7 . For typical deep learning models, where the entire model is activated for every example, this leads to a roughly quadratic blow-up in training costs, as both the model size and the number of training examples increase. Unfortunately, the advances in computing power and distributed computation fall short of meeting such demand.", + "Various forms of conditional computation have been proposed as a way to increase model capacity without a proportional increase in computational costs BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 . In these schemes, large parts of a network are active or inactive on a per-example basis. The gating decisions may be binary or sparse and continuous, stochastic or deterministic. Various forms of reinforcement learning and back-propagation are proposed for trarining the gating decisions.", + "While these ideas are promising in theory, no work to date has yet demonstrated massive improvements in model capacity, training time, or model quality. We blame this on a combination of the following challenges:", + "Modern computing devices, especially GPUs, are much faster at arithmetic than at branching. Most of the works above recognize this and propose turning on/off large chunks of the network with each gating decision.", + "Large batch sizes are critical for performance, as they amortize the costs of parameter transfers and updates. Conditional computation reduces the batch sizes for the conditionally active chunks of the network.", + "Network bandwidth can be a bottleneck. A cluster of GPUs may have computational power thousands of times greater than the aggregate inter-device network bandwidth. To be computationally efficient, the relative computational versus network demands of an algorithm must exceed this ratio. Embedding layers, which can be seen as a form of conditional computation, are handicapped by this very problem. Since the embeddings generally need to be sent across the network, the number of (example, parameter) interactions is limited by network bandwidth instead of computational capacity.", + "Depending on the scheme, loss terms may be necessary to achieve the desired level of sparsity per-chunk and/or per example. BIBREF13 use three such terms. These issues can affect both model quality and load-balancing.", + "Model capacity is most critical for very large data sets. The existing literature on conditional computation deals with relatively small image recognition data sets consisting of up to 600,000 images. It is hard to imagine that the labels of these images provide a sufficient signal to adequately train a model with millions, let alone billions of parameters.", + "In this work, we for the first time address all of the above challenges and finally realize the promise of conditional computation. We obtain greater than 1000x improvements in model capacity with only minor losses in computational efficiency and significantly advance the state-of-the-art results on public language modeling and translation data sets." + ], + [ + "Our approach to conditional computation is to introduce a new type of general purpose neural network component: a Sparsely-Gated Mixture-of-Experts Layer (MoE). The MoE consists of a number of experts, each a simple feed-forward neural network, and a trainable gating network which selects a sparse combination of the experts to process each input (see Figure FIGREF8 ). All parts of the network are trained jointly by back-propagation.", + "While the introduced technique is generic, in this paper we focus on language modeling and machine translation tasks, which are known to benefit from very large models. In particular, we apply a MoE convolutionally between stacked LSTM layers BIBREF15 , as in Figure FIGREF8 . The MoE is called once for each position in the text, selecting a potentially different combination of experts at each position. The different experts tend to become highly specialized based on syntax and semantics (see Appendix SECREF84 Table TABREF92 ). On both language modeling and machine translation benchmarks, we improve on best published results at a fraction of the computational cost." + ], + [ + "Since its introduction more than two decades ago BIBREF16 , BIBREF17 , the mixture-of-experts approach has been the subject of much research. Different types of expert architectures hae been proposed such as SVMs BIBREF18 , Gaussian Processes BIBREF19 , BIBREF20 , BIBREF21 , Dirichlet Processes BIBREF22 , and deep networks. Other work has focused on different expert configurations such as a hierarchical structure BIBREF23 , infinite numbers of experts BIBREF24 , and adding experts sequentially BIBREF25 . BIBREF26 suggest an ensemble model in the format of mixture of experts for machine translation. The gating network is trained on a pre-trained ensemble NMT model.", + "The works above concern top-level mixtures of experts. The mixture of experts is the whole model. BIBREF10 introduce the idea of using multiple MoEs with their own gating networks as parts of a deep model. It is intuitive that the latter approach is more powerful, since complex problems may contain many sub-problems each requiring different experts. They also allude in their conclusion to the potential to introduce sparsity, turning MoEs into a vehicle for computational computation.", + "Our work builds on this use of MoEs as a general purpose neural network component. While BIBREF10 uses two stacked MoEs allowing for two sets of gating decisions, our convolutional application of the MoE allows for different gating decisions at each position in the text. We also realize sparse gating and demonstrate its use as a practical way to massively increase model capacity." + ], + [ + "The Mixture-of-Experts (MoE) layer consists of a set of INLINEFORM0 \u201cexpert networks\" INLINEFORM1 , and a \u201cgating network\" INLINEFORM2 whose output is a sparse INLINEFORM3 -dimensional vector. Figure FIGREF8 shows an overview of the MoE module. The experts are themselves neural networks, each with their own parameters. Although in principle we only require that the experts accept the same sized inputs and produce the same-sized outputs, in our initial investigations in this paper, we restrict ourselves to the case where the models are feed-forward networks with identical architectures, but with separate parameters.", + "Let us denote by INLINEFORM0 and INLINEFORM1 the output of the gating network and the output of the INLINEFORM2 -th expert network for a given input INLINEFORM3 . The output INLINEFORM4 of the MoE module can be written as follows: DISPLAYFORM0 ", + "We save computation based on the sparsity of the output of INLINEFORM0 . Wherever INLINEFORM1 , we need not compute INLINEFORM2 . In our experiments, we have up to thousands of experts, but only need to evaluate a handful of them for every example. If the number of experts is very large, we can reduce the branching factor by using a two-level hierarchical MoE. In a hierarchical MoE, a primary gating network chooses a sparse weighted combination of \u201cexperts\", each of which is itself a secondary mixture-of-experts with its own gating network. In the following we focus on ordinary MoEs. We provide more details on hierarchical MoEs in Appendix SECREF60 .", + "Our implementation is related to other models of conditional computation. A MoE whose experts are simple weight matrices is similar to the parameterized weight matrix proposed in BIBREF12 . A MoE whose experts have one hidden layer is similar to the block-wise dropout described in BIBREF13 , where the dropped-out layer is sandwiched between fully-activated layers." + ], + [ + "A simple choice of non-sparse gating function BIBREF17 is to multiply the input by a trainable weight matrix INLINEFORM0 and then apply the INLINEFORM1 function. DISPLAYFORM0 ", + "We add two components to the Softmax gating network: sparsity and noise. Before taking the softmax function, we add tunable Gaussian noise, then keep only the top k values, setting the rest to INLINEFORM0 (which causes the corresponding gate values to equal 0). The sparsity serves to save computation, as described above. While this form of sparsity creates some theoretically scary discontinuities in the output of gating function, we have not yet observed this to be a problem in practice. The noise term helps with load balancing, as will be discussed in Appendix SECREF51 . The amount of noise per component is controlled by a second trainable weight matrix INLINEFORM1 . DISPLAYFORM0 DISPLAYFORM1 ", + "We train the gating network by simple back-propagation, along with the rest of the model. If we choose INLINEFORM0 , the gate values for the top k experts have nonzero derivatives with respect to the weights of the gating network. This type of occasionally-sensitive behavior is described in BIBREF9 with respect to noisy rectifiers. Gradients also back-propagate through the gating network to its inputs. Our method differs here from BIBREF13 who use boolean gates and a REINFORCE-style approach to train the gating network." + ], + [ + "On modern CPUs and GPUs, large batch sizes are necessary for computational efficiency, so as to amortize the overhead of parameter loads and updates. If the gating network chooses INLINEFORM0 out of INLINEFORM1 experts for each example, then for a batch of INLINEFORM2 examples, each expert receives a much smaller batch of approximately INLINEFORM3 examples. This causes a naive MoE implementation to become very inefficient as the number of experts increases. The solution to this shrinking batch problem is to make the original batch size as large as possible. However, batch size tends to be limited by the memory necessary to store activations between the forwards and backwards passes. We propose the following techniques for increasing the batch size:", + "In a conventional distributed training setting, multiple copies of the model on different devices asynchronously process distinct batches of data, and parameters are synchronized through a set of parameter servers. In our technique, these different batches run synchronously so that they can be combined for the MoE layer. We distribute the standard layers of the model and the gating network according to conventional data-parallel schemes, but keep only one shared copy of each expert. Each expert in the MoE layer receives a combined batch consisting of the relevant examples from all of the data-parallel input batches. The same set of devices function as data-parallel replicas (for the standard layers and the gating networks) and as model-parallel shards (each hosting a subset of the experts). If the model is distributed over INLINEFORM0 devices, and each device processes a batch of size INLINEFORM1 , each expert receives a batch of approximately INLINEFORM2 examples. Thus, we achieve a factor of INLINEFORM3 improvement in expert batch size.", + "In the case of a hierarchical MoE (Section SECREF60 ), the primary gating network employs data parallelism, and the secondary MoEs employ model parallelism. Each secondary MoE resides on one device.", + "This technique allows us to increase the number of experts (and hence the number of parameters) by proportionally increasing the number of devices in the training cluster. The total batch size increases, keeping the batch size per expert constant. The memory and bandwidth requirements per device also remain constant, as do the step times, as does the amount of time necessary to process a number of training examples equal to the number of parameters in the model. It is our goal to train a trillion-parameter model on a trillion-word corpus. We have not scaled our systems this far as of the writing of this paper, but it should be possible by adding more hardware.", + "In our language models, we apply the same MoE to each time step of the previous layer. If we wait for the previous layer to finish, we can apply the MoE to all the time steps together as one big batch. Doing so increases the size of the input batch to the MoE layer by a factor of the number of unrolled time steps.", + "We suspect that even more powerful models may involve applying a MoE recurrently. For example, the weight matrices of a LSTM or other RNN could be replaced by a MoE. Sadly, such models break the convolutional trick from the last paragraph, since the input to the MoE at one timestep depends on the output of the MoE at the previous timestep. BIBREF27 describe a technique for drastically reducing the number of stored activations in an unrolled RNN, at the cost of recomputing forward activations. This would allow for a large increase in batch size." + ], + [ + "Another major performance concern in distributed computing is network bandwidth. Since the experts are stationary (see above) and the number of gating parameters is small, most of the communication involves sending the inputs and outputs of the experts across the network. To maintain computational efficiency, the ratio of an expert's computation to the size of its input and output must exceed the ratio of computational to network capacity of the computing device. For GPUs, this may be thousands to one. In our experiments, we use experts with one hidden layer containing thousands of RELU-activated units. Since the weight matrices in the expert have sizes INLINEFORM0 _ INLINEFORM1 _ INLINEFORM2 and INLINEFORM3 _ INLINEFORM4 _ INLINEFORM5 , the ratio of computation to input and output is equal to the size of the hidden layer. Conveniently, we can increase computational efficiency simply by using a larger hidden layer, or more hidden layers." + ], + [ + "We have observed that the gating network tends to converge to a state where it always produces large weights for the same few experts. This imbalance is self-reinforcing, as the favored experts are trained more rapidly and thus are selected even more by the gating network. BIBREF10 describe the same phenomenon, and use a hard constraint at the beginning of training to avoid this local minimum. BIBREF13 include a soft constraint on the batch-wise average of each gate.", + "We take a soft constraint approach. We define the importance of an expert relative to a batch of training examples to be the batchwise sum of the gate values for that expert. We define an additional loss INLINEFORM0 , which is added to the overall loss function for the model. This loss is equal to the square of the coefficient of variation of the set of importance values, multiplied by a hand-tuned scaling factor INLINEFORM1 . This additional loss encourages all experts to have equal importance. DISPLAYFORM0 DISPLAYFORM1 ", + "While this loss function can ensure equal importance, experts may still receive very different numbers of examples. For example, one expert may receive a few examples with large weights, and another may receive many examples with small weights. This can cause memory and performance problems on distributed hardware. To solve this problem, we introduce a second loss function, INLINEFORM0 , which ensures balanced loads. Appendix SECREF51 contains the definition of this function, along with experimental results." + ], + [ + "This dataset, introduced by BIBREF28 consists of shuffled unique sentences from news articles, totaling approximately 829 million words, with a vocabulary of 793,471 words.", + "The best previously published results BIBREF2 use models consisting of one or more stacked Long Short-Term Memory (LSTM) layers BIBREF15 , BIBREF29 . The number of parameters in the LSTM layers of these models vary from 2 million to 151 million. Quality increases greatly with parameter count, as do computational costs. Results for these models form the top line of Figure FIGREF32 -right.", + "Our models consist of two stacked LSTM layers with a MoE layer between them (see Figure FIGREF8 ). We vary the sizes of the layers and the number of experts. For full details on model architecture, training regimen, additional baselines and results, see Appendix SECREF65 .", + "To investigate the effects of adding capacity, we trained a series of MoE models all with roughly equal computational costs: about 8 million multiply-and-adds per training example per timestep in the forwards pass, excluding the softmax layer. We call this metric (ops/timestep). We trained models with flat MoEs containing 4, 32, and 256 experts, and models with hierarchical MoEs containing 256, 1024, and 4096 experts. Each expert had about 1 million parameters. For all the MoE layers, 4 experts were active per input.", + "The results of these models are shown in Figure FIGREF32 -left. The model with 4 always-active experts performed (unsurprisingly) similarly to the computationally-matched baseline models, while the largest of the models (4096 experts) achieved an impressive 24% lower perplexity on the test set.", + "In addition to the largest model from the previous section, we trained two more MoE models with similarly high capacity (4 billion parameters), but higher computation budgets. These models had larger LSTMs, and fewer but larger and experts. Details can be found in Appendix UID77 . Results of these three models form the bottom line of Figure FIGREF32 -right. Table TABREF33 compares the results of these models to the best previously-published result on this dataset . Even the fastest of these models beats the best published result (when controlling for the number of training epochs), despite requiring only 6% of the computation.", + "We trained our models using TensorFlow BIBREF30 on clusters containing 16-32 Tesla K40 GPUs. For each of our models, we determine computational efficiency in TFLOPS/GPU by dividing the number of floating point operations required to process one training batch by the observed step time and the number of GPUs in the cluster. The operation counts used here are higher than the ones we report in our ops/timestep numbers in that we include the backwards pass, we include the importance-sampling-based training of the softmax layer, and we count a multiply-and-add as two separate operations. For all of our MoE models, the floating point operations involved in the experts represent between 37% and 46% of the total.", + "For our baseline models wtih no MoE, observed computational efficiency ranged from 1.07-1.29 TFLOPS/GPU. For our low-computation MoE models, computation efficiency ranged from 0.74-0.90 TFLOPS/GPU, except for the 4-expert model which did not make full use of the available parallelism. Our highest-computation MoE model was more efficient at 1.56 TFLOPS/GPU, likely due to the larger matrices. These numbers represent a significant fraction of the theoretical maximum of 4.29 TFLOPS/GPU claimed by NVIDIA. Detailed results are in Appendix SECREF65 , Table TABREF76 ." + ], + [ + "On the 1-billion-word corpus, adding additional capacity seems to produce diminishing returns as the number of parameters in the MoE layer exceeds 1 billion, as can be seen in Figure FIGREF32 -left. We hypothesized that for a larger training set, even higher capacities would produce significant quality improvements.", + "We constructed a similar training set consisting of shuffled unique sentences from Google's internal news corpus, totalling roughly 100 billion words. Similarly to the previous section, we tested a series of models with similar computational costs of about 8 million ops/timestep. In addition to a baseline LSTM model, we trained models augmented with MoE layers containing 32, 256, 1024, 4096, 16384, 65536, and 131072 experts. This corresponds to up to 137 billion parameters in the MoE layer. Details on architecture, training, and results are given in Appendix SECREF78 .", + "Figure FIGREF37 shows test perplexity as a function of capacity after training on 10 billion words (top line) and 100 billion words (bottom line). When training over the full 100 billion words, test perplexity improves significantly up to 65536 experts (68 billion parameters), dropping 39% lower than the computationally matched baseline, but degrades at 131072 experts, possibly a result of too much sparsity. The widening gap between the two lines demonstrates (unsurprisingly) that increased model capacity helps more on larger training sets.", + "Even at 65536 experts (99.994% layer sparsity), computational efficiency for the model stays at a respectable 0.72 TFLOPS/GPU." + ], + [ + "Our model was a modified version of the GNMT model described in BIBREF3 . To reduce computation, we decreased the number of LSTM layers in the encoder and decoder from 9 and 8 to 3 and 2 respectively. We inserted MoE layers in both the encoder (between layers 2 and 3) and the decoder (between layers 1 and 2). Each MoE layer contained up to 2048 experts each with about two million parameters, adding a total of about 8 billion parameters to the models. Further details on model architecture, testing procedure and results can be found in Appendix SECREF84 .", + "We benchmarked our method on the WMT'14 En INLINEFORM0 Fr and En INLINEFORM1 De corpora, whose training sets have 36M sentence pairs and 5M sentence pairs, respectively. The experimental protocols were also similar to those in BIBREF3 : newstest2014 was used as the test set to compare against previous work BIBREF31 , BIBREF32 , BIBREF3 , while the combination of newstest2012 and newstest2013 was used as the development set. We also tested the same model on a Google's Production English to French data.", + "Tables TABREF42 , TABREF43 , and TABREF44 show the results of our largest models, compared with published results. Our approach achieved BLEU scores of 40.56 and 26.03 on the WMT'14 En INLINEFORM0 Fr and En INLINEFORM1 De benchmarks. As our models did not use RL refinement, these results constitute significant gains of 1.34 and 1.12 BLEU score on top of the strong baselines in BIBREF3 . The perplexity scores are also better. On the Google Production dataset, our model achieved 1.01 higher test BLEU score even after training for only one sixth of the time." + ], + [ + " BIBREF35 train a single GNMT BIBREF3 model on a very large combined dataset of twelve language pairs. Results are somewhat worse than those for 12 separately trained single-pair GNMT models. This is not surprising, given that the twelve models have 12 times the capacity and twelve times the aggregate training of the one model. We repeat this experiment with a single MoE-augmented model. See Appendix SECREF84 for details on model architecture. We train our model on the same dataset as BIBREF35 and process the same number of training examples (about 3 billion sentence pairs). Our training time was shorter due to the lower computational budget of our model.", + "Results for the single-pair GNMT models, the multilingual GNMT model and the multilingual MoE model are given in Table TABREF50 . The MoE model achieves 19% lower perplexity on the dev set than the multilingual GNMT model. On BLEU score, the MoE model significantly beats the multilingual GNMT model on 11 of the 12 language pairs (by as much as 5.84 points), and even beats the monolingual GNMT models on 8 of 12 language pairs. The poor performance on English INLINEFORM0 Korean seems to be a result of severe overtraining, as for the rarer language pairs a small number of real examples were highly oversampled in the training corpus.", + "" + ], + [ + "This work is the first to demonstrate major wins from conditional computation in deep networks. We carefully identified the design considerations and challenges of conditional computing and addressed them with a combination of algorithmic and engineering solutions. While we focused on text, conditional computation may help in other domains as well, provided sufficiently large training sets. We look forward to seeing many novel implementations and applications of conditional computation in the years to come." + ], + [ + "tocsectionAppendices" + ], + [ + "As discussed in section SECREF4 , for load-balancing purposes, we want to define an additional loss function to encourage experts to receive roughly equal numbers of training examples. Unfortunately, the number of examples received by an expert is a discrete quantity, so it can not be used in back-propagation. Instead, we define a smooth estimator INLINEFORM0 of the number of examples assigned to each expert for a batch INLINEFORM1 of inputs. The smoothness allows us to back-propagate gradients through the estimator. This is the purpose of the noise term in the gating function. We define INLINEFORM2 as the probability that INLINEFORM3 is nonzero, given a new random choice of noise on element INLINEFORM4 , but keeping the already-sampled choices of noise on the other elements. To compute INLINEFORM5 , we note that the INLINEFORM6 is nonzero if and only if INLINEFORM7 is greater than the INLINEFORM8 -greatest element of INLINEFORM9 excluding itself. The probability works out to be: DISPLAYFORM0 ", + "Where INLINEFORM0 means the kth highest component of INLINEFORM1 , excluding component INLINEFORM2 . Simplifying, we get: DISPLAYFORM0 ", + "Where INLINEFORM0 is the CDF of the standard normal distribution. DISPLAYFORM0 ", + "We can now define the load loss to be the square of the coefficient of variation of the load vector, multiplied by a hand-tuned scaling factor INLINEFORM0 . DISPLAYFORM0 ", + "To avoid out-of-memory errors, we need to initialize the network in a state of approximately equal expert load (since the soft constraints need some time to work). To accomplish this, we initialize the matrices INLINEFORM0 and INLINEFORM1 to all zeros, which yields no signal and some noise.", + "We trained a set of models with identical architecture (the MoE-256 model described in Appendix SECREF65 ), using different values of INLINEFORM0 and INLINEFORM1 . We trained each model for 10 epochs, then measured perplexity on the test set. We also measured the coefficients of variation in INLINEFORM2 and INLINEFORM3 , as well as ratio of the load on the most overloaded expert to the average load. This last value is significant for load balancing purposes on distributed hardware. All of these metrics were averaged over several training batches.", + "Results are reported in Table TABREF58 . All the combinations containing at least one the two losses led to very similar model quality, where having no loss was much worse. Models with higher values of INLINEFORM0 had lower loads on the most overloaded expert." + ], + [ + "If the number of experts is very large, we can reduce the branching factor by using a two-level hierarchical MoE. In a hierarchical MoE, a primary gating network chooses a sparse weighted combination of \u201cexperts\", each of which is itself a secondary mixture-of-experts with its own gating network. If the hierarchical MoE consists of INLINEFORM0 groups of INLINEFORM1 experts each, we denote the primary gating network by INLINEFORM2 , the secondary gating networks by INLINEFORM3 , and the expert networks by INLINEFORM4 . The output of the MoE is given by: DISPLAYFORM0 ", + "Our metrics of expert utilization change to the following: DISPLAYFORM0 DISPLAYFORM1 ", + " INLINEFORM0 and INLINEFORM1 deonte the INLINEFORM2 functions for the primary gating network and INLINEFORM3 secondary gating network respectively. INLINEFORM4 denotes the subset of INLINEFORM5 for which INLINEFORM6 .", + "It would seem simpler to let INLINEFORM0 , but this would not have a gradient with respect to the primary gating network, so we use the formulation above." + ], + [ + "Our model consists of five layers: a word embedding layer, a recurrent Long Short-Term Memory (LSTM) layer BIBREF15 , BIBREF29 , a MoE layer, a second LSTM layer, and a softmax layer. The dimensionality of the embedding layer, the number of units in each LSTM layer, and the input and output dimensionality of the MoE layer are all equal to 512. For every layer other than the softmax, we apply drouput BIBREF43 to the layer output, dropping each activation with probability INLINEFORM0 , otherwise dividing by INLINEFORM1 . After dropout, the output of the previous layer is added to the layer output. This residual connection encourages gradient flow BIBREF37 .", + "Each expert in the MoE layer is a feed forward network with one ReLU-activated hidden layer of size 1024 and an output layer of size 512. Thus, each expert contains INLINEFORM0 parameters. The output of the MoE layer is passed through a sigmoid function before dropout. We varied the number of experts between models, using ordinary MoE layers with 4, 32 and 256 experts and hierarchical MoE layers with 256, 1024 and 4096 experts. We call the resulting models MoE-4, MoE-32, MoE-256, MoE-256-h, MoE-1024-h and MoE-4096-h. For the hierarchical MoE layers, the first level branching factor was 16, corresponding to the number of GPUs in our cluster. We use Noisy-Top-K Gating (see Section UID14 ) with INLINEFORM1 for the ordinary MoE layers and INLINEFORM2 at each level of the hierarchical MoE layers. Thus, each example is processed by exactly 4 experts for a total of 4M ops/timestep. The two LSTM layers contribute 2M ops/timestep each for the desired total of 8M.", + "The MoE-4 model does not employ sparsity, since all 4 experts are always used. In addition, we trained four more computationally-matched baseline models with no sparsity:", + "MoE-1-Wide: The MoE layer consists of a single \"expert\" containing one ReLU-activated hidden layer of size 4096.", + "MoE-1-Deep: The MoE layer consists of a single \"expert\" containing four ReLU-activated hidden layers, each with size 1024.", + "4xLSTM-512: We replace the MoE layer with two additional 512-unit LSTM layers.", + "LSTM-2048-512: The model contains one 2048-unit LSTM layer (and no MoE). The output of the LSTM is projected down to 512 dimensions BIBREF41 . The next timestep of the LSTM receives the projected output. This is identical to one of the models published in BIBREF2 . We re-ran it to account for differences in training regimen, and obtained results very similar to the published ones.", + "The models were trained on a cluster of 16 K40 GPUs using the synchronous method described in Section SECREF3 . Each batch consisted of a set of sentences totaling roughly 300,000 words. In the interest of time, we limited training to 10 epochs, (27,000 steps). Training took 12-16 hours for all models, except for MoE-4, which took 18 hours (since all the expert computation was performed on only 4 of 16 GPUs). We used the Adam optimizer BIBREF39 . The base learning rate was increased linearly for the first 1000 training steps, and decreased after that so as to be proportional to the inverse square root of the step number. The Softmax output layer was trained efficiently using importance sampling similarly to the models in BIBREF2 . For each model, we performed a hyper-parmeter search to find the best dropout probability, in increments of 0.1.", + "To ensure balanced expert utilization we set INLINEFORM0 and INLINEFORM1 , as described in Section SECREF4 and Appendix SECREF51 .", + "We evaluate our model using perplexity on the holdout dataset, used by BIBREF28 , BIBREF2 . We follow the standard procedure and sum over all the words including the end of sentence symbol. Results are reported in Table TABREF76 . For each model, we report the test perplexity, the computational budget, the parameter counts, the value of INLINEFORM0 , and the computational efficiency.", + "We ran two additional models (MoE-34M and MoE-143M) to investigate the effects of adding more computation in the presence of a large MoE layer. These models have computation budgets of 34M and 143M ops/timestep. Similar to the models above, these models use a MoE layer between two LSTM layers. The dimensionality of the embedding layer, and the input and output dimensionality of the MoE layer are set to 1024 instead of 512. For MoE-34M, the LSTM layers have 1024 units. For MoE-143M, the LSTM layers have 4096 units and an output projection of size 1024 BIBREF41 . MoE-34M uses a hierarchical MoE layer with 1024 experts, each with a hidden layer of size 2048. MoE-143M uses a hierarchical MoE layer with 256 experts, each with a hidden layer of size 8192. Both models have 4B parameters in the MoE layers. We searched for the best INLINEFORM0 for each model, and trained each model for 10 epochs.", + "The two models achieved test perplexity of INLINEFORM0 and INLINEFORM1 respectively, showing that even in the presence of a large MoE, more computation is still useful. Results are reported at the bottom of Table TABREF76 . The larger of the two models has a similar computational budget to the best published model from the literature, and training times are similar. Comparing after 10 epochs, our model has a lower test perplexity by INLINEFORM2 ." + ], + [ + "The models are similar in structure to the 8-million-operations-per-timestep models described in the previous section. We vary the number of experts between models, using an ordinary MoE layer with 32 experts and hierarchical MoE layers with 256, 1024, 4096, 16384, 65536 and 131072 experts. For the hierarchical MoE layers, the first level branching factors are 32, 32, 64, 128, 256 and 256, respectively.", + "Models are trained on a cluster of 32 Tesla K40 GPUs, except for the last two models, which are trained on clusters of 64 and 128 GPUs so as to have enough memory for all the parameters. For all models, training batch sizes are approximately 2.5 million words. Models are trained once-through over about 100 billion words.", + "We implement several memory optimizations in order to fit up to 1 billion parameters per GPU. First, we do not store the activations of the hidden layers of the experts, but instead recompute them on the backwards pass. Secondly, we modify the optimizer on the expert parameters to require less auxiliary storage:", + "The Adam optimizer BIBREF39 keeps first and second moment estimates of the per-parameter gradients. This triples the required memory. To avoid keeping a first-moment estimator, we set INLINEFORM0 . To reduce the size of the second moment estimator, we replace it with a factored approximation. For a matrix of parameters, instead of maintaining a full matrix of second-moment estimators, we maintain vectors of row-wise and column-wise averages of that matrix. At each step, the matrix of estimators is taken to be the outer product of those two vectors divided by the mean of either one. This technique could similarly be applied to Adagrad BIBREF36 .", + "We evaluate our model using perplexity on a holdout dataset. Results are reported in Table TABREF81 . Perplexity after 100 billion training words is 39% lower for the 68-billion-parameter MoE model than for the baseline model. It is notable that the measured computational efficiency of the largest model (0.30 TFLOPS/GPU) is very low compared to the other models. This is likely a result of the fact that, for purposes of comparison to the other models, we did not increase the training batch size proportionally to the number of GPUs. For comparison, we include results for a computationally matched baseline model consisting of 4 LSTMs, and for an unpruned 5-gram model with Kneser-Ney smoothing BIBREF40 ." + ], + [ + "Our model is a modified version of the GNMT model described in BIBREF3 . To reduce computation, we decrease the number of LSTM layers in the encoder and decoder from 9 and 8 to 3 and 2 respectively. We insert MoE layers in both the encoder (between layers 2 and 3) and the decoder (between layers 1 and 2). We use an attention mechanism between the encoder and decoder, with the first decoder LSTM receiving output from and providing input for the attention . All of the layers in our model have input and output dimensionality of 512. Our LSTM layers have 2048 hidden units, with a 512-dimensional output projection. We add residual connections around all LSTM and MoE layers to encourage gradient flow BIBREF37 . Similar to GNMT, to effectively deal with rare words, we used sub-word units (also known as \u201cwordpieces\") BIBREF42 for inputs and outputs in our system.", + "We use a shared source and target vocabulary of 32K wordpieces. We also used the same beam search technique as proposed in BIBREF3 .", + "We train models with different numbers of experts in the MoE layers. In addition to a baseline model with no MoE layers, we train models with flat MoE layers containing 32 experts, and models with hierarchical MoE layers containing 512 and 2048 experts. The flat MoE layers use INLINEFORM0 and the hierarchical MoE models use INLINEFORM1 at each level of the gating network. Thus, each input is processed by exactly 4 experts in each MoE layer. Each expert in the MoE layer is a feed forward network with one hidden layer of size 2048 and ReLU activation. Thus, each expert contains INLINEFORM2 parameters. The output of the MoE layer is passed through a sigmoid function. We use the strictly-balanced gating function described in Appendix SECREF93 .", + "We used the same model architecture as for the single-language-pair models, with the following exceptions: We used noisy-top-k gating as described in Section UID14 , not the scheme from Appendix SECREF93 . The MoE layers in the encoder and decoder are non-hierarchical MoEs with INLINEFORM0 experts, and INLINEFORM1 . Each expert has a larger hidden layer of size 8192. This doubles the amount of computation in the MoE layers, raising the computational budget of the entire model from 85M to 102M ops/timestep.", + "We trained our networks using the Adam optimizer BIBREF39 . The base learning rate was increased linearly for the first 2000 training steps, held constant for an additional 8000 steps, and decreased after that so as to be proportional to the inverse square root of the step number. For the single-language-pair models, similarly to BIBREF3 , we applied dropout BIBREF43 to the output of all embedding, LSTM and MoE layers, using INLINEFORM0 . Training was done synchronously on a cluster of up to 64 GPUs as described in section SECREF3 . Each training batch consisted of a set of sentence pairs containing roughly 16000 words per GPU.", + "To ensure balanced expert utilization we set INLINEFORM0 and INLINEFORM1 , as described in Section SECREF4 and Appendix SECREF51 .", + "We evaluated our models using the perplexity and the standard BLEU score metric. We reported tokenized BLEU score as computed by the multi-bleu.pl script, downloaded from the public implementation of Moses (on Github), which was also used in BIBREF31 .", + "Tables TABREF42 , TABREF43 and TABREF44 in Section SECREF39 show comparisons of our results to other published methods. Figure FIGREF91 shows test perplexity as a function of number of words in the (training data's) source sentences processed for models with different numbers of experts. As can be seen from the Figure, as we increased the number of experts to approach 2048, the test perplexity of our model continued to improve.", + "We found that the experts indeed become highly specialized by syntax and/or semantics, as can be seen in Table TABREF92 . For example, one expert is used when the indefinite article \u201ca\" introduces the direct object in a verb phrase indicating importance or leadership." + ], + [ + "Due to some peculiarities in our infrastructure which have since been fixed, at the time we ran some of the machine translation experiments, our models ran faster if every expert received exactly the same batch size. To accommodate this, we used a different gating function which we describe below.", + "Recall that we define the softmax gating function to be: DISPLAYFORM0 ", + "To obtain a sparse gating vector, we multiply INLINEFORM0 component-wise with a sparse mask INLINEFORM1 and normalize the output. The mask itself is a function of INLINEFORM2 and specifies which experts are assigned to each input example: DISPLAYFORM0 ", + "To implement top-k gating in this formulation, we would let INLINEFORM0 , where: DISPLAYFORM0 ", + "To force each expert to receive the exact same number of examples, we introduce an alternative mask function, INLINEFORM0 , which operates over batches of input vectors. Instead of keeping the top INLINEFORM1 values per example, we keep the top INLINEFORM2 values per expert across the training batch, where INLINEFORM3 , so that each example is sent to an average of INLINEFORM4 experts. DISPLAYFORM0 ", + "As our experiments suggest and also observed in BIBREF38 , using a batchwise function during training (such as INLINEFORM0 ) requires modifications to the inference when we may not have a large batch of examples. Our solution to this is to train a vector INLINEFORM1 of per-expert threshold values to approximate the effects of the batchwise mask. We use the following mask at inference time: DISPLAYFORM0 ", + "To learn the threshold values, we apply an additional loss at training time which is minimized when the batchwise mask and the threshold mask are identical. DISPLAYFORM0 " + ], + [ + "The attention mechanism described in GNMT BIBREF3 involves a learned \u201cAttention Function\" INLINEFORM0 which takes a \u201csource vector\" INLINEFORM1 and a \u201ctarget vector\" INLINEFORM2 , and must be computed for every source time step INLINEFORM3 and target time step INLINEFORM4 . In GNMT, the attention function is implemented as a feed forward neural network with a hidden layer of size INLINEFORM5 . It can be expressed as: DISPLAYFORM0 ", + "Where INLINEFORM0 and INLINEFORM1 are trainable weight matrices and INLINEFORM2 is a trainable weight vector.", + "For performance reasons, in our models, we used a slightly different attention function: DISPLAYFORM0 ", + "With our attention function, we can simultaneously compute the attention function on multiple source time steps and multiple target time steps using optimized matrix multiplications. We found little difference in quality between the two functions." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1110/instruction.md b/qasper-1110/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..500c73c94c9b914b770cf46a59b382d1a21fc0a8 --- /dev/null +++ b/qasper-1110/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Evaluation of basic modules for isolated spelling error correction in Polish texts + +Question: What methods are tested in PIEWi? \ No newline at end of file diff --git a/qasper-1111/instruction.md b/qasper-1111/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ea31ea5cdb5c8b541a5a111a3f92431f0b48c351 --- /dev/null +++ b/qasper-1111/instruction.md @@ -0,0 +1,67 @@ +Name of Paper: Evaluation of basic modules for isolated spelling error correction in Polish texts + +Question: Which specific error correction solutions have been proposed for specialized corpora in the past? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Problems of spelling correction for Polish", + "Baseline methods", + "Vector distance", + "Recurrent neural networks", + "Experimental setup", + "Results", + "Conclusion" + ], + "paragraphs": [ + [ + "Spelling error correction is a fundamental NLP task. Most language processing applications benefit greatly from being provided clean texts for their best performance. Human users of computers also often expect competent help in making spelling of their texts correct.", + "Because of the lack of tests of many common spelling correction methods for Polish, it is useful to establish how they perform in a simple scenario. We constrain ourselves to the pure task of isolated correction of non-word errors. They are traditionally separated in error correction literature BIBREF0 . Non-word errors are here incorrect word forms that not only differ from what was intended, but also do not constitute another, existing word themselves. Much of the initial research on error correction focused on this simple task, tackled without means of taking the context of the nearest words into account.", + "It is true that, especially in the case of neural networks, it is often possible and desirable to combine problems of error detection, correction and context awareness into one task trained with a supervised training procedure. In language correction research for English language also grammatical and regular spelling errors have been treated uniformly with much success BIBREF1 .", + "However, when more traditional methods are used, because of their predictability and interpretability for example, one can mix and match various approaches to dealing with the subproblems of detection, correction and context handling (often equivalent to employing some kind of a language model). We call it a modular approach to building spelling error correction systems. There is recent research where this paradigm was applied, interestingly, to convolutional networks trained separately for various subtasks BIBREF2 . In similar setups it is more useful to assess abilities of various solutions in isolation. The exact architecture of a spelling correction system should depend on characteristics of texts it will work on.", + "Similar considerations eliminated from our focus handcrafted solutions for the whole spelling correction pipeline, primarily the LanguageTool BIBREF3 . Its performance in fixing spelling of Polish tweets was already tested BIBREF4 . For our purposes it would be given an unfair advantage, since it is a rule-based system making heavy use of words in context of the error." + ], + [ + "Published work on language correction for Polish dates back at least to 1970s, when simplest Levenshtein distance solutions were used for cleaning mainframe inputs BIBREF5 , BIBREF6 . Spelling correction tests described in literature have tended to focus on one approach applied to a specific corpus. Limited examples include works on spellchecking mammography reports and tweets BIBREF7 , BIBREF4 . These works emphasized the importance of tailoring correction systems to specific problems of corpora they are applied to. For example, mammography reports suffer from poor typing, which in this case is a repetitive work done in relative hurry. Tweets, on the other hand, tend to contain emoticons and neologisms that can trick solutions based on rules and dictionaries, such as LanguageTool. The latter is, by itself, fairly well suited for Polish texts, since a number of extensions to the structure of this application was inspired by problems with morphology of Polish language BIBREF3 .", + "These existing works pointed out more general, potentially useful qualities specific to spelling errors in Polish language texts. It is, primarily, the problem of leaving out diacritical signs, or, more rarely, adding them in wrong places. This phenomenon stems from using a variant of the US keyboard layout, where combinations of AltGr with some alphabetic keys produces characters unique to Polish. When the user forgets or neglects to press the AltGr key, typos such as writing *olowek instead of o\u0142\u00f3wek appear. In fact, BIBREF4 managed to get substantial performance on Twitter corpus by using this \u201ddiacritical swapping\u201d alone." + ], + [ + "The methods that we evaluated are baselines are the ones we consider to be basic and with moderate potential of yielding particularly good results. Probably the most straightforward approach to error correction is selecting known words from a dictionary that are within the smallest edit distance from the error. We used the Levenshtein distance metric BIBREF8 implemented in Apache Lucene library BIBREF9 . It is a version of edit distance that treats deletions, insertions and replacements as adding one unit distance, without giving a special treatment to character swaps. The SGJP \u2013 Grammatical Dictionary of Polish BIBREF10 was used as the reference vocabulary.", + "Another simple approach is the aforementioned diacritical swapping, which is a term that we introduce here for referring to a solution inspired by the work of BIBREF4 . Namely, from the incorrect form we try to produce all strings obtainable by either adding or removing diacritical marks from characters. We then exclude options that are not present in SGJP, and select as the correction the one within the smallest edit distance from the error. It is possible for the number of such diacritically-swapped options to become very big. For example, the token Modlin-Zegrze-Pultusk-R\u00f3\u017can-Ostro\u0142\u0119ka-\u0141om\u017ca-Osowiec (taken from PlEWi corpus of spelling errors, see below) can yield over INLINEFORM0 states with this method, such as M\u00f3d\u0142i\u0144-\u017b\u0119grz\u0119-Pu\u0142tu\u015bk-Ro\u017a\u0105\u0144-\u00d3\u015btr\u00f3lek\u0105-L\u00f3mz\u0105-\u00d3\u015b\u00f3wi\u0119\u0107. The actual correction here is just fixing the \u0142 in Pu\u0142tusk. Hence we only try to correct in this way tokens that are shorter than 17 characters." + ], + [ + "A promising method, adapted from work on correcting texts by English language learners BIBREF11 , expands on the concept of selecting a correction nearest to the spelling error according to some notion of distance. Here, the Levenshtein distance is used in a weighted sum to cosine distance between word vectors. This is based on the observation that trained vectors models of distributional semantics contain also representations of spelling errors, if they were not pruned. Their representations tend to be similar to those of their correct counterparts. For example, the token enginir will appear in similar contexts as engineer, and therefore will be assigned a similar vector embedding.", + "The distance between two tokens INLINEFORM0 and INLINEFORM1 is thus defined as INLINEFORM2 ", + "Here INLINEFORM0 is just Levenshtein distance between strings, and INLINEFORM1 \u2013 cosine distance between vectors. INLINEFORM2 denotes the word vector for INLINEFORM3 . Both distance metrics are in our case roughly in the range [0,1] thanks to the scaling of edit distance performed automatically by Apache Lucene. We used a pretrained set of word embeddings of Polish BIBREF12 , obtained with the flavor word2vec procedure using skipgrams and negative sampling BIBREF13 ." + ], + [ + "Another powerful approach, if conceptually simple in linguistic terms, is using a character-based recurrent neural network. Here, we test uni- and bidirectional Long Short-Term Memory networks BIBREF14 that are fed characters of the error as their input and are expected to output its correct form, character after character. This is similar to traditional solutions conceptualizing the spelling error as a chain of characters, which are used as evidence to predict the most likely chain of replacements (original characters). This was done with n-gram methods, Markov chains and other probabilistic models BIBREF15 . Since nowadays neural networks enjoy a large awareness as an element of software infrastructure, with actively maintained packages readily available, their evaluation seems to be the most practically useful. We used the PyTorch BIBREF16 implementation of LSTM in particular.", + "The bidirectional version BIBREF17 of LSTM reads the character chains forward and backwards at the same time. Predictions from networks running in both directions are averaged.", + "In order to provide the network an additional, broad picture peek at the whole error form we also evaluated a setup where the internal state of LSTM cells, instead of being initialized randomly, is computed from an ELMo embedding BIBREF18 of the token. The ELMo embedder is capable of integrating linguistic information carried by the whole form (probably often not much in case of errors), as well as the string as a character chain. The latter is processed with a convolutional neural network. How this representation is constructed is informed by the whole corpus on which the embedder was trained. The pretrained ELMo model that we used BIBREF19 was trained on Wikipedia and Common Crawl corpora of Polish.", + "The ELMo embedding network outputs three layers as matrices, which are supposed to reflect subsequent compositional layers of language, from phonetic phenomena at the bottom to lexical ones at the top. A weighted sum of these layers is computed, with weights trained along with the LSTM error-correcting network. Then we apply a trained linear transformation, followed by INLINEFORM0 non-linearity: INLINEFORM1 ", + "(applied cellwise) in order to obtain the initial setting of parameters for the main LSTM. Our ELMo-augmented LSTM is bidirectional." + ], + [ + "PlEWi BIBREF20 is an early version of WikEd BIBREF21 error corpus, containing error type annotations allowing us to select only non-word errors for evaluation. Specifically, PlEWi supplied 550,755 [error, correction] pairs, from which 298,715 were unique. The corpus contains data extracted from histories of page versions of Polish Wikipedia. An algorithm designed by the corpus author determined where the changes were correcting spelling errors, as opposed to expanding content and disagreements among Wikipedia editors.", + "The corpus features texts that are descriptive rather than conversational, contain relatively many proper names and are more likely to have been at least skimmed by the authors before submitting for online publication. Error cases provided by PlEWi are, therefore, not a balanced representation of spelling errors in written Polish language. PlEWi does have the advantage of scale in comparison to existing literature, such as BIBREF4 operating on a set of only 740 annotated errors in tweets.", + "All methods were tested on a test subset of 25% of cases, with 75% left for training (where needed) and 5% for development.", + "The methods that required training \u2013 namely recurrent neural networks \u2013 had their loss measured as cross-entropy loss measure between correct character labels and predictions. This value was minimized with Adam algorithm BIBREF22 . The networks were trained for 35 epochs." + ], + [ + "The experimental results are presented in Table TABREF4 . Diacritic swapping showed a remarkably poor performance, despite promising mentions in existing literature. This might be explained by the already mentioned feature of Wikipedia edits, which can be expected to be to some degree self-reviewed before submission. This can very well limit the number of most trivial mistakes.", + "On the other hand, the vector distance method was able to bring a discernible improvement over pure Levenshtein distance, comparable even with the most basic LSTM. It is possible that assigning more fine-tuned weights to edit distance and semantic distance would make the quality of predictions even higher. The idea of using vector space measurements explicitly can be also expanded if we were to consider the problem of contextualizing corrections. For example, the semantic distance of proposed corrections to the nearest words is likely to carry much information about their appropriateness. Looking from another angle, searching for words that seem semantically off in context may be a good heuristic for detecting errors that are not nonword (that is, they lead to wrong forms appearing in text which are nevertheless in-vocabulary).", + "The good performance of recurrent network methods is hardly a surprise, given observed effectiveness of neural networks in many NLP tasks in the recent decade. It seems that bidirectional LSTM augmented with ELMo may already hit the limit for correcting Polish spelling errors without contextual information. While it improves accuracy in comparison to LSTM initialized withrandom noise, it makes the test cross-entropy slightly worse, which hints at overfitting. The perplexity measures actually increase sharply for more sophisticated architectures. Perplexity should show how little probability is assigned by the model to true answers. We measure it as INLINEFORM0 ", + "where INLINEFORM0 is a sequence of INLINEFORM1 characters, forming the correct version of the word, and INLINEFORM2 is the estimated probability of the INLINEFORM3 th character, given previous predicted characters and the incorrect form. The observed increase of perplexity for increasingly accurate models is most likely due to more refined predicted probability distributions, which go beyond just assigning the bulk of probability to the best answer.", + "Interesting insights can be gained from weights assigned by optimization to layers of ELMo network, which are taken as the word form embedding (Table TABREF5 ). The first layer, and the one that is nearest to input of the network, is given relatively the least importance, while the middle one dominates both others taken together. This suggests that in error correction, at least for Polish, the middle level of morphemes and other characteristic character chunks is more important than phenomena that are low-level or tied to some specific words. This observation should be taken into account in further research on practical solutions for spelling correction." + ], + [ + "Among the methods tested the bidirectional LSTM, especially initialized by ELMo embeddings, offers the best accuracy and raw performance. Adding ELMo to a straightforward PyTorch implementation of LSTM may be easier now than at the time of performing our tests, as since then the authors of ELMoForManyLangs package BIBREF19 improved their programmatic interface. However, if a more interpretable and explainable output is required, some version of vector distance combined with edit distance may be the best direction. It should be noted that this method produces multiple candidate corrections with their similarity scores, as opposed to only one \u201cbest guess\u201c correction that can be obtained from a character-based LSTM. This is important in applications where it is up to humans to the make the final decision, and they are only to be aided by a machine.", + "It is desirable for further reasearch to expand the corpus material into a wider and more representative set of texts. Nevertheless, the solution for any practical case has to be tailored to its characteristic error patterns. Works on language correction for English show that available corpora can be \u201dboosted\u201d BIBREF1 , i.e. expanded by generating new errors consistent with a generative model inferred from the data. This may greatly aid in developing models that are dependent on learning from error corpora.", + "A deliberate omission in this paper are the elements accompanying most real-word error correction solutions. Some fairly obvious approaches to integrating evidence from context include n-grams and Markov chains, although the possibility of using measurements in spaces of semantic vectors was already mentioned in this article. Similarly, non-word errors can be easily detected with comparing tokens against reference vocabulary, but in practice one should have ways of detecting mistakes masquerading as real words and fixing bad segmentation (tokens that are glued together or improperly separated). Testing how performant are various methods for dealing with these problems in Polish language is left for future research." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1116/instruction.md b/qasper-1116/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ea17cc58730436913827dee252d52253fe2da2b9 --- /dev/null +++ b/qasper-1116/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Using Whole Document Context in Neural Machine Translation + +Question: Which datasets were used in the experiment? \ No newline at end of file diff --git a/qasper-1117/instruction.md b/qasper-1117/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..99b8b529c317f5c889c51336d901d7d394c69256 --- /dev/null +++ b/qasper-1117/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Using Whole Document Context in Neural Machine Translation + +Question: What evaluation metrics did they use? \ No newline at end of file diff --git a/qasper-1118/instruction.md b/qasper-1118/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..f1096aa357991038b496579040cba36760eff424 --- /dev/null +++ b/qasper-1118/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Finding Street Gang Members on Twitter + +Question: Do they evaluate only on English datasets? \ No newline at end of file diff --git a/qasper-1120/instruction.md b/qasper-1120/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..7452cc423207041989888dbed412d0cc73b024e0 --- /dev/null +++ b/qasper-1120/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Finding Street Gang Members on Twitter + +Question: What are the differences in the use of YouTube links between gang member and the rest of the Twitter population? \ No newline at end of file diff --git a/qasper-1121/instruction.md b/qasper-1121/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ebe3bb420650a1abe2550768e5964cad17e88444 --- /dev/null +++ b/qasper-1121/instruction.md @@ -0,0 +1,88 @@ +Name of Paper: Finding Street Gang Members on Twitter + +Question: What are the differences in the use of images between gang member and the rest of the Twitter population? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction and Motivation", + "Related Work", + "Discovering Gang Member Profiles", + "Data collection", + "Data analysis", + "Learning algorithms", + "Evaluation", + "Experimental results", + "Evaluation Over Unseen Profiles", + "Conclusion and Future Work", + "Acknowledgement" + ], + "paragraphs": [ + [ + "The crime and violence street gangs introduce into neighborhoods is a growing epidemic in cities around the world. Today, over 1.23 million people in the United States are members of a street gang BIBREF0 , BIBREF1 , which is a coalition of peers, united by mutual interests, with identifiable leadership and internal organization, who act collectively to conduct illegal activity and to control a territory, facility, or enterprise BIBREF2 . They promote criminal activities such as drug trafficking, assault, robbery, and threatening or intimidating a neighborhood BIBREF1 . Moreover, data from the Centers for Disease Control in the United States suggests that the victims of at least 1.3% of all gang-related homicides are merely innocent bystanders who live in gang occupied neighborhoods BIBREF3 .", + "Street gang members have established online presences coinciding with their physical occupation of neighborhoods. The National Gang Threat Assessment Report confirms that at least tens of thousands of gang members are using social networking websites such as Twitter and video sharing websites such as YouTube in their daily life BIBREF0 . They are very active online; the 2007 National Assessment Center's survey of gang members found that 25% of individuals in gangs use the Internet for at least 4 hours a week BIBREF4 . Gang members typically use social networking sites and social media to develop online respect for their street gang BIBREF5 and to post intimidating, threatening images or videos BIBREF6 . This \u201cCyber-\u201d or \u201cInternet banging\u201d BIBREF7 behavior is precipitated by the fact that an increasing number of young members of the society are joining gangs BIBREF8 , and these young members have become enamored with technology and with the notion of sharing information quickly and publicly through social media. Stronger police surveillance in the physical spaces where gangs congregate further encourages gang members to seek out virtual spaces such as social media to express their affiliation, to sell drugs, and to celebrate their illegal activities BIBREF9 .", + "Gang members are able to post publicly on Twitter without fear of consequences because there are few tools law enforcement can use to surveil this medium BIBREF10 . Police departments across the United States instead rely on manual processes to search social media for gang member profiles and to study their posts. For example, the New York City police department employs over 300 detectives to combat teen violence triggered by insults, dares, and threats exchanged on social media, and the Toronto police department teaches officers about the use of social media in investigations BIBREF11 . Officer training is broadly limited to understanding policies on using Twitter in investigations and best practices for data storage BIBREF12 . The safety and security of city neighborhoods can thus be improved if law enforcement were equipped with intelligent tools to study social media for gang activity.", + "The need for better tools for law enforcement cannot be underscored enough. Recent news reports have shown that many incidents involving gangs start on Twitter, escalate over time, and lead to an offline event that could have been prevented by an early warning. For example, the media reported on a possible connection between the death of a teenage rapper from Illinois and the final set of tweets he posted. One of his last tweets linked to a video of him shouting vulgar words at a rival gang member who, in return, replied \u201cI'ma kill you\u201d on social media. In a following tweet, the teenage rapper posted \u201cim on 069\u201d, revealing his location, and was shot dead soon after that post. Subsequent investigation revealed that the rivalry leading to his death began and was carried out entirely on social media. Other reporting has revealed how innocent bystanders have also become targets in online fights, leaving everyone in a neighborhood at risk.", + "This paper investigates whether gang member profiles can be identified automatically on Twitter, which can enable better surveillance of gang members on social media. Classifying Twitter profiles into particular types of users has been done in other contexts BIBREF13 , BIBREF14 , BIBREF15 , but gang member profiles pose unique challenges. For example, many Twitter profile classifiers search for contextual clues in tweets and profile descriptions BIBREF16 , but gang member profiles use a rapidly changing lexicon of keywords and phrases that often have only a local, geographic context. This is illustrated in Figure FIGREF6 , which shows the Twitter profile descriptions of two verified deceased gang members. The profile of @OsoArrogantJoJo provides evidence that he belongs to a rival gang of the Black Disciples by #BDK, a hashtag that is only known to those involved with gang culture in Chicago. @PappyNotPapi's profile mentions #PBG and our investigations revealed that this hashtag is newly founded and stands for the Pooh Bear Gang, a gang that was formerly known as the Insane Cutthroat Gangsters. Given the very local, rapidly changing lexicon of gang members on social media, building a database of keywords, phrases, and other identifiers to find gang members nationally is not feasible. Instead, this study proposes heterogeneous sets of features derived not only from profile and tweet text but also from the emoji usage, profile images, and links to YouTube videos reflecting their music culture. A large set of gang member profiles, obtained through a careful data collection process, is compared against non-gang member profiles to find contrasting features. Experimental results show that using these sets of features, we can build a classifier that has a low false positive rate and a promising INLINEFORM0 -score of 0.7755.", + "This paper is organized as follows. Section SECREF2 discusses the related literature and positions how this work differs from other related works. Section SECREF3 discusses the data collection, manual feature selection and our approach to identify gang member profiles. Section SECREF4 gives a detailed explanation for evaluation of the proposed method and the results in detail. Section SECREF5 concludes the work reported while discussing the future work planned." + ], + [ + "Gang violence is a well studied social science topic dating back to 1927 BIBREF17 . However, the notions of \u201cCyber-\u201d or \u201cInternet banging\u201d, which is defined as \u201cthe phenomenon of gang affiliates using social media sites to trade insults or make violent threats that lead to homicide or victimization\u201d BIBREF7 , was only recently introduced BIBREF18 , BIBREF10 . Patton et al. introduced the concept of \u201cInternet banging\u201d and studied how social media is now being used as a tool for gang self-promotion and as a way for gang members to gain and maintain street credibility BIBREF7 . They also discussed the relationship between gang-related crime and hip-hop culture, giving examples on how hip-hop music shared on social media websites targeted at harassing rival gang members often ended up in real-world collisions among those gangs. Decker et al. and Patton et al. have also reported that street gangs perform Internet banging with social media posts of videos depicting their illegal behaviors, threats to rival gangs, and firearms BIBREF19 , BIBREF20 .", + "The ability to take action on these discoveries is limited by the tools available to discover gang members on social media and to analyze the content they post BIBREF18 . Recent attempts to improve our abilities include a proposed architecture for a surveillance system that can learn the structure, function, and operation of gangs through what they post on social media BIBREF10 . However, the architecture requires a set of gang member profiles for input, thus assuming that they have already been discovered. Patton et al. BIBREF20 devised a method to automatically collect tweets from a group of gang members operating in Detroit, MI. However, their approach required the profile names of the gang members to be known beforehand, and data collection was localized to a single city in the country.", + "This work builds upon existing methods to automatically discover gang member profiles on Twitter. This type of user profile classification problem has been explored in a diverse set of applications such as political affiliation BIBREF13 , ethnicity BIBREF13 , gender BIBREF15 , predicting brand loyalty BIBREF13 , and user occupations BIBREF16 . However, these approaches may utilize an abundance of positive examples in their training data, and only rely on a single feature type (typically, tweet text). Whereas most profile classifiers focus on a single type of feature (e.g. profile text), we consider the use of a variety of feature types, including emoji, YouTube links, and photo features." + ], + [ + "This section discusses the methodology we followed to study and classify the Twitter profiles of gang members automatically. It includes a semi-automatic data collection process to discover a large set of verifiable gang member profiles, an evaluation of the tweets of gang and non-gang member posts to identify promising features, and the deployment of multiple supervised learning algorithms to perform the classification." + ], + [ + "Discovering gang member profiles on Twitter to build training and testing datasets is a challenging task. Past strategies to find these profiles were to search for keywords, phrases, and events that are known to be related to gang activity in a particular city a priori BIBREF10 , BIBREF20 . However, such approaches are unlikely to yield adequate data to train an automatic classifier since gang members from different geographic locations and cultures use local languages, location-specific hashtags, and share information related to activities in a local region BIBREF10 . Such region-specific tweets and profiles may be used to train a classifier to find gang members within a small region but not across the Twitterverse. To overcome these limitations, we adopted a semi-automatic workflow, illustrated in Figure FIGREF7 , to build a dataset of gang member profiles suitable for training a classifier. The steps of the workflow are:", + "1. Seed Term Discovery: Following the success of identifying gang member profiles from Chicago BIBREF10 , we began our data collection with discovering universal terms used by gang members. We first searched for profiles with hashtags for Chicago gangs noted in BIBREF10 , namely #BDK (Black Disciple Killers) and #GDK (Gangster Disciples Killers). Those profiles were analyzed and manually verified as explained in Step 3. Analysis of these profiles identified a small set of hashtags they all use in their profile descriptions. Searching Twitter profiles using those hashtags, we observed that gang members across the U.S. use them, thus we consider those terms to be location neutral. For example, gang members post #FreeDaGuys in their profile to support their fellow members who are in jail, #RIPDaGuys to convey the grieving for fallen gang members, and #FuckDaOpps to show their hatred towards police officers. We used these terms as keywords to discover Twitter profiles irrespective of geographical location. We used the Followerwonk Web service API and Twitter REST API to search Twitter profile descriptions by keywords #FreeDaGuys, #FreeMyNigga, #RIPDaGuys, and #FuckDaOpps. Since there are different informal ways people spell a word in social media, we also considered variations on the spelling of each keyword; for example, for #FreeDaGuys, we searched both #FreeDaGuys, and #FreeTheGuys.", + "2. Gang Affiliated Rappers' Twitter Profile Discovery: Finding profiles by a small set of keywords is unlikely to yield sufficient data. Thus, we sought additional gang member profiles with an observation from Patton et al. BIBREF7 that the influence of hip-hop music and culture on offline gang member activities can also be seen in their social media posts. We thus also consider the influence of hip-hop culture on Twitter by exploring the Twitter network of known gangster rappers who were murdered in 2015 due to gang-related incidents. We searched for these rapper profiles on Twitter and manually checked that the rapper was affiliated to a gang.", + "3. Manual verification of Twitter profiles: We verified each profile discovered manually by examining the profile picture, profile background image, recent tweets, and recent pictures posted by a user. During these checks, we searched for terms, activities, and symbols that we believed could be associated with a gang. For example, profiles whose image or background included guns in a threatening way, stacks of money, showing gang hand signs and gestures, and humans holding or posing with a gun, appeared likely to be from a gang member. Such images were often identified in profiles of users who submitted tweets that contain messages of support or sadness for prisoners or recently fallen gang members, or used a high volume of threatening and intimidating slang language. Only profiles where the images, words, and tweets all suggested gang affiliation were labeled as gang affiliates and added to our dataset. Although this manual verification does have a degree of subjectivity, in practice, the images and words used by gang members on social media are so pronounced that we believe any reasonable analyst would agree that they are gang members. We found that not all the profiles collected belonged to gang members; we observed relatives and followers of gang members posting the same hashtags as in Step 1 to convey similar feelings in their profile descriptions.", + "4. Using Retweets to discover more profiles: From the set of verified profiles, we explored their retweet and follower networks as a way to expand the dataset. We first considered authors of tweets which were retweeted by a gang member in our seed set. In Twitter, \u201cretweeting\u201d is a mechanism by which a user can share someone else's tweet to their follower audience. Assuming that a user only retweets things that they believe or their audience would be interested in, it may be reasonable to assume that gang members would only be interested in sharing what other gang members have to say, and hence, the authors of gang members' retweets could also be gang members.", + "5. Using Followers and Followees to discover more profiles: We analyzed followers and followees of our seed gang member profiles to find more gang member profiles. A Twitter user can follow other Twitter users so that the individual will be subscribed to their tweets as a follower and they will be able to start a private conversation by sending direct messages to the individual. Motivated by the sociological concept of homophily, which claims that individuals have a tendency to associate and bond with similar others, we hypothesized that the followers and followees of Twitter profiles from the seed set may also be gang members. Manual verification of Twitter profiles collected from retweets, followers, and followees of gang members showed that a majority of those profiles are non-gang members who are either family members, hip-hop artists, women or profiles with pornographic content. To ensure that our dataset is not biased towards a specific gang or geographic location, only a limited number of profiles were collected via retweets, followers and followees.", + "Table TABREF8 summarizes the number of profiles manually verified as gang members from Twitter profiles collected in step 1, 2, 4 and 5. Altogether we collected 400 gang member's Twitter profiles. This is a large number compared to previous studies of gang member activities on social media that curated a maximum of 91 profiles BIBREF10 . Moreover, we believe the profiles collected represent a diverse set of gang members that are not biased toward a particular geographic area or lingo as our data collection process used location-independent terms proven to be used by gang members when they express themselves." + ], + [ + "We next explore differences between gang and non-gang member Twitter usage to find promising features for classifying profiles. For this purpose, profiles of non-gang members were collected from the Twitter Streaming API. We collected a random sample of tweets and the profiles of the users who authored the tweets in the random sample. We manually verified that all Twitter profiles collected in this approach belong to non-gang members. The profiles selected were then filtered by location to remove non-U.S. profiles by reverse geo-coding the location stated in their profile description by the Google Maps API. Profiles with location descriptions that were unspecified or did not relate to a location in the U.S. were discarded. We collected 2,000 non-gang member profiles in this manner. In addition, we added 865 manually verified non-gang member profiles collected using the location neutral keywords discussed in Section SECREF3 . Introducing these profiles, which have some characteristics of gang members (such as cursing frequently or cursing at law enforcement) but are not, captures local languages used by family/friends of gang members and ordinary people in a neighborhood where gangs operate.", + "With the Twitter REST API, we collected the maximum number of most recent tweets that can be retrieved (3,200) along with profile descriptions and images (profile and cover photos) of every gang and non-gang member profile. The resulting dataset consists of 400 gang member Twitter profiles and 2,865 non-gang member Twitter profiles. The dataset has a total of 821,412 tweets from gang member profiles and 7,238,758 tweets from non-gang member profiles. Prior to analyzing any text content, we removed all of the seed words used to find gang member profiles, all stop words, and performed stemming across all tweets and profile descriptions.", + "Figure FIGREF14 summarizes the words seen most often in the gang and non-gang members' tweets as clouds. They show a clear difference in language. For example, we note that gang members more frequently use curse words in comparison to ordinary users. Although cursing is frequent in tweets, they represent just 1.15% of all words used BIBREF21 . In contrast, we found 5.72% of all words posted by gang member accounts to be classified as a curse word, which is nearly five times more than the average curse word usage on Twitter. The clouds also reflect the fact that gang members often talk about drugs and money with terms such as smoke, high, hit, and money, while ordinary users hardly speak about finances and drugs. We also noticed that gang members talk about material things with terms such as got, money, make, real, need whereas ordinary users tend to vocalize their feelings with terms such as new, like, love, know, want, look, make, us. These differences make it clear that the individual words used by gang and non-gang members will be relevant features for gang profile classification.", + "On Twitter, a user can give a self-description as a part of the user's profile. A comparison of the top 10 words in gang members' and non-gang members' Twitter profile descriptions is shown in Figure FIGREF21 . The first 10 words are the most frequently used words in non-gang members' profiles and the latter 10 words are the most frequently used words in gang members' profiles. Word comparison shows that gang members prefer to use curse words (nigga, fuck, shit) in their profile descriptions while non-gang members use words related to their feelings or interests (love, life, live, music, book). The terms rip and free which appear in approximately INLINEFORM0 of all gang member Twitter profiles, suggest that gang members use their profile descriptions as a space to grieve for their fallen or incarcerated gang members. The term gang in gang members' profile descriptions suggest that gang members like to self-identify themselves on Twitter. Such lexical features may therefore be of great importance for automatically identifying gang member profiles. We take counts of unigrams from gang and non-gang members' Twitter profile descriptions as classification features.", + "It has been recognized that music is a key cultural component in an urban lifestyle and that gang members often want to emulate the scenarios and activities the music conveys BIBREF7 . Our analysis confirms that the influence of gangster rap is expressed in gang members' Twitter posts. We found that 51.25% of the gang members collected have a tweet that links to a YouTube video. Following these links, a simple keyword search for the terms gangsta and hip-hop in the YouTube video description found that 76.58% of the shared links are related to hip-hop music, gangster rap, and the culture that surrounds this music genre. Moreover, this high proportion is not driven by a small number of profiles that prolifically share YouTube links; eight YouTube links are shared on average by a gang member.", + "Recognizing the frequency with which gang members post YouTube links on gangster rap and hip-hop, we consider the YouTube videos posted in a user's tweets as features for the classifier. In particular, for each YouTube video tweeted, we used the YouTube API to retrieve the video's description and its comments. Further analysis of YouTube data showed a difference between terms in gang members' YouTube data and non-gang members' YouTube data. For example, the top 5 terms (after stemming and stop word removal) used in YouTube videos shared by gang members are shit, like, nigga, fuck, lil while like, love, peopl, song, get are the top 5 terms in non-gang member video data. To represent a user profile based on their music interests, we generated a bag of words from the video descriptions and comments from all shared videos.", + "Motivated by recent work involving the use of emojis by gang members BIBREF22 , we also studied if and how gang and non-gang members use emoji symbols in their tweets. Our analysis found that gang members have a penchant for using just a small set of emoji symbols that convey their anger and violent behavior through their tweets. Figure FIGREF24 illustrates the emoji distribution for the top 20 most frequent emojis used by gang member profiles in our dataset. The fuel pump emoji was the most frequently used emoji by the gang members, which is often used in the context of selling or consuming marijuana. The pistol emoji is the second most frequent in our dataset, which is often used with the guardsman emoji or the police cop emoji in an `emoji chain'. Figure FIGREF28 presents some prototypical `chaining' of emojis used by gang members. The chains may reflect their anger at law enforcement officers, as a cop emoji is often followed up with the emoji of a weapon, bomb, or explosion. We found that 32.25% of gang members in our dataset have chained together the police and the pistol emoji, compared to just 1.14% of non-gang members. Moreover, only 1.71% of non-gang members have used the hundred points emoji and pistol emoji together in tweets while 53% of gang members have used them. A variety of the angry face emoji such as devil face emoji and imp emoji were also common in gang member tweets. The frequency of each emoji symbol used across the set of user's tweets are thus considered as features for our classifier.", + "In our profile verification process, we observed that most gang member profiles portray a context representative of gang culture. Some examples of these profile pictures are shown in Figure FIGREF32 , where the user holds or points weapons, is seen in a group fashion which displays a gangster culture, or is showing off graffiti, hand signs, tattoos and bulk cash. Descriptions of these images may thus empower our classifier. Thus, we translated profile images into features with the Clarifai web service. Clarifai offers a free API to query a deep learning system that tags images with a set of scored keywords that reflect what is seen in the image. We tagged the profile image and cover image for each profile using 20 tags identified by Clarifai. Figure FIGREF36 offers the 20 most often used tags applied to gang and non-gang member profiles. Since we take all the tags returned for an image, we see common words such as people and adult coming up in the top 20 tag set. However, gang member profile images were assigned unique tags such as trigger, bullet, worship while non-gang images were uniquely tagged with beach, seashore, dawn, wildlife, sand, pet. The set of tags returned by Clarifai were thus considered as features for the classifier." + ], + [ + "The unigrams of tweets, profile text, and linked YouTube video descriptions and comments, along with the distribution of emoji symbols and the profile image tags were used to train four different classification models: a Naive Bayes net, a Logistic Regression, a Random Forest, and a Support Vector Machine (SVM). These four models were chosen because they are known to perform well over text features, which is the dominant type of feature considered. The performance of the models are empirically compared to determine the most suitable classification technique for this problem. Data for the models are represented as a vector of term frequencies where the terms were collected from one or more feature sets described above." + ], + [ + "We next evaluate the performance of classifiers that use the above features to discover gang member profiles on Twitter. For this purpose, we use the training set discussed in Section SECREF3 with 400 gang member profiles (the `positive'/`gang' class) and 2,865 non-gang member profiles (the `negative'/`non-gang' class). We trained and evaluated the performance of the classifiers mentioned in Section SECREF31 under a 10-fold cross validation scheme. For each of the four learning algorithms, we consider variations involving only tweet text, emoji, profile, image, or music interest (YouTube comments and video description) features, and a final variant that considers all types of features together. The classifiers that use a single feature type were intended to help us study the quality of their predictive power by itself. When building these single-feature classifiers, we filtered the training dataset based on the availability of the single feature type in the training data. For example, we only used the twitter profiles that had at least a single emoji in their tweets to train classifiers that consider emoji features. We found 3,085 such profiles out of the 3,265 profiles in the training set. When all feature types were considered, we developed two different models:", + "Because a Twitter profile may not have every feature type, Model(1) represents a practical scenario where not every Twitter profile contains every type of feature. In this model, the non-occurrence of a feature is represented by `zeroing out' the feature value during model training. Model(2) represents the ideal scenario where all profiles contain every feature type. For this model, we used 1,358 training instances (42% of all training instances), out of which 172 were gang members (43% of all gang members) and 1,186 were non-gang members (41% of all non-gang members). We used version 0.17.1 of scikit-learn machine learning library to implement the classifiers.", + "For each 10-fold cross validation experiment, we report three evaluation metrics for the `gang' and `non-gang' classes, namely, the Precision = INLINEFORM0 , Recall = INLINEFORM1 , and INLINEFORM2 -score = INLINEFORM3 , where INLINEFORM4 is the number of true positives, INLINEFORM5 is the number of false positives, INLINEFORM6 is the number of true negatives, and INLINEFORM7 is the number of false negatives. We report these metrics for the positive `gang' and negative `non-gang' classes separately because of class imbalance in our dataset." + ], + [ + "Table TABREF37 presents the average precision, recall, and INLINEFORM0 -score over the 10 folds for the single-feature and combined feature classifiers. The table includes, in braces (`{ }'), the number of gang and non-gang profiles that contain a particular feature type, and hence the number of profiles used for the 10-fold cross validation. It is reasonable to expect that any Twitter profile is not that of a gang member, predicting a Twitter user as a non-gang member is much easier than predicting a Twitter user as a gang member. Moreover false positive classifications of the `gang' class may be detrimental to law enforcement investigations, which may go awry as they surveil an innocent person based on the classifier's suggestion. We thus believe that a small false positive rate of the `gang' class to be an especially important evaluation metric. We say that a classifier is `ideal' if it demonstrates high precision, recall, and INLINEFORM1 -score for the `gang' class while performing well on the `non-gang' class as well.", + "The best performing classifier that considers single features is a Random Forest model over tweet features (T), with a reasonable INLINEFORM0 -score of 0.7229 for the `gang' class. It also features the highest INLINEFORM1 -score for the `non-gang' class (0.9671). Its strong performance is intuitive given the striking differences in language as shown in Figure FIGREF14 and discussed in Section UID22 . We also noted that music features offer promising results, with an INLINEFORM2 -score of 0.6505 with a Naive Bayes classifier, as well as emoji features with an INLINEFORM3 -score of 0.6067 also achieved by a Naive Bayes classifier. However, the use of profile data and image tags by themselves yield relatively poor INLINEFORM4 -scores no matter which classifier considered. There may be two reasons for this despite the differences we observed in Section SECREF17 . First, these two feature types did not generate a large number of specific features for learning. For example, descriptions are limited to just 160 characters per profile, leading to a limited number of unigrams (in our dataset, 10 on average) that can be used to train the classifiers. Second, the profile images were tagged by a third party Web service which is not specifically designed to identify gang hand signs, drugs and guns, which are often shared by gang members. This led to a small set of image tags in their profiles that were fairly generic, i.e., the image tags in Figure FIGREF36 such as `people', `man', and `adult'.", + "Combining these diverse sets of features into a single classifier yields even better results. Our results for Model(1) show that the Random Forest achieves the highest INLINEFORM0 -scores for both `gang' (0.7364) and `non-gang' (0.9690) classes and yields the best precision of 0.8792, which corresponds to a low false positive rate when labeling a profile as a gang member. Despite the fact that it has lower positive recall compared to the second best performing classifier (a Random Forest trained over only tweet text features (T)), for this problem setting, we should be willing to increase the chance that a gang member will go unclassified if it means reducing the chance of applying a `gang' label to a non-gang member. When we tested Model(2), a Random Forrest classifier achieved an INLINEFORM1 -score of 0.7755 (improvement of 7.28% with respect to the best performing single feature type classifier (T)) for `gang' class with a precision of 0.8961 (improvement of 6.26% with respect to (T)) and a recall of 0.6994 (improvement of 9.26% with respect to (T)). Model(2) thus outperforms Model(1), and we expect its performance to improve with the availability of more training data with all feature types.", + "px" + ], + [ + "We also tested the trained classifiers using a set of Twitter profiles from a separate data collection process that may emulate the classifier's operation in a real-time setting. For this experiment, we captured real-time tweets from Los Angeles, CA and from ten South Side, Chicago neighborhoods that are known for gang-related activities BIBREF10 using the Twitter streaming API. We consider these areas with known gang presence on social media to ensure that some positive profiles would appear in our test set. We ultimately collected 24,162 Twitter profiles: 15,662 from Los Angeles, and 8,500 from Chicago. We populated data for each profile by using the 3,200 most recent tweets (the maximum that can be collected from Twitter's API) for each profile. Since the 24,162 profiles are far too many to label manually, we qualitatively study those profiles the classifier placed into the `gang' class.", + "We used the training dataset to train our best performing random forest classifier (which use all feature types) and tested it on the test dataset. We then analyzed the Twitter profiles that our classifier labeled as belonging to the `gang' class. Each of those profiles had several features which overlap with gang members such as displaying hand signs and weapons in their profile images or in videos posted by them, gang names or gang-related hashtags in their profile descriptions, frequent use of curse words, and the use of terms such as \u201cmy homie\" to refer to self-identified gang members. Representative tweets extracted from those profiles are depicted in Figure FIGREF41 . The most frequent words found in tweets from those profiles were shit, nigga, got, bitch, go, fuck etc. and their user profiles had terms such as free, artist, shit, fuck, freedagang, and ripthefallen. They had frequently used emojis such as face with tears of joy, hundred points symbol, fire, skull, money bag, and pistol. For some profiles, it was less obvious that the classifier correctly identified a gang member. Such profiles used the same emojis and curse words commonly found in gang members profiles, but their profile picture and tweet content was not indicative of a gang affiliation. In conclusion, we find that in a real-time-like setting, the classifier to be able to extract profiles with features that strongly suggest gang affiliation. Of course, these profiles demand further investigation and extensive evidence from other sources in order to draw a concrete conclusion, especially in the context of a law enforcement investigation. We refrain from reporting any profile names or specific details about the profiles labeled as a `gang' member to comply with the applicable IRB governing this human subject research.", + "px" + ], + [ + "This paper presented an approach to address the problem of automatically identifying gang member profiles on Twitter. Despite the challenges in developing such automated systems, mainly due to difficulties in finding online gang member profiles for developing training datasets, we proposed an approach that uses features extracted from textual descriptions, emojis, images and videos shared on Twitter (textual features extracted from images, and videos). Exploratory analysis of these types of features revealed interesting, and sometimes striking differences in the ways gang and non-gang members use Twitter. Classifiers trained over features that highlight these differences, were evaluated under 10-fold cross validation. Our best classifier achieved a promising INLINEFORM0 -score of 0.7755 over the `gang' profiles when all types of features were considered.", + "Future work will strengthen our training dataset by including more gang member Twitter profiles by searching for more location-independent keywords. We also plan to develop our own image classification system specifically designed to classify images found on gang member profiles. We would also like to experiment with building dictionaries that contain gang names to understand whether \u201chaving a gang name in the profile description\u201d as a feature can improve our results. Finally, we would also like to study how can we further improve our classifier models using word embeddings BIBREF23 and social networks of known gang members.", + "px" + ], + [ + "We are thankful to Uday Kiran Yeda for helping us with data collection. We acknowledge partial support from the National Science Foundation (NSF) award: CNS-1513721: \u201cContext-Aware Harassment Detection on Social Media\u201d, National Institutes of Health (NIH) award: MH105384-01A1: \u201cModeling Social Behavior for Healthcare Utilization in Depression\u201d and Grant No. 2014-PS-PSN-00006 awarded by the Bureau of Justice Assistance. The Bureau of Justice Assistance is a component of the U.S. Department of Justice's Office of Justice Programs, which also includes the Bureau of Justice Statistics, the National Institute of Justice, the Office of Juvenile Justice and Delinquency Prevention, the Office for Victims of Crime, and the SMART Office. Points of view or opinions in this document are those of the authors and do not necessarily represent the official position or policies of the U.S. Department of Justice, NSF or NIH.", + "px" + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1127/instruction.md b/qasper-1127/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..971dc6b00c75694671c360a69a6cf7aade6b6cdc --- /dev/null +++ b/qasper-1127/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A Unified System for Aggression Identification in English Code-Mixed and Uni-Lingual Texts + +Question: How have the differences in communication styles between Twitter and Facebook increase the complexity of the problem? \ No newline at end of file diff --git a/qasper-1128/instruction.md b/qasper-1128/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c8db5594426c23387c38f76f5efcb73559d42265 --- /dev/null +++ b/qasper-1128/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A Unified System for Aggression Identification in English Code-Mixed and Uni-Lingual Texts + +Question: What are the key differences in communication styles between Twitter and Facebook? \ No newline at end of file diff --git a/qasper-1129/instruction.md b/qasper-1129/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..efcc47091f7cbb3ec6cf8832b8fc7ccb79a45cdd --- /dev/null +++ b/qasper-1129/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A Unified System for Aggression Identification in English Code-Mixed and Uni-Lingual Texts + +Question: What data/studies do the authors provide to support the assertion that the majority of aggressive conversations contain code-mixed languages? \ No newline at end of file diff --git a/qasper-1151/instruction.md b/qasper-1151/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..5a9efa311886c3603ce4151e033ddc5d1e3b4cf8 --- /dev/null +++ b/qasper-1151/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Logician: A Unified End-to-End Neural Approach for Open-Domain Information Extraction + +Question: What's the size of the previous largest OpenIE dataset? \ No newline at end of file diff --git a/qasper-1173/instruction.md b/qasper-1173/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..5cfdd875dde2a7c2eb94a0985fe31b35be62c4dc --- /dev/null +++ b/qasper-1173/instruction.md @@ -0,0 +1,145 @@ +Name of Paper: Exploring Hierarchical Interaction Between Review and Summary for Better Sentiment Analysis + +Question: Which review dataset do they use? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Method", + "Method ::: Problem Formulation", + "Method ::: Model Overview", + "Method ::: Summary Encoder", + "Method ::: Hierarchically-Refined Review Encoder", + "Method ::: Hierarchically-Refined Review Encoder ::: Sequence Encoding Layer", + "Method ::: Hierarchically-Refined Review Encoder ::: Attention Inference Layer", + "Method ::: Output Layer", + "Method ::: Training", + "Experiments", + "Experiments ::: Datasets", + "Experiments ::: Experimental Settings", + "Experiments ::: Baselines ::: HSSC @!START@BIBREF6@!END@.", + "Experiments ::: Baselines ::: SAHSSC @!START@BIBREF7@!END@.", + "Experiments ::: Baselines ::: BiLSTM+Pooling.", + "Experiments ::: Baselines ::: BiLSTM+Self-attention @!START@BIBREF13@!END@.", + "Experiments ::: Baselines ::: BiLSTM+Hard Attention", + "Experiments ::: Development Experiments", + "Experiments ::: Development Experiments ::: Self-attention Baseline", + "Experiments ::: Development Experiments ::: Hidden Size", + "Experiments ::: Development Experiments ::: Number of Layers", + "Experiments ::: Results", + "Experiments ::: Results ::: Review Length", + "Experiments ::: Results ::: Case Study", + "Conclusion" + ], + "paragraphs": [ + [ + "Sentiment analysis BIBREF0, BIBREF1 is a fundamental task in natural language processing. In particular, sentiment analysis of user reviews has wide applicationsBIBREF2, BIBREF3, BIBREF4, BIBREF5. In many review websites such as Amazon and IMDb, the user is allowed to give a summary in addition to their review. Summaries usually contain more abstract information about the review. As shown in Figure FIGREF3, two screenshots of reviews were taken from Amazon and IMDb websites, respectively. The user-written summaries of these reviews can be highly indicative of the final polarity. As a result, it is worth considering them together with the review itself for making sentiment classification.", + "To this end, some recent work BIBREF6, BIBREF7 exploits joint modeling. The model structure can be illustrated by Figure FIGREF4. In particular, given a review input, a model is trained to simultaneously predict the sentiment and summary. As a result, both summary information and review information are integrated in the review encoder through back-propagation training. However, one limitation of this method is that it does not explicitly encode a summary during test time.", + "One solution, as shown in Figure FIGREF4, is to train a separate summary generator, which learns to predict a summary given a review. This allows a sentiment classifier to simultaneously encode the review and its summary, before making a prediction using both representations. One further advantage of this model is that it can make use of a user-given summary if it is available with the review, which is the case for the review websites shown in Figure 1. We therefore investigate such a model. One limitation of this method, however, is that it does not capture interaction of review and summary information as thoroughly as the method shown in Figure FIGREF4, since the review and the summary are encoded using two separate encoders.", + "To address this issue, we further investigate a joint encoder for review and summary, which is demonstrated in Figure FIGREF4. The model works by jointly encoding the review and the summary in a multi-layer structure, incrementally updating the representation of the review by consulting the summary representation at each layer. As shown in Figure FIGREF5, our model consists of a summary encoder, a hierarchically-refined review encoder and an output layer. The review encoder is composed of multiple attention layers, each consisting of a sequence encoding layer and an attention inference layer. Summary information is integrated into the representation of the review content at each attention layer, thus, a more abstract review representation is learned in subsequent layers based on a lower-layer representation. This mechanism allows the summary to better guide the representation of the review in a bottom-up manner for improved sentiment classification.", + "We evaluate our proposed model on the SNAP (Stanford Network Analysis Project) Amazon review datasets BIBREF8, which contain not only reviews and ratings, but also golden summaries. In scenarios where there is no user-written summary for a review, we use pointer-generator network BIBREF9 to generate abstractive summaries. Empirical results show that our model significantly outperforms all strong baselines, including joint modeling, separate encoder and joint encoder methods. In addition, our model achieves new state-of-the-art performance, attaining 2.1% (with generated summary) and 4.8% (with golden summary) absolutely improvements compared to the previous best method on SNAP Amazon review benchmark." + ], + [ + "The majority of recent sentiment analysis models are based on either convolutional or recurrent neural networks to encode sequences BIBREF10, BIBREF11.", + "In particular, attention-based models have been widely explored, which assign attention weights to hidden states to generate a representation of the input sequence. A hierarchical model with two levels of attention mechanisms was proposed for document classification BIBREF12. Self-attention mechanism has also been used in sentiment analysis BIBREF13, BIBREF14. However, BIBREF15 empirically showed that self-attention mechanism does not consistently agree with the most salient features, which means that self-attention models may suffer from attending on explicit but irrelevant sentimental words.", + "Rationales were also introduced to sentiment analysis task. BIBREF16 proposed a unsupervised latent model that selects a rationale and then uses the rationale for sentiment analysis. A rationale-augmented CNN model BIBREF17 was proposed, which regards golden rationales as additional input and uses the probability as rationale-level attention weights to generate the final representation for text classification.", + "There has also been work focusing on joint summarization and sentiment classification BIBREF6, BIBREF7, whose general structures are illustrated in Figure FIGREF4. These models can predict sentiment label and summary simultaneously. However, they do not encode summaries explicitly during test time, which makes their performance be limited to some extent." + ], + [ + "In this section, we introduce our proposed model in details. We first give the problem formulation, followed by an overview of the proposed model, and explain each layer of our model in details, before finally giving the loss function and training methods." + ], + [ + "The input to our task is a pair $(X^w, X^s)$, where $X^w = x^w_1, x^w_2, ..., x^w_n$ is a summary and $X^s = x^s_1, x^s_2,...,x^s_m$ is a review, the task is to predict the sentiment label $y \\in [1, 5]$, where 1 denotes the most negative sentiment and 5 denotes the most positive sentiment. $n$ and $m$ denote the size of the review and summary in the number of words, respectively. The training set is $D=\\lbrace (X^w_i, X^s_i, y_i)\\rbrace |_{i=1}^M$ where $M$ is the total number of training examples." + ], + [ + "Figure FIGREF5 gives the architecture of the proposed model, which consists of three modules: a summary encoder, a hierarchically-refined review encoder and an output layer. The summary encoder encodes the summary into a hidden state matrix. The review encoder consists of several layers for representing $\\mathbf {x}^w$, each containing a sequence encoding sublayer and an attention inference sublayer. The sequence encoding sublayer encodes the review text as a word sequence. The attention inference layer acts as a key component, which takes the hidden states from both the original review and the summary as input calculating dot-product attention weights for original review under additional supervision from summary information. Multi-head attention BIBREF18 as well as residual connection are also adopted. The output layer predicts the potential sentiment label according to hidden states from the previous layer." + ], + [ + "Input for the summary encoder is a sequence of summary word representations $\\mathbf {x}^s = \\mathbf {x}^s_1, \\mathbf {x}^s_2, ..., \\mathbf {x}^s_m = \\lbrace emb(x_1^s), ..., emb(x_m^s)\\rbrace $, where $emb$ denotes a word embedding lookup table. Word representations are fed into a standard BiLSTM. We adopt a standard LSTM formulation, where a sequence of hidden states $\\mathbf {h}_t$ are calculated from a sequence of $\\mathbf {x}_t$($t \\in [1,...,m]$).", + "A forward left-to-right LSTM layer and a backward right-to-left LSTM yield a sequence of forward hidden states $\\lbrace {\\stackrel{\\rightarrow }{\\mathbf {h}_1^s}},...,{\\stackrel{\\rightarrow }{\\mathbf {h}_n^s}}\\rbrace $ and a sequence of backward hidden states $\\lbrace {\\stackrel{\\leftarrow }{\\mathbf {h}_1^s}},...,{\\stackrel{\\leftarrow }{\\mathbf {h}_n^s}}\\rbrace $, respectively. The two hidden states are concatenated to form a final representation:", + "We then apply an average-pooling operation over the hidden and take $\\mathbf {h}^s = avg\\_pooling(\\mathbf {h}^s_1, \\mathbf {h}^s_2,...,\\mathbf {h}^s_n)$ as the final representation of summary text." + ], + [ + "The hierarchically-refined review encoder consists of several review encoder layers, each of which is composed of a sequence encoding layer and an attention inference layer." + ], + [ + "Given a review $\\mathbf {x}^w = \\lbrace emb(x_1^w),...,emb(x_n^w)\\rbrace $, another BiLSTM is adopted (the same equation with different parameters compared to the one used in the summary encoder), deriving a sequence of review hidden states $\\mathbf {H}^w=\\lbrace \\mathbf {h}^w_1, \\mathbf {h}^w_2,...,\\mathbf {h}^s_n \\rbrace $." + ], + [ + "In the attention inference layer, we model the dependencies between the original review and the summary with multi-head dot-product attention.Each head produces an attention matrix $\\mathbf {\\alpha } \\in \\mathbb {R}^{d_h \\times 1}$ consisting of a set of similarity scores between the hidden state of each token of the review text and the summary representation. The hidden state outputs are calculated by", + "where $\\mathbf {W}_i^Q \\in \\mathbb {R}^{d_{h} \\times \\frac{d_{h}}{k}}$, $\\mathbf {W}_i^K \\in \\mathbb {R}^{d_{h} \\times \\frac{d_{h}}{k}}$ and $\\mathbf {W}_i^V \\in \\mathbb {R}^{d_{h} \\times \\frac{d_{h}}{k}}$ are model parameters. $Q$, $K$ and $V$ represent Query, Key and Value, respectively. $k$ is the number of parallel heads and $i \\in [1,k]$ indicates which head is being processed.", + "Following BIBREF18, we adopt a residual connection around each attention inference layer, followed by layer normalization BIBREF19 :", + "$\\mathbf {H}$ is then fed to the subsequent sequence encoding layer as input, if any.", + "According to the equations of standard LSTM and Equation DISPLAY_FORM13, tokens of the original review that are the most relevant to the summary are focused on more by consulting summary representation. The hidden states $\\mathbf {H}^{w,s}$ are thus a representation matrix of the review text that encompass key features of summary representation. Multi-head attention mechanism ensures that multi-faced semantic dependency features can be captured during the process, which is beneficial for scenarios where several key points exist in one review.", + "Note also that our design of the review encoding part of the hierarchically-refined attention network is similar to the Transformer architecture in the use of multi-head attention, residual connection and layer normalization BIBREF18. However, our experiments show that bi-directional LSTM works better compared to self-attention network as a basic layer structure. This may result from the fact that Transformer requires a larger amount of training data for the most effectiveness." + ], + [ + "Finally, global average pooling is applied after the previous layer, and then followed by a classifier layer:", + "where $\\hat{y}$ is the predicted sentiment label; $\\mathbf {W}$ and $\\mathbf {b}$ are parameters to be learned." + ], + [ + "Given a dataset $D={\\lbrace (X^w_t,X^s_t,y_t)\\rbrace }|^{|T|}_{t=1}$, our model can be trained by minimizing the cross-entropy loss between", + "where $\\mathbf {p}^{y_t}$ denotes the value of the label in $\\mathbf {p}$ that corresponds to $y_t$." + ], + [ + "We compare our model with several strong baselines and previous state-of-the-art methods, investigating its main effects." + ], + [ + "We empirically compare different methods using Amazon SNAP Review Dataset BIBREF20, which is a part of Stanford Network Analysis Project. The raw dataset consists of around 34 millions Amazon reviews in different domains, such as books, games, sports and movies. Each review mainly contains a product ID, a piece of user information, a plain text review, a review summary and an overall sentiment rating which ranges from 1 to 5. The statistics of our adopted dataset is shown in Table TABREF20. For fair comparison with previous work, we adopt the same partitions used by previous work BIBREF6, BIBREF7, which is, for each domain, the first 1000 samples are taken as the development set, the following 1000 samples as the test set, and the rest as the training set." + ], + [ + "We use GloVe BIBREF22 300-dimensional embeddings as pretrained word vectors. A LSTM hidden size of 256 and four heads for multi-head attention mechanism are adopted. We use Adam BIBREF23 to optimize our model, with an initial learning rate of 0.0003, a decay rate of 0.97, momentum parameters $\\beta _1 = 0.9$, $\\beta _2 = 0.999$, and $\\epsilon = 1 \\times 10^{-8}$. The dropout rate is set depending on the size of each dataset, which is 0.5 for both Toys & Games and Sports & Outdoors and 0.2 for Movies & TV.", + "We conduct experiments with both golden summaries and generated summaries. For generating automatic-decoded summaries, we train a pointer-generator network (PG-Net) with coverage mechanism BIBREF9, which is a specially designed sequence-to-sequence attention-based model that can generate the summary by copying words from the text document or generating words from a fixed vocabulary set at the same time. We generally follow the experimental settings in the original paper except for some minor adjustments specially made for our datasets. Noted that in our work PG-Net can be replaced by any other summarization model." + ], + [ + "This model adopts encoder parameter sharing for jointly sentiment classification and summarization. It predicts the sentiment label using a highway layer, concatenating the hidden state in summary decoder and the original text representation in encoder." + ], + [ + "This work also adopts encoder parameter sharing for jointly sentiment classification and summarization. They use two separate BiLSTMs with self-attention mechanism for generating review and summary representations." + ], + [ + "For this baseline, we use a BiLSTM with hidden sizes of 256 in both directions, and average pooling across all hidden states to form the representation. This method serves as a naive baseline for making use of both review and summary in sentiment classification. It can also be used to compare the effectiveness of the review itself, the summary itself and the combination of both when used as inputs to the problem." + ], + [ + "This baseline uses a BiLSTM with hidden size of 256 in both directions. On the top of BiLSTM, self-attention is used to provide a set of summation weight vectors for the final representation. This method is conceptually simple yet gives the state-of-the-art results for many classification and text matching tasks. Its main difference to our model lies in the fact that attention is performed only in the top hidden layer in this method, yet in every layer in ours." + ], + [ + "To demonstrate the efficiency of our model structure, we also adopt hard attention BIBREF24 for comparison, which is supervised using an extractive summarization objective. In particular, words in the original review that match to the corresponding summary are treated as the summary in their original order. In the case of Figure FIGREF3, the extractive summaries for the review are \u201cJames Cameron's Titanic is easily the most overrated film in history\u201d, which corresponds to the user-written summary \u201cJames Cameron's 1997 Titanic is easily the most overrated film in history!\u201d. The model also calculates another loss between attention weights and extractive summary labels, so that the hard attention weights are trained to strictly follow the extractive summary.", + "For baselines that adopt the separate encoder structure, we generally calculate the representations of review and summary separately with two encoders that hold their own parameters, and then concatenate the two representations alongside the hidden-size dimension. For the joint encoder baselines, we first concatenate the review and summary text, and then encode the concatenated text with one single encoder." + ], + [ + "We use the Toys & Games development set to investigate different key configurations of our model. The results are shown in Table TABREF29." + ], + [ + "We compare different numbers of BiLSTM layers and hidden sizes in BiLSTM self-attention. As can be seen, with more layers a stacked BiLSTM with larger hidden sizes does not give better results compared to a hidden size of 256 either." + ], + [ + "We see an evident improvement of our model when the hidden size increases from 128 to 256. However, the improvement becomes relatively small compared to a large increase in the number of parameters when the hidden size is further increased to 360. Therefore, we adopt 256 as the hidden size in our experiments." + ], + [ + "As Table TABREF29 shows, the accuracy increases when increasing layer numbers from 1 to 2. More layers do not increase the accuracy on development set. We thus set 2 as the number of review encoder layers in the experiments. The best performing model size is comparable to that of the BiLSTM self-attention, demonstrating that the number of parameters is not the key factor to models' performance." + ], + [ + "Table TABREF34 and Table TABREF35 show the final results. Our model outperforms all the baseline models and the top-performing models with both generated summary and golden summary, for all the three datasets. In the scenario where golden summaries are used, BiLSTM+self-attention performs the best among all the baselines, which shows that attention is a useful way to integrate summary and review information. Hard-attention receives more supervision information compared with soft-attention, by supervision signals from extractive summaries. However, it underperforms the soft attention model, which indicates that the most salient words for making sentiment classification may not strictly overlap with extractive summaries. This justifies the importance of user written or automatic-generated summary.", + "A comparison between models that use summary information and those that do not use summary information shows that the review summary is useful for sentiment classification. In addition, the same models work consistently better when the user written gold summary is used compared to a system generated summary, which is intuitively reasonable since the current state-of-the-art abstractive summarization models are far from perfect. Interestingly, as shown in the second section of the table, the gold summary itself does not lead to better sentiment accuracy compared with the review itself, which shows that summaries better serve as auxiliary information sources to review contents.", + "With both gold summaries and automatic-generated summaries, our model gives better results as compared to BiLSTM+self-attention. The latter integrates information from reviews and summaries only in the top representation layer, which is also the standard practice in question answering BIBREF25 and machine translation BIBREF26 models. In contrast, our model integrates summary information into the review representation in each layer, thereby allowing the integrated representation to be hierarchically refined, leading to more abstract hidden states.", + "Finally, the fact that with gold summary, our baseline and final models outperforms the state-of-the-art methods by jointly training shows the importance of making use of user written summaries when they are available. Even with system summary, out models still outperforms HSSC and SAHSSC, showing that our network is more effective than parameter sharing under the same setting without input summaries." + ], + [ + "Figure FIGREF37 consists of line graphs on the accuracy of BiLSTM+self-attention, BiLSTM+pooling and our model against the review length. As the review length increases, the performance of all models decreases. BiLSTM+self-attention does not outperform BiLSTM+pooling on long text. Our method gives better results compared to two baseline models for long reviews, demonstrating that our model is effective for capturing long-term dependency. This is likely because hierarchically-refined attention maintains the most salient information while ignoring the redundant parts of the original review text. Our model can thus be more robust when review has irrelevant sentimental words, which usually exists in larger reviews such as the example in Figure FIGREF3. The hierarchical architecture allows the lower layers to encode local information, while the higher layers can capture long-term dependency and thus better encode global information." + ], + [ + "Our model has a natural advantage of interpretability thanks to the use of attention inference layer. We visualize the hierarchically-refined attention of two samples from the test set of Toys & Games. We also visualize self-attention distribution for fair comparison. To make the visualizations clear and to avoid confusion, we choose to visualize the most salient parts, by rescaling all attention weights into an interval of $[0, 100]$ and adopting 50 as a threshold for attention visualization, showing only attention weights $\\ge 50$.", + "As shown in Figure FIGREF38, the example with generated summary has 5 stars as its golden rating score. The summary text is \u201cfun for the whole new game in all ages ! ! ! fun ! ! !\", which suggests that the game is (1) fun (from word \u201cfun\") and (2) not difficult to learn (from phrase \u201call ages\"). It can be seen that both the self-attention model and the first layer of our model attend to the strongly positive phrase \u201cquite fun\", which is relevant to the word \u201cfun\" in the summary. In comparisons the second layer attends to the phrase \u201cmuch easier\", which is relevant to the phrase \u201cin all ages\" in the summary. This verifies our model's effectiveness of leveraging abstractive summary information.", + "Figure FIGREF38 illustrates a 5-star-rating example with golden summary. The summary text is \u201cFavorite Game to Teach to Newbies\". As shown in the heatmap, self-attention can only attend to some general sentimental words, such as \u201chard\", \u201cfun\", \u201cimmensely\" and \u201cmost\", which deviates from the main idea of the document text. In comparison, the first layer of our model attends to phrases like \u201ceasy to teach\", which is a perfect match of the phrase \u201cteach to newbies\" in the summary. This shows that the shallow sequence inference layer can learn direct similarity matching information under the supervision of summarization. In addition, the second layer of our model attends to phrases including \u201cwould recommend this to anyone\", which links to \u201ceasy to teach\" and \u201cTeach to Newbies\", showing that the deeper sequence inference layer of our model can learn potential connections between the review and the summary." + ], + [ + "We investigated a hierarchically-refined attention network for better sentiment prediction. Our model allows multi-interaction between summary and review representation in a hierarchical manner. Empirical results show that the proposed method outperforms all strong baselines and previous work and achieves new state-of-the-art performance on SNAP Amazon Review dataset." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1174/instruction.md b/qasper-1174/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..56c9e5dd92ead4e3809fb25ee77a2f5fde1f1299 --- /dev/null +++ b/qasper-1174/instruction.md @@ -0,0 +1,148 @@ +Name of Paper: Generaci\'on autom\'atica de frases literarias en espa\~nol + +Question: What evaluation metrics did they look at? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introducci\u00f3n", + "Trabajos previos", + "Corpus utilizados ::: Corpus 5KL", + "Corpus utilizados ::: Corpus 8KF", + "Modelos propuestos", + "Modelos propuestos ::: Modelo generativo estoc\u00e1stico usando cadenas de Markov", + "Modelos propuestos ::: Modelo generativo basado en Texto enlatado", + "Modelos propuestos ::: Modelo 1: Markov y aprendizaje profundo", + "Modelos propuestos ::: Modelo 2: Texto enlatado, aprendizaje profundo y an\u00e1lisis morfosint\u00e1ctico", + "Modelos propuestos ::: Modelo 3: Texto enlatado, aprendizaje profundo e interpretaci\u00f3n geom\u00e9trica", + "Experimentos y resultados", + "Experimentos y resultados ::: Resultados", + "Conclusi\u00f3n y trabajo futuro", + "Agradecimientos" + ], + "paragraphs": [ + [ + "Los investigadores en Procesamiento de Lenguaje Natural (PLN) durante mucho tiempo han utilizado corpus constituidos por documentos enciclop\u00e9dicos (notablemente Wikipedia), period\u00edsticos (peri\u00f3dicos o revistas) o especializados (documentos legales, cient\u00edficos o t\u00e9cnicos) para el desarrollo y pruebas de sus modelos BIBREF0, BIBREF1, BIBREF2.", + "La utilizaci\u00f3n y estudios de corpora literarios sistem\u00e1ticamente han sido dejados a un lado por varias razones. En primer lugar, el nivel de discurso literario es m\u00e1s complejo que los otros g\u00e9neros. En segundo lugar, a menudo, los documentos literarios hacen referencia a mundos o situaciones imaginarias o aleg\u00f3ricas, a diferencia de los otros g\u00e9neros que describen sobre todo situaciones o hechos factuales. Estas y otras caracter\u00edsticas presentes en los textos literarios, vuelven sumamente compleja la tarea de an\u00e1lisis autom\u00e1tico de este tipo de textos. En este trabajo nos proponemos utilizar corpora literarios, a fin de generar realizaciones literarias (frases nuevas) no presentes en dichos corpora.", + "La producci\u00f3n de textos literarios es el resultado de un proceso donde una persona hace uso de aptitudes creativas. Este proceso, denominado \u201cproceso creativo\u201d, ha sido analizado por BIBREF3, quien propone tres tipos b\u00e1sicos de creatividad: la primera, Creatividad Combinatoria (CCO), donde se fusionan elementos conocidos para la generaci\u00f3n de nuevos elementos. La segunda, Creatividad Exploratoria (CE), donde la generaci\u00f3n ocurre a partir de la observaci\u00f3n o exploraci\u00f3n. La tercera, Creatividad Transformacional (CT), donde los elementos generados son producto de alteraciones o experimentaciones aplicadas al dominio de la CE.", + "Sin embargo, cuando se pretende automatizar el proceso creativo, la tarea debe ser adaptada a m\u00e9todos formales que puedan ser realizados en un algoritmo. Este proceso automatizado da lugar a un nuevo concepto denominado Creatividad Computacional (CC), introducido por BIBREF4, quien retoma para ello la CT y la CE propuestas por BIBREF3.", + "La definici\u00f3n de literatura no tiene un consenso universal, y muchas variantes de la definici\u00f3n pueden ser encontradas. En este trabajo optaremos por introducir una definici\u00f3n pragm\u00e1tica de frase literaria, que servir\u00e1 para nuestros modelos y experimentos.", + "Definici\u00f3n. Una frase literaria es una frase que se diferencia de las frases en lengua general, porque contiene elementos (nombres, verbos, adjetivos, adverbios) que son percibidos como elegantes o menos coloquiales que sus equivalentes en lengua general.", + "En particular, proponemos crear artificialmente frases literarias utilizando modelos generativos y aproximaciones sem\u00e1nticas basados en corpus de lengua literaria. La combinaci\u00f3n de esos modelos da lugar a una homosintaxis, es decir, la producci\u00f3n de texto nuevo a partir de formas de discurso de diversos autores. La homosintaxis no tiene el mismo contenido sem\u00e1ntico, ni siquiera las mismas palabras, aunque guarda la misma estructura sint\u00e1ctica.", + "En este trabajo proponemos estudiar el problema de la generaci\u00f3n de texto literario original en forma de frases aisladas, no a nivel de p\u00e1rrafos. La generaci\u00f3n de p\u00e1rrafos puede ser objeto de trabajos futuros. Una evaluaci\u00f3n de la calidad de las frases generadas por nuestro sistema ser\u00e1 presentada.", + "Este art\u00edculo est\u00e1 estructurado como sigue. En la Secci\u00f3n SECREF2 presentamos un estado del arte de la creatividad computacional. En la Secci\u00f3n SECREF3 describimos los corpus utilizados. Nuestros modelos son descritos en la Secci\u00f3n SECREF4. Los resultados y su interpretaci\u00f3n se encuentran en la Secci\u00f3n SECREF5. Finalmente la Secci\u00f3n SECREF6 presenta algunas ideas de trabajos futuros antes de concluir." + ], + [ + "La generaci\u00f3n de texto es una tarea relativamente cl\u00e1sica, que ha sido estudiada en diversos trabajos. Por ejemplo, BIBREF5 presentan un modelo basado en cadenas de Markov para la generaci\u00f3n de texto en idioma polaco. Los autores definen un conjunto de estados actuales y calculan la probabilidad de pasar al estado siguiente. La ecuaci\u00f3n (DISPLAY_FORM1) calcula la probabilidad de pasar al estado $X_{i}$ a partir de $X_{j}$,", + "Para ello, se utiliza una matriz de transici\u00f3n, la cual contiene las probabilidades de transici\u00f3n de un estado actual $X_i$ a los posibles estados futuros $X_{i+1}$. Cada estado puede estar definido por $n$-gramas de letras o de palabras.", + "La tarea inicia en un estado $X_i$ dado por el usuario. Posteriormente, usando la matriz de transici\u00f3n, se calcula la probabilidad de pasar al estado siguiente $X_{i+1}$. En ese momento el estado predicho $X_{i+1}$ se convierte en el estado actual $X_i$, repitiendo este proceso hasta satisfacer una condici\u00f3n. Este m\u00e9todo tiene un buen comportamiento al generar palabras de 4 o 5 letras. En polaco esta longitud corresponde a la longitud media de la mayor parte de las palabras BIBREF6.", + "Tambi\u00e9n hay trabajos que realizan an\u00e1lisis m\u00e1s profundos para generar no solamente palabras, sino p\u00e1rrafos completos. BIBREF7 presentan un algoritmo que genera autom\u00e1ticamente comentarios descriptivos para bloques de c\u00f3digo (m\u00e9todos) en Java. Para ello, se toma el nombre del m\u00e9todo y se usa como la acci\u00f3n o idea central de la descripci\u00f3n a generar. Posteriormente se usan un conjunto de heur\u00edsticas, para seleccionar las l\u00edneas de c\u00f3digo del m\u00e9todo que puedan aportar mayor informaci\u00f3n, y se procesan para generar la descripci\u00f3n. La tarea consiste en construir sintagmas, a partir de la idea central dada por el nombre del m\u00e9todo, y enriquecerlos con la informaci\u00f3n de los elementos extra\u00eddos. Por ejemplo, si hay un m\u00e9todo removeWall(Wall x) y se encuentra la llamada al m\u00e9todo removeWall(oldWall), la descripci\u00f3n generada podr\u00eda ser: \u201cRemove old Wall\u201d. Obteni\u00e9ndose la acci\u00f3n (verbo) y el objeto (sustantivo) directamente del nombre del m\u00e9todo y el adjetivo a partir de la llamada. Estas ideas permiten a los autores la generaci\u00f3n de comentarios extensos sin perder la coherencia y la gramaticalidad.", + "Tambi\u00e9n se encuentran trabajos de generaci\u00f3n textual que se proponen como meta resultados con un valor m\u00e1s art\u00edstico. BIBREF8 presentan un conjunto de algoritmos para la generaci\u00f3n de una gu\u00eda narrativa basada en la idea de Creatividad Exploratoria BIBREF3. El modelo establece i/ un conjunto universal U de conceptos relevantes relacionados a un dominio; ii/ un modelo generador de texto; iii/ un subconjunto de conceptos S que pertenecen al conjunto universal U; y iv/ algoritmos encargados de establecer las relaciones entre U y S para generar nuevos conceptos. Estos nuevos conceptos ser\u00e1n posteriormente comparados con los conceptos ya existentes en U para verificar la coherencia y relaci\u00f3n con la idea principal. Si los resultados son adecuados, estos nuevos conceptos se utilizan para dar continuaci\u00f3n a la narrativa.", + "Son diversos los trabajos que est\u00e1n orientados a la generaci\u00f3n de una narrativa ficticia como cuentos o historias. BIBREF9 proponen un modelo de generaci\u00f3n de texto narrativo a partir del an\u00e1lisis de entidades. Dichas entidades son palabras (verbos, sustantivos o adjetivos) dentro de un texto que ser\u00e1n usados para generar la frase siguiente. El modelo recupera las entidades obtenidas de tres fuentes principales: la frase actual, la frase previa y el documento completo (contexto), y las procesa con una red neuronal para seleccionar las mejores de acuerdo a diversos criterios. A partir de un conjunto de heur\u00edsticas, se analizaron las frases generadas para separar aquellas que expresaran una misma idea (par\u00e1frasis), de aquellas que tuvieran una relaci\u00f3n entre sus entidades pero con ideas diferentes.", + "La generaci\u00f3n de texto literario es un proceso muy diferente a la generaci\u00f3n de texto aleatorio BIBREF10, BIBREF11 y tampoco se limita a una idea o concepto general. El texto literario est\u00e1 destinado a ser un documento elegante y agradable a la lectura, haciendo uso de figuras literarias y un vocabulario distinto al empleado en la lengua general. Esto da a la obra una autenticidad y define el estilo del autor. El texto literario tambi\u00e9n debe diferenciarse de las estructuras r\u00edgidas o estereotipadas de los g\u00e9neros period\u00edstico, enciclop\u00e9dico o cient\u00edfico.", + "BIBREF12 proponen un modelo para la generaci\u00f3n de poemas y se basa en dos premisas b\u00e1sicas: \u00bfqu\u00e9 decir? y \u00bfc\u00f3mo decirlo? La propuesta parte de la selecci\u00f3n de un conjunto de frases tomando como gu\u00eda una lista de palabras dadas por el usuario. Las frases son procesadas por un modelo de red neuronal BIBREF13, para construir combinaciones coherentes y formular un contexto. Este contexto es analizado para identificar sus principales elementos y generar las l\u00edneas del poema, que tambi\u00e9n pasar\u00e1n a formar parte del contexto. El modelo fue evaluado manualmente por 30 expertos en una escala de 1 a 5, analizando legibilidad, coherencia y significatividad en frases de 5 palabras, obteniendo una precisi\u00f3n de 0.75. Sin embargo, la coherencia entre frases result\u00f3 ser muy pobre.", + "BIBREF14, BIBREF15 proponen un modelo de generaci\u00f3n de poemas a base de plantillas. El algoritmo inicia con un conjunto de frases relacionadas a partir de palabras clave. Las palabras clave sirven para generar un contexto. Las frases son procesadas usando el sistema PEN para obtener su informaci\u00f3n gramatical. Esta informaci\u00f3n es empleada para la generaci\u00f3n de nuevas platillas gramaticales y finalmente la construcci\u00f3n de las l\u00edneas del poema, tratando de mantener la coherencia y la gramaticalidad.", + "El modelo sentiGAN BIBREF16 pretende generar texto con un contexto emocional. Se trata de una actualizaci\u00f3n del modelo GAN (Generative Adversarial Net) BIBREF17 que ha producido resultados alentadores en la generaci\u00f3n textual, aunque con ciertos problemas de calidad y coherencia. Se utiliza el an\u00e1lisis sem\u00e1ntico de una entrada proporcionada por el usuario que sirve para la creaci\u00f3n del contexto. La propuesta principal de SentiGAN sugiere establecer un n\u00famero definido de generadores textuales que deber\u00e1n producir texto relacionado a una emoci\u00f3n definida. Los generadores son entrenados bajo dos esquemas: i/ una serie de elementos ling\u00fc\u00edsticos que deben ser evitados para la generaci\u00f3n del texto; y ii/ un conjunto de elementos relacionados con la emoci\u00f3n ligada al generador. A trav\u00e9s de c\u00e1lculos de distancia, heur\u00edsticas y modelos probabil\u00edsticos, el generador crea un texto lo m\u00e1s alejado del primer esquema y lo m\u00e1s cercano al segundo.", + "Tambi\u00e9n existen trabajos con un alcance m\u00e1s corto pero de mayor precisi\u00f3n. BIBREF18 proponen la evaluaci\u00f3n de un conjunto de datos con un modelo basado en redes neuronales para la generaci\u00f3n de subconjuntos de multi-palabras. Este mismo an\u00e1lisis, se considera en BIBREF19, en donde se busca establecer o detectar la relaci\u00f3n hiper\u00f3nimo-hip\u00f3nimo con la ayuda del modelo de Deep Learning Word2vec BIBREF20. La propuesta de BIBREF19 reporta una precisi\u00f3n de 0.70 al ser evaluado sobre un corpus manualmente etiquetado.", + "La literatura es una actividad art\u00edstica que exige capacidades creativas importantes y que ha llamado la atenci\u00f3n de cient\u00edficos desde hace cierto tiempo. BIBREF4 realiza un estado del arte interesante donde menciona algunos trabajos que tuvieron un primer acercamiento a la obra literaria desde una perspectiva superficial. Por ejemplo, el modelo \u201cThrough the park\u201d BIBREF21, es capaz de generar narraciones hist\u00f3ricas empleando la elipsis. Esta t\u00e9cnica es empleada para manipular, entre otras cosas, el ritmo de la narraci\u00f3n. En los trabajos \u201cAbout So Many Things\u201d BIBREF22 y \u201cTaroko Gorge\u201d BIBREF23 se muestran textos generados autom\u00e1ticamente. El primero de ellos genera estrofas de 4 l\u00edneas estrechamente relacionadas entre ellas. Eso se logra a trav\u00e9s de un an\u00e1lisis gramatical que establece conexiones entre entidades de distintas l\u00edneas. El segundo trabajo muestra algunos poemas cortos generados autom\u00e1ticamente con una estructura m\u00e1s compleja que la de las estrofas. El inconveniente de ambos enfoques es el uso de una estructura inflexible, lo que genera textos repetitivos con una gramaticalidad limitada.", + "El proyecto MEXICA modela la generaci\u00f3n colaborativa de narraciones BIBREF4. El prop\u00f3sito es la generaci\u00f3n de narraciones completas utilizando obras de la \u00e9poca Precolombina. MEXICA genera narraciones simulando el proceso creativo de E-R (Engaged y Reflexive) BIBREF24. Este proceso se describe como la acci\u00f3n, donde el autor trae a su mente un conjunto de ideas y contextos y establece una conexi\u00f3n coherente entre estas (E). Posteriormente se reflexiona sobre las conexiones establecidas y se eval\u00faa el resultado final para considerar si este realmente satisface lo esperado (R). El proceso itera hasta que el autor lo considera concluido." + ], + [ + "Este corpus fue constituido con aproximadamente 5 000 documentos (en su mayor parte libros) en espa\u00f1ol. Los documentos originales, en formatos heterog\u00e9neos, fueron procesados para crear un \u00fanico documento codificado en utf8. Las frases fueron segmentadas autom\u00e1ticamente, usando un programa en PERL 5.0 y expresiones regulares, para obtener una frase por l\u00ednea.", + "Las caracter\u00edsticas del corpus 5KL se encuentran en la Tabla TABREF4. Este corpus es empleado para el entrenamiento de los modelos de aprendizaje profundo (Deep Learning, Secci\u00f3n SECREF4).", + "El corpus literario 5KL posee la ventaja de ser muy extenso y adecuado para el aprendizaje autom\u00e1tico. Tiene sin embargo, la desventaja de que no todas las frases son necesariamente \u201cfrases literarias\u201d. Muchas de ellas son frases de lengua general: estas frases a menudo otorgan una fluidez a la lectura y proporcionan los enlaces necesarios a las ideas expresadas en las frases literarias.", + "Otra desventaja de este corpus es el ruido que contiene. El proceso de segmentaci\u00f3n puede producir errores en la detecci\u00f3n de fronteras de frases. Tambi\u00e9n los n\u00fameros de p\u00e1gina, cap\u00edtulos, secciones o \u00edndices producen errores. No se realiz\u00f3 ning\u00fan proceso manual de verificaci\u00f3n, por lo que a veces se introducen informaciones indeseables: copyrights, datos de la edici\u00f3n u otros. Estas son, sin embargo, las condiciones que presenta un corpus literario real." + ], + [ + "Un corpus heterog\u00e9neo de casi 8 000 frases literarias fue constituido manualmente a partir de poemas, discursos, citas, cuentos y otras obras. Se evitaron cuidadosamente las frases de lengua general, y tambi\u00e9n aquellas demasiado cortas ($N \\le 3$ palabras) o demasiado largas ($N \\ge 30$ palabras). El vocabulario empleado es complejo y est\u00e9tico, adem\u00e1s que el uso de ciertas figuras literarias como la rima, la an\u00e1fora, la met\u00e1fora y otras pueden ser observadas en estas frases.", + "Las caracter\u00edsticas del corpus 8KF se muestran en la Tabla TABREF6. Este corpus fue utilizado principalmente en los dos modelos generativos: modelo basado en cadenas de Markov (Secci\u00f3n SECREF13) y modelo basado en la generaci\u00f3n de Texto enlatado (Canned Text, Secci\u00f3n SECREF15)." + ], + [ + "En este trabajo proponemos tres modelos h\u00edbridos (combinaciones de modelos generativos cl\u00e1sicos y aproximaciones sem\u00e1nticas) para la producci\u00f3n de frases literarias. Hemos adaptado dos modelos generativos, usando an\u00e1lisis sint\u00e1ctico superficial (shallow parsing) y un modelo de aprendizaje profundo (Deep Learning) BIBREF25, combinados con tres modelos desarrollados de aproximaci\u00f3n sem\u00e1ntica.", + "En una primera fase, los modelos generativos recuperan la informaci\u00f3n gramatical de cada palabra del corpus 8KF (ver Secci\u00f3n SECREF3), en forma de etiquetas POS (Part of Speech), a trav\u00e9s de un an\u00e1lisis morfosint\u00e1ctico. Utilizamos Freeling BIBREF26 que permite an\u00e1lisis ling\u00fc\u00edsticos en varios idiomas. Por ejemplo, para la palabra \u201cProfesor\u201d Freeling genera la etiqueta POS [NCMS000]. La primera letra indica un sustantivo (Noun), la segunda un sustantivo com\u00fan (Common); la tercera indica el g\u00e9nero masculino (Male) y la cuarta da informaci\u00f3n de n\u00famero (Singular). Los 3 \u00faltimos caracteres dan informaci\u00f3n detallada del campo sem\u00e1ntico, entidades nombradas, etc. En nuestro caso usaremos solamente los 4 primeros niveles de las etiquetas.", + "Con los resultados del an\u00e1lisis morfosint\u00e1ctico, se genera una salida que llamaremos Estructura gramatical vac\u00eda (EGV): compuesta exclusivamente de una secuencia de etiquetas POS; o Estructura gramatical parcialmente vac\u00eda (EGP), compuesta de etiquetas POS y de palabras funcionales (art\u00edculos, pronombres, conjunciones, etc.).", + "En la segunda fase, las etiquetas POS (en la EGV y la EGP) ser\u00e1n reemplazadas por un vocabulario adecuado usando ciertas aproximaciones sem\u00e1nticas.", + "La producci\u00f3n de una frase $f(Q,N)$ es guiada por dos par\u00e1metros: un contexto representado por un t\u00e9rmino $Q$ (o query) y una longitud $3 \\le N \\le 15$, dados por el usuario. Los corpus 5KL y 8KF son utilizados en varias fases de la producci\u00f3n de las frases $f$.", + "El Modelo 1 est\u00e1 compuesto por: i/ un modelo generativo estoc\u00e1stico basado en cadenas de Markov para la selecci\u00f3n de la pr\u00f3xima etiqueta POS usando el algoritmo de Viterbi; y ii/ un modelo de aprendizaje profundo (Word2vec), para recuperar el vocabulario que reemplazar\u00e1 la secuencia de etiquetas POS.", + "El Modelo 2 es una combinaci\u00f3n de: i/ el modelo generativo de Texto enlatado; y ii/ un modelo Word2vec, con un c\u00e1lculo de distancias entre diversos vocabularios que han sido constituidos manualmente.", + "El Modelo 3 utiliza: i/ la generaci\u00f3n de Texto enlatado; y ii/ una interpretaci\u00f3n geom\u00e9trica del aprendizaje profundo. Esta interpretaci\u00f3n est\u00e1 basada en una b\u00fasqueda de informaci\u00f3n iterativa (Information Retrieval, IR), que realiza simult\u00e1neamente un alejamiento de la sem\u00e1ntica original y un acercamiento al query $Q$ del usuario." + ], + [ + "Este modelo generativo, que llamaremos Modelo de Markov, est\u00e1 basado en el algoritmo de Viterbi y las cadenas de Markov BIBREF27, donde se selecciona una etiqueta POS con la m\u00e1xima probabilidad de ocurrencia, para ser agregada al final de la secuencia actual.", + "Utilizamos el corpus de frases literarias 8KF (ver Secci\u00f3n SECREF5), que fue convenientemente filtrado para eliminar tokens indeseables: n\u00fameros, siglas, horas y fechas. El corpus filtrado se analiz\u00f3 usando Freeling, que recibe en entrada una cadena de texto y entrega el texto con una etiqueta POS para cada palabra. El corpus es analizado frase a frase, reemplazando cada palabra por su respectiva etiqueta POS. Al final del an\u00e1lisis, se obtiene un nuevo corpus 8KPOS con $s = 7~679$ secuencias de etiquetas POS, correspondientes al mismo n\u00famero de frases del corpus 8KF. Las secuencias del corpus 8KPOS sirven como conjunto de entrenamiento para el algoritmo de Viterbi, que calcula las probabilidades de transici\u00f3n, que ser\u00e1n usadas para generar cadenas de Markov.", + "Las $s$ estructuras del corpus 8KPOS procesadas con el algoritmo de Viterbi son representadas en una matriz de transici\u00f3n $P_{[s \\times s]}$. $P$ ser\u00e1 utilizada para crear nuevas secuencias de etiquetas POS no existentes en el corpus 8KPOS, simulando un proceso creativo. Nosotros hemos propuesto el algoritmo Creativo-Markov que describe este procedimiento.", + "En este algoritmo, $X_i$ representa el estado de una etapa de la creaci\u00f3n de una frase, en el instante $i$, que corresponde a una secuencia de etiquetas POS. Siguiendo un procedimiento de Markov, en un instante $i$ se selecciona la pr\u00f3xima etiqueta POS$_{i+1}$, con m\u00e1xima probabilidad de ocurrencia, dada la \u00faltima etiqueta POS$_i$ de la secuencia $X_{i}$. La etiqueta POS$_{i+1}$ ser\u00e1 agregada al final de $X_{i}$ para generar el estado $X_{i+1}$. $P(X_{i+1}=Y|X_{i}=Z)$ es la probabilidad de transici\u00f3n de un estado a otro, obtenido con el algoritmo de Viterbi. Se repiten las transiciones, hasta alcanzar una longitud deseada.", + "El resultado es una EGV, donde cada cuadro vac\u00edo representa una etiqueta POS que ser\u00e1 remplazada por una palabra en la etapa final de generaci\u00f3n de la nueva frase. El remplazo se realiza usando un modelo de aprendizaje profundo (Secci\u00f3n SECREF19). La arquitectura general de este modelo se muestra en la Figura FIGREF14." + ], + [ + "El algoritmo creativo-Markov del Modelo de Markov logra reproducir patrones ling\u00fc\u00edsticos (secuencias POS) detectados en el corpus 8KPOS, pero de corta longitud. Cuando se intent\u00f3 extender la longitud de las frases a $N>6$ palabras, no fue posible mantener la coherencia y legibilidad (como se ver\u00e1 en la Secci\u00f3n SECREF19). Decidimos entonces utilizar m\u00e9todos de generaci\u00f3n textual guiados por estructuras morfosint\u00e1cticas fijas: el Texto enlatado. BIBREF28 argumentan que el uso de estas estructuras ahorran tiempo de an\u00e1lisis sint\u00e1ctico y permite concentrarse directamente en el vocabulario.", + "La t\u00e9cnica de Texto enlatado ha sido empleada tambi\u00e9n en varios trabajos, con objetivos espec\u00edficos. BIBREF29, BIBREF30 desarrollaron modelos para la generaci\u00f3n de di\u00e1logos y frases simples. Esta t\u00e9cnica es llamada \u201cGeneraci\u00f3n basada en plantillas\u201d (Template-based Generation) o de manera intuitiva, Texto enlatado.", + "Decidimos emplear texto enlatado para la generaci\u00f3n textual usando un corpus de plantillas (templates), construido a partir del corpus 8KF (Secci\u00f3n SECREF3). Este corpus contiene estructuras gramaticales flexibles que pueden ser manipuladas para crear nuevas frases. Estas plantillas pueden ser seleccionadas aleatoriamente o a trav\u00e9s de heur\u00edsticas, seg\u00fan un objetivo predefinido.", + "Una plantilla es construida a partir de las palabras de una frase $f$, donde se reemplazan \u00fanicamente las palabras llenas de las clases verbo, sustantivo o adjetivo $\\lbrace V, S, A \\rbrace $, por sus respectivas etiquetas POS. Las otras palabras, en particular las palabras funcionales, son conservadas. Esto producir\u00e1 una estructura gramatical parcialmente vac\u00eda, EGP. Posteriormente las etiquetas podr\u00e1n ser reemplazadas por palabras (t\u00e9rminos), relacionadas con el contexto definido por el query $Q$ del usuario.", + "El proceso inicia con la selecci\u00f3n aleatoria de una frase original $f_{o} \\in $ corpus 8KF de longitud $|f_{o}|=N$. $f_{o}$ ser\u00e1 analizada con Freeling para identificar los sintagmas. Los elementos $\\lbrace V, S, A \\rbrace $ de los sintagmas de $f_{o}$ ser\u00e1n reemplazados por sus respectivas etiquetas POS. Estos elementos son los que mayor informaci\u00f3n aportan en cualquier texto, independientemente de su longitud o g\u00e9nero BIBREF31. Nuestra hip\u00f3tesis es que al cambiar solamente estos elementos, simulamos la generaci\u00f3n de frases por homosintaxis: sem\u00e1ntica diferente, misma estructura.", + "La salida de este proceso es una estructura h\u00edbrida parcialmente vac\u00eda (EGP) con palabras funcionales que dan un soporte gramatical y las etiquetas POS. La arquitectura general de este modelo se ilustra en la Figura FIGREF18. Los cuadros llenos representan palabras funcionales y los cuadros vac\u00edos etiquetas POS a ser reemplazadas." + ], + [ + "Los modelos generativos generan estructuras gramaticales vac\u00edas (EGV) o parcialmente vac\u00edas (EGP) que pueden ser manipuladas para generar nuevas frases $f(Q,N)$. La idea es que las frases $f$ sean generadas por homosintaxis. En esta secci\u00f3n, proponemos un modelo de aproximaci\u00f3n sem\u00e1ntica que utiliza el algoritmo Word2vec (basado en aprendizaje profundo), combinado con el modelo generativo de Markov descrito en la Secci\u00f3n SECREF13. El proceso se describe a continuaci\u00f3n.", + "El corpus 5KL es pre-procesado para uniformizar el formato del texto, eliminando caracteres que no son importantes para el an\u00e1lisis sem\u00e1ntico: puntuaci\u00f3n, n\u00fameros, etc. Esta etapa prepara los datos de entrenamiento del algoritmo de aprendizaje profundo que utiliza una representaci\u00f3n vectorial del corpus 5KL. Para el aprendizaje profundo utilizamos la biblioteca Gensim, la versi\u00f3n en Python de Word2vec. Con este algoritmo se obtiene un conjunto de palabras asociadas (embeddings) a un contexto definido por un query $Q$. Word2vec recibe un t\u00e9rmino $Q$ y devuelve un l\u00e9xico $L(Q)=(w_1,w_2,...,w_m)$ que representa un conjunto de $m$ palabras sem\u00e1nticamente pr\u00f3ximas a $Q$. Formalmente, Word2vec: $Q \\rightarrow L(Q)$.", + "El pr\u00f3ximo paso consiste en procesar la EGV producida por Markov. Las etiquetas POS ser\u00e1n identificadas y clasificadas como POS$_{\\Phi }$ funcionales (correspondientes a puntuaci\u00f3n y palabras funcionales) y POS$_\\lambda $ llenas $\\in \\lbrace V, S, A \\rbrace $ (verbos, sustantivos, adjetivos).", + "Las etiquetas POS$_\\Phi $ ser\u00e1n reemplazadas por palabras obtenidas de recursos ling\u00fc\u00edsticos (diccionarios) constru\u00eddos con la ayuda de Freeling. Los diccionarios consisten en entradas de pares: POS$_\\Phi $ y una lista de palabras y signos asociados, formalmente POS$_\\Phi $ $\\rightarrow $ $l$(POS$_\\Phi )=(l_1,l_2,...,l_j)$. Se reemplaza aleatoriamente cada POS$_\\Phi $ por una palabra de $l$ que corresponda a la misma clase gramatical.", + "Las etiquetas POS$_\\lambda $ ser\u00e1n reemplazadas por las palabras producidas por Word2vec $L(Q)$. Si ninguna de las palabras de $L(Q)$ tiene la forma sint\u00e1ctica exigida por POS$_\\lambda $, empleamos la biblioteca PATTERN para realizar conjugaciones o conversiones de g\u00e9nero y/o n\u00famero y reemplazar correctamente POS$_\\lambda $.", + "Si el conjunto de palabras $L(Q)$, no contiene ning\u00fan tipo de palabra llena, que sea adecuada o que pueda manipularse con la biblioteca PATTERN, para reemplazar las etiquetas POS$_\\lambda $, se toma otra palabra, $w_i \\in L(Q)$, lo m\u00e1s cercana a $Q$ (en funci\u00f3n de la distancia producida por Word2vec). Se define un nuevo $Q*=w_i$ que ser\u00e1 utilizado para generar un nuevo conjunto de palabras $L(Q*)$. Este procedimiento se repite hasta que $L(Q*)$ contenga una palabra que pueda reemplazar la POS$_{\\lambda }$ en cuesti\u00f3n. El resultado de este procedimiento es una nueva frase $f$ que no existe en los corpora 5KL y 8KF. La Figura FIGREF23 muestra el proceso descrito." + ], + [ + "En este modelo proponemos una combinaci\u00f3n entre el modelo de Texto enlatado (Secci\u00f3n SECREF15) y un algoritmo de aprendizaje profundo con Word2vec entrenado sobre el corpus 5KL. El objetivo es eliminar las iteraciones del Modelo 1, que son necesarias cuando las etiquetas POS no pueden ser reemplazadas con el l\u00e9xico $L(Q)$.", + "Se efect\u00faa un an\u00e1lisis morfosint\u00e1ctico del corpus 5KL usando Freeling y se usan las etiquetas POS para crear conjuntos de palabras que posean la misma informaci\u00f3n gramatical (etiquetas POS id\u00e9nticas). Una Tabla Asociativa (TA) es generada como resultado de este proceso. La TA consiste en $k$ entradas de pares POS$_k$ y una lista de palabras asociadas. Formalmente POS$_k \\rightarrow V_k =\\lbrace v_{k,1},v_{k,2},...,v_{k,i}\\rbrace $.", + "El Modelo 2 es ejecutado una sola vez para cada etiqueta POS$_k$. La EGP no ser\u00e1 reemplazada completamente: las palabras funcionales y los signos de puntuaci\u00f3n son conservados.", + "Para generar una nueva frase se reemplaza cada etiqueta POS$_k \\in $ EGP, $k=1,2,...$, por una palabra adecuada. Para cada etiqueta POS$_k$, se recupera el l\u00e9xico $V_k$ a partir de TA.", + "El vocabulario es procesado por el algoritmo Word2vec, que calcula el valor de proximidad (distancia) entre cada palabra del vocabulario $v_{k,i}$ y el query $Q$ del usuario, $dist(Q,v_{k,i})$. Despu\u00e9s se ordena el vocabulario $V_k$ en forma descendente seg\u00fan los valores de proximidad $dist(Q,v_{k,i})$ y se escoge aleatoriamente uno de los primeros tres elementos para reemplazar la etiqueta POS$_k$ de la EGP.", + "El resultado es una nueva frase $f_2(Q,N)$ que no existe en los corpora 5KL y 8KF. El proceso se ilustra en la figura FIGREF26." + ], + [ + "El Modelo 3 reutiliza varios de los recursos anteriores: el algoritmo Word2vec, la Tabla Asociativa TA y la estructura gramatical parcialmente vac\u00eda (EGP) obtenida del modelo de Texto enlatado. El modelo utiliza distancias vectoriales para determinar las palabras m\u00e1s adecuadas que sustituir\u00e1n las etiquetas POS de una EGP y as\u00ed generar una nueva frase. Para cada etiqueta POS$_k$, $k=1,2,...$ $\\in $ EGP, que se desea sustituir, usamos el algoritmo descrito a continuaci\u00f3n.", + "Se construye un vector para cada una de las tres palabras siguientes:", + "$o$: es la palabra $k$ de la frase $f_{o}$ (Secci\u00f3n SECREF15), correspondiente a la etiqueta POS$_k$. Esta palabra permite recrear un contexto del cual la nueva frase debe alejarse, evitando producir una par\u00e1frasis.", + "$Q$: palabra que define al query proporcionado por el usuario.", + "$w$: palabra candidata que podr\u00eda reemplazar POS$_k$, $w \\in V_k$. El vocabulario posee un tama\u00f1o $|V_k| = m$ palabras y es recuperado de la TA correspondiente a la POS$_k$.", + "Las 10 palabras $o_i$ m\u00e1s pr\u00f3ximas a $o$, las 10 palabras $Q_i$ m\u00e1s pr\u00f3ximas a $Q$ y las 10 palabras $w_i$ m\u00e1s pr\u00f3ximas a $w$ (en este orden y obtenidas con Word2vec), son concatenadas y representadas en un vector simb\u00f3lico $\\vec{U}$ de 30 dimensiones. El n\u00famero de dimensiones fue fijado a 30 de manera emp\u00edrica, como un compromiso razonable entre diversidad l\u00e9xica y tiempo de procesamiento. El vector $\\vec{U}$ puede ser escrito como:", + "donde cada elemento $u_j, j=1,...,10$, representa una palabra pr\u00f3xima a $o$; $u_j, j=11,...,20$, representa una palabra pr\u00f3xima a $Q$; y $u_j, j=21,...,30$, es una palabra pr\u00f3xima a $w$. $\\vec{U}$ puede ser re-escrito de la siguiente manera (ecuaci\u00f3n DISPLAY_FORM32):", + "$o$, $Q$ y $w$ generan respectivamente tres vectores num\u00e9ricos de 30 dimensiones:", + "donde los valores de $\\vec{X}$ son obtenidos tomando la distancia entre la palabra $o$ y cada palabra $u_j \\in \\vec{U}, j=1,...,30$. La distancia, $x_j=dist(o,u_j)$ es proporcionada por Word2vec y adem\u00e1s $x_j \\in [0,1]$. Evidentemente la palabra $o$ estar\u00e1 m\u00e1s pr\u00f3xima a las 10 primeras palabras $u_j$ que a las restantes.", + "Un proceso similar permite obtener los valores de $\\vec{Q}$ y $\\vec{W}$ a partir de $Q$ y $w$, respectivamente. En estos casos, el $query$ $Q$ estar\u00e1 m\u00e1s pr\u00f3ximo a las palabras $u_j$ en las posiciones $j=11,...,20$ y la palabra candidata $w$ estar\u00e1 m\u00e1s pr\u00f3xima a las palabras $u_j$ en las posiciones $j=21,...30$.", + "Enseguida, se calculan las similitudes coseno entre $\\vec{Q}$ y $\\vec{W}$ (ecuaci\u00f3n DISPLAY_FORM34) y entre $\\vec{X}$ y $\\vec{W}$ (ecuaci\u00f3n DISPLAY_FORM35). Estos valores tambi\u00e9n est\u00e1n normalizados entre [0,1].", + "El proceso se repite para todas las palabras $w$ del l\u00e9xico $V_k$. Esto genera otro conjunto de vectores $\\vec{X}, \\vec{Q}$ y $\\vec{W}$ para los cuales se deber\u00e1n calcular nuevamente las similitudes. Al final se obtienen $m$ valores de similitudes $\\theta _i$ y $\\beta _i$, $ i= 1,..., m$, y se calculan los promedios $\\langle \\theta \\rangle $ y $\\langle \\beta \\rangle $.", + "El cociente normalizado $\\left( \\frac{\\langle \\theta \\rangle }{\\theta _i} \\right)$ indica qu\u00e9 tan grande es la similitud de $\\theta _i$ con respecto al promedio $\\langle \\theta \\rangle $ (interpretaci\u00f3n de tipo maximizaci\u00f3n); es decir, que tan pr\u00f3xima se encuentra la palabra candidata $w$ al query $Q$.", + "El cociente normalizado $\\left( \\frac{\\beta _i}{\\langle \\beta \\rangle } \\right)$ indica qu\u00e9 tan reducida es la similitud de $\\beta _i$ con respecto a $\\langle \\beta \\rangle $ (interpretaci\u00f3n de tipo minimizaci\u00f3n); es decir, qu\u00e9 tan lejos se encuentra la palabra candidata $w$ de la palabra $o$ de $f_{o}$.", + "Estas fracciones se obtienen en cada par $(\\theta _i, \\beta _i)$ y se combinan (minimizaci\u00f3n-maximizaci\u00f3n) para calcular un score $S_i$, seg\u00fan la ecuaci\u00f3n (DISPLAY_FORM36):", + "Mientras m\u00e1s elevado sea el valor $S_i$, mejor obedece a nuestros objetivos: acercarse al $query$ y alejarse de la sem\u00e1ntica original.", + "Finalmente ordenamos en forma decreciente la lista de valores de $S_i$ y se escoge, de manera aleatoria, entre los 3 primeros, la palabra candidata $w$ que reemplazar\u00e1 la etiqueta POS$_k$ en cuesti\u00f3n. El resultado es una nueva frase $f_3(Q,N)$ que no existe en los corpora utilizados para construir el modelo.", + "En la Figura FIGREF37 se muestra una representaci\u00f3n del modelo descrito." + ], + [ + "Dado la especificidad de nuestros experimentos (idioma, corpora disponibles, homosintaxis), no es posible compararse directamente con otros m\u00e9todos.", + "Tampoco consideramos la utilizaci\u00f3n de un baseline de tipo aleatorio, porque los resultados carecer\u00edan de la homosintaxis y ser\u00eda sumamente f\u00e1cil obtener mejores resultados. Dicho lo anterior, el Modelo 1 podr\u00eda ser considerado como nuestro propio baseline." + ], + [ + "A continuaci\u00f3n presentamos un protocolo de evaluaci\u00f3n manual de los resultados obtenidos. El experimento consisti\u00f3 en la generaci\u00f3n de 15 frases por cada uno de los tres modelos propuestos. Para cada modelo, se consideraron tres queries: $Q=$ {AMOR, GUERRA, SOL}, generando 5 frases con cada uno. Las 15 frases fueron mezcladas entre s\u00ed y reagrupadas por queries, antes de presentarlas a los evaluadores.", + "Para la evaluaci\u00f3n, se pidi\u00f3 a 7 personas leer cuidadosamente las 45 frases (15 frases por query). Todos los evaluadores poseen estudios universitarios y son hispanohablantes nativos. Se les pidi\u00f3 anotar en una escala de [0,1,2] (donde 0=mal, 1=aceptable y 2=correcto) los criterios siguientes:", + "Gramaticalidad: ortograf\u00eda, conjugaciones correctas, concordancia en g\u00e9nero y n\u00famero.", + "Coherencia: legibilidad, percepci\u00f3n de una idea general.", + "Contexto: relaci\u00f3n de la frase con respecto al query.", + "Los resultados de la evaluaci\u00f3n se presentan en la Tabla TABREF42, en la forma de promedios normalizados entre [0,1] y de su desviaci\u00f3n est\u00e1ndar $\\sigma $.", + "Las frases generadas por los modelos propuestos presentan caracter\u00edsticas particulares.", + "El Modelo 1 produce generalmente frases con un contexto estrechamente relacionado con el query del usuario, pero a menudo carecen de coherencia y gramaticalidad. Este modelo presenta el valor m\u00e1s alto para el contexto, pero tambi\u00e9n la desviaci\u00f3n est\u00e1ndar m\u00e1s elevada. Se puede inferir que existe cierta discrepancia entre los evaluadores. Los valores altos para el contexto se explican por el grado de libertad de la EGV generada por el modelo de Markov. La EGV permite que todos los elementos de la estructura puedan ser sustituidos por un l\u00e9xico guiado \u00fanicamente por los resultados del algoritmo Word2vec.", + "El Modelo 2 genera frases razonablemente coherentes y gramaticalmente correctas, pero en ocasiones el contexto se encuentra m\u00e1s pr\u00f3ximo a la frase original que al query. Esto puede ser interpretado como una par\u00e1frasis elemental, que no es lo que deseamos.", + "Finalmente, el Modelo 3 genera frases coherentes, gramaticalmente correctas y mejor relacionadas al query que el Modelo 2. Esto se logra siguiendo una intuici\u00f3n opuesta a la par\u00e1frasis: buscamos conservar la estructura sint\u00e1ctica de la frase original, generando una sem\u00e1ntica completamente diferente.", + "Por otro lado, la m\u00ednima dispersi\u00f3n se observa en el Modelo 1, es decir, hay una gran concordancia entre las percepciones de los evaluadores para este criterio." + ], + [ + "En este art\u00edculo hemos presentado tres modelos de producci\u00f3n de frases literarias. La generaci\u00f3n de este g\u00e9nero textual necesita sistemas espec\u00edficos que deben considerar el estilo, la sintaxis y una sem\u00e1ntica que no necesariamente respeta la l\u00f3gica de los documentos de g\u00e9neros factuales, como el period\u00edstico, enciclop\u00e9dico o cient\u00edfico. Los resultados obtenidos son alentadores para el Modelo 3, utilizando Texto enlatado, aprendizaje profundo y una interpretaci\u00f3n del tipo IR. El trabajo a futuro necesita la implementaci\u00f3n de m\u00f3dulos para procesar los $queries$ multi-t\u00e9rmino del usuario. Tambi\u00e9n se tiene contemplada la generaci\u00f3n de frases ret\u00f3ricas utilizando los modelos aqu\u00ed propuestos u otros con un enfoque probabil\u00edstico BIBREF32. Los modelos aqu\u00ed presentados pueden ser enriquecidos a trav\u00e9s de la integraci\u00f3n de otros componentes, como caracter\u00edsticas de una personalidad y/o las emociones BIBREF33, BIBREF34, BIBREF35, BIBREF36. Finalmente, un protocolo de evaluaci\u00f3n semi-autom\u00e1tico (y a gran escala) est\u00e1 igualmente previsto." + ], + [ + "Los autores agradecen a Eric SanJuan respecto a las ideas y el concepto de la homosintaxis." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1180/instruction.md b/qasper-1180/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..33ef4c6b7cf3c8d3ff2b8dca8622181163b83d83 --- /dev/null +++ b/qasper-1180/instruction.md @@ -0,0 +1,60 @@ +Name of Paper: A Latent Morphology Model for Open-Vocabulary Neural Machine Translation + +Question: What are the three languages studied in the paper? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Evaluation ::: Models", + "Evaluation ::: Data and Languages", + "Evaluation ::: Training Settings", + "Evaluation ::: Results ::: The Effect of Morphological Typology", + "Evaluation ::: Results ::: The Effect of Data Size", + "Evaluation ::: Results ::: Predicting Unseen Words", + "Evaluation ::: Results ::: Feature Variations", + "Conclusion", + "Acknowledgments", + "Appendix ::: The Effect of Feature Dimensions" + ], + "paragraphs": [ + [ + "Neural machine translation (NMT) systems are conventionally trained based on the approach of maximizing the log-likelihood on a training corpus in order to learn distributed representations of words according to their sentence context, which is highly demanding in terms of training data as well as the network capacity. Under conditions of lexical sparsity, which may include the cases when the amount of training examples is insufficient to observe words in different context, and particularly in translation of morphologically-rich languages, where the same word can have exponentially many different surface realizations due to syntactic conditions, which are often rarely or ever observed in any set of collected examples, the model may suffer in learning accurate representations of words. The standard approach to overcome this limitation is to replace the word representations in the model with subword units that are shared among words, which are, in principle, more reliable as they are observed more frequently in varying context BIBREF0, BIBREF1. One drawback related to this approach, however, is that the estimation of the subword vocabulary relies on word segmentation methods optimized using corpus-dependent statistics, disregarding any linguistic notion and the translation objective, which may result in morphological errors during splitting, resulting in subword units that are semantically ambiguous as they might be used in far too many lexical contexts BIBREF2. Moreover, the words are generated predicting multiple subword units, which makes generalizing to unseen word forms more difficult, where some of the subword units that could be used to reconstruct a given word may be unlikely in the given context. To alleviate the sub-optimal effects of using explicit segmentation and generalize better to new morphological forms, recent studies explored the idea of extending the same approach to model translation directly at the level of characters BIBREF3, BIBREF4, which, in turn, have demonstrated the requirement of using comparably deeper networks, as the network would then need to learn longer distance grammatical dependencies BIBREF5.", + "In this paper, we explore the benefit of explicitly modeling variations in the surface forms of words using methods from deep latent variable modeling in order to improve the translation accuracy in low-resource and morphologically-rich languages. Latent variable models allow us to inject inductive biases relevant to the task, which, in our case, is word formation, and we believe that follows a certain hierarchical procedure. Our model translates words one character at a time based on word representations learned compositionally from sub-lexical components, which are parameterized by a hierarchical latent variable model mimicking the process of morphological inflection, consisting of a continuous-space dense vector capturing the lexical semantics, and a set of (approximately) discrete features, representing the morphosyntactic role of the word in a given sentence. Each word representation during decoding is reformulated based on the shared latent morphological features, aiding in learning more reliable representations of words under sparse settings by generalizing across their different surface forms. We evaluate our method in translating English into three morphologically-rich languages each with a distinct morphological typology: Arabic, Czech and Turkish, and show that our model is able to obtain better translation accuracy and generalization capacity than conventional approaches to open-vocabulary NMT." + ], + [ + "We evaluate our model by comparing it in machine translation against three baselines which constitute the conventional open-vocabulary NMT methods, including architectures using atomic parameterization either with subword units segmented with BPE BIBREF0 or characters, and the hierarchical parameterization method employed for generating all words in the output. We implement all architectures using Pytorch BIBREF6 within the OpenNMT-py framework BIBREF7." + ], + [ + "In order to evaluate our model we design two sets of experiments. The experiments in \u00a7SECREF8 aim to evaluate different methods under low-resource settings, for languages with different morphological typology. We model the machine translation task from English into three languages with distinct morphological characteristics: Arabic (templatic), Czech (fusional), and Turkish (agglutinative). We use the TED Talks corpora BIBREF8 for training the NMT models for these experiments. In \u00a7SECREF10, we conduct more experiments in Turkish to demonstrate the case of increased data sparsity using multi-domain training corpora, where we extend the training set using corpora from EU Bookshop BIBREF9, Global Voices, Gnome, Tatoeba, Ubuntu BIBREF10, KDE4 BIBREF11, Open Subtitles BIBREF12 and SETIMES BIBREF13. The statistical characteristics of the training sets are given in Tables TABREF16 and TABREF17. We use the official evaluation sets of the IWSLT for validating and testing the accuracy of the models. In order to increase the number of unknown and rare words in the evaluation sets we measure accuracy on large test sets combining evaluation sets from many years (Table TABREF18 presents the evaluation sets used for development and testing). The accuracy of each model output is measured using BLEU BIBREF15 and chrF3 BIBREF16 metrics, whereas the significance of the improvements are computed using bootstrap hypothesis testing BIBREF17." + ], + [ + "All models are implemented using gated recurrent units (GRU) BIBREF18, and have a single-layer bi-RNN encoder. The source sides of the data used for training all NMT models, and the target sides of the data used in training the subword-level NMT models are segmented using BPE with 16,000 merge rules. We implement all decoders using a comparable number of GRU parameters, including 3-layer stacked-GRU subword and character-level decoders, where the attention is computed after the 1st layer BIBREF19 and a 3-layer hierarchical decoder which implements the attention mechanism after the 2nd layer. All models use an embedding dimension and GRU size of 512. The latent morphology model uses the same hierarchical GRU architecture, where the middle layer is augmented using 4 multi-layer perceptrons with 256 hidden units. We use a lemma vector dimension of 150, 10 inflectional features (See \u00a7SECREF21 for experiments conducted to tune the feature dimensions) and set the regularization constant to $\\rho =0.4$. All models are trained using the Adam optimizer BIBREF20 with a batch size of 100, dropout rate of 0.2, learning rate of 0.0004 and learning rate decay of 0.8, applied when the perplexity does not decrease at a given epoch. Translations are generated with beam search with a beam size of 5, where the hierarchical models implement the hierarchical beam search BIBREF21." + ], + [ + "The experiment results given in Table TABREF9 shows the performance of each model in translating English into Arabic, Czech and Turkish. In Turkish, the most sparse target language in our benchmark, using character-based decoding shows to be more advantageous compared to the subword-level and hierarchical models, due to the fact that reduced granularity in the vocabulary units might aid in better predicting words under conditions of high data sparsity. In Arabic, on the other hand, using a hierarchical decoding model shows to be advantageous compared to the character-level decoder, as it might be useful in better learning syntactic dependencies, whereas it also outperforms the subword-level decoder. Using the latent morphology model provides improvements of 0.51 and 0.30 BLEU points in Arabic and Turkish over the best performing baselines, respectively. The fact that our model can efficiently work in both Arabic and Turkish suggests that it can handle the generation of both concatenative and non-concatenative morphological transformations. The results in the English-to-Czech translation direction do not indicate a specific advantage of using either method for generating fusional morphology, where morphemes are already optimized at the surface level, although our model is still able to achieve translation accuracy comparable to the character-level model." + ], + [ + "The experiment conducted in the English-to-Turkish translation direction by increasing the amount of training data with multi-domain corpora demonstrates a more challenging case, where there is a greater possibility of observing rare words, either in the form of morphological inflections due to the complex agglutinative morphology of Turkish, or ambiguous terminology raising from the multi-domain characteristics. In this experiment, the character-level model experiences a drop in performance and its accuracy is much lower than the subword-level one, suggesting that its capacity cannot cope with the increased amount of sparsity. Empirical results suggest that with increased capacity, character-level models carry the potential to reach comparable performance to subword-level models BIBREF4. Our model reaches a much larger improvement of 0.82 BLEU points over the subword-level and 2.54 BLEU points over the character-level decoders, suggesting that it could make use of the increased sparsity in learning more accurate representations." + ], + [ + "In addition to general evaluation using automatic metrics, we perform a more focused analysis to illustrate the performance of different methods in predicting unseen words. We sample the sentences from the development sets which contain out-of-vocabulary words, and compute the average perplexity per character on these sentences using different NMT models, as suggested by BIBREF22. In general, the highest perplexities are obtained using the subword-based model, suggesting that generating unseen words using subword units is indeed increasing the difficulty of prediction, compared to the character-level which obtains the lowest perplexity. This result indicates that increased granularity aids in reducing the uncertainty during prediction. Similar to the results in \u00a7SECREF8, in Czech the values are almost comparable. Due to its stochastic nature, our model yields higher perplexity values compared to the hierarchical model, whereas the values range between subword and character-based models, possibly finding an optimal level of granularity between the two solutions." + ], + [ + "In order to understand whether the latent inflectional features in fact capture information about variations related to morphological transformations, we try generating different surface forms of the same lemma by assigning different values to the inflectional features. We use the latent morphology model based decoder to translate the English word `go', and after sampling the lemma, we fix its value and vary the values of the inflectional features at random positions for generating different outputs. Table TABREF14 presents different sets of feature values and the corresponding outputs generated by the decoder.", + "The model generates different surface forms for different sets of features, confirming that latent variables encode information related to the infinitive form of the verb, as well as its formality conditions, prepositions, person, number and tense. We also observe that many trials based on different feature combinations may result in the same outputs, although some feature values may not be set in a single-word context. Varying the features individually does not necessarily yield distinct changes in the output, suggesting that some features may act jointly in determining the word form." + ], + [ + "In this paper we presented a novel decoding architecture for NMT employing a hierarchical latent variable model to promote sparsity in lexical representations, which demonstrated promising application for morphologically-rich and low-resource languages. Our model generates words one character at a time by composing two latent features representing their lemmas and inflectional features. We evaluate our model against conventional open-vocabulary NMT solutions such as subword and character-level decoding methods in translationg English into three morphologically-rich languages with different morphological typologies under low to mid-resource settings. Our results show that our model can significantly outperform subword-level NMT models, whereas demonstrates better capacity than character-level models in coping with increased amounts of data sparsity. We also conduct ablation studies on the effect of feature variations to the predictions, which prove that despite being completely unsupervised, our model can in fact capture morphosyntactic information and generalize to different surface forms of words." + ], + [ + "This project received funding from the European Union\u2019s Horizon 2020 research and innovation programme under grant agreements 825299 (GoURMET) and 688139 (SUMMA)." + ], + [ + "We investigate the optimal lemma and inflectional feature sizes by measuring the accuracy in English-to-Turkish translation using different feature vector dimensions. The results given in Figure FIGREF22 show that gradually compressing the word representations computed by recurrent hidden states, with an original dimension of 512, from 500 to 100, leads to increased output accuracy, suggesting that encoding more compact representations might provide the model with a better generalization capability. Our results also show that using a feature dimension of 10 is sufficient in reaching the best accuracy." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1187/instruction.md b/qasper-1187/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c8796db927148b62eb5a81b458384f8700961883 --- /dev/null +++ b/qasper-1187/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Quantifying Similarity between Relations with Fact Distribution + +Question: Which competitive relational classification models do they test? \ No newline at end of file diff --git a/qasper-1308/instruction.md b/qasper-1308/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..62460a469d05fb77a6813ad165d7e0f141f95c55 --- /dev/null +++ b/qasper-1308/instruction.md @@ -0,0 +1,78 @@ +Name of Paper: HateMonitors: Language Agnostic Abuse Detection in Social Media + +Question: What are the languages used to test the model? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related works", + "Dataset and Task description", + "Dataset and Task description ::: Datasets", + "Dataset and Task description ::: Tasks", + "System Description", + "System Description ::: Feature Generation ::: Preprocessing:", + "System Description ::: Feature Generation ::: Feature vectors:", + "System Description ::: Our Model", + "Results", + "Discussion", + "Conclusion" + ], + "paragraphs": [ + [ + "In social media, abusive language denotes a text which contains any form of unacceptable language in a post or a comment. Abusive language can be divided into hate speech, offensive language and profanity. Hate speech is a derogatory comment that hurts an entire group in terms of ethnicity, race or gender. Offensive language is similar to derogatory comment, but it is targeted towards an individual. Profanity refers to any use of unacceptable language without a specific target. While profanity is the least threatening, hate speech has the most detrimental effect on the society.", + "Social media moderators are having a hard time in combating the rampant spread of hate speech as it is closely related to the other forms of abusive language. The evolution of new slangs and multilingualism, further adding to the complexity.", + "Recently, there has been a sharp rise in hate speech related incidents in India, the lynchings being the clear indication BIBREF1. Arun et al. BIBREF1 suggests that hate speech in India is very complicated as people are not directly spreading hate but are spreading misinformation against a particular community. Hence, it has become imperative to study hate speech in Indian language.", + "For the first time, a shared task on abusive content detection has been released for Hindi language at HASOC 2019. This will fuel the hate speech and offensive language research for Indian languages. The inclusion of datasets for English and German language will give a performance comparison for detection of abusive content in high and low resource language.", + "In this paper, we focus on the detection of multilingual hate speech detection that are written in Hindi, English, and German and describe our submission (HateMonitors) for HASOC at FIRE 2019 competition. Our system concatenates two types of sentence embeddings to represent each tweet and use machine learning models for classification." + ], + [ + "Analyzing abusive language in social media is a daunting task. Waseem et al. BIBREF2 categorizes abusive language into two sub-classes \u2013 hate speech and offensive language. In their analysis of abusive language, Classifying abusive language into these two subtypes is more challenging due to the correlation between offensive language and hate speech BIBREF3. Nobata et al. BIBREF4 uses predefined language element and embeddings to train a regression model. With the introduction of better classification models BIBREF5, BIBREF6 and newer features BIBREF7, BIBREF3, BIBREF8, the research in hate and offensive speech detection has gained momentum.", + "Silva et al. BIBREF9 performed a large scale study to understand the target of such hate speech on two social media platforms: Twitter and Whisper. These target could be the Refugees and Immigrants BIBREF10, Jews BIBREF11, BIBREF12 and Muslims BIBREF13, BIBREF14. People could become the target of hate speech based on Nationality BIBREF15, sex BIBREF16, BIBREF17, and gender BIBREF18, BIBREF19 as well. Public expressions of hate speech affects the devaluation of minority members BIBREF20, the exclusion of minorities from the society BIBREF21, and tend to diffuse through the network at a faster rate BIBREF22.", + "One of the key issues with the current state of the hate and offensive language research is that the majority of the research is dedicated to the English language on BIBREF23. Few researchers have tried to solve the problem of abusive language in other languages BIBREF10, BIBREF24, but the works are mostly monolingual. Any online social media platform contains people of different ethnicity, which results in the spread of information in multiple languages. Hence, a robust classifier is needed, which can deal with abusive language in the multilingual domain. Several shared tasks like HASOC BIBREF0, HaSpeeDe BIBREF25, GermEval BIBREF26, AMI BIBREF27, HatEval BIBREF28 have focused on detection of abusive text in multiple languages recently." + ], + [ + "The dataset at HASOC 2019 were given in three languages: Hindi, English, and German. Dataset in Hindi and English had three subtasks each, while German had only two subtasks. We participated in all the tasks provided by the organisers and decided to develop a single model that would be language agnostic. We used the same model architecture for all the three languages." + ], + [ + "We present the statistics for HASOC dataset in Table TABREF5. From the table, we can observe that the dataset for the German language is highly unbalanced, English and Hindi are more or less balanced for sub-task A. For sub-task B German dataset is balanced but others are unbalanced. For sub-task C both the datasets are highly unbalanced." + ], + [ + "Sub-task A consists of building a binary classification model which can predict if a given piece of text is hateful and offensive (HOF) or not (NOT). A data point is annotated as HOF if it contains any form of non-acceptable language such as hate speech, aggression, profanity. Each of the three languages had this subtask.", + "Sub-task B consists of building a multi-class classification model which can predict the three different classes in the data points annotated as HOF: Hate speech (HATE), Offensive language (OFFN), and Profane (PRFN). Again all three languages have this sub-task.", + "Sub-task C consists of building a binary classification model which can predict the type of offense: Targeted (TIN) and Untargeted (UNT). Sub-task C was not conducted for the German dataset." + ], + [ + "In this section, we will explain the details about our system, which comprises of two sub-parts- feature generation and model selection. Figure FIGREF15 shows the architecture of our system." + ], + [ + "We preprocess the tweets before performing the feature extraction. The following steps were followed:", + "We remove all the URLs.", + "Convert text to lowercase. This step was not applied to the Hindi language since Devanagari script does not have lowercase and uppercase characters.", + "We did not normalize the mentions in the text as they could potentially reveal important information for the embeddings encoders.", + "Any numerical figure was normalized to a string `number'.", + "We did not remove any punctuation and stop-words since the context of the sentence might get lost in such a process. Since we are using sentence embedding, it is essential to keep the context of the sentence intact." + ], + [ + "The preprocessed posts are then used to generate features for the classifier. For our model, we decided to generate two types of feature vector: BERT Embeddings and LASER Embeddings. For each post, we generate the BERT and LASER Embedding, which are then concatenated and fed as input to the final classifier.", + "Multilingual BERT embeddings: Bidirectional Encoder Representations from Transformers(BERT) BIBREF29 has played a key role in the advancement of natural language processing domain (NLP). BERT is a language model which is trained to predict the masked words in a sentence. To generate the sentence embedding for a post, we take the mean of the last 11 layers (out of 12) to get a sentence vector with length of 768.", + "LASER embeddings: Researchers at Facebook released a language agnostic sentence embeddings representations (LASER) BIBREF30, where the model jointly learns on 93 languages. The model takes the sentence as input and produces a vector representation of length 1024. The model is able to handle code mixing as well BIBREF31.", + "We pass the preprocessed sentences through each of these embedding models and got two separate sentence representation. Further, we concatenate the embeddings into one single feature vector of length 1792, which is then passed to the final classification model." + ], + [ + "The amount of data in each category was insufficient to train a deep learning model. Building such deep models would lead to overfitting. So, we resorted to using simpler models such as SVM and Gradient boosted trees. Gradient boosted trees BIBREF32 are often the choice for systems where features are pre-extracted from the raw data. In the category of gradient boosted trees, Light Gradient Boosting Machine (LGBM) BIBREF33 is considered one of the most efficient in terms of memory footprint. Moreover, it has been part of winning solutions of many competition . Hence, we used LGBM as model for the downstream tasks in this competition." + ], + [ + "The performance of our models across different languages for sub-task A are shown in table TABREF19. Our model got the first position in the German sub-task with a macro F1 score of 0.62. The results of sub-task B and sub-task C is shown in table TABREF20 and TABREF21 respectively." + ], + [ + "In the results of subtask A, models are mainly affected by imbalance of the dataset. The training dataset of Hindi dataset was more balanced than English or German dataset. Hence, the results were around 0.78. As the dataset in German language was highly imbalanced, the results drops to 0.62. In subtask B, the highest F1 score reached was by the profane class for each language in table TABREF20. The model got confused between OFFN, HATE and PRFN labels which suggests that these models are not able to capture the context in the sentence. The subtask C was again a case of imbalanced dataset as targeted(TIN) label gets the highest F1 score in table TABREF21." + ], + [ + "In this shared task, we experimented with zero-shot transfer learning on abusive text detection with pre-trained BERT and LASER sentence embeddings. We use an LGBM model to train the embeddings to perform downstream task. Our model for German language got the first position. The results provided a strong baseline for further research in multilingual hate speech. We have also made the models public for use by other researchers." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1312/instruction.md b/qasper-1312/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..900e2723f72674c9b4eb731cedd06fd0c6d0098f --- /dev/null +++ b/qasper-1312/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: BERT has a Moral Compass: Improvements of ethical and moral values of machines + +Question: What is the Moral Choice Machine? \ No newline at end of file diff --git a/qasper-1313/instruction.md b/qasper-1313/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..9df1048ab0ed82c926474e46791825082094bfe0 --- /dev/null +++ b/qasper-1313/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: BERT has a Moral Compass: Improvements of ethical and moral values of machines + +Question: How is moral bias measured? \ No newline at end of file diff --git a/qasper-1315/instruction.md b/qasper-1315/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4144c7a1ddb41e6f8bbb237663da999950da1209 --- /dev/null +++ b/qasper-1315/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: BERT has a Moral Compass: Improvements of ethical and moral values of machines + +Question: How do the authors define deontological ethical reasoning? \ No newline at end of file diff --git a/qasper-1322/instruction.md b/qasper-1322/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..37601253b711b8eee4f29d6829d59a94890a3ff5 --- /dev/null +++ b/qasper-1322/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Incorporating Priors with Feature Attribution on Text Classification + +Question: Which datasets do they use? \ No newline at end of file diff --git a/qasper-1323/instruction.md b/qasper-1323/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..5d4c69c03a1b1b197109f0c9381fbd0539aeef38 --- /dev/null +++ b/qasper-1323/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A Density Ratio Approach to Language Model Fusion in End-to-End Automatic Speech Recognition + +Question: What metrics are used for evaluation? \ No newline at end of file diff --git a/qasper-1324/instruction.md b/qasper-1324/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0bd9420b3a9057e8166898193bef14ce67746661 --- /dev/null +++ b/qasper-1324/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A Density Ratio Approach to Language Model Fusion in End-to-End Automatic Speech Recognition + +Question: How much training data is used? \ No newline at end of file diff --git a/qasper-1325/instruction.md b/qasper-1325/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ba17d66953af395d17151af4aa2d99644e8ced0b --- /dev/null +++ b/qasper-1325/instruction.md @@ -0,0 +1,136 @@ +Name of Paper: A Density Ratio Approach to Language Model Fusion in End-to-End Automatic Speech Recognition + +Question: How is the training data collected? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "A Brief History of Language Model incorporation in ASR", + "Language Model incorporation into End-to-end ASR, using Bayes' rule ::: A Sequence-level Hybrid Pseudo-Generative Model", + "Language Model incorporation into End-to-end ASR, using Bayes' rule ::: Top-down fundamentals of RNN-T", + "Language Model incorporation into End-to-end ASR, using Bayes' rule ::: Application of Shallow Fusion to RNN-T", + "Language Model incorporation into End-to-end ASR, using Bayes' rule ::: Application of the Density Ratio Method to RNN-T", + "Language Model incorporation into End-to-end ASR, using Bayes' rule ::: Implementation", + "Training, development and evaluation data ::: Training data", + "Training, development and evaluation data ::: Dev and Eval Sets", + "Cross-domain evaluation: YouTube-trained RNN-T @!START@$\\rightarrow $@!END@ Voice Search", + "Cross-domain evaluation: YouTube-trained RNN-T @!START@$\\rightarrow $@!END@ Voice Search ::: RNN-T and RNN-LM model settings", + "Cross-domain evaluation: YouTube-trained RNN-T @!START@$\\rightarrow $@!END@ Voice Search ::: Experiments and results", + "Fine-tuning a YouTube-trained RNN-T using limited Voice Search audio data", + "Discussion", + "Summary", + "Summary ::: Acknowledgments" + ], + "paragraphs": [ + [ + "End-to-end models such as Listen, Attend & Spell (LAS) BIBREF0 or the Recurrent Neural Network Transducer (RNN-T) BIBREF1 are sequence models that directly define $P(W | X)$, the posterior probability of the word or subword sequence $W$ given an audio frame sequence $X$, with no chaining of sub-module probabilities. State-of-the-art, or near state-of-the-art results have been reported for these models on challenging tasks BIBREF2, BIBREF3.", + "End-to-end ASR models in essence do not include independently trained symbols-only or acoustics-only sub-components. As such, they do not provide a clear role for language models $P(W)$ trained only on text/transcript data. There are, however, many situations where we would like to use a separate LM to complement or modify a given ASR system. In particular, no matter how plentiful the paired {audio, transcript} training data, there are typically orders of magnitude more text-only data available. There are also many practical applications of ASR where we wish to adapt the language model, e.g., biasing the recognition grammar towards a list of specific words or phrases for a specific context.", + "The research community has been keenly aware of the importance of this issue, and has responded with a number of approaches, under the rubric of \u201cFusion\u201d. The most popular of these is \u201cShallow Fusion\u201d BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, which is simple log-linear interpolation between the scores from the end-to-end model and the separately-trained LM. More structured approaches, \u201cDeep Fusion\u201d BIBREF9, \u201cCold Fusion\u201d BIBREF10 and \u201cComponent Fusion\u201d BIBREF11 jointly train an end-to-end model with a pre-trained LM, with the goal of learning the optimal combination of the two, aided by gating mechanisms applied to the set of joint scores. These methods have not replaced the simple Shallow Fusion method as the go-to method in most of the ASR community. Part of the appeal of Shallow Fusion is that it does not require model retraining \u2013 it can be applied purely at decoding time. The Density Ratio approach proposed here can be seen as an extension of Shallow Fusion, sharing some of its simplicity and practicality, but offering a theoretical grounding in Bayes' rule.", + "After describing the historical context, theory and practical implementation of the proposed Density Ratio method, this article describes experiments comparing the method to Shallow Fusion in a cross-domain scenario. An RNN-T model was trained on large-scale speech data with semi-supervised transcripts from YouTube videos, and then evaluated on data from a live Voice Search service, using an RNN-LM trained on Voice Search transcripts to try to boost performance. Then, exploring the transition between cross-domain and in-domain, limited amounts of Voice Search speech data were used to fine-tune the YouTube-trained RNN-T model, followed by LM fusion via both the Density Ratio method and Shallow Fusion. The ratio method was found to produce consistent gains over Shallow Fusion in all scenarios examined." + ], + [ + "Generative models and Bayes' rule. The Noisy Channel Model underlying the origins of statistical ASR BIBREF12 used Bayes' rule to combine generative models of both the acoustics $p(X|W)$ and the symbol sequence $P(W)$:", + "for an acoustic feature vector sequence $X = {\\mbox{\\bf x}}_1, ..., {\\mbox{\\bf x}}_T$ and a word or sub-word sequence $W = s_1, ..., s_U$ with possible time alignments $S_W = \\lbrace ..., {\\bf s}, ...\\rbrace $. ASR decoding then uses the posterior probability $P(W|X)$. A prior $p({\\bf s}| W)$ on alignments can be implemented e.g. via a simple 1st-order state transition model. Though lacking in discriminative power, the paradigm provides a clear theoretical framework for decoupling the acoustic model (AM) $p(X|W)$ and LM $P(W)$.", + "Hybrid model for DNNs/LSTMs within original ASR framework. The advent of highly discriminative Deep Neural Networks (DNNs) BIBREF13, BIBREF14, BIBREF15, BIBREF16, BIBREF17 and Long Short Term Memory models (LSTMs) BIBREF18, BIBREF19 posed a challenge to the original Noisy Channel Model, as they produce phoneme- or state- level posteriors $P({\\bf s}(t) | {\\mbox{\\bf x}}_t)$, not acoustic likelihoods $p({\\mbox{\\bf x}}_t | {\\bf s}(t))$. The \u201chybrid\u201d model BIBREF20 proposed the use of scaled likelihoods, i.e. posteriors divided by separately estimated state priors $P(w)$. For bidirectional LSTMs, the scaled-likelihood over a particular alignment ${\\bf s}$ is taken to be", + "using $k(X)$ to represent a $p(X)$-dependent term shared by all hypotheses $W$, that does not affect decoding. This \u201cpseudo-generative\u201d score can then be plugged into the original model of Eq. (DISPLAY_FORM2) and used for ASR decoding with an arbitrary LM $P(W)$. For much of the ASR community, this approach still constitutes the state-of-the-art BIBREF2, BIBREF21, BIBREF22.", + "Shallow Fusion. The most popular approach to LM incorporation for end-to-end ASR is a linear interpolation,", + "with no claim to direct interpretability according to probability theory, and often a reward for sequence length $|W|$, scaled by a factor $\\beta $ BIBREF5, BIBREF7, BIBREF8, BIBREF23." + ], + [ + "The model makes the following assumptions:", + "The source domain $\\psi $ has some true joint distribution $P_{\\psi }(W, X)$ over text and audio;", + "The target domain $\\tau $ has some other true joint distribution $P_{\\tau }(W, X)$;", + "A source domain end-to-end model (e.g. RNN-T) captures $P_{\\psi }(W | X)$ reasonably well;", + "Separately trained LMs (e.g. RNN-LMs) capture $P_{\\psi }(W)$ and $P_{\\tau }(W)$ reasonably well;", + "$p_{\\psi }(X | W)$ is roughly equal to $p_{\\tau }(X | W)$, i.e. the two domains are acoustically consistent; and", + "The target domain posterior, $P_{\\tau }(W | X)$, is unknown.", + "The starting point for the proposed Density Ratio Method is then to express a \u201chybrid\u201d scaled acoustic likelihood for the source domain, in a manner paralleling the original hybrid model BIBREF20:", + "Similarly, for the target domain:", + "Given the stated assumptions, one can then estimate the target domain posterior as:", + "with $k(X) = p_{\\psi }(X) / p_{\\tau }(X)$ shared by all hypotheses $W$, and the ratio $P_{\\tau }(W) / {P_{\\psi }(W)}$ (really a probablity mass ratio) giving the proposed method its name.", + "In essence, this model is just an application of Bayes' rule to end-to-end models and separate LMs. The approach can be viewed as the sequence-level version of the classic hybrid model BIBREF20. Similar use of Bayes' rule to combine ASR scores with RNN-LMs has been described elsewhere, e.g. in work connecting grapheme-level outputs with word-level LMs BIBREF6, BIBREF24, BIBREF25. However, to our knowledge this approach has not been applied to end-to-end models in cross-domain settings, where one wishes to leverage a language model from the target domain. For a perspective on a \u201cpure\u201d (non-hybrid) deep generative approach to ASR, see BIBREF26." + ], + [ + "The RNN Transducer (RNN-T) BIBREF1 defines a sequence-level posterior $P(W|X)$ for a given acoustic feature vector sequence $X = {\\mbox{\\bf x}}_1, ..., {\\mbox{\\bf x}}_T$ and a given word or sub-word sequence $W = s_1, ..., s_U$ in terms of possible alignments $S_W = \\lbrace ..., ({\\bf s}, {\\bf t}), ... \\rbrace $ of $W$ to $X$. The tuple $({\\bf s}, {\\bf t})$ denotes a specific alignment sequence, a symbol sequence and corresponding sequence of time indices, consistent with the sequence $W$ and utterance $X$. The symbols in ${\\bf s}$ are elements of an expanded symbol space that includes optional, repeatable blank symbols used to represent acoustics-only path extensions, where the time index is incremented, but no non-blank symbols are added. Conversely, non-blank symbols are only added to a partial path time-synchronously. (I.e., using $i$ to index elements of ${\\bf s}$ and ${\\bf t}$, $t_{i+1} = t_i + 1$ if $s_{i+1}$ is blank, and $t_{i + 1} = t_i$ if $s_{i+1}$ is non-blank). $P(W|X)$ is defined by summing over alignment posteriors:", + "Finally, $P(s_{i+1} | X, t_i, s_{1:i})$ is defined using an LSTM-based acoustic encoder with input $X$, an LSTM-based label encoder with non-blank inputs $s$, and a feed-forward joint network combining outputs from the two encoders to produce predictions for all symbols $s$, including the blank symbol.", + "The Forward-Backward algorithm can be used to calculate Eq. (DISPLAY_FORM16) efficiently during training, and Viterbi-based beam search (based on the argmax over possible alignments) can be used for decoding when $W$ is unknown BIBREF1, BIBREF27." + ], + [ + "Shallow Fusion (Eq. (DISPLAY_FORM4)) can be implemented in RNN-T for each time-synchronous non-blank symbol path extension. The LM score corresponding to the same symbol extension can be \u201cfused\u201d into the log-domain score used for decoding:", + "This is only done when the hypothesized path extension $s_{i+1}$ is a non-blank symbol; the decoding score for blank symbol path extensions is the unmodified $\\log P(s_{i+1} | X, t_i, s_{1:i})$." + ], + [ + "Eq. (DISPLAY_FORM14) can be implemented via an estimated RNN-T \u201cpseudo-posterior\u201d, when $s_{i+1}$ is a non-blank symbol:", + "This estimate is not normalized over symbol outputs, but it plugs into Eq. () and Eq. (DISPLAY_FORM16) to implement the RNN-T version of Eq. (DISPLAY_FORM14). In practice, scaling factors $\\lambda _\\psi $ and $\\lambda _\\tau $ on the LM scores, and a non-blank reward $\\beta $, are used in the final decoding score:" + ], + [ + "The ratio method is very simple to implement. The procedure is essentially to:", + "Train an end-to-end model such as RNN-T on a given source domain training set $\\psi $ (paired audio/transcript data);", + "Train a neural LM such as RNN-LM on text transcripts from the same training set $\\psi $;", + "Train a second RNN-LM on the target domain $\\tau $;", + "When decoding on the target domain, modify the RNN-T output by the ratio of target/training RNN-LMs, as defined in Eq. (DISPLAY_FORM21), and illustrated in Fig. FIGREF1.", + "The method is purely a decode-time method; no joint training is involved, but it does require tuning of the LM scaling factor(s) (as does Shallow Fusion). A held-out set can be used for that purpose." + ], + [ + "The following data sources were used to train the RNN-T and associated RNN-LMs in this study.", + "Source-domain baseline RNN-T: approximately 120M segmented utterances (190,000 hours of audio) from YouTube videos, with associated transcripts obtained from semi-supervised caption filtering BIBREF28.", + "Source-domain normalizing RNN-LM: transcripts from the same 120M utterance YouTube training set. This corresponds to about 3B tokens of the sub-word units used (see below, Section SECREF30).", + "Target-domain RNN-LM: 21M text-only utterance-level transcripts from anonymized, manually transcribed audio data, representative of data from a Voice Search service. This corresponds to about 275M sub-word tokens.", + "Target-domain RNN-T fine-tuning data: 10K, 100K, 1M and 21M utterance-level {audio, transcript} pairs taken from anonymized, transcribed Voice Search data. These fine-tuning sets roughly correspond to 10 hours, 100 hours, 1000 hours and 21,000 hours of audio, respectively." + ], + [ + "The following data sources were used to choose scaling factors and/or evaluate the final model performance.", + "Source-domain Eval Set (YouTube). The in-domain performance of the YouTube-trained RNN-T baseline was measured on speech data taken from Preferred Channels on YouTube BIBREF29. The test set is taken from 296 videos from 13 categories, with each video averaging 5 minutes in length, corresponding to 25 hours of audio and 250,000 word tokens in total.", + "Target-domain Dev & Eval sets (Voice Search). The Voice Search dev and eval sets each consist of approximately 7,500 anonymized utterances (about 33,000 words and corresponding to about 8 hours of audio), distinct from the fine-tuning data described earlier, but representative of the same Voice Search service." + ], + [ + "The first set of experiments uses an RNN-T model trained on {audio, transcript} pairs taken from segmented YouTube videos, and evaluates the cross-domain generalization of this model to test utterances taken from a Voice Search dataset, with and without fusion to an external LM." + ], + [ + "The overall structure of the models used here is as follows:", + "", + "RNN-T:", + "Acoustic features: 768-dimensional feature vectors obtained from 3 stacked 256-dimensional logmel feature vectors, extracted every 20 msec from 16 kHz waveforms, and sub-sampled with a stride of 3, for an effective final feature vector step size of 60 msec.", + "Acoustic encoder: 6 LSTM layers x (2048 units with 1024-dimensional projection); bidirectional.", + "Label encoder (aka \u201cdecoder\u201d in end-to-end ASR jargon): 1 LSTM layer x (2048 units with 1024-dimensional projection).", + "RNN-T joint network hidden dimension size: 1024.", + "Output classes: 10,000 sub-word \u201cmorph\u201d units BIBREF30 , input via a 512-dimensional embedding.", + "Total number of parameters: approximately 340M", + "RNN-LMs for both source and target domains were set to match the RNN-T decoder structure and size:", + "1 layer x (2048 units with 1024-dimensional projection).", + "Output classes: 10,000 morphs (same as the RNN-T).", + "Total number of parameters: approximately 30M.", + "The RNN-T and the RNN-LMs were independently trained on 128-core tensor processing units (TPUs) using full unrolling and an effective batch size of 4096. All models were trained using the Adam optimization method BIBREF31 for 100K-125K steps, corresponding to about 4 passes over the 120M utterance YouTube training set, and 20 passes over the 21M utterance Voice Search training set. The trained RNN-LM perplexities (shown in Table TABREF28) show the benefit to Voice Search test perplexity of training on Voice Search transcripts." + ], + [ + "In the first set of experiments, the constraint $\\lambda _\\psi = \\lambda _\\tau $ was used to simplify the search for the LM scaling factor in Eq. DISPLAY_FORM21. Fig. FIGREF40 and Fig. FIGREF41 illustrate the different relative sensitivities of WER to the LM scaling factor(s) for Shallow Fusion and the Density Ratio method, as well as the effect of the RNN-T sequence length scaling factor, measured on the dev set.", + "The LM scaling factor affects the relative value of the symbols-only LM score vs. that of the acoustics-aware RNN-T score. This typically alters the balance of insertion vs. deletion errors. In turn, this effect can be offset (or amplified) by the sequence length scaling factor $\\beta $ in Eq. (DISPLAY_FORM4), in the case of RNN-T, implemented as a non-blank symbol emission reward. (The blank symbol only consumes acoustic frames, not LM symbols BIBREF1). Given that both factors have related effects on overall WER, the LM scaling factor(s) and the sequence length scaling factor need to be tuned jointly.", + "Fig. FIGREF40 and Fig. FIGREF41 illustrate the different relative sensitivities of WER to these factors for Shallow Fusion and the Density Ratio method, measured on the dev set.", + "In the second set of experiments, $\\beta $ was fixed at -0.1, but the constraint $\\lambda _\\psi = \\lambda _\\tau $ was lifted, and a range of combinations was evaluated on the dev set. The results are shown in Fig. FIGREF43. The shading in Figs. FIGREF40, FIGREF41 and FIGREF43 uses the same midpoint value of 15.0 to highlight the results.", + "The best combinations of scaling factors from the dev set evaluations (see Fig. FIGREF40, Fig. FIGREF41 and Fig. FIGREF43) were used to generate the final eval set results, WERs and associated deletion, insertion and substitution rates, shown in Table TABREF44. These results are summarized in Table TABREF45, this time showing the exact values of LM scaling factor(s) used." + ], + [ + "The experiments in Section SECREF5 showed that an LM trained on text from the target Voice Search domain can boost the cross-domain performance of an RNN-T. The next experiments examined fine-tuning the original YouTube-trained RNN-T on varied, limited amounts of Voice Search {audio, transcript} data. After fine-tuning, LM fusion was applied, again comparing Shallow Fusion and the Density Ratio method.", + "Fine-tuning simply uses the YouTube-trained RNN-T model to warm-start training on the limited Voice Search {audio, transcript} data. This is an effective way of leveraging the limited Voice Search audio data: within a few thousand steps, the fine-tuned model reaches a decent level of performance on the fine-tuning task \u2013 though beyond that, it over-trains. A held-out set can be used to gauge over-training and stop training for varying amounts of fine-tuning data.", + "The experiments here fine-tuned the YouTube-trained RNN-T baseline using 10 hours, 100 hours and 1000 hours of Voice Search data, as described in Section SECREF27. (The source domain RNN-LM was not fine-tuned). For each fine-tuned model, Shallow Fusion and the Density Ratio method were used to evaluate incorporation of the Voice Search RNN-LM, described in Section SECREF5, trained on text transcripts from the much larger set of 21M Voice Search utterances. As in Section SECREF5, the dev set was used to tune the LM scaling factor(s) and the sequence length scaling factor $\\beta $. To ease parameter tuning, the constraint $\\lambda _\\psi = \\lambda _\\tau $ was used for the Density Ratio method. The best combinations of scaling factors from the dev set were then used to generate the final eval results, which are shown in Table TABREF45" + ], + [ + "The experiments described here examined the generalization of a YouTube-trained end-to-end RNN-T model to Voice Search speech data, using varying quantities (from zero to 100%) of Voice Search audio data, and 100% of the available Voice Search text data. The results show that in spite of the vast range of acoustic and linguistic patterns covered by the YouTube-trained model, it is still possible to improve performance on Voice Search utterances significantly via Voice Search specific fine-tuning and LM fusion. In particular, LM fusion significantly boosts performance when only a limited quantity of Voice Search fine-tuning data is used.", + "The Density Ratio method consistently outperformed Shallow Fusion for the cross-domain scenarios examined, with and without fine-tuning to audio data from the target domain. Furthermore, the gains in WER over the baseline are significantly larger for the Density Ratio method than for Shallow Fusion, with up to 28% relative reduction in WER (17.5% $\\rightarrow $ 12.5%) compared to up to 17% relative reduction (17.5% $\\rightarrow $ 14.5%) for Shallow Fusion, in the no fine-tuning scenario.", + "Notably, the \u201csweet spot\u201d of effective combinations of LM scaling factor and sequence length scaling factor is significantly larger for the Density Ratio method than for Shallow Fusion (see Fig. FIGREF40 and Fig. FIGREF41). Compared to Shallow Fusion, larger absolute values of the scaling factor can be used.", + "A full sweep of the LM scaling factors ($\\lambda _\\psi $ and $\\lambda _\\tau $) can improve over the constrained setting $\\lambda _\\psi = \\lambda _\\tau $, though not by much. Fig. FIGREF43 shows that the optimal setting of the two factors follows a roughly linear pattern along an off-diagonal band.", + "Fine-tuning using transcribed Voice Search audio data leads to a large boost in performance over the YouTube-trained baseline. Nonetheless, both fusion methods give gains on top of fine-tuning, especially for the limited quantities of fine-tuning data. With 10 hours of fine-tuning, the Density Ratio method gives a 20% relative gain in WER, compared to 12% relative for Shallow Fusion. For 1000 hours of fine-tuning data, the Density Ratio method gives a 10.5% relative gave over the fine-tuned baseline, compared to 7% relative for Shallow Fusion. Even for 21,000 hours of fine-tuning data, i.e. the entire Voice Search training set, the Density Ratio method gives an added boost, from 7.8% to 7.4% WER, a 5% relative improvement.", + "A clear weakness of the proposed method is the apparent need for scaling factors on the LM outputs. In addition to the assumptions made (outlined in Section SECREF5), it is possible that this is due to the implicit LM in the RNN-T being more limited than the RNN-LMs used." + ], + [ + "This article proposed and evaluated experimentally an alternative to Shallow Fusion for incorporation of an external LM into an end-to-end RNN-T model applied to a target domain different from the source domain it was trained on. The Density Ratio method is simple conceptually, easy to implement, and grounded in Bayes' rule, extending the classic hybrid ASR model to end-to-end models. In contrast, the most commonly reported approach to LM incorporation, Shallow Fusion, has no clear interpretation from probability theory. Evaluated on a YouTube $\\rightarrow $ Voice Search cross-domain scenario, the method was found to be effective, with up to 28% relative gains in word error over the non-fused baseline, and consistently outperforming Shallow Fusion by a significant margin. The method continues to produce gains when fine-tuning to paired target domain data, though the gains diminish as more fine-tuning data is used. Evaluation using a variety of cross-domain evaluation scenarios is needed to establish the general effectiveness of the method." + ], + [ + "The authors thank Matt Shannon and Khe Chai Sim for valuable feedback regarding this work." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1330/instruction.md b/qasper-1330/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4801b25490fb47b4380a40e89e900e441fb31505 --- /dev/null +++ b/qasper-1330/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Attention Is (not) All You Need for Commonsense Reasoning + +Question: Which datasets do they evaluate on? \ No newline at end of file diff --git a/qasper-1337/instruction.md b/qasper-1337/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..319eb3d589c011b6780cff867aa98308d2dcbd60 --- /dev/null +++ b/qasper-1337/instruction.md @@ -0,0 +1,116 @@ +Name of Paper: Seshat: A tool for managing and verifying annotation campaigns of audio data + +Question: Can it be used for any language? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Overview of Seshat", + "Development ::: Engineering choices", + "Development ::: Engineering choices ::: Back-end Choices", + "Development ::: Engineering choices ::: Front-end Choices", + "Development ::: UX/UI Choices", + "Using Seshat ::: Installation and Setup", + "Using Seshat ::: Launching and monitoring an annotation campaign", + "Using Seshat ::: Scripting API", + "Using Seshat ::: Annotation Parser Customisation", + "Using Seshat ::: Inter-rater agreement: the @!START@$\\gamma $@!END@ measure", + "Use cases", + "Use cases ::: Clinical interviews", + "Use cases ::: In-the-wild child-centered recordings", + "Conclusion and Future work", + "Acknowledgements" + ], + "paragraphs": [ + [ + "Large corpora of speech, obtained in the laboratory and in naturalistic conditions, become easier to collect. This new trend broadens the scope of scientific questions on speech and language that can be answered. However, this poses an important challenge for the construction of reliable and usable annotations. Managing annotators and ensuring the quality of their annotations are highly demanding tasks for research endeavours and industrial projects BIBREF0. When organised manually, the manager of annotation campaigns usually faces three major problems: the mishandling of files (e.g., character-encoding problems, incorrect naming of files), the non-conformity of the annotations BIBREF1, and the inconsistency of the annotations BIBREF2.", + "In this paper, we introduce Seshat, a system for the automated management of annotation campaigns for audio/speech data which addresses these challenges. It is built on two components that communicate via a Restful API: a back-end (server) written in Flask and a front-end (client) in Angular Typescript. Seshat is easy to install for non-developers and easy to use for researchers and annotators while having some extension capabilities for developers.", + "In Section SECREF2, we describe the related work on annotations tools, which do not provide solutions to all the aforementioned challenges during corpus creation. In Section SECREF3, we make an overview of the different functionalities of the software. Then, we explain, in Section SECREF4, the architecture of the software, and also the several UX/UI design and engineering choices that have been made to facilitate the usage of the platform. We describe how to use of Seshat in Section SECREF5 and Section SECREF6 presents two specific use-cases. Finally, we conclude and describe future plans for Seshat in Section SECREF7." + ], + [ + "Self-hosted annotation systems. There are many standalone solutions for the transcription of speech data that are already used by researchers: Transcriber BIBREF3, Wavesurfer BIBREF4, Praat BIBREF5, ELAN BIBREF6, XTrans BIBREF7. These systems allow the playback of sound data and the construction of different layers of annotations with various specifications, with some advanced capabilities (such as annotations with hierarchical or no relationship between layers, number of audio channels, video support). Yet, these solutions lack a management system: each researcher must track the files assigned to annotators and build a pipeline to parse (and eventually check) the output annotation files. Moreover, checking can only be done once the annotations have been submitted to the researchers. This task becomes quickly untraceable as the number of files and annotators grow. In addition, most of these transcription systems do not provide a way to evaluate consistency (intra- and inter-annotator agreement) that would be appropriate for speech data BIBREF8.", + "Web-based annotations systems. There are several web-based annotation systems for the annotation of audio data. Among them we find light-weight systems, like the VIA software BIBREF9 or Praat on the web BIBREF10 that allow to build simple layers of annotations. However, they do not provide a proper management system for a pool of annotators nor do they integrate annotation checking.", + "On the other side of the spectrum, there are more sophisticated systems with various capabilities. Camomille BIBREF11 and the EMU-SDMS system (that can also be used offline) BIBREF12 allow to work with speech data and to distribute the tasks to several annotators. But these systems require expertise in web hosting and technologies to deploy and modify them.", + "Finally, WebAnno BIBREF13 and GATE Teamware BIBREF14 are the tools that most closely match our main contributions regarding quality control (conformity and consistency checking), annotators' management and flexibility. WebAnno includes consistency checking with the integration of different metrics BIBREF15. However, these tools have only been built for text data. The format and all the custom layers have been designed for Natural Language Processing tasks. Porting WebAnno to support speech data seemed a major engineering challenge. That is why it appeared necessary to develop a new and user-friendly tool addressed to the speech community." + ], + [ + "Seshat is a user-friendly web-based interface whose objective is to smoothly manage large campaigns of audio data annotation, see Figure FIGREF8. Below, we describe the several terms used in Seshat's workflow:", + "[font=, leftmargin=1cm, style=nextline]", + "A set of audio/speech files that a Campaign Manager wants to annotate. It is indicated either by a folder containing sound files, or by a CSV summarizing a set of files. We support the same formats as Praat so far: WAV, Flac and MP3.", + "An object that enables the Campaign Manager to assign Annotation Tasks to the Annotators. It references a Corpus, and allows the Manager to track the annotation's tasks progress and completion in real time. At its creation, a Textgrid Checking Scheme can also be defined for that campaign.", + "It is contained in an Annotation Campaign, it references an audio file from the campaign's designated Audio Corpus, and assigned to Annotators. It can either be a Single Annotator Task (assigned to one Annotator) or a Double Annotator Task (assigned to two annotators, who will annotatote the assigned task in parallel).", + "A set of rules defining the TextGrid files' structure and content of the annotations. It is set at the beginning of the Annotation Campaign's creation, and is used to enforce that all TextGrids from the campaign contain the same amount of Tiers, with the same names. It can also enforce, for certain chosen tiers, a set of valid annotations.", + "Users with the rights to create Annotation Campaigns and Annotators user accounts, and assign Annotation Tasks to Annotators.", + "Users who are assigned a set of Annotation Tasks. Their job is to complete the annotation of the audio files with the Praat software.", + "If the TextGrid file they submit does not comply with their Annotation Task's TextGrid Checking Scheme, Seshat pinpoint their annotation errors with detailed messages. The annotator can re-submit the concerned file to the platform based on these different feedbacks.", + "Once they they connected to their instance of Seshat, campaign managers can access ongoing annotation campaigns or create new ones. Campaign managers are able to add annotators, assign annotation tasks and track progress. Annotator see a list of assigned tasks. The first step for them is to download the sound file with its corresponding auto-generated template TextGrid. In the current implementation, the annotation work has to be done locally with Praat. An upcoming version will use of web tools like Praat on the web BIBREF10. Once the task is completed, the TextGrid file is to be uploaded to Seshat via the web interface. We used the TextGrid format because of the wide acceptance of the Praat software in the speech science community (e.g., language acquisition research, clinical linguistics, phonetics and phonology).", + "The Textgrid Checking Scheme that encompasses rules on the tier's naming, file structure, and the content of the annotations, is associated with a specific campaign and defined at the creation of the campaign. Seshat back-end will automatically check that the submitted TextGrid file conforms to the Annotation Campaign's Textgrid Checking Scheme.", + "Seshat allows the campaign manager to create two type of tasks: single annotator, and double annotator. Regarding the first task, one audio file is attributed to one annotator. Once the annotation is completed, Sesha automatically checks the conformity of the annotation, and only declares a tasks completed if the conformity checks is passed. Regarding the second task, one audio file is attributed to two annotators. The two annotators annotate the same file independently, then the two versions are merged and the annotators are guided through a compare and review process to agree one final version. We summarise in the Figure FIGREF7 the different steps for the double-annotator task. At each step during merging, the two annotators are provided feedbacks to focus on where are the disagreements. This process also results in the computation of an Inter-annotator agreement for each file. The double annotator task can be used to train new annotators alongside experts.", + "Annotating speech data is a joint task of segmentation and categorisation of audio events. That is why we adopted the $\\gamma $ measure BIBREF8 to evaluate the inter- or intra- annotator agreement in each individual tier. Campaign manager can customise the distance used by $\\gamma $ by inserting a custom distance along their own parser (See short snippet of code for a parser of French Phonetics with the SAMPA alphabet in Algorithm )." + ], + [ + "Our utmost priority when building Seshat was to make it as easy as possible for others to deploy, use, administer and eventually contribute to. To do so, we chose the most common frameworks that are free and open-source, all of which are detailed in the following sections. Additionally, to match the current trend in web development, we decided to use the so-called \"web-app\" architecture for Seshat, i.e., we separated the application into two distinct entities: a front-end, running on the browser, and a back-end, serving data to the front-end and interacting with the database." + ], + [ + "The back-end system runs on a server. It holds and updates the campaign databases and runs the annotation checking and inter-rater agreement evaluation services. We chose Python, given its widespread use in the scientific community, with a wide array of speech and linguistic packages. Moreover, its usage on the back-end side will allow the future integration of powerful speech processing tools like Pyannote BIBREF16 to semi-automatize annotations. We thus went for Python3.6 for Seshat's server back-end. We used the Flask-Smorest extension (which is based on Flask) to clearly and thoroughly document our API, which can be exported to the popular OpenAPI 3.0.2 RESTful API description format.", + "The files and server data are stored on a MongoDB database, chosen for its flexible document model and general ease of use. We used the Object-Relational Mapping (ORM) MongoEngine to define our database schemas and interact with that database. MongoDB's GridFS system also allowed us to directly store annotation files (which are usually very light-weight) directly in the database, instead of going through the file system." + ], + [ + "The front-end handles all of the interactions between the users (campaing manager or annotator) with the databses. It is implemented as an App within their browser. We decided to base Seshat's front-end on the Angular Typescript framework. Despite its' steep learning curve, it enforces strict design patterns that guarantee that others can make additions to our code without jeopardising the stability of the App. Angular Typescript has a wide community support in the web development industry and is backed by Google and Microsoft. Moreover, the fact that it is based on TypeScript alleviates the numerous shortcomings of JavaScript, ensuring our implementation's readability and stability." + ], + [ + "The interface and the features we selected for our implementation are the process of a year-long iterative process involving a team of annotators, two campaign managers and software engineers. We followed some guiding principles from the recent Material design language. Our goal while designing our interface (with the help of a professional designer) was to make it fully usable by non-technical people. We also put some extra care into the annotators' interface to give them a clear sense of what is to be done, how they should follow the annotation protocol, and how to correct potential errors in their annotations (See Figure FIGREF21) The goal was to reduce the number of actions to perform for annotators and enable to focus only on the annotations content." + ], + [ + "Setting up a modern fully-fledged web service is a arduous task, usually requiring a seasoned system administrator as well as sometimes having very precise system requirements. Luckily, the Docker virtualisation platform ensures that anyone with a recent-enough install of that software can set up Seshat in about one command (while still allowing some flexibility via a configuration file). For those willing to have a more tightly-controlled installation of Seshat on their system, we also fully specify the manual installation steps in our online documentation).", + "Importing an audio corpus that you are willing to annotate is easy as dropping files into a default `corpora/` folder. It is possible to either drop a folder containing audio files (with no constraints on the folder's structure), or a CSV file listing audio filenames along with their durations (in case the files are sensitive and you're not willing to risk them being hosted on the server). It is then possible to review the automatically imported files via the web interface." + ], + [ + "The Campaign manager can easily define and monitor annotation campaign. As shown in Figure FIGREF33, the online form enable to choose corpora, pre-define and pre-configure the annotations scheme (tiers and parsers). There are 2 types of tiers already implemented by default: one with no check at all, and one with pre-defined categories. For the latter, these categories are pre-defined when the campaign is created.", + "Only Campaign managers can access and build new campaigns. If Campaign manager have several campaigns they can easily switch between them via the menu bar or get a full overview with the dashboard (See Figure FIGREF26). The campaign managers can visualise the progress of the assigned tasks at the campaign level or more precisely at the task level. They can retrieve all the intermediate files that have been created for each task. For instance, the campaign manager can examine qualitatively and quantitatively what are the annotation differences before the merge phases of the double annotator task." + ], + [ + "For those willing to interact with Seshat using code, it is possible to interact with Seshat using either its RESTful API or its command-line interface (CLI). The API endpoints that can be called are all listed in a simple interface, and can be made from any programming language able to make HTTP requests. The CLI interface can be used via your terminal, and therefore can be interacted with using Bash scripts.", + "A typical usage of these features would be to assign annotation tasks from a large speech corpus (spoken by several speakers) to a large pool of annotators, all the while making sure each annotator has a similar number of tasks, with each speaker being evenly distributed among annotators as well. This would be tedious to do manually via the user interface, but easy to program in any scripting language." + ], + [ + "We aimed at a reasonable trade-off between simplicity and flexibility for the TextGrid annotations checking component. However, we understand (from our own experience in particular) that sometimes annotations can follow a very specific and complex standard (for instance, parsing SAMPA phonemes strings). To allow users to define their own annotation standards, we added the possibility for users to define an annotation parser, via a simple package-based extension system (taking inspiration from pyannote's extension system). Anyone willing to create a new annotation parser has to be able to program in Python and have a minimal understanding of its packaging system.", + "As presented in our example French SAMPA Parser (Algorithm ), implementing a custom annotation parsers only requires the overload of two methods from Seshat's BaseCustomParser class:", + "check-annotation: takes an annotation string as input and raises an error if and only if the annotation is deemed to be invalid. It doesn't return anything.", + "distance: takes two annotations as input and should return a float corresponding to the distance between these two annotations." + ], + [ + "It is necessary have a measure of confidence to obtain high-quality datasets and therefore to draw valid conclusions from annotations. Annotations tasks of audio and speech data usually have some specificities. The items to annotate have to be both segmented in time and categorised. The segments can be hierarchically defined or overlapping. In addition, the audio stream may require only sparse annotations (especially in-the-wild recordings which contain a lot of non-speech segments). To evaluate speech annotations, the measure needs to take these characteristics into account. That is why we decided to re-implement and compute the $\\gamma $ measure (see mathet2015unified for its design and the advantages of this measure over previous agreement measures).", + "First, the $\\gamma $ software aligns (tier-wise) the annotations of the different annotators. To align the two sets of annotations the $\\gamma $ measure the distance between all the individual units. The difference of position of two annotated units $u$ and $v$ is measured with the positional distance:", + "", + "If the tiers are categorical, the distance for the content of the annotated units $u$ and $v$ is defined as:", + "", + "This distance can be over-written by the custom parser as mentioned above. These two distance are summed with equal weights to obtain the distance between every annotated units from 2 annotators. Then, it is possible to obtain the disorder $\\delta (a)$ of a specific alignment $a$ by summing the distance of all the aligned units in $a$. All possible alignments $a$ are considered and the one that minimises the disorder $\\delta (a)$ is kept.", + "To get the value of $\\gamma $, the disorder is chance-corrected to obtain an expected disorder. It is obtained by re-sampling randomly the annotations of the annotators. This means that real annotations are drawn from the annotators, and one position in the audio is randomly chosen. The annotation is split at this random position and the two parts are permuted. It is then possible to obtain an approximation of the expected disorder $\\delta _e$. The final agreement measure is defined as:", + "This $\\gamma $ measure is automatically computed by the back-end server for the double-annotator tasks. The Campaign manager can retrieve these measures in Seshat by downloading a simple CSV file." + ], + [ + "We present two use cases on which Seshat was developped: clinical interviews, and daylong child-centered recordings." + ], + [ + "Seshat was intially developped to study the impact of Huntington's Disease BIBREF17 on speech and language production. One hundred and fifty two interviews between a neuropsychologist and a patient with the Huntington's Disease (HD) were recorded between June 2018 and November 2019. The campaign manager created a campaign with multiple tiers to annotate the turn takings and the speech/non speech boundaries of the utterances of the patient. For both tasks, the annotations did not need to cover completely the audio (sparsity property mentioned above). For the Turn-taking annotations, there are 3 pre-defined tiers, each one with a single class ('Patient', 'Non-Patient', and 'Noise'), which results in possible overlap between these classes. For the Utterance annotations, there is only one pre-defined class ('Utterance').", + "To this date, a total of 67 files have been fully annotated with the help of Seshat by a cohort of 18 speech pathologist students (see Figure FIGREF33). Among these, 16 have been done by 2 different annotators independently with the Double-annotator task. The results are summarised in Table TABREF34.", + "Even though there are more categories for Turn-Takings than Utterance (gut2004measuring reported that the more categories the more the task is difficult in speech annotations), the mean $\\gamma $ for the Turn-Takings $\\gamma = 0.64$ is slightly higher than the one for Utterance $\\gamma = 0.61$. And the range of values for the Turn-Takings is smaller than the Utterance. Indeed, the speech pathologists reported the difficulty to annotate the boundary of utterances in spontaneous speech, with several ambiguous cases due to pauses. These results will help us to redefine the protocol and be more precise on the given instructions." + ], + [ + "The Seshat software is also currently used to annotate audio files in a study of day-long audio-recordings captured by two devices (LENA BIBREF18, and a BabyCloud baby-logger device) worn by young children growing up in remote Papua New Guinea. The project aims at establishing language input and outcomes in this seldom-studied population. To establish reliability levels, 20 1-min files were double-annotated by 2 speech pathology students. Among the tasks given to the annotators there was: (1) locating the portions of Speech (Speech activity), (2) locating the speech produced by an adult that is directed to a child or not (Adult-Directed Speech versus Child-Directed Speech). As in the previous example, the annotations do not need to cover the full audio file. The Speech Activity task has only 1 class ('Speech') and the Addressee task has 2 classes ('ADS', 'CDS').", + "These recordings have been done in naturalistic and noisy conditions; moreover, the annotators do not understand the language. Probably as a result of these challenges, agreement between annotators is lower than in the Clinical interviews use case. This information is nonetheless valuable to the researchers, as it can help them appropriately lower their confidence in the ensuing speech quantity estimates." + ], + [ + "Seshat is a new tool for the management of audio annotation efforts. Seshat enables users to define their own campaign of annotations. Based on this configuration, Seshat automatically enforces the format of the annotations returned by the annotators. Besides, we also add the capability to finely tailor the parsing of the annotations. Finally, Seshat provides automatic routines to compute the inter-rate agreements that are specifically designed for audio annotations. Seshat lays some foundations for more advanced features, either for the interface or the annotation capabilities. In future work, we plan to implement an automatic task assignments and an integration of a diarization processing step to reduce human effort. Another planned feature is to add possibility for the campaign manager to design more complex annotation workflows such as, for instance, dependencies between tiers or more intermediate steps of annotations." + ], + [ + "This research was conducted thanks to Agence Nationale de la Recherche (ANR-17-CE28-0007 LangAge, ANR-16-DATA-0004 ACLEW, ANR-14-CE30-0003 MechELex, ANR-17-EURE-0017, ANR-10-IDEX-0001-02 PSL*) and grants from Facebook AI Research (Research Grant), Google (Faculty Research Award), and Microsoft Research (Azure Credits and Grant), and a J. S. McDonnell Foundation Understanding Human Cognition Scholar Award." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1341/instruction.md b/qasper-1341/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ab0c46f1ec9f55d5aaeeff94709f782c42753494 --- /dev/null +++ b/qasper-1341/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Discriminating between similar languages in Twitter using label propagation + +Question: What information is contained in the social graph of tweet authors? \ No newline at end of file diff --git a/qasper-1346/instruction.md b/qasper-1346/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..a0efa8f5b0faa090133de8c8cc993e97a1b0b2d9 --- /dev/null +++ b/qasper-1346/instruction.md @@ -0,0 +1,147 @@ +Name of Paper: Classifying movie genres by analyzing text reviews + +Question: what evaluation metrics are discussed? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Theory", + "Preprocessing", + "Models", + "Evaluation", + "Data", + "Method", + "Data collection", + "Data preprocessing", + "Model", + "Result", + "Discussion", + "Conclusion", + "All genres" + ], + "paragraphs": [ + [ + "By only reading a single text review of a movie it can be difficult to say what the genre of that movie is, but by using text mining techniques on thousands of movie reviews is it possible to predict the genre?", + "This paper explores the possibility of classifying genres of a movie based only on a text review of that movie. This is an interesting problem because to the naked eye it may seem difficult to predict the genre by only looking at a text review. One example of a review can be seen in the following example:", + "I liked the film. Some of the action scenes were very interesting, tense and well done. I especially liked the opening scene which had a semi truck in it. A very tense action scene that seemed well done. Some of the transitional scenes were filmed in interesting ways such as time lapse photography, unusual colors, or interesting angles. Also the film is funny is several parts. I also liked how the evil guy was portrayed too. I'd give the film an 8 out of 10.", + "http://www.imdb.com/title/tt0211938/reviews", + "From the quoted review, one could probably predict the movie falls in the action genre; however, it would be difficult to predict all three of the genres (action, comedy, crime) that International Movie Database (IMDB) lists. With the use of text mining techniques it is feasible to predict multiple genres based on a review.", + "There are numerous previous works on classifying the sentiment of reviews, e.g., maas-EtAl:2011:ACL-HLT2011 by BIBREF0 . There are fewer scientific papers available on specifically classifying movie genres based on reviews; therefore, inspiration for this paper comes from papers describing classification of text for other or general contexts. One of those papers is DBLP:journals/corr/cmp-lg-9707002 where BIBREF1 describe how to use a multilayer perceptron (MLP) for genre classification.", + "All data, in the form of reviews and genres, used in this paper originates from IMDb." + ], + [ + "In this section all relevant theory and methodology is described. Table TABREF1 lists basic terminology and a short description of their meaning." + ], + [ + "Data preprocessing is important when working with text data because it can reduce the number of features and it formats the data into the desired form BIBREF2 .", + "Removing stop words is a common type of filtering in text mining. Stop words are words that usually contain little or no information by itself and therefore it is better to remove them. Generally words that occur often can be considered stop words such as the, a and it. BIBREF2 ", + "Lemmatization is the process of converting verbs into their infinitive tense form and nouns into their singular form. The reason for doing this is to reduce words into their basic forms and thus simplify the data. For example am, are and is are converted to be. BIBREF2 ", + "A way of representing a large corpus is to calculate the Term Frequency Inverse Document Frequency (tf-idf) of the corpus and then feed the models the tf-idf. As described in ramos2003using by BIBREF3 tf-idf is both efficient and simple for matching a query of words with a document in a corpus. Tf-idf is calculated by multiplying the Term Frequency (tf) with the Inverse Document Frequency (idf) , which is formulated as DISPLAYFORM0 ", + "where INLINEFORM0 is a document in corpus INLINEFORM1 and INLINEFORM2 is a term. INLINEFORM3 is defined as DISPLAYFORM0 ", + "and INLINEFORM0 is defined as DISPLAYFORM0 ", + "where INLINEFORM0 is the number of times INLINEFORM1 occurs in INLINEFORM2 and INLINEFORM3 total number of documents in the corpus." + ], + [ + "MLP is a class of feedforward neural network built up by a layered acyclic graph. An MLP consists of at least three layers and non-linear activations. The first layer is called input layer, the second layer is called hidden layer and the third layer is called output layer. The three layers are fully connected which means that every node in the hidden layer is connected to every node in the other layers. MLP is trained using backpropagation, where the weights are updated by calculating the gradient descent with respect to an error function. BIBREF4 ", + "K-nearest Neighbors (KNN) works by evaluating similarities between entities, where INLINEFORM0 stands for how many neighbors are taken into account during the classification. KNN is different from MLP in the sense that it does not require a computationally heavy training step; instead, all of the computation is done at the classification step. There are multiple ways of calculating the similarity, one way is to calculate the Minkowski distance. The Minkowski distance between two points DISPLAYFORM0 ", + "and DISPLAYFORM0 ", + "is defined by DISPLAYFORM0 ", + "where INLINEFORM0 which is equal to the Euclidean distance. BIBREF2 " + ], + [ + "When evaluating classifiers it is common to use accuracy, precision and recall as well as Hamming loss. Accuracy, precision and recall are defined by the the four terms true positive ( INLINEFORM0 ), true negative ( INLINEFORM1 ), false positive ( INLINEFORM2 ) and false negative ( INLINEFORM3 ) which can be seen in table TABREF16 .", + "Accuracy is a measurement of how correct a model's predictions are and is defined as DISPLAYFORM0 ", + ".", + "Precision is a ratio of how often positive predictions actually are positve and is defined as DISPLAYFORM0 ", + ".", + "Recall is a measurement of how good the model is to find all true positives and is defined as DISPLAYFORM0 ", + ". BIBREF5 ", + "It has been shown that when calculating precision and recall on multi-label classifiers, it can be advantageous to use micro averaged precision and recall BIBREF6 . The formulas for micro averaged precision are expressed as DISPLAYFORM0 DISPLAYFORM1 ", + "where INLINEFORM0 is label index and INLINEFORM1 is number of labels.", + "Hamming loss is different in the sense that it is a loss and it is defined as the fraction of wrong labels to the total number of labels. Hamming loss can be a good measurement when it comes to evaluating multi-label classifiers. the hamming loss is expressed as DISPLAYFORM0 ", + "where INLINEFORM0 is number of documents, INLINEFORM1 number of labels, INLINEFORM2 is the target value and INLINEFORM3 is predicted value. BIBREF7 ", + "For evaluation the INLINEFORM0 and INLINEFORM1 was calculated as defined in section SECREF15 for both the MLP model and the KNN model. For precision and recall formulas EQREF20 and EQREF21 were used because of their advantage in multi-label classification. The distribution of predicted genres was also shown in a histogram and compared to the target distribution of genres.", + "Furthermore the ratio of reviews that got zero genres predicted was also calculated and can be expressed as DISPLAYFORM0 ", + "where INLINEFORM0 is the number of reviews without any predicted genre and INLINEFORM1 is the total amount of predicted reviews." + ], + [ + "Data used in this paper comes from two separate sources. The first source was Large Movie Review Dataset v1.0 BIBREF0 which is a dataset for binary sentiment analysis of moview reviews. The dataset contains a total of 50000 reviews in raw text together with information on whether the review is positive or negative and a URL to the movie on IMDb. The sentiment information was not used in this paper. Out of the 50000, reviews only 7000 were used because of limitations on computational power, resulting in a corpus of 7000 documents.", + "The second source of data was the genres for all reviews which were scraped from the IMDb site. A total of 27 different genres were scraped. A list of all genres can be find in Appendix SECREF8 . A review can have one genre or multiple genres. For example a review can be for a movie that is both Action, Drama and Thriller at the same time while another move only falls into Drama." + ], + [ + "This section presents all steps needed to reproduce the results presented in this paper." + ], + [ + "In this paper the data comes from two sources where the first is a collection of text reviews. Those reviews were downloaded from Large Movie Review Datasets website . Because only 7000 reviews was used in this paper all of them were from the `train` folder and split evenly between positive reviews and negative reviews.", + "The genres for the reviews where obtained by iterating through all reviews and doing the following steps:", + "Save the text of the review.", + "Retrieve IMDb URL to the movie from the Large Movie Review Datasets data.", + "Scrape that movie website for all genres and download the genres.", + "The distribution of genres was plotted in a histogram to check that the scraped data looked reasonable and can be seen in figure FIGREF27 . All genres with less than 50 reviews corresponding to that genre were removed.", + "The number of genres per review can be seen in figure FIGREF28 and it shows that it is most common for a review to have three different genres; furthermore, it shows that no review has more than three genres.", + "http://ai.stanford.edu/ amaas/data/sentiment" + ], + [ + "All reviews were preprocessed according to the following steps:", + "Remove all non-alphanumeric characters.", + "Lower case all tokens.", + "Remove all stopwords.", + "Lemmatize all tokens.", + "Both the removal of stopwords and lemmatization were done with Python's Natural Language Toolkit (NLTK). Next the reviews and corresponding genres were split into a training set and a test set with INLINEFORM0 devided into the train set and INLINEFORM1 into the test set.", + "The preprocessed corpus was then used to calculate a tf-idf representing all reviews. The calculation of the tf-idf was done using scikit-learn'smodule TfidfVectorizer. Both transform and fit were run on the training set and only the transform was run on the test set. The decision to use tf-idf as a data representation is supported by BIBREF3 in ramos2003using which concludes that tf-idf is both simple and effective at categorizing relevant words.", + "https://www.python.org http://www.nltk.org http://scikit-learn.org" + ], + [ + "This paper experimented with two different models and compared them against each other. The inspiration for the first model comes from BIBREF1 in their paper DBLP:journals/corr/cmp-lg-9707002 where they used an MLP for text genre detection. The model used in this paper comes from scikit-learn's neural_network module and is called MLPClassifier. Table TABREF35 shows all parameters that were changed from the default values.", + "The second model was a KNN which was chosen because of it is simple and does not require the pre-training that the MLP needs. The implementation of this model comes from scikit-learn's neighbors module and is called KNeighborsClassifier. The only parameter that was changed after some trial and error was the k-parameter which was set to 3.", + "Both models were fitted using the train set and then predictions were done for the test set." + ], + [ + "Table TABREF38 shows the INLINEFORM0 , INLINEFORM1 and INLINEFORM2 for the models. The KNN model had a higher accuracy of INLINEFORM3 compared to MPL's accuracy of INLINEFORM4 and the KNN model had a higher recall but slightly lower precision than the MLP model.", + "Table TABREF39 shows the INLINEFORM0 and INLINEFORM1 for the models, it shows that the KNN model had lower values for both the INLINEFORM2 and INLINEFORM3 compared to the MLP model.", + "Figure FIGREF40 shows the distribution of the genres for the predicted values when using MLP and the test set. The same comparison between KNN and the test set can be seen in figure FIGREF41 ." + ], + [ + "When looking at the results it is apparent that KNN is better than MLP in these experiments. In particular, the INLINEFORM0 stands out between KNN and MLP where KNN got INLINEFORM1 and MLP got INLINEFORM2 which is considered a significant difference. Given that the INLINEFORM3 was relatively high for both models, this result hints that the models only predicted genres when the confidence was high, which resulted in fewer genres being predicted than the target. This can also be confirmed by looking at the figures FIGREF40 and FIGREF41 where the absolute number of reviews predicted for most genres was lower than the target. This unsatisfyingly low INLINEFORM4 can be explained by the multi-label nature of the problem in this paper. Even if the model correctly predicted 2 out of three genres it is considered a misclassification. A reason for the low accuracy could be that the models appeared to be on the conservative side when predicting genres.", + "Another factor that affected the performance of the models was the INLINEFORM0 which confirmed that over INLINEFORM1 of the reviews for the KNN model and over INLINEFORM2 of the reviews for the MLP model did not receive any predicted genre. Because no review had zero genres all predictions with zero genres are misclassified and this could be a good place to start when improving the models.", + "Furthermore, when looking at the INLINEFORM0 it shows that when looking at the individual genres for all reviews the number of wrong predictions are very low which is promising when trying to answer this paper's main question: whether it is possible to predict the genre of the movie associated with a text review. It should be taken into account that this paper only investigated about 7000 movie reviews and the results could change significantly, for better or for worse, if a much larger data set was used. In this paper, some of the genres had very low amounts of training data, which could be why those genres were not predicted in the same frequency as the target. An example of that can be seen by looking at genre Sci-Fi in figure FIGREF40 ." + ], + [ + "This paper demonstrates that by only looking at text reviews of a movie, there is enough information to predict its genre with an INLINEFORM0 of INLINEFORM1 . This result implies that movie reviews carry latent information about genres. This paper also shows the complexity of doing prediction on multi-label problems, both in implementation and data processing but also when it comes to evaluation. Regular metrics typically work, but they mask the entire picture and the depth of how good a model is.", + "Finally this paper provides an explanation of the whole process needed to conduct an experiment like this. The process includes downloading a data set, web scraping for extra information, data preprocessing, model tuning and evaluation of the results." + ], + [ + "Action", + "Adult", + "Adventure", + "Animation", + "Biography", + "Comedy", + "Crime", + "Documentary", + "Drama", + "Family", + "Fantasy", + "Film-Noir", + "Game-Show", + "History", + "Horror", + "Music", + "Musical", + "Mystery", + "Reality-TV", + "Romance", + "Sci-Fi", + "Short", + "Sport", + "Talk-Show", + "Thriller", + "War", + "Western" + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1363/instruction.md b/qasper-1363/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..2182c09b288e3958877fe376e1941d3146305894 --- /dev/null +++ b/qasper-1363/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Towards Language Agnostic Universal Representations + +Question: Did they experiment with tasks other than word problems in math? \ No newline at end of file diff --git a/qasper-1379/instruction.md b/qasper-1379/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..74c80296247a68643786fba6b4a5b0226f8a9f90 --- /dev/null +++ b/qasper-1379/instruction.md @@ -0,0 +1,121 @@ +Name of Paper: Adversarial Examples with Difficult Common Words for Paraphrase Identification + +Question: What are benhmark datasets for paraphrase identification? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work ::: Deep Paraphrase Identification", + "Related Work ::: Adversarial Examples for NLP", + "Related Work ::: Adversarial Example Generation", + "Methodology ::: Task Definition", + "Methodology ::: Algorithm Framework", + "Methodology ::: Original Example Sampling", + "Methodology ::: Replaceable Position Pairs", + "Methodology ::: Candidate Substitution Word Generation", + "Methodology ::: Beam Search for Finding Adversarial Examples", + "Experiments ::: Datasets", + "Experiments ::: Target Models", + "Experiments ::: Implementation Details", + "Experiments ::: Main Results", + "Experiments ::: Manual Evaluation", + "Experiments ::: Adversarial Training", + "Experiments ::: Sentence Pair BOW Similarity", + "Experiments ::: Effectiveness of Paired Common Words", + "Conclusion" + ], + "paragraphs": [ + [ + "Paraphrase identification is to determine whether a pair of sentences are paraphrases of each other BIBREF0. It is important for applications such as duplicate post matching on social media BIBREF1, plagiarism detection BIBREF2, and automatic evaluation for machine translation BIBREF3 or text summarization BIBREF4.", + "Paraphrase identification can be viewed as a sentence matching problem. Many deep models have recently been proposed and their performance has been greatly advanced on benchmark datasets BIBREF5, BIBREF6, BIBREF7. However, previous research shows that deep models are vulnerable to adversarial examples BIBREF8, BIBREF9 which are particularly constructed to make models fail. Adversarial examples are of high value for revealing the weakness and robustness issues of models, and can thereby be utilized to improve the model performance for challenging cases, robustness, and also security.", + "In this paper, we propose a novel algorithm to generate a new type of adversarial examples for paraphrase identification. To generate an adversarial example that consists of a sentence pair, we first sample an original sentence pair from the dataset, and then adversarially replace some word pairs with difficult common words respectively. Here each pair of words consists of two words from the two sentences respectively. And difficult common words are words that we adversarially select to appear in both sentences such that the example becomes harder for the target model. The target model is likely to be distracted by difficult common words and fail to judge the similarity or difference in the context, thereby making a wrong prediction.", + "Our adversarial examples are motivated by two observations. Firstly, for a sentence pair with a label matched, when some common word pairs are replaced with difficult common words respectively, models can be fooled to predict an incorrect label unmatched. As the first example in Figure FIGREF1 shows, we can replace two pairs of common words, \u201cpurpose\u201d and \u201clife\u201d, with another common words \u201cmeasure\u201d and \u201cvalue\u201d respectively. The modified sentence pair remains matched but fools the target model. It is mainly due to the bias between different words and some words are more difficult for the model. When such words appear in the example, the model fails to combine them with the unmodified context and judge the overall similarity of the sentence pair. Secondly, for an unmatched sentence pair, when some word pairs, not necessarily common words, are replaced with difficult common words, models can be fooled to predict an incorrect label matched. As the second example in Figure FIGREF1 shows, we can replace words \u201cGmail\u201d and \u201cschool\u201d with a common word \u201ccredit\u201d, and replace words \u201caccount\u201d and \u201cmanagement\u201d with \u201dscore\u201d. The modified sentences remain unmatched, but the target model can be fooled to predict matched for being distracted by the common words while ignoring the difference in the unmodified context.", + "Following these observations, we focus on robustness issues regarding capturing semantic similarity or difference in the unmodified part when distracted by difficult common words in the modified part. We try to modify an original example into an adversarial one with multiple steps. In each step, for a matched example, we replace some pair of common words together, with another word adversarially selected from the vocabulary; and for an unmatched example, we replace some word pair, not necessarily a common word pair, with a common word. In this way, we replace a pair of words together from two sentences respectively with an adversarially selected word in each step. To preserve the original label and grammaticality, we impose a few heuristic constraints on replaceable positions, and apply a language model to generate substitution words that are compatible with the context. We aim to adversarially find a word replacement solution that maximizes the target model loss and makes the model fail, using beam search.", + "We generate valid adversarial examples that are substantially different from those in previous work for paraphrase identification. Our adversarial examples are not limited to be semantically equivalent to original sentences and the unmodified parts of the two sentences are of low lexical similarity. To the best of our knowledge, none of previous work is able to generate such kind of adversarial examples. We further discuss our difference with previous work in Section 2.2.", + "In summary, we mainly make the following contributions:", + "We propose an algorithm to generate new adversarial examples for paraphrase identification. Our adversarial examples focus on robustness issues that are substantially different from those in previous work.", + "We reveal a new type of robustness issues in deep paraphrase identification models regarding difficult common words. Experiments show that the target models have a severe performance drop on the adversarial examples, while human annotators are much less affected and most modified sentences retain a good grammaticality.", + "Using our adversarial examples in adversarial training can mitigate the robustness issues, and these examples can foster future research." + ], + [ + "Paraphrase identification can be viewed as a problem of sentence matching. Recently, many deep models for sentence matching have been proposed and achieved great advancements on benchmark datasets. Among those, some approaches encode each sentence independently and apply a classifier on the embeddings of two sentences BIBREF10, BIBREF11, BIBREF12. In addition, some models make strong interactions between two sentences by jointly encoding and matching sentences BIBREF5, BIBREF13, BIBREF14 or hierarchically extracting matching features from the interaction space of the sentence pair BIBREF15, BIBREF16, BIBREF6. Notably, BERT pre-trained on large-scale corpora achieved even better results BIBREF7. In this paper, we study the robustness of recent typical deep models for paraphrase identification and generate new adversarial examples for revealing their robustness issues and improving their robustness." + ], + [ + "Many methods have been proposed to find different types of adversarial examples for NLP tasks. We focus on those that can be applied to paraphrase identification. Some of them generate adversarial examples by adding semantic-preserving perturbations to the input sentences. BIBREF17 added perturbations to word embeddings. BIBREF18, BIBREF19, BIBREF20, BIBREF21, BIBREF22 employed several character-level or word-level manipulations. BIBREF23 used syntactically controlled paraphrasing, and BIBREF24 paraphrased sentences with extracted rules. However, for some tasks including paraphrase identification, adversarial examples can be semantically different from original sentences, to study other robustness issues tailored to the corresponding tasks.", + "For sentence matching and paraphrase identification, other types of adversarial examples can be obtained by considering the relation and the correspondence between two sentences. BIBREF25 considered logical rules of sentence relations but can only generate unlabelled adversarial examples. BIBREF26 and BIBREF27 generated a sentence pair by modifying a single original sentence. They combined both original and modified sentences to form a pair. They modified the original sentence using back translation, word swapping, or single word replacement with lexical knowledge. Among them, back translation still aimed to produce semantically equivalent sentences; the others generated pairs of sentences with large Bag-of-Words (BOW) similarities, and the unmodified parts of the two sentences are exactly the same, so these same unmodified parts required little matching by target models. By contrast, we generate new adversarial examples with targeted labels by modifying a pair of original sentences together, using difficult common words. The modified sentences can be semantically different from original ones but still valid. The generated sentence pairs have much lower BOW similarities, and the unmodified parts are lexically diverse to reveal robustness issues regarding matching these parts when distracted by difficult common words in the modified parts. Thereby we study a new kind of robustness issues in paraphrase identification." + ], + [ + "For a certain type of adversarial examples, adversarial attacks or adversarial example generation aim to find examples that are within the defined type and make existing models fail. Some work has no access to the target model until an adversarial dataset is generated BIBREF28, BIBREF26, BIBREF23, BIBREF24, BIBREF29, BIBREF27. However, in many cases including ours, finding successful adversarial examples, i.e. examples on which the target model fails, is challenging, and employing an attack algorithm with access to the target model during generation is often necessary to ensure a high success rate.", + "Some prior work used gradient-based methods BIBREF30, BIBREF19, BIBREF31, requiring the model gradients to be accessible in addition to the output, and thus are inapplicable in black-box settings BIBREF21 where only model outputs are accessible. Though, the beam search in BIBREF19 can be adapted to black-box settings.", + "Gradient-free methods for NLP generally construct adversarial examples by querying the target model for output scores and making generation decisions to maximize the model loss. BIBREF25 searched in the solution space. One approach in BIBREF28 greedily made word replacements and queried the target model in several steps. BIBREF21 employed a genetic algorithm. BIBREF32 proposed a two-stage greedy algorithm and a method with gumbel softmax to improve the efficiency. In this work, we also focus on a black-box setting, which is more challenging than white-box settings. We use a two-stage beam search to find adversarial examples in multiple steps. We clarify that the major focus of this work is on studying new robustness issues and a new type of adversarial examples, instead of attack algorithms for an existing certain type of adversarial examples. Therefore, the choice of the attack algorithm is minor for this work as long as the success rates are sufficiently high." + ], + [ + "Paraphrase identification can be formulated as follows: given two sentences $P=p_1p_2\\cdots p_n$ and $Q=q_1q_2\\cdots q_m$, the goal is to predict whether $P$ and $Q$ are paraphrases of each other, by estimating a probability distribution", + "where $y\\in \\mathcal {Y} = \\lbrace matched, unmatched \\rbrace $. For each label $y$, the model outputs a score $[Z (P, Q)]_{y}$ which is the predicted probability of this label.", + "We aim to generate an adversarial example by adversarially modifying an original sentence pair $(P, Q)$ while preserving the label and grammaticality. The goal is to make the target model fail on the adversarially modified example $(\\hat{P}, \\hat{Q})$:", + "where $y$ indicates the gold label and $\\overline{y}$ is the wrong label opposite to the gold one." + ], + [ + "Figure FIGREF12 illustrates the work flow of our algorithm. We generate an adversarial example by firstly sampling an original example from the corpus and then constructing adversarial modifications. We use beam search and take multiple steps to modify the example, until the target model fails or the step number limit is reached. In each step, we modify the sentences by replacing a word pair with a difficult common word. There are two stages in deciding the word replacements. We first determine the best replaceable position pairs in the sentence pair, and next determine the best substitution words for the corresponding positions. We evaluate different options according to the target model loss they raise, and we retain $B$ best options after each stage of each step during beam search. Finally, the adversarially modified example is returned." + ], + [ + "To sample an original example from the dataset for subsequent adversarial modifications, we consider two different cases regarding whether the label is unmatched or matched. For the unmatched case, we sample two different sentence pairs $(P_1, Q_1)$ and $(P_2, Q_2)$ from the original data, and then form an unmatched example $(P_1, Q_2, unmatched)$ with sentences from two sentence pairs respectively. We also limit the length difference $||P_1|-|Q_2||$ and resample until the limit is satisfied, since sentence pairs with large length difference inherently tend to be unmatched and are too easy for models. By sampling two sentences from different examples, the two sentences tend to have less in common originally, which can help better preserve the label during adversarial modifications, while this also makes it more challenging for our algorithm to make the target model fail. On the other hand, matched examples cannot be sampled in this way, and thus for the matched case, we simply sample an example with a matched label from the dataset, namely, $(P, Q, matched)$." + ], + [ + "During adversarial modifications, we replace a word pair at each step. We set heuristic rules on replaceable position pairs to preserve the label and grammaticality. First of all, we require the words on the replaceable positions to be one of nouns, verbs, or adjectives, and not stopwords meanwhile. We also require a pair of replaceable words to have similar Part-of-Speech (POS) tags, i.e. the two words are both nouns, both verbs, or both adjectives. For a matched example, we further require the two words on each replaceable position pair to be exactly the same.", + "Figure FIGREF15 shows two examples of determining replaceable positions. For the first example (matched), only common words \u201cpurpose\u201d and \u201clife\u201d can be replaced. And since they are replaced simultaneously with another common words, the modified sentences are likely to talk about another same thing, e.g. changing from \u201cpurpose of life\u201d to \u201cmeasure of value\u201d, and thereby the new sentences tend to remain matched. As for the second example (unmatched), each noun in the first sentence, \u201cGmail\u201d and \u201caccount\u201d, can form replaceable word pairs with each noun in the second sentence, \u201cschool\u201d, \u201cmanagement\u201d and \u201csoftware\u201d. The irreplaceable part determines that the modified sentences are \u201cHow can I get $\\cdots $ back ? \u201d and \u201cWhat is the best $\\cdots $ ?\u201d respectively. Sentences based on these two templates are likely to discuss about different things or different aspects, even when filled with common words, and thus they are likely to remain unmatched. In this way, the labels can be preserved in most cases." + ], + [ + "For a pair of replaceable positions, we generate candidate substitution words that can replace the current words on the two positions. To preserve the grammaticality and keep the modified sentences like human language, substitution words should be compatible with the context. Therefore, we apply a BERT language model BIBREF7 to generate candidate substitution words. Specifically, when some words in a text are masked, the BERT masked language model can predict the masked words based on the context. For a sentence $x_1x_2\\cdots x_l$ where the $k$-th token is masked, the BERT masked language model gives the following probability distribution:", + "Thereby, to replace word $p_i$ and $q_j$ from the two sentences respectively, we mask $p_i$ and $q_j$ and present each sentence to the BERT masked language model. We aim to replace $p_i$ and $q_j$ with a common word $w$, which can be regarded as the masked word to be predicted. From the language model output, we obtain a joint probability distribution as follows:", + "We rank all the words within the vocabulary of the target model and choose top $K$ words with the largest probabilities, as the candidate substitution words for the corresponding positions." + ], + [ + "Once the replaceable positions and candidate substitution words can be determined, we use beam search with beam size $B$ to find optimal adversarial modifications in multiple steps. At step $t$, we perform a modification in two stages to determine replaceable positions and the corresponding substitution words respectively, based on the two-stage greedy framework by BIBREF32.", + "To determine the best replaceable positions, we enumerate all the possible position pairs, and obtain a set of candidate intermediate examples, $C_{pos}^{(t)}$, by replacing words on each position pair with a special token [PAD] respectively. We then query the target model with the examples in $C_{pos}^{(t)}$ to obtain the model output. We take top $B$ examples that maximize the output score of the opposite label $\\overline{y}$ (we define this operation as $\\mathop {\\arg {\\rm top}B}$), obtaining a set of intermediate examples $\\lbrace (\\hat{P}_{pos}^{(t,k)}, \\hat{Q}_{pos}^{(t,k)}) \\rbrace _{k=1}^{B}$, as follows:", + "We then determine difficult common words to replace the [PAD] placeholders. For each example in $\\lbrace (\\hat{P}_{pos}^{(t, k)}, \\hat{Q}_{pos}^{(t, k)}) \\rbrace _{k=1}^B$, we enumerate all the words in the candidate substitution word set of the corresponding positions with [PAD]. We obtain a set of candidate examples, $C^{(t)}$, by replacing the [PAD] placeholders with each candidate substitution word respectively. Similarly to the first stage, we take top $B$ examples that maximize the output score of the opposite label $\\overline{y}$. This yields a set of modified example after step $t$, $\\lbrace (\\hat{P}^{(t, k)}, \\hat{Q}^{(t, k)}) \\rbrace _{k=1}^{B}$, as follows:", + "After $t$ steps, for some modified example $(\\hat{P}^{(t,k)}, \\hat{Q}^{(t,k)})$, if the label predicted by the target model is already $\\overline{y}$, i.e. $[Z(\\hat{P}^{(t,k)}, \\hat{Q}^{(t,k)})]_{\\overline{y}} > [Z(\\hat{P}^{(t,k)},\\hat{Q}^{(t,k)})]_y$, this example is a successful adversarial example and thus we terminate the modification process. Otherwise, we continue taking another step, until the step number limit $S$ is reached and in case of that an unsuccessful adversarial example is returned." + ], + [ + "We adopt the following two datasets:", + "Quora BIBREF1: The Quora Question Pairs dataset contains question pairs annotated with labels indicating whether the two questions are paraphrases. We use the same dataset partition as BIBREF5, with 384,348/10,000/10,000 pairs in the training/development/test set respectively.", + "MRPC BIBREF34: The Microsoft Research Paraphrase Corpus consists of sentence pairs collected from online news. Each pair is annotated with a label indicating whether the two sentences are semantically equivalent. There are 4,076/1,725 pairs in the training/test set respectively." + ], + [ + "We adopt the following typical deep models as the target models in our experiments:", + "BiMPM BIBREF5, the Bilateral Multi-Perspective Matching model, matches two sentences on all combinations of time stamps from multiple perspectives, with BiLSTM layers to encode the sentences and aggregate matching results.", + "DIIN BIBREF6, the Densely Interactive Inference Network, creates a word-by-word interaction matrix by computing similarities on sentence representations encoded by a highway network and self-attention, and then adopts DenseNet BIBREF35 to extract interaction features for matching.", + "BERT BIBREF7, the Bidirectional Encoder Representations from Transformers, is pre-trained on large-scale corpora, and then fine-tuned on this task. The matching result is obtained by applying a classifier on the encoded hidden states of the two sentences." + ], + [ + "We adopt existing open source codes for target models BiMPM, DIIN and BERT, and also the BERT masked language model. For Quora, the step number limit $S$ is set to 5; the number of candidate substitution words generated using the language model $K$ and the beam size $B$ are both set to 25. $S$, $K$ and $B$ are doubled for MRPC where sentences are generally longer. The length difference between unmatched sentence pairs is limited to be no more than 3." + ], + [ + "We train each target model on the original training data, and then generate adversarial examples for the target models. For each dataset, we sample 1,000 original examples with balanced labels from the corresponding test set, and adversarially modify them for each target model. We evaluate the accuracies of target models on the corresponding adversarial examples, compared with their accuracies on the original examples. Let $s$ be the success rate of generating adversarial examples that the target model fails, the accuracy of the target model on the returned adversarial examples is $1-s$. Table TABREF18 presents the results.", + "The target models have high overall accuracies on the original examples, especially on the sampled ones since we form an unmatched original example with independently sampled sentences. The models have relatively lower accuracies on the unmatched examples in the full original test set of MRPC because MRPC is relatively small while the two labels are imbalanced in the original data (3,900 matched examples and 1,901 unmatched examples). Therefore, we generate adversarial examples with balanced labels instead of following the original distribution.", + "After adversarial modifications, the performance of the original target models (those without the \u201c-adv\u201d suffix) drops dramatically (e.g. the overall accuracy of BERT on Quora drops from 94.6% to 24.1%), revealing that the target models are vulnerable to our adversarial examples. Particularly, even though our generation is constrained by a BERT language model, BERT is still vulnerable to our adversarial examples. These results demonstrate the effectiveness of our algorithm for generating adversarial examples and also revealing the corresponding robustness issues. Moreover, we present some generated adversarial examples in the appendix.", + "We notice that the original models are more vulnerable to unmatched adversarial examples, because there are generally more replaceable position choices during the generation. Nevertheless, the results of the matched case are also sufficiently strong to reveal the robustness issues. We do not quantitatively compare the performance drop of the target models on the adversarial examples with previous work, because we generate a new type of adversarial examples that previous methods are not capable of. We have different experiment settings, including original example sampling and constraints on adversarial modifications, which are tailored to the robustness issues we study. Performance drop on different kinds of adversarial examples with little overlap is not comparable, and thus surpassing other adversarial examples on model performance drop is unnecessary and irrelevant to support our contributions. Therefore, such comparisons are not included in this paper." + ], + [ + "To verify the validity our generated adversarial examples, we further perform a manual evaluation. For each dataset, using BERT as the target model, we randomly sample 100 successful adversarial examples on which the target model fails, with balanced labels. We blend these adversarial examples with the corresponding original examples, and present each example to three workers on Amazon Mechanical Turk. We ask the workers to label the examples and also rate the grammaticality of the sentences with a scale of 1/2/3 (3 for no grammar error, 2 for minor errors, and 1 for vital errors). We integrate annotations from different workers with majority voting for labels and averaging for grammaticality.", + "Table TABREF35 shows the results. Unlike target models whose performance drops dramatically on adversarial examples, human annotators retain high accuracies with a much smaller drop, while the accuracies of the target models are 0 on these adversarial examples. This demonstrates that the labels of most adversarial examples are successfully preserved to be consistent with original examples. Results also show that the grammaticality difference between the original examples and adversarial examples is also small, suggesting that most adversarial examples retain a good grammaticality. This verifies the validity of our adversarial examples." + ], + [ + "Adversarial training can often improve model robustness BIBREF25, BIBREF27. We also fine-tune the target models using adversarial training. At each training step, we train the model with a batch of original examples along with adversarial examples with balanced labels. The adversarial examples account for around 10% in a batch. During training, we generate adversarial examples with the current model as the target and update the model parameters with the hybrid batch iteratively. The beam size for generation is set to 1 to reduce the computation cost, since the generation success rate is minor in adversarial training. We evaluate the adversarially trained models, as shown in Table TABREF18.", + "After adversarial training, the performance of all the target models raises significantly, while that on the original examples remain comparable. Note that since the focus of this paper is on model robustness which can hardly be reflected in original data, we do not expect performance improvement on original data. The results demonstrate that adversarial training with our adversarial examples can significantly improve the robustness we focus on without remarkably hurting the performance on original data. Moreover, although the adversarial example generation is constrained by a BERT language model, BiMPM and DIIN which do not use the BERT language model can also significantly benefit from the adversarial examples, further demonstrating the effectiveness of our method." + ], + [ + "To quantitatively demonstrate the difference between the adversarial examples we generate and those by previous work BIBREF26, BIBREF27, we compute the average BOW cosine similarity between the generated pairs of sentences. We only compare with previous methods that also aim to generate labeled adversarial examples that are not limited to be semantically equivalent to original sentences. Results are shown in Table TABREF38. Each pair of adversarial sentences by BIBREF26 differ by only one word. And in BIBREF27, sentence pairs generated with word swapping have exactly the same BOW. These two approaches both have high BOW similarities. By contrast, our method generates sentence pairs with much lower BOW similarities. This demonstrates a significant difference between our examples and the others. Unlike previous methods, we generate adversarial examples that can focus on robustness issues regarding the distraction from modified words that are the same for both sentences, towards matching the unmodified parts that are diverse for two sentences." + ], + [ + "We further analyse the necessity and effectiveness of modifying sentences with paired common words. We consider another version that replaces one single word independently at each step without using paired common words, namely the unpaired version. Firstly, for matched adversarial examples that can be semantically different from original sentences, the unpaired version is inapplicable, because the matched label can be easily broken if common words from two sentences are changed into other words independently. And for the unmatched case, we show that the unpaired version is much less effective. For a more fair comparison, we double the step number limit for the unpaired version. As shown in Table TABREF41, the performance of target models on unmatched examples generated by the unpaired version, particularly that of BERT, is mostly much higher than those by our full algorithm, except for BiMPM on MRPC but its accuracies have almost reached 0 (0.0% for unpaired and 0.2% for paired). This demonstrates that our algorithm using paired common words are more effective in generating adversarial examples, on which the performance of the target model is generally much lower. An advantage of using difficult common words for unmatched examples is that such words tend to make target models over-confident about common words and distract the models on recognizing the semantic difference in the unmodified part. Our algorithm explicitly utilizes this property and thus can well reveal such a robustness issue. Moreover, although there is no such a property for the matched case, replacing existing common words with more difficult ones can still distract the target model on judging the semantic similarity in the unmodified part, due to the bias between different words learned by the model, and thus our algorithm for generating adversarial examples with difficult common words works for both matched and unmatched cases." + ], + [ + "In this paper, we propose a novel algorithm to generate new adversarial examples for paraphrase identification, by adversarially modifying original examples with difficult common words. We generate labeled adversarial examples that can be semantically different from original sentences and the BOW similarity between each pair of sentences is generally low. Such examples reveal robustness issues that previous methods are not able for. The accuracies of the target models drop dramatically on our adversarial examples, while human annotators are much less affected and the modified sentences retain a good grammarticality. We also show that model robustness can be improved using adversarial training with our adversarial examples. Moreover, our adversarial examples can foster future research for further improving model robustness." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1390/instruction.md b/qasper-1390/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..737053bff9188581c69241ce08f07f56abac58d6 --- /dev/null +++ b/qasper-1390/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Citation Text Generation + +Question: How was the evaluation corpus collected? \ No newline at end of file diff --git a/qasper-1397/instruction.md b/qasper-1397/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..06b73c95e7cca5c2b1f86eb213dec869a1248647 --- /dev/null +++ b/qasper-1397/instruction.md @@ -0,0 +1,103 @@ +Name of Paper: SocialNLP EmotionX 2019 Challenge Overview: Predicting Emotions in Spoken Dialogues and Chats + +Question: How large is the first dataset? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Datasets", + "Datasets ::: Augmentation", + "Challenge Details", + "Submissions", + "Submissions ::: IDEA", + "Submissions ::: KU", + "Submissions ::: HSU", + "Submissions ::: Podlab", + "Submissions ::: AlexU", + "Submissions ::: Antenna", + "Submissions ::: CYUT", + "Results", + "Evaluation & Discussion", + "Evaluation & Discussion ::: Deep Learning Models.", + "Evaluation & Discussion ::: Unbalanced Labels.", + "Evaluation & Discussion ::: Emotional Model and Annotation Challenges.", + "Evaluation & Discussion ::: Data Sources.", + "Acknowledgment" + ], + "paragraphs": [ + [ + "Emotions are a central component of our existence as human beings, and are manifested by physiological and psychological changes that often affect behavior and action. Emotions involve a complicated interplay of mind, body, language, and culture BIBREF0.", + "Detecting and recognizing emotions is a difficult task for machines. Nevertheless, following the successful use of computational linguistics to analyze sentiment in texts, there is growing interest in the more difficult task of the automatic detection and classification of emotions in texts.", + "The detection of emotions in text is a complicated challenge for multiple reasons: first, emotions are complex entities, and no universally-agreed upon psychological model of emotions exists. Second, isolated texts convey less information compared to a complete human interaction in which emotions can be detected from the other person's facial expressions, listening to their tone of voice, etc. However, due to important applications in fields such as psychology, marketing, and political science, research in this topic is now expanding rapidly BIBREF1.", + "In particular, dialogue systems such as those available on social media or instant messaging services are rich sources of textual data and have become the focus of much attention. Emotions of utterances within dialogues can be detected more precisely due to the presence of more context. For example, a single utterance (\u201cOK!\u201d) might convey different emotions (happiness, anger, surprise), depending on its context. Taking all this into consideration, in 2018 the EmotionX Challenge asked participants to detect emotions in complete dialogues BIBREF2. Participants were challenged to classify utterances using Ekman's well-known theory of six basic emotions (sadness, happiness, anger, fear, disgust, and surprise) BIBREF3.", + "For the 2019 challenge, we built and expanded upon the 2018 challenge. We provided an additional 20% of data for training, as well as augmenting the dataset using two-way translation. The metric used was micro-F1 score, and we also report the macro-F1 score.", + "A total of thirty-six teams registered to participate in the challenge. Eleven of the teams successfully submitted their data for performance evaluation, and seven of them submitted technical papers for the workshop. Approaches used by the teams included deep neural networks and SVM classifiers. In the following sections we expand on the challenge and the data. We then briefly describe the various approaches used by the teams, and conclude with a summary and some notes. Detailed descriptions of the various submissions are available in the teams' technical reports." + ], + [ + "The two datasets used for the challenge are Friends and EmotionPush, part of the EmotionLines corpus BIBREF4. The datasets contain English-language dialogues of varying lengths. For the competition, we provided 1,000 labeled dialogues from each dataset for training, and 240 unlabeled dialogues from each dataset for evaluation. The Friends dialogues are scripts taken from the American TV sitcom (1994-2004). The EmotionPush dialogues are from Facebook Messenger chats by real users which have been anonymized to ensure user privacy. For both datasets, dialogue lengths range from 5 to 24 lines each. A breakdown of the lengths of the dialogues is shown in Table .", + "We employed workers using Amazon Mechanical Turk (aka AMT or MTurk) to annotate the dialogues BIBREF5. Each complete dialogue was offered as a single MTurk Human Intelligence Task (HIT), within which each utterance was read and annotated for emotions by the worker. Each HIT was assigned to five workers. To ensure workers were qualified for the annotation task, we set up a number of requirements: workers had to be from an English-speaking country (Australia, Canada, Great Britain, Ireland, New Zealand, or the US), have a high HIT approval rate (at least 98%), and have already performed a minimum of 2,000 HITs.", + "In the datasets, each utterance is accompanied by an annotation and emotion. The annotation contains the raw count of votes for each emotion by the five annotators, with the order of the emotions being Neutral, Joy, Sadness, Fear, Anger, Surprise, Disgust. For example, an annotation of \u201c2000030\u201d denotes that two annotators voted for \u201cneutral\u201d, and three voted for \u201csurprise\u201d.", + "The labeled emotion is calculated using the absolute majority of votes. Thus, if a specific emotion received three or more votes, then that utterance is labeled with that emotion. If there is no majority vote, the utterance is labeled with \u201cnon-neutral\u201d label. In addition to the utterance, annotation, and label, each line in each dialogue includes the speaker's name (in the case of EmotionPush, a speaker ID was used). The emotion distribution for Friends and EmotionPush, for both training and evaluation data, is shown in Table .", + "We used Fleiss' kappa measure to assess the reliability of agreement between the annotators BIBREF6. The value for $\\kappa $-statistic is $0.326$ and $0.342$ for Friends and EmotionPush, respectively. For the combined datasets the value of the $\\kappa $-statistic is $0.345$.", + "Sample excerpts from the two datasets, with their annotations and labels, are given in Table ." + ], + [ + "NLP tasks require plenty of data. Due to the relatively small number of samples in our datasets, we added more labeled data using a technique developed in BIBREF7 that was used by the winning team in Kaggle's Toxic Comment Classification Challenge BIBREF8. The augmented datasets are similar to the original data files, but include additional machine-computed utterances for each original utterance. We created the additional utterances using the Google Translate API. Each original utterance was first translated from English into three target languages (German, French, and Italian), and then translated back into English. The resulting utterances were included together in the same object with the original utterance. These \u201cduplex translations\u201d can sometimes result in the original sentence, but many times variations are generated that convey the same emotions. Table shows an example utterance (labeled with \u201cJoy\u201d) after augmentation." + ], + [ + "A dedicated website for the competition was set up. The website included instructions, the registration form, schedule, and other relevant details. Following registration, participants were able to download the training datasets.", + "The label distribution of emotions in our data are highly unbalanced, as can be seen in Figure FIGREF6. Due to the small number of three of the labels, participants were instructed to use only four emotions for labels: joy, sadness, anger, and neutral. Evaluation of submissions was done using only utterances with these four labels. Utterances with labels other than the above four (i.e., surprise, disgust, fear or non-neutral) were discarded and not used in the evaluation.", + "Scripts for verifying and evaluating the submissions were made available online. We used micro-F1 as the comparison metric." + ], + [ + "A total of eleven teams submitted their evaluations, and are presented in the online leaderboard. Seven of the teams also submitted technical reports, the highlights of which are summarized below. More details are available in the relevant reports." + ], + [ + "BIBREF9 Two different BERT models were developed. For Friends, pre-training was done using a sliding window of two utterances to provide dialogue context. Both Next Sentence Prediction (NSP) phase on the complete unlabeled scripts from all 10 seasons of Friends, which are available for download. In addition, the model learned the emotional disposition of each of six main six main characters in Friends (Rachel, Monica, Phoebe, Joey, Chandler and Ross) by adding a special token to represent the speaker. For EmotionPush, pre-training was performed on Twitter data, as it is similar in nature to chat based dialogues. In both cases, special attention was given to the class imbalance issue by applying \u201cweighted balanced warming\u201d on the loss function." + ], + [ + "BIBREF10 BERT is post-trained via Masked Language Model (MLM) and Next Sentence Prediction (NSP) on a corpus consisting of the complete and augmented dialogues of Friends, and the EmotionPush training data. The resulting token embeddings are max-pooled and fed into a dense network for classification. A $K$-fold cross-validation ensemble with majority voting was used for prediction. To deal with the class imbalance problem, weighted cross entropy was used as a training loss function." + ], + [ + "BIBREF11 A pre-trained BERT is fine-tuned using filtered training data which only included the desired labels. Additional augmented data with joy, sadness, and anger labels are also used. BERT is then fed into a standard feed-forward-network with a softmax layer used for classification." + ], + [ + "BIBREF12 A support vector machine (SVM) was used for classification. Words are ranked using a per-emotion TF-IDF score. Experiments were performed to verify whether the previous utterance would improve classification performance. Input to the Linear SVM was done using one-hot-encoding of top ranking words." + ], + [ + "BIBREF13 The classifier uses a pre-trained BERT model followed by a feed-forward neural network with a softmax output. Due to the overwhelming presence of the neutral label, a classifying cascade is employed, where the majority classifier is first used to decide whether the utterance should be classified with \u201cneutral\u201d or not. A second classifier is used to focus on the other emotions (joy, sadness, and anger). Dealing with the imbalanced classes is done through the use of a weighted loss function." + ], + [ + "BIBREF14 BERT is first used to generate word and sentence embeddings for all utterances. The resulting calculated word embeddings are fed into a Convolutional Neural Network (CNN), and its output is then concatenated with the BERT-generated sentence embeddings. The concatenated vectors are then used to train a bi-directional GRU with a residual connection followed by a fully-connected layer, and finally a softmax layer produces predictions. Class imbalance is tackled using focal loss BIBREF15." + ], + [ + "BIBREF16 A word embedding layer followed by a bi-directional GRU-based RNN. Output from the RNN was fed into a single-node classifier. The augmented dataset was used for training the model, but \u201cneutral\u201d-labeled utterances were filtered to deal with class imbalance." + ], + [ + "The submissions and the final results are summarized in Tables and . Two of the submissions did not follow up with technical papers and thus they do not appear in this summary. We note that the top-performing models used BERT, reflecting the recent state-of-the-art performance of this model in many NLP tasks. For Friends and EmotionPush the top micro-F1 scores were 81.5% and 88.5% respectively." + ], + [ + "An evaluation summary of the submissions is available in Tables and . We only present the teams that submitted technical reports. A full leaderboard that includes all the teams is available on the challenge website. This section highlights some observations related to the challenge. Identical utterances can convey different emotions in different contexts. A few of the models incorporated the dialogue context into the model, such as the models proposed by teams IDEA and KU." + ], + [ + "Most of the submissions used deep learning models. Five of the models were based on the BERT architecture, with some using pre-trained BERT. Some of the submissions enhanced the model by adding context and speaker related encoding to improve performance. We also received submissions using more traditional networks such as CNN, as well as machine learning classics such as SVM. The results demonstrate that domain knowledge, feature engineering, and careful application of existing methodologies is still paramount for building successful machine learning models." + ], + [ + "Emotion detection in text often suffers from a data imbalance problem, our datasets included. The teams used two approaches to deal with this issue. Some used a class-balanced loss functions while others under-sampled classes with majority label \u201cneutral\u201d. Classification performance of underrepresented emotions, especially sadness and anger, is low compared to the others. This is still a challenge, especially as some real-world applications are dependent on detection of specific emotions such as anger and sadness." + ], + [ + "The discrete 6-emotion model and similar models are often used in emotion detection tasks. However, such 1-out-of-n models are limited in a few ways: first, expressed emotions are often not discrete but mixed (for example, surprise and joy or surprise and anger are often manifested in the same utterance). This leads to more inter-annotator disagreement, as annotators can only select one emotion. Second, there are additional emotional states that are not covered by the basic six emotions but are often conveyed in speech and physical expressions, such as desire, embarrassment, relief, and sympathy. This is reflected in feedback we received from one of the AMT workers: \u201cI am doing my best on your HITs. However, the emotions given (7 of them) are a lot of times not the emotion I'm reading (such as questioning, happy, excited, etc). Your emotions do not fit them all...\u201d.", + "To further investigate, we calculated the per-emotion $\\kappa $-statistic for our datasets in Table . We see that for some emotions, such as disgust and fear (and anger for EmotionPush), the $\\kappa $-statistic is poor, indicating ambiguity in annotation and thus an opportunity for future improvement. We also note that there is an interplay between the emotion label distribution, per-emotion classification performance, and their corresponding $\\kappa $ scores, which calls for further investigation." + ], + [ + "One of the main requirements of successful training of deep learning models is the availability of high-quality labeled data. Using AMT to label data has proved to be useful. However, current data is limited in quantity. In addition, more work needs to be done in order to measure, evaluate, and guarantee annotation quality. In addition, the Friends data is based on an American TV series which emphasizes certain emotions, and it remains to be seen how to transfer learning of emotions to other domains." + ], + [ + "This research is partially supported by Ministry of Science and Technology, Taiwan, under Grant no. MOST108-2634-F-001-004- and MOST107-2218-E-002-009-." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1510/instruction.md b/qasper-1510/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..3050f8b2eaf6ad7b1dc2e2714edc7e0c96caa23c --- /dev/null +++ b/qasper-1510/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: The Effect of Context on Metaphor Paraphrase Aptness Judgments + +Question: Do they report results only on English data? \ No newline at end of file diff --git a/qasper-1517/instruction.md b/qasper-1517/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c520278be44554b70dc73a139f6ed3f915e9128f --- /dev/null +++ b/qasper-1517/instruction.md @@ -0,0 +1,104 @@ +Name of Paper: Drug-drug Interaction Extraction via Recurrent Neural Network with Multiple Attention Layers + +Question: What are the existing methods mentioned in the paper? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Proposed Model", + "Preprocessing", + "Embedding Layer", + "Bidirectional RNN Encoding Layer", + "Word Level Attention", + "Sentence Level Attention", + "Classification and Training", + "DDI Prediction", + "Datasets and Evaluation Metrics", + "Hyperparameter Settings and Training", + "Experimental Results", + "Conclusion and Future Work", + "Acknowledgment" + ], + "paragraphs": [ + [ + "Drug-drug interaction (DDI) is a situation when one drug increases or decreases the effect of another drug BIBREF0 . Adverse drug reactions may cause severe side effect, if two or more medicines were taken and their DDI were not investigated in detail. DDI is a common cause of illness, even a cause of death BIBREF1 . Thus, DDI databases for clinical medication decisions are proposed by some researchers. These databases such as SFINX BIBREF2 , KEGG BIBREF3 , CredibleMeds BIBREF4 help physicians and pharmacists avoid most adverse drug reactions.", + "Traditional DDI databases are manually constructed according to clinical records, scientific research and drug specifications. For instance, The sentence \u201cWith combined use, clinicians should be aware, when phenytoin is added, of the potential for reexacerbation of pulmonary symptomatology due to lowered serum theophylline concentrations BIBREF5 \u201d, which is from a pharmacotherapy report, describe the side effect of phenytoin and theophylline's combined use. Then this information on specific medicines will be added to DDI databases. As drug-drug interactions have being increasingly found, manually constructing DDI database would consume a lot of manpower and resources.", + "There has been many efforts to automatically extract DDIs from natural language BIBREF0 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , mainly medical literature and clinical records. These works can be divided into the following categories:", + "To avoid complex feature engineering and NLP toolkits' usage, we employ deep learning approaches for sentence comprehension as a whole. Our model takes in a sentence from biomedical literature which contains a drug pair and outputs what kind of DDI this drug pair belongs. This assists physicians refrain from improper combined use of drugs. In addition, the word and sentence level attentions are introduced to our model for better DDI predictions.", + "We train our language comprehension model with labeled instances. Figure FIGREF5 shows partial records in DDI corpus BIBREF16 . We extract the sentence and drug pairs in the records. There are 3 drug pairs in this example thus we have 3 instances. The DDI corpus annotate each drug pair in the sentence with a DDI type. The DDI type, which is the most concerned information, is described in table TABREF4 . The details about how we train our model and extract the DDI type from text are described in the remaining sections." + ], + [ + "In DDI extraction task, NLP methods or machine learning approaches are proposed by most of the work. Chowdhury BIBREF14 and Thomas et al. BIBREF11 proposed methods that use linguistic phenomenons and two-stage SVM to classify DDIs. FBK-irst BIBREF10 is a follow-on work which applies kernel method to the existing model and outperforms it.", + "Neural network based approaches have been proposed by several works. Liu et al. BIBREF9 employ CNN for DDI extraction for the first time which outperforms the traditional machine learning based methods. Limited by the convolutional kernel size, the CNN can only extracted features of continuous 3 to 5 words rather than distant words. Liu et al. BIBREF8 proposed dependency-based CNN to handle distant but relevant words. Sahu et al. BIBREF12 proposed LSTM based DDI extraction approach and outperforms CNN based approach, since LSTM handles sentence as a sequence instead of slide windows. To conclude, Neural network based approaches have advantages of 1) less reliance on extra NLP toolkits, 2) simpler preprocessing procedure, 3) better performance than text analysis and machine learning methods.", + "Drug-drug interaction extraction is a relation extraction task of natural language processing. Relation extraction aims to determine the relation between two given entities in a sentence. In recent years, attention mechanism and various neural networks are applied to relation extraction BIBREF17 , BIBREF18 , BIBREF19 , BIBREF20 , BIBREF21 . Convolutional deep neural network are utilized for extracting sentence level features in BIBREF19 . Then the sentence level features are concatenated with lexical level features, which are obtained by NLP toolkit WordNet BIBREF22 , followed by a multilayer perceptron (MLP) to classify the entities' relation. A fixed work is proposed by Nguyen et al. BIBREF21 . The convolutional kernel is set various size to capture more n-gram features. In addition, the word and position embedding are trained automatically instead of keeping constant as in BIBREF19 . Wang et al. BIBREF20 introduce multi-level attention mechanism to CNN in order to emphasize the keywords and ignore the non-critical words during relation detection. The attention CNN model outperforms previous state-of-the-art methods.", + "Besides CNN, Recurrent neural network (RNN) has been applied to relation extraction as well. Zhang et al. BIBREF18 utilize long short-term memory network (LSTM), a typical RNN model, to represent sentence. The bidirectional LSTM chronologically captures the previous and future information, after which a pooling layer and MLP have been set to extract feature and classify the relation. Attention mechanism is added to bidirectional LSTM in BIBREF17 for relation extraction. An attention layer gives each memory cell a weight so that classifier can catch the principal feature for the relation detection. The Attention based bidirectional LSTM has been proven better than previous work." + ], + [ + "In this section, we present our bidirectional recurrent neural network with multiple attention layer model. The overview of our architecture is shown in figure FIGREF15 . For a given instance, which describes the details about two or more drugs, the model represents each word as a vector in embedding layer. Then the bidirectional RNN layer generates a sentence matrix, each column vector in which is the semantic representation of the corresponding word. The word level attention layer transforms the sentence matrix to vector representation. Then sentence level attention layer generates final representation for the instance by combining several relevant sentences in view of the fact that these sentences have the same drug pair. Followed by a softmax classifier, the model classifies the drug pair in the given instance as specific DDI." + ], + [ + "The DDI corpus contains thousands of XML files, each of which are constructed by several records. For a sentence containing INLINEFORM0 drugs, there are INLINEFORM1 drug pairs. We replace the interested two drugs with \u201cdrug1\u201d and \u201cdrug2\u201d while the other drugs are replaced by \u201cdurg0\u201d, as in BIBREF9 did. This step is called drug blinding. For example, the sentence in figure FIGREF5 generates 3 instances after drug blinding: \u201cdrug1: an increased risk of hepatitis has been reported to result from combined use of drug2 and drug0\u201d, \u201cdrug1: an increased risk of hepatitis has been reported to result from combined use of drug0 and drug2\u201d, \u201cdrug0: an increased risk of hepatitis has been reported to result from combined use of drug1 and drug2\u201d. The drug blinded sentences are the instances that are fed to our model.", + "We put the sentences with the same drug pairs together as a set, since the sentence level attention layer (will be described in Section SECREF21 ) will use the sentences which contain the same drugs." + ], + [ + "Given an instance INLINEFORM0 which contains specified two drugs INLINEFORM1 , INLINEFORM2 , each word is embedded in a INLINEFORM3 dimensional space ( INLINEFORM4 , INLINEFORM5 are the dimension of word embedding and position embedding). The look up table function INLINEFORM6 maps a word or a relative position to a column vector. After embedding layer the sentence is represented by INLINEFORM7 , where DISPLAYFORM0 ", + "The INLINEFORM0 function is usually implemented with matrix-vector product. Let INLINEFORM1 , INLINEFORM2 denote the one-hot representation (column vector) of word and relative distance. INLINEFORM3 , INLINEFORM4 are word and position embedding query matrix. The look up functions are implemented by DISPLAYFORM0 ", + "Then the word sequence INLINEFORM0 is fed to the RNN layer. Note that the sentence will be filled with INLINEFORM1 if its length is less than INLINEFORM2 ." + ], + [ + "The words in the sequence are read by RNN's gated recurrent unit (GRU) one by one. The GRU takes the current word INLINEFORM0 and the previous GRU's hidden state INLINEFORM1 as input. The current GRU encodes INLINEFORM2 and INLINEFORM3 into a new hidden state INLINEFORM4 (its dimension is INLINEFORM5 , a hyperparameter), which can be regarded as informations the GRU remembered.", + "Figure FIGREF25 shows the details in GRU. The reset gate INLINEFORM0 selectively forgets informations delivered by previous GRU. Then the hidden state becomes INLINEFORM1 . The update gate INLINEFORM2 updates the informations according to INLINEFORM3 and INLINEFORM4 . The equations below describe these procedures. Note that INLINEFORM5 stands for element wise multiplication. DISPLAYFORM0 DISPLAYFORM1 ", + "The bidirectional RNN contains forward RNN and backward RNN. Forward RNN reads sentence from INLINEFORM0 to INLINEFORM1 , generating INLINEFORM2 , INLINEFORM3 , ..., INLINEFORM4 . Backward RNN reads sentence from INLINEFORM5 to INLINEFORM6 , generating INLINEFORM7 , INLINEFORM8 , ..., INLINEFORM9 . Then the encode result of this layer is DISPLAYFORM0 ", + "We apply dropout technique in RNN layer to avoid overfitting. Each GRU have a probability (denoted by INLINEFORM0 , also a hyperparameter) of being dropped. The dropped GRU has no output and will not affect the subsequent GRUs. With bidirectional RNN and dropout technique, the input INLINEFORM1 is encoded into sentence matrix INLINEFORM2 ." + ], + [ + "The purpose of word level attention layer is to extract sentence representation (also known as feature vector) from encoded matrix. We use word level attention instead of max pooling, since attention mechanism can determine the importance of individual encoded word in each row of INLINEFORM0 . Let INLINEFORM1 denotes the attention vector (column vector), INLINEFORM2 denotes the filter that gives each element in the row of INLINEFORM3 a weight. The following equations shows the attention operation, which is also illustrated in figure FIGREF15 . DISPLAYFORM0 DISPLAYFORM1 ", + "The softmax function takes a vector INLINEFORM0 as input and outputs a vector, DISPLAYFORM0 ", + " INLINEFORM0 denotes the feature vector captured by this layer. Several approaches BIBREF12 , BIBREF17 use this vector and softmax classifier for classification. Inspired by BIBREF23 we propose the sentence level attention to combine the information of other sentences for a improved DDI classification." + ], + [ + "The previous layers captures the features only from the given sentence. However, other sentences may contains informations that contribute to the understanding of this sentence. It is reasonable to look over other relevant instances when determine two drugs' interaction from the given sentence. In our implementation, the instances that have the same drug pair are believed to be relevant. The relevant instances set is denoted by INLINEFORM0 , where INLINEFORM1 is the sentence feature vector. INLINEFORM2 stands for how well the instance INLINEFORM3 matches its DDI INLINEFORM4 (Vector representation of a specific DDI). INLINEFORM5 is a diagonal attention matrix, multiplied by which the feature vector INLINEFORM6 can concentrate on those most representative features. DISPLAYFORM0 DISPLAYFORM1 ", + " INLINEFORM0 is the softmax result of INLINEFORM1 . The final sentence representation is decided by all of the relevant sentences' feature vector, as Equation EQREF24 shows. DISPLAYFORM0 ", + "Note that the set INLINEFORM0 is gradually growing as new sentence with the same drugs pairs is found when training. An instance INLINEFORM1 is represented by INLINEFORM2 before sentence level attention. The sentence level attention layer finds the set INLINEFORM3 , instances in which have the same drug pair as in INLINEFORM4 , and put INLINEFORM5 in INLINEFORM6 . Then the final sentence representation INLINEFORM7 is calculated in this layer." + ], + [ + "A given sentence INLINEFORM0 is finally represented by the feature vector INLINEFORM1 . Then we feed it to a softmax classifier. Let INLINEFORM2 denotes the set of all kinds of DDI. The output INLINEFORM3 is the probabilities of each class INLINEFORM4 belongs. DISPLAYFORM0 ", + "We use cross entropy cost function and INLINEFORM0 regularization as the optimization objective. For INLINEFORM1 -th instance, INLINEFORM2 denotes the one-hot representation of it's label, where the model outputs INLINEFORM3 . The cross entropy cost is: DISPLAYFORM0 ", + "For a mini-batch INLINEFORM0 , the optimization objective is: DISPLAYFORM0 ", + "All parameters in this model is: DISPLAYFORM0 ", + "We optimize the parameters of objective function INLINEFORM0 with Adam BIBREF24 , which is a variant of mini-batch stochastic gradient descent. During each train step, the gradient of INLINEFORM1 is calculated. Then INLINEFORM2 is adjusted according to the gradient. After the end of training, we have a model that is able to predict two drugs' interactions when a sentence about these drugs is given." + ], + [ + "The model is trained for DDI classification. The parameters in list INLINEFORM0 are tuned during the training process. Given a new sentence with two drugs, we can use this model to classify the DDI type.", + "The DDI prediction follows the procedure described in Section SECREF6 - SECREF26 . The given sentence is eventually represented by feature vector INLINEFORM0 . Then INLINEFORM1 is classified to a specific DDI type with a softmax classifier. In next section, we will evaluate our model's DDI prediction performance and see the advantages and shortcomings of our model." + ], + [ + "We use the DDI corpus of the 2013 DDIExtraction challenge BIBREF16 to train and test our model. The DDIs in this corpus are classified as five types. We give the definitions of these types and their example sentences, as shown in table TABREF4 . This standard dataset is made up of training set and testing set. We use the same metrics as in other drug-drug interaction extraction literature BIBREF11 , BIBREF10 , BIBREF25 , BIBREF9 , BIBREF8 , BIBREF12 : the overall precision, recall, and F1 score on testing set. INLINEFORM0 denotes the set of {False, Mechanism, Effect, Advise, Int}. The precision and recall of each INLINEFORM1 are calculated by DISPLAYFORM0 DISPLAYFORM1 ", + "Then the overall precision, recall, and F1 score are calculated by DISPLAYFORM0 ", + "Besides, we evaluate the captured feature vectors with t-SNE BIBREF26 , a visualizing and intuitive way to map a high dimensional vector into a 2 or 3-dimensional space. If the points in a low dimensional space are easy to be split, the feature vectors are believed to be more distinguishable." + ], + [ + "We use TensorFlow BIBREF27 r0.11 to implement the proposed model. The input of each word is an ordered triple (word, relative distance from drug1, relative distance from drug2). The sentence, which is represented as a matrix, is fed to the model. The output of the model is a INLINEFORM0 -dimensional vector representing the probabilities of being corresponding DDI. It is the network, parameters, and hyperparameters which decides the output vector. The network's parameters are adjusted during training, where the hyperparameters are tuned by hand. The hyperparameters after tuning are as follows. The word embedding's dimension INLINEFORM1 , the position embedding's dimension INLINEFORM2 , the hidden state's dimension INLINEFORM3 , the probability of dropout INLINEFORM4 , other hyperparameters which are not shown here are set to TensorFlow's default values.", + "The word embedding is initialized by pre-trained word vectors using GloVe BIBREF28 , while other parameters are initialized randomly. During each training step, a mini-batch (the mini-batch size INLINEFORM0 in our implementation) of sentences is selected from training set. The gradient of objective function is calculated for parameters updating (See Section SECREF26 ).", + "Figure FIGREF32 shows the training process. The objective function INLINEFORM0 is declining as the training mini-batches continuously sent to the model. As the testing mini-batches, the INLINEFORM1 function is fluctuating while its overall trend is descending. The instances in testing set are not participated in training so that INLINEFORM2 function is not descending so fast. However, training and testing instances have similar distribution in sample space, causing that testing instances' INLINEFORM3 tends to be smaller along with the training process. INLINEFORM4 has inverse relationship with the performance measurement. The F1 score is getting fluctuating around a specific value after enough training steps. The reason why fluctuating range is considerable is that only a tiny part of the whole training or testing set has been calculated the F1 score. Testing the whole set during every step is time consuming and not necessary. We will evaluate the model on the whole testing set in Section SECREF47 ." + ], + [ + "We save our model every 100 step and predict all the DDIs of the instances in the testing set. These predictions' F1 score is shown in figure FIGREF40 . To demonstrate the sentence level attention layer is effective, we drop this layer and then directly use INLINEFORM0 for softmax classification (See figure FIGREF15 ). The result is shown with \u201cRNN + dynamic word embedding + ATT\u201d curve, which illustrates that the sentence level attention layer contributes to a more accurate model.", + "Whether a dynamic or static word embedding is better for a DDI extraction task is under consideration. Nguyen et al. BIBREF21 shows that updating word embedding at the time of other parameters being trained makes a better performance in relation extraction task. We let the embedding be static when training, while other conditions are all the same. The \u201cRNN + static word embedding + 2ATT\u201d curve shows this case. We can draw a conclusion that updating the initialized word embedding trains more suitable word vectors for the task, which promotes the performance.", + "We compare our best F1 score with other state-of-the-art approaches in table TABREF39 , which shows our model has competitive advantage in dealing with drug-drug interaction extraction. The predictions confusion matrix is shown in table TABREF46 . The DDIs other than false being classified as false makes most of the classification error. It may perform better if a classifier which can tells true and false DDI apart is trained. We leave this two-stage classifier to our future work. Another phenomenon is that the \u201cInt\u201d type is often classified as \u201cEffect\u201d. The \u201cInt\u201d sentence describes there exists interaction between two drugs and this information implies the two drugs' combination will have good or bed effect. That's the reason why \u201cInt\u201d and \u201cEffect\u201d are often obfuscated.", + "To evaluate the features our model captured, we employ scikit-learn BIBREF29 's t-SNE class to map high dimensional feature vectors to 2-dimensional vectors, which can be depicted on a plane. We depict all the features of the instances in testing set, as shown in figure FIGREF41 . The RNN model using dynamic word embedding and 2 layers of attention is the most distinguishable one. Unfortunately, the classifier can not classify all the instances into correct classes. Comparing table TABREF46 with figure UID44 , both of which are from the best performed model, we can observe some conclusions. The \u201cInt\u201d DDIs are often misclassified as \u201cEffect\u201d, for the reason that some of the \u201cInt\u201d points are in the \u201cEffect\u201d cluster. The \u201cEffect\u201d points are too scattered so that plenty of \u201cEffect\u201d DDIs are classified to other types. The \u201cMechanism\u201d points are gathered around two clusters, causing that most of the \u201cmechanism\u201d DDIs are classified to two types: \u201cFalse\u201d and \u201cMechanism\u201d. In short, the visualizability of feature mapping gives better explanations for the prediction results and the quality of captured features." + ], + [ + "To conclude, we propose a recurrent neural network with multiple attention layers to extract DDIs from biomedical text. The sentence level attention layer, which combines other sentences containing the same drugs, has been added to our model. The experiments shows that our model outperforms the state-of-the-art DDI extraction systems. Task relevant word embedding and two attention layers improved the performance to some extent.", + "The imbalance of the classes and the ambiguity of semantics cause most of the misclassifications. We consider that instance generation using generative adversarial networks would cover the instance shortage in specific category. It is also reasonable to use distant supervision learning (which utilize other relevant material) for knowledge supplement and obtain a better performed DDI extraction system." + ], + [ + "This work is supported by the NSFC under Grant 61303191, 61303190, 61402504, 61103015." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1520/instruction.md b/qasper-1520/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..b66fb8484dd6da04969dc12dd9c552e5e1f6509e --- /dev/null +++ b/qasper-1520/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Compositional Neural Machine Translation by Removing the Lexicon from Syntax + +Question: How do they damage different neural modules? \ No newline at end of file diff --git a/qasper-1528/instruction.md b/qasper-1528/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..7ca975020ea25419ab81b11b422d0745a086642c --- /dev/null +++ b/qasper-1528/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: $\rho$-hot Lexicon Embedding-based Two-level LSTM for Sentiment Analysis + +Question: What is the agreement value for each dataset? \ No newline at end of file diff --git a/qasper-1535/instruction.md b/qasper-1535/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..5008cfa61f4b8b851617057ef0d71afbde9f1ea2 --- /dev/null +++ b/qasper-1535/instruction.md @@ -0,0 +1,101 @@ +Name of Paper: Unsupervised Pre-training for Natural Language Generation: A Literature Review + +Question: How strategy-based methods handle obstacles in NLG? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Background: Unsupervised Pre-training for NLU", + "Unsupervised Pre-training and Parameter Initialization for NLG", + "Architecture-based Methods ::: Inducing Task-Specific Architecture in Pre-training", + "Architecture-based Methods ::: Inducing Task-Specific Architecture in Pre-training ::: Denoising Autoencoder", + "Architecture-based Methods ::: Inducing Task-Specific Architecture in Pre-training ::: Conditional Masked Language Model", + "Architecture-based Methods ::: Inducing Task-Specific Architecture in Pre-training ::: Sequence to Sequence Language Model", + "Architecture-based Methods ::: Encoder-Agnostic Architectures for Adaptation", + "Strategy-based Methods ::: Fine-tuning Schedules for Adaption", + "Strategy-based Methods ::: Proxy Tasks for Adaption", + "Strategy-based Methods ::: Knowledge Distillation for Adaption", + "Discussions ::: The Relationship between Architecture- and Strategy-based Methods", + "Discussions ::: Experimental Phenomenons", + "Discussions ::: Future Directions", + "Conclusion" + ], + "paragraphs": [ + [ + "Unsupervised pre-training has sparked a sensational research interest in the natural language processing (NLP) community. This technology provides a promising way to exploit linguistic information from large-scale unlabelled textual data, which can serve as an auxiliary prior knowledge to benefit a wide range of NLP applications. In the literature, language modeling (LM) is a prevalent task for pre-training, where the target words are predicted conditioned on a given context. Therefore, it is intuitive to employ the pre-trained LMs for natural language generation, as the pre-training objective naturally accords with the goal of NLG. However, revolutionary improvements are only observed in the field of NLU.", + "The primary factor that impedes the progress of unsupervised pre-training in NLG is an idiosyncratic nature of text generation: Basically, we do not write words from scratch, but instead based on particular context, e.g., the source language sentences for translation, the dialog histories for response generation, and the visual scenes for image captioning, among others. In unsupervised pre-training, the task-specific context is not available, which leads to a discrepancy between pre-training and training in the target task. More precisely, the challenges posed by the discrepancy can be reflected in two aspects: First, the diverse context makes it intractable to design a universal representation extractor as in the case of NLU, and the pre-trained language generators may have to modify their inner structures to deal with the task-specific context. Second, the mismatch in data distribution and objective between the two training stages might result in the performance on the pre-training tasks being compromised during fine-tuning, which is dubbed as the catastrophic forgetting problem BIBREF0.", + "In response to the above challenges, two lines of work are proposed by resorting to architecture-based and strategy-based solutions, respectively. Architecture-based methods either try to induce task-specific architecture during pre-training (task-specific methods), or aim at building a general pre-training architecture to fit all downstream tasks (task-agnostic methods). Strategy-based methods depart from the pre-training stage, seeking to take advantage of the pre-trained models during the process of target task learning. The approaches include fine-tuning schedules that elaborately design the control of learning rates for optimization, proxy tasks that leverage labeled data to help the pre-trained model better fit the target data distribution, and knowledge distillation approaches that ditch the paradigm of initialization with pre-trained parameters by adopting the pre-trained model as a teacher network.", + "The remainder of this review is organized as follows: In Section SECREF2, we will introduce the background knowledge about unsupervised pre-training for NLU, followed by a sketch of how the pre-trained models are employed through parameter initialization for NLG in Section SECREF3. In Section SECREF4, we will describe the architecture-based methods, and the strategy-based methods are presented in Section SECREF5. Section SECREF6 provides some in-depth discussions, and Section SECREF7 concludes this review." + ], + [ + "Learning fine-grained language representations is a perennial topic in natural language understanding. In restrospect, compelling evidences suggest that good representations can be learned through unsupervised pre-training.", + "Early work focused on word-level representations BIBREF1, BIBREF2, which encodes each word independently. For sentence-level representations, there are roughly two kinds of pre-training objectives, namely discriminative pre-training and generative pre-training. Discriminative pre-training distinguishes context sentence(s) for a given sentence from non-context sentence(s) BIBREF3, BIBREF4, with the aim to capture inter-sentence relationships. Generative pre-training follows the language model paradigm:", + "where $x_{t}$ is the $t^{th}$ word in the textual sequence to generate, $T$ indicates sequence length, $\\theta $ stands for learnable parameters, and $C$ is the context information, which is defined by the pre-training objective. ELMo BIBREF5 and GPT (short for Generative Pre-training) BIBREF6 adopt uni-directional Transformer BIBREF7 and bi-directional LSTM BIBREF8 language models, respectively. In this case, the context is defined as $x_{1:t}$ or $x_{t+1:T}$. BERT BIBREF3 is trained with a novel masked language model (MLM), which is a non-autoregressive way of generation. Specifically, MLM randomly replaces a fixed proportion of tokens in each sentence with a special [MASK] token or a random token, which results in a corrupted sentence $X_{\\text{mask}}$, and predicts each replaced token based on the same context $X_{\\text{mask}}$. To alleviate the inconsistency with target tasks caused by the introduction of [MASK] token, XLNet BIBREF9 introduces permutation-based language model, which conducts autoregressive language modeling over all possible permutations of the original word sequence. This gives rise to a context $C=X_{\\mathbf {z}_{1:t-1}}$, where $\\mathbf {z}$ is a certain permutation of $[1,2, \\ldots , T]$, according to the definitions in BIBREF9. BIBREF10 and BIBREF11 pre-trained an encoder-decoder framework to reconstruct the input sentence and the surrounding sentence, respectively, and the encoded input sentence thereby is included in the context $C$.", + "The sentence representations learned by LMs can be used to perform many NLU tasks by adding a simple linear classifier. Despite the objective of language modeling, the pre-trained representations and have successfuly pushed the state-of-the-art on multiple benchmarks ." + ], + [ + "NLG systems are usually built with an encoder-decoder framework, where the encoder reads the context information and the decoder generates the target text from the encoded vectorial representations. A direct way to utilize the pre-trained models is to initialize part of the encoder (when dealing with textual context) and/or the decoder with pre-trained parameters. For the encoder, pre-training is expected to provide better sentence representations, as we discussed in Section SECREF2. For the decoder, the intuition is to endow the model with some rudimentary ability for text generation.", + "BIBREF12 employed BERT as the encoder for abstractive text summarization, with some additional techniques to help integrate the BERT-initialized encoder with the randomly initialized decoder, which we will explicate in Section SECREF12. GPT-2 BIBREF13 inherited the left-to-right LM pre-training objective from GPT and extended the application to NLG, where the pre-trained LM directly serves as the language generator, with some special symbols to identify task-specific contexts. In the case of zero-shot task transfer, preliminary experiments showed that straightforward adaption of GPT-2 compares unfavorably with other unsupervised baselines.", + "BIBREF14 is among the first attempts to investigate unsupervised pre-training for sequence to sequence (Seq2Seq) learning. They used pre-trained LSTM-based LMs to initialize the first layer of the encoder and the decoder, which act as representation extractors. An additional LSTM layer, which is randomly initialized, is then added on top of the pre-trained LMs to build the Seq2Seq framework. To make use of the text generation ability of LMs, the output softmax layer of the decoder LM is also retained. Some recent endeavours BIBREF15, BIBREF16 explored multiple combinations of GPT- and BERT-based models to initialize the encoder and the decoder, respectively. Although remarkable results are observed, the separately pre-trained LMs are still inconsistent with the Seq2Seq framework." + ], + [ + "Separately initializing the encoder and the decoder with LMs neglects the interaction between the two modules at the pre-training stage, which is sub-optimal. For NLG tasks that can be modeled as Seq2Seq learning, it is feasible to jointly pre-train the encoder and the decoder. Existing approaches for this sake can be categorized into three variants: Denoising autoencoders (DAEs), conditional masked language models (CMLMs) and sequence to sequence language models (Seq2Seq LMs)." + ], + [ + "An intuitive way to conduct unsupervised Seq2Seq learning is to train an autoencoder (AE) based on encoder-decoder framework. Different from AEs, DAEs take a corrupted sentence as input and reconstruct the original sentence. The advantage is that the corrupted input will force the decoder to extract relevant information from the source side for text generation. To obtain the corrupted sentence, BIBREF17 designed three noising functions: shuffle, delete and replace (the left plot of Figure FIGREF4 gives an illustration), each of which is controlled by a pre-defined probability distribution. To be more specific, each token in the raw sequence is assigned with a new index based on a gaussion distribution $N(0, \\sigma )$; the delete and replace operations of a token are determined by a Bernoulli distribution $B(p)$ with Beta distribution as prior. The three functions are applied to the raw sequences in random order." + ], + [ + "CMLM BIBREF18 extends the single model MLM proposed by BIBREF3 to the encoder-decoder setting, where the masked text sequence is read by the encoder, and the decoder only reconstructs the masked tokens, in construct to the entire sequence in DAEs. As the middle plot of Figure FIGREF4 shows, CMLM masks consecutive tokens , and the unmasked tokens in the encoder side are masked when being feed to the decoder. Following the notations in BIBREF18, let us assume that the tokens with index from $u$ to $v$ are masked from the raw sentence $X$, which results in $X^{\\backslash u: v}$, and $X^{u: v}$ denotes the decoder input. Then, when predicting each masked token $x_{t}$ ($u \\le t \\le v$), the context is $X^{u: v}_{