diff --git a/qasper-0117/instruction.md b/qasper-0117/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..aa3352a50c7010ae466ef7c5c13fbbd588aee7ef --- /dev/null +++ b/qasper-0117/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Diversity, Density, and Homogeneity: Quantitative Characteristic Metrics for Text Collections + +Question: Did they propose other metrics? \ No newline at end of file diff --git a/qasper-0118/instruction.md b/qasper-0118/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..41667e04676cc06e39fc0da797b3fd5bc71c1a80 --- /dev/null +++ b/qasper-0118/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Diversity, Density, and Homogeneity: Quantitative Characteristic Metrics for Text Collections + +Question: Which real-world datasets did they use? \ No newline at end of file diff --git a/qasper-0127/instruction.md b/qasper-0127/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..bce413d22d63e181b8fb39aa184b024ed9076cca --- /dev/null +++ b/qasper-0127/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A simple discriminative training method for machine translation with large-scale features + +Question: Is new method inferior in terms of robustness to MIRAs in experiments? \ No newline at end of file diff --git a/qasper-0128/instruction.md b/qasper-0128/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..63014ed19fe32a7fe4c8bcd6e8c9455e4e385de7 --- /dev/null +++ b/qasper-0128/instruction.md @@ -0,0 +1,72 @@ +Name of Paper: A simple discriminative training method for machine translation with large-scale features + +Question: What experiments with large-scale features are performed? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Plackett-Luce Model", + "Plackett-Luce Model in Statistical Machine Translation", + "Plackett-Luce Model in Statistical Machine Translation ::: N-best Hypotheses Resample", + "Evaluation", + "Evaluation ::: Plackett-Luce Model for SMT Tuning", + "Evaluation ::: Plackett-Luce Model for SMT Reranking" + ], + "paragraphs": [ + [ + "Since Och BIBREF0 proposed minimum error rate training (MERT) to exactly optimize objective evaluation measures, MERT has become a standard model tuning technique in statistical machine translation (SMT). Though MERT performs better by improving its searching algorithm BIBREF1, BIBREF2, BIBREF3, BIBREF4, it does not work reasonably when there are lots of features. As a result, margin infused relaxed algorithms (MIRA) dominate in this case BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10.", + "In SMT, MIRAs consider margin losses related to sentence-level BLEUs. However, since the BLEU is not decomposable into each sentence, these MIRA algorithms use some heuristics to compute the exact losses, e.g., pseudo-document BIBREF8, and document-level loss BIBREF9.", + "Recently, another successful work in large-scale feature tuning include force decoding basedBIBREF11, classification based BIBREF12.", + "We aim to provide a simpler tuning method for large-scale features than MIRAs. Out motivation derives from an observation on MERT. As MERT considers the quality of only top1 hypothesis set, there might have more-than-one set of parameters, which have similar top1 performances in tuning, but have very different topN hypotheses. Empirically, we expect an ideal model to benefit the total N-best list. That is, better hypotheses should be assigned with higher ranks, and this might decrease the error risk of top1 result on unseen data.", + "PlackettBIBREF13 offered an easy-to-understand theory of modeling a permutation. An N-best list is assumedly generated by sampling without replacement. The $i$th hypothesis to sample relies on those ranked after it, instead of on the whole list. This model also supports a partial permutation which accounts for top $k$ positions in a list, regardless of the remaining. When taking $k$ as 1, this model reduces to a standard conditional probabilistic training, whose dual problem is actual the maximum entropy based BIBREF14. Although Och BIBREF0 substituted direct error optimization for a maximum entropy based training, probabilistic models correlate with BLEU well when features are rich enough. The similar claim also appears in BIBREF15. This also make the new method be applicable in large-scale features." + ], + [ + "Plackett-Luce was firstly proposed to predict ranks of horses in gambling BIBREF13. Let $\\mathbf {r}=(r_{1},r_{2}\\ldots r_{N})$ be $N$ horses with a probability distribution $\\mathcal {P}$ on their abilities to win a game, and a rank $\\mathbf {\\pi }=(\\pi (1),\\pi (2)\\ldots \\pi (|\\mathbf {\\pi }|))$ of horses can be understood as a generative procedure, where $\\pi (j)$ denotes the index of the horse in the $j$th position.", + "In the 1st position, there are $N$ horses as candidates, each of which $r_{j}$ has a probability $p(r_{j})$ to be selected. Regarding the rank $\\pi $, the probability of generating the champion is $p(r_{\\pi (1)})$. Then the horse $r_{\\pi (1)}$ is removed from the candidate pool.", + "In the 2nd position, there are only $N-1$ horses, and their probabilities to be selected become $p(r_{j})/Z_{2}$, where $Z_{2}=1-p(r_{\\pi (1)})$ is the normalization. Then the runner-up in the rank $\\pi $, the $\\pi (2)$th horse, is chosen at the probability $p(r_{\\pi (2)})/Z_{2}$. We use a consistent terminology $Z_{1}$ in selecting the champion, though $Z_{1}$ equals 1 trivially.", + "This procedure iterates to the last rank in $\\pi $. The key idea for the Plackett-Luce model is the choice in the $i$th position in a rank $\\mathbf {\\pi }$ only depends on the candidates not chosen at previous stages. The probability of generating a rank $\\pi $ is given as follows", + "where $Z_{j}=1-\\sum _{t=1}^{j-1}p(r_{\\pi (t)})$.", + "We offer a toy example (Table TABREF3) to demonstrate this procedure.", + "Theorem 1 The permutation probabilities $p(\\mathbf {\\pi })$ form a probability distribution over a set of permutations $\\Omega _{\\pi }$. For example, for each $\\mathbf {\\pi }\\in \\Omega _{\\pi }$, we have $p(\\mathbf {\\pi })>0$, and $\\sum _{\\pi \\in \\Omega _{\\pi }}p(\\mathbf {\\pi })=1$.", + "We have to note that, $\\Omega _{\\pi }$ is not necessarily required to be completely ranked permutations in theory and in practice, since gamblers might be interested in only the champion and runner-up, and thus $|\\mathbf {\\pi }|\\le N$. In experiments, we would examine the effects on different length of permutations, systems being termed $PL(|\\pi |)$.", + "Theorem 2 Given any two permutations $\\mathbf {\\pi }$ and $\\mathbf {\\pi }\\prime $, and they are different only in two positions $p$ and $q$, $pp(\\pi (q))$, then $p(\\pi )>p(\\pi \\prime )$.", + "In other words, exchanging two positions in a permutation where the horse more likely to win is not ranked before the other would lead to an increase of the permutation probability.", + "This suggests the ground-truth permutation, ranked decreasingly by their probabilities, owns the maximum permutation probability on a given distribution. In SMT, we are motivated to optimize parameters to maximize the likelihood of ground-truth permutation of an N-best hypotheses.", + "Due to the limitation of space, see BIBREF13, BIBREF16 for the proofs of the theorems." + ], + [ + "In SMT, let $\\mathbf {f}=(f_{1},f_{2}\\ldots )$ denote source sentences, and $\\mathbf {e}=(\\lbrace e_{1,1},\\ldots \\rbrace ,\\lbrace e_{2,1},\\ldots \\rbrace \\ldots )$ denote target hypotheses. A set of features are defined on both source and target side. We refer to $h(e_{i,*})$ as a feature vector of a hypothesis from the $i$th source sentence, and its score from a ranking function is defined as the inner product $h(e_{i,*})^{T}w$ of the weight vector $w$ and the feature vector.", + "We first follow the popular exponential style to define a parameterized probability distribution over a list of hypotheses.", + "The ground-truth permutation of an $n$best list is simply obtained after ranking by their sentence-level BLEUs. Here we only concentrate on their relative ranks which are straightforward to compute in practice, e.g. add 1 smoothing. Let $\\pi _{i}^{*}$ be the ground-truth permutation of hypotheses from the $i$th source sentences, and our optimization objective is maximizing the log-likelihood of the ground-truth permutations and penalized using a zero-mean and unit-variance Gaussian prior. This results in the following objective and gradient:", + "where $Z_{i,j}$ is defined as the $Z_{j}$ in Formula (1) of the $i$th source sentence.", + "The log-likelihood function is smooth, differentiable, and concave with the weight vector $w$, and its local maximal solution is also a global maximum. Iteratively selecting one parameter in $\\alpha $ for tuning in a line search style (or MERT style) could also converge into the global global maximum BIBREF17. In practice, we use more fast limited-memory BFGS (L-BFGS) algorithm BIBREF18." + ], + [ + "The log-likelihood of a Plackett-Luce model is not a strict upper bound of the BLEU score, however, it correlates with BLEU well in the case of rich features. The concept of \u201crich\u201d is actually qualitative, and obscure to define in different applications. We empirically provide a formula to measure the richness in the scenario of machine translation.", + "The greater, the richer. In practice, we find a rough threshold of r is 5.", + "In engineering, the size of an N-best list with unique hypotheses is usually less than several thousands. This suggests that, if features are up to thousands or more, the Plackett-Luce model is quite suitable here. Otherwise, we could reduce the size of N-best lists by sampling to make $r$ beyond the threshold.", + "Their may be other efficient sampling methods, and here we adopt a simple one. If we want to $m$ samples from a list of hypotheses $\\mathbf {e}$, first, the $\\frac{m}{3}$ best hypotheses and the $\\frac{m}{3}$ worst hypotheses are taken by their sentence-level BLEUs. Second, we sample the remaining hypotheses on distribution $p(e_{i})\\propto \\exp (h(e_{i})^{T}w)$, where $\\mathbf {w}$ is an initial weight from last iteration." + ], + [ + "We compare our method with MERT and MIRA in two tasks, iterative training, and N-best list rerank. We do not list PRO BIBREF12 as our baseline, as Cherry et al.BIBREF10 have compared PRO with MIRA and MERT massively.", + "In the first task, we align the FBIS data (about 230K sentence pairs) with GIZA++, and train a 4-gram language model on the Xinhua portion of Gigaword corpus. A hierarchical phrase-based (HPB) model (Chiang, 2007) is tuned on NIST MT 2002, and tested on MT 2004 and 2005. All features are eight basic ones BIBREF20 and extra 220 group features. We design such feature templates to group grammars by the length of source side and target side, (feat-type,a$\\le $src-side$\\le $b,c$\\le $tgt-side$\\le $d), where the feat-type denotes any of the relative frequency, reversed relative frequency, lexical probability and reversed lexical probability, and [a, b], [c, d] enumerate all possible subranges of [1, 10], as the maximum length on both sides of a hierarchical grammar is limited to 10. There are 4 $\\times $ 55 extra group features.", + "In the second task, we rerank an N-best list from a HPB system with 7491 features from a third party. The system uses six million parallel sentence pairs available to the DARPA BOLT Chinese-English task. This system includes 51 dense features (translation probabilities, provenance features, etc.) and up to 7440 sparse features (mostly lexical and fertility-based). The language model is a 6-gram model trained on a 10 billion words, including the English side of our parallel corpora plus other corpora such as Gigaword (LDC2011T07) and Google News. For the tuning and test sets, we use 1275 and 1239 sentences respectively from the LDC2010E30 corpus." + ], + [ + "We conduct a full training of machine translation models. By default, a decoder is invoked for at most 40 times, and each time it outputs 200 hypotheses to be combined with those from previous iterations and sent into tuning algorithms.", + "In getting the ground-truth permutations, there are many ties with the same sentence-level BLEU, and we just take one randomly. In this section, all systems have only around two hundred features, hence in Plackett-Luce based training, we sample 30 hypotheses in an accumulative $n$best list in each round of training.", + "All results are shown in Table TABREF10, we can see that all PL($k$) systems does not perform well as MERT or MIRA in the development data, this maybe due to that PL($k$) systems do not optimize BLEU and the features here are relatively not enough compared to the size of N-best lists (empirical Formula DISPLAY_FORM9). However, PL($k$) systems are better than MERT in testing. PL($k$) systems consider the quality of hypotheses from the 2th to the $k$th, which is guessed to act the role of the margin like SVM in classification . Interestingly, MIRA wins first in training, and still performs quite well in testing.", + "The PL(1) system is equivalent to a max-entropy based algorithm BIBREF14 whose dual problem is actually maximizing the conditional probability of one oracle hypothesis. When we increase the $k$, the performances improve at first. After reaching a maximum around $k=5$, they decrease slowly. We explain this phenomenon as this, when features are rich enough, higher BLEU scores could be easily fitted, then longer ground-truth permutations include more useful information." + ], + [ + "After being de-duplicated, the N-best list has an average size of around 300, and with 7491 features. Refer to Formula DISPLAY_FORM9, this is ideal to use the Plackett-Luce model. Results are shown in Figure FIGREF12. We observe some interesting phenomena.", + "First, the Plackett-Luce models boost the training BLEU very greatly, even up to 2.5 points higher than MIRA. This verifies our assumption, richer features benefit BLEU, though they are optimized towards a different objective.", + "Second, the over-fitting problem of the Plackett-Luce models PL($k$) is alleviated with moderately large $k$. In PL(1), the over-fitting is quite obvious, the portion in which the curve overpasses MIRA is the smallest compared to other $k$, and its convergent performance is below the baseline. When $k$ is not smaller than 5, the curves are almost above the MIRA line. After 500 L-BFGS iterations, their performances are no less than the baseline, though only by a small margin.", + "This experiment displays, in large-scale features, the Plackett-Luce model correlates with BLEU score very well, and alleviates overfitting in some degree." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0173/instruction.md b/qasper-0173/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0639e2cfd24f7a3470cb1ceac0c3cb131ebac34b --- /dev/null +++ b/qasper-0173/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Towards Multimodal Emotion Recognition in German Speech Events in Cars using Transfer Learning + +Question: How is face and audio data analysis evaluated? \ No newline at end of file diff --git a/qasper-0174/instruction.md b/qasper-0174/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..2c807ae3ed6ea6f19c3f000f07b892c71b3d6d7b --- /dev/null +++ b/qasper-0174/instruction.md @@ -0,0 +1,112 @@ +Name of Paper: Towards Multimodal Emotion Recognition in German Speech Events in Cars using Transfer Learning + +Question: What is the baseline method for the task? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work ::: Facial Expressions", + "Related Work ::: Acoustic", + "Related Work ::: Text", + "Data set Collection", + "Data set Collection ::: Study Setup and Design", + "Data set Collection ::: Procedure", + "Data set Collection ::: Data Analysis", + "Methods ::: Emotion Recognition from Facial Expressions", + "Methods ::: Emotion Recognition from Audio Signal", + "Methods ::: Emotion Recognition from Transcribed Utterances", + "Results ::: Facial Expressions and Audio", + "Results ::: Text from Transcribed Utterances", + "Results ::: Text from Transcribed Utterances ::: Experiment 1: In-Domain application", + "Results ::: Text from Transcribed Utterances ::: Experiment 2: Simple Out-Of-Domain application", + "Results ::: Text from Transcribed Utterances ::: Experiment 3: Transfer Learning application", + "Summary & Future Work", + "Acknowledgment" + ], + "paragraphs": [ + [ + "Automatic emotion recognition is commonly understood as the task of assigning an emotion to a predefined instance, for example an utterance (as audio signal), an image (for instance with a depicted face), or a textual unit (e.g., a transcribed utterance, a sentence, or a Tweet). The set of emotions is often following the original definition by Ekman Ekman1992, which includes anger, fear, disgust, sadness, joy, and surprise, or the extension by Plutchik Plutchik1980 who adds trust and anticipation.", + "Most work in emotion detection is limited to one modality. Exceptions include Busso2004 and Sebe2005, who investigate multimodal approaches combining speech with facial information. Emotion recognition in speech can utilize semantic features as well BIBREF0. Note that the term \u201cmultimodal\u201d is also used beyond the combination of vision, audio, and text. For example, Soleymani2012 use it to refer to the combination of electroencephalogram, pupillary response and gaze distance.", + "In this paper, we deal with the specific situation of car environments as a testbed for multimodal emotion recognition. This is an interesting environment since it is, to some degree, a controlled environment: Dialogue partners are limited in movement, the degrees of freedom for occurring events are limited, and several sensors which are useful for emotion recognition are already integrated in this setting. More specifically, we focus on emotion recognition from speech events in a dialogue with a human partner and with an intelligent agent.", + "Also from the application point of view, the domain is a relevant choice: Past research has shown that emotional intelligence is beneficial for human computer interaction. Properly processing emotions in interactions increases the engagement of users and can improve performance when a specific task is to be fulfilled BIBREF1, BIBREF2, BIBREF3, BIBREF4. This is mostly based on the aspect that machines communicating with humans appear to be more trustworthy when they show empathy and are perceived as being natural BIBREF3, BIBREF5, BIBREF4.", + "Virtual agents play an increasingly important role in the automotive context and the speech modality is increasingly being used in cars due to its potential to limit distraction. It has been shown that adapting the in-car speech interaction system according to the drivers' emotional state can help to enhance security, performance as well as the overall driving experience BIBREF6, BIBREF7.", + "With this paper, we investigate how each of the three considered modalitites, namely facial expressions, utterances of a driver as an audio signal, and transcribed text contributes to the task of emotion recognition in in-car speech interactions. We focus on the five emotions of joy, insecurity, annoyance, relaxation, and boredom since terms corresponding to so-called fundamental emotions like fear have been shown to be associated to too strong emotional states than being appropriate for the in-car context BIBREF8. Our first contribution is the description of the experimental setup for our data collection. Aiming to provoke specific emotions with situations which can occur in real-world driving scenarios and to induce speech interactions, the study was conducted in a driving simulator. Based on the collected data, we provide baseline predictions with off-the-shelf tools for face and speech emotion recognition and compare them to a neural network-based approach for emotion recognition from text. Our second contribution is the introduction of transfer learning to adapt models trained on established out-of-domain corpora to our use case. We work on German language, therefore the transfer consists of a domain and a language transfer." + ], + [ + "A common approach to encode emotions for facial expressions is the facial action coding system FACS BIBREF9, BIBREF10, BIBREF11. As the reliability and reproducability of findings with this method have been critically discussed BIBREF12, the trend has increasingly shifted to perform the recognition directly on images and videos, especially with deep learning. For instance, jung2015joint developed a model which considers temporal geometry features and temporal appearance features from image sequences. kim2016hierarchical propose an ensemble of convolutional neural networks which outperforms isolated networks.", + "In the automotive domain, FACS is still popular. Ma2017 use support vector machines to distinguish happy, bothered, confused, and concentrated based on data from a natural driving environment. They found that bothered and confused are difficult to distinguish, while happy and concentrated are well identified. Aiming to reduce computational cost, Tews2011 apply a simple feature extraction using four dots in the face defining three facial areas. They analyze the variance of the three facial areas for the recognition of happy, anger and neutral. Ihme2018 aim at detecting frustration in a simulator environment. They induce the emotion with specific scenarios and a demanding secondary task and are able to associate specific face movements according to FACS. Paschero2012 use OpenCV (https://opencv.org/) to detect the eyes and the mouth region and track facial movements. They simulate different lightning conditions and apply a multilayer perceptron for the classification task of Ekman's set of fundamental emotions.", + "Overall, we found that studies using facial features usually focus on continuous driver monitoring, often in driver-only scenarios. In contrast, our work investigates the potential of emotion recognition during speech interactions." + ], + [ + "Past research on emotion recognition from acoustics mainly concentrates on either feature selection or the development of appropriate classifiers. rao2013emotion as well as ververidis2004automatic compare local and global features in support vector machines. Next to such discriminative approaches, hidden Markov models are well-studied, however, there is no agreement on which feature-based classifier is most suitable BIBREF13. Similar to the facial expression modality, recent efforts on applying deep learning have been increased for acoustic speech processing. For instance, lee2015high use a recurrent neural network and palaz2015analysis apply a convolutional neural network to the raw speech signal. Neumann2017 as well as Trigeorgis2016 analyze the importance of features in the context of deep learning-based emotion recognition.", + "In the automotive sector, Boril2011 approach the detection of negative emotional states within interactions between driver and co-driver as well as in calls of the driver towards the automated spoken dialogue system. Using real-world driving data, they find that the combination of acoustic features and their respective Gaussian mixture model scores performs best. Schuller2006 collects 2,000 dialog turns directed towards an automotive user interface and investigate the classification of anger, confusion, and neutral. They show that automatic feature generation and feature selection boost the performance of an SVM-based classifier. Further, they analyze the performance under systematically added noise and develop methods to mitigate negative effects. For more details, we refer the reader to the survey by Schuller2018. In this work, we explore the straight-forward application of domain independent software to an in-car scenario without domain-specific adaptations." + ], + [ + "Previous work on emotion analysis in natural language processing focuses either on resource creation or on emotion classification for a specific task and domain. On the side of resource creation, the early and influential work of Pennebaker2015 is a dictionary of words being associated with different psychologically relevant categories, including a subset of emotions. Another popular resource is the NRC dictionary by Mohammad2012b. It contains more than 10000 words for a set of discrete emotion classes. Other resources include WordNet Affect BIBREF14 which distinguishes particular word classes. Further, annotated corpora have been created for a set of different domains, for instance fairy tales BIBREF15, Blogs BIBREF16, Twitter BIBREF17, BIBREF18, BIBREF19, BIBREF20, BIBREF21, Facebook BIBREF22, news headlines BIBREF23, dialogues BIBREF24, literature BIBREF25, or self reports on emotion events BIBREF26 (see BIBREF27 for an overview).", + "To automatically assign emotions to textual units, the application of dictionaries has been a popular approach and still is, particularly in domains without annotated corpora. Another approach to overcome the lack of huge amounts of annotated training data in a particular domain or for a specific topic is to exploit distant supervision: use the signal of occurrences of emoticons or specific hashtags or words to automatically label the data. This is sometimes referred to as self-labeling BIBREF21, BIBREF28, BIBREF29, BIBREF30.", + "A variety of classification approaches have been tested, including SNoW BIBREF15, support vector machines BIBREF16, maximum entropy classification, long short-term memory network, and convolutional neural network models BIBREF18. More recently, the state of the art is the use of transfer learning from noisy annotations to more specific predictions BIBREF29. Still, it has been shown that transferring from one domain to another is challenging, as the way emotions are expressed varies between areas BIBREF27. The approach by Felbo2017 is different to our work as they use a huge noisy data set for pretraining the model while we use small high quality data sets instead.", + "Recently, the state of the art has also been pushed forward with a set of shared tasks, in which the participants with top results mostly exploit deep learning methods for prediction based on pretrained structures like embeddings or language models BIBREF21, BIBREF31, BIBREF20.", + "Our work follows this approach and builds up on embeddings with deep learning. Furthermore, we approach the application and adaption of text-based classifiers to the automotive domain with transfer learning." + ], + [ + "The first contribution of this paper is the construction of the AMMER data set which we describe in the following. We focus on the drivers' interactions with both a virtual agent as well as a co-driver. To collect the data in a safe and controlled environment and to be able to consider a variety of predefined driving situations, the study was conducted in a driving simulator." + ], + [ + "The study environment consists of a fixed-base driving simulator running Vires's VTD (Virtual Test Drive, v2.2.0) simulation software (https://vires.com/vtd-vires-virtual-test-drive/). The vehicle has an automatic transmission, a steering wheel and gas and brake pedals. We collect data from video, speech and biosignals (Empatica E4 to record heart rate, electrodermal activity, skin temperature, not further used in this paper) and questionnaires. Two RGB cameras are fixed in the vehicle to capture the drivers face, one at the sun shield above the drivers seat and one in the middle of the dashboard. A microphone is placed on the center console. One experimenter sits next to the driver, the other behind the simulator. The virtual agent accompanying the drive is realized as Wizard-of-Oz prototype which enables the experimenter to manually trigger prerecorded voice samples playing trough the in-car speakers and to bring new content to the center screen. Figure FIGREF4 shows the driving simulator.", + "The experimental setting is comparable to an everyday driving task. Participants are told that the goal of the study is to evaluate and to improve an intelligent driving assistant. To increase the probability of emotions to arise, participants are instructed to reach the destination of the route as fast as possible while following traffic rules and speed limits. They are informed that the time needed for the task would be compared to other participants. The route comprises highways, rural roads, and city streets. A navigation system with voice commands and information on the screen keeps the participants on the predefined track.", + "To trigger emotion changes in the participant, we use the following events: (i) a car on the right lane cutting off to the left lane when participants try to overtake followed by trucks blocking both lanes with a slow overtaking maneuver (ii) a skateboarder who appears unexpectedly on the street and (iii) participants are praised for reaching the destination unexpectedly quickly in comparison to previous participants.", + "Based on these events, we trigger three interactions (Table TABREF6 provides examples) with the intelligent agent (Driver-Agent Interactions, D\u2013A). Pretending to be aware of the current situation, e. g., to recognize unusual driving behavior such as strong braking, the agent asks the driver to explain his subjective perception of these events in detail. Additionally, we trigger two more interactions with the intelligent agent at the beginning and at the end of the drive, where participants are asked to describe their mood and thoughts regarding the (upcoming) drive. This results in five interactions between the driver and the virtual agent.", + "Furthermore, the co-driver asks three different questions during sessions with light traffic and low cognitive demand (Driver-Co-Driver Interactions, D\u2013Co). These questions are more general and non-traffic-related and aim at triggering the participants' memory and fantasy. Participants are asked to describe their last vacation, their dream house and their idea of the perfect job. In sum, there are eight interactions per participant (5 D\u2013A, 3 D\u2013Co)." + ], + [ + "At the beginning of the study, participants were welcomed and the upcoming study procedure was explained. Subsequently, participants signed a consent form and completed a questionnaire to provide demographic information. After that, the co-driving experimenter started with the instruction in the simulator which was followed by a familiarization drive consisting of highway and city driving and covering different driving maneuvers such as tight corners, lane changing and strong braking. Subsequently, participants started with the main driving task. The drive had a duration of 20 minutes containing the eight previously mentioned speech interactions. After the completion of the drive, the actual goal of improving automatic emotional recognition was revealed and a standard emotional intelligence questionnaire, namely the TEIQue-SF BIBREF32, was handed to the participants. Finally, a retrospective interview was conducted, in which participants were played recordings of their in-car interactions and asked to give discrete (annoyance, insecurity, joy, relaxation, boredom, none, following BIBREF8) was well as dimensional (valence, arousal, dominance BIBREF33 on a 11-point scale) emotion ratings for the interactions and the according situations. We only use the discrete class annotations in this paper." + ], + [ + "Overall, 36 participants aged 18 to 64 years ($\\mu $=28.89, $\\sigma $=12.58) completed the experiment. This leads to 288 interactions, 180 between driver and the agent and 108 between driver and co-driver. The emotion self-ratings from the participants yielded 90 utterances labeled with joy, 26 with annoyance, 49 with insecurity, 9 with boredom, 111 with relaxation and 3 with no emotion. One example interaction per interaction type and emotion is shown in Table TABREF7. For further experiments, we only use joy, annoyance/anger, and insecurity/fear due to the small sample size for boredom and no emotion and under the assumption that relaxation brings little expressivity." + ], + [ + "We preprocess the visual data by extracting the sequence of images for each interaction from the point where the agent's or the co-driver's question was completely uttered until the driver's response stops. The average length is 16.3 seconds, with the minimum at 2.2s and the maximum at 54.7s. We apply an off-the-shelf tool for emotion recognition (the manufacturer cannot be disclosed due to licensing restrictions). It delivers frame-by-frame scores ($\\in [0;100]$) for discrete emotional states of joy, anger and fear. While joy corresponds directly to our annotation, we map anger to our label annoyance and fear to our label insecurity. The maximal average score across all frames constitutes the overall classification for the video sequence. Frames where the software is not able to detect the face are ignored." + ], + [ + "We extract the audio signal for the same sequence as described for facial expressions and apply an off-the-shelf tool for emotion recognition. The software delivers single classification scores for a set of 24 discrete emotions for the entire utterance. We consider the outputs for the states of joy, anger, and fear, mapping analogously to our classes as for facial expressions. Low-confidence predictions are interpreted as \u201cno emotion\u201d. We accept the emotion with the highest score as the discrete prediction otherwise." + ], + [ + "For the emotion recognition from text, we manually transcribe all utterances of our AMMER study. To exploit existing and available data sets which are larger than the AMMER data set, we develop a transfer learning approach. We use a neural network with an embedding layer (frozen weights, pre-trained on Common Crawl and Wikipedia BIBREF36), a bidirectional LSTM BIBREF37, and two dense layers followed by a soft max output layer. This setup is inspired by BIBREF38. We use a dropout rate of 0.3 in all layers and optimize with Adam BIBREF39 with a learning rate of $10^{-5}$ (These parameters are the same for all further experiments). We build on top of the Keras library with the TensorFlow backend. We consider this setup our baseline model.", + "We train models on a variety of corpora, namely the common format published by BIBREF27 of the FigureEight (formally known as Crowdflower) data set of social media, the ISEAR data BIBREF40 (self-reported emotional events), and, the Twitter Emotion Corpus (TEC, weakly annotated Tweets with #anger, #disgust, #fear, #happy, #sadness, and #surprise, Mohammad2012). From all corpora, we use instances with labels fear, anger, or joy. These corpora are English, however, we do predictions on German utterances. Therefore, each corpus is preprocessed to German with Google Translate. We remove URLs, user tags (\u201c@Username\u201d), punctuation and hash signs. The distributions of the data sets are shown in Table TABREF12.", + "To adapt models trained on these data, we apply transfer learning as follows: The model is first trained until convergence on one out-of-domain corpus (only on classes fear, joy, anger for compatibility reasons). Then, the parameters of the bi-LSTM layer are frozen and the remaining layers are further trained on AMMER. This procedure is illustrated in Figure FIGREF13" + ], + [ + "Table TABREF16 shows the confusion matrices for facial and audio emotion recognition on our complete AMMER data set and Table TABREF17 shows the results per class for each method, including facial and audio data and micro and macro averages. The classification from facial expressions yields a macro-averaged $\\text{F}_1$ score of 33 % across the three emotions joy, insecurity, and annoyance (P=0.31, R=0.35). While the classification results for joy are promising (R=43 %, P=57 %), the distinction of insecurity and annoyance from the other classes appears to be more challenging.", + "Regarding the audio signal, we observe a macro $\\text{F}_1$ score of 29 % (P=42 %, R=22 %). There is a bias towards negative emotions, which results in a small number of detected joy predictions (R=4 %). Insecurity and annoyance are frequently confused." + ], + [ + "The experimental setting for the evaluation of emotion recognition from text is as follows: We evaluate the BiLSTM model in three different experiments: (1) in-domain, (2) out-of-domain and (3) transfer learning. For all experiments we train on the classes anger/annoyance, fear/insecurity and joy. Table TABREF19 shows all results for the comparison of these experimental settings." + ], + [ + "We first set a baseline by validating our models on established corpora. We train the baseline model on 60 % of each data set listed in Table TABREF12 and evaluate that model with 40 % of the data from the same domain (results shown in the column \u201cIn-Domain\u201d in Table TABREF19). Excluding AMMER, we achieve an average micro $\\text{F}_1$ of 68 %, with best results of F$_1$=73 % on TEC. The model trained on our AMMER corpus achieves an F1 score of 57%. This is most probably due to the small size of this data set and the class bias towards joy, which makes up more than half of the data set. These results are mostly in line with Bostan2018." + ], + [ + "Now we analyze how well the models trained in Experiment 1 perform when applied to our data set. The results are shown in column \u201cSimple\u201d in Table TABREF19. We observe a clear drop in performance, with an average of F$_1$=48 %. The best performing model is again the one trained on TEC, en par with the one trained on the Figure8 data. The model trained on ISEAR performs second best in Experiment 1, it performs worst in Experiment 2." + ], + [ + "To adapt models trained on previously existing data sets to our particular application, the AMMER corpus, we apply transfer learning. Here, we perform leave-one-out cross validation. As pre-trained models we use each model from Experiment 1 and further optimize with the training subset of each crossvalidation iteration of AMMER. The results are shown in the column \u201cTransfer L.\u201d in Table TABREF19. The confusion matrix is also depicted in Table TABREF16.", + "With this procedure we achieve an average performance of F$_1$=75 %, being better than the results from the in-domain Experiment 1. The best performance of F$_1$=76 % is achieved with the model pre-trained on each data set, except for ISEAR. All transfer learning models clearly outperform their simple out-of-domain counterpart.", + "To ensure that this performance increase is not only due to the larger data set, we compare these results to training the model without transfer on a corpus consisting of each corpus together with AMMER (again, in leave-one-out crossvalidation). These results are depicted in column \u201cJoint C.\u201d. Thus, both settings, \u201ctransfer learning\u201d and \u201cjoint corpus\u201d have access to the same information.", + "The results show an increase in performance in contrast to not using AMMER for training, however, the transfer approach based on partial retraining the model shows a clear improvement for all models (by 7pp for Figure8, 10pp for EmoInt, 8pp for TEC, 13pp for ISEAR) compared to the \u201dJoint\u201d setup." + ], + [ + "We described the creation of the multimodal AMMER data with emotional speech interactions between a driver and both a virtual agent and a co-driver. We analyzed the modalities of facial expressions, acoustics, and transcribed utterances regarding their potential for emotion recognition during in-car speech interactions. We applied off-the-shelf emotion recognition tools for facial expressions and acoustics. For transcribed text, we developed a neural network-based classifier with transfer learning exploiting existing annotated corpora. We find that analyzing transcribed utterances is most promising for classification of the three emotional states of joy, annoyance and insecurity.", + "Our results for facial expressions indicate that there is potential for the classification of joy, however, the states of annoyance and insecurity are not well recognized. Future work needs to investigate more sophisticated approaches to map frame predictions to sequence predictions. Furthermore, movements of the mouth region during speech interactions might negatively influence the classification from facial expressions. Therefore, the question remains how facial expressions can best contribute to multimodal detection in speech interactions.", + "Regarding the classification from the acoustic signal, the application of off-the-shelf classifiers without further adjustments seems to be challenging. We find a strong bias towards negative emotional states for our experimental setting. For instance, the personalization of the recognition algorithm (e. g., mean and standard deviation normalization) could help to adapt the classification for specific speakers and thus to reduce this bias. Further, the acoustic environment in the vehicle interior has special properties and the recognition software might need further adaptations.", + "Our transfer learning-based text classifier shows considerably better results. This is a substantial result in its own, as only one previous method for transfer learning in emotion recognition has been proposed, in which a sentiment/emotion specific source for labels in pre-training has been used, to the best of our knowledge BIBREF29. Other applications of transfer learning from general language models include BIBREF41, BIBREF42. Our approach is substantially different, not being trained on a huge amount of noisy data, but on smaller out-of-domain sets of higher quality. This result suggests that emotion classification systems which work across domains can be developed with reasonable effort.", + "For a productive application of emotion detection in the context of speech events we conclude that a deployed system might perform best with a speech-to-text module followed by an analysis of the text. Further, in this work, we did not explore an ensemble model or the interaction of different modalities. Thus, future work should investigate the fusion of multiple modalities in a single classifier." + ], + [ + "We thank Laura-Ana-Maria Bostan for discussions and data set preparations. This research has partially been funded by the German Research Council (DFG), project SEAT (KL 2869/1-1)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0187/instruction.md b/qasper-0187/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..f5b8dcb8e140c284d71b7803cd758b0d93035cdf --- /dev/null +++ b/qasper-0187/instruction.md @@ -0,0 +1,107 @@ +Name of Paper: "Hinglish"Language -- Modeling a Messy Code-Mixed Language + +Question: How is the dataset collected? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Introduction ::: Modeling challenges", + "Related Work ::: Transfer learning based approaches", + "Related Work ::: Hybrid models", + "Dataset and Features", + "Dataset and Features ::: Challenges", + "Model Architecture", + "Model Architecture ::: Loss function", + "Model Architecture ::: Models", + "Model Architecture ::: Hyper parameters", + "Results", + "Conclusion and Future work", + "References" + ], + "paragraphs": [ + [ + "Hinglish is a linguistic blend of Hindi (very widely spoken language in India) and English (an associate language of urban areas) and is spoken by upwards of 350 million people in India. While the name is based on the Hindi language, it does not refer exclusively to Hindi, but is used in India, with English words blending with Punjabi, Gujarati, Marathi and Hindi. Sometimes, though rarely, Hinglish is used to refer to Hindi written in English script and mixing with English words or phrases. This makes analyzing the language very interesting. Its rampant usage in social media like Twitter, Facebook, Online blogs and reviews has also led to its usage in delivering hate and abuses in similar platforms. We aim to find such content in the social media focusing on the tweets. Hypothetically, if we can classify such tweets, we might be able to detect them and isolate them for further analysis before it reaches public. This will a great application of AI to the social cause and thus is motivating. An example of a simple, non offensive message written in Hinglish could be:", + "\"Why do you waste your time with . Aapna ghar sambhalta nahi(). Chale dusro ko basane..!!\"", + "The second part of the above sentence is written in Hindi while the first part is in English. Second part calls for an action to a person to bring order to his/her home before trying to settle others." + ], + [ + "From the modeling perspective there are couple of challenges introduced by the language and the labelled dataset. Generally, Hinglish follows largely fuzzy set of rules which evolves and is dependent upon the users preference. It doesn't have any formal definitions and thus the rules of usage are ambiguous. Thus, when used by different users the text produced may differ. Overall the challenges posed by this problem are:", + "Geographical variation: Depending upon the geography of origination, the content may be be highly influenced by the underlying region.", + "Language and phonetics variation: Based on a census in 2001, India has 122 major languages and 1599 other languages. The use of Hindi and English in a code switched setting is highly influenced by these language.", + "No grammar rules: Hinglish has no fixed set of grammar rules. The rules are inspired from both Hindi and English and when mixed with slur and slang produce large variation.", + "Spelling variation: There is no agreement on the spellings of the words which are mixed with English. For example to express love, a code mixed spelling, specially when used social platforms might be pyaar, pyar or pyr.", + "Dataset: Based on some earlier work, only available labelled dataset had 3189 rows of text messages of average length of 116 words and with a range of 1, 1295. Prior work addresses this concern by using Transfer Learning on an architecture learnt on about 14,500 messages with an accuracy of 83.90. We addressed this concern using data augmentation techniques applied on text data." + ], + [ + "Mathur et al. in their paper for detecting offensive tweets proposed a Ternary Trans-CNN model where they train a model architecture comprising of 3 layers of Convolution 1D having filter sizes of 15, 12 and 10 and kernel size of 3 followed by 2 dense fully connected layer of size 64 and 3. The first dense FC layer has ReLU activation while the last Dense layer had Softmax activation. They were able to train this network on a parallel English dataset provided by Davidson et al. The authors were able to achieve Accuracy of 83.9%, Precision of 80.2%, Recall of 69.8%.", + "The approach looked promising given that the dataset was merely 3189 sentences divided into three categories and thus we replicated the experiment but failed to replicate the results. The results were poor than what the original authors achieved. But, most of the model hyper-parameter choices where inspired from this work." + ], + [ + "In another localized setting of Vietnamese language, Nguyen et al. in 2017 proposed a Hybrid multi-channel CNN and LSTM model where they build feature maps for Vietnamese language using CNN to capture shorterm dependencies and LSTM to capture long term dependencies and concatenate both these feature sets to learn a unified set of features on the messages. These concatenated feature vectors are then sent to a few fully connected layers. They achieved an accuracy rate of 87.3% with this architecture." + ], + [ + "We used dataset, HEOT obtained from one of the past studies done by Mathur et al. where they annotated a set of cleaned tweets obtained from twitter for the conversations happening in Indian subcontinent. A labelled dataset for a corresponding english tweets were also obtained from a study conducted by Davidson et al. This dataset was important to employ Transfer Learning to our task since the number of labeled dataset was very small. Basic summary and examples of the data from the dataset are below:" + ], + [ + "The obtained data set had many challenges and thus a data preparation task was employed to clean the data and make it ready for the deep learning pipeline. The challenges and processes that were applied are stated below:", + "Messy text messages: The tweets had urls, punctuations, username mentions, hastags, emoticons, numbers and lots of special characters. These were all cleaned up in a preprocessing cycle to clean the data.", + "Stop words: Stop words corpus obtained from NLTK was used to eliminate most unproductive words which provide little information about individual tweets.", + "Transliteration: Followed by above two processes, we translated Hinglish tweets into English words using a two phase process", + "Transliteration: In phase I, we used translation API's provided by Google translation services and exposed via a SDK, to transliteration the Hinglish messages to English messages.", + "Translation: After transliteration, words that were specific to Hinglish were translated to English using an Hinglish-English dictionary. By doing this we converted the Hinglish message to and assortment of isolated words being presented in the message in a sequence that can also be represented using word to vector representation.", + "Data augmentation: Given the data set was very small with a high degree of imbalance in the labelled messages for three different classes, we employed a data augmentation technique to boost the learning of the deep network. Following techniques from the paper by Jason et al. was utilized in this setting that really helped during the training phase.Thsi techniques wasnt used in previous studies. The techniques were:", + "Synonym Replacement (SR):Randomly choose n words from the sentence that are not stop words. Replace each of these words with one of its synonyms chosen at random.", + "Random Insertion (RI):Find a random synonym of a random word in the sentence that is not a stop word. Insert that synonym into a random position in the sentence. Do this n times.", + "Random Swap (RS):Randomly choose two words in the sentence and swap their positions. Do this n times.", + "Random Deletion (RD):For each word in the sentence, randomly remove it with probability p.", + "Word Representation: We used word embedding representations by Glove for creating word embedding layers and to obtain the word sequence vector representations of the processed tweets. The pre-trained embedding dimension were one of the hyperparamaters for model. Further more, we introduced another bit flag hyperparameter that determined if to freeze these learnt embedding.", + "Train-test split: The labelled dataset that was available for this task was very limited in number of examples and thus as noted above few data augmentation techniques were applied to boost the learning of the network. Before applying augmentation, a train-test split of 78%-22% was done from the original, cleansed data set. Thus, 700 tweets/messages were held out for testing. All model evaluation were done in on the test set that got generated by this process. The results presented in this report are based on the performance of the model on the test set. The training set of 2489 messages were however sent to an offline pipeline for augmenting the data. The resulting training dataset was thus 7934 messages. the final distribution of messages for training and test was thus below:" + ], + [ + "We tested the performance of various model architectures by running our experiment over 100 times on a CPU based compute which later as migrated to GPU based compute to overcome the slow learning progress. Our universal metric for minimizing was the validation loss and we employed various operational techniques for optimizing on the learning process. These processes and its implementation details will be discussed later but they were learning rate decay, early stopping, model checkpointing and reducing learning rate on plateau." + ], + [ + "For the loss function we chose categorical cross entropy loss in finding the most optimal weights/parameters of the model. Formally this loss function for the model is defined as below:", + "The double sum is over the number of observations and the categories respectively. While the model probability is the probability that the observation i belongs to category c." + ], + [ + "Among the model architectures we experimented with and without data augmentation were:", + "Fully Connected dense networks: Model hyperparameters were inspired from the previous work done by Vo et al and Mathur et al. This was also used as a baseline model but we did not get appreciable performance on such architecture due to FC networks not being able to capture local and long term dependencies.", + "Convolution based architectures: Architecture and hyperparameter choices were chosen from the past study Deon on the subject. We were able to boost the performance as compared to only FC based network but we noticed better performance from architectures that are suitable to sequences such as text messages or any timeseries data.", + "Sequence models: We used SimpleRNN, LSTM, GRU, Bidirectional LSTM model architecture to capture long term dependencies of the messages in determining the class the message or the tweet belonged to.", + "Based on all the experiments we conducted below model had best performance related to metrics - Recall rate, F1 score and Overall accuracy." + ], + [ + "Choice of model parameters were in the above models were inspired from previous work done but then were tuned to the best performance of the Test dataset. Following parameters were considered for tuning.", + "Learning rate: Based on grid search the best performance was achieved when learning rate was set to 0.01. This value was arrived by a grid search on lr parameter.", + "Number of Bidirectional LSTM units: A set of 32, 64, 128 hidden activation units were considered for tuning the model. 128 was a choice made by Vo et al in modeling for Vietnamese language but with our experiments and with a small dataset to avoid overfitting to train dataset, a smaller unit sizes were considered.", + "Embedding dimension: 50, 100 and 200 dimension word representation from Glove word embedding were considered and the best results were obtained with 100d representation, consistent with choices made in the previous work.", + "Transfer learning on Embedding; Another bit flag for training the embedding on the train data or freezing the embedding from Glove was used. It was determined that set of pre-trained weights from Glove was best when it was fine tuned with Hinglish data. It provides evidence that a separate word or sentence level embedding when learnt for Hinglish text analysis will be very useful.", + "Number of dense FC layers.", + "Maximum length of the sequence to be considered: The max length of tweets/message in the dataset was 1265 while average was 116. We determined that choosing 200 resulted in the best performance." + ], + [ + "During our experimentation, it was evident that this is a hard problem especially detecting the hate speech, text in a code- mixed language. The best recall rate of 77 % for hate speech was obtained by a Bidirectional LSTM with 32 units with a recurrent drop out rate of 0.2. Precision wise GRU type of RNN sequence model faired better than other kinds for hate speech detection. On the other hand for detecting offensive and non offensive tweets, fairly satisfactory results were obtained. For offensive tweets, 92 % precision was and recall rate of 88% was obtained with GRU versus BiLSTM based models. Comparatively, Recall of 85 % and precision of 76 % was obtained by again GRU and BiLSTM based models as shown and marked in the results." + ], + [ + "The results of the experiments are encouraging on detective offensive vs non offensive tweets and messages written in Hinglish in social media. The utilization of data augmentation technique in this classification task was one of the vital contributions which led us to surpass results obtained by previous state of the art Hybrid CNN-LSTM based models. However, the results of the model for predicting hateful tweets on the contrary brings forth some shortcomings of the model. The biggest shortcoming on the model based on error analysis indicates less than generalized examples presented by the dataset. We also note that the embedding learnt from the Hinglish data set may be lacking and require extensive training to have competent word representations of Hinglish text. Given this learning's, we identify that creating word embeddings on much larger Hinglish corpora may have significant results. We also hypothesize that considering alternate methods than translation and transliteration may prove beneficial." + ], + [ + "[1] Mathur, Puneet and Sawhney, Ramit and Ayyar, Meghna and Shah, Rajiv, Did you offend me? classification of offensive tweets in hinglish language, Proceedings of the 2nd Workshop on Abusive Language Online (ALW2)", + "[2] Mathur, Puneet and Shah, Rajiv and Sawhney, Ramit and Mahata, Debanjan Detecting offensive tweets in hindi-english code-switched language Proceedings of the Sixth International Workshop on Natural Language Processing for Social Media", + "[3] Vo, Quan-Hoang and Nguyen, Huy-Tien and Le, Bac and Nguyen, Minh-Le Multi-channel LSTM-CNN model for Vietnamese sentiment analysis 2017 9th international conference on knowledge and systems engineering (KSE)", + "[4] Hochreiter, Sepp and Schmidhuber, J\u00fcrgen Long short-term memory Neural computation 1997", + "[5] Sinha, R Mahesh K and Thakur, Anil Multi-channel LSTM-CNN model for Vietnamese sentiment analysis 2017 9th international conference on knowledge and systems engineering (KSE)", + "[6] Pennington, Jeffrey and Socher, Richard and Manning, Christopher Glove: Global vectors for word representation Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP)", + "[7] Zhang, Lei and Wang, Shuai and Liu, Bing Deep learning for sentiment analysis: A survey Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery", + "[8] Caruana, Rich and Lawrence, Steve and Giles, C Lee Overfitting in neural nets: Backpropagation, conjugate gradient, and early stopping Advances in neural information processing systems", + "[9] Beale, Mark Hudson and Hagan, Martin T and Demuth, Howard B Neural network toolbox user\u2019s guide The MathWorks Incs", + "[10] Chollet, Fran\u00e7ois and others Keras: The python deep learning library Astrophysics Source Code Library", + "[11] Wei, Jason and Zou, Kai EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)" + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0315/instruction.md b/qasper-0315/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..d1882f89aea4764ca5aa36c11f430d0731a2f91c --- /dev/null +++ b/qasper-0315/instruction.md @@ -0,0 +1,64 @@ +Name of Paper: Filling Gender&Number Gaps in Neural Machine Translation with Black-box Context Injection + +Question: What type of syntactic analysis is performed? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Morphological Ambiguity in Translation", + "Black-Box Knowledge Injection", + "Experiments & Results", + "Quantitative Results", + "Qualitative Results", + "Comparison to vanmassenhove-hardmeier-way:2018:EMNLP", + "Other Languages", + "Related Work", + "Conclusions" + ], + "paragraphs": [ + [ + "A common way for marking information about gender, number, and case in language is morphology, or the structure of a given word in the language. However, different languages mark such information in different ways \u2013 for example, in some languages gender may be marked on the head word of a syntactic dependency relation, while in other languages it is marked on the dependent, on both, or on none of them BIBREF0 . This morphological diversity creates a challenge for machine translation, as there are ambiguous cases where more than one correct translation exists for the same source sentence. For example, while the English sentence \u201cI love language\u201d is ambiguous with respect to the gender of the speaker, Hebrew marks verbs for the gender of their subject and does not allow gender-neutral translation. This allows two possible Hebrew translations \u2013 one in a masculine and the other in a feminine form. As a consequence, a sentence-level translator (either human or machine) must commit to the gender of the speaker, adding information that is not present in the source. Without additional context, this choice must be done arbitrarily by relying on language conventions, world knowledge or statistical (stereotypical) knowledge.", + "Indeed, the English sentence \u201cI work as a doctor\u201d is translated into Hebrew by Google Translate using the masculine verb form oved, indicating a male speaker, while \u201cI work as a nurse\u201d is translated with the feminine form ovedet, indicating a female speaker (verified on March 2019). While this is still an issue, there have been recent efforts to reduce it for specific language pairs.", + "We present a simple black-box method to influence the interpretation chosen by an NMT system in these ambiguous cases. More concretely, we construct pre-defined textual hints about the gender and number of the speaker and the audience (the interlocutors), which we concatenate to a given input sentence that we would like to translate accordingly. We then show that a black-box NMT system makes the desired morphological decisions according to the given hint, even when no other evidence is available on the source side. While adding those hints results in additional text on the target side, we show that it is simple to remove, leaving only the desired translation.", + "Our method is appealing as it only requires simple pre-and-post processing of the inputs and outputs, without considering the system internals, or requiring specific annotated data and training procedure as in previous work BIBREF1 . We show that in spite of its simplicity, it is effective in resolving many of the ambiguities and improves the translation quality in up to 2.3 BLEU when given the correct hints, which may be inferred from text metadata or other sources. Finally, we perform a fine-grained syntactic analysis of the translations generated using our method which shows its effectiveness." + ], + [ + "Different languages use different morphological features marking different properties on different elements. For example, English marks for number, case, aspect, tense, person, and degree of comparison. However, English does not mark gender on nouns and verbs. Even when a certain property is marked, languages differ in the form and location of the marking BIBREF0 . For example, marking can occur on the head of a syntactic dependency construction, on its argument, on both (requiring agreement), or on none of them. Translation systems must generate correct target-language morphology as part of the translation process. This requires knowledge of both the source-side and target-side morphology. Current state-of-the-art translation systems do capture many aspects of natural language, including morphology, when a relevant context is available BIBREF2 , BIBREF3 , but resort to \u201cguessing\u201d based on the training-data statistics when it is not. Complications arise when different languages convey different kinds of information in their morphological systems. In such cases, a translation system may be required to remove information available in the source sentence, or to add information not available in it, where the latter can be especially tricky." + ], + [ + "Our goal is to supply an NMT system with knowledge regarding the speaker and interlocutor of first-person sentences, in order to produce the desired target-side morphology when the information is not available in the source sentence. The approach we take in the current work is that of black-box injection, in which we attempt to inject knowledge to the input in order to influence the output of a trained NMT system, without having access to its internals or its training procedure as proposed by vanmassenhove-hardmeier-way:2018:EMNLP.", + "We are motivated by recent work by BIBREF4 who showed that NMT systems learn to track coreference chains when presented with sufficient discourse context. We conjecture that there are enough sentence-internal pronominal coreference chains appearing in the training data of large-scale NMT systems, such that state-of-the-art NMT systems can and do track sentence-internal coreference. We devise a wrapper method to make use of this coreference tracking ability by introducing artificial antecedents that unambiguously convey the desired gender and number properties of the speaker and audience.", + "More concretely, a sentence such as \u201cI love you\u201d is ambiguous with respect to the gender of the speaker and the gender and number of the audience. However, sentences such as \u201cI love you, she told him\u201d are unambiguous given the coreference groups {I, she} and {you, him} which determine I to be feminine singular and you to be masculine singular. We can thus inject the desired information by prefixing a sentence with short generic sentence fragment such as \u201cShe told him:\u201d or \u201cShe told them that\u201d, relying on the NMT system's coreference tracking abilities to trigger the correctly marked translation, and then remove the redundant translated prefix from the generated target sentence. We observed that using a parataxis construction (i.e. \u201cshe said to him:\u201d) almost exclusively results in target-side parataxis as well (in 99.8% of our examples), making it easy to identify and strip the translated version from the target side. Moreover, because the parataxis construction is grammatically isolated from the rest of the sentence, it can be stripped without requiring additional changes or modification to the rest of the sentence, ensuring grammaticality." + ], + [ + "To demonstrate our method in a black-box setting, we focus our experiments on Google's machine translation system (GMT), accessed through its Cloud API. To test the method on real-world sentences, we consider a monologue from the stand-up comedy show \u201cSarah Silverman: A Speck of Dust\u201d. The monologue consists of 1,244 English sentences, all by a female speaker conveyed to a plural, gender-neutral audience. Our parallel corpora consists of the 1,244 English sentences from the transcript, and their corresponding Hebrew translations based on the Hebrew subtitles. We translate the monologue one sentence at a time through the Google Cloud API. Eyeballing the results suggest that most of the translations use the incorrect, but default, masculine and singular forms for the speaker and the audience, respectively. We expect that by adding the relevant condition of \u201cfemale speaking to an audience\u201d we will get better translations, affecting both the gender of the speaker and the number of the audience.", + "To verify this, we experiment with translating the sentences with the following variations: No Prefix\u2014The baseline translation as returned by the GMT system. \u201cHe said:\u201d\u2014Signaling a male speaker. We expect to further skew the system towards masculine forms. \u201cShe said:\u201d\u2014Signaling a female speaker and unknown audience. As this matches the actual speaker's gender, we expect an improvement in translation of first-person pronouns and verbs with first-person pronouns as subjects. \u201cI said to them:\u201d\u2014Signaling an unknown speaker and plural audience. \u201cHe said to them:\u201d\u2014Masculine speaker and plural audience. \u201cShe said to them:\u201d\u2014Female speaker and plural audience\u2014the complete, correct condition. We expect the best translation accuracy on this setup. \u201cHe/she said to him/her\u201d\u2014Here we set an (incorrect) singular gender-marked audience, to investigate our ability to control the audience morphology." + ], + [ + "We compare the different conditions by comparing BLEU BIBREF5 with respect to the reference Hebrew translations. We use the multi-bleu.perl script from the Moses toolkit BIBREF6 . Table shows BLEU scores for the different prefixes. The numbers match our expectations: Generally, providing an incorrect speaker and/or audience information decreases the BLEU scores, while providing the correct information substantially improves it - we see an increase of up to 2.3 BLEU over the baseline. We note the BLEU score improves in all cases, even when given the wrong gender of either the speaker or the audience. We hypothesise this improvement stems from the addition of the word \u201csaid\u201d which hints the model to generate a more \u201cspoken\u201d language which matches the tested scenario. Providing correct information for both speaker and audience usually helps more than providing correct information to either one of them individually. The one outlier is providing \u201cShe\u201d for the speaker and \u201cher\u201d for the audience. While this is not the correct scenario, we hypothesise it gives an improvement in BLEU as it further reinforces the female gender in the sentence." + ], + [ + "The BLEU score is an indication of how close the automated translation is to the reference translation, but does not tell us what exactly changed concerning the gender and number properties we attempt to control. We perform a finer-grained analysis focusing on the relation between the injected speaker and audience information, and the morphological realizations of the corresponding elements. We parse the translations and the references using a Hebrew dependency parser. In addition to the parse structure, the parser also performs morphological analysis and tagging of the individual tokens. We then perform the following analysis.", + "Speaker's Gender Effects: We search for first-person singular pronouns with subject case (ani, unmarked for gender, corresponding to the English I), and consider the gender of its governing verb (or adjectives in copular constructions such as `I am nice'). The possible genders are `masculine', `feminine' and `both', where the latter indicates a case where the none-diacriticized written form admits both a masculine and a feminine reading. We expect the gender to match the ones requested in the prefix.", + "Interlocutors' Gender and Number Effects: We search for second-person pronouns and consider their gender and number. For pronouns in subject position, we also consider the gender and number of their governing verbs (or adjectives in copular constructions). For a singular audience, we expect the gender and number to match the requested ones. For a plural audience, we expect the masculine-plural forms.", + "Results: Speaker. Figure FIGREF3 shows the result for controlling the morphological properties of the speaker ({he, she, I} said). It shows the proportion of gender-inflected verbs for the various conditions and the reference. We see that the baseline system severely under-predicts the feminine form of verbs as compared to the reference. The \u201cHe said\u201d conditions further decreases the number of feminine verbs, while the \u201cI said\u201d conditions bring it back to the baseline level. Finally, the \u201cShe said\u201d prefixes substantially increase the number of feminine-marked verbs, bringing the proportion much closer to that of the reference (though still under-predicting some of the feminine cases).", + "Results: Audience. The chart in Figure FIGREF3 shows the results for controlling the number of the audience (...to them vs nothing). It shows the proportion of singular vs. plural second-person pronouns on the various conditions. It shows a similar trend: the baseline system severely under-predicts the plural forms with respect to the reference translation, while adding the \u201cto them\u201d condition brings the proportion much closer to that of the reference." + ], + [ + "Closely related to our work, vanmassenhove-hardmeier-way:2018:EMNLP proposed a method and an English-French test set to evaluate gender-aware translation, based on the Europarl corpus BIBREF7 . We evaluate our method (using Google Translate and the given prefixes) on their test set to see whether it is applicable to another language pair and domain. Table shows the results of our approach vs. their published results and the Google Translate baseline. As may be expected, Google Translate outperforms their system as it is trained on a different corpus and may use more complex machine translation models. Using our method improves the BLEU score even further." + ], + [ + "To test our method\u2019s outputs on multiple languages, we run our pre-and post-processing steps with Google Translate using examples we sourced from native speakers of different languages. For every example we have an English sentence and two translations in the corresponding language, one in masculine and one in feminine form. Not all examples are using the same source English sentence as different languages mark different information. Table shows that for these specific examples our method worked on INLINEFORM0 of the languages we had examples for, while for INLINEFORM1 languages both translations are masculine, and for 1 language both are feminine." + ], + [ + "E17-1101 showed that given input with author traits like gender, it is possible to retain those traits in Statistical Machine Translation (SMT) models. W17-4727 showed that incorporating morphological analysis in the decoder improves NMT performance for morphologically rich languages. burlot:hal-01618387 presented a new protocol for evaluating the morphological competence of MT systems, indicating that current translation systems only manage to capture some morphological phenomena correctly. Regarding the application of constraints in NMT, N16-1005 presented a method for controlling the politeness level in the generated output. DBLP:journals/corr/FiclerG17aa showed how to guide a neural text generation system towards style and content parameters like the level of professionalism, subjective/objective, sentiment and others. W17-4811 showed that incorporating more context when translating subtitles can improve the coherence of the generated translations. Most closely to our work, vanmassenhove-hardmeier-way:2018:EMNLP also addressed the missing gender information by training proprietary models with a gender-indicating-prefix. We differ from this work by treating the problem in a black-box manner, and by addressing additional information like the number of the speaker and the gender and number of the audience." + ], + [ + "We highlight the problem of translating between languages with different morphological systems, in which the target translation must contain gender and number information that is not available in the source. We propose a method for injecting such information into a pre-trained NMT model in a black-box setting. We demonstrate the effectiveness of this method by showing an improvement of 2.3 BLEU in an English-to-Hebrew translation setting where the speaker and audience gender can be inferred. We also perform a fine-grained syntactic analysis that shows how our method enables to control the morphological realization of first and second-person pronouns, together with verbs and adjectives related to them. In future work we would like to explore automatic generation of the injected context, or the use of cross-sentence context to infer the injected information." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0325/instruction.md b/qasper-0325/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..9aa6b22ad0fcf941472bf60b6b522411b6078322 --- /dev/null +++ b/qasper-0325/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Tag-based Multi-Span Extraction in Reading Comprehension + +Question: What is difference in peformance between proposed model and state-of-the art on other question types? \ No newline at end of file diff --git a/qasper-0341/instruction.md b/qasper-0341/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..f94a0a483d53da676570aaeae4c3b579cb2229f3 --- /dev/null +++ b/qasper-0341/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Exploring Hate Speech Detection in Multimodal Publications + +Question: How large is the dataset? \ No newline at end of file diff --git a/qasper-0510/instruction.md b/qasper-0510/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..47431830442931f5f3459e72e74e263d64367989 --- /dev/null +++ b/qasper-0510/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Cross-lingual, Character-Level Neural Morphological Tagging + +Question: On which dataset is the experiment conducted? \ No newline at end of file diff --git a/qasper-0517/instruction.md b/qasper-0517/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..09f61a1f5b53cb6f43d7c68c134922432b3dca30 --- /dev/null +++ b/qasper-0517/instruction.md @@ -0,0 +1,50 @@ +Name of Paper: Visual Natural Language Query Auto-Completion for Estimating Instance Probabilities + +Question: How they complete a user query prefix conditioned upon an image? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Methods", + "Methods ::: Modifying FactorCell LSTM for Image Query Auto-Completion", + "Methods ::: Fine Tuning BERT for Instance Probability Estimation", + "Methods ::: Data and Training Details", + "Results", + "Results ::: Conclusions" + ], + "paragraphs": [ + [ + "This work focuses on the problem of finding objects in an image based on natural language descriptions. Existing solutions take into account both the image and the query BIBREF0, BIBREF1, BIBREF2. In our problem formulation, rather than having the entire text, we are given only a prefix of the text which requires completing the text based on a language model and the image, and finding a relevant object in the image. We decompose the problem into three components: (i) completing the query from text prefix and an image; (ii) estimating probabilities of objects based on the completed text, and (iii) segmenting and classifying all instances in the image. We combine, extend, and modify state of the art components: (i) we extend a FactorCell LSTM BIBREF3, BIBREF4 which conditionally completes text to complete a query from both a text prefix and an image; (ii) we fine tune a BERT embedding to compute instance probabilities from a complete sentence, and (iii) we use Mask-RCNN BIBREF5 for instance segmentation.", + "Recent natural language embeddings BIBREF6 have been trained with the objectives of predicting masked words and determining whether sentences follow each other, and are efficiently used across a dozen of natural language processing tasks. Sequence models have been conditioned to complete text from a prefix and index BIBREF3, however have not been extended to take into account an image. Deep neural networks have been trained to segment all instances in an image at very high quality BIBREF5, BIBREF7. We propose a novel method of natural language query auto-completion for estimating instance probabilities conditioned on the image and a user query prefix. Our system combines and modifies state of the art components used in query completion, language embedding, and masked instance segmentation. Estimating a broad set of instance probabilities enables selection which is agnostic to the segmentation procedure." + ], + [ + "Figure FIGREF2 shows the architecture of our approach. First, we extract image features with a pre-trained CNN. We incorporate the image features into a modified FactorCell LSTM language model along with the user query prefix to complete the query. The completed query is then fed into a fine-tuned BERT embedding to estimate instance probabilities, which in turn are used for instance selection.", + "We denote a set of objects $o_k \\in O$ where O is the entire set of recognizable object classes. The user inputs a prefix, $p$, an incomplete query on an image, $I$. Given $p$, we auto-complete the intended query $q$. We define the auto-completion query problem in equation DISPLAY_FORM3 as the maximization of the probability of a query conditioned on an image where $w_i \\in A$ is the word in position $i$.", + "", + "We pose our instance probability estimation problem given an auto-completed query $\\mathbf {q^*}$ as a multilabel problem where each class can independently exist. Let $O_{q*}$ be the set of instances referred to in $\\mathbf {q^*}$. Given $\\hat{p}_k$ is our estimate of $P(o_k \\in O_{q*})$ and $y_k = \\mathbb {1}[o_k \\in O_{q*}]$, the instance selection model minimizes the sigmoid cross-entropy loss function:" + ], + [ + "We utilize the FactorCell (FC) adaptation of an LSTM with coupled input and forget gates BIBREF4 to autocomplete queries. The FactorCell is an LSTM with a context-dependent weight matrix $\\mathbf {W^{\\prime }} = \\mathbf {W} + \\mathbf {A}$ in place of $\\mathbf {W}$. Given a character embedding $w_t \\in \\mathbb {R}^e$, a previous hidden state $h_{t-1} \\in \\mathbb {R}^h$ , the adaptation matrix, $\\mathbf {A}$, is formed by taking the product of the context, c, with two basis tensors $\\mathbf {Z_L} \\in \\mathbb {R}^{m\\times (e+h)\\times r}$ and $\\mathbf {Z_R} \\in \\mathbb {R}^{r\\times h \\times m}$.", + "To adapt the FactorCell BIBREF4 for our purposes, we replace user embeddings with a low-dimensional image representation. Thus, we are able to modify each query completion to be personalized to a specific image representation. We extract features from an input image using a CNN pretrained on ImageNet, retraining only the last two fully connected layers. The image feature vector is fed into the FactorCell through the adaptation matrix. We perform beam search over the sequence of predicted characters to chose the optimal completion for the given prefix." + ], + [ + "We fine tune a pre-trained BERT embedding to perform transfer learning for our instance selection task. We use a 12-layer implementation which has been shown to generalize and perform well when fine-tuned for new tasks such as question answering, text classification, and named entity recognition. To apply the model to our task, we add an additional dense layer to the BERT architecture with 10% dropout, mapping the last pooled layer to the object classes in our data." + ], + [ + "We use the Visual Genome (VG) BIBREF8 and ReferIt BIBREF9 datasets which are suitable for our purposes. The VG data contains images, region descriptions, relationships, question-answers, attributes, and object instances. The region descriptions provide a replacement for queries since they mention various objects in different regions of each image. However, while some region descriptions are referring phrases, some are more similar to descriptions (see examples in Table TABREF10). The large number of examples makes the Visual Genome dataset particularly useful for our task. The smaller ReferIt dataset consists of referring expressions attached to images which more closely resemble potential user queries of images. We train separate models using both datasets.", + "For training, we aggregated (query, image) pairs using the region descriptions from the VG dataset and referring expressions from the ReferIt dataset. Our VG training set consists of 85% of the data: 16k images and 740k corresponding region descriptions. The Referit training data consists of 9k images and 54k referring expressions.", + "The query completion models are trained using a 128 dimensional image representation, a rank $r=64$ personalized matrix, 24 dimensional character embeddings, 512 dimensional LSTM hidden units, and a max length of 50 characters per query, with Adam at a 5e-4 learning rate, and a batch size of 32 for 80K iterations. The instance selection model is trained using (region description, object set) pairs from the VG dataset resulting in a training set of approximately 1.73M samples. The remaining 300K samples are split into validation and testing. Our training procedure for the instance selection model fine tunes all 12 layers of BERT with 32 sample batch sizes for 250K iterations, using Adam and performing learning rate warm-up for the first 10% of iterations with a target 5e-5 learning rate. The entire training processes takes around a day on an NVIDIA Tesla P100 GPU." + ], + [ + "Figure 3 shows example results. We evaluate query completion by language perplexity and mean reciprocal rank (MRR) and evaluate instance selection by F1-score. We compare the perplexity on both sets of test queries using corresponding images vs. random noise as context. Table TABREF11 shows perplexity on the VG and ReferIt test queries with both corresponding images and random noise. The VG and ReferIt datasets have character vocabulary sizes of 89 and 77 respectively.", + "Given the matching index $t$ of the true query in the top 10 completions we compute the MRR as $\\sum _{n}{\\frac{1}{t}}$ where we replace the reciprocal rank with 0 if the true query does not appear in the top ten completions. We evaluate the VG and ReferIt test queries with varying prefix sizes and compare performance with the corresponding image and random noise as context. MRR is influenced by the length of the query, as longer queries are more difficult to match. Therefore, as expected we observe better performance on the ReferIt dataset for all prefix lengths. Finally, our instance selection achieves an F1-score of 0.7618 over all 2,909 instance classes." + ], + [ + "Our results demonstrate that auto-completion based on both language and vision performs better than by using only language, and that fine tuning a BERT embedding allows to efficiently rank instances in the image. In future work we would like to extract referring expressions using simple grammatical rules to differentiate between referring and non-referring region descriptions. We would also like to combine the VG and ReferIt datasets to train a single model and scale up our datasets to improve query completions." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0528/instruction.md b/qasper-0528/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..3dcd2148a64adc5410ab0c480307888ed5672b7b --- /dev/null +++ b/qasper-0528/instruction.md @@ -0,0 +1,83 @@ +Name of Paper: Analysis of Risk Factor Domains in Psychosis Patient Health Records + +Question: What datasets did the authors use? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Data", + "Annotation Task", + "Inter-Annotator Agreement", + "Topic Extraction", + "Results and Discussion", + "Future Work and Conclusion", + "Acknowledgments" + ], + "paragraphs": [ + [ + "Psychotic disorders typically emerge in late adolescence or early adulthood BIBREF0 , BIBREF1 and affect approximately 2.5-4% of the population BIBREF2 , BIBREF3 , making them one of the leading causes of disability worldwide BIBREF4 . A substantial proportion of psychiatric inpatients are readmitted after discharge BIBREF5 . Readmissions are disruptive for patients and families, and are a key driver of rising healthcare costs BIBREF6 , BIBREF7 . Reducing readmission risk is therefore a major unmet need of psychiatric care. Developing clinically implementable machine learning tools to enable accurate assessment of risk factors associated with readmission offers opportunities to inform the selection of treatment interventions and implement appropriate preventive measures.", + "In psychiatry, traditional strategies to study readmission risk factors rely on clinical observation and manual retrospective chart review BIBREF8 , BIBREF9 . This approach, although benefitting from clinical expertise, does not scale well for large data sets, is effort-intensive, and lacks automation. An efficient, more robust, and cheaper NLP-based alternative approach has been developed and met with some success in other medical fields BIBREF10 . However, this approach has seldom been applied in psychiatry because of the unique characteristics of psychiatric medical record content.", + "There are several challenges for topic extraction when dealing with clinical narratives in psychiatric EHRs. First, the vocabulary used is highly varied and context-sensitive. A patient may report \u201cfeeling `really great and excited'\" \u2013 symptoms of mania \u2013 without any explicit mention of keywords that differ from everyday vocabulary. Also, many technical terms in clinical narratives are multiword expressions (MWEs) such as `obsessive body image', `linear thinking', `short attention span', or `panic attack'. These phrasemes are comprised of words that in isolation do not impart much information in determining relatedness to a given topic but do in the context of the expression.", + "Second, the narrative structure in psychiatric clinical narratives varies considerably in how the same phenomenon can be described. Hallucinations, for example, could be described as \u201cthe patient reports auditory hallucinations,\" or \u201cthe patient has been hearing voices for several months,\" amongst many other possibilities.", + "Third, phenomena can be directly mentioned without necessarily being relevant to the patient specifically. Psychosis patient discharge summaries, for instance, can include future treatment plans (e.g. \u201cPrevent relapse of a manic or major depressive episode.\", \u201cPrevent recurrence of psychosis.\") containing vocabulary that at the word-level seem strongly correlated with readmission risk. Yet at the paragraph-level these do not indicate the presence of a readmission risk factor in the patient and in fact indicate the absence of a risk factor that was formerly present.", + "Lastly, given the complexity of phenotypic assessment in psychiatric illnesses, patients with psychosis exhibit considerable differences in terms of illness and symptom presentation. The constellation of symptoms leads to various diagnoses and comorbidities that can change over time, including schizophrenia, schizoaffective disorder, bipolar disorder with psychosis, and substance use induced psychosis. Thus, the lexicon of words and phrases used in EHRs differs not only across diagnoses but also across patients and time.", + "Taken together, these factors make topic extraction a difficult task that cannot be accomplished by keyword search or other simple text-mining techniques.", + "To identify specific risk factors to focus on, we not only reviewed clinical literature of risk factors associated with readmission BIBREF11 , BIBREF12 , but also considered research related to functional remission BIBREF13 , forensic risk factors BIBREF14 , and consulted clinicians involved with this project. Seven risk factor domains \u2013 Appearance, Mood, Interpersonal, Occupation, Thought Content, Thought Process, and Substance \u2013 were chosen because they are clinically relevant, consistent with literature, replicable across data sets, explainable, and implementable in NLP algorithms.", + "In our present study, we evaluate multiple approaches to automatically identify which risk factor domains are associated with which paragraphs in psychotic patient EHRs. We perform this study in support of our long-term goal of creating a readmission risk classifier that can aid clinicians in targeting individual treatment interventions and assessing patient risk of harm (e.g. suicide risk, homicidal risk). Unlike other contemporary approaches in machine learning, we intend to create a model that is clinically explainable and flexible across training data while maintaining consistent performance.", + "To incorporate clinical expertise in the identification of risk factor domains, we undertake an annotation project, detailed in section 3.1. We identify a test set of over 1,600 EHR paragraphs which a team of three domain-expert clinicians annotate paragraph-by-paragraph for relevant risk factor domains. Section 3.2 describes the results of this annotation task. We then use the gold standard from the annotation project to assess the performance of multiple neural classification models trained exclusively on Term Frequency \u2013 Inverse Document Frequency (TF-IDF) vectorized EHR data, described in section 4. To further improve the performance of our model, we incorporate domain-relevant MWEs identified using all in-house data." + ], + [ + "McCoy et al. mccoy2015clinical constructed a corpus of web data based on the Research Domain Criteria (RDoC) BIBREF15 , and used this corpus to create a vector space document similarity model for topic extraction. They found that the `negative valence' and `social' RDoC domains were associated with readmission. Using web data (in this case data retrieved from the Bing API) to train a similarity model for EHR texts is problematic since it differs from the target data in both structure and content. Based on reconstruction of the procedure, we conclude that many of the informative MWEs critical to understanding the topics of paragraphs in EHRs are not captured in the web data. Additionally, RDoC is by design a generalized research construct to describe the entire spectrum of mental disorders and does not include domains that are based on observation or causes of symptoms. Important indicators within EHRs of patient health, like appearance or occupation, are not included in the RDoC constructs.", + "Rumshisky et al. rumshisky2016predicting used a corpus of EHRs from patients with a primary diagnosis of major depressive disorder to create a 75-topic LDA topic model that they then used in a readmission prediction classifier pipeline. Like with McCoy et al. mccoy2015clinical, the data used to train the LDA model was not ideal as the generalizability of the data was narrow, focusing on only one disorder. Their model achieved readmission prediction performance with an area under the curve of .784 compared to a baseline of .618. To perform clinical validation of the topics derived from the LDA model, they manually evaluated and annotated the topics, identifying the most informative vocabulary for the top ten topics. With their training data, they found the strongest coherence occurred in topics involving substance use, suicidality, and anxiety disorders. But given the unsupervised nature of the LDA clustering algorithm, the topic coherence they observed is not guaranteed across data sets." + ], + [ + "[2]The vast majority of patients in our target cohort are", + "dependents on a parental private health insurance plan.", + "Our target data set consists of a corpus of discharge summaries, admission notes, individual encounter notes, and other clinical notes from 220 patients in the OnTrackTM program at McLean Hospital. OnTrackTM is an outpatient program, focusing on treating adults ages 18 to 30 who are experiencing their first episodes of psychosis. The length of time in the program varies depending on patient improvement and insurance coverage, with an average of two to three years. The program focuses primarily on early intervention via individual therapy, group therapy, medication evaluation, and medication management. See Table TABREF2 for a demographic breakdown of the 220 patients, for which we have so far extracted approximately 240,000 total EHR paragraphs spanning from 2011 to 2014 using Meditech, the software employed by McLean for storing and organizing EHR data.", + "These patients are part of a larger research cohort of approximately 1,800 psychosis patients, which will allow us to connect the results of this EHR study with other ongoing research studies incorporating genetic, cognitive, neurobiological, and functional outcome data from this cohort.", + "We also use an additional data set for training our vector space model, comprised of EHR texts queried from the Research Patient Data Registry (RPDR), a centralized regional data repository of clinical data from all institutions in the Partners HealthCare network. These records are highly comparable in style and vocabulary to our target data set. The corpus consists of discharge summaries, encounter notes, and visit notes from approximately 30,000 patients admitted to the system's hospitals with psychiatric diagnoses and symptoms. This breadth of data captures a wide range of clinical narratives, creating a comprehensive foundation for topic extraction.", + "After using the RPDR query tool to extract EHR paragraphs from the RPDR database, we created a training corpus by categorizing the extracted paragraphs according to their risk factor domain using a lexicon of 120 keywords that were identified by the clinicians involved in this project. Certain domains \u2013 particularly those involving thoughts and other abstract concepts \u2013 are often identifiable by MWEs rather than single words. The same clinicians who identified the keywords manually examined the bigrams and trigrams with the highest TF-IDF scores for each domain in the categorized paragraphs, identifying those which are conceptually related to the given domain. We then used this lexicon of 775 keyphrases to identify more relevant training paragraphs in RPDR and treat them as (non-stemmed) unigrams when generating the matrix. By converting MWEs such as `shortened attention span', `unusual motor activity', `wide-ranging affect', or `linear thinking' to non-stemmed unigrams, the TF-IDF score (and therefore the predictive value) of these terms is magnified. In total, we constructed a corpus of roughly 100,000 paragraphs consisting of 7,000,000 tokens for training our model." + ], + [ + "In order to evaluate our models, we annotated 1,654 paragraphs selected from the 240,000 paragraphs extracted from Meditech with the clinically relevant domains described in Table TABREF3 . The annotation task was completed by three licensed clinicians. All paragraphs were removed from the surrounding EHR context to ensure annotators were not influenced by the additional contextual information. Our domain classification models consider each paragraph independently and thus we designed the annotation task to mirror the information available to the models.", + "The annotators were instructed to label each paragraph with one or more of the seven risk factor domains. In instances where more than one domain was applicable, annotators assigned the domains in order of prevalence within the paragraph. An eighth label, `Other', was included if a paragraph was ambiguous, uninterpretable, or about a domain not included in the seven risk factor domains (e.g. non-psychiatric medical concerns and lab results). The annotations were then reviewed by a team of two clinicians who adjudicated collaboratively to create a gold standard. The gold standard and the clinician-identified keywords and MWEs have received IRB approval for release to the community. They are available as supplementary data to this paper." + ], + [ + "Inter-annotator agreement (IAA) was assessed using a combination of Fleiss's Kappa (a variant of Scott's Pi that measures pairwise agreement for annotation tasks involving more than two annotators) BIBREF16 and Cohen's Multi-Kappa as proposed by Davies and Fleiss davies1982measuring. Table TABREF6 shows IAA calculations for both overall agreement and agreement on the first (most important) domain only. Following adjudication, accuracy scores were calculated for each annotator by evaluating their annotations against the gold standard.", + "Overall agreement was generally good and aligned almost exactly with the IAA on the first domain only. Out of the 1,654 annotated paragraphs, 671 (41%) had total agreement across all three annotators. We defined total agreement for the task as a set-theoretic complete intersection of domains for a paragraph identified by all annotators. 98% of paragraphs in total agreement involved one domain. Only 35 paragraphs had total disagreement, which we defined as a set-theoretic null intersection between the three annotators. An analysis of the 35 paragraphs with total disagreement showed that nearly 30% included the term \u201cblunted/restricted\". In clinical terminology, these terms can be used to refer to appearance, affect, mood, or emotion. Because the paragraphs being annotated were extracted from larger clinical narratives and examined independently of any surrounding context, it was difficult for the annotators to determine the most appropriate domain. This lack of contextual information resulted in each annotator using a different `default' label: Appearance, Mood, and Other. During adjudication, Other was decided as the most appropriate label unless the paragraph contained additional content that encompassed other domains, as it avoids making unnecessary assumptions. [3]Suicidal ideation [4]Homicidal ideation [5]Ethyl alcohol and ethanol", + "A Fleiss's Kappa of 0.575 lies on the boundary between `Moderate' and `Substantial' agreement as proposed by Landis and Koch landis1977measurement. This is a promising indication that our risk factor domains are adequately defined by our present guidelines and can be employed by clinicians involved in similar work at other institutions.", + "The fourth column in Table TABREF6 , Mean Accuracy, was calculated by averaging the three annotator accuracies as evaluated against the gold standard. This provides us with an informative baseline of human parity on the domain classification task.", + "[6]Rectified Linear Units, INLINEFORM0 BIBREF17 [7]Adaptive Moment Estimation BIBREF18 " + ], + [ + "Figure FIGREF8 illustrates the data pipeline for generating our training and testing corpora, and applying them to our classification models.", + "We use the TfidfVectorizer tool included in the scikit-learn machine learning toolkit BIBREF19 to generate our TF-IDF vector space models, stemming tokens with the Porter Stemmer tool provided by the NLTK library BIBREF20 , and calculating TF-IDF scores for unigrams, bigrams, and trigrams. Applying Singular Value Decomposition (SVD) to the TF-IDF matrix, we reduce the vector space to 100 dimensions, which Zhang et al. zhang2011comparative found to improve classifier performance.", + "Starting with the approach taken by McCoy et al. mccoy2015clinical, who used aggregate cosine similarity scores to compute domain similarity directly from their TF-IDF vector space model, we extend this method by training a suite of three-layer multilayer perceptron (MLP) and radial basis function (RBF) neural networks using a variety of parameters to compare performance. We employ the Keras deep learning library BIBREF21 using a TensorFlow backend BIBREF22 for this task. The architectures of our highest performing MLP and RBF models are summarized in Table TABREF7 . Prototype vectors for the nodes in the hidden layer of our RBF model are selected via k-means clustering BIBREF23 on each domain paragraph megadocument individually. The RBF transfer function for each hidden layer node is assigned the same width, which is based off the maximum Euclidean distance between the centroids that were computed using k-means.", + "To prevent overfitting to the training data, we utilize a dropout rate BIBREF24 of 0.2 on the input layer of all models and 0.5 on the MLP hidden layer.", + "Since our classification problem is multiclass, multilabel, and open-world, we employ seven nodes with sigmoid activations in the output layer, one for each risk factor domain. This allows us to identify paragraphs that fall into more than one of the seven domains, as well as determine paragraphs that should be classified as Other. Unlike the traditionally used softmax activation function, which is ideal for single-label, closed-world classification tasks, sigmoid nodes output class likelihoods for each node independently without the normalization across all classes that occurs in softmax.", + "We find that the risk factor domains vary in the degree of homogeneity of language used, and as such certain domains produce higher similarity scores, on average, than others. To account for this, we calculate threshold similarity scores for each domain using the formula min=avg(sim)+ INLINEFORM0 * INLINEFORM1 (sim), where INLINEFORM2 is standard deviation and INLINEFORM3 is a constant, which we set to 0.78 for our MLP model and 1.2 for our RBF model through trial-and-error. Employing a generalized formula as opposed to manually identifying threshold similarity scores for each domain has the advantage of flexibility in regards to the target data, which may vary in average similarity scores depending on its similarity to the training data. If a paragraph does not meet threshold on any domain, it is classified as Other." + ], + [ + "Table TABREF9 shows the performance of our models on classifying the paragraphs in our gold standard. To assess relative performance of feature representations, we also include performance metrics of our models without MWEs. Because this is a multilabel classification task we use macro-averaging to compute precision, recall, and F1 scores for each paragraph in the testing set. In identifying domains individually, our models achieved the highest per-domain scores on Substance (F1 INLINEFORM0 0.8) and the lowest scores on Interpersonal and Mood (F1 INLINEFORM1 0.5). We observe a consistency in per-domain performance rankings between our MLP and RBF models.", + "The wide variance in per-domain performance is due to a number of factors. Most notably, the training examples we extracted from RPDR \u2013 while very comparable to our target OnTrackTM data \u2013 may not have an adequate variety of content and range of vocabulary. Although using keyword and MWE matching to create our training corpus has the advantage of being significantly less labor intensive than manually labeling every paragraph in the corpus, it is likely that the homogeneity of language used in the training paragraphs is higher than it would be otherwise. Additionally, all of the paragraphs in the training data are assigned exactly one risk factor domain even if they actually involve multiple risk factor domains, making the clustering behavior of the paragraphs more difficult to define. Figure FIGREF10 illustrates the distribution of paragraphs in vector space using 2-component Linear Discriminant Analysis (LDA) BIBREF26 .", + "Despite prior research indicating that similar classification tasks to ours are more effectively performed by RBF networks BIBREF27 , BIBREF28 , BIBREF29 , we find that a MLP network performs marginally better with significantly less preprocessing (i.e. k-means and width calculations) involved. We can see in Figure FIGREF10 that Thought Process, Appearance, Substance, and \u2013 to a certain extent \u2013 Occupation clearly occupy specific regions, whereas Interpersonal, Mood, and Thought Content occupy the same noisy region where multiple domains overlap. Given that similarity is computed using Euclidean distance in an RBF network, it is difficult to accurately classify paragraphs that fall in regions occupied by multiple risk factor domain clusters since prototype centroids from the risk factor domains will overlap and be less differentiable. This is confirmed by the results in Table TABREF9 , where the differences in performance between the RBF and MLP models are more pronounced in the three overlapping domains (0.496 vs 0.448 for Interpersonal, 0.530 vs 0.496 for Mood, and 0.721 vs 0.678 for Thought Content) compared to the non-overlapping domains (0.564 vs 0.566 for Appearance, 0.592 vs 0.598 for Occupation, 0.797 vs 0.792 for Substance, and 0.635 vs 0.624 for Thought Process). We also observe a similarity in the words and phrases with the highest TF-IDF scores across the overlapping domains: many of the Thought Content words and phrases with the highest TF-IDF scores involve interpersonal relations (e.g. `fear surrounding daughter', `father', `family history', `familial conflict') and there is a high degree of similarity between high-scoring words for Mood (e.g. `meets anxiety criteria', `cope with mania', `ocd'[8]) and Thought Content (e.g. `mania', `feels anxious', `feels exhausted').", + "[8]Obsessive-compulsive disorder", + "MWEs play a large role in correctly identifying risk factor domains. Factoring them into our models increased classification performance by 15%, a marked improvement over our baseline model. This aligns with our expectations that MWEs comprised of a quotidian vocabulary hold much more clinical significance than when the words in the expressions are treated independently.", + "Threshold similarity scores also play a large role in determining the precision and recall of our models: higher thresholds lead to a smaller number of false positives and a greater number of false negatives for each risk factor domain. Conversely, more paragraphs are incorrectly classified as Other when thresholds are set higher. Since our classifier will be used in future work as an early step in a data analysis pipeline for determining readmission risk, misclassifying a paragraph with an incorrect risk factor domain at this stage can lead to greater inaccuracies at later stages. Paragraphs misclassified as Other, however, will be discarded from the data pipeline. Therefore, we intentionally set a conservative threshold where only the most confidently labeled paragraphs are assigned membership in a particular domain." + ], + [ + "To achieve our goal of creating a framework for a readmission risk classifier, the present study performed necessary evaluation steps by updating and adding to our model iteratively. In the first stage of the project, we focused on collecting the data necessary for training and testing, and on the domain classification annotation task. At the same time, we began creating the tools necessary for automatically extracting domain relevance scores at the paragraph and document level from patient EHRs using several forms of vectorization and topic modeling. In future versions of our risk factor domain classification model we will explore increasing robustness through sequence modeling that considers more contextual information.", + "Our current feature set for training a machine learning classifier is relatively small, consisting of paragraph domain scores, bag-of-words, length of stay, and number of previous admissions, but we intend to factor in many additional features that extend beyond the scope of the present study. These include a deeper analysis of clinical narratives in EHRs: our next task will be to extend our EHR data pipeline by distinguishing between clinically positive and negative phenomena within each risk factor domain. This will involve a series of annotation tasks that will allow us to generate lexicon-based and corpus-based sentiment analysis tools. We can then use these clinical sentiment scores to generate a gradient of patient improvement or deterioration over time.", + "We will also take into account structured data that have been collected on the target cohort throughout the course of this study such as brain based electrophysiological (EEG) biomarkers, structural brain anatomy from MRI scans (gray matter volume, cortical thickness, cortical surface-area), social and role functioning assessments, personality assessment (NEO-FFI[9]), and various symptom scales (PANSS[10], MADRS[11], YMRS[12]). For each feature we consider adding, we will evaluate the performance of the classifier with and without the feature to determine its contribution as a predictor of readmission." + ], + [ + "This work was supported by a grant from the National Institute of Mental Health (grant no. 5R01MH109687 to Mei-Hua Hall). We would also like to thank the LOUHI 2018 Workshop reviewers for their constructive and helpful comments.", + "[9]NEO Five-Factor Inventory BIBREF30 [10]Positive and Negative Syndrome Scale BIBREF31 [11]Montgomery-Asperg Depression Rating Scale BIBREF32 [12]Young Mania Rating Scale BIBREF33 " + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0543/instruction.md b/qasper-0543/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..f2d174fe2d1e9629c94cc12e369501463c80fc70 --- /dev/null +++ b/qasper-0543/instruction.md @@ -0,0 +1,41 @@ +Name of Paper: How Does Language Influence Documentation Workflow? Unsupervised Word Discovery Using Translations in Multiple Languages + +Question: How is the performance of the model evaluated? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Methodology ::: The Multilingual Mboshi Parallel Corpus:", + "Methodology ::: Bilingual Unsupervised Word Segmentation/Discovery Approach:", + "Methodology ::: Multilingual Leveraging:", + "Experiments", + "Conclusion" + ], + "paragraphs": [ + [ + "The Cambridge Handbook of Endangered Languages BIBREF3 estimates that at least half of the 7,000 languages currently spoken worldwide will no longer exist by the end of this century. For these endangered languages, data collection campaigns have to accommodate the challenge that many of them are from oral tradition, and producing transcriptions is costly. This transcription bottleneck problem can be handled by translating into a widely spoken language to ensure subsequent interpretability of the collected recordings, and such parallel corpora have been recently created by aligning the collected audio with translations in a well-resourced language BIBREF1, BIBREF2, BIBREF4. Moreover, some linguists suggested that more than one translation should be collected to capture deeper layers of meaning BIBREF5.", + "This work is a contribution to the Computational Language Documentation (CLD) research field, that aims to replace part of the manual steps performed by linguists during language documentation initiatives by automatic approaches. Here we investigate the unsupervised word discovery and segmentation task, using the bilingual-rooted approach from BIBREF6. There, words in the well-resourced language are aligned to unsegmented phonemes in the endangered language in order to identify group of phonemes, and to cluster them into word-like units. We experiment with the Mboshi-French parallel corpus, translating the French text into four other well-resourced languages in order to investigate language impact in this CLD approach. Our results hint that this language impact exists, and that models based on different languages will output different word-like units." + ], + [ + "In this work we extend the bilingual Mboshi-French parallel corpus BIBREF2, fruit of the documentation process of Mboshi (Bantu C25), an endangered language spoken in Congo-Brazzaville. The corpus contains 5,130 utterances, for which it provides audio, transcriptions and translations in French. We translate the French into four other well-resourced languages through the use of the $DeepL$ translator. The languages added to the dataset are: English, German, Portuguese and Spanish. Table shows some statistics for the produced Multilingual Mboshi parallel corpus." + ], + [ + "We use the bilingual neural-based Unsupervised Word Segmentation (UWS) approach from BIBREF6 to discover words in Mboshi. In this approach, Neural Machine Translation (NMT) models are trained between language pairs, using as source language the translation (word-level) and as target, the language to document (unsegmented phonemic sequence). Due to the attention mechanism present in these networks BIBREF7, posterior to training, it is possible to retrieve soft-alignment probability matrices between source and target sequences. These matrices give us sentence-level source-to-target alignment information, and by using it for clustering neighbor phonemes aligned to the same translation word, we are able to create segmentation in the target side. The product of this approach is a set of (discovered-units, translation words) pairs." + ], + [ + "In this work we apply two simple methods for including multilingual information into the bilingual models from BIBREF6. The first one, Multilingual Voting, consists of merging the information learned by models trained with different language pairs by performing a voting over the final discovered boundaries. The voting is performed by applying an agreement threshold $T$ over the output boundaries. This threshold balances between accepting all boundaries from all the bilingual models (zero agreement) and accepting only input boundaries discovered by all these models (total agreement). The second method is ANE Selection. For every language pair and aligned sentence in the dataset, a soft-alignment probability matrix is generated. We use Average Normalized Entropy (ANE) BIBREF8 computed over these matrices for selecting the most confident one for segmenting each phoneme sequence. This exploits the idea that models trained on different language pairs will have language-related behavior, thus differing on the resulting alignment and segmentation over the same phoneme sequence." + ], + [ + "The experiment settings from this paper and evaluation protocol for the Mboshi corpus (Boundary F-scores using the ZRC speech reference) are the same from BIBREF8. Table presents the results for bilingual UWS and multilingual leveraging. For the former, we reach our best result by using as aligned information the French, the original aligned language for this dataset. Languages closely related to French (Spanish and Portuguese) ranked better, while our worst result used German. English also performs notably well in our experiments. We believe this is due to the statistics features of the resulting text. We observe in Table that the English portion of the dataset contains the smallest vocabulary among all languages. Since we train our systems in very low-resource settings, vocabulary-related features can impact greatly the system's capacity to language-model, and consequently the final quality of the produced alignments. Even in high-resource settings, it was already attested that some languages are more difficult to model than others BIBREF9.", + "For the multilingual selection experiments, we experimented combining the languages from top to bottom as they appear Table (ranked by performance; e.g. 1-3 means the combination of FR(1), EN(2) and PT(3)). We observe that the performance improvement is smaller than the one observed in previous work BIBREF10, which we attribute to the fact that our dataset was artificially augmented. This could result in the available multilingual form of supervision not being as rich as in a manually generated dataset. Finally, the best boundary segmentation result is obtained by performing multilingual voting with all the languages and an agreement of 50%, which indicates that the information learned by different languages will provide additional complementary evidence.", + "Lastly, following the methodology from BIBREF8, we extract the most confident alignments (in terms of ANE) discovered by the bilingual models. Table presents the top 10 most confident (discovered type, translation) pairs. Looking at the pairs the bilingual models are most confident about, we observe there are some types discovered by all the bilingual models (e.g. Mboshi word itua, and the concatenation obo\u00e1+ng\u00e1). However, the models still differ for most of their alignments in the table. This hints that while a portion of the lexicon might be captured independently of the language used, other structures might be more dependent of the chosen language. On this note, BIBREF11 suggests the notion of word cannot always be meaningfully defined cross-linguistically." + ], + [ + "In this work we train bilingual UWS models using the endangered language Mboshi as target and different well-resourced languages as aligned information. Results show that similar languages rank better in terms of segmentation performance, and that by combining the information learned by different models, segmentation is further improved. This might be due to the different language-dependent structures that are captured by using more than one language. Lastly, we extend the bilingual Mboshi-French parallel corpus, creating a multilingual corpus for the endangered language Mboshi that we make available to the community." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0573/instruction.md b/qasper-0573/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..f28819e3a0cc50ac6d1380429454e272720d2b7c --- /dev/null +++ b/qasper-0573/instruction.md @@ -0,0 +1,115 @@ +Name of Paper: Representation of Constituents in Neural Language Models: Coordination Phrase as a Case Study + +Question: What is the size of the datasets employed? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Methods ::: Psycholinguistics Paradigm", + "Methods ::: Models Tested ::: Recurrent Neural Network (RNN) Language Models", + "Methods ::: Models Tested ::: ActionLSTM", + "Methods ::: Models Tested ::: Generative Recurrent Neural Network Grammars (RNNG)", + "Experiment 1: Non-coordination Agreement", + "Experiment 2: Simple Coordination", + "Experiment 2: Simple Coordination ::: Number Agreement", + "Experiment 2: Simple Coordination ::: Gender Agreement", + "Experiment 3: Complex Coordination", + "Experiment 3: Complex Coordination ::: Complex Coordination Control", + "Experiment 3: Complex Coordination ::: Complex Coordination Critical", + "Experiment 4: Inverted Coordination", + "Discussion", + "Acknowledgments", + "The Effect of Annotation Schemes", + "PTB/FTB Agreement Patterns" + ], + "paragraphs": [ + [ + "Humans deploy structure-sensitive expectations to guide processing during natural language comprehension BIBREF0. While it has been shown that neural language models show similar structure-sensitivity in their predictions about upcoming material BIBREF1, BIBREF2, previous work has focused on dependencies that are conditioned by features attached to a single word, such as subject number BIBREF3, BIBREF4 or wh-question words BIBREF5. There has been no systematic investigation into models' ability to compute phrase-level features\u2014features that are attached to a set of words\u2014and whether models can deploy these more abstract properties to drive downstream expectations.", + "In this work, we assess whether state-of-the-art neural models can compute and employ phrase-level gender and number features of coordinated subject Noun Phrases (CoordNPs) with two nouns. Typical syntactic phrases are endocentric: they are headed by a single child, whose features determine the agreement requirements for the entire phrase. In Figure FIGREF1, for example, the word star heads the subject NP The star; since star is singular, the verb must be singular. CoordNPs lack endocentricity: neither conjunct NP solely determines the features of the NP as a whole. Instead, these feature values are determined by compositional rules sensitive to the features of the conjuncts and the identity of the coordinator. In Figure FIGREF1, because the coordinator is and, the subject NP number is plural even though both conjuncts (the star and the moon) are singular. As this case demonstrates, the agreement behavior for CoordNPs must be driven by more abstract, constituent-level representations, and cannot be reduced to features hosted on a single lexical item.", + "We use four suites of experiments to assess whether neural models are able to build up phrase-level representations of CoordNPs on the fly and deploy them to drive humanlike behavior. First, we present a simple control experiment to show that models can represent number and gender features of non-coordinate NPs (Non-coordination Agreement). Second, we show that models modulate their expectations for downstream verb number based on the CoordNP's coordinating conjunction combined with the features of the coordinated nouns (Simple Coordination). We rule out the possibility that models are using simple heuristics by designing a set of stimuli where a simple heuristic would fail due to structural ambiguity (Complex Coordination). The striking success for all models in this experiment indicates that even neural models with no explicit hierarchical bias, trained on a relatively small amount of text are able to learn fine-grained and robust generalizations about the interaction between CoordNPs and local syntactic context. Finally, we use subject\u2013auxiliary inversion to test whether an upstream lexical item modulates model expectation for the phrasal-level features of a downstream CoordNP (Inverted Coordination). Here, we find that all models are insensitive to the fine-grained features of this particular syntactic context. Overall, our results indicate that neural models can learn fine-grained information about the interaction of Coordinated NPs and local syntactic context, but their behavior remains unhumanlike in many key respects." + ], + [ + "To determine whether state-of-the-art neural architectures are capable of learning humanlike CoordNP/verb agreement properties, we adopt the psycholinguistics paradigm for model assessment. In this paradigm the models are tested using hand-crafted sentences designed to test underlying network knowledge. The assumption here is that if a model implicitly learns humanlike linguistic knowledge during training, its expectations for upcoming words should qualitatively match human expectations in novel contexts. For example, BIBREF1 and BIBREF6 assessed how well neural models had learned the subject/verb number agreement by feeding them with the prefix The keys to the cabinet .... If the models predicted the grammatical continuation are over the ungrammatical continuation is, they can be said to have learned the number agreement insofar as the number of the head noun and not the number of the distractor noun, cabinet, drives expectations about the number of the matrix verb.", + "If models are able to robustly modulate their expectations based on the internal components of the CoordNP, this will provide evidence that the networks are building up a context-sensitive phrase-level representation. We quantify model expectations as surprisal values. Surprisal is the negative log-conditional probability $S(x_i) = -\\log _2 p(x_i|x_1 \\dots x_{i-1})$ of a sentence's $i^{th}$ word $x_i$ given the previous words. Surprisal tells us how strongly $x_i$ is expected in context and is known to correlate with human processing difficulty BIBREF7, BIBREF0, BIBREF8. In the CoordNP/Verb agreement studies presented here, cases where the proceeding context sets high expectation for a number-inflected verb form $w_i$, (e.g. singular `is') we would expect $S(w_i)$ to be lower than its number-mismatched counterpart (e.g. plural `are')." + ], + [ + "are trained to output the probability distribution of the upcoming word given a context, without explicitly representing the structure of the context BIBREF9, BIBREF10. We trained two two-layer recurrent neural language models with long short-term memory architecture BIBREF11 on a relatively small corpus. The first model, referred as `LSTM (PTB)' in the following sections, was trained on the sentences from Penn Treebank BIBREF12. The second model, referred as `LSTM (FTB)', was trained on the sentences from French Treebank BIBREF13. We set the size of input word embedding and LSTM hidden layer of both models as 256.", + "We also compare LSTM language models trained on large corpora. We incorporate two pretrained English language models: one trained on the Billion Word benchmark (referred as `LSTM (1B)') from BIBREF14, and the other trained on English Wikipedia (referred as `LSTM (enWiki)') from BIBREF3. For French, we trained a large LSTM language model (referred as `LSTM (frWaC)') on a random subset (about 4 million sentences, 138 million word tokens) of the frWaC dataset BIBREF15. We set the size of the input embeddings and hidden layers to 400 for the LSTM (frWaC) model since it is trained on a large dataset." + ], + [ + "models the linearized bracketed tree structure of a sentence by learning to predict the next action required to construct a phrase-structure parse BIBREF16. The action space consists of three possibilities: open a new non-terminal node and opening bracket; generate a terminal node; and close a bracket. To compute surprisal values for a given token, we approximate $P(w_i|w_{1\\cdots i-1)}$ by marginalizing over the most-likely partial parses found by word-synchronous beam search BIBREF17." + ], + [ + "jointly model the word sequence as well as the underlying syntactic structure BIBREF18. Following BIBREF19, we estimate surprisal using word-synchronous beam search BIBREF17. We use the same hyper-parameter settings as BIBREF18.", + "The annotation schemes used to train the syntactically-supervised models differ slightly between French and English. In the PTB (English) CoordNPs are flat structures bearing an `NP' label. In FTB (French), CoordNPs are binary-branching, labeled as NPs, except for the phrasal node dominating the coordinating conjunction, which is labeled `COORD'. We examine the effects of annotation schemes on model performance in Appendix SECREF8." + ], + [ + "In order to provide a baseline for following experiments, here we assess whether the models tested have learned basic representations of number and gender features for non-coordinated Noun Phrases. We test number agreement in English and French as well as gender agreement in French. Both English and French have two grammatical number feature: singular (sg) and plural (pl). French has two grammatical gender features: masculine (m) and feminine (f).", + "The experimental materials include sentences where the subject NPs contain a single noun which can either match with the matrix verb (in the case of number agreement) or a following predicative adjective (in the case of gender agreement). Conditions are given in Table TABREF9 and Table TABREF10. We measure model behavior by computing the plural expectation, or the surprisal of the singular continuation minus the surprisal of the plural continuation for each condition and took the average for each condition. We expect a positive plural expectation in the Npl conditions and a negative plural expectation in the Nsg conditions. For gender expectation we compute a gender expectation, which is S(feminine continuation) $-$ S(masculine continuation). We measure surprisal at the verbs and predicative adjectives themselves.", + "The results for this experiment are in Figure FIGREF11, with the plural expectation and gender expectation on the y-axis and conditions on the x-axis. For this and subsequent experiments error bars represent 95% confidence intervals for across-item means. For number agreement, all the models in English and French show positive plural expectation when the head noun is plural and negative plural expectation when it is singular. For gender agreement, however, only the LSTM (frWaC) shows modulation of gender expectation based on the gender of the head noun. This is most likely due to the lower frequency of predicative adjectives compared to matrix verbs in the corpus." + ], + [ + "In this section, we test whether neural language models can use grammatical features hosted on multiple components of a coordination phrase\u2014the coordinated nouns as well as the coordinating conjunction\u2014to drive downstream expectations. We test number agreement in both English and French and gender agreement in French." + ], + [ + "In simple subject/verb number agreement, the number features of the CoordNP are determined by the coordinating conjunction and the number features of the two coordinated NPs. CoordNPs formed by and are plural and thus require plural verbs; CoordNPs formed by or allow either plural or singular verbs, often with the number features of the noun linearly closest to the verb playing a more important role, although this varies cross-linguistically BIBREF20. Forced-choice preference experiments in BIBREF21 reveal that English native speakers prefer singular agreement when the closest conjunct in an or-CoordNP is singular and plural agreement when the closest conjunct is plural. In French, both singular and plural verbs are possible when two singular NPs are joined via disjunction BIBREF22.", + "In order to assess whether the neural models learn the basic CoordNP licensing for English, we adapted 37 items from BIBREF21, along the 16 conditions outlined in Table TABREF14. Test items consist of the sentence preamble, followed by either the singular or plural BE verb, half the time in present tense (is/are) and half the time in past tense (was/were). We measured the plural expectation, following the procedure in Section SECREF3. We created 24 items using the same conditions as the English experiment to test the models trained in French, using the 3rd person singular and plural form of verb aller, `to go' (va, vont). Within each item, nouns match in gender; across all conditions half the nouns are masculine, half feminine.", + "The results for this experiment can be seen in Figure FIGREF12, with the results for English on the left and French on the right. The results for and are on the top row, or on the bottom row. For all figures the y-axis shows the plural expectation, or the difference in surprisal between the singular condition and the plural condition. Turning first to English-and (Figure FIGREF12), all models show plural expectation (the bars are significantly greater than zero) in the pl_and_pl and sg_and_pl conditions, as expected. For the pl_and_sg condition, only the LSTM (enWiki) and ActionLSTM are greater than zero, indicating humanlike behavior. For the sg_and_sg condition, only the LSTM (enWiki) model shows the correct plural expectation. For the French-and (Figure FIGREF12), all models show positive plural expectation in all conditions, as expected, except for the LSTM (FTB) in the sg_and_sg condition.", + "Examining the results for English-or, we find that all models demonstrate humanlike expectation in the pl_or_pl and sg_or_pl conditions. The LSTM (1B), LSTM (PTB), and RNNG models show zero or negative singular expectation for the pl_or_sg conditions, as expected. However the LSTM (enWiki) and ActionLSTM models show positive plural expectation in this condition, indicating that they have not learned the humanlike generalizations. All models show significantly negative plural expectation in the sg_or_sg condition, as expected. In the French-or cases, models show almost identical behavior to the and conditions, except the LSTM (frWaC) shows smaller plural expectation when singular nouns are linearly proximal to the verb.", + "These results indicate moderate success at learning coordinate NP agreement, however this success may be the result of an overly simple heuristic. It appears that expectation for both plural and masculine continuations are driven by a linear combination of the two nominal number/gender features transferred into log-probability space, with the earlier noun mattering less than the later noun. A model that optimally captures human grammatical preferences should show no or only slight difference across conditions in the surprisal differential for the and conditions, and be greater than zero in all cases. Yet, all the models tested show gradient performance based on the number of plural conjuncts." + ], + [ + "In French, if two nouns are coordinated with et (and-coordination), agreement must be masculine if there is one masculine element in the coordinate structure. If the nouns are coordinated with ou (or-coordination), both masculine and feminine agreement is acceptable BIBREF23, BIBREF24. Although linear proximity effects have been tested for a number of languages that employ grammatical gender, as in e.g. Slavic languages BIBREF25, there is no systematic study for French.", + "To assess whether the French neural models learned humanlike gender agreement, we created 24 test items, following the examples in Table TABREF16, and measured the masculine expectation. In our test items, the coordinated subject NP is followed by a predicative adjective, which either takes on masculine or feminine gender morphology.", + "Results from the experiment can be seen in Figure FIGREF17. No models shows qualitative difference based on the coordinator, and only the LSTM (frWaC) shows significant behavior difference between conditions. Here, we find positive masculine expectation in the m_and_m and f_and_m conditions, and negative masculine expectation in the f_and_f condition, as expected. However, in the m_and_f condition, the masculine expectation is not significantly different from zero, where we would expect it to be positive. In the or-coordination conditions, following our expectation, masculine expectation is positive when both conjuncts are masculine and negative when both are feminine. For the LSTM (FTB) and ActionLSTM models, the masculine expectation is positive (although not significantly so) in all conditions, consistent with results in Section SECREF3." + ], + [ + "One possible explanation for the results presented in the previous section is that the models are using a `bag of features' approach to plural and masculine licensing that is opaque to syntactic context: Following a coordinating conjunction surrounded by nouns, models simply expect the following verb to be plural, proportionally to the number of plural nouns.", + "In this section, we control for this potential confound by conducting two experiments: In the Complex Coordination Control experiments we assess models' ability to extend basic CoordNP licensing into sententially-embedded environments, where the CoordNP can serve as an embedded subject. In the Complex Coordination Critical experiments, we leverage the sentential embedding environment to demonstrate that when the CoordNPs cannot plausibly serve as the subject of the embedded phrase, models are able to suppress the previously-demonstrated expectations set up by these phrases. These results demonstrate that models are not following a simple strategy for predicting downstream number and gender features, but are building up CoordNP representations on the fly, conditioned on the local syntactic context." + ], + [ + "Following certain sentential-embedding verbs, CoordNPs serve unambiguously as the subject of the verb's sentence complement and should trigger number agreement behavior in the main verb of the embedded clause, similar to the behavior presented in SECREF13. To assess this, we use the 37 test items in English and 24 items in French in section SECREF13, following the conditions in Table TABREF19 (for number agreement), testing only and coordination. For gender agreement, we use the same test items and conditions for and coordination in Section SECREF15, but with the Coordinated NPs embedded in a context similar to SECREF18. As before, we derived the plural expectation by measuring the difference in surprisal between the singular and plural continuations and the gender expectation by computing the difference in surprisal between the masculine and feminine predicates.", + ". Je croyais que les prix et les d\u00e9penses \u00e9taient importants/importantes.", + "I thought that the.pl price.mpl and the.pl expense.fpl were important.mpl/fpl", + "I thought that the prices and the expenses were important.", + "The results for the control experiments can be seen in Figure FIGREF20, with English number agreement on the top row, French number agreement in the middle row and French gender agreement on the bottom. The y-axis shows either plural or masculine expectation, with the various conditions along the x-axis. For English number agreement, we find that the models behave similarly as they do for simple coordination contexts. All models show significant plural expectation when the closest noun is plural, with only two models demonstrating plural expectation in the sg_and_sg case. The French number agreement tests show similar results, with all models except LSTM (FTB) demonstrating significant plural prediction in all cases. Turning to French gender agreement, only the LSTM (frWaC) shows sensitivity to the various conditions, with positive masculine expectation in the m_and_m condition and negative expectation in the f_and_f condition, as expected. These results indicate that the behavior shown in Section SECREF13 extends to more complex syntactic environments\u2014in this case to sentential embeddings. Interestingly, for some models, such as the LSTM (1B), behavior is more humanlike when the CoordNP serves as the subject of an embedded sentence. This may be because the model, which has a large number of hidden states and may be extra sensitive to fine-grained syntactic information carried on lexical items BIBREF2, is using the complementizer, that, to drive more robust expectations." + ], + [ + "In order to assess whether the models' strategy for CoordNP/verb number agreement is sensitive to syntactic context, we contrast the results presented above to those from a second, critical experiment. Here, two coordinated nouns follow a verb that cannot take a sentential complement, as in the examples given in Table TABREF23. Of the two possible continuations\u2014are or is\u2014the plural is only grammatically licensed when the second of the two conjuncts is plural. In these cases, the plural continuation may lead to a final sentence where the first noun serves as the verb's object and the second introduces a second main clause coordinated with the first, as in I fixed the doors and the windows are still broken. For the same reason, the singular-verb continuation is only licensed when the noun immediately following and is singular.", + "We created 37 test items in both English and French, and calculated the plural expectation. If the models were following a simple strategy to drive CoordNP/verb number agreement, then we should see either no difference in plural expectation across the four conditions or behavior no different from the control experiment. If, however, the models are sensitive to the licensing context, we should see a contrast based solely on the number features of the second conjunct, where plural expectation is positive when the second conjunct is plural, and negative otherwise.", + "Experimental items for a critical gender test were created similarly, as in Example SECREF22. As with plural agreement, gender expectation should be driven solely by the second conjunct: For the f_and_m and m_and_m conditions, the only grammatical continuation is one where the adjectival predicate bears masculine gender morphology. Conversely, for the m_and_f or f_and_f conditions, the only grammatical continuation is one where the adjectival predicate bears feminine morphology. As in SECREF13, we created 24 test items and measured the gender expectation by calculating the difference in surprisal between the masculine and feminine continuations.", + ". Nous avons accept\u00e9 les prix et les d\u00e9penses \u00e9taient importants/importantes.", + "we have accepted the.pl price.mpl and the expense.fpl were important.mpl/fpl", + "We have accepted the prices and the expenses were important.", + "The results from the critical experiments are in Figure FIGREF21, with the English number agreement on the top row, French number agreement in the middle and gender expectation on the bottom row. Here the y-axis shows either plural expectation or masculine expectation, with the various conditions are on the x-axis. The results here are strikingly different from those in the control experiments. For number agreement, all models in both languages show strong plural expectation in conditions where the second noun is plural (blue and green bars), as they do in the control experiments. Crucially, when the second noun is singular, the plural expectation is significantly negative for all models (save for the French LSTM (FTB) pl_and_sg condition). Turning to gender agreement, only the LSTM (frWaC) model shows differentiation between the four conditions tested. However, whereas the f_and_m and m_and_f gender expectations are not significantly different from zero in the control condition, in the critical condition they pattern with the purely masculine and purely feminine conditions, indicating that, in this syntactic context, the model has successfully learned to base gender expectation solely off of the second noun.", + "These results are inconsistent with a simple `bag of features' strategy that is insensitive to local syntactic context. They indicate that both models can interpret the same string as either a coordinated noun phrase, or as an NP object and the start of a coordinated VP with the second NP as its subject." + ], + [ + "In addition to using phrase-level features to drive expectation about downstream lexical items, human processors can do the inverse\u2014use lexical features to drive expectations about upcoming syntactic chunks. In this experiment, we assess whether neural models use number features hosted on a verb to modulate their expectations for upcoming CoordNPs.", + "To assess whether neural language models learn inverted coordination rules, we adapted items from Section SECREF13 in both English (37 items) and French (24 items), following the paradigm in Table TABREF24. The first part of the phrase contains either a plural or singular verb and a plural or singular noun. In this case, we sample the surprisal for the continuations and (or is grammatical in all conditions, so it is omitted from this study). Our expectation is that `and' is less surprising in the Vpl_Nsg condition than in the Vsg_Nsg condition, where a CoordNP is not licensed by the grammar in either French or English (as in *What is the pig and the cat eating?). We also expect lower surprisal for and in the Vpl_Nsg condition, where it is obligatory for a grammatical continuation, than in the Vpl_Npl condition, where it is optional.", + "For French experimental items, the question is embedded into a sentential-complement taking verb, following Example SECREF6, due to the fact that unembedded subject-verb inverted questions sound very formal and might be relatively rare in the training data.", + ". Je me demande o\u00f9 vont le maire et", + "I myself ask where go.3PL the.MSG mayor.MSG and", + "The results for both languages are shown in Figure FIGREF25, with the surprisal at the coordinator on the y-axis and the various conditions on the x-axis. No model in either language shows a signficant difference in surprisal between the Vpl_Nsg and Vpl_Npl conditions or between the Vpl_Nsg and Vsg_Nsg conditions. The LSTM (1B) shows significant difference between the Vpl_Nsg and Vpl_Npl conditions, but in the opposite direction than expected, with the coordinator less surprising in the latter condition. These results indicate that the models are unable to use the fine-grained context sensitivity to drive expectations for CoordNPs, at least in the inversion setting." + ], + [ + "The experiments presented here extend and refine a line of research investigating what linguistic knowledge is acquired by neural language models. Previous studies have demonstrated that sequential models trained on a simple regime of optimizing the next word can learn long-distance syntactic dependencies in impressive detail. Our results provide complimentary insights, demonstrating that a range of model architectures trained on a variety of datasets can learn fine-grained information about the interaction of CoordNPs and local syntactic context, but their behavior remains unhumanlike in many key ways. Furthermore, to our best knowledge, this work presents the first psycholinguistic analysis of neural language models trained on French, a high-resource language that has so far been under-investigated in this line of research.", + "In the simple coordination experiment, we demonstrated that models were able to capture some of the agreement behaviors of humans, although their performance deviated in crucial aspects. Whereas human behavior is best modeled as a `percolation' process, the neural models appear to be using a linear combination of NP constituent number to drive CoordNP/verb number agreement, with the second noun weighted more heavily than the first. In these experiments, supervision afforded by the RNNG and ActionLSTM models did not translate into more robust or humanlike learning outcomes. The complex coordination experiments provided evidence that the neural models tested were not using a simple `bag of features' strategy, but were sensitive to syntactic context. All models tested were able to interpret material that had similar surface form in ways that corresponded to two different tree-structural descriptions, based on local context. The inverted coordination experiment provided a contrasting example, in which models were unable to modulate expectations based on subtleties in the syntactic environment.", + "Across all our experiments, the French models performed consistently better on subject/verb number agreement than on subject/predicate gender agreement. Although there are likely more examples of subject/verb number agreement in the French training data, gender agreement is syntactically mandated and widespread in French. It remains an open question why all but one of the models tested were unable to leverage the numerous examples of gender agreement seen in various contexts during training to drive correct subject/predicate expectations." + ], + [ + "This project is supported by a grant of Labex EFL ANR-10-LABX-0083 (and Idex ANR-18-IDEX-0001) for AA and MIT\u2013IBM AI Laboratory and the MIT\u2013SenseTimeAlliance on Artificial Intelligence for RPL. We would like to thank the anonymous reviewers for their comments and Anne Abeill\u00e9 for her advice and feedback." + ], + [ + "This section further investigates the effects of CoordNP annotation schemes on the behaviors of structurally-supervised models. We test whether an explicit COORD phrasal tag improves model performance. We trained two additional RNNG models on 38,546 sentences from the Penn Treebank annotated with two different schemes: The first, RNNG (PTB-control) was trained with the original Penn Treebank annotation. The second, RNNG (PTB-coord), was trained on the same sentences, but with an extended coordination annotation scheme, meant to employ the scheme employed in the FTB, adapted from BIBREF26. We stripped empty categories from their scheme and only kept the NP-COORD label for constituents inside a coordination structure. Figure FIGREF26 illustrates the detailed annotation differences between two datasets. We tested both models on all the experiments presented in Sections SECREF3-SECREF6 above.", + "Turning to the results of these six experiments: We see little difference between the two models in the Non-coordination agreement experiment. For the Complex coordination control and Complex coordination critical experiments, both models are largely the same as well. However, in the Simple and-coordination and Simple or-coordination experiments the values for all conditions are shifted upwards for the RNNG PTB-coord model, indicating higher over-all preference for the plural continuation. Furthermore, the range of values is reduced in the RNNG PTB-coord model, compared to the RNNG PTB-control model. These results indicate that adding an explicit COORD phrasal label does not drastically change model performance: Both models still appear to be using a linear combination of number features to drive plural vs. singular expectation. However, the explicit representation has made the interior of the coordination phrase more opaque to the model (each feature matters less) and has slightly shifted model preference towards plural continuations. In this sense, the PTB-coord model may have learned a generalization about CoordNPs, but this generalization remains unlike the ones learned by humans." + ], + [ + "We present statistics of subject/predicate agreement patterns in the Penn Treebank (PTB) and French Treebank (FTB) in Table TABREF28 and TABREF29." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0574/instruction.md b/qasper-0574/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4e7a9ce7b17c19825fb02966f51229dbcf9be55a --- /dev/null +++ b/qasper-0574/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Representation of Constituents in Neural Language Models: Coordination Phrase as a Case Study + +Question: What are the baseline models? \ No newline at end of file diff --git a/qasper-0587/instruction.md b/qasper-0587/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..1156a4877a6a9030a9e8374c134c98c02d845f3e --- /dev/null +++ b/qasper-0587/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: How Far are We from Effective Context Modeling ? An Exploratory Study on Semantic Parsing in Context + +Question: What are two datasets models are tested on? \ No newline at end of file diff --git a/qasper-0588/instruction.md b/qasper-0588/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..e3b7b3d4890bed4e1fd11426a544587c1c7f333f --- /dev/null +++ b/qasper-0588/instruction.md @@ -0,0 +1,149 @@ +Name of Paper: A Novel Aspect-Guided Deep Transition Model for Aspect Based Sentiment Analysis + +Question: How big is the improvement over the state-of-the-art results? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Model Description", + "Model Description ::: Aspect-Guided Encoder", + "Model Description ::: Aspect-Reconstruction", + "Model Description ::: Training Objective", + "Experiments ::: Datasets and Metrics ::: Data Preparation.", + "Experiments ::: Datasets and Metrics ::: Aspect-Category Sentiment Analysis.", + "Experiments ::: Datasets and Metrics ::: Aspect-Term Sentiment Analysis.", + "Experiments ::: Datasets and Metrics ::: Metrics.", + "Experiments ::: Implementation Details", + "Experiments ::: Baselines", + "Experiments ::: Main Results and Analysis ::: Aspect-Category Sentiment Analysis Task", + "Experiments ::: Main Results and Analysis ::: Aspect-Term Sentiment Analysis Task", + "Experiments ::: Main Results and Analysis ::: Ablation Study", + "Experiments ::: Main Results and Analysis ::: Impact of Model Depth", + "Experiments ::: Main Results and Analysis ::: Effectiveness of Aspect-reconstruction Approach", + "Experiments ::: Main Results and Analysis ::: Impact of Loss Weight @!START@$\\lambda $@!END@", + "Experiments ::: Main Results and Analysis ::: Comparison on Three-Class for the Aspect-Term Sentiment Analysis Task", + "Analysis and Discussion ::: Case Study and Visualization.", + "Analysis and Discussion ::: Error Analysis.", + "Related Work ::: Sentiment Analysis.", + "Related Work ::: Deep Transition.", + "Conclusions", + "Acknowledgments" + ], + "paragraphs": [ + [ + "Aspect based sentiment analysis (ABSA) is a fine-grained task in sentiment analysis, which can provide important sentiment information for other natural language processing (NLP) tasks. There are two different subtasks in ABSA, namely, aspect-category sentiment analysis and aspect-term sentiment analysis BIBREF0, BIBREF1. Aspect-category sentiment analysis aims at predicting the sentiment polarity towards the given aspect, which is in predefined several categories and it may not appear in the sentence. For instance, in Table TABREF2, the aspect-category sentiment analysis is going to predict the sentiment polarity towards the aspect \u201cfood\u201d, which is not appeared in the sentence. By contrast, the goal of aspect-term sentiment analysis is to predict the sentiment polarity over the aspect term which is a subsequence of the sentence. For instance, the aspect-term sentiment analysis will predict the sentiment polarity towards the aspect term \u201cThe appetizers\u201d, which is a subsequence of the sentence. Additionally, the number of categories of the aspect term is more than one thousand in the training corpus.", + "As shown in Table TABREF2, sentiment polarity may be different when different aspects are considered. Thus, the given aspect (term) is crucial to ABSA tasks BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6. Besides, BIBREF7 show that not all words of a sentence are useful for the sentiment prediction towards a given aspect (term). For instance, when the given aspect is the \u201cservice\u201d, the words \u201cappetizers\u201d and \u201cok\u201d are irrelevant for the sentiment prediction. Therefore, an aspect-independent (weakly associative) encoder may encode such background words (e.g., \u201cappetizers\u201d and \u201cok\u201d) into the final representation, which may lead to an incorrect prediction.", + "Numerous existing models BIBREF8, BIBREF9, BIBREF10, BIBREF1 typically utilize an aspect-independent encoder to generate the sentence representation, and then apply the attention mechanism BIBREF11 or gating mechanism to conduct feature selection and extraction, while feature selection and extraction may base on noised representations. In addition, some models BIBREF12, BIBREF13, BIBREF14 simply concatenate the aspect embedding with each word embedding of the sentence, and then leverage conventional Long Short-Term Memories (LSTMs) BIBREF15 to generate the sentence representation. However, it is insufficient to exploit the given aspect and conduct potentially complex feature selection and extraction.", + "To address this issue, we investigate a novel architecture to enhance the capability of feature selection and extraction with the guidance of the given aspect from scratch. Based on the deep transition Gated Recurrent Unit (GRU) BIBREF16, BIBREF17, BIBREF18, BIBREF19, an aspect-guided GRU encoder is thus proposed, which utilizes the given aspect to guide the sentence encoding procedure at the very beginning stage. In particular, we specially design an aspect-gate for the deep transition GRU to control the information flow of each token input, with the aim of guiding feature selection and extraction from scratch, i.e. sentence representation generation. Furthermore, we design an aspect-oriented objective to enforce our model to reconstruct the given aspect, with the sentence representation generated by the aspect-guided encoder. We name this Aspect-Guided Deep Transition model as AGDT. With all the above contributions, our AGDT can accurately generate an aspect-specific representation for a sentence, and thus conduct more accurate sentiment predictions towards the given aspect.", + "We evaluate the AGDT on multiple datasets of two subtasks in ABSA. Experimental results demonstrate the effectiveness of our proposed approach. And the AGDT significantly surpasses existing models with the same setting and achieves state-of-the-art performance among the models without using additional features (e.g., BERT BIBREF20). Moreover, we also provide empirical and visualization analysis to reveal the advantages of our model. Our contributions can be summarized as follows:", + "We propose an aspect-guided encoder, which utilizes the given aspect to guide the encoding of a sentence from scratch, in order to conduct the aspect-specific feature selection and extraction at the very beginning stage.", + "We propose an aspect-reconstruction approach to further guarantee that the aspect-specific information has been fully embedded into the sentence representation.", + "Our AGDT substantially outperforms previous systems with the same setting, and achieves state-of-the-art results on benchmark datasets compared to those models without leveraging additional features (e.g., BERT)." + ], + [ + "As shown in Figure FIGREF6, the AGDT model mainly consists of three parts: aspect-guided encoder, aspect-reconstruction and aspect concatenated embedding. The aspect-guided encoder is specially designed to guide the encoding of a sentence from scratch for conducting the aspect-specific feature selection and extraction at the very beginning stage. The aspect-reconstruction aims to guarantee that the aspect-specific information has been fully embedded in the sentence representation for more accurate predictions. The aspect concatenated embedding part is used to concatenate the aspect embedding and the generated sentence representation so as to make the final prediction." + ], + [ + "The aspect-guided encoder is the core module of AGDT, which consists of two key components: Aspect-guided GRU and Transition GRU BIBREF16.", + "A-GRU: Aspect-guided GRU (A-GRU) is a specially-designed unit for the ABSA tasks, which is an extension of the L-GRU proposed by BIBREF19. In particular, we design an aspect-gate to select aspect-specific representations through controlling the transformation scale of token embeddings at each time step.", + "At time step $t$, the hidden state $\\mathbf {h}_{t}$ is computed as follows:", + "where $\\odot $ represents element-wise product; $\\mathbf {z}_{t}$ is the update gate BIBREF16; and $\\widetilde{\\mathbf {h}}_{t}$ is the candidate activation, which is computed as:", + "where $\\mathbf {g}_{t}$ denotes the aspect-gate; $\\mathbf {x}_{t}$ represents the input word embedding at time step $t$; $\\mathbf {r}_{t}$ is the reset gate BIBREF16; $\\textbf {H}_1(\\mathbf {x}_{t})$ and $\\textbf {H}_2(\\mathbf {x}_{t})$ are the linear transformation of the input $\\mathbf {x}_{t}$, and $\\mathbf {l}_{t}$ is the linear transformation gate for $\\mathbf {x}_{t}$ BIBREF19. $\\mathbf {r}_{t}$, $\\mathbf {z}_{t}$, $\\mathbf {l}_{t}$, $\\mathbf {g}_{t}$, $\\textbf {H}_{1}(\\mathbf {x}_{t})$ and $\\textbf {H}_{2}(\\mathbf {x}_{t})$ are computed as:", + "where \u201c$\\mathbf {a}$\" denotes the embedding of the given aspect, which is the same at each time step. The update gate $\\mathbf {z}_t$ and reset gate $\\mathbf {r}_t$ are the same as them in the conventional GRU.", + "In Eq. (DISPLAY_FORM9) $\\sim $ (), the aspect-gate $\\mathbf {g}_{t}$ controls both nonlinear and linear transformations of the input $\\mathbf {x}_{t}$ under the guidance of the given aspect at each time step. Besides, we also exploit a linear transformation gate $\\mathbf {l}_{t}$ to control the linear transformation of the input, according to the current input $\\mathbf {x}_t$ and previous hidden state $\\mathbf {h}_{t-1}$, which has been proved powerful in the deep transition architecture BIBREF19.", + "As a consequence, A-GRU can control both non-linear transformation and linear transformation for input $\\mathbf {x}_{t}$ at each time step, with the guidance of the given aspect, i.e., A-GRU can guide the encoding of aspect-specific features and block the aspect-irrelevant information at the very beginning stage.", + "T-GRU: Transition GRU (T-GRU) BIBREF17 is a crucial component of deep transition block, which is a special case of GRU with only \u201cstate\u201d as an input, namely its input embedding is zero embedding. As in Figure FIGREF6, a deep transition block consists of an A-GRU followed by several T-GRUs at each time step. For the current time step $t$, the output of one A-GRU/T-GRU is fed into the next T-GRU as the input. The output of the last T-GRU at time step $t$ is fed into A-GRU at the time step $t+1$. For a T-GRU, each hidden state at both time step $t$ and transition depth $i$ is computed as:", + "where the update gate $\\mathbf {z}_{t}^i$ and the reset gate $\\mathbf {r}_{t}^i$ are computed as:", + "The AGDT encoder is based on deep transition cells, where each cell is composed of one A-GRU at the bottom, followed by several T-GRUs. Such AGDT model can encode the sentence representation with the guidance of aspect information by utilizing the specially designed architecture." + ], + [ + "We propose an aspect-reconstruction approach to guarantee the aspect-specific information has been fully embedded in the sentence representation. Particularly, we devise two objectives for two subtasks in ABSA respectively. In terms of aspect-category sentiment analysis datasets, there are only several predefined aspect categories. While in aspect-term sentiment analysis datasets, the number of categories of term is more than one thousand. In a real-life scenario, the number of term is infinite, while the words that make up terms are limited. Thus we design different loss-functions for these two scenarios.", + "For the aspect-category sentiment analysis task, we aim to reconstruct the aspect according to the aspect-specific representation. It is a multi-class problem. We take the softmax cross-entropy as the loss function:", + "where C1 is the number of predefined aspects in the training example; ${y}_{i}^{c}$ is the ground-truth and ${p}_{i}^{c}$ is the estimated probability of a aspect.", + "For the aspect-term sentiment analysis task, we intend to reconstruct the aspect term (may consist of multiple words) according to the aspect-specific representation. It is a multi-label problem and thus the sigmoid cross-entropy is applied:", + "where C2 denotes the number of words that constitute all terms in the training example, ${y}_{i}^{t}$ is the ground-truth and ${p}_{i}^{t}$ represents the predicted value of a word.", + "Our aspect-oriented objective consists of $\\mathcal {L}_{c}$ and $\\mathcal {L}_{t}$, which guarantee that the aspect-specific information has been fully embedded into the sentence representation." + ], + [ + "The final loss function is as follows:", + "where the underlined part denotes the conventional loss function; C is the number of sentiment labels; ${y}_{i}$ is the ground-truth and ${p}_{i}$ represents the estimated probability of the sentiment label; $\\mathcal {L}$ is the aspect-oriented objective, where Eq. DISPLAY_FORM14 is for the aspect-category sentiment analysis task and Eq. DISPLAY_FORM15 is for the aspect-term sentiment analysis task. And $\\lambda $ is the weight of $\\mathcal {L}$.", + "As shown in Figure FIGREF6, we employ the aspect reconstruction approach to reconstruct the aspect (term), where \u201csoftmax\u201d is for the aspect-category sentiment analysis task and \u201csigmoid\u201d is for the aspect-term sentiment analysis task. Additionally, we concatenate the aspect embedding on the aspect-guided sentence representation to predict the sentiment polarity. Under that loss function (Eq. DISPLAY_FORM17), the AGDT can produce aspect-specific sentence representations." + ], + [ + "We conduct experiments on two datasets of the aspect-category based task and two datasets of the aspect-term based task. For these four datasets, we name the full dataset as \u201cDS\". In each \u201cDS\", there are some sentences like the example in Table TABREF2, containing different sentiment labels, each of which associates with an aspect (term). For instance, Table TABREF2 shows the customer's different attitude towards two aspects: \u201cfood\u201d (\u201cThe appetizers\") and \u201cservice\u201d. In order to measure whether a model can detect different sentiment polarities in one sentence towards different aspects, we extract a hard dataset from each \u201cDS\u201d, named \u201cHDS\u201d, in which each sentence only has different sentiment labels associated with different aspects. When processing the original sentence $s$ that has multiple aspects ${a}_{1},{a}_{2},...,{a}_{n}$ and corresponding sentiment labels ${l}_{1},{l}_{2},...,{l}_{n}$ ($n$ is the number of aspects or terms in a sentence), the sentence will be expanded into (s, ${a}_{1}$, ${l}_{1}$), (s, ${a}_{2}$, ${l}_{2}$), ..., (s, ${a}_{n}$, ${l}_{n}$) in each dataset BIBREF21, BIBREF22, BIBREF1, i.e, there will be $n$ duplicated sentences associated with different aspects and labels." + ], + [ + "For comparison, we follow BIBREF1 and use the restaurant reviews dataset of SemEval 2014 (\u201crestaurant-14\u201d) Task 4 BIBREF0 to evaluate our AGDT model. The dataset contains five predefined aspects and four sentiment labels. A large dataset (\u201crestaurant-large\u201d) involves restaurant reviews of three years, i.e., 2014 $\\sim $ 2016 BIBREF0. There are eight predefined aspects and three labels in that dataset. When creating the \u201crestaurant-large\u201d dataset, we follow the same procedure as in BIBREF1. Statistics of datasets are shown in Table TABREF19." + ], + [ + "We use the restaurant and laptop review datasets of SemEval 2014 Task 4 BIBREF0 to evaluate our model. Both datasets contain four sentiment labels. Meanwhile, we also conduct a three-class experiment, in order to compare with some work BIBREF13, BIBREF3, BIBREF7 which removed \u201cconflict\u201d labels. Statistics of both datasets are shown in Table TABREF20." + ], + [ + "The evaluation metrics are accuracy. All instances are shown in Table TABREF19 and Table TABREF20. Each experiment is repeated five times. The mean and the standard deviation are reported." + ], + [ + "We use the pre-trained 300d Glove embeddings BIBREF23 to initialize word embeddings, which is fixed in all models. For out-of-vocabulary words, we randomly sample their embeddings by the uniform distribution $U(-0.25, 0.25)$. Following BIBREF8, BIBREF24, BIBREF25, we take the averaged word embedding as the aspect representation for multi-word aspect terms. The transition depth of deep transition model is 4 (see Section SECREF30). The hidden size is set to 300. We set the dropout rate BIBREF26 to 0.5 for input token embeddings and 0.3 for hidden states. All models are optimized using Adam optimizer BIBREF27 with gradient clipping equals to 5 BIBREF28. The initial learning rate is set to 0.01 and the batch size is set to 4096 at the token level. The weight of the reconstruction loss $\\lambda $ in Eq. DISPLAY_FORM17 is fine-tuned (see Section SECREF30) and respectively set to 0.4, 0.4, 0.2 and 0.5 for four datasets." + ], + [ + "To comprehensively evaluate our AGDT, we compare the AGDT with several competitive models.", + "ATAE-LSTM. It is an attention-based LSTM model. It appends the given aspect embedding with each word embedding, and then the concatenated embedding is taken as the input of LSTM. The output of LSTM is appended aspect embedding again. Furthermore, attention is applied to extract features for final predictions.", + "CNN. This model focuses on extracting n-gram features to generate sentence representation for the sentiment classification.", + "TD-LSTM. This model uses two LSTMs to capture the left and right context of the term to generate target-dependent representations for the sentiment prediction.", + "IAN. This model employs two LSTMs and interactive attention mechanism to learn representations of the sentence and the aspect, and concatenates them for the sentiment prediction.", + "RAM. This model applies multiple attentions and memory networks to produce the sentence representation.", + "GCAE. It uses CNNs to extract features and then employs two Gated Tanh-Relu units to selectively output the sentiment information flow towards the aspect for predicting sentiment labels." + ], + [ + "We present the overall performance of our model and baseline models in Table TABREF27. Results show that our AGDT outperforms all baseline models on both \u201crestaurant-14\u201d and \u201crestaurant-large\u201d datasets. ATAE-LSTM employs an aspect-weakly associative encoder to generate the aspect-specific sentence representation by simply concatenating the aspect, which is insufficient to exploit the given aspect. Although GCAE incorporates the gating mechanism to control the sentiment information flow according to the given aspect, the information flow is generated by an aspect-independent encoder. Compared with GCAE, our AGDT improves the performance by 2.4% and 1.6% in the \u201cDS\u201d part of the two dataset, respectively. These results demonstrate that our AGDT can sufficiently exploit the given aspect to generate the aspect-guided sentence representation, and thus conduct accurate sentiment prediction. Our model benefits from the following aspects. First, our AGDT utilizes an aspect-guided encoder, which leverages the given aspect to guide the sentence encoding from scratch and generates the aspect-guided representation. Second, the AGDT guarantees that the aspect-specific information has been fully embedded in the sentence representation via reconstructing the given aspect. Third, the given aspect embedding is concatenated on the aspect-guided sentence representation for final predictions.", + "The \u201cHDS\u201d, which is designed to measure whether a model can detect different sentiment polarities in a sentence, consists of replicated sentences with different sentiments towards multiple aspects. Our AGDT surpasses GCAE by a very large margin (+11.4% and +4.9% respectively) on both datasets. This indicates that the given aspect information is very pivotal to the accurate sentiment prediction, especially when the sentence has different sentiment labels, which is consistent with existing work BIBREF2, BIBREF3, BIBREF4. Those results demonstrate the effectiveness of our model and suggest that our AGDT has better ability to distinguish the different sentiments of multiple aspects compared to GCAE." + ], + [ + "As shown in Table TABREF28, our AGDT consistently outperforms all compared methods on both domains. In this task, TD-LSTM and ATAE-LSTM use a aspect-weakly associative encoder. IAN, RAM and GCAE employ an aspect-independent encoder. In the \u201cDS\u201d part, our AGDT model surpasses all baseline models, which shows that the inclusion of A-GRU (aspect-guided encoder), aspect-reconstruction and aspect concatenated embedding has an overall positive impact on the classification process.", + "In the \u201cHDS\u201d part, the AGDT model obtains +3.6% higher accuracy than GCAE on the restaurant domain and +4.2% higher accuracy on the laptop domain, which shows that our AGDT has stronger ability for the multi-sentiment problem against GCAE. These results further demonstrate that our model works well across tasks and datasets." + ], + [ + "We conduct ablation experiments to investigate the impacts of each part in AGDT, where the GRU is stacked with 4 layers. Here \u201cAC\u201d represents aspect concatenated embedding , \u201cAG\u201d stands for A-GRU (Eq. (DISPLAY_FORM8) $\\sim $ ()) and \u201cAR\u201d denotes the aspect-reconstruction (Eq. (DISPLAY_FORM14) $\\sim $ (DISPLAY_FORM17)).", + "From Table TABREF31 and Table TABREF32, we can conclude:", + "Deep Transition (DT) achieves superior performances than GRU, which is consistent with previous work BIBREF18, BIBREF19 (2 vs. 1).", + "Utilizing \u201cAG\u201d to guide encoding aspect-related features from scratch has a significant impact for highly competitive results and particularly in the \u201cHDS\u201d part, which demonstrates that it has the stronger ability to identify different sentiment polarities towards different aspects. (3 vs. 2).", + "Aspect concatenated embedding can promote the accuracy to a degree (4 vs. 3).", + "The aspect-reconstruction approach (\u201cAR\u201d) substantially improves the performance, especially in the \u201cHDS\" part (5 vs. 4).", + "the results in 6 show that all modules have an overall positive impact on the sentiment classification." + ], + [ + "We have demonstrated the effectiveness of the AGDT. Here, we investigate the impact of model depth of AGDT, varying the depth from 1 to 6. Table TABREF39 shows the change of accuracy on the test sets as depth increases. We find that the best results can be obtained when the depth is equal to 4 at most case, and further depth do not provide considerable performance improvement." + ], + [ + "Here, we investigate how well the AGDT can reconstruct the aspect information. For the aspect-term reconstruction, we count the construction is correct when all words of the term are reconstructed. Table TABREF40 shows all results on four test datasets, which shows the effectiveness of aspect-reconstruction approach again." + ], + [ + "We randomly sample a temporary development set from the \u201cHDS\" part of the training set to choose the lambda for each dataset. And we investigate the impact of $\\lambda $ for aspect-oriented objectives. Specifically, $\\lambda $ is increased from 0.1 to 1.0. Figure FIGREF33 illustrates all results on four \u201cHDS\" datasets, which show that reconstructing the given aspect can enhance aspect-specific sentiment features and thus obtain better performances." + ], + [ + "We also conduct a three-class experiment to compare our AGDT with previous models, i.e., IARM, TNet, VAE, PBAN, AOA and MGAN, in Table TABREF41. These previous models are based on an aspect-independent (weakly associative) encoder to generate sentence representations. Results on all domains suggest that our AGDT substantially outperforms most competitive models, except for the TNet on the laptop dataset. The reason may be TNet incorporates additional features (e.g., position features, local ngrams and word-level features) compared to ours (only word-level features)." + ], + [ + "To give an intuitive understanding of how the proposed A-GRU works from scratch with different aspects, we take a review sentence as an example. As the example \u201cthe appetizers are ok, but the service is slow.\u201d shown in Table TABREF2, it has different sentiment labels towards different aspects. The color depth denotes the semantic relatedness level between the given aspect and each word. More depth means stronger relation to the given aspect.", + "Figure FIGREF43 shows that the A-GRU can effectively guide encoding the aspect-related features with the given aspect and identify corresponding sentiment. In another case, \u201coverpriced Japanese food with mediocre service.\u201d, there are two extremely strong sentiment words. As the above of Figure FIGREF44 shows, our A-GRU generates almost the same weight to the word \u201coverpriced\u201d and \u201cmediocre\u201d. The bottom of Figure FIGREF44 shows that reconstructing the given aspect can effectively enhance aspect-specific sentiment features and produce correct sentiment predictions." + ], + [ + "We further investigate the errors from AGDT, which can be roughly divided into 3 types. 1) The decision boundary among the sentiment polarity is unclear, even the annotators can not sure what sentiment orientation over the given aspect in the sentence. 2) The \u201cconflict/neutral\u201d instances are extremely easily misclassified as \u201cpositive\u201d or \u201cnegative\u201d, due to the imbalanced label distribution in training corpus. 3) The polarity of complex instances is hard to predict, such as the sentence that express subtle emotions, which are hardly effectively captured, or containing negation words (e.g., never, less and not), which easily affect the sentiment polarity." + ], + [ + "There are kinds of sentiment analysis tasks, such as document-level BIBREF34, sentence-level BIBREF35, BIBREF36, aspect-level BIBREF0, BIBREF37 and multimodal BIBREF38, BIBREF39 sentiment analysis. For the aspect-level sentiment analysis, previous work typically apply attention mechanism BIBREF11 combining with memory network BIBREF40 or gating units to solve this task BIBREF8, BIBREF41, BIBREF42, BIBREF1, BIBREF43, BIBREF44, BIBREF45, BIBREF46, where an aspect-independent encoder is used to generate the sentence representation. In addition, some work leverage the aspect-weakly associative encoder to generate aspect-specific sentence representation BIBREF12, BIBREF13, BIBREF14. All of these methods make insufficient use of the given aspect information. There are also some work which jointly extract the aspect term (and opinion term) and predict its sentiment polarity BIBREF47, BIBREF48, BIBREF49, BIBREF50, BIBREF51, BIBREF52, BIBREF53, BIBREF54, BIBREF55. In this paper, we focus on the latter problem and leave aspect extraction BIBREF56 to future work. And some work BIBREF57, BIBREF58, BIBREF59, BIBREF30, BIBREF60, BIBREF51 employ the well-known BERT BIBREF20 or document-level corpora to enhance ABSA tasks, which will be considered in our future work to further improve the performance." + ], + [ + "Deep transition has been proved its superiority in language modeling BIBREF17 and machine translation BIBREF18, BIBREF19. We follow the deep transition architecture in BIBREF19 and extend it by incorporating a novel A-GRU for ABSA tasks." + ], + [ + "In this paper, we propose a novel aspect-guided encoder (AGDT) for ABSA tasks, based on a deep transition architecture. Our AGDT can guide the sentence encoding from scratch for the aspect-specific feature selection and extraction. Furthermore, we design an aspect-reconstruction approach to enforce AGDT to reconstruct the given aspect with the generated sentence representation. Empirical studies on four datasets suggest that the AGDT outperforms existing state-of-the-art models substantially on both aspect-category sentiment analysis task and aspect-term sentiment analysis task of ABSA without additional features." + ], + [ + "We sincerely thank the anonymous reviewers for their thorough reviewing and insightful suggestions. Liang, Xu, and Chen are supported by the National Natural Science Foundation of China (Contract 61370130, 61976015, 61473294 and 61876198), and the Beijing Municipal Natural Science Foundation (Contract 4172047), and the International Science and Technology Cooperation Program of the Ministry of Science and Technology (K11F100010)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0722/instruction.md b/qasper-0722/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..825ab217013431f08493760056cd0f4d52633e9a --- /dev/null +++ b/qasper-0722/instruction.md @@ -0,0 +1,97 @@ +Name of Paper: Using word embeddings to improve the discriminability of co-occurrence text networks + +Question: What other natural processing tasks authors think could be studied by using word embeddings? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related works", + "Material and Methods", + "Results and Discussion", + "Results and Discussion ::: Performance analysis", + "Results and Discussion ::: Effects of considering stopwords and local thresholding", + "Conclusion", + "Acknowledgments", + "Supplementary Information ::: Stopwords", + "Supplementary Information ::: List of books", + "Supplementary Information ::: Additional results" + ], + "paragraphs": [ + [ + "The ability to construct complex and diverse linguistic structures is one of the main features that set us apart from all other species. Despite its ubiquity, some language aspects remain unknown. Topics such as language origin and evolution have been studied by researchers from diverse disciplines, including Linguistic, Computer Science, Physics and Mathematics BIBREF0, BIBREF1, BIBREF2. In order to better understand the underlying language mechanisms and universal linguistic properties, several models have been developed BIBREF3, BIBREF4. A particular language representation regards texts as complex systems BIBREF5. Written texts can be considered as complex networks (or graphs), where nodes could represent syllables, words, sentences, paragraphs or even larger chunks BIBREF5. In such models, network edges represent the proximity between nodes, e.g. the frequency of the co-occurrence of words. Several interesting results have been obtained from networked models, such as the explanation of Zipf's Law as a consequence of the least effort principle and theories on the nature of syntactical relationships BIBREF6, BIBREF7.", + "In a more practical scenario, text networks have been used in text classification tasks BIBREF8, BIBREF9, BIBREF10. The main advantage of the model is that it does not rely on deep semantical information to obtain competitive results. Another advantage of graph-based approaches is that, when combined with other approaches, it yields competitive results BIBREF11. A simple, yet recurrent text model is the well-known word co-occurrence network. After optional textual pre-processing steps, in a co-occurrence network each different word becomes a node and edges are established via co-occurrence in a desired window. A common strategy connects only adjacent words in the so called word adjacency networks.", + "While the co-occurrence representation yields good results in classification scenarios, some important features are not considered in the model. For example, long-range syntactical links, though less frequent than adjacent syntactical relationships, might be disregarded from a simple word adjacency approach BIBREF12. In addition, semantically similar words not sharing the same lemma are mapped into distinct nodes. In order to address these issues, here we introduce a modification of the traditional network representation by establishing additional edges, referred to as \u201cvirtual\u201d edges. In the proposed model, in addition to the co-occurrence edges, we link two nodes (words) if the corresponding word embedding representation is similar. While this approach still does not merge similar nodes into the same concept, similar nodes are explicitly linked via virtual edges.", + "Our main objective here is to evaluate whether such an approach is able to improve the discriminability of word co-occurrence networks in a typical text network classification task. We evaluate the methodology for different embedding techniques, including GloVe, Word2Vec and FastText. We also investigated different thresholding strategies to establish virtual links. Our results revealed, as a proof of principle, that the proposed approach is able to improve the discriminability of the classification when compared to the traditional co-occurrence network. While the gain in performance depended upon the text length being considered, we found relevant gains for intermediary text lengths. Additional results also revealed that a simple thresholding strategy combined with the use of stopwords tends to yield the best results.", + "We believe that the proposed representation could be applied in other text classification tasks, which could lead to potential gains in performance. Because the inclusion of virtual edges is a simple technique to make the network denser, such an approach can benefit networked representations with a limited number of nodes and edges. This representation could also shed light into language mechanisms in theoretical studies relying on the representation of text as complex networks. Potential novel research lines leveraging the adopted approach to improve the characterization of texts in other applications are presented in the conclusion." + ], + [ + "Complex networks have been used in a wide range of fields, including in Social Sciences BIBREF13, Neuroscience BIBREF14, Biology BIBREF15, Scientometry BIBREF16 and Pattern Recognition BIBREF17, BIBREF18, BIBREF19, BIBREF20. In text analysis, networks are used to uncover language patterns, including the origins of the ever present Zipf's Law BIBREF21 and the analysis of linguistic properties of natural and unknown texts BIBREF22, BIBREF23. Applications of network science in text mining and text classification encompasses applications in semantic analysis BIBREF24, BIBREF25, BIBREF26, BIBREF27, authorship attribution BIBREF28, BIBREF29 and stylometry BIBREF28, BIBREF30, BIBREF31. Here we focus in the stylometric analysis of texts using complex networks.", + "In BIBREF28, the authors used a co-occurrence network to study a corpus of English and Polish books. They considered a dataset of 48 novels, which were written by 8 different authors. Differently from traditional co-occurrence networks, some punctuation marks were considered as words when mapping texts as networks. The authors also decided to create a methodology to normalize the obtained network metrics, since they considered documents with variations in length. A similar approach was adopted in a similar study BIBREF32, with a focus on comparing novel measurements and measuring the effect of considering stopwords in the network structure.", + "A different approach to analyze co-occurrence networks was devised in BIBREF33. Whilst most approaches only considered traditional network measurements or devised novel topological and dynamical measurements, the authors combined networked and semantic information to improve the performance of network-based classification. Interesting, the combined use of network motifs and node labels (representing the corresponding words) allowed an improvement in performance in the considered task. A similar combination of techniques using a hybrid approach was proposed in BIBREF8. Networked-based approaches has also been applied to the authorship recognition tasks in other languages, including Persian texts BIBREF9.", + "Co-occurrence networks have been used in other contexts other than stylometric analysis. The main advantage of this approach is illustrated in the task aimed at diagnosing diseases via text analysis BIBREF11. Because the topological analysis of co-occurrence language networks do not require deep semantic analysis, this model is able to model text created by patients suffering from cognitive impairment BIBREF11. Recently, it has been shown that the combination of network and traditional features could be used to improve the diagnosis of patients with cognitive impairment BIBREF11. Interestingly, this was one of the first approaches suggesting the use of embeddings to address the particular problem of lack of statistics to create a co-occurrence network in short documents BIBREF34.", + "While many of the works dealing with word co-occurrence networks have been proposed in the last few years, no systematic study of the effects of including information from word embeddings in such networks has been analyzed. This work studies how links created via embeddings information modify the underlying structure of networks and, most importantly, how it can improve the model to provide improved classification performance in the stylometry task." + ], + [ + "To represent texts as networks, we used the so-called word adjacency network representation BIBREF35, BIBREF28, BIBREF32. Typically, before creating the networks, the text is pre-processed. An optional pre-processing step is the removal of stopwords. This step is optional because such words include mostly article and prepositions, which may be artlessly represented by network edges. However, in some applications \u2013 including the authorship attribution task \u2013 stopwords (or function words) play an important role in the stylistic characterization of texts BIBREF32. A list of stopwords considered in this study is available in the Supplementary Information.", + "The pre-processing step may also include a lemmatization procedure. This step aims at mapping words conveying the same meaning into the same node. In the lemmatization process, nouns and verbs are mapped into their singular and infinite forms. Note that, while this step is useful to merge words sharing a lemma into the same node, more complex semantical relationships are overlooked. For example, if \u201ccar\u201d and \u201cvehicle\u201d co-occur in the same text, they are considered as distinct nodes, which may result in an inaccurate representation of the text.", + "Such a drawback is addressed by including \u201cvirtual\u201d edges connecting nodes. In other words, even if two words are not adjacent in the text, we include \u201cvirtual\u201d edges to indicate that two distant words are semantically related. The inclusion of such virtual edges is illustrated in Figure FIGREF1. In order to measure the semantical similarity between two concepts, we use the concept of word embeddings BIBREF36, BIBREF37. Thus, each word is represented using a vector representation encoding the semantical and contextual characteristics of the word. Several interesting properties have been obtained from distributed representation of words. One particular property encoded in the embeddings representation is the fact the semantical similarity between concepts is proportional to the similarity of vectors representing the words. Similarly to several other works, here we measure the similarity of the vectors via cosine similarity BIBREF38.", + "The following strategies to create word embedding were considered in this paper:", + "GloVe: the Global Vectors (GloVe) algorithm is an extension of the Word2vec model BIBREF39 for efficient word vector learning BIBREF40. This approach combines global statistics from matrix factorization techniques (such as latent semantic analysis) with context-based and predictive methods like Word2Vec. This method is called as Global Vector method because the global corpus statistics are captured by GloVe. Instead of using a window to define the local context, GloVe constructs an explicit word-context matrix (or co-occurrence matrix) using statistics across the entire corpus. The final result is a learning model that oftentimes yields better word vector representations BIBREF40.", + "Word2Vec: this is a predictive model that finds dense vector representations of words using a three-layer neural network with a single hidden layer BIBREF39. It can be defined in a two-fold way: continuous bag-of-words and skip-gram model. In the latter, the model analyzes the words of a set of sentences (or corpus) and attempts to predict the neighbors of such words. For example, taking as reference the word \u201cRobin\u201d, the model decides that \u201cHood\u201d is more likely to follow the reference word than any other word. The vectors are obtained as follows: given the vocabulary (generated from all corpus words), the model trains a neural network with the sentences of the corpus. Then, for a given word, the probabilities that each word follows the reference word are obtained. Once the neural network is trained, the weights of the hidden layer are used as vectors of each corpus word.", + "FastText: this method is another extension of the Word2Vec model BIBREF41. Unlike Word2Vec, FastText represents each word as a bag of character n-grams. Therefore, the neural network not only trains individual words, but also several n-grams of such words. The vector for a word is the sum of vectors obtained for the character n-grams composing the word. For example, the embedding obtained for the word \u201ccomputer\u201d with $n\\le 3$ is the sum of the embeddings obtained for \u201cco\u201d, \u201ccom\u201d, \u201comp\u201d, \u201cmpu\u201d, \u201cput\u201d, \u201cute\u201d, \u201cter\u201d and \u201cer\u201d. In this way, this method obtains improved representations for rare words, since n-grams composing rare words might be present in other words. The FastText representation also allows the model to understand suffixes and prefixes. Another advantage of FastText is its efficiency to be trained in very large corpora.", + "Concerning the thresholding process, we considered two main strategies. First, we used a global strategy: in addition to the co-occurrence links (continuous lines in Figure FIGREF1), only \u201cvirtual\u201d edges stronger than a given threshold are left in the network. Thus only the most similar concepts are connected via virtual links. This strategy is hereafter referred to as global strategy. Unfortunately, this method may introduce an undesired bias towards hubs BIBREF42.", + "To overcome the potential disadvantages of the global thresholding method, we also considered a more refined thresholding approach that takes into account the local structure to decide whether a weighted link is statistically significant BIBREF42. This method relies on the idea that the importance of an edge should be considered in the the context in which it appears. In other words, the relevance of an edge should be evaluated by analyzing the nodes connected to its ending points. Using the concept of disparity filter, the method devised in BIBREF42 defines a null model that quantifies the probability of a node to be connected to an edge with a given weight, based on its other connections. This probability is used to define the significance of the edge. The parameter that is used to measure the significance of an edge $e_{ij}$ is $\\alpha _{ij}$, defined as:", + "where $w_{ij}$ is the weight of the edge $e_{ij}$ and $k_i$ is the degree of the $i$-th node. The obtained network corresponds to the set of nodes and edges obtained by removing all edges with $\\alpha $ higher than the considered threshold. Note that while the similarity between co-occurrence links might be considered to compute $\\alpha _{ij}$, only \u201cvirtual\u201d edges (i.e. the dashed lines in Figure FIGREF1) are eligible to be removed from the network in the filtering step. This strategy is hereafter referred to as local strategy.", + "After co-occurrence networks are created and virtual edges are included, in the next step we used a characterization based on topological analysis. Because a global topological analysis is prone to variations in network size, we focused our analysis in the local characterization of complex networks. In a local topological analysis, we use as features the value of topological/dynamical measurements obtained for a set of words. In this case, we selected as feature the words occurring in all books of the dataset. For each word, we considered the following network measurements: degree, betweenness, clustering coefficient, average shortest path length, PageRank, concentric symmetry (at the second and third hierarchical level) BIBREF32 and accessibility BIBREF43, BIBREF44 (at the second and third hierarchical level). We chose these measurements because all of them capture some particular linguistic feature of texts BIBREF45, BIBREF46, BIBREF47, BIBREF48. After network measurements are extracted, they are used in machine learning algorithms. In our experiments, we considered Decision Trees (DT), nearest neighbors (kNN), Naive Bayes (NB) and Support Vector Machines (SVM). We used some heuristics to optimize classifier parameters. Such techniques are described in the literature BIBREF49. The accuracy of the pattern recognition methods were evaluated using cross-validation BIBREF50.", + "In summary, the methodology used in this paper encompasses the following steps:", + "Network construction: here texts are mapped into a co-occurrence networks. Some variations exists in the literature, however here we focused in the most usual variation, i.e. the possibility of considering or disregarding stopwords. A network with co-occurrence links is obtained after this step.", + "Network enrichment: in this step, the network is enriched with virtual edges established via similarity of word embeddings. After this step, we are given a complete network with weighted links. Virtually, any embedding technique could be used to gauge the similarity between nodes.", + "Network filtering: in order to eliminate spurious links included in the last step, the weakest edges are filtered. Two approaches were considered: a simple approach based on a global threshold and a local thresholding strategy that preserves network community structure. The outcome of this network filtering step is a network with two types of links: co-occurrence and virtual links (as shown in Figure FIGREF1).", + "Feature extraction: In this step, topological and dynamical network features are extracted. Here, we do not discriminate co-occurrence from virtual edges to compute the network metrics.", + "Pattern classification: once features are extracted from complex networks, they are used in pattern classification methods. This might include supervised, unsupervised and semi-supervised classification. This framework is exemplified in the supervised scenario.", + "The above framework is exemplified with the most common technique(s). It should be noted that the methods used, however, can be replaced by similar techniques. For example, the network construction could consider stopwords or even punctuation marks BIBREF51. Another possibility is the use of different strategies of thresholding. While a systematic analysis of techniques and parameters is still required to reveal other potential advantages of the framework based on the addition of virtual edges, in this paper we provide a first analysis showing that virtual edges could be useful to improve the discriminability of texts modeled as complex networks.", + "Here we used a dataset compatible with datasets used recently in the literature (see e.g. BIBREF28, BIBREF10, BIBREF52). The objective of the studied stylometric task is to identify the authorship of an unknown document BIBREF53. All data and some statistics of each book are shown in the Supplementary Information." + ], + [ + "In Section SECREF13, we probe whether the inclusion of virtual edges is able to improve the performance of the traditional co-occurrence network-based classification in a usual stylometry task. While the focus of this paper is not to perform a systematic analysis of different methods comprising the adopted network, we consider two variations in the adopted methodology. In Section SECREF19, we consider the use of stopwords and the adoption of a local thresholding process to establish different criteria to create new virtual edges." + ], + [ + "In Figure FIGREF14, we show some of the improvements in performance obtained when including a fixed amount of virtual edges using GloVe as embedding method. In each subpanel, we show the relative improvement in performance obtained as a function of the fraction of additional edges. In this section, we considered the traditional co-occurrence as starting point. In other words, the network construction disregarded stopwords. The list of stopwords considered in this paper is available in the Supplementary Information. We also considered the global approach to filter edges.", + "The relative improvement in performance is given by $\\Gamma _+{(p)}/\\Gamma _0$, where $\\Gamma _+{(p)}$ is the accuracy rate obtained when $p\\%$ additional edges are included and $\\Gamma _0 = \\Gamma _+{(p=0)}$, i.e. $\\Gamma _0$ is the accuracy rate measured from the traditional co-occurrence model. We only show the highest relative improvements in performance for each classifier. In our analysis, we considered also samples of text with distinct length, since the performance of network-based methods is sensitive to text length BIBREF34. In this figure, we considered samples comprising $w=\\lbrace 1.0, 2.5, 5.0, 10.0\\rbrace $ thousand words.", + "The results obtained for GloVe show that the highest relative improvements in performance occur for decision trees. This is apparent specially for the shortest samples. For $w=1,000$ words, the decision tree accuracy is enhanced by a factor of almost 50% when $p=20\\%$. An excellent gain in performance is also observed for both Naive Bayes and SVM classifiers, when $p=18\\%$ and $p=12\\%$, respectively. When $w=2,500$ words, the highest improvements was observed for the decision tree algorithm. A minor improvement was observed for the kNN method. A similar behavior occurred for $w=5,000$ words. Interestingly, SVM seems to benefit from the use of additional edges when larger documents are considered. When only 5% virtual edges are included, the relative gain in performance is about 45%.", + "The relative gain in performance obtained for Word2vec is shown in Figure FIGREF15. Overall, once again decision trees obtained the highest gain in performance when short texts are considered. Similar to the analysis based on the GloVe method, the gain for kNN is low when compared to the benefit received by other methods. Here, a considerable gain for SVM in only clear for $w=2,500$ and $p=10\\%$. When large texts are considered, Naive Bayes obtained the largest gain in performance.", + "Finally, the relative gain in performance obtained for FastText is shown in Figure FIGREF16. The prominent role of virtual edges in decision tree algorithm in the classification of short texts once again is evident. Conversely, the classification of large documents using virtual edges mostly benefit the classification based on the Naive Bayes classifier. Similarly to the results observed for Glove and Word2vec, the gain in performance obtained for kNN is low compared when compared to other methods.", + "While Figures FIGREF14 \u2013 FIGREF16 show the relative behavior in the accuracy, it still interesting to observe the absolute accuracy rate obtained with the classifiers. In Table TABREF17, we show the best accuracy rate (i.e. $\\max \\Gamma _+ = \\max _p \\Gamma _+(p)$) for GloVe. We also show the average difference in performance ($\\langle \\Gamma _+ - \\Gamma _0 \\rangle $) and the total number of cases in which an improvement in performance was observed ($N_+$). $N_+$ ranges in the interval $0 \\le N_+ \\le 20$. Table TABREF17 summarizes the results obtained for $w = \\lbrace 1.0, 5.0, 10.0\\rbrace $ thousand words. Additional results for other text length are available in Tables TABREF28\u2013TABREF30 of the Supplementary Information.", + "In very short texts, despite the low accuracy rates, an improvement can be observed in all classifiers. The best results was obtained with SVM when virtual edges were included. For $w=5,000$ words, the inclusion of new edges has no positive effect on both kNN and Naive Bayes algorithms. On the other hand, once again SVM could be improved, yielding an optimized performance. For $w=10,000$ words, SVM could not be improved. However, even without improvement it yielded the maximum accuracy rate. The Naive Bayes algorithm, in average, could be improved by a margin of about 10%.", + "The results obtained for Word2vec are summarized in Table TABREF29 of the Supplementary Information. Considering short documents ($w=1,000$ words), here the best results occurs only with the decision tree method combined with enriched networks. Differently from the GloVe approach, SVM does not yield the best results. Nonetheless, the highest accuracy across all classifiers and values of $p$ is the same. For larger documents ($w=5,000$ and $w=10,000$ words), no significant difference in performance between Word2vec and GloVe is apparent.", + "The results obtained for FastText are shown in Table TABREF18. In short texts, only kNN and Naive Bayes have their performance improved with virtual edges. However, none of the optimized results for these classifiers outperformed SVM applied to the traditional co-occurrence model. Conversely, when $w=5,000$ words, the optimized results are obtained with virtual edges in the SVM classifier. Apart from kNN, the enriched networks improved the traditional approach in all classifiers. For large chunks of texts ($w=10,000$), once again the approach based on SVM and virtual edges yielded optimized results. All classifiers benefited from the inclusion of additional edges. Remarkably, Naive Bayes improved by a margin of about $13\\%$." + ], + [ + "While in the previous section we focused our analysis in the traditional word co-occurrence model, here we probe if the idea of considering virtual edges can also yield optimized results in particular modifications of the framework described in the methodology. The first modification in the co-occurrence model is the use of stopwords. While in semantical application of network language modeling stopwords are disregarded, in other application it can unravel interesting linguistic patterns BIBREF10. Here we analyzed the effect of using stopwords in enriched networks. We summarize the obtained results in Table TABREF20. We only show the results obtained with SVM, as it yielded the best results in comparison to other classifiers. The accuracy rate for other classifiers is shown in the Supplementary Information.", + "The results in Table TABREF20 reveals that even when stopwords are considered in the original model, an improvement can be observed with the addition of virtual edges. However, the results show that the degree of improvement depends upon the text length. In very short texts ($w=1,000$), none of the embeddings strategy was able to improve the performance of the classification. For $w=1,500$, a minor improvement was observed with FastText: the accuracy increased from $\\Gamma _0 = 37.18\\%$ to $38.46\\%$. A larger improvement could be observed for $w=2,000$. Both Word2vec and FastText approaches allowed an increase of more than 5% in performance. A gain higher than 10% was observed for $w=2,500$ with Word2vec. For larger pieces of texts, the gain is less expressive or absent. All in all, the results show that the use of virtual edges can also benefit the network approach based on stopwords. However, no significant improvement could be observed with very short and very large documents. The comparison of all three embedding methods showed that no method performed better than the others in all cases.", + "We also investigated if more informed thresholding strategies could provide better results. While the simple global thresholding approach might not be able to represent more complex structures, we also tested a more robust approach based on the local approach proposed by Serrano et al. BIBREF42. In Table TABREF21, we summarize the results obtained with this thresholding strategies. The table shows $\\max \\Gamma _+^{(L)} / \\max \\Gamma _+^{(G)}$, where $\\Gamma _+^{(L)}$ and $\\Gamma _+^{(G)}$ are the accuracy obtained with the local and global thresholding strategy, respectively. The results were obtained with the SVM classifier, as it turned to be the most efficient classification method. We found that there is no gain in performance when the local strategy is used. In particular cases, the global strategy is considerably more efficient. This is the case e.g. when GloVe is employed in texts with $w=1,500$ words. The performance of the global strategy is $12.2\\%$ higher than the one obtained with the global method. A minor difference in performance was found in texts comprising $w=1,000$ words, yet the global strategy is still more efficient than the global one.", + "To summarize all results obtained in this study we show in Table TABREF22 the best results obtained for each text length. We also show the relative gain in performance with the proposed approach and the embedding technique yielding the best result. All optimized results were obtained with the use of stopwords, global thresholding strategy and SVM as classification algorithm. A significant gain is more evident for intermediary text lengths." + ], + [ + "Textual classification remains one of the most important facets of the Natural Language Processing area. Here we studied a family of classification methods, the word co-occurrence networks. Despite this apparent simplicity, this model has been useful in several practical and theoretical scenarios. We proposed a modification of the traditional model by establishing virtual edges to connect nodes that are semantically similar via word embeddings. The reasoning behind this strategy is the fact the similar words are not properly linked in the traditional model and, thus, important links might be overlooked if only adjacent words are linked.", + "Taking as reference task a stylometric problem, we showed \u2013 as a proof of principle \u2013 that the use of virtual edges might improve the discriminability of networks. When analyzing the best results for each text length, apart from very short and long texts, the proposed strategy yielded optimized results in all cases. The best classification performance was always obtained with the SVM classifier. In addition, we found an improved performance when stopwords are used in the construction of the enriched co-occurrence networks. Finally, a simple global thresholding strategy was found to be more efficient than a local approach that preserves the community structure of the networks. Because complex networks are usually combined with other strategies BIBREF8, BIBREF11, we believe that the proposed could be used in combination with other methods to improve the classification performance of other text classification tasks.", + "Our findings paves the way for research in several new directions. While we probed the effectiveness of virtual edges in a specific text classification task, we could extend this approach for general classification tasks. A systematic comparison of embeddings techniques could also be performed to include other recent techniques BIBREF54, BIBREF55. We could also identify other relevant techniques to create virtual edges, allowing thus the use of the methodology in other networked systems other than texts. For example, a network could be enriched with embeddings obtained from graph embeddings techniques. A simpler approach could also consider link prediction BIBREF56 to create virtual edges. Finally, other interesting family of studies concerns the discrimination between co-occurrence and virtual edges, possibly by creating novel network measurements considering heterogeneous links." + ], + [ + "The authors acknowledge financial support from FAPESP (Grant no. 16/19069-9), CNPq-Brazil (Grant no. 304026/2018-2). This study was financed in part by the Coordena\u00e7\u00e3o de Aperfei\u00e7oamento de Pessoal de N\u00edvel Superior - Brasil (CAPES) - Finance Code 001." + ], + [ + "The following words were considered as stopwords in our analysis: all, just, don't, being, over, both, through, yourselves, its, before, o, don, hadn, herself, ll, had, should, to, only, won, under, ours,has, should've, haven't, do, them, his, very, you've, they, not, during, now, him, nor, wasn't, d, did, didn, this, she, each, further, won't, where, mustn't, isn't, few, because, you'd, doing, some, hasn, hasn't, are, our, ourselves, out, what, for, needn't, below, re, does, shouldn't, above, between, mustn, t, be, we, who, mightn't, doesn't, were, here, shouldn, hers, aren't, by, on, about, couldn, of, wouldn't, against, s, isn, or, own, into, yourself, down, hadn't, mightn, couldn't, wasn, your, you're, from, her, their, aren, it's, there, been, whom, too, wouldn, themselves, weren, was, until, more, himself, that, didn't, but, that'll, with, than, those, he, me, myself, ma, weren't, these, up, will, while, ain, can, theirs, my, and, ve, then, is, am, it, doesn, an, as, itself, at, have, in, any, if, again, no, when, same, how, other, which, you, shan't, shan, needn, haven, after, most, such, why, a, off i, m, yours, you'll, so, y, she's, the, having, once." + ], + [ + "The list of books is shown in Tables TABREF25 and TABREF26. For each book we show the respective authors (Aut.) and the following quantities: total number of words ($N_W$), total number of sentences ($N_S$), total number of paragraphs ($N_P$) and the average sentence length ($\\langle S_L \\rangle $), measured in number of words. The following authors were considered: Hector Hugh (HH), Thomas Hardy (TH), Daniel Defoe (DD), Allan Poe (AP), Bram Stoker (BS), Mark Twain (MT), Charles Dickens (CD), Pelham Grenville (PG), Charles Darwin (CD), Arthur Doyle (AD), George Eliot (GE), Jane Austen (JA), and Joseph Conrad (JC)." + ], + [ + "In this section we show additional results obtained for different text length. More specifically, we show the results obtained for GloVe, Word2vec and FastText when stopwords are either considered in the text or disregarded from the analysis." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0725/instruction.md b/qasper-0725/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..d0a7215c815ea7d0ce6fad07953a86c609718e9a --- /dev/null +++ b/qasper-0725/instruction.md @@ -0,0 +1,97 @@ +Name of Paper: Using word embeddings to improve the discriminability of co-occurrence text networks + +Question: On what model architectures are previous co-occurence networks based? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related works", + "Material and Methods", + "Results and Discussion", + "Results and Discussion ::: Performance analysis", + "Results and Discussion ::: Effects of considering stopwords and local thresholding", + "Conclusion", + "Acknowledgments", + "Supplementary Information ::: Stopwords", + "Supplementary Information ::: List of books", + "Supplementary Information ::: Additional results" + ], + "paragraphs": [ + [ + "The ability to construct complex and diverse linguistic structures is one of the main features that set us apart from all other species. Despite its ubiquity, some language aspects remain unknown. Topics such as language origin and evolution have been studied by researchers from diverse disciplines, including Linguistic, Computer Science, Physics and Mathematics BIBREF0, BIBREF1, BIBREF2. In order to better understand the underlying language mechanisms and universal linguistic properties, several models have been developed BIBREF3, BIBREF4. A particular language representation regards texts as complex systems BIBREF5. Written texts can be considered as complex networks (or graphs), where nodes could represent syllables, words, sentences, paragraphs or even larger chunks BIBREF5. In such models, network edges represent the proximity between nodes, e.g. the frequency of the co-occurrence of words. Several interesting results have been obtained from networked models, such as the explanation of Zipf's Law as a consequence of the least effort principle and theories on the nature of syntactical relationships BIBREF6, BIBREF7.", + "In a more practical scenario, text networks have been used in text classification tasks BIBREF8, BIBREF9, BIBREF10. The main advantage of the model is that it does not rely on deep semantical information to obtain competitive results. Another advantage of graph-based approaches is that, when combined with other approaches, it yields competitive results BIBREF11. A simple, yet recurrent text model is the well-known word co-occurrence network. After optional textual pre-processing steps, in a co-occurrence network each different word becomes a node and edges are established via co-occurrence in a desired window. A common strategy connects only adjacent words in the so called word adjacency networks.", + "While the co-occurrence representation yields good results in classification scenarios, some important features are not considered in the model. For example, long-range syntactical links, though less frequent than adjacent syntactical relationships, might be disregarded from a simple word adjacency approach BIBREF12. In addition, semantically similar words not sharing the same lemma are mapped into distinct nodes. In order to address these issues, here we introduce a modification of the traditional network representation by establishing additional edges, referred to as \u201cvirtual\u201d edges. In the proposed model, in addition to the co-occurrence edges, we link two nodes (words) if the corresponding word embedding representation is similar. While this approach still does not merge similar nodes into the same concept, similar nodes are explicitly linked via virtual edges.", + "Our main objective here is to evaluate whether such an approach is able to improve the discriminability of word co-occurrence networks in a typical text network classification task. We evaluate the methodology for different embedding techniques, including GloVe, Word2Vec and FastText. We also investigated different thresholding strategies to establish virtual links. Our results revealed, as a proof of principle, that the proposed approach is able to improve the discriminability of the classification when compared to the traditional co-occurrence network. While the gain in performance depended upon the text length being considered, we found relevant gains for intermediary text lengths. Additional results also revealed that a simple thresholding strategy combined with the use of stopwords tends to yield the best results.", + "We believe that the proposed representation could be applied in other text classification tasks, which could lead to potential gains in performance. Because the inclusion of virtual edges is a simple technique to make the network denser, such an approach can benefit networked representations with a limited number of nodes and edges. This representation could also shed light into language mechanisms in theoretical studies relying on the representation of text as complex networks. Potential novel research lines leveraging the adopted approach to improve the characterization of texts in other applications are presented in the conclusion." + ], + [ + "Complex networks have been used in a wide range of fields, including in Social Sciences BIBREF13, Neuroscience BIBREF14, Biology BIBREF15, Scientometry BIBREF16 and Pattern Recognition BIBREF17, BIBREF18, BIBREF19, BIBREF20. In text analysis, networks are used to uncover language patterns, including the origins of the ever present Zipf's Law BIBREF21 and the analysis of linguistic properties of natural and unknown texts BIBREF22, BIBREF23. Applications of network science in text mining and text classification encompasses applications in semantic analysis BIBREF24, BIBREF25, BIBREF26, BIBREF27, authorship attribution BIBREF28, BIBREF29 and stylometry BIBREF28, BIBREF30, BIBREF31. Here we focus in the stylometric analysis of texts using complex networks.", + "In BIBREF28, the authors used a co-occurrence network to study a corpus of English and Polish books. They considered a dataset of 48 novels, which were written by 8 different authors. Differently from traditional co-occurrence networks, some punctuation marks were considered as words when mapping texts as networks. The authors also decided to create a methodology to normalize the obtained network metrics, since they considered documents with variations in length. A similar approach was adopted in a similar study BIBREF32, with a focus on comparing novel measurements and measuring the effect of considering stopwords in the network structure.", + "A different approach to analyze co-occurrence networks was devised in BIBREF33. Whilst most approaches only considered traditional network measurements or devised novel topological and dynamical measurements, the authors combined networked and semantic information to improve the performance of network-based classification. Interesting, the combined use of network motifs and node labels (representing the corresponding words) allowed an improvement in performance in the considered task. A similar combination of techniques using a hybrid approach was proposed in BIBREF8. Networked-based approaches has also been applied to the authorship recognition tasks in other languages, including Persian texts BIBREF9.", + "Co-occurrence networks have been used in other contexts other than stylometric analysis. The main advantage of this approach is illustrated in the task aimed at diagnosing diseases via text analysis BIBREF11. Because the topological analysis of co-occurrence language networks do not require deep semantic analysis, this model is able to model text created by patients suffering from cognitive impairment BIBREF11. Recently, it has been shown that the combination of network and traditional features could be used to improve the diagnosis of patients with cognitive impairment BIBREF11. Interestingly, this was one of the first approaches suggesting the use of embeddings to address the particular problem of lack of statistics to create a co-occurrence network in short documents BIBREF34.", + "While many of the works dealing with word co-occurrence networks have been proposed in the last few years, no systematic study of the effects of including information from word embeddings in such networks has been analyzed. This work studies how links created via embeddings information modify the underlying structure of networks and, most importantly, how it can improve the model to provide improved classification performance in the stylometry task." + ], + [ + "To represent texts as networks, we used the so-called word adjacency network representation BIBREF35, BIBREF28, BIBREF32. Typically, before creating the networks, the text is pre-processed. An optional pre-processing step is the removal of stopwords. This step is optional because such words include mostly article and prepositions, which may be artlessly represented by network edges. However, in some applications \u2013 including the authorship attribution task \u2013 stopwords (or function words) play an important role in the stylistic characterization of texts BIBREF32. A list of stopwords considered in this study is available in the Supplementary Information.", + "The pre-processing step may also include a lemmatization procedure. This step aims at mapping words conveying the same meaning into the same node. In the lemmatization process, nouns and verbs are mapped into their singular and infinite forms. Note that, while this step is useful to merge words sharing a lemma into the same node, more complex semantical relationships are overlooked. For example, if \u201ccar\u201d and \u201cvehicle\u201d co-occur in the same text, they are considered as distinct nodes, which may result in an inaccurate representation of the text.", + "Such a drawback is addressed by including \u201cvirtual\u201d edges connecting nodes. In other words, even if two words are not adjacent in the text, we include \u201cvirtual\u201d edges to indicate that two distant words are semantically related. The inclusion of such virtual edges is illustrated in Figure FIGREF1. In order to measure the semantical similarity between two concepts, we use the concept of word embeddings BIBREF36, BIBREF37. Thus, each word is represented using a vector representation encoding the semantical and contextual characteristics of the word. Several interesting properties have been obtained from distributed representation of words. One particular property encoded in the embeddings representation is the fact the semantical similarity between concepts is proportional to the similarity of vectors representing the words. Similarly to several other works, here we measure the similarity of the vectors via cosine similarity BIBREF38.", + "The following strategies to create word embedding were considered in this paper:", + "GloVe: the Global Vectors (GloVe) algorithm is an extension of the Word2vec model BIBREF39 for efficient word vector learning BIBREF40. This approach combines global statistics from matrix factorization techniques (such as latent semantic analysis) with context-based and predictive methods like Word2Vec. This method is called as Global Vector method because the global corpus statistics are captured by GloVe. Instead of using a window to define the local context, GloVe constructs an explicit word-context matrix (or co-occurrence matrix) using statistics across the entire corpus. The final result is a learning model that oftentimes yields better word vector representations BIBREF40.", + "Word2Vec: this is a predictive model that finds dense vector representations of words using a three-layer neural network with a single hidden layer BIBREF39. It can be defined in a two-fold way: continuous bag-of-words and skip-gram model. In the latter, the model analyzes the words of a set of sentences (or corpus) and attempts to predict the neighbors of such words. For example, taking as reference the word \u201cRobin\u201d, the model decides that \u201cHood\u201d is more likely to follow the reference word than any other word. The vectors are obtained as follows: given the vocabulary (generated from all corpus words), the model trains a neural network with the sentences of the corpus. Then, for a given word, the probabilities that each word follows the reference word are obtained. Once the neural network is trained, the weights of the hidden layer are used as vectors of each corpus word.", + "FastText: this method is another extension of the Word2Vec model BIBREF41. Unlike Word2Vec, FastText represents each word as a bag of character n-grams. Therefore, the neural network not only trains individual words, but also several n-grams of such words. The vector for a word is the sum of vectors obtained for the character n-grams composing the word. For example, the embedding obtained for the word \u201ccomputer\u201d with $n\\le 3$ is the sum of the embeddings obtained for \u201cco\u201d, \u201ccom\u201d, \u201comp\u201d, \u201cmpu\u201d, \u201cput\u201d, \u201cute\u201d, \u201cter\u201d and \u201cer\u201d. In this way, this method obtains improved representations for rare words, since n-grams composing rare words might be present in other words. The FastText representation also allows the model to understand suffixes and prefixes. Another advantage of FastText is its efficiency to be trained in very large corpora.", + "Concerning the thresholding process, we considered two main strategies. First, we used a global strategy: in addition to the co-occurrence links (continuous lines in Figure FIGREF1), only \u201cvirtual\u201d edges stronger than a given threshold are left in the network. Thus only the most similar concepts are connected via virtual links. This strategy is hereafter referred to as global strategy. Unfortunately, this method may introduce an undesired bias towards hubs BIBREF42.", + "To overcome the potential disadvantages of the global thresholding method, we also considered a more refined thresholding approach that takes into account the local structure to decide whether a weighted link is statistically significant BIBREF42. This method relies on the idea that the importance of an edge should be considered in the the context in which it appears. In other words, the relevance of an edge should be evaluated by analyzing the nodes connected to its ending points. Using the concept of disparity filter, the method devised in BIBREF42 defines a null model that quantifies the probability of a node to be connected to an edge with a given weight, based on its other connections. This probability is used to define the significance of the edge. The parameter that is used to measure the significance of an edge $e_{ij}$ is $\\alpha _{ij}$, defined as:", + "where $w_{ij}$ is the weight of the edge $e_{ij}$ and $k_i$ is the degree of the $i$-th node. The obtained network corresponds to the set of nodes and edges obtained by removing all edges with $\\alpha $ higher than the considered threshold. Note that while the similarity between co-occurrence links might be considered to compute $\\alpha _{ij}$, only \u201cvirtual\u201d edges (i.e. the dashed lines in Figure FIGREF1) are eligible to be removed from the network in the filtering step. This strategy is hereafter referred to as local strategy.", + "After co-occurrence networks are created and virtual edges are included, in the next step we used a characterization based on topological analysis. Because a global topological analysis is prone to variations in network size, we focused our analysis in the local characterization of complex networks. In a local topological analysis, we use as features the value of topological/dynamical measurements obtained for a set of words. In this case, we selected as feature the words occurring in all books of the dataset. For each word, we considered the following network measurements: degree, betweenness, clustering coefficient, average shortest path length, PageRank, concentric symmetry (at the second and third hierarchical level) BIBREF32 and accessibility BIBREF43, BIBREF44 (at the second and third hierarchical level). We chose these measurements because all of them capture some particular linguistic feature of texts BIBREF45, BIBREF46, BIBREF47, BIBREF48. After network measurements are extracted, they are used in machine learning algorithms. In our experiments, we considered Decision Trees (DT), nearest neighbors (kNN), Naive Bayes (NB) and Support Vector Machines (SVM). We used some heuristics to optimize classifier parameters. Such techniques are described in the literature BIBREF49. The accuracy of the pattern recognition methods were evaluated using cross-validation BIBREF50.", + "In summary, the methodology used in this paper encompasses the following steps:", + "Network construction: here texts are mapped into a co-occurrence networks. Some variations exists in the literature, however here we focused in the most usual variation, i.e. the possibility of considering or disregarding stopwords. A network with co-occurrence links is obtained after this step.", + "Network enrichment: in this step, the network is enriched with virtual edges established via similarity of word embeddings. After this step, we are given a complete network with weighted links. Virtually, any embedding technique could be used to gauge the similarity between nodes.", + "Network filtering: in order to eliminate spurious links included in the last step, the weakest edges are filtered. Two approaches were considered: a simple approach based on a global threshold and a local thresholding strategy that preserves network community structure. The outcome of this network filtering step is a network with two types of links: co-occurrence and virtual links (as shown in Figure FIGREF1).", + "Feature extraction: In this step, topological and dynamical network features are extracted. Here, we do not discriminate co-occurrence from virtual edges to compute the network metrics.", + "Pattern classification: once features are extracted from complex networks, they are used in pattern classification methods. This might include supervised, unsupervised and semi-supervised classification. This framework is exemplified in the supervised scenario.", + "The above framework is exemplified with the most common technique(s). It should be noted that the methods used, however, can be replaced by similar techniques. For example, the network construction could consider stopwords or even punctuation marks BIBREF51. Another possibility is the use of different strategies of thresholding. While a systematic analysis of techniques and parameters is still required to reveal other potential advantages of the framework based on the addition of virtual edges, in this paper we provide a first analysis showing that virtual edges could be useful to improve the discriminability of texts modeled as complex networks.", + "Here we used a dataset compatible with datasets used recently in the literature (see e.g. BIBREF28, BIBREF10, BIBREF52). The objective of the studied stylometric task is to identify the authorship of an unknown document BIBREF53. All data and some statistics of each book are shown in the Supplementary Information." + ], + [ + "In Section SECREF13, we probe whether the inclusion of virtual edges is able to improve the performance of the traditional co-occurrence network-based classification in a usual stylometry task. While the focus of this paper is not to perform a systematic analysis of different methods comprising the adopted network, we consider two variations in the adopted methodology. In Section SECREF19, we consider the use of stopwords and the adoption of a local thresholding process to establish different criteria to create new virtual edges." + ], + [ + "In Figure FIGREF14, we show some of the improvements in performance obtained when including a fixed amount of virtual edges using GloVe as embedding method. In each subpanel, we show the relative improvement in performance obtained as a function of the fraction of additional edges. In this section, we considered the traditional co-occurrence as starting point. In other words, the network construction disregarded stopwords. The list of stopwords considered in this paper is available in the Supplementary Information. We also considered the global approach to filter edges.", + "The relative improvement in performance is given by $\\Gamma _+{(p)}/\\Gamma _0$, where $\\Gamma _+{(p)}$ is the accuracy rate obtained when $p\\%$ additional edges are included and $\\Gamma _0 = \\Gamma _+{(p=0)}$, i.e. $\\Gamma _0$ is the accuracy rate measured from the traditional co-occurrence model. We only show the highest relative improvements in performance for each classifier. In our analysis, we considered also samples of text with distinct length, since the performance of network-based methods is sensitive to text length BIBREF34. In this figure, we considered samples comprising $w=\\lbrace 1.0, 2.5, 5.0, 10.0\\rbrace $ thousand words.", + "The results obtained for GloVe show that the highest relative improvements in performance occur for decision trees. This is apparent specially for the shortest samples. For $w=1,000$ words, the decision tree accuracy is enhanced by a factor of almost 50% when $p=20\\%$. An excellent gain in performance is also observed for both Naive Bayes and SVM classifiers, when $p=18\\%$ and $p=12\\%$, respectively. When $w=2,500$ words, the highest improvements was observed for the decision tree algorithm. A minor improvement was observed for the kNN method. A similar behavior occurred for $w=5,000$ words. Interestingly, SVM seems to benefit from the use of additional edges when larger documents are considered. When only 5% virtual edges are included, the relative gain in performance is about 45%.", + "The relative gain in performance obtained for Word2vec is shown in Figure FIGREF15. Overall, once again decision trees obtained the highest gain in performance when short texts are considered. Similar to the analysis based on the GloVe method, the gain for kNN is low when compared to the benefit received by other methods. Here, a considerable gain for SVM in only clear for $w=2,500$ and $p=10\\%$. When large texts are considered, Naive Bayes obtained the largest gain in performance.", + "Finally, the relative gain in performance obtained for FastText is shown in Figure FIGREF16. The prominent role of virtual edges in decision tree algorithm in the classification of short texts once again is evident. Conversely, the classification of large documents using virtual edges mostly benefit the classification based on the Naive Bayes classifier. Similarly to the results observed for Glove and Word2vec, the gain in performance obtained for kNN is low compared when compared to other methods.", + "While Figures FIGREF14 \u2013 FIGREF16 show the relative behavior in the accuracy, it still interesting to observe the absolute accuracy rate obtained with the classifiers. In Table TABREF17, we show the best accuracy rate (i.e. $\\max \\Gamma _+ = \\max _p \\Gamma _+(p)$) for GloVe. We also show the average difference in performance ($\\langle \\Gamma _+ - \\Gamma _0 \\rangle $) and the total number of cases in which an improvement in performance was observed ($N_+$). $N_+$ ranges in the interval $0 \\le N_+ \\le 20$. Table TABREF17 summarizes the results obtained for $w = \\lbrace 1.0, 5.0, 10.0\\rbrace $ thousand words. Additional results for other text length are available in Tables TABREF28\u2013TABREF30 of the Supplementary Information.", + "In very short texts, despite the low accuracy rates, an improvement can be observed in all classifiers. The best results was obtained with SVM when virtual edges were included. For $w=5,000$ words, the inclusion of new edges has no positive effect on both kNN and Naive Bayes algorithms. On the other hand, once again SVM could be improved, yielding an optimized performance. For $w=10,000$ words, SVM could not be improved. However, even without improvement it yielded the maximum accuracy rate. The Naive Bayes algorithm, in average, could be improved by a margin of about 10%.", + "The results obtained for Word2vec are summarized in Table TABREF29 of the Supplementary Information. Considering short documents ($w=1,000$ words), here the best results occurs only with the decision tree method combined with enriched networks. Differently from the GloVe approach, SVM does not yield the best results. Nonetheless, the highest accuracy across all classifiers and values of $p$ is the same. For larger documents ($w=5,000$ and $w=10,000$ words), no significant difference in performance between Word2vec and GloVe is apparent.", + "The results obtained for FastText are shown in Table TABREF18. In short texts, only kNN and Naive Bayes have their performance improved with virtual edges. However, none of the optimized results for these classifiers outperformed SVM applied to the traditional co-occurrence model. Conversely, when $w=5,000$ words, the optimized results are obtained with virtual edges in the SVM classifier. Apart from kNN, the enriched networks improved the traditional approach in all classifiers. For large chunks of texts ($w=10,000$), once again the approach based on SVM and virtual edges yielded optimized results. All classifiers benefited from the inclusion of additional edges. Remarkably, Naive Bayes improved by a margin of about $13\\%$." + ], + [ + "While in the previous section we focused our analysis in the traditional word co-occurrence model, here we probe if the idea of considering virtual edges can also yield optimized results in particular modifications of the framework described in the methodology. The first modification in the co-occurrence model is the use of stopwords. While in semantical application of network language modeling stopwords are disregarded, in other application it can unravel interesting linguistic patterns BIBREF10. Here we analyzed the effect of using stopwords in enriched networks. We summarize the obtained results in Table TABREF20. We only show the results obtained with SVM, as it yielded the best results in comparison to other classifiers. The accuracy rate for other classifiers is shown in the Supplementary Information.", + "The results in Table TABREF20 reveals that even when stopwords are considered in the original model, an improvement can be observed with the addition of virtual edges. However, the results show that the degree of improvement depends upon the text length. In very short texts ($w=1,000$), none of the embeddings strategy was able to improve the performance of the classification. For $w=1,500$, a minor improvement was observed with FastText: the accuracy increased from $\\Gamma _0 = 37.18\\%$ to $38.46\\%$. A larger improvement could be observed for $w=2,000$. Both Word2vec and FastText approaches allowed an increase of more than 5% in performance. A gain higher than 10% was observed for $w=2,500$ with Word2vec. For larger pieces of texts, the gain is less expressive or absent. All in all, the results show that the use of virtual edges can also benefit the network approach based on stopwords. However, no significant improvement could be observed with very short and very large documents. The comparison of all three embedding methods showed that no method performed better than the others in all cases.", + "We also investigated if more informed thresholding strategies could provide better results. While the simple global thresholding approach might not be able to represent more complex structures, we also tested a more robust approach based on the local approach proposed by Serrano et al. BIBREF42. In Table TABREF21, we summarize the results obtained with this thresholding strategies. The table shows $\\max \\Gamma _+^{(L)} / \\max \\Gamma _+^{(G)}$, where $\\Gamma _+^{(L)}$ and $\\Gamma _+^{(G)}$ are the accuracy obtained with the local and global thresholding strategy, respectively. The results were obtained with the SVM classifier, as it turned to be the most efficient classification method. We found that there is no gain in performance when the local strategy is used. In particular cases, the global strategy is considerably more efficient. This is the case e.g. when GloVe is employed in texts with $w=1,500$ words. The performance of the global strategy is $12.2\\%$ higher than the one obtained with the global method. A minor difference in performance was found in texts comprising $w=1,000$ words, yet the global strategy is still more efficient than the global one.", + "To summarize all results obtained in this study we show in Table TABREF22 the best results obtained for each text length. We also show the relative gain in performance with the proposed approach and the embedding technique yielding the best result. All optimized results were obtained with the use of stopwords, global thresholding strategy and SVM as classification algorithm. A significant gain is more evident for intermediary text lengths." + ], + [ + "Textual classification remains one of the most important facets of the Natural Language Processing area. Here we studied a family of classification methods, the word co-occurrence networks. Despite this apparent simplicity, this model has been useful in several practical and theoretical scenarios. We proposed a modification of the traditional model by establishing virtual edges to connect nodes that are semantically similar via word embeddings. The reasoning behind this strategy is the fact the similar words are not properly linked in the traditional model and, thus, important links might be overlooked if only adjacent words are linked.", + "Taking as reference task a stylometric problem, we showed \u2013 as a proof of principle \u2013 that the use of virtual edges might improve the discriminability of networks. When analyzing the best results for each text length, apart from very short and long texts, the proposed strategy yielded optimized results in all cases. The best classification performance was always obtained with the SVM classifier. In addition, we found an improved performance when stopwords are used in the construction of the enriched co-occurrence networks. Finally, a simple global thresholding strategy was found to be more efficient than a local approach that preserves the community structure of the networks. Because complex networks are usually combined with other strategies BIBREF8, BIBREF11, we believe that the proposed could be used in combination with other methods to improve the classification performance of other text classification tasks.", + "Our findings paves the way for research in several new directions. While we probed the effectiveness of virtual edges in a specific text classification task, we could extend this approach for general classification tasks. A systematic comparison of embeddings techniques could also be performed to include other recent techniques BIBREF54, BIBREF55. We could also identify other relevant techniques to create virtual edges, allowing thus the use of the methodology in other networked systems other than texts. For example, a network could be enriched with embeddings obtained from graph embeddings techniques. A simpler approach could also consider link prediction BIBREF56 to create virtual edges. Finally, other interesting family of studies concerns the discrimination between co-occurrence and virtual edges, possibly by creating novel network measurements considering heterogeneous links." + ], + [ + "The authors acknowledge financial support from FAPESP (Grant no. 16/19069-9), CNPq-Brazil (Grant no. 304026/2018-2). This study was financed in part by the Coordena\u00e7\u00e3o de Aperfei\u00e7oamento de Pessoal de N\u00edvel Superior - Brasil (CAPES) - Finance Code 001." + ], + [ + "The following words were considered as stopwords in our analysis: all, just, don't, being, over, both, through, yourselves, its, before, o, don, hadn, herself, ll, had, should, to, only, won, under, ours,has, should've, haven't, do, them, his, very, you've, they, not, during, now, him, nor, wasn't, d, did, didn, this, she, each, further, won't, where, mustn't, isn't, few, because, you'd, doing, some, hasn, hasn't, are, our, ourselves, out, what, for, needn't, below, re, does, shouldn't, above, between, mustn, t, be, we, who, mightn't, doesn't, were, here, shouldn, hers, aren't, by, on, about, couldn, of, wouldn't, against, s, isn, or, own, into, yourself, down, hadn't, mightn, couldn't, wasn, your, you're, from, her, their, aren, it's, there, been, whom, too, wouldn, themselves, weren, was, until, more, himself, that, didn't, but, that'll, with, than, those, he, me, myself, ma, weren't, these, up, will, while, ain, can, theirs, my, and, ve, then, is, am, it, doesn, an, as, itself, at, have, in, any, if, again, no, when, same, how, other, which, you, shan't, shan, needn, haven, after, most, such, why, a, off i, m, yours, you'll, so, y, she's, the, having, once." + ], + [ + "The list of books is shown in Tables TABREF25 and TABREF26. For each book we show the respective authors (Aut.) and the following quantities: total number of words ($N_W$), total number of sentences ($N_S$), total number of paragraphs ($N_P$) and the average sentence length ($\\langle S_L \\rangle $), measured in number of words. The following authors were considered: Hector Hugh (HH), Thomas Hardy (TH), Daniel Defoe (DD), Allan Poe (AP), Bram Stoker (BS), Mark Twain (MT), Charles Dickens (CD), Pelham Grenville (PG), Charles Darwin (CD), Arthur Doyle (AD), George Eliot (GE), Jane Austen (JA), and Joseph Conrad (JC)." + ], + [ + "In this section we show additional results obtained for different text length. More specifically, we show the results obtained for GloVe, Word2vec and FastText when stopwords are either considered in the text or disregarded from the analysis." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0741/instruction.md b/qasper-0741/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..77cbedc70b31df72b34bab2de9d126ea1cb0f1bf --- /dev/null +++ b/qasper-0741/instruction.md @@ -0,0 +1,43 @@ +Name of Paper: Learning Twitter User Sentiments on Climate Change with Limited Labeled Data + +Question: Do the authors mention any confounds to their study? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Background", + "Data", + "Labeling Methodology", + "Outcome Analysis", + "Results & Discussion" + ], + "paragraphs": [ + [ + "Much prior work has been done at the intersection of climate change and Twitter, such as tracking climate change sentiment over time BIBREF2 , finding correlations between Twitter climate change sentiment and seasonal effects BIBREF3 , and clustering Twitter users based on climate mentalities using network analysis BIBREF4 . Throughout, Twitter has been accepted as a powerful tool given the magnitude and reach of samples unattainable from standard surveys. However, the aforementioned studies are not scalable with regards to training data, do not use more recent sentiment analysis tools (such as neural nets), and do not consider unbiased comparisons pre- and post- various climate events (which would allow for a more concrete evaluation of shocks to climate change sentiment). This paper aims to address these three concerns as follows.", + "First, we show that machine learning models formed using our labeling technique can accurately predict tweet sentiment (see Section SECREF2 ). We introduce a novel method to intuit binary sentiments of large numbers of tweets for training purposes. Second, we quantify unbiased outcomes from these predicted sentiments (see Section SECREF4 ). We do this by comparing sentiments within the same cohort of Twitter users tweeting both before and after specific natural disasters; this removes bias from over-weighting Twitter users who are only compelled to compose tweets after a disaster." + ], + [ + "We henceforth refer to a tweet affirming climate change as a \u201cpositive\" sample (labeled as 1 in the data), and a tweet denying climate change as a \u201cnegative\" sample (labeled as -1 in the data). All data were downloaded from Twitter in two separate batches using the \u201ctwint\" scraping tool BIBREF5 to sample historical tweets for several different search terms; queries always included either \u201cclimate change\" or \u201cglobal warming\", and further included disaster-specific search terms (e.g., \u201cbomb cyclone,\" \u201cblizzard,\" \u201csnowstorm,\" etc.). We refer to the first data batch as \u201cinfluential\" tweets, and the second data batch as \u201cevent-related\" tweets.", + "The first data batch consists of tweets relevant to blizzards, hurricanes, and wildfires, under the constraint that they are tweeted by \u201cinfluential\" tweeters, who we define as individuals certain to have a classifiable sentiment regarding the topic at hand. For example, we assume that any tweet composed by Al Gore regarding climate change is a positive sample, whereas any tweet from conspiracy account @ClimateHiJinx is a negative sample. The assumption we make in ensuing methods (confirmed as reasonable in Section SECREF2 ) is that influential tweeters can be used to label tweets in bulk in the absence of manually-labeled tweets. Here, we enforce binary labels for all tweets composed by each of the 133 influential tweeters that we identified on Twitter (87 of whom accept climate change), yielding a total of 16,360 influential tweets.", + "The second data batch consists of event-related tweets for five natural disasters occurring in the U.S. in 2018. These are: the East Coast Bomb Cyclone (Jan. 2 - 6); the Mendocino, California wildfires (Jul. 27 - Sept. 18); Hurricane Florence (Aug. 31 - Sept. 19); Hurricane Michael (Oct. 7 - 16); and the California Camp Fires (Nov. 8 - 25). For each disaster, we scraped tweets starting from two weeks prior to the beginning of the event, and continuing through two weeks after the end of the event. Summary statistics on the downloaded event-specific tweets are provided in Table TABREF1 . Note that the number of tweets occurring prior to the two 2018 sets of California fires are relatively small. This is because the magnitudes of these wildfires were relatively unpredictable, whereas blizzards and hurricanes are often forecast weeks in advance alongside public warnings. The first (influential tweet data) and second (event-related tweet data) batches are de-duplicated to be mutually exclusive. In Section SECREF2 , we perform geographic analysis on the event-related tweets from which we can scrape self-reported user city from Twitter user profile header cards; overall this includes 840 pre-event and 5,984 post-event tweets.", + "To create a model for predicting sentiments of event-related tweets, we divide the first data batch of influential tweets into training and validation datasets with a 90%/10% split. The training set contains 49.2% positive samples, and the validation set contains 49.0% positive samples. We form our test set by manually labeling a subset of 500 tweets from the the event-related tweets (randomly chosen across all five natural disasters), of which 50.0% are positive samples." + ], + [ + "Our first goal is to train a sentiment analysis model (on training and validation datasets) in order to perform classification inference on event-based tweets. We experimented with different feature extraction methods and classification models. Feature extractions examined include Tokenizer, Unigram, Bigram, 5-char-gram, and td-idf methods. Models include both neural nets (e.g. RNNs, CNNs) and standard machine learning tools (e.g. Naive Bayes with Laplace Smoothing, k-clustering, SVM with linear kernel). Model accuracies are reported in Table FIGREF3 .", + "The RNN pre-trained using GloVe word embeddings BIBREF6 achieved the higest test accuracy. We pass tokenized features into the embedding layer, followed by an LSTM BIBREF7 with dropout and ReLU activation, and a dense layer with sigmoid activation. We apply an Adam optimizer on the binary crossentropy loss. Implementing this simple, one-layer LSTM allows us to surpass the other traditional machine learning classification methods. Note the 13-point spread between validation and test accuracies achieved. Ideally, the training, validation, and test datasets have the same underlying distribution of tweet sentiments; the assumption made with our labeling technique is that the influential accounts chosen are representative of all Twitter accounts. Critically, when choosing the influential Twitter users who believe in climate change, we highlighted primarily politicians or news sources (i.e., verifiably affirming or denying climate change); these tweets rarely make spelling errors or use sarcasm. Due to this skew, the model yields a high rate of false negatives. It is likely that we could lessen the gap between validation and test accuracies by finding more \u201creal\" Twitter users who are climate change believers, e.g. by using the methodology found in BIBREF4 ." + ], + [ + "Our second goal is to compare the mean values of users' binary sentiments both pre- and post- each natural disaster event. Applying our highest-performing RNN to event-related tweets yields the following breakdown of positive tweets: Bomb Cyclone (34.7%), Mendocino Wildfire (80.4%), Hurricane Florence (57.2%), Hurricane Michael (57.6%), and Camp Fire (70.1%). As sanity checks, we examine the predicted sentiments on a subset with geographic user information and compare results to the prior literature.", + "In Figure FIGREF3 , we map 4-clustering results on three dimensions: predicted sentiments, latitude, and longitude. The clusters correspond to four major regions of the U.S.: the Northeast (green), Southeast (yellow), Midwest (blue), and West Coast (purple); centroids are designated by crosses. Average sentiments within each cluster confirm prior knowledge BIBREF1 : the Southeast and Midwest have lower average sentiments ( INLINEFORM0 and INLINEFORM1 , respectively) than the West Coast and Northeast (0.22 and 0.09, respectively). In Figure FIGREF5 , we plot predicted sentiment averaged by U.S. city of event-related tweeters. The majority of positive tweets emanate from traditionally liberal hubs (e.g. San Francisco, Los Angeles, Austin), while most negative tweets come from the Philadelphia metropolitan area. These regions aside, rural areas tended to see more negative sentiment tweeters post-event, whereas urban regions saw more positive sentiment tweeters; however, overall average climate change sentiment pre- and post-event was relatively stable geographically. This map further confirms findings that coastal cities tend to be more aware of climate change BIBREF8 .", + "From these mapping exercises, we claim that our \u201cinfluential tweet\" labeling is reasonable. We now discuss our final method on outcomes: comparing average Twitter sentiment pre-event to post-event. In Figure FIGREF8 , we display these metrics in two ways: first, as an overall average of tweet binary sentiment, and second, as a within-cohort average of tweet sentiment for the subset of tweets by users who tweeted both before and after the event (hence minimizing awareness bias). We use Student's t-test to calculate the significance of mean sentiment differences pre- and post-event (see Section SECREF4 ). Note that we perform these mean comparisons on all event-related data, since the low number of geo-tagged samples would produce an underpowered study." + ], + [ + "In Figure FIGREF8 , we see that overall sentiment averages rarely show movement post-event: that is, only Hurricane Florence shows a significant difference in average tweet sentiment pre- and post-event at the 1% level, corresponding to a 0.12 point decrease in positive climate change sentiment. However, controlling for the same group of users tells a different story: both Hurricane Florence and Hurricane Michael have significant tweet sentiment average differences pre- and post-event at the 1% level. Within-cohort, Hurricane Florence sees an increase in positive climate change sentiment by 0.21 points, which is contrary to the overall average change (the latter being likely biased since an influx of climate change deniers are likely to tweet about hurricanes only after the event). Hurricane Michael sees an increase in average tweet sentiment of 0.11 points, which reverses the direction of tweets from mostly negative pre-event to mostly positive post-event. Likely due to similar bias reasons, the Mendocino wildfires in California see a 0.06 point decrease in overall sentiment post-event, but a 0.09 point increase in within-cohort sentiment. Methodologically, we assert that overall averages are not robust results to use in sentiment analyses.", + "We now comment on the two events yielding similar results between overall and within-cohort comparisons. Most tweets regarding the Bomb Cyclone have negative sentiment, though sentiment increases by 0.02 and 0.04 points post-event for overall and within-cohort averages, respectively. Meanwhile, the California Camp Fires yield a 0.11 and 0.27 point sentiment decline in overall and within-cohort averages, respectively. This large difference in sentiment change can be attributed to two factors: first, the number of tweets made regarding wildfires prior to the (usually unexpected) event is quite low, so within-cohort users tend to have more polarized climate change beliefs. Second, the root cause of the Camp Fires was quickly linked to PG&E, bolstering claims that climate change had nothing to do with the rapid spread of fire; hence within-cohort users were less vocally positive regarding climate change post-event.", + "There are several caveats in our work: first, tweet sentiment is rarely binary (this work could be extended to a multinomial or continuous model). Second, our results are constrained to Twitter users, who are known to be more negative than the general U.S. population BIBREF9 . Third, we do not take into account the aggregate effects of continued natural disasters over time. Going forward, there is clear demand in discovering whether social networks can indicate environmental metrics in a \u201cnowcasting\" fashion. As climate change becomes more extreme, it remains to be seen what degree of predictive power exists in our current model regarding climate change sentiments with regards to natural disasters." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0771/instruction.md b/qasper-0771/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..f0a32c691df2d52891c70b7ddbcff7843f6493d1 --- /dev/null +++ b/qasper-0771/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: The First Evaluation of Chinese Human-Computer Dialogue Technology + +Question: What collection steps do they mention? \ No newline at end of file diff --git a/qasper-0779/instruction.md b/qasper-0779/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..05a1c8be0b240258bfb866eed86d30a402ab485c --- /dev/null +++ b/qasper-0779/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Multi-style Generative Reading Comprehension + +Question: What are the baselines that Masque is compared against? \ No newline at end of file diff --git a/qasper-0782/instruction.md b/qasper-0782/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..8df1621816f2447bc61aca4fd2f58b0b0b1383e2 --- /dev/null +++ b/qasper-0782/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: A Cascade Sequence-to-Sequence Model for Chinese Mandarin Lip Reading + +Question: What was the previous state of the art model for this task? \ No newline at end of file diff --git a/qasper-0785/instruction.md b/qasper-0785/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4e7ec79955e6dca68a425f5ce24d08b88af3299a --- /dev/null +++ b/qasper-0785/instruction.md @@ -0,0 +1,84 @@ +Name of Paper: Dissecting Content and Context in Argumentative Relation Analysis + +Question: Do they report results only on English data? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Argumentative Relation Prediction: Models and Features", + "Models", + "Feature implementation", + "Results", + "Discussion", + "Conclusion", + "Acknowledgments" + ], + "paragraphs": [ + [ + "In recent years we have witnessed a great surge in activity in the area of computational argument analysis (e.g. BIBREF0 , BIBREF1 , BIBREF2 , BIBREF3 ), and the emergence of dedicated venues such as the ACL Argument Mining workshop series starting in 2014 BIBREF4 .", + "Argumentative relation classification is a sub-task of argument analysis that aims to determine relations between argumentative units A and B, for example, A supports B; A attacks B. Consider the following argumentative units (1) and (2), given the topic (0) \u201cMarijuana should be legalized\u201d:", + "This example is modeled in Figure FIGREF3 .", + "It is clear that (1) has a negative stance towards the topic and (2) has a positive stance towards the topic. Moreover, we can say that (2) attacks (1). In discourse, such a relation is often made explicit through discourse markers: (1). However, (2); On the one hand (1), on the other (2); (1), although (2); Admittedly, (2); etc. In the absence of such markers we must determine this relation by assessing the semantics of the individual argumentative units, including (often implicit) world knowledge about how they are related to each other.", + "In this work, we show that argumentative relation classifiers \u2013 when provided with textual context surrounding an argumentative unit's span \u2013 are very prone to neglect the actual textual content of the EAU span. Instead they heavily rely on contextual markers, such as conjunctions or adverbials, as a basis for prediction. We argue that a system's capacity of predicting the correct relation based on the argumentative units' content is important in many circumstances, e.g., when an argumentative debate crosses document boundaries. For example, the prohibition of marijuana debate extends across populations and countries \u2013 argumentative units for this debate can be recovered from thousands of documents scattered across the world wide web. As a consequence, argumentative relation classification systems should not be (immensely) dependent on contextual clues \u2013 in the discussed cross-document setting these clues may even be misleading for such a system, since source and target arguments can be embedded in different textual contexts (e.g., when (1) and (2) stem from different documents it is easy to imagine a textual context where (2) is not introduced by however but instead by an `inverse' form such as e.g. moreover)." + ], + [ + "It is well-known that the rhetorical and argumentative structure of texts bear great similarities. For example, BIBREF5 , BIBREF6 , BIBREF0 observe that elementary discourse units (EDUs) in RST BIBREF7 share great similarity with elementary argumentative units (EAUs) in argumentation analysis. BIBREF8 experiment with a modified version of the Microtext corpus BIBREF9 , which is an extensively annotated albeit small corpus. Similar to us, they separate argumentative units from discursive contextual markers. While BIBREF8 conduct a human evaluation to investigate the separation of Logos and Pathos aspects of arguments, our work investigates how (de-)contextualization of argumentative units affects automatic argumentative relation classification models." + ], + [ + "In this section, we describe different formulations of the argumentative relation classification task and describe features used by our replicated model. In order to test our hypotheses, we propose to group all features into three distinct types." + ], + [ + "Now, we introduce a classification of three different prediction models used in the argumentative relation prediction literature. We will inspect all of them and show that all can suffer from severe issues when focusing (too much) on the context.", + "The model INLINEFORM0 adopts a discourse parsing view on argumentative relation prediction and predicts one outgoing edge for an argumentative unit (one-outgoing edge). Model INLINEFORM1 assumes a connected graph with argumentative units and is tasked with predicting edge labels for unit tuples (labeling relations in a graph). Finally, a model INLINEFORM2 is given two (possibly) unrelated argumentative units and is tasked with predicting connections as well as edge labels (joint edge prediction and labeling).", + " BIBREF13 divide the task into relation prediction INLINEFORM0 and relation class assignment INLINEFORM1 : DISPLAYFORM0 ", + " DISPLAYFORM0 ", + "which the authors describe as argumentative relation identification ( INLINEFORM0 ) and stance detection ( INLINEFORM1 ). In their experiments, INLINEFORM2 , i.e., no distinction is made between features that access only the argument content (EAU span) or only the EAU's embedding context, and some features also consider both (e.g., discourse features). This model adopts a parsing view on argumentative relation classification: every unit is allowed to have only one type of outgoing relation (this follows trivially from the fact that INLINEFORM3 has only one input). Applying such a model to argumentative attack and support relations might impose unrealistic constraints on the resulting argumentation graph: A given premise might in fact attack or support several other premises. The approach may suffice for the case of student argumentative essays, where EAUs are well-framed in a discourse structure, but seems overly restrictive for many other scenarios.", + "Another way of framing the task, is to learn a function DISPLAYFORM0 ", + "Here, an argumentative unit is allowed to be in a attack or support relation to multiple other EAUs. Yet, both INLINEFORM0 and INLINEFORM1 assume that inputs are already linked and only the class of the link is unknown.", + "Thus, we might also model the task in a three-class classification setting to learn a more general function that performs relation prediction and classification jointly (see also, e.g., BIBREF10 ): DISPLAYFORM0 ", + "The model described by Eq. EQREF22 is the most general one: not only does it assume a graph view on argumentative units and their relations (as does Eq. EQREF20 ); in model formulation (Eq. EQREF22 ), an argumentative unit can have no or multiple support or attack relations. It naturally allows for cases where an argumentative unit INLINEFORM0 (supports INLINEFORM1 INLINEFORM2 attacks INLINEFORM3 INLINEFORM4 is-unrelated-to INLINEFORM5 ). Given a set of EAUs mined from different documents, this model enables us to construct a full-fledged argumentation graph." + ], + [ + "Our feature implementation follows the feature descriptions for Stance recognition and link identification in BIBREF13 . These features and variations of them have been used successfully in several successive works (cf. BIBREF1 , BIBREF16 , BIBREF15 ).", + "For any model the features are indexed by INLINEFORM0 . We create a function INLINEFORM1 which maps from feature indices to feature types. In other words, INLINEFORM2 tells us, for any given feature, whether it is content-based ( INLINEFORM3 ), content-ignorant ( INLINEFORM4 ) or full access ( INLINEFORM5 ). The features for, e.g., the joint prediction model INLINEFORM6 of type INLINEFORM7 ( INLINEFORM8 ) can then simply be described as INLINEFORM9 . Recall that features computed on the basis of the EAU span are content-based ( INLINEFORM10 ), features from the EAU-surrounding text are content-ignorant ( INLINEFORM11 ) and features computed from both are denoted by full-access ( INLINEFORM12 ). Details on the extraction of features are provided below.", + "These consist of boolean values indicating whether a certain word appears in the argumentative source or target EAU or both (and separately, their contexts). More precisely, for any classification instance we extract uni-grams from within the span of the EAU (if INLINEFORM0 ) or solely from the sentence-context surrounding the EAUs (if INLINEFORM1 ). Words which occur in both bags are only visible in the full-access setup INLINEFORM2 and are modeled as binary indicators.", + "Such features consist of syntactic production rules extracted from constituency trees \u2013 they are modelled analogously to the lexical features as a bag of production rules. To make a clear division between features derived from the EAU embedding context and features derived from within the EAU span, we divide the constituency tree in two parts, as is illustrated in Figure FIGREF26 .", + "If the EAU is embedded in a covering sentence, we cut the syntax tree at the corresponding edge ( in Figure FIGREF26 ). In this example, the content-ignorant ( INLINEFORM0 ) bag-of-word production rule representation includes the rules INLINEFORM1 and INLINEFORM2 . Analogously to the lexical features, the production rules are modeled as binary indicator features.", + "These features describe shallow statistics such as the ratio of argumentative unit tokens compared to sentence tokens or the position of the argumentative unit in the paragraph. We set these features to zero for the content representation of the argumentative unit and replicate those features that allow us to treat the argumentative unit as a black-box. For example, in the content-based ( INLINEFORM0 ) system that has access only to the EAU, we can compute the #tokens in the EAU, but not the #tokens in EAU divided by #tokens in the sentence. The latter feature is only accessible in the full access system variants. Hence, in the content-based ( INLINEFORM1 ) system most of these statistics are set to zero since they cannot be computed by considering only the EAU span.", + "For the content-based representation we retrieve only discourse relations that are confined within the span of the argumentative unit. In the very frequent case that discourse features cross the boundaries of embedding context and EAU span, we only take them into account for INLINEFORM0 .", + "We use the element-wise sum of 300-dimensional pre-trained GloVe vectors BIBREF24 corresponding to the words within the EAU span ( INLINEFORM0 ) and the words of the EAU-surrounding context ( INLINEFORM1 ). Additionally, we compute the element-wise subtraction of the source EAU vector from the target EAU vector, with the aim of modelling directions in distributional space, similarly to BIBREF25 . Words with no corresponding pre-trained word vector and empty sequences (e.g., no preceding context available) are treated as a zero-vector.", + "Tree-based sentiment annotations are sentiment scores assigned to nodes in constituency parse trees BIBREF26 . We represent these scores by a one-hot vector of dimension 5 (5 is very positive, 1 is very negative). We determine the contextual ( INLINEFORM0 ) sentiment by looking at the highest possible node of the context which does not contain the EAU (ADVP in Figure FIGREF26 ). The sentiment for an EAU span ( INLINEFORM1 ) is assigned to the highest possible node covering the EAU span which does not contain the context sub-tree (S in Figure FIGREF26 ). The full-access ( INLINEFORM2 ) score is assigned to the lowest possible node which covers both the EAU span and its surrounding context (S' in Figure FIGREF26 ). Next to the sentiment scores for the selected tree nodes and analogously to the word embeddings, we also calculate the element-wise subtraction of the one-hot sentiment source vectors from the one-hot sentiment target vectors. This results in three additional vectors corresponding to INLINEFORM3 , INLINEFORM4 and INLINEFORM5 difference vectors." + ], + [ + "Our first step towards our main experiments is to replicate the competitive argumentative relation classifier of BIBREF13 , BIBREF1 . Hence, for comparison purposes, we first formulate the task exactly as it was done in this prior work, using the model formulation in Eq. EQREF17 , which determines the type of outgoing edge from a source (i.e., tree-like view).", + "The results in Table TABREF38 confirm the results of BIBREF13 and suggest that we successfully replicated a large proportion of their features.", + "The results for all three prediction settings (one outgoing edge: INLINEFORM0 , support/attack: INLINEFORM1 and support/attack/neither: INLINEFORM2 ) across all type variables ( INLINEFORM3 , INLINEFORM4 and INLINEFORM5 ) are displayed in Table TABREF39 . All models significantly outperform the majority baseline with respect to macro F1. Intriguingly, the content-ignorant models ( INLINEFORM6 ) always perform significantly better than the models which only have access to the EAUs' content ( INLINEFORM7 , INLINEFORM8 ). In the most general task formulation ( INLINEFORM9 ), we observe that INLINEFORM10 even significantly outperforms the model which has maximum access (seeing both EAU spans and surrounding contexts: INLINEFORM11 ).", + "At first glance, the results of the purely EAU focused systems ( INLINEFORM0 ) are disappointing, since they fall far behind their competitors. On the other hand, their F1 scores are not devastatingly bad. The strong most-frequent-class baseline is significantly outperformed by the content-based ( INLINEFORM1 ) system, across all three prediction settings.", + "In summary our findings are as follows: (i) models which see the EAU span (content-based, INLINEFORM0 ) are significantly outperformed by models that have no access to the span itself (content-ignorant, INLINEFORM1 ) across all settings; (ii) in two of three prediction settings ( INLINEFORM2 and INLINEFORM3 ), the model which only has access to the context even outperforms the model that has access to all information in the input. The fact that using features derived exclusively from the EAU embedding context ( INLINEFORM4 ) can lead to better results than using a full feature-system ( INLINEFORM5 ) suggests that some information from the EAU can even be harmful. Why this is the case, we cannot answer exactly. A plausible cause might be related to the smaller dimension of the feature space, which makes the SVM less likely to overfit. Still, this finding comes as a surprise and calls for further investigation in future work.", + "A system for argumentative relation classification can be applied in one of two settings: single-document or cross-document, as illustrated in Figure FIGREF42 :", + "in the first case (top), a system is tasked to classify EAUs that appear linearly in one document \u2013 here contextual clues can often highlight the relationship between two units. This is the setting we have been considering up to now. However, in the second scenario (bottom), we have moved away from the closed single-document setting and ask the system to classify two EAUs extracted from different document contexts. This setting applies, for instance, when we are mining arguments from multiple sources.", + "In both cases, however, a system that relies more on contextual clues than on the content expressed in the EAUs is problematic: in the single-document setting, such a system will rely on discourse indicators \u2013 whether or not they are justified by content \u2013 and can thus easily be fooled.", + "In the cross-document setting, discourse-based indicators \u2013 being inherently defined with respect to their internal document context \u2013 do not have a defined rhetorical function with respect to EAUs in a separate document and thus a system that has learned to rely on such markers within a single-document setting can be seriously misled.", + "We believe that the cross-document setting should be an important goal in argumentation analysis, since it generalizes better to many debates of interest, where EAUs can be found scattered across thousands of documents. For example, for the topic of legalizing marijuana, EAUs may be mined from millions of documents and thus their relations may naturally extend across document boundaries. If a system learns to over-proportionally attend to the EAUs' surrounding contexts it is prone to making many errors.", + "In what follows we are simulating the effects that an overly context-sensitive classifier could have in a cross-document setting, by modifying our experimental setting, and study the effects on the different model types: In one setup \u2013 we call it randomized-context \u2013 we systematically distort the context of our testing instances by exchanging the context in a randomized manner; in the other setting \u2013 called no-context, we are deleting the context around the ADUs to be classified. Randomized-context simulates an open world debate where argumentative units may occur in different contexts, sometimes with discourse markers indicating an opposite class. In other words, in this setting we want to examine effects when porting a context-sensitive system to a multi-document setting. For example, as seen in Figure FIGREF42 , the context of an argumentative unit may change from \u201cHowever\u201d to \u201cMoreover\u201d \u2013 which can happen naturally in open debates.", + "The results are displayed in Figure FIGREF43 . In the standard setting (Figure FIGREF43 ), the models that have access to the context besides the content ( INLINEFORM0 ) and the models that are only allowed to access the context ( INLINEFORM1 ), always perform better than the content-based models ( INLINEFORM2 ) (bars above zero). However, when we randomly flip contexts of the test instances (Figure FIGREF43 ), or suppress them entirely (Figure FIGREF43 ), the opposite picture emerges: the content-based models always outperform the other models. For some classes (support, INLINEFORM3 ) the difference can exceed 50 F1 percentage points. These two studies, where testing examples are varied regarding their context (randomized-context or no-context) simulates what can be expected if we apply our systems for relation class assignment to EAUs stemming from heterogeneous sources. While the performances of a purely content-based model naturally stays stable, the performance of the other systems decrease notably \u2013 they perform worse than the content-based model.", + "We calculate the ANOVA classification F scores of the features with respect to our three task formulations INLINEFORM0 and INLINEFORM1 . The F percentiles of features extracted from the EAU surrounding text ( INLINEFORM2 ) and features extracted from the EAU span ( INLINEFORM3 ), are displayed in Figure FIGREF50 .", + "It clearly stands out that features obtained from the EAU surrounding context ( INLINEFORM0 ) are assigned much higher scores compared to features stemming from the EAU span ( INLINEFORM1 ). This holds true for all three task formulations and provides further evidence that models \u2013 when given the option \u2013 put a strong focus on contextual clues while neglecting the information provided by the EAU span itself." + ], + [ + "While competitive systems for argumentative relation classification are considered to be robust, our experiments have shown that despite confidence-inspiring scores on unseen testing data, such systems can easily be fooled \u2013 they can deliver strong performance scores although the classifier does not have access to the content of the EAUs. In this respect, we have provided evidence that there is a danger in case models focus too much on rhetorical indicators, in detriment of the context. Thus, the following question arises: How can we prevent argumentation models from modeling arguments or argumentative units and their relations in overly na\u00efve ways? A simple and intuitive way is to dissect EAUs from their surrounding document context. Models trained on data that is restricted to the EAUs' content will be forced to focus on the content of EAUs. We believe that this will enhance the robustness of such models and allows them to generalize to cross-document argument relation classification. The corpus of student essays makes such transformations straightforward: only the EAUs were annotated (e.g., \u201cHowever, INLINEFORM0 A INLINEFORM1 \u201d). If annotations extend over the EAUs (e.g., only full sentences are annotated, \u201c INLINEFORM2 However, A INLINEFORM3 \u201d), such transformations could be performed automatically after a discourse parsing step. When inspecting the student essays corpus, we further observed that an EAU mining step should involve coreference resolution to better capture relations between EAUs that involve anaphors (e.g., \u201cExercising makes you feel better\u201d and \u201cIt INLINEFORM4 increases endorphin levels\u201d).", + "Thus, in order to conduct real-world end-to-end argumentation relation mining for a given topic, we envision a system that addresses three steps: (i) mining of EAUs and (ii) replacement of pronouns in EAUs with referenced entities (e.g., INLINEFORM0 ). Finally (iii), given the cross product of mined EAUs we can apply a model of type INLINEFORM1 to construct a full-fledged argumentation graph, possibly spanning multiple documents. We have shown that in order to properly perform step (iii), we need stronger models that are able to better model EAU contents. Hence, we encourage the argumentation community to test their systems on a decontextualized version of the student essays, including the proposed \u2013 and possibly further extended \u2013 testing setups, to challenge the semantic representation and reasoning capacities of argument analysis models. This will lead to more realistic performance estimates and increased robustness of systems when addressing desirable multi-document tasks." + ], + [ + "We have shown that systems which put too much focus on discourse information may be easily fooled \u2013 an issue which has severe implications when systems are applied to cross-document argumentative relation classification tasks. The strong reliance on contextual clues is also problematic in single-document contexts, where systems can run a risk of assigning relation labels relying on contextual and rhetorical effects \u2013 instead of focusing on content. Hence, we propose that researchers test their argumentative relation classification systems on two alternative versions of the StudentEssay data that reflect different access levels. (i) EAU-span only, where systems only see the EAU spans and (ii) context-only, where systems can only see the EAU-surrounding context. These complementary settings will (i) challenge the semantic capacities of a system, and (ii) unveil the extent to which a system is focusing on the discourse context when making decisions. We will offer our testing environments to the research community through a platform that provides datasets and scripts and a table to trace the results of content-based systems." + ], + [ + "This work has been supported by the German Research Foundation as part of the Research Training Group Adaptive Preparation of Information from Heterogeneous Sources (AIPHES) under grant no. GRK 1994/1 and by the Leibniz ScienceCampus \u201cEmpirical Linguistics and Computational Language Modeling\u201d, supported by the Leibniz Association under grant no. SAS-2015-IDS-LWC and by the Ministry of Science, Research, and Art of Baden-W\u00fcrttemberg. " + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0913/instruction.md b/qasper-0913/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..686361f1807cbd58fcbb78c327b86a7c88e7991d --- /dev/null +++ b/qasper-0913/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Identifying and Understanding User Reactions to Deceptive and Trusted Social News Sources + +Question: What are the nine types? \ No newline at end of file diff --git a/qasper-0914/instruction.md b/qasper-0914/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..9822dc4703e3c20ab95e66d8fc5ff3776754f4ec --- /dev/null +++ b/qasper-0914/instruction.md @@ -0,0 +1,112 @@ +Name of Paper: Discriminative Acoustic Word Embeddings: Recurrent Neural Network-Based Approaches + +Question: How do they represent input features of their model to train embeddings? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related work", + "Approach", + "Training", + "EXPERIMENTS", + "Classification network details", + "Siamese network details", + "Results", + "Effect of model structure", + "Effect of embedding dimensionality", + "Effect of training vocabulary", + "Visualization of embeddings", + "Conclusion" + ], + "paragraphs": [ + [ + "Many speech processing tasks \u2013 such as automatic speech recognition or spoken term detection \u2013 hinge on associating segments of speech signals with word labels. In most systems developed for such tasks, words are broken down into sub-word units such as phones, and models are built for the individual units. An alternative, which has been considered by some researchers, is to consider each entire word segment as a single unit, without assigning parts of it to sub-word units. One motivation for the use of whole-word approaches is that they avoid the need for sub-word models. This is helpful since, despite decades of work on sub-word modeling BIBREF0 , BIBREF1 , it still poses significant challenges. For example, speech processing systems are still hampered by differences in conversational pronunciations BIBREF2 . A second motivation is that considering whole words at once allows us to consider a more flexible set of features and reason over longer time spans.", + "Whole-word approaches typically involve, at some level, template matching. For example, in template-based speech recognition BIBREF3 , BIBREF4 , word scores are computed from dynamic time warping (DTW) distances between an observed segment and training segments of the hypothesized word. In query-by-example search, putative matches are typically found by measuring the DTW distance between the query and segments of the search database BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 . In other words, whole-word approaches often boil down to making decisions about whether two segments are examples of the same word or not.", + "An alternative to DTW that has begun to be explored is the use of acoustic word embeddings (AWEs), or vector representations of spoken word segments. AWEs are representations that can be learned from data, ideally such that the embeddings of two segments corresponding to the same word are close, while embeddings of segments corresponding to different words are far apart. Once word segments are represented via fixed-dimensional embeddings, computing distances is as simple as measuring a cosine or Euclidean distance between two vectors.", + "There has been some, thus far limited, work on acoustic word embeddings, focused on a number of embedding models, training approaches, and tasks BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , BIBREF15 , BIBREF16 . In this paper we explore new embedding models based on recurrent neural networks (RNNs), applied to a word discrimination task related to query-by-example search. RNNs are a natural model class for acoustic word embeddings, since they can handle arbitrary-length sequences. We compare several types of RNN-based embeddings and analyze their properties. Compared to prior embeddings tested on the same task, our best models achieve sizable improvements in average precision." + ], + [ + "We next briefly describe the most closely related prior work.", + "Maas et al. BIBREF9 and Bengio and Heigold BIBREF10 used acoustic word embeddings, based on convolutional neural networks (CNNs), to generate scores for word segments in automatic speech recognition. Maas et al. trained CNNs to predict (continuous-valued) embeddings of the word labels, and used the resulting embeddings to define feature functions in a segmental conditional random field BIBREF17 rescoring system. Bengio and Heigold also developed CNN-based embeddings for lattice rescoring, but with a contrastive loss to separate embeddings of a given word from embeddings of other words.", + "Levin et al. BIBREF11 developed unsupervised embeddings based on representing each word as a vector of DTW distances to a collection of reference word segments. This representation was subsequently used in several applications: a segmental approach for query-by-example search BIBREF12 , lexical clustering BIBREF18 , and unsupervised speech recognition BIBREF19 . Voinea et al. BIBREF15 developed a representation also based on templates, in their case phone templates, designed to be invariant to specific transformations, and showed their robustness on digit classification.", + "Kamper et al. BIBREF13 compared several types of acoustic word embeddings for a word discrimination task related to query-by-example search, finding that embeddings based on convolutional neural networks (CNNs) trained with a contrastive loss outperformed the reference vector approach of Levin et al. BIBREF11 as well as several other CNN and DNN embeddings and DTW using several feature types. There have now been a number of approaches compared on this same task and data BIBREF11 , BIBREF20 , BIBREF21 , BIBREF22 . For a direct comparison with this prior work, in this paper we use the same task and some of the same training losses as Kamper et al., but develop new embedding models based on RNNs.", + "The only prior work of which we are aware using RNNs for acoustic word embeddings is that of Chen et al. BIBREF16 and Chung et al. BIBREF14 . Chen et al. learned a long short-term memory (LSTM) RNN for word classification and used the resulting hidden state vectors as a word embedding in a query-by-example task. The setting was quite specific, however, with a small number of queries and speaker-dependent training. Chung et al. BIBREF14 worked in an unsupervised setting and trained single-layer RNN autoencoders to produce embeddings for a word discrimination task. In this paper we focus on the supervised setting, and compare a variety of RNN-based structures trained with different losses.", + "" + ], + [ + "", + "An acoustic word embedding is a function that takes as input a speech segment corresponding to a word, INLINEFORM0 , where each INLINEFORM1 is a vector of frame-level acoustic features, and outputs a fixed-dimensional vector representing the segment, INLINEFORM2 . The basic embedding model structure we use is shown in Fig. FIGREF1 . The model consists of a deep RNN with some number INLINEFORM3 of stacked layers, whose final hidden state vector is passed as input to a set of INLINEFORM4 of fully connected layers; the output of the final fully connected layer is the embedding INLINEFORM5 .", + "The RNN hidden state at each time frame can be viewed as a representation of the input seen thus far, and its value in the last time frame INLINEFORM0 could itself serve as the final word embedding. The fully connected layers are added to account for the fact that some additional transformation may improve the representation. For example, the hidden state may need to be larger than the desired word embedding dimension, in order to be able to \"remember\" all of the needed intermediate information. Some of that information may not be needed in the final embedding. In addition, the information maintained in the hidden state may not necessarily be discriminative; some additional linear or non-linear transformation may help to learn a discriminative embedding.", + "Within this class of embedding models, we focus on Long Short-Term Memory (LSTM) networks BIBREF23 and Gated Recurrent Unit (GRU) networks BIBREF24 . These are both types of RNNs that include a mechanism for selectively retaining or discarding information at each time frame when updating the hidden state, in order to better utilize long-term context. Both of these RNN variants have been used successfully in speech recognition BIBREF25 , BIBREF26 , BIBREF27 , BIBREF28 .", + "In an LSTM RNN, at each time frame both the hidden state INLINEFORM0 and an associated \u201ccell memory\" vector INLINEFORM1 , are updated and passed on to the next time frame. In other words, each forward edge in Figure FIGREF1 can be viewed as carrying both the cell memory and hidden state vectors. The updates are modulated by the values of several gating vectors, which control the degree to which the cell memory and hidden state are updated in light of new information in the current frame. For a single-layer LSTM network, the updates are as follows:", + " INLINEFORM0 ", + "where INLINEFORM0 , and INLINEFORM1 are all vectors of the same dimensionality, INLINEFORM2 , and INLINEFORM3 are learned weight matrices of the appropriate sizes, INLINEFORM4 and INLINEFORM5 are learned bias vectors, INLINEFORM6 is a componentwise logistic activation, and INLINEFORM7 refers to the Hadamard (componentwise) product.", + "Similarly, in a GRU network, at each time step a GRU cell determines what components of old information are retained, overwritten, or modified in light of the next step in the input sequence. The output from a GRU cell is only the hidden state vector. A GRU cell uses a reset gate INLINEFORM0 and an update gate INLINEFORM1 as described below for a single-layer network: INLINEFORM2 ", + "where INLINEFORM0 , and INLINEFORM1 are all the same dimensionality, INLINEFORM2 , and INLINEFORM3 are learned weight matrices of the appropriate size, and INLINEFORM4 , INLINEFORM5 and INLINEFORM6 are learned bias vectors.", + "All of the above equations refer to single-layer networks. In a deep network, with multiple stacked layers, the same update equations are used in each layer, with the state, cell, and gate vectors replaced by layer-specific vectors INLINEFORM0 and so on for layer INLINEFORM1 . For all but the first layer, the input INLINEFORM2 is replaced by the hidden state vector from the previous layer INLINEFORM3 .", + "For the fully connected layers, we use rectified linear unit (ReLU) BIBREF29 activation, except for the final layer which depends on the form of supervision and loss used in training.", + "" + ], + [ + "We train the RNN-based embedding models using a set of pre-segmented spoken words. We use two main training approaches, inspired by prior work but with some differences in the details. As in BIBREF13 , BIBREF10 , our first approach is to use the word labels of the training segments and train the networks to classify the word. In this case, the final layer of INLINEFORM0 is a log-softmax layer. Here we are limited to the subset of the training set that has a sufficient number of segments per word to train a good classifier, and the output dimensionality is equal to the number of words (but see BIBREF13 for a study of varying the dimensionality in such a classifier-based embedding model by introducing a bottleneck layer). This model is trained end-to-end and is optimized with a cross entropy loss. Although labeled data is necessarily limited, the hope is that the learned models will be useful even when applied to spoken examples of words not previously seen in the training data. For words not seen in training, the embeddings should correspond to some measure of similarity of the word to the training words, measured via the posterior probabilities of the previously seen words. In the experiments below, we examine this assumption by analyzing performance on words that appear in the training data compared to those that do not.", + "The second training approach, based on earlier work of Kamper et al. BIBREF13 , is to train \"Siamese\" networks BIBREF30 . In this approach, full supervision is not needed; rather, we use weak supervision in the form of pairs of segments labeled as same or different. The base model remains the same as before\u2014an RNN followed by a set of fully connected layers\u2014but the final layer is no longer a softmax but rather a linear activation layer of arbitrary size. In order to learn the parameters, we simultaneously feed three word segments through three copies of our model (i.e. three networks with shared weights). One input segment is an \u201canchor\", INLINEFORM0 , the second is another segment with the same word label, INLINEFORM1 , and the third is a segment corresponding to a different word label, INLINEFORM2 . Then, the network is trained using a \u201ccos-hinge\" loss:", + " DISPLAYFORM0 ", + "where INLINEFORM0 is the cosine distance between INLINEFORM1 . Unlike cross entropy training, here we directly aim to optimize relative (cosine) distance between same and different word pairs. For tasks such as query-by-example search, this training loss better respects our end objective, and can use more data since neither fully labeled data nor any minimum number of examples of each word should be needed.", + "" + ], + [ + "", + "Our end goal is to improve performance on downstream tasks requiring accurate word discrimination. In this paper we use an intermediate task that more directly tests whether same- and different-word pairs have the expected relationship. and that allows us to compare to a variety of prior work. Specifically, we use the word discrimination task of Carlin et al. BIBREF20 , which is similar to a query-by-example task where the word segmentations are known. The evaluation consists of determining, for each pair of evaluation segments, whether they are examples of the same or different words, and measuring performance via the average precision (AP). We do this by measuring the cosine similarity between their acoustic word embeddings and declaring them to be the same if the distance is below a threshold. By sweeping the threshold, we obtain a precision-recall curve from which we compute the AP.", + "The data used for this task is drawn from the Switchboard conversational English corpus BIBREF31 . The word segments range from 50 to 200 frames in length. The acoustic features in each frame (the input to the word embedding models INLINEFORM0 ) are 39-dimensional MFCCs+ INLINEFORM1 + INLINEFORM2 . We use the same train, development, and test partitions as in prior work BIBREF13 , BIBREF11 , and the same acoustic features as in BIBREF13 , for as direct a comparison as possible. The train set contains approximately 10k example segments, while dev and test each contain approximately 11k segments (corresponding to about 60M pairs for computing the dev/test AP). As in BIBREF13 , when training the classification-based embeddings, we use a subset of the training set containing all word types with a minimum of 3 occurrences, reducing the training set size to approximately 9k segments.", + "When training the Siamese networks, the training data consists of all of the same-word pairs in the full training set (approximately 100k pairs). For each such training pair, we randomly sample a third example belonging to a different word type, as required for the INLINEFORM0 loss.", + "" + ], + [ + "Our classifier-based embeddings use LSTM or GRU networks with 2\u20134 stacked layers and 1\u20133 fully connected layers. The final embedding dimensionality is equal to the number of unique word labels in the training set, which is 1061. The recurrent hidden state dimensionality is fixed at 512 and dropout BIBREF32 between stacked recurrent layers is used with probability INLINEFORM0 . The fully connected hidden layer dimensionality is fixed at 1024. Rectified linear unit (ReLU) non-linearities and dropout with INLINEFORM1 are used between fully-connected layers. However, between the final recurrent hidden state output and the first fully-connected layer no non-linearity or dropout is applied. These settings were determined through experiments on the development set.", + "The classifier network is trained with a cross entropy loss and optimized using stochastic gradient descent (SGD) with Nesterov momentum BIBREF33 . The learning rate is initialized at 0.1 and is reduced by a factor of 10 according to the following heuristic: If 99% of the current epoch's average batch loss is greater than the running average of batch losses over the last 3 epochs, this is considered a plateau; if there are 3 consecutive plateau epochs, then the learning rate is reduced. Training stops when reducing the learning rate no longer improves dev set AP. Then, the model from the epoch corresponding to the the best dev set AP is chosen. Several other optimizers\u2014Adagrad BIBREF34 , Adadelta BIBREF35 , and Adam BIBREF36 \u2014were explored in initial experiments on the dev set, but all reported results were obtained using SGD with Nesterov momentum.", + "" + ], + [ + "For experiments with Siamese networks, we initialize (warm-start) the networks with the tuned classification network, removing the final log-softmax layer and replacing it with a linear layer of size equal to the desired embedding dimensionality. We explored embeddings with dimensionalities between 8 and 2048. We use a margin of 0.4 in the cos-hinge loss.", + "In training the Siamese networks, each training mini-batch consists of INLINEFORM0 triplets. INLINEFORM1 triplets are of the form INLINEFORM2 where INLINEFORM3 and INLINEFORM4 are examples of the same class (a pair from the 100k same-word pair set) and INLINEFORM5 is a randomly sampled example from a different class. Then, for each of these INLINEFORM6 triplets INLINEFORM7 , an additional triplet INLINEFORM8 is added to the mini-batch to allow all segments to serve as anchors. This is a slight departure from earlier work BIBREF13 , which we found to improve stability in training and performance on the development set.", + "In preliminary experiments, we compared two methods for choosing the negative examples INLINEFORM0 during training, a uniform sampling approach and a non-uniform one. In the case of uniform sampling, we sample INLINEFORM1 uniformly at random from the full set of training examples with labels different from INLINEFORM2 . This sampling method requires only word-pair supervision. In the case of non-uniform sampling, INLINEFORM3 is sampled in two steps. First, we construct a distribution INLINEFORM4 over word labels INLINEFORM5 and sample a different label from it. Second, we sample an example uniformly from within the subset with the chosen label. The goal of this method is to speed up training by targeting pairs that violate the margin constraint. To construct the multinomial PMF INLINEFORM6 , we maintain an INLINEFORM7 matrix INLINEFORM8 , where INLINEFORM9 is the number of unique word labels in training. Each word label corresponds to an integer INLINEFORM10 INLINEFORM11 [1, INLINEFORM12 ] and therefore a row in INLINEFORM13 . The values in a row of INLINEFORM14 are considered similarity scores, and we can retrieve the desired PMF for each row by normalizing by its sum.", + "At the start of each epoch, we initialize INLINEFORM0 with 0's along the diagonal and 1's elsewhere (which reduces to uniform sampling). For each training pair INLINEFORM1 , we update INLINEFORM2 for both INLINEFORM3 and INLINEFORM4 :", + " INLINEFORM0 ", + "The PMFs INLINEFORM0 are updated after the forward pass of an entire mini-batch. The constant INLINEFORM1 enforces a potentially stronger constraint than is used in the INLINEFORM2 loss, in order to promote diverse sampling. In all experiments, we set INLINEFORM3 . This is a heuristic approach, and it would be interesting to consider various alternatives. Preliminary experiments showed that the non-uniform sampling method outperformed uniform sampling, and in the following we report results with non-uniform sampling.", + "We optimize the Siamese network model using SGD with Nesterov momentum for 15 epochs. The learning rate is initialized to 0.001 and dropped every 3 epochs until no improvement is seen on the dev set. The final model is taken from the epoch with the highest dev set AP. All models were implemented in Torch BIBREF37 and used the rnn library of BIBREF38 .", + "" + ], + [ + " Based on development set results, our final embedding models are LSTM networks with 3 stacked layers and 3 fully connected layers, with output dimensionality of 1024 in the case of Siamese networks. Final test set results are given in Table TABREF7 . We include a comparison with the best prior results on this task from BIBREF13 , as well as the result of using standard DTW on the input MFCCs (reproduced from BIBREF13 ) and the best prior result using DTW, obtained with frame features learned with correlated autoencoders BIBREF21 . Both classifier and Siamese LSTM embedding models outperform all prior results on this task of which we are aware.", + "We next analyze the effects of model design choices, as well as the learned embeddings themselves.", + "" + ], + [ + "Table TABREF10 shows the effect on development set performance of the number of stacked layers INLINEFORM0 , the number of fully connected layers INLINEFORM1 , and LSTM vs. GRU cells, for classifier-based embeddings. The best performance in this experiment is achieved by the LSTM network with INLINEFORM2 . However, performance still seems to be improving with additional layers, suggesting that we may be able to further improve performance by adding even more layers of either type. However, we fixed the model to INLINEFORM3 in order to allow for more experimentation and analysis within a reasonable time.", + "Table TABREF10 reveals an interesting trend. When only one fully connected layer is used, the GRU networks outperform the LSTMs given a sufficient number of stacked layers. On the other hand, once we add more fully connected layers, the LSTMs outperform the GRUs. In the first few lines of Table TABREF10 , we use 2, 3, and 4 layer stacks of LSTMs and GRUs while holding fixed the number of fully-connected layers at INLINEFORM0 . There is clear utility in stacking additional layers; however, even with 4 stacked layers the RNNs still underperform the CNN-based embeddings of BIBREF13 until we begin adding fully connected layers.", + "After exploring a variety of stacked RNNs, we fixed the stack to 3 layers and varied the number of fully connected layers. The value of each additional fully connected layer is clearly greater than that of adding stacked layers. All networks trained with 2 or 3 fully connected layers obtain more than 0.4 AP on the development set, while stacked RNNs with 1 fully connected layer are at around 0.3 AP or less. This may raise the question of whether some simple fully connected model may be all that is needed; however, previous work has shown that this approach is not competitive BIBREF13 , and convolutional or recurrent layers are needed to summarize arbitrary-length segments into a fixed-dimensional representation.", + "" + ], + [ + "For the Siamese networks, we varied the output embedding dimensionality, as shown in Fig. FIGREF11 . This analysis shows that the embeddings learned by the Siamese RNN network are quite robust to reduced dimensionality, outperforming the classifier model for all dimensionalities 32 or higher and outperforming previously reported dev set performance with CNN-based embeddings BIBREF13 for all dimensionalities INLINEFORM0 .", + "" + ], + [ + "We might expect the learned embeddings to be more accurate for words that are seen in training than for ones that are not. Fig. FIGREF11 measures this effect by showing performance as a function of the number of occurrences of the dev words in the training set. Indeed, both model types are much more successful for in-vocabulary words, and their performance improves the higher the training frequency of the words. However, performance increases more quickly for the Siamese network than for the classifier as training frequency increases. This may be due to the fact that, if a word type occurs at least INLINEFORM0 times in the classifier training set, then it occurs at least INLINEFORM1 times in the Siamese paired training data.", + "" + ], + [ + "In order to gain a better qualitative understanding of the differences between clasiffier and Siamese-based embeddings, and of the learned embedding space more generally, we plot a two-dimensional visualization of some of our learned embeddings via t-SNE BIBREF40 in Fig. FIGREF12 . For both classifier and Siamese embeddings, there is a marked difference in the quality of clusters formed by embeddings of words that were previously seen vs. previously unseen in training. However, the Siamese network embeddings appear to have better relative distances between word clusters with similar and dissimilar pronunciations. For example, the word programs appears equidistant from problems and problem in the classifier-based embedding space, but in the Siamese embedding space problems falls between problem and programs. Similarly, the cluster for democracy shifts with respect to actually and especially to better respect differences in pronunciation. More study of learned embeddings, using more data and word types, is needed to confirm such patterns in general. Improvements in unseen word embeddings from the classifier embedding space to the Siamese embedding space (such as for democracy, morning, and basketball) are a likely result of optimizing the model for relative distances between words.", + "" + ], + [ + "", + "Our main finding is that RNN-based acoustic word embeddings outperform prior approaches, as measured via a word discrimination task related to query-by-example search. Our best results are obtained with deep LSTM RNNs with a combination of several stacked layers and several fully connected layers, optimized with a contrastive Siamese loss. Siamese networks have the benefit that, for any given training data set, they are effectively trained on a much larger set, in the sense that they measure a loss and gradient for every possible pair of data points. Our experiments suggest that the models could still be improved with additional layers. In addition, we have found that, for the purposes of acoustic word embeddings, fully connected layers are very important and have a more significant effect per layer than stacked layers, particularly when trained with the cross entropy loss function.", + "These experiments represent an initial exploration of sequential neural models for acoustic word embeddings. There are a number of directions for further work. For example, while our analyses suggest that Siamese networks are better than classifier-based models at embedding previously unseen words, our best embeddings are still much poorer for unseen words. Improvements in this direction may come from larger training sets, or may require new models that better model the shared structure between words. Other directions for future work include additional forms of supervision and training, as well as application to downstream tasks." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0924/instruction.md b/qasper-0924/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0b14b97836fe155d30a354e1ed5befd993db6343 --- /dev/null +++ b/qasper-0924/instruction.md @@ -0,0 +1,75 @@ +Name of Paper: Characterizing Diabetes, Diet, Exercise, and Obesity Comments on Twitter + +Question: Do they evaluate only on English data? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Methods", + "Data Collection", + "Topic Discovery", + "Topic Content Analysis", + "Results", + "Discussion", + "Conclusion", + "Conflict of interest", + "Acknowledgement" + ], + "paragraphs": [ + [ + "The global prevalence of obesity has doubled between 1980 and 2014, with more than 1.9 billion adults considered as overweight and over 600 million adults considered as obese in 2014 BIBREF0 . Since the 1970s, obesity has risen 37 percent affecting 25 percent of the U.S. adults BIBREF1 . Similar upward trends of obesity have been found in youth populations, with a 60% increase in preschool aged children between 1990 and 2010 BIBREF2 . Overweight and obesity are the fifth leading risk for global deaths according to the European Association for the Study of Obesity BIBREF0 . Excess energy intake and inadequate energy expenditure both contribute to weight gain and diabetes BIBREF3 , BIBREF4 .", + "Obesity can be reduced through modifiable lifestyle behaviors such as diet and exercise BIBREF4 . There are several comorbidities associated with being overweight or obese, such as diabetes BIBREF5 . The prevalence of diabetes in adults has risen globally from 4.7% in 1980 to 8.5% in 2014. Current projections estimate that by 2050, 29 million Americans will be diagnosed with type 2 diabetes, which is a 165% increase from the 11 million diagnosed in 2002 BIBREF6 . Studies show that there are strong relations among diabetes, diet, exercise, and obesity (DDEO) BIBREF7 , BIBREF4 , BIBREF8 , BIBREF9 ; however, the general public's perception of DDEO remains limited to survey-based studies BIBREF10 .", + "The growth of social media has provided a research opportunity to track public behaviors, information, and opinions about common health issues. It is estimated that the number of social media users will increase from 2.34 billion in 2016 to 2.95 billion in 2020 BIBREF11 . Twitter has 316 million users worldwide BIBREF12 providing a unique opportunity to understand users' opinions with respect to the most common health issues BIBREF13 . Publicly available Twitter posts have facilitated data collection and leveraged the research at the intersection of public health and data science; thus, informing the research community of major opinions and topics of interest among the general population BIBREF14 , BIBREF15 , BIBREF16 that cannot otherwise be collected through traditional means of research (e.g., surveys, interviews, focus groups) BIBREF17 , BIBREF18 . Furthermore, analyzing Twitter data can help health organizations such as state health departments and large healthcare systems to provide health advice and track health opinions of their populations and provide effective health advice when needed BIBREF13 .", + "Among computational methods to analyze tweets, computational linguistics is a well-known developed approach to gain insight into a population, track health issues, and discover new knowledge BIBREF19 , BIBREF20 , BIBREF21 , BIBREF22 . Twitter data has been used for a wide range of health and non-health related applications, such as stock market BIBREF23 and election analysis BIBREF24 . Some examples of Twitter data analysis for health-related topics include: flu BIBREF25 , BIBREF26 , BIBREF27 , BIBREF28 , BIBREF29 , BIBREF30 , mental health BIBREF31 , Ebola BIBREF32 , BIBREF33 , Zika BIBREF34 , medication use BIBREF35 , BIBREF36 , BIBREF37 , diabetes BIBREF38 , and weight loss and obesity BIBREF39 , BIBREF40 , BIBREF41 , BIBREF42 , BIBREF21 .", + "The previous Twitter studies have dealt with extracting common topics of one health issue discussed by the users to better understand common themes; however, this study utilizes an innovative approach to computationally analyze unstructured health related text data exchanged via Twitter to characterize health opinions regarding four common health issues, including diabetes, diet, exercise and obesity (DDEO) on a population level. This study identifies the characteristics of the most common health opinions with respect to DDEO and discloses public perception of the relationship among diabetes, diet, exercise and obesity. These common public opinions/topics and perceptions can be used by providers and public health agencies to better understand the common opinions of their population denominators in regard to DDEO, and reflect upon those opinions accordingly." + ], + [ + "Our approach uses semantic and linguistics analyses for disclosing health characteristics of opinions in tweets containing DDEO words. The present study included three phases: data collection, topic discovery, and topic-content analysis." + ], + [ + "This phase collected tweets using Twitter's Application Programming Interfaces (API) BIBREF43 . Within the Twitter API, diabetes, diet, exercise, and obesity were selected as the related words BIBREF4 and the related health areas BIBREF19 . Twitter's APIs provides both historic and real-time data collections. The latter method randomly collects 1% of publicly available tweets. This paper used the real-time method to randomly collect 10% of publicly available English tweets using several pre-defined DDEO-related queries (Table TABREF6 ) within a specific time frame. We used the queries to collect approximately 4.5 million related tweets between 06/01/2016 and 06/30/2016. The data will be available in the first author's website. Figure FIGREF3 shows a sample of collected tweets in this research." + ], + [ + "To discover topics from the collected tweets, we used a topic modeling approach that fuzzy clusters the semantically related words such as assigning \u201cdiabetes\", \u201ccancer\", and \u201cinfluenza\" into a topic that has an overall \u201cdisease\" theme BIBREF44 , BIBREF45 . Topic modeling has a wide range of applications in health and medical domains such as predicting protein-protein relationships based on the literature knowledge BIBREF46 , discovering relevant clinical concepts and structures in patients' health records BIBREF47 , and identifying patterns of clinical events in a cohort of brain cancer patients BIBREF48 .", + "Among topic models, Latent Dirichlet Allocation (LDA) BIBREF49 is the most popular effective model BIBREF50 , BIBREF19 as studies have shown that LDA is an effective computational linguistics model for discovering topics in a corpus BIBREF51 , BIBREF52 . LDA assumes that a corpus contains topics such that each word in each document can be assigned to the topics with different degrees of membership BIBREF53 , BIBREF54 , BIBREF55 .", + "Twitter users can post their opinions or share information about a subject to the public. Identifying the main topics of users' tweets provides an interesting point of reference, but conceptualizing larger subtopics of millions of tweets can reveal valuable insight to users' opinions. The topic discovery component of the study approach uses LDA to find main topics, themes, and opinions in the collected tweets.", + "We used the Mallet implementation of LDA BIBREF49 , BIBREF56 with its default settings to explore opinions in the tweets. Before identifying the opinions, two pre-processing steps were implemented: (1) using a standard list for removing stop words, that do not have semantic value for analysis (such as \u201cthe\"); and, (2) finding the optimum number of topics. To determine a proper number of topics, log-likelihood estimation with 80% of tweets for training and 20% of tweets for testing was used to find the highest log-likelihood, as it is the optimum number of topics BIBREF57 . The highest log-likelihood was determined 425 topics." + ], + [ + "The topic content analysis component used an objective interpretation approach with a lexicon-based approach to analyze the content of topics. The lexicon-based approach uses dictionaries to disclose the semantic orientation of words in a topic. Linguistic Inquiry and Word Count (LIWC) is a linguistics analysis tool that reveals thoughts, feelings, personality, and motivations in a corpus BIBREF58 , BIBREF59 , BIBREF60 . LIWC has accepted rate of sensitivity, specificity, and English proficiency measures BIBREF61 . LIWC has a health related dictionary that can help to find whether a topic contains words associated with health. In this analysis, we used LIWC to find health related topics." + ], + [ + "Obesity and Diabetes showed the highest and the lowest number of tweets (51.7% and 8.0%). Diet and Exercise formed 23.7% and 16.6% of the tweets (Table TABREF6 ).", + "Out of all 4.5 million DDEO-related tweets returned by Tweeter's API, the LDA found 425 topics. We used LIWC to filter the detected 425 topics and found 222 health-related topics. Additionally, we labeled topics based on the availability of DDEO words. For example, if a topic had \u201cdiet\", we labeled it as a diet-related topic. As expected and driven by the initial Tweeter API's query, common topics were Diabetes, Diet, Exercise, and Obesity (DDEO). (Table TABREF7 ) shows that the highest and the lowest number of topics were related to exercise and diabetes (80 and 21 out of 222). Diet and Obesity had almost similar rates (58 and 63 out of 222).", + "Each of the DDEO topics included several common subtopics including both DDEO and non-DDEO terms discovered by the LDA algorithm (Table TABREF7 ). Common subtopics for \u201cDiabetes\", in order of frequency, included type 2 diabetes, obesity, diet, exercise, blood pressure, heart attack, yoga, and Alzheimer. Common subtopics for \u201cDiet\" included obesity, exercise, weight loss [medicine], celebrities, vegetarian, diabetes, religious diet, pregnancy, and mental health. Frequent subtopics for \u201cExercise\" included fitness, obesity, daily plan, diet, brain, diabetes, and computer games. And finally, the most common subtopics for \u201cObesity\" included diet, exercise, children, diabetes, Alzheimer, and cancer (Table TABREF7 ). Table TABREF8 provides illustrative examples for each of the topics and subtopics.", + "Further exploration of the subtopics revealed additional patterns of interest (Tables TABREF7 and TABREF8 ). We found 21 diabetes-related topics with 8 subtopics. While type 2 diabetes was the most frequent of the sub-topics, heart attack, Yoga, and Alzheimer are the least frequent subtopics for diabetes. Diet had a wide variety of emerging themes ranging from celebrity diet (e.g., Beyonce) to religious diet (e.g., Ramadan). Diet was detected in 63 topics with 10 subtopics; obesity, and pregnancy and mental health were the most and the least discussed obesity-related topics, respectively. Exploring the themes for Exercise subtopics revealed subjects such as computer games (e.g., Pokemon-Go) and brain exercises (e.g., memory improvement). Exercise had 7 subtopics with fitness as the most discussed subtopic and computer games as the least discussed subtopic. Finally, Obesity themes showed topics such as Alzheimer (e.g., research studies) and cancer (e.g., breast cancer). Obesity had the lowest diversity of subtopics: six with diet as the most discussed subtopic, and Alzheimer and cancer as the least discussed subtopics.", + "Diabetes subtopics show the relation between diabetes and exercise, diet, and obesity. Subtopics of diabetes revealed that users post about the relationship between diabetes and other diseases such as heart attack (Tables TABREF7 and TABREF8 ). The subtopic Alzheimer is also shown in the obesity subtopics. This overlap between categories prompts the discussion of research and linkages among obesity, diabetes, and Alzheimer's disease. Type 2 diabetes was another subtopic expressed by users and scientifically documented in the literature.", + "The main DDEO topics showed some level of interrelationship by appearing as subtopics of other DDEO topics. The words with italic and underline styles in Table 2 demonstrate the relation among the four DDEO areas. Our results show users' interest about posting their opinions, sharing information, and conversing about exercise & diabetes, exercise & diet, diabetes & diet, diabetes & obesity, and diet & obesity (Figure FIGREF9 ). The strongest correlation among the topics was determined to be between exercise and obesity ( INLINEFORM0 ). Other notable correlations were: diabetes and obesity ( INLINEFORM1 ), and diet and obesity ( INLINEFORM2 )." + ], + [ + "Diabetes, diet, exercise, and obesity are common public health related opinions. Analyzing individual- level opinions by automated algorithmic techniques can be a useful approach to better characterize health opinions of a population. Traditional public health polls and surveys are limited by a small sample size; however, Twitter provides a platform to capture an array of opinions and shared information a expressed in the words of the tweeter. Studies show that Twitter data can be used to discover trending topics, and that there is a strong correlation between Twitter health conversations and Centers for Disease Control and Prevention (CDC) statistics BIBREF62 .", + "This research provides a computational content analysis approach to conduct a deep analysis using a large data set of tweets. Our framework decodes public health opinions in DDEO related tweets, which can be applied to other public health issues. Among health-related subtopics, there are a wide range of topics from diseases to personal experiences such as participating in religious activities or vegetarian diets.", + "Diabetes subtopics showed the relationship between diabetes and exercise, diet, and obesity (Tables TABREF7 and TABREF8 ). Subtopics of diabetes revealed that users posted about the relation between diabetes and other diseases such as heart attack. The subtopic Alzheimer is also shown in the obesity subtopics. This overlap between categories prompts the discussion of research and linkages among obesity, diabetes, and Alzheimer's disease. Type 2 diabetes was another subtopic that was also expressed by users and scientifically documented in the literature. The inclusion of Yoga in posts about diabetes is interesting. While yoga would certainly be labeled as a form of fitness, when considering the post, it was insightful to see discussion on the mental health benefits that yoga offers to those living with diabetes BIBREF63 .", + "Diet had the highest number of subtopics. For example, religious diet activities such as fasting during the month of Ramadan for Muslims incorporated two subtopics categorized under the diet topic (Tables TABREF7 and TABREF8 ). This information has implications for the type of diets that are being practiced in the religious community, but may help inform religious scholars who focus on health and psychological conditions during fasting. Other religions such as Judaism, Christianity, and Taoism have periods of fasting that were not captured in our data collection, which may have been due to lack of posts or the timeframe in which we collected data. The diet plans of celebrities were also considered influential to explaining and informing diet opinions of Twitter users BIBREF64 .", + "Exercise themes show the Twitter users' association of exercise with \u201cbrain\" benefits such as increased memory and cognitive performance (Tables TABREF7 and TABREF8 ) BIBREF65 . The topics also confirm that exercising is associated with controlling diabetes and assisting with meal planning BIBREF66 , BIBREF9 , and obesity BIBREF67 . Additionally, we found the Twitter users mentioned exercise topics about the use of computer games that assist with exercising. The recent mobile gaming phenomenon Pokeman-Go game BIBREF68 was highly associated with the exercise topic. Pokemon-Go allows users to operate in a virtual environment while simultaneously functioning in the real word. Capturing Pokemons, battling characters, and finding physical locations for meeting other users required physically activity to reach predefined locations. These themes reflect on the potential of augmented reality in increasing patients' physical activity levels BIBREF69 .", + "Obesity had the lowest number of subtopics in our study. Three of the subtopics were related to other diseases such as diabetes (Tables TABREF7 and TABREF8 ). The scholarly literature has well documented the possible linkages between obesity and chronic diseases such as diabetes BIBREF1 as supported by the study results. The topic of children is another prominent subtopic associated with obesity. There has been an increasing number of opinions in regard to child obesity and national health campaigns that have been developed to encourage physical activity among children BIBREF70 . Alzheimer was also identified as a topic under obesity. Although considered a perplexing finding, recent studies have been conducted to identify possible correlation between obesity and Alzheimer's disease BIBREF71 , BIBREF72 , BIBREF73 . Indeed, Twitter users have expressed opinions about the study of Alzheimer's disease and the linkage between these two topics.", + "This paper addresses a need for clinical providers, public health experts, and social scientists to utilize a large conversational dataset to collect and utilize population level opinions and information needs. Although our framework is applied to Twitter, the applications from this study can be used in patient communication devices monitored by physicians or weight management interventions with social media accounts, and support large scale population-wide initiatives to promote healthy behaviors and preventative measures for diabetes, diet, exercise, and obesity.", + "This research has some limitations. First, our DDEO analysis does not take geographical location of the Twitter users into consideration and thus does not reveal if certain geographical differences exists. Second, we used a limited number of queries to select the initial pool of tweets, thus perhaps missing tweets that may have been relevant to DDEO but have used unusual terms referenced. Third, our analysis only included tweets generated in one month; however, as our previous work has demonstrated BIBREF42 , public opinion can change during a year. Additionally, we did not track individuals across time to detect changes in common themes discussed. Our future research plans includes introducing a dynamic framework to collect and analyze DDEO related tweets during extended time periods (multiple months) and incorporating spatial analysis of DDEO-related tweets." + ], + [ + "This study represents the first step in developing routine processes to collect, analyze, and interpret DDEO-related posts to social media around health-related topics and presents a transdisciplinary approach to analyzing public discussions around health topics. With 2.34 billion social media users in 2016, the ability to collect and synthesize social media data will continue to grow. Developing methods to make this process more streamlined and robust will allow for more rapid identification of public health trends in real time.", + "Note: Amir Karami will handle correspondence at all stages of refereeing and publication." + ], + [ + "The authors state that they have no conflict of interest." + ], + [ + "This research was partially supported by the first author's startup research funding provided by the University of South Carolina, School of Library and Information Science. We thank Jill Chappell-Fail and Jeff Salter at the University of South Carolina College of Information and Communications for assistance with technical support.", + "References" + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-0970/instruction.md b/qasper-0970/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..2b0a6c5f2b46a5d8b1237db584fcde8906d45a35 --- /dev/null +++ b/qasper-0970/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Recognizing Musical Entities in User-generated Content + +Question: What are their results on the entity recognition task? \ No newline at end of file diff --git a/qasper-0984/instruction.md b/qasper-0984/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..a67ec9a9c0fde584d234ecf14e89c896f4463374 --- /dev/null +++ b/qasper-0984/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: QuaRel: A Dataset and Models for Answering Questions about Qualitative Relationships + +Question: Which off-the-shelf tools do they use on QuaRel? \ No newline at end of file diff --git a/qasper-1001/instruction.md b/qasper-1001/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..27ba2f6c4f46a887314ce5a42f7006e0d822aee0 --- /dev/null +++ b/qasper-1001/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Multi-Task Bidirectional Transformer Representations for Irony Detection + +Question: Why is being feature-engineering free an advantage? \ No newline at end of file diff --git a/qasper-1006/instruction.md b/qasper-1006/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..2e9a4050bbcdb0a59a5c47488a8b7a6495fa5207 --- /dev/null +++ b/qasper-1006/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Evaluating Rewards for Question Generation Models + +Question: What human evaluation metrics were used in the paper? \ No newline at end of file diff --git a/qasper-1009/instruction.md b/qasper-1009/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..767c20c23d34ad7aa3c6076fe46ca9ce17e34945 --- /dev/null +++ b/qasper-1009/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Gated Convolutional Neural Networks for Domain Adaptation + +Question: Are there conceptual benefits to using GCNs over more complex architectures like attention? \ No newline at end of file diff --git a/qasper-1031/instruction.md b/qasper-1031/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..30262adccacdbd0d394eea95771517c4edd9e591 --- /dev/null +++ b/qasper-1031/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Analysing Coreference in Transformer Outputs + +Question: Which coreference phenomena are analyzed? \ No newline at end of file diff --git a/qasper-1055/instruction.md b/qasper-1055/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..51045c0b16433248adb712b62d1e9d19f2c464cb --- /dev/null +++ b/qasper-1055/instruction.md @@ -0,0 +1,155 @@ +Name of Paper: Cohesion and Coalition Formation in the European Parliament: Roll-Call Votes and Twitter Activities + +Question: Do the authors mention any possible confounds in their study? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Abstract", + "Introduction", + "Related work", + "Methods", + "Co-voting measured by agreement", + "A network-based measure of co-voting", + "Measuring cohesion and coalitions on Twitter", + "Cohesion of political groups", + "Coalitions in the European Parliament", + "Discussion", + "Conclusions", + "Acknowledgments" + ], + "paragraphs": [ + [ + "We study the cohesion within and the coalitions between political groups in the Eighth European Parliament (2014\u20132019) by analyzing two entirely different aspects of the behavior of the Members of the European Parliament (MEPs) in the policy-making processes. On one hand, we analyze their co-voting patterns and, on the other, their retweeting behavior. We make use of two diverse datasets in the analysis. The first one is the roll-call vote dataset, where cohesion is regarded as the tendency to co-vote within a group, and a coalition is formed when the members of several groups exhibit a high degree of co-voting agreement on a subject. The second dataset comes from Twitter; it captures the retweeting (i.e., endorsing) behavior of the MEPs and implies cohesion (retweets within the same group) and coalitions (retweets between groups) from a completely different perspective.", + "We employ two different methodologies to analyze the cohesion and coalitions. The first one is based on Krippendorff's Alpha reliability, used to measure the agreement between raters in data-analysis scenarios, and the second one is based on Exponential Random Graph Models, often used in social-network analysis. We give general insights into the cohesion of political groups in the European Parliament, explore whether coalitions are formed in the same way for different policy areas, and examine to what degree the retweeting behavior of MEPs corresponds to their co-voting patterns. A novel and interesting aspect of our work is the relationship between the co-voting and retweeting patterns." + ], + [ + "Social-media activities often reflect phenomena that occur in other complex systems. By observing social networks and the content propagated through these networks, we can describe or even predict the interplay between the observed social-media activities and another complex system that is more difficult, if not impossible, to monitor. There are numerous studies reported in the literature that successfully correlate social-media activities to phenomena like election outcomes BIBREF0 , BIBREF1 or stock-price movements BIBREF2 , BIBREF3 .", + "In this paper we study the cohesion and coalitions exhibited by political groups in the Eighth European Parliament (2014\u20132019). We analyze two entirely different aspects of how the Members of the European Parliament (MEPs) behave in policy-making processes. On one hand, we analyze their co-voting patterns and, on the other, their retweeting (i.e., endorsing) behavior.", + "We use two diverse datasets in the analysis: the roll-call votes and the Twitter data. A roll-call vote (RCV) is a vote in the parliament in which the names of the MEPs are recorded along with their votes. The RCV data is available as part of the minutes of the parliament's plenary sessions. From this perspective, cohesion is seen as the tendency to co-vote (i.e., cast the same vote) within a group, and a coalition is formed when members of two or more groups exhibit a high degree of co-voting on a subject. The second dataset comes from Twitter. It captures the retweeting behavior of MEPs and implies cohesion (retweets within the same group) and coalitions (retweets between the groups) from a completely different perspective.", + "With over 300 million monthly active users and 500 million tweets posted daily, Twitter is one of the most popular social networks. Twitter allows its users to post short messages (tweets) and to follow other users. A user who follows another user is able to read his/her public tweets. Twitter also supports other types of interaction, such as user mentions, replies, and retweets. Of these, retweeting is the most important activity as it is used to share and endorse content created by other users. When a user retweets a tweet, the information about the original author as well as the tweet's content are preserved, and the tweet is shared with the user's followers. Typically, users retweet content that they agree with and thus endorse the views expressed by the original tweeter.", + "We apply two different methodologies to analyze the cohesion and coalitions. The first one is based on Krippendorff's INLINEFORM0 BIBREF4 which measures the agreement among observers, or voters in our case. The second one is based on Exponential Random Graph Models (ERGM) BIBREF5 . In contrast to the former, ERGM is a network-based approach and is often used in social-network analyses. Even though these two methodologies come with two different sets of techniques and are based on different assumptions, they provide consistent results.", + "The main contributions of this paper are as follows:", + "(i) We give general insights into the cohesion of political groups in the Eighth European Parliament, both overall and across different policy areas.", + "(ii) We explore whether coalitions are formed in the same way for different policy areas.", + "(iii) We explore to what degree the retweeting behavior of MEPs corresponds to their co-voting patterns.", + "(iv) We employ two statistically sound methodologies and examine the extent to which the results are sensitive to the choice of methodology. While the results are mostly consistent, we show that the difference are due to the different treatment of non-attending and abstaining MEPs by INLINEFORM0 and ERGM.", + "The most novel and interesting aspect of our work is the relationship between the co-voting and the retweeting patterns. The increased use of Twitter by MEPs on days with a roll-call vote session (see Fig FIGREF1 ) is an indicator that these two processes are related. In addition, the force-based layouts of the co-voting network and the retweet network reveal a very similar structure on the left-to-center side of the political spectrum (see Fig FIGREF2 ). They also show a discrepancy on the far-right side of the spectrum, which calls for a more detailed analysis." + ], + [ + "In this paper we study and relate two very different aspects of how MEPs behave in policy-making processes. First, we look at their co-voting behavior, and second, we examine their retweeting patterns. Thus, we draw related work from two different fields of science. On one hand, we look at how co-voting behavior is analyzed in the political-science literature and, on the other, we explore how Twitter is used to better understand political and policy-making processes. The latter has been more thoroughly explored in the field of data mining (specifically, text mining and network analysis).", + "To the best of our knowledge, this is the first paper that studies legislative behavior in the Eighth European Parliament. The legislative behavior of the previous parliaments was thoroughly studied by Hix, Attina, and others BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 . These studies found that voting behavior is determined to a large extent\u2014and when viewed over time, increasingly so\u2014by affiliation to a political group, as an organizational reflection of the ideological position. The authors found that the cohesion of political groups in the parliament has increased, while nationality has been less and less of a decisive factor BIBREF12 . The literature also reports that a split into political camps on the left and right of the political spectrum has recently replaced the `grand coalition' between the two big blocks of Christian Conservatives (EPP) and Social Democrats (S&D) as the dominant form of finding majorities in the parliament. The authors conclude that coalitions are to a large extent formed along the left-to-right axis BIBREF12 .", + "In this paper we analyze the roll-call vote data published in the minutes of the parliament's plenary sessions. For a given subject, the data contains the vote of each MEP present at the respective sitting. Roll-call vote data from the European Parliament has already been extensively studied by other authors, most notably by Hix et al. BIBREF10 , BIBREF13 , BIBREF11 . To be able to study the cohesion and coalitions, authors like Hix, Attina, and Rice BIBREF6 , BIBREF13 , BIBREF14 defined and employed a variety of agreement measures. The most prominent measure is the Agreement Index proposed by Hix et al. BIBREF13 . This measure computes the agreement score from the size of the majority class for a particular vote. The Agreement Index, however, exhibits two drawbacks: (i) it does not account for co-voting by chance, and (ii) without a proper adaptation, it does not accommodate the scenario in which the agreement is to be measured between two different political groups.", + "We employ two statistically sound methodologies developed in two different fields of science. The first one is based on Krippendorff's INLINEFORM0 BIBREF4 . INLINEFORM1 is a measure of the agreement among observers, coders, or measuring instruments that assign values to items or phenomena. It compares the observed agreement to the agreement expected by chance. INLINEFORM2 is used to measure the inter- and self-annotator agreement of human experts when labeling data, and the performance of classification models in machine learning scenarios BIBREF15 . In addition to INLINEFORM3 , we employ Exponential Random Graph Models (ERGM) BIBREF5 . In contrast to the former, ERGM is a network-based approach, often used in social-network analyses. ERGM can be employed to investigate how different network statistics (e.g., number of edges and triangles) or external factors (e.g., political group membership) govern the network-formation process.", + "The second important aspect of our study is related to analyzing the behavior of participants in social networks, specifically Twitter. Twitter is studied by researchers to better understand different political processes, and in some cases to predict their outcomes. Eom et al. BIBREF1 consider the number of tweets by a party as a proxy for the collective attention to the party, explore the dynamics of the volume, and show that this quantity contains information about an election's outcome. Other studies BIBREF16 reach similar conclusions. Conover et al. BIBREF17 predicted the political alignment of Twitter users in the run-up to the 2010 US elections based on content and network structure. They analyzed the polarization of the retweet and mention networks for the same elections BIBREF18 . Borondo et al. BIBREF19 analyzed user activity during the Spanish presidential elections. They additionally analyzed the 2012 Catalan elections, focusing on the interplay between the language and the community structure of the network BIBREF20 . Most existing research, as Larsson points out BIBREF21 , focuses on the online behavior of leading political figures during election campaigns.", + "This paper continues our research on communities that MEPs (and their followers) form on Twitter BIBREF22 . The goal of our research was to evaluate the role of Twitter in identifying communities of influence when the actual communities are known. We represent the influence on Twitter by the number of retweets that MEPs \u201creceive\u201d. We construct two networks of influence: (i) core, which consists only of MEPs, and (ii) extended, which also involves their followers. We compare the detected communities in both networks to the groups formed by the political, country, and language membership of MEPs. The results show that the detected communities in the core network closely match the political groups, while the communities in the extended network correspond to the countries of residence. This provides empirical evidence that analyzing retweet networks can reveal real-world relationships and can be used to uncover hidden properties of the networks.", + "Lazer BIBREF23 highlights the importance of network-based approaches in political science in general by arguing that politics is a relational phenomenon at its core. Some researchers have adopted the network-based approach to investigate the structure of legislative work in the US Congress, including committee and sub-committee membership BIBREF24 , bill co-sponsoring BIBREF25 , and roll-call votes BIBREF26 . More recently, Dal Maso et al. BIBREF27 examined the community structure with respect to political coalitions and government structure in the Italian Parliament. Scherpereel et al. BIBREF28 examined the constituency, personal, and strategic characteristics of MEPs that influence their tweeting behavior. They suggested that Twitter's characteristics, like immediacy, interactivity, spontaneity, personality, and informality, are likely to resonate with political parties across Europe. By fitting regression models, the authors find that MEPs from incohesive groups have a greater tendency to retweet.", + "In contrast to most of these studies, we focus on the Eighth European Parliament, and more importantly, we study and relate two entirely different behavioral aspects, co-voting and retweeting. The goal of this research is to better understand the cohesion and coalition formation processes in the European Parliament by quantifying and comparing the co-voting patterns and social behavior." + ], + [ + "In this section we present the methods to quantify cohesion and coalitions from the roll-call votes and Twitter activities." + ], + [ + "We first show how the co-voting behaviour of MEPs can be quantified by a measure of the agreement between them. We treat individual RCVs as observations, and MEPs as independent observers or raters. When they cast the same vote, there is a high level of agreement, and when they vote differently, there is a high level of disagreement. We define cohesion as the level of agreement within a political group, a coalition as a voting agreement between political groups, and opposition as a disagreement between different groups.", + "There are many well-known measures of agreement in the literature. We selected Krippendorff's Alpha-reliability ( INLINEFORM0 ) BIBREF4 , which is a generalization of several specialized measures. It works for any number of observers, and is applicable to different variable types and metrics (e.g., nominal, ordered, interval, etc.). In general, INLINEFORM1 is defined as follows: INLINEFORM2 ", + "where INLINEFORM0 is the actual disagreement between observers (MEPs), and INLINEFORM1 is disagreement expected by chance. When observers agree perfectly, INLINEFORM2 INLINEFORM3 , when the agreement equals the agreement by chance, INLINEFORM4 INLINEFORM5 , and when the observers disagree systematically, INLINEFORM6 INLINEFORM7 .", + "The two disagreement measures are defined as follows: INLINEFORM0 ", + " INLINEFORM0 ", + "The arguments INLINEFORM0 , and INLINEFORM1 are defined below and refer to the values in the coincidence matrix that is constructed from the RCVs data. In roll-call votes, INLINEFORM2 (and INLINEFORM3 ) is a nominal variable with two possible values: yes and no. INLINEFORM4 is a difference function between the values of INLINEFORM5 and INLINEFORM6 , defined as: INLINEFORM7 ", + "The RCVs data has the form of a reliability data matrix: INLINEFORM0 ", + "where INLINEFORM0 is the number of RCVs, INLINEFORM1 is the number of MEPs, INLINEFORM2 is the number of votes cast in the voting INLINEFORM3 , and INLINEFORM4 is the actual vote of an MEP INLINEFORM5 in voting INLINEFORM6 (yes or no).", + "A coincidence matrix is constructed from the reliability data matrix, and is in general a INLINEFORM0 -by- INLINEFORM1 square matrix, where INLINEFORM2 is the number of possible values of INLINEFORM3 . In our case, where only yes/no votes are relevant, the coincidence matrix is a 2-by-2 matrix of the following form: INLINEFORM4 ", + "A cell INLINEFORM0 accounts for all coincidences from all pairs of MEPs in all RCVs where one MEP has voted INLINEFORM1 and the other INLINEFORM2 . INLINEFORM3 and INLINEFORM4 are the totals for each vote outcome, and INLINEFORM5 is the grand total. The coincidences INLINEFORM6 are computed as: INLINEFORM7 ", + "where INLINEFORM0 is the number of INLINEFORM1 pairs in vote INLINEFORM2 , and INLINEFORM3 is the number of MEPs that voted in INLINEFORM4 . When computing INLINEFORM5 , each pair of votes is considered twice, once as a INLINEFORM6 pair, and once as a INLINEFORM7 pair. The coincidence matrix is therefore symmetrical around the diagonal, and the diagonal contains all the equal votes.", + "The INLINEFORM0 agreement is used to measure the agreement between two MEPs or within a group of MEPs. When applied to a political group, INLINEFORM1 corresponds to the cohesion of the group. The closer INLINEFORM2 is to 1, the higher the agreement of the MEPs in the group, and hence the higher the cohesion of the group.", + "We propose a modified version of INLINEFORM0 to measure the agreement between two different groups, INLINEFORM1 and INLINEFORM2 . In the case of a voting agreement between political groups, high INLINEFORM3 is interpreted as a coalition between the groups, whereas negative INLINEFORM4 indicates political opposition.", + "Suppose INLINEFORM0 and INLINEFORM1 are disjoint subsets of all the MEPs, INLINEFORM2 , INLINEFORM3 . The respective number of votes cast by both group members in vote INLINEFORM4 is INLINEFORM5 and INLINEFORM6 . The coincidences are then computed as: INLINEFORM7 ", + "where the INLINEFORM0 pairs come from different groups, INLINEFORM1 and INLINEFORM2 . The total number of such pairs in vote INLINEFORM3 is INLINEFORM4 . The actual number INLINEFORM5 of the pairs is multiplied by INLINEFORM6 so that the total contribution of vote INLINEFORM7 to the coincidence matrix is INLINEFORM8 ." + ], + [ + "In this section we describe a network-based approach to analyzing the co-voting behavior of MEPs. For each roll-call vote we form a network, where the nodes in the network are MEPs, and an undirected edge between two MEPs is formed when they cast the same vote.", + "We are interested in the factors that determine the cohesion within political groups and coalition formation between political groups. Furthermore, we investigate to what extent communication in a different social context, i.e., the retweeting behavior of MEPs, can explain the co-voting of MEPs. For this purpose we apply an Exponential Random Graph Model BIBREF5 to individual roll-call vote networks, and aggregate the results by means of the meta-analysis.", + "ERGMs allow us to investigate the factors relevant for the network-formation process. Network metrics, as described in the abundant literature, serve to gain information about the structural properties of the observed network. A model investigating the processes driving the network formation, however, has to take into account that there can be a multitude of alternative networks. If we are interested in the parameters influencing the network formation we have to consider all possible networks and measure their similarity to the originally observed network. The family of ERGMs builds upon this idea.", + "Assume a random graph INLINEFORM0 , in the form of a binary adjacency matrix, made up of a set of INLINEFORM1 nodes and INLINEFORM2 edges INLINEFORM3 where, similar to a binary choice model, INLINEFORM4 if the nodes INLINEFORM5 are connected and INLINEFORM6 if not. Since network data is by definition relational and thus violates assumptions of independence, classical binary choice models, like logistic regression, cannot be applied in this context. Within an ERGM, the probability for a given network is modelled by DISPLAYFORM0 ", + " where INLINEFORM0 is the vector of parameters and INLINEFORM1 is the vector of network statistics (counts of network substructures), which are a function of the adjacency matrix INLINEFORM2 . INLINEFORM3 is a normalization constant corresponding to the sample of all possible networks, which ensures a proper probability distribution. Evaluating the above expression allows us to make assertions if and how specific nodal attributes influence the network formation process. These nodal attributes can be endogenous (dyad-dependent parameters) to the network, like the in- and out-degrees of a node, or exogenous (dyad-independent parameters), as the party affiliation, or the country of origin in our case.", + "An alternative formulation of the ERGM provides the interpretation of the coefficients. We introduce the change statistic, which is defined as the change in the network statistics when an edge between nodes INLINEFORM0 and INLINEFORM1 is added or not. If INLINEFORM2 and INLINEFORM3 denote the vectors of counts of network substructures when the edge is added or not, the change statistics is defined as follows: INLINEFORM4 ", + "With this at hand it can be shown that the distribution of the variable INLINEFORM0 , conditional on the rest of the graph INLINEFORM1 , corresponds to: INLINEFORM2 ", + "This implies on the one hand that the probability depends on INLINEFORM0 via the change statistic INLINEFORM1 , and on the other hand, that each coefficient within the vector INLINEFORM2 represents an increase in the conditional log-odds ( INLINEFORM3 ) of the graph when the corresponding element in the vector INLINEFORM4 increases by one. The need to condition the probability on the rest of the network can be illustrated by a simple example. The addition (removal) of a single edge alters the network statistics. If a network has only edges INLINEFORM5 and INLINEFORM6 , the creation of an edge INLINEFORM7 would not only add an additional edge but would also alter the count for other network substructures included in the model. In this example, the creation of the edge INLINEFORM8 also increases the number of triangles by one. The coefficients are transformed into probabilities with the logistic function: INLINEFORM9 ", + "For example, in the context of roll-call votes, the probability that an additional co-voting edge is formed between two nodes (MEPs) of the same political group is computed with that equation. In this context, the nodematch (nodemix) coefficients of the ERGM (described in detail bellow) therefore refer to the degree of homophilous (heterophilous) matching of MEPs with regard to their political affiliation, or, expressed differently, the propensity of MEPs to co-vote with other MEPs of their respective political group or another group.", + "A positive coefficient reflects an increased chance that an edge between two nodes with respective properties, like group affiliation, given all other parameters unchanged, is formed. Or, put differently, a positive coefficient implies that the probability of observing a network with a higher number of corresponding pairs relative to the hypothetical baseline network, is higher than to observe the baseline network itself BIBREF31 . For an intuitive interpretation, log-odds value of 0 corresponds to the even chance probability of INLINEFORM0 . Log-odds of INLINEFORM1 correspond to an increase of probability by INLINEFORM2 , whereas log-odds of INLINEFORM3 correspond to a decrease of probability by INLINEFORM4 .", + "The computational challenges of estimating ERGMs is to a large degree due to the estimation of the normalizing constant. The number of possible networks is already extremely large for very small networks and the computation is simply not feasible. Therefore, an appropriate sample has to be found, ideally covering the most probable areas of the probability distribution. For this we make use of a method from the Markov Chain Monte Carlo (MCMC) family, namely the Metropolis-Hastings algorithm.", + "The idea behind this algorithm is to generate and sample highly weighted random networks departing from the observed network. The Metropolis-Hastings algorithm is an iterative algorithm which samples from the space of possible networks by randomly adding or removing edges from the starting network conditional on its density. If the likelihood, in the ERGM context also denoted as weights, of the newly generated network is higher than that of the departure network it is retained, otherwise it is discarded. In the former case, the algorithm starts anew from the newly generated network. Otherwise departure network is used again. Repeating this procedure sufficiently often and summing the weights associated to the stored (sampled) networks allows to compute an approximation of the denominator in equation EQREF18 (normalizing constant).", + "The algorithm starts sampling from the originally observed network INLINEFORM0 . The optimization of the coefficients is done simultaneously, equivalently with the Metropolis-Hastings algorithm. At the beginning starting values have to be supplied. For the study at hand we used the \u201cergm\u201d library from the statistical R software package BIBREF5 implementing the Gibbs-Sampling algorithm BIBREF32 which is a special case of the Metropolis-Hastings algorithm outlined.", + "In order to answer our question of the importance of the factors which drive the network formation process in the roll-call co-voting network, the ERGM is specified with the following parameters:", + "nodematch country: This parameter adds one network statistic to the model, i.e., the number of edges INLINEFORM0 where INLINEFORM1 . The coefficient indicates the homophilious mixing behavior of MEPs with respect to their country of origin. In other words, this coefficient indicates how relevant nationality is in the formation of edges in the co-voting network.", + "nodematch national party: This parameter adds one network statistic to the model: the number of edges INLINEFORM0 with INLINEFORM1 . The coefficient indicates the homophilious mixing behavior of the MEPs with regard to their party affiliation at the national level. In the context of this study, this coefficient can be interpreted as an indicator for within-party cohesion at the national level.", + "nodemix EP group: This parameter adds one network statistic for each pair of European political groups. These coefficients shed light on the degree of coalitions between different groups as well as the within group cohesion . Given that there are nine groups in the European Parliament, this coefficient adds in total 81 statistics to the model.", + "edge covariate Twitter: This parameter corresponds to a square matrix with the dimension of the adjacency matrix of the network, which corresponds to the number of mutual retweets between the MEPs. It provides an insight about the extent to which communication in one social context (Twitter), can explain cooperation in another social context (co-voting in RCVs).", + "An ERGM as specified above is estimated for each of the 2535 roll-call votes. Each roll-call vote is thereby interpreted as a binary network and as an independent study. It is assumed that a priori each MEP could possibly form an edge with each other MEP in the context of a roll-call vote. Assumptions over the presence or absence of individual MEPs in a voting session are not made. In other words the dimensions of the adjacency matrix (the node set), and therefore the distribution from which new networks are drawn, is kept constant over all RCVs and therefore for every ERGM. The ERGM results therefore implicitly generalize to the case where potentially all MEPs are present and could be voting. Not voting is incorporated implicitly by the disconnectedness of a node.", + "The coefficients of the 2535 roll-call vote studies are aggregated by means of a meta-analysis approach proposed by Lubbers BIBREF33 and Snijders et al. BIBREF34 . We are interested in average effect sizes of different matching patterns over different topics and overall. Considering the number of RCVs, it seems straightforward to interpret the different RCV networks as multiplex networks and collapse them into one weighted network, which could then be analysed by means of a valued ERGM BIBREF35 . There are, however, two reasons why we chose the meta-analysis approach instead. First, aggregating the RCV data results into an extremely dense network, leading to severe convergence (degeneracy) problems for the ERGM. Second, the RCV data contains information about the different policy areas the individual votes were about. Since we are interested in how the coalition formation in the European Parliament differs over different areas, a method is needed that allows for an ex-post analysis of the corresponding results. We therefore opted for the meta-analysis approach by Lubbers and Snijders et al. This approach allows us to summarize the results by decomposing the coefficients into average effects and (class) subject-specific deviations. The different ERGM runs for each RCV are thereby regarded as different studies with identical samples that are combined to obtain a general overview of effect sizes. The meta-regression model is defined as: INLINEFORM0 ", + "Here INLINEFORM0 is a parameter estimate for class INLINEFORM1 , and INLINEFORM2 is the average coefficient. INLINEFORM3 denotes the normally distributed deviation of the class INLINEFORM4 with a mean of 0 and a variance of INLINEFORM5 . INLINEFORM6 is the estimation error of the parameter value INLINEFORM7 from the ERGM. The meta-analysis model is fitted by an iterated, weighted, least-squares model in which the observations are weighted by the inverse of their variances. For the overall nodematch between political groups, we weighted the coefficients by group sizes. The results from the meta analysis can be interpreted as if they stemmed from an individual ERGM run.", + "In our study, the meta-analysis was performed using the RSiena library BIBREF36 , which implements the method proposed by Lubbers and Snijders et al. BIBREF33 , BIBREF34 ." + ], + [ + "The retweeting behavior of MEPs is captured by their retweet network. Each MEP active on Twitter is a node in this network. An edge in the network between two MEPs exists when one MEP retweeted the other. The weight of the edge is the number of retweets between the two MEPs. The resulting retweet network is an undirected, weighted network.", + "We measure the cohesion of a political group INLINEFORM0 as the average retweets, i.e., the ratio of the number of retweets between the MEPs in the group INLINEFORM1 to the number of MEPs in the group INLINEFORM2 . The higher the ratio, the more each MEP (on average) retweets the MEPs from the same political group, hence, the higher the cohesion of the political group. The definition of the average retweeets ( INLINEFORM3 ) of a group INLINEFORM4 is: INLINEFORM5 ", + "This measure of cohesion captures the aggregate retweeting behavior of the group. If we consider retweets as endorsements, a larger number of retweets within the group is an indicator of agreement between the MEPs in the group. It does not take into account the patterns of retweeting within the group, thus ignoring the social sub-structure of the group. This is a potentially interesting direction and we leave it for future work.", + "We employ an analogous measure for the strength of coalitions in the retweet network. The coalition strength between two groups INLINEFORM0 and INLINEFORM1 is the ratio of the number of retweets from one group to the other (but not within groups) INLINEFORM2 to the total number of MEPs in both groups, INLINEFORM3 . The definition of the average retweeets ( INLINEFORM4 ) between groups INLINEFORM5 and INLINEFORM6 is: INLINEFORM7 " + ], + [ + "In this section we first report on the level of cohesion of the European Parliament's groups by analyzing the co-voting through the agreement and ERGM measures. Next, we explore two important policy areas, namely Economic and monetary system and State and evolution of the Union. Finally, we analyze the cohesion of the European Parliament's groups on Twitter.", + "Existing research by Hix et al. BIBREF10 , BIBREF13 , BIBREF11 shows that the cohesion of the European political groups has been rising since the 1990s, and the level of cohesion remained high even after the EU's enlargement in 2004, when the number of MEPs increased from 626 to 732.", + "We measure the co-voting cohesion of the political groups in the Eighth European Parliament using Krippendorff's Alpha\u2014the results are shown in Fig FIGREF30 (panel Overall). The Greens-EFA have the highest cohesion of all the groups. This finding is in line with an analysis of previous compositions of the Fifth and Sixth European Parliaments by Hix and Noury BIBREF11 , and the Seventh by VoteWatch BIBREF37 . They are closely followed by the S&D and EPP. Hix and Noury reported on the high cohesion of S&D in the Fifth and Sixth European Parliaments, and we also observe this in the current composition. They also reported a slightly less cohesive EPP-ED. This group split in 2009 into EPP and ECR. VoteWatch reports EPP to have cohesion on a par with Greens-EFA and S&D in the Seventh European Parliament. The cohesion level we observe in the current European Parliament is also similar to the level of Greens-EFA and S&D.", + "The catch-all group of the non-aligned (NI) comes out as the group with the lowest cohesion. In addition, among the least cohesive groups in the European Parliament are the Eurosceptics EFDD, which include the British UKIP led by Nigel Farage, and the ENL whose largest party are the French National Front, led by Marine Le Pen. Similarly, Hix and Noury found that the least cohesive groups in the Seventh European Parliament are the nationalists and Eurosceptics. The Eurosceptic IND/DEM, which participated in the Sixth European Parliament, transformed into the current EFDD, while the nationalistic UEN was dissolved in 2009.", + "We also measure the voting cohesion of the European Parliament groups using an ERGM, a network-based method\u2014the results are shown in Fig FIGREF31 (panel Overall). The cohesion results obtained with ERGM are comparable to the results based on agreement. In this context, the parameters estimated by the ERGM refer to the matching of MEPs who belong to the same political group (one parameter per group). The parameters measure the homophilous matching between MEPs who have the same political affiliation. A positive value for the estimated parameter indicates that the co-voting of MEPs from that group is greater than what is expected by chance, where the expected number of co-voting links by chance in a group is taken to be uniformly random. A negative value indicates that there are fewer co-voting links within a group than expected by chance.", + "Even though INLINEFORM0 and ERGM compute scores relative to what is expected by chance, they refer to different interpretations of chance. INLINEFORM1 's concept of chance is based on the number of expected pair-wise co-votes between MEPs belonging to a group, knowing the votes of these MEPs on all RCVs. ERGM's concept of chance is based on the number of expected pair-wise co-votes between MEPs belonging to a group on a given RCV, knowing the network-related properties of the co-voting network on that particular RCV. The main difference between INLINEFORM2 and ERGM, though, is the treatment of non-voting and abstained MEPs. INLINEFORM3 considers only the yes/no votes, and consequently, agreements by the voting MEPs of the same groups are considerably higher than co-voting by chance. ERGM, on the other hand, always considers all MEPs, and non-voting and abstained MEPs are treated as disconnected nodes. The level of co-voting by chance is therefore considerably lower, since there is often a large fraction of MEPs that do not attand or abstain.", + "As with INLINEFORM0 , Greens-EFA, S&D, and EPP exhibit the highest cohesion, even though their ranking is permuted when compared to the ranking obtained with INLINEFORM1 . At the other end of the scale, we observe the same situation as with INLINEFORM2 . The non-aligned members NI have the lowest cohesion, followed by EFDD and ENL.", + "The only place where the two methods disagree is the level of cohesion of GUE-NGL. The Alpha attributes GUE-NGL a rather high level of cohesion, on a par with ALDE, whereas the ERGM attributes them a much lower cohesion. The reason for this difference is the relatively high abstention rate of GUE-NGL. Whereas the overall fraction of non-attending and abstaining MEPs across all RCVs and all political groups is 25%, the GUE-NGL abstention rate is 34%. This is reflected in an above average cohesion by INLINEFORM0 where only yes/no votes are considered, and in a relatively lower, below average cohesion by ERGM. In the later case, the non-attendance is interpreted as a non-cohesive voting of a political groups as a whole.", + "In addition to the overall cohesion, we also focus on two selected policy areas. The cohesion of the political groups related to these two policy areas is shown in the first two panels in Fig FIGREF30 ( INLINEFORM0 ) and Fig FIGREF31 (ERGM).", + "The most important observation is that the level of cohesion of the political groups is very stable across different policy areas. These results are corroborated by both methodologies. Similar to the overall cohesion, the most cohesive political groups are the S&D, Greens-EFA, and EPP. The least cohesive group is the NI, followed by the ENL and EFDD. The two methodologies agree on the level of cohesion for all the political groups, except for GUE-NGL, due to a lower attendance rate.", + "We determine the cohesion of political groups on Twitter by using the average number of retweets between MEPs within the same group. The results are shown in Fig FIGREF33 . The right-wing ENL and EFDD come out as the most cohesive groups, while all the other groups have a far lower average number of retweets. MEPs from ENL and EFDD post by far the largest number of retweets (over 240), and at the same time over 94% of their retweets are directed to MEPs from the same group. Moreover, these two groups stand out in the way the retweets are distributed within the group. A large portion of the retweets of EFDD (1755) go to Nigel Farage, the leader of the group. Likewise, a very large portion of retweets of ENL (2324) go to Marine Le Pen, the leader of the group. Farage and Le Pen are by far the two most retweeted MEPs, with the third one having only 666 retweets." + ], + [ + "Coalition formation in the European Parliament is largely determined by ideological positions, reflected in the degree of cooperation of parties at the national and European levels. The observation of ideological inclinations in the coalition formation within the European Parliament was already made by other authors BIBREF11 and is confirmed in this study. The basic patterns of coalition formation in the European Parliament can already be seen in the co-voting network in Fig FIGREF2 A. It is remarkable that the degree of attachment between the political groups, which indicates the degree of cooperation in the European Parliament, nearly exactly corresponds to the left-to-right seating order.", + "The liberal ALDE seems to have an intermediator role between the left and right parts of the spectrum in the parliament. Between the extreme (GUE-NGL) and center left (S&D) groups, this function seems to be occupied by Greens-EFA. The non-aligned members NI, as well as the Eurosceptic EFFD and ENL, seem to alternately tip the balance on both poles of the political spectrum. Being ideologically more inclined to vote with other conservative and right-wing groups (EPP, ECR), they sometimes also cooperate with the extreme left-wing group (GUE-NGL) with which they share their Euroscepticism as a common denominator.", + "Figs FIGREF36 and FIGREF37 give a more detailed understanding of the coalition formation in the European Parliament. Fig FIGREF36 displays the degree of agreement or cooperation between political groups measured by Krippendorff's INLINEFORM0 , whereas Fig FIGREF37 is based on the result from the ERGM. We first focus on the overall results displayed in the right-hand plots of Figs FIGREF36 and FIGREF37 .", + "The strongest degrees of cooperation are observed, with both methods, between the two major parties (EPP and S&D) on the one hand, and the liberal ALDE on the other. Furthermore, we see a strong propensity for Greens-EFA to vote with the Social Democrats (5th strongest coalition by INLINEFORM0 , and 3rd by ERGM) and the GUE-NGL (3rd strongest coalition by INLINEFORM1 , and 5th by ERGM). These results underline the role of ALDE and Greens-EFA as intermediaries for the larger groups to achieve a majority. Although the two largest groups together have 405 seats and thus significantly more than the 376 votes needed for a simple majority, the degree of cooperation between the two major groups is ranked only as the fourth strongest by both methods. This suggests that these two political groups find it easier to negotiate deals with smaller counterparts than with the other large group. This observation was also made by Hix et al. BIBREF12 , who noted that alignments on the left and right of the political spectrum have in recent years replaced the \u201cGrand Coalition\u201d between the two large blocks of Christian Conservatives (EPP) and Social Democrats (S&D) as the dominant form of finding majorities in the parliament.", + "Next, we focus on the coalition formation within the two selected policy areas. The area State and Evolution of the Union is dominated by cooperation between the two major groups, S&D and EPP, as well as ALDE. We also observe a high degree of cooperation between groups that are generally regarded as integration friendly, like Greens-EFA and GUE-NGL. We see, particularly in Fig FIGREF36 , a relatively high degree of cooperation between groups considered as Eurosceptic, like ECR, EFFD, ENL, and the group of non-aligned members.", + "The dichotomy between supporters and opponents of European integration is even more pronounced within the policy area Economic and Monetary System. In fact, we are taking a closer look specifically at these two areas as they are, at the same time, both contentious and important. Both methods rank the cooperation between S&D and EPP on the one hand, and ALDE on the other, as the strongest.", + "We also observe a certain degree of unanimity among the Eurosceptic and right-wing groups (EFDD, ENL, and NI) in this policy area. This seems plausible, as these groups were (especially in the aftermath of the global financial crisis and the subsequent European debt crisis) in fierce opposition to further payments to financially troubled member states. However, we also observe a number of strong coalitions that might, at first glance, seem unusual, specifically involving the left-wing group GUE-NGL on the one hand, and the right-wing EFDD, ENL, and NI on the other. These links also show up in the network plot in Fig FIGREF2 A. This might be attributable to a certain degree of Euroscepticism on both sides: rooted in criticism of capitalism on the left, and at least partly a raison d'\u00eatre on the right. Hix et al. BIBREF11 discovered this pattern as well, and proposed an additional explanation\u2014these coalitions also relate to a form of government-opposition dynamic that is rooted at the national level, but is reflected in voting patterns at the European level.", + "In general, we observe two main differences between the INLINEFORM0 and ERGM results: the baseline cooperation as estimated by INLINEFORM1 is higher, and the ordering of coalitions from the strongest to the weakest is not exactly the same. The reason is the same as for the cohesion, namely different treatment of non-voting and abstaining MEPs. When they are ignored, as by INLINEFORM2 , the baseline level of inter-group co-voting is higher. When non-attending and abstaining is treated as voting differently, as by ERGM, it is considerably more difficult to achieve co-voting coalitions, specially when there are on average 25% MEPs that do not attend or abstain. Groups with higher non-attendance rates, such as GUE-NGL (34%) and NI (40%) are less likely to form coalitions, and therefore have relatively lower ERGM coefficients (Fig FIGREF37 ) than INLINEFORM3 scores (Fig FIGREF36 ).", + "The first insight into coalition formation on Twitter can be observed in the retweet network in Fig FIGREF2 B. The ideological left to right alignment of the political groups is reflected in the retweet network. Fig FIGREF40 shows the strength of the coalitions on Twitter, as estimated by the number of retweets between MEPs from different groups. The strongest coalitions are formed between the right-wing groups EFDD and ECR, as well as ENL and NI. At first, this might come as a surprise, since these groups do not form strong coalitions in the European Parliament, as can be seen in Figs FIGREF36 and FIGREF37 . On the other hand, the MEPs from these groups are very active Twitter users. As previously stated, MEPs from ENL and EFDD post the largest number of retweets. Moreover, 63% of the retweets outside of ENL are retweets of NI. This effect is even more pronounced with MEPs from EFDD, whose retweets of ECR account for 74% of their retweets from other groups.", + "In addition to these strong coalitions on the right wing, we find coalition patterns to be very similar to the voting coalitions observed in the European Parliament, seen in Figs FIGREF36 and FIGREF37 . The strongest coalitions, which come immediately after the right-wing coalitions, are between Greens-EFA on the one hand, and GUE-NGL and S&D on the other, as well as ALDE on the one hand, and EPP and S&D on the other. These results corroborate the role of ALDE and Greens-EFA as intermediaries in the European Parliament, not only in the legislative process, but also in the debate on social media.", + "To better understand the formation of coalitions in the European Parliament and on Twitter, we examine the strongest cooperation between political groups at three different thresholds. For co-voting coalitions in the European Parliament we choose a high threshold of INLINEFORM0 , a medium threshold of INLINEFORM1 , and a negative threshold of INLINEFORM2 (which corresponds to strong oppositions). In this way we observe the overall patterns of coalition and opposition formation in the European Parliament and in the two specific policy areas. For cooperation on Twitter, we choose a high threshold of INLINEFORM3 , a medium threshold of INLINEFORM4 , and a very low threshold of INLINEFORM5 .", + "The strongest cooperations in the European Parliament over all policy areas are shown in Fig FIGREF42 G. It comes as no surprise that the strongest cooperations are within the groups (in the diagonal). Moreover, we again observe GUE-NGL, S&D, Greens-EFA, ALDE, and EPP as the most cohesive groups. In Fig FIGREF42 H, we observe coalitions forming along the diagonal, which represents the seating order in the European Parliament. Within this pattern, we observe four blocks of coalitions: on the left, between GUE-NGL, S&D, and Greens-EFA; in the center, between S&D, Greens-EFA, ALDE, and EPP; on the right-center between ALDE, EPP, and ECR; and finally, on the far-right between ECR, EFDD, ENL, and NI. Fig FIGREF42 I shows the strongest opposition between groups that systematically disagree in voting. The strongest disagreements are between left- and right-aligned groups, but not between the left-most and right-most groups, in particular, between GUE-NGL and ECR, but also between S&D and Greens-EFA on one side, and ENL and NI on the other.", + "In the area of Economic and monetary system we see a strong cooperation between EPP and S&D (Fig FIGREF42 A), which is on a par with the cohesion of the most cohesive groups (GUE-NGL, S&D, Greens-EFA, ALDE, and EPP), and is above the cohesion of the other groups. As pointed out in the section \u201csec:coalitionpolicy\u201d, there is a strong separation in two blocks between supporters and opponents of European integration, which is even more clearly observed in Fig FIGREF42 B. On one hand, we observe cooperation between S&D, ALDE, EPP, and ECR, and on the other, cooperation between GUE-NGL, Greens-EFA, EFDD, ENL, and NI. This division in blocks is seen again in Fig FIGREF42 C, which shows the strongest disagreements. Here, we observe two blocks composed of S&D, EPP, and ALDE on one hand, and GUE-NGL, EFDD, ENL, and NI on the other, which are in strong opposition to each other.", + "In the area of State and Evolution of the Union we again observe a strong division in two blocks (see Fig FIGREF42 E). This is different to the Economic and monetary system, however, where we observe a far-left and far-right cooperation, where the division is along the traditional left-right axis.", + "The patterns of coalitions forming on Twitter closely resemble those in the European Parliament. In Fig FIGREF42 J we see that the strongest degrees of cooperation on Twitter are within the groups. The only group with low cohesion is the NI, whose members have only seven retweets between them. The coalitions on Twitter follow the seating order in the European Parliament remarkably well (see Fig FIGREF42 K). What is striking is that the same blocks form on the left, center, and on the center-right, both in the European Parliament and on Twitter. The largest difference between the coalitions in the European Parliament and on Twitter is on the far-right, where we observe ENL and NI as isolated blocks.", + "The results shown in Fig FIGREF44 quantify the extent to which communication in one social context (Twitter) can explain cooperation in another social context (co-voting in the European Parliament). A positive value indicates that the matching behavior in the retweet network is similar to the one in the co-voting network, specific for an individual policy area. On the other hand, a negative value implies a negative \u201ccorrelation\u201d between the retweeting and co-voting of MEPs in the two different contexts.", + "The bars in Fig FIGREF44 correspond to the coefficients from the edge covariate terms of the ERGM, describing the relationship between the retweeting and co-voting behavior of MEPs. The coefficients are aggregated for individual policy areas by means of a meta-analysis.", + "Overall, we observe a positive correlation between retweeting and co-voting, which is significantly different from zero. The strongest positive correlations are in the areas Area of freedom, security and justice, External relations of the Union, and Internal markets. Weaker, but still positive, correlations are observed in the areas Economic, social and territorial cohesion, European citizenship, and State and evolution of the Union. The only exception, with a significantly negative coefficient, is the area Economic and monetary system. This implies that in the area Economic and monetary system we observe a significant deviation from the usual co-voting patterns. Results from section \u201csec:coalitionpolicy\u201d, confirm that this is indeed the case. Especially noteworthy are the coalitions between GUE-NGL and Greens-EFA on the left wing, and EFDD and ENL on the right wing. In the section \u201csec:coalitionpolicy\u201d we interpret these results as a combination of Euroscepticism on both sides, motivated on the left by a skeptical attitude towards the market orientation of the EU, and on the right by a reluctance to give up national sovereignty." + ], + [ + "We study cohesion and coalitions in the Eighth European Parliament by analyzing, on one hand, MEPs' co-voting tendencies and, on the other, their retweeting behavior.", + "We reveal that the most cohesive political group in the European Parliament, when it comes to co-voting, is Greens-EFA, closely followed by S&D and EPP. This is consistent with what VoteWatch BIBREF37 reported for the Seventh European Parliament. The non-aligned (NI) come out as the least cohesive group, followed by the Eurosceptic EFDD. Hix and Noury BIBREF11 also report that nationalists and Eurosceptics form the least cohesive groups in the Sixth European Parliament. We reaffirm most of these results with both of the two employed methodologies. The only point where the two methodologies disagree is in the level of cohesion for the left-wing GUE-NGL, which is portrayed by ERGM as a much less cohesive group, due to their relatively lower attendance rate. The level of cohesion of the political groups is quite stable across different policy areas and similar conclusions apply.", + "On Twitter we can see results that are consistent with the RCV results for the left-to-center political spectrum. The exception, which clearly stands out, is the right-wing groups ENL and EFDD that seem to be the most cohesive ones. This is the direct opposite of what was observed in the RCV data. We speculate that this phenomenon can be attributed to the fact that European right-wing groups, on a European but also on a national level, rely to a large degree on social media to spread their narratives critical of European integration. We observed the same phenomenon recently during the Brexit campaign BIBREF38 . Along our interpretation the Brexit was \u201cwon\u201d to some extent due to these social media activities, which are practically non-existent among the pro-EU political groups. The fact that ENL and EFDD are the least cohesive groups in the European Parliament can be attributed to their political focus. It seems more important for the group to agree on its anti-EU stance and to call for independence and sovereignty, and much less important to agree on other issues put forward in the parliament.", + "The basic pattern of coalition formation, with respect to co-voting, can already be seen in Fig FIGREF2 A: the force-based layout almost completely corresponds to the seating order in the European Parliament (from the left- to the right-wing groups). A more thorough examination shows that the strongest cooperation can be observed, for both methodologies, between EPP, S&D, and ALDE, where EPP and S&D are the two largest groups, while the liberal ALDE plays the role of an intermediary in this context. On the other hand, the role of an intermediary between the far-left GUE-NGL and its center-left neighbor, S&D, is played by the Greens-EFA. These three parties also form a strong coalition in the European Parliament. On the far right of the spectrum, the non-aligned, EFDD, and ENL form another coalition. This behavior was also observed by Hix et al. BIBREF12 , stating that alignments on the left and right have in recent years replaced the \u201cGrand Coalition\u201d between the two large blocks of Christian Conservatives (EPP) and Social Democrats (S&D) as the dominant form of finding majorities in the European Parliament. When looking at the policy area Economic and monetary system, we see the same coalitions. However, interestingly, EFDD, ENL, and NI often co-vote with the far-left GUE-NGL. This can be attributed to a certain degree of Euroscepticism on both sides: as a criticism of capitalism, on one hand, and as the main political agenda, on the other. This pattern was also discovered by Hix et al. BIBREF12 , who argued that these coalitions emerge from a form of government-opposition dynamics, rooted at the national level, but also reflected at the European level.", + "When studying coalitions on Twitter, the strongest coalitions can be observed on the right of the spectrum (between EFDD, ECR, ENL, and NI). This is, yet again, in contrast to what was observed in the RCV data. The reason lies in the anti-EU messages they tend to collectively spread (retweet) across the network. This behavior forms strong retweet ties, not only within, but also between, these groups. For example, MEPs of EFDD mainly retweet MEPs from ECR (with the exception of MEPs from their own group). In contrast to these right-wing coalitions, we find the other coalitions to be consistent with what is observed in the RCV data. The strongest coalitions on the left-to-center part of the axis are those between GUE-NGL, Greens-EFA, and S&D, and between S&D, ALDE, and EPP. These results reaffirm the role of Greens-EFA and ALDE as intermediaries, not only in the European Parliament but also in the debates on social media.", + "Last, but not least, with the ERGM methodology we measure the extent to which the retweet network can explain the co-voting activities in the European Parliament. We compute this for each policy area separately and also over all RCVs. We conclude that the retweet network indeed matches the co-voting behavior, with the exception of one specific policy area. In the area Economic and monetary system, the links in the (overall) retweet network do not match the links in the co-voting network. Moreover, the negative coefficients imply a radically different formation of coalitions in the European Parliament. This is consistent with the results in Figs FIGREF36 and FIGREF37 (the left-hand panels), and is also observed in Fig FIGREF42 (the top charts). From these figures we see that in this particular case, the coalitions are also formed between the right-wing groups and the far-left GUE-NGL. As already explained, we attribute this to the degree of Euroscepticism that these groups share on this particular policy issue." + ], + [ + "In this paper we analyze (co-)voting patterns and social behavior of members of the European Parliament, as well as the interaction between these two systems. More precisely, we analyze a set of 2535 roll-call votes as well as the tweets and retweets of members of the MEPs in the period from October 2014 to February 2016. The results indicate a considerable level of correlation between these two complex systems. This is consistent with previous findings of Cherepnalkoski et al. BIBREF22 , who reconstructed the adherence of MEPs to their respective political or national group solely from their retweeting behavior.", + "We employ two different methodologies to quantify the co-voting patterns: Krippendorff's INLINEFORM0 and ERGM. They were developed in different fields of research, use different techniques, and are based on different assumptions, but in general they yield consistent results. However, there are some differences which have consequences for the interpretation of the results.", + " INLINEFORM0 is a measure of agreement, designed as a generalization of several specialized measures, that can compare different numbers of observations, in our case roll-call votes. It only considers yes/no votes. Absence and abstention by MEPs is ignored. Its baseline ( INLINEFORM1 ), i.e., co-voting by chance, is computed from the yes/no votes of all MEPs on all RCVs.", + "ERGMs are used in social-network analyses to determine factors influencing the edge formation process. In our case an edge between two MEPs is formed when they cast the same yes/no vote within a RCV. It is assumed that a priori each MEP can form a link with any other MEP. No assumptions about the presence or absence of individual MEPs in a voting session are made. Each RCV is analyzed as a separate binary network. The node set is thereby kept constant for each RCV network. While the ERGM departs from the originally observed network, where MEPs who didn't vote or abstained appear as isolated nodes, links between these nodes are possible within the network sampling process which is part of the ERGM optimization process. The results of several RCVs are aggregated by means of the meta-analysis approach. The baseline (ERGM coefficients INLINEFORM0 ), i.e., co-voting by chance, is computed from a large sample of randomly generated networks.", + "These two different baselines have to be taken into account when interpreting the results of INLINEFORM0 and ERGM. In a typical voting session, 25% of the MEPs are missing or abstaining. When assessing cohesion of political groups, all INLINEFORM1 values are well above the baseline, and the average INLINEFORM2 . The average ERGM cohesion coefficients, on the other hand, are around the baseline. The difference is even more pronounced for groups with higher non-attendance/abstention rates like GUE-NGL (34%) and NI (40%). When assessing strength of coalitions between pairs of groups, INLINEFORM3 values are balanced around the baseline, while the ERGM coefficients are mostly negative. The ordering of coalitions from the strongest to the weakest is therefor different when groups with high non-attendance/abstention rates are involved.", + "The choice of the methodology to asses cohesion and coalitions is not obvious. Roll-call voting is used for decisions which demand a simple majority only. One might however argue that non-attendance/abstention corresponds to a no vote, or that absence is used strategically. Also, the importance of individual votes, i.e., how high on the agenda of a political group is the subject, affects their attendance, and consequently the perception of their cohesion and the potential to act as a reliable coalition partner." + ], + [ + "This work was supported in part by the EC projects SIMPOL (no. 610704) and DOLFINS (no. 640772), and by the Slovenian ARRS programme Knowledge Technologies (no. P2-103)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1203/instruction.md b/qasper-1203/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c6f43351e86186d2244a62cf9674ff4d7e18c8b6 --- /dev/null +++ b/qasper-1203/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Harry Potter and the Action Prediction Challenge from Natural Language + +Question: Why do they think this task is hard? What is the baseline performance? \ No newline at end of file diff --git a/qasper-1233/instruction.md b/qasper-1233/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..8354437b60bd77f705b69ce261cfaa39ef308350 --- /dev/null +++ b/qasper-1233/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Generative Dialog Policy for Task-oriented Dialog Systems + +Question: How much is proposed model better than baselines in performed experiments? \ No newline at end of file diff --git a/qasper-1234/instruction.md b/qasper-1234/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..d21fff972c8ef4adeaa00214890ac269eb06fa4c --- /dev/null +++ b/qasper-1234/instruction.md @@ -0,0 +1,132 @@ +Name of Paper: Generative Dialog Policy for Task-oriented Dialog Systems + +Question: What are state-of-the-art baselines? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Technical Background ::: Encoder-Decoder Seq2Seq Models", + "Technical Background ::: Attention Mechanism", + "Generative Dialogue Policy", + "Generative Dialogue Policy ::: Notations and Task Formulation", + "Generative Dialogue Policy ::: Utterance Encoder", + "Generative Dialogue Policy ::: Dialogue State Tracker", + "Generative Dialogue Policy ::: Knowledge Base", + "Generative Dialogue Policy ::: Dialogue Policy Maker", + "Generative Dialogue Policy ::: Nature Language Generator", + "Generative Dialogue Policy ::: Training", + "Experiments", + "Experiments ::: Datasets", + "Experiments ::: Baselines", + "Experiments ::: Parameters settings", + "Experiments ::: Experimental Results", + "Experiments ::: Case Study", + "Conclusion" + ], + "paragraphs": [ + [ + "Task-oriented dialogue system is an important tool to build personal virtual assistants, which can help users to complete most of the daily tasks by interacting with devices via natural language. It's attracting increasing attention of researchers, and lots of works have been proposed in this area BIBREF0, BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7.", + "The existing task-oriented dialogue systems usually consist of four components: (1) natural language understanding (NLU), it tries to identify the intent of a user; (2) dialogue state tracker (DST), it keeps the track of user goals and constraints in every turn; (3) dialogue policy maker (DP), it aims to generate the next available dialogue action; and (4) natural language generator (NLG), it generates a natural language response based on the dialogue action. Among the four components, dialogue policy maker plays a key role in order to complete dialogues effectively, because it decides the next dialogue action to be executed.", + "As far as we know, the dialogue policy makers in most existing task-oriented dialogue systems just use the classifiers of the predefined acts to obtain dialogue policy BIBREF0, BIBREF2, BIBREF4, BIBREF8, BIBREF9. The classification-based dialogue policy learning methods can assign either only a dialogue act and its corresponding parameters BIBREF10, BIBREF2, BIBREF0 or multiple dialogue acts without their corresponding parameters for a dialogue action BIBREF11. However, all these existing methods cannot obtain multiple dialogue acts and their corresponding parameters for a dialogue action at the same time.", + "Intuitively, it will be more reasonable to construct multiple dialogue acts and their corresponding parameters for a dialogue action at the same time. For example, it can be shown that there are 49.4% of turns in the DSTC2 dataset and 61.5% of turns in the Maluuba dataset have multiple dialogue acts and their corresponding parameters as the dialogue action. If multiple dialogue acts and their corresponding parameters can be obtained at the same time, the final response of task-oriented dialogue systems will become more accurate and effective. For example, as shown in Figure FIGREF3, a user wants to get the name of a cheap french restaurant. The correct dialogue policy should generate three acts in current dialogue turn: offer(name=name_slot), inform(food=french) and inform(food=cheap). Thus, the user's real thought may be: \u201cname_slot is a cheap french restaurant\". If losing the act offer, the system may generate a response like \u201cThere are some french restaurants\", which will be far from the user's goal.", + "To address this challenge, we propose a Generative Dialogue Policy model (GDP) by casting the dialogue policy learning problem as a sequence optimization problem. The proposed model generates a series of acts and their corresponding parameters by the learned dialogue policy. Specifically, our proposed model uses a recurrent neural network (RNN) as action decoder to construct dialogue policy maker instead of traditional classifiers. Attention mechanism is used to help the decoder decode dialogue acts and their corresponding parameters, and then the template-based natural language generator uses the results of the dialogue policy maker to choose an appropriate sentence template as the final response to the user.", + "Extensive experiments conducted on two benchmark datasets verify the effectiveness of our proposed method. Our contributions in this work are three-fold.", + "The existing methods cannot construct multiple dialogue acts and their corresponding parameters at the same time. In this paper, We propose a novel generative dialogue policy model to solve the problem.", + "The extensive experiments demonstrate that the proposed model significantly outperforms the state-of-the-art baselines on two benchmarks.", + "We publicly release the source code." + ], + [ + "Usually, the existing task-oriented dialogue systems use a pipeline of four separate modules: natural language understanding, dialogue belief tracker, dialogue policy and natural language generator. Among these four modules, dialogue policy maker plays a key role in task-oriented dialogue systems, which generates the next dialogue action.", + "As far as we know, nearly all the existing approaches obtain the dialogue policy by using the classifiers of all predefined dialogue acts BIBREF12, BIBREF13. There are usually two kinds of dialogue policy learning methods. One constructs a dialogue act and its corresponding parameters for a dialogue action. For example, BIBREF0 constructs a simple classifier for all the predefined dialogue acts. BIBREF2 build a complex classifier for some predefined dialogue acts, addtionally BIBREF2 adds two acts for each parameter: one to inform its value and the other to request it. The other obtains the dialogue policy by using multi-label classification to consider multiple dialogue acts without their parameters. BIBREF11 performs multi-label multi-class classification for dialogue policy learning and then the multiple acts can be decided based on a threshold. Based on these classifiers, the reinforcement learning can be used to further update the dialogue policy of task-oriented dialogue systems BIBREF3, BIBREF14, BIBREF9.", + "In the real scene, an correct dialogue action usually consists of multiple dialogue acts and their corresponding parameters. However, it is very hard for existing classification-based dialogue policy maker to achieve this goal. Thus, in this paper we propose a novel generative dialogue policy maker to address this issue by casting the dialogue policy learning problem as a sequence optimization problem." + ], + [ + "Seq2Seq model was first introduced by BIBREF15 for statistical machine translation. It uses two recurrent neural networks (RNN) to solve the sequence-to-sequence mapping problem. One called encoder encodes the user utterance into a dense vector representing its semantics, the other called decoder decodes this vector to the target sentence. Now Seq2Seq framework has already been used in task-oriented dialog systems such as BIBREF4 and BIBREF1, and shows the challenging performance. In the Seq2Seq model, given the user utterance $Q=(x_1, x_2, ..., x_n)$, the encoder squeezes it into a context vector $C$ and then used by decoder to generate the response $R=(y_1, y_2, ..., y_m)$ word by word by maximizing the generation probability of $R$ conditioned on $Q$. The objective function of Seq2Seq can be written as:", + "In particular, the encoder RNN produces the context vector $C$ by doing calculation below:", + "The $h_t$ is the hidden state of the encoder RNN at time step $t$ and $f$ is the non-linear transformation which can be a long-short term memory unit LSTM BIBREF16 or a gated recurrent unit GRU BIBREF15. In this paper, we implement $f$ by using GRU.", + "The decoder RNN generates each word in reply conditioned on the context vector $C$. The probability distribution of candidate words at every time step $t$ is calculated as:", + "The $s_t$ is the hidden state of decoder RNN at time step $t$ and $y_{t-1}$ is the generated word in the reply at time $t-1$ calculated by softmax operations." + ], + [ + "Attention mechanisms BIBREF17 have been proved to improved effectively the generation quality for the Seq2Seq framework. In Seq2Seq with attention, each $y_i$ corresponds to a context vector $C_i$ which is calculated dynamically. It is a weighted average of all hidden states of the encoder RNN. Formally, $C_i$ is defined as $C_i=\\sum _{j=1}^{n} \\alpha _{ij}h_j$, where $\\alpha _{ij}$ is given by:", + "where $s_{i-1}$ is the last hidden state of the decoder, the $\\eta $ is often implemented as a multi-layer-perceptron (MLP) with tanh as the activation function." + ], + [ + "Figure FIGREF13 shows the overall system architecture of the proposed GDP model. Our model contains five main components: (1) utterance encoder; (2) dialogue belief tracker; (3) dialogue policy maker; (4) knowledge base; (5) template-based natural language generator. Next, we will describe each component of our proposed GDP model in detail." + ], + [ + "Given the user utterance $U_t$ at turn $t$ and the dialogue context $C_{t-1}$ which contains the result of the dialogue belief tracker at turn $t-1$, the task-oriented dialog system needs to generate user's intents $C_t$ by dialogue belief tracker and then uses this information to get the knowledge base query result $k_t \\in \\mathbb {R}^k$. Then the model needs to generate the next dialogue action $A_t$ based on $k_t$, $U_t$ and $C_t$. The natural language generator provides the template-based response $R_t$ as the final reply by using $A_t$. The $U_t$ and $C_t$ are the sequences, $k_t$ is a one-hot vector representing the number of the query results. For baselines, in this paper, the $A_t$ is the classification result of the next dialogue action, but in our proposed model it's a sequence which contains multiple acts and their corresponding parameters." + ], + [ + "A bidirectional GRU is used to encode the user utterance $U_t$, the last turn response $R_{t-1}$ made by the system and the dialogue context $C_{t-1}$ into a continuous representation. The vector is generated by concatenating the last forward and backward GRU states. $U_t = (w_1, w_2, ..., w_{T_m})$ is the user utterance at turn $t$. $C_{t-1}=(c_1, c_2, ..., c_{T_n})$ is the dialogue context made by dialogue belief tracker at $t-1$ turn. $R_{t-1}$ is the response made by our task-oriented dialogue system at last turn. Then the words of $[C_{t-1}, R_{t-1}, U_t]$ are firstly mapped into an embedding space and further serve as the inputs of each step to the bidirectional GRU. Let $n$ denotes the number of words in the sequence $[C_{t-1}, R_{t-1}, U_t]$. The $\\overrightarrow{h_{t^{\\prime }}^u}$ and $\\overleftarrow{h_{t^{\\prime }}^u}$ represent the forward and backward GRU state outputs at time step $t^{\\prime }$. The encoder output of timestep $i$ denote as $\\overline{h_i^u}$.", + "where $e([C_{t-1}, R_{t-1}, U_t])$ is the embedding of the input sequence, $d_h$ is the hidden size of the GRU. $H_u$ contains the encoder hidden state of each timestep, which will be used by attention mechanism in dialogue policy maker." + ], + [ + "Dialogue state tracker maintains the state of a conversation and collects the user's goals during the dialogue. Recent work successfully represents this component as discriminative classifiers. BIBREF5 verified that the generation is a better way to model the dialogue state tracker.", + "Specifically, we use a GRU as the generator to decode the $C_t$ of current turn. In order to capture user intent information accurately, the basic attention mechanism is calculated when the decoder decodes the $C_t$ at each step, which is the same as the Eq. (DISPLAY_FORM12).", + "where $m$ is the length of $C_t$, $e(y_i)$ is the embedding of the token, $d_h$ is the hidden size of the GRU and the hidden state at $i$ timestep of the RNN in dialogue state tracker denote as $h_i^d$. The decoded token at step $i$ denotes as $y_i^d$." + ], + [ + "Knowledge base is a database that stores information about the related task. For example, in the restaurant reservation, a knowledge base stores the information of all the restaurants, such as location and price. After dialogue belief tracker, the $C_t$ will be used as the constraints to search the results in knowledge base. Then the one-hot vector $k_t$ will be produced when the system gets the number of the results.", + "The search result $k_t$ has a great influence on dialogue policy. For example, if the result has multiple matches, the system should request more constraints of the user. In practice, let $k_t$ be an one-hot vector of 20 dimensions to represent the number of query results. Then $k_t$ will be used as the cue for dialogue policy maker." + ], + [ + "In task-oriented dialogue systems, supervised classification is a straightforward solution for dialogue policy modeling. However, we observe that classification cannot hold enough information for dialogue policy modeling. The generative approach is another way to model the dialogue policy maker for task-oriented dialogue systems, which generates the next dialogue acts and their corresponding parameters based on the dialogue context word by word. Thus the generative approach converts the dialogue policy learning problem into a sequence optimization problem.", + "The dialogue policy maker generates the next dialogue action $A_t$ based on $k_t$ and $[H_u, H_d]$. Our proposed model uses the GRU as the action decoder to decode the acts and their parameters for the response. Particularly, at step $i$, for decoding $y_i^p$ of $A_t$, the decoder GRU takes the embedding of $y_{i-1}^p$ to generate a hidden vector $h_i^p$. Basic attention mechanism is calculated.", + "where $e$ is the embedding of the token, $c_u$ is the context vector of the input utterance and $c_d$ is the context vector of the dialogue state tracker. $h_i^p$ is the hidden state of the GRU in dialogue policy maker at $i$ timestep.", + "where $y_i^p$ is the token decoded at $i$ timestep. And the final results of dialogue policy maker denote as $A_t$, and the $k$ is the length of it. In our proposed model, the dialogue policy maker can be viewed as a decoder of the seq2seq model conditioned on $[C_{t-1},R_{t-1},U_t]$ and $k_t$." + ], + [ + "After getting the dialogue action $A_t$ by the learned dialogue policy maker, the task-oriented dialogue system needs to generate an appropriate response $R_t$ for users. We construct the natural language generator by using template sentences. For each dataset, we extract all the system responses, then we manually modify responses to construct the sentence templates for task-oriented dialogue systems. In our proposed model, the sequence of the acts and parameters $A_t$ will be used for searching appropriate template. However, the classification-based baselines use the categories of acts and their corresponding parameters to search the corresponding template." + ], + [ + "In supervised learning, because our proposed model is built in a seq2seq way, the standard cross entropy is adopted as our objective function to train dialogue belief tracker and dialogue policy maker.", + "After supervised learning, the dialogue policy can be further updated by using reinforcement learning. In the context of reinforcement learning, the decoder of dialogue policy maker can be viewed as a policy network, denoted as $\\pi _{\\theta }(y_j)$ for decoding $y_j$, $\\theta $ is the parameters of the decoder. Accordingly, the hidden state created by GRU is the corresponding state, and the choice of the current token $y_j$ is an action.", + "Reward function is also very important for reinforcement learning when decoding every token. To encourage our policy maker to generate correct acts and their corresponding parameters, we set the reward function as follows: once the dialogue acts and their parameters are decoded correctly, the reward is 2; otherwise, the reward is -5; only the label of the dialogue act is decoded correctly but parameters is wrong, the reward is 1; $\\lambda $ is a decay parameter. More details are shown in Sec SECREF41. In our proposed model, rewards can only be obtained at the end of decoding $A_t$. In order to get the rewards at each decoding step, we sample some results $A_t$ after choosing $y_j$, and the reward of $y_j$ is set as the average of all the sampled results' rewards.", + "In order to ensure that the model's performance is stable during the fine-tuning phase of reinforcement learning, we freeze the parameters of user utterance and dialogue belief tracker, only the parameters of the dialogue policy maker will be optimized by reinforcement learning. Policy gradient algorithm REINFORCE BIBREF18 is used for pretrained dialogue policy maker:", + "where the $m$ is the length of the decoded action. The objective function $J$ can be optimized by gradient descent." + ], + [ + "We evaluate the performance of the proposed model in three aspects: (1) the accuracy of the dialogue state tracker, it aims to show the impact of the dialogue state tracker on the dialogue policy maker; (2) the accuracy of dialogue policy maker, it aims to explain the performance of different methods of constructing dialogue policy; (3) the quality of the final response, it aims to explain the impact of the dialogue policy on the final dialogue response. The evaluation metrics are listed as follows:", + "BPRA: Belief Per-Response Accuracy (BPRA) tests the ability to generate the correct user intents during the dialogue. This metric is used to evaluate the accuracy of dialogue belief tracker BIBREF1.", + "APRA: Action Per-Response Accuracy (APRA) evaluates the per-turn accuracy of the dialogue actions generated by dialogue policy maker. For baselines, APRA evaluates the classification accuracy of the dialogue policy maker. But our model actually generates each individual token of actions, and we consider a prediction to be correct only if every token of the model output matches the corresponding token in the ground truth.", + "BLEU BIBREF19: The metric evaluates the quality of the final response generated by natural language generator. The metric is usually used to measure the performance of the task-oriented dialogue system.", + "We also choose the following metrics to evaluate the efficiency of training the model:", + "$\\mathbf {Time_{full}}$: The time for training the whole model, which is important for industry settings.", + "$\\mathbf {Time_{DP}}$: The time for training the dialogue policy maker in a task-oriented dialogue system." + ], + [ + "We adopt the DSTC2 BIBREF20 dataset and Maluuba BIBREF21 dataset to evaluate our proposed model. Both of them are the benchmark datasets for building the task-oriented dialog systems. Specifically, the DSTC2 is a human-machine dataset in the single domain of restaurant searching. The Maluuba is a very complex human-human dataset in travel booking domain which contains more slots and values than DSTC2. Detailed slot information in each dataset is shown in Table TABREF34." + ], + [ + "For comparison, we choose two state-of-the-art baselines and their variants.", + "E2ECM BIBREF11: In dialogue policy maker, it adopts a classic classification for skeletal sentence template. In our implement, we construct multiple binary classifications for each act to search the sentence template according to the work proposed by BIBREF11.", + "CDM BIBREF10: This approach designs a group of classifications (two multi-class classifications and some binary classifications) to model the dialogue policy.", + "E2ECM+RL: It fine tunes the classification parameters of the dialogue policy by REINFORCE BIBREF18.", + "CDM+RL: It fine tunes the classification of the act and corresponding parameters by REINFORCE BIBREF18.", + "In order to verify the performance of the dialogue policy maker, the utterance encoder and dialogue belief tracker of our proposed model and baselines is the same, only dialogue policy maker is different." + ], + [ + "For all models, the hidden size of dialogue belief tracker and utterance encoder is 350, and the embedding size $d_{emb}$ is set to 300. For our proposed model, the hidden size of decoder in dialogue policy maker is 150. The vocabulary size $|V|$ is 540 for DSTC2 and 4712 for Maluuba. And the size of $k_t$ is set to 20. An Adam optimizer BIBREF22 is used for training our models and baselines, with a learning rate of 0.001 for supervised training and 0.0001 for reinforcement learning. In reinforcement learning, the decay parameter $\\lambda $ is set to 0.8. The weight decay is set to 0.001. And early stopping is performed on developing set." + ], + [ + "The experimental results of the proposed model and baselines will be analyzed from the following aspects.", + "BPRA Results: As shown in Table TABREF35, most of the models have similar performance on BPRA on these two datasets, which can guarantee a consistent impact on the dialogue policy maker. All the models perform very well in BPRA on DSTC2 dataset. On Maluuba dataset, the BPRA decreases because of the complex domains. We can notice that BPRA of CDM is slightly worse than other models on Maluuba dataset, the reason is that the CDM's dialogue policy maker contains lots of classifications and has the bigger loss than other models because of complex domains, which affects the training of the dialogue belief tracker.", + "APRA Results: Compared with baselines, GDP achieves the best performance in APRA on two datasets. It can be noted that we do not compare with the E2ECM baseline in APRA. E2ECM only uses a simple classifier to recognize the label of the acts and ignores the parameters information. In our experiment, APRA of E2ECM is slightly better than our method. Considering the lack of parameters of the acts, it's unfair for our GDP method. Furthermore, the CDM baseline considers the parameters of the act. But GDP is far better than CDM in supervised learning and reinforcement learning.", + "BLEU Results: GDP significantly outperforms the baselines on BLEU. As mentioned above, E2ECM is actually slightly better than GDP in APRA. But in fact, we can find that the language quality of the response generated by GDP is still better than E2ECM, which proves that lack of enough parameters information makes it difficult to find the appropriate sentence template in NLG. It can be found that the BLEU of all models is very poor on Maluuba dataset. The reason is that Maluuba is a human-human task-oriented dialogue dataset, the utterances are very flexible, the natural language generator for all methods is difficult to generate an accurate utterance based on the context. And DSTC2 is a human-machine dialog dataset. The response is very regular so the effectiveness of NLG will be better than that of Maluuba. But from the results, the GDP is still better than the baselines on Maluuba dataset, which also verifies that our proposed method is more accurate in modeling dialogue policy on complex domains than the classification-based methods.", + "Time and Model Size: In order to obtain more accurate and complete dialogue policy for task-oriented dialogue systems, the proposed model has more parameters on the dialogue policy maker than baselines. As shown in Figure FIGREF44, E2ECM has the minimal dialogue policy parameters because of the simple classification. It needs minimum training time, but the performance of E2ECM is bad. The number of parameters in the CDM model is slightly larger than E2ECM. However, because both of them are classification methods, they all lose some important information about dialogue policy. Therefore, we can see from the experimental results that the quality of CDM's dialogue policy is as bad as E2ECM. The number of dialogue policy maker's parameters in GDP model is much larger than baselines. Although the proposed model need more time to be optimized by supervised learning and reinforcement learning, the performance is much better than all baselines." + ], + [ + "Table TABREF43 illustrates an example of our proposed model and baselines on DSTC2 dataset. In this example, a user's goal is to find a cheap restaurant in the east part of the town. In the current turn, the user wants to get the address of the restaurant.", + "E2ECM chooses the inform and offer acts accurately, but the lack of the inform's parameters makes the final output deviate from the user's goal. CDM generates the parameters of offer successfully, but the lack of the information of inform also leads to a bad result. By contrast, the proposed model GDP can generate all the acts and their corresponding parameters as the dialogue action. Interestingly, the final result of GDP is exactly the same as the ground truth, which verifies that the proposed model is better than the state-of-the-art baselines." + ], + [ + "In this paper, we propose a novel model named GDP. Our proposed model treats the dialogue policy modeling as the generative task instead of the discriminative task which can hold more information for dialogue policy modeling. We evaluate the GDP on two benchmark task-oriented dialogue datasets. Extensive experiments show that GDP outperforms the existing classification-based methods on both action accuracy and BLEU." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1258/instruction.md b/qasper-1258/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..96055f86578019c1dcea99aa02f9ba0b99c09108 --- /dev/null +++ b/qasper-1258/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: "How May I Help You?": Modeling Twitter Customer Service Conversations Using Fine-Grained Dialogue Acts + +Question: Which Twitter customer service industries are investigated? \ No newline at end of file diff --git a/qasper-1267/instruction.md b/qasper-1267/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..cb95524b9807b363b06d43ae15a12138074d2061 --- /dev/null +++ b/qasper-1267/instruction.md @@ -0,0 +1,86 @@ +Name of Paper: SIM: A Slot-Independent Neural Model for Dialogue State Tracking + +Question: How do they measure model size? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Problem Formulation", + "Slot-Independent Model", + "Slot-Independent Model ::: Input Representation", + "Slot-Independent Model ::: Contextual Representation", + "Slot-Independent Model ::: Inter-Attention", + "Experiment ::: Dataset", + "Experiment ::: Metrics", + "Experiment ::: Training Details", + "Experiment ::: Baseline models and result", + "Conclusion", + "Acknowledgement" + ], + "paragraphs": [ + [ + "With the rapid development in deep learning, there is a recent boom of task-oriented dialogue systems in terms of both algorithms and datasets. The goal of task-oriented dialogue is to fulfill a user's requests such as booking hotels via communication in natural language. Due to the complexity and ambiguity of human language, previous systems have included semantic decoding BIBREF0 to project natural language input into pre-defined dialogue states. These states are typically represented by slots and values: slots indicate the category of information and values specify the content of information. For instance, the user utterance \u201ccan you help me find the address of any hotel in the south side of the city\u201d can be decoded as $inform(area, south)$ and $request(address)$, meaning that the user has specified the value south for slot area and requested another slot address.", + "Numerous methods have been put forward to decode a user's utterance into slot values. Some use hand-crafted features and domain-specific delexicalization methods to achieve strong performance BIBREF1, BIBREF2. BIBREF0 employs CNN and pretrained embeddings to further improve the state tracking accuracy. BIBREF3 extends this work by using two additional statistical update mechanisms. BIBREF4 uses human teaching and feedback to boost the state tracking performance. BIBREF5 utilizes both global and local attention mechanism in the proposed GLAD model which obtains state-of-the-art results on WoZ and DSTC2 datasets. However, most of these methods require slot-specific neural structures for accurate prediction. For example, BIBREF5 defines a parametrized local attention matrix for each slot. Slot-specific mechanisms become unwieldy when the dialogue task involves many topics and slots, as is typical in a complex conversational setting like product troubleshooting. Furthermore, due to the sparsity of labels, there may not be enough data to thoroughly train each slot-specific network structure. BIBREF6, BIBREF7 both propose to remove the model's dependency on dialogue slots but there's no modification to the representation part, which could be crucial to textual understanding as we will show later.", + "To solve this problem, we need a state tracking model independent of dialogue slots. In other words, the network should depend on the semantic similarity between slots and utterance instead of slot-specific modules. To this end, we propose the Slot-Independent Model (SIM). Our model complexity does not increase when the number of slots in dialogue tasks go up. Thus, SIM has many fewer parameters than existing dialogue state tracking models. To compensate for the exclusion of slot-specific parameters, we incorporate better feature representation of user utterance and dialogue states using syntactic information and convolutional neural networks (CNN). The refined representation, in addition to cross and self-attention mechanisms, make our model achieve even better performance than slot-specific models. For instance, on Wizard-of-Oz (WOZ) 2.0 dataset BIBREF8, the SIM model obtains a joint-accuracy score of 89.5%, 1.4% higher than the previously best model GLAD, with only 22% of the number of parameters. On DSTC2 dataset, SIM achieves comparable performance with previous best models with only 19% of the model size." + ], + [ + "As outlined in BIBREF9, the dialogue state tracking task is formulated as follows: at each turn of dialogue, the user's utterance is semantically decoded into a set of slot-value pairs. There are two types of slots. Goal slots indicate the category, e.g. area, food, and the values specify the constraint given by users for the category, e.g. South, Mediterranean. Request slots refer to requests, and the value is the category that the user demands, e.g. phone, area. Each user's turn is thus decoded into turn goals and turn requests. Furthermore, to summarize the user's goals so far, the union of all previous turn goals up to the current turn is defined as joint goals.", + "Similarly, the dialogue system's reply from the previous round is labeled with a set of slot-value pairs denoted as system actions. The dialogue state tracking task requires models to predict turn goal and turn request given user's utterance and system actions from previous turns.", + "Formally, the ontology of dialogue, $O$, consists of all possible slots $S$ and the set of values for each slot, $V(s), \\forall s \\in S$. Specifically, req is the name for request slot and its values include all the requestable category information. The dialogue state tracking task is that, given the user's utterance in the $i$-th turn, $U$, and system actions from the $(i-1)$-th turn, $A=\\lbrace (s_1, v_1), ..., (s_q, v_q)\\rbrace $, where $s_j \\in S, v_j \\in V(s_j)$, the model should predict:", + "Turn goals: $\\lbrace (s_1, v_1), ..., (s_b, v_b)\\rbrace $, where $s_j \\in S, v_j \\in V(s_j)$,", + "Turn requests: $\\lbrace (req, v_1), ..., (req, v_t)\\rbrace $, where $v_j \\in V(req)$.", + "The joint goals at turn $i$ are then computed by taking the union of all the predicted turn goals from turn 1 to turn $i$.", + "Usually this prediction task is cast as a binary classification problem: for each slot-value pair $(s, v)$, determine whether it should be included in the predicted turn goals/requests. Namely, the model is to learn a mapping function $f(U, A, (s, v))\\rightarrow \\lbrace 0,1\\rbrace $." + ], + [ + "To predict whether a slot-value pair should be included in the turn goals/requests, previous models BIBREF0, BIBREF5 usually define network components for each slot $s\\in S$. This can be cumbersome when the ontology is large, and it suffers from the insufficient data problem: the labelled data for a single slot may not suffice to effectively train the parameters for the slot-specific neural networks structure.", + "Therefore, we propose that in the classification process, the model needs to rely on the semantic similarity between the user's utterance and slot-value pair, with system action information. In other words, the model should have only a single global neural structure independent of slots. We heretofore refer to this model as Slot-Independent Model (SIM) for dialogue state tracking." + ], + [ + "Suppose the user's utterance in the $i$-th turn contains $m$ words, $U=(w_1, w_2, ..., w_m)$. For each word $w_i$, we use GloVe word embedding $e_i$, character-CNN embedding $c_i$, Part-Of-Speech (POS) embedding $\\operatorname{POS}_i$, Named-Entity-Recognition (NER) embedding $\\operatorname{NER}_i$ and exact match feature $\\operatorname{EM}_i$. The POS and NER tags are extracted by spaCy and then mapped into a fixed-length vector. The exact matching feature has two bits, indicating whether a word and its lemma can be found in the slot-value pair representation, respectively. This is the first step to establish a semantic relationship between user utterance and slots. To summarize, we represent the user utterance as $X^U=\\lbrace {u}_1, {u}_2, ..., {u}_m\\rbrace \\in \\mathbb {R}^{m\\times d_u}, {u}_i=[e_i; c_i; \\operatorname{POS}_i; \\operatorname{NER}_i; \\operatorname{EM}_i]$.", + "For each slot-value pair $(s, v)$ either in system action or in the ontology, we get its text representation by concatenating the contents of slot and value. We use GloVe to embed each word in the text. Therefore, each slot-value pair in system actions is represented as $X^A\\in \\mathbb {R}^{a\\times d}$ and each slot-value pair in ontology is represented as $X^O\\in \\mathbb {R}^{o\\times d}$, where $a$ and $o$ is the number of words in the corresponding text." + ], + [ + "To incorporate contextual information, we employ a bi-directional RNN layer on the input representation. For instance, for user utterance,", + "We apply variational dropout BIBREF10 for RNN inputs, i.e. the dropout mask is shared over different timesteps.", + "After RNN, we use linear self-attention to get a single summarization vector for user utterance, using weight vector $w\\in \\mathbb {R}^{d_{rnn}}$ and bias scalar $b$:", + "For each slot-value pair in the system actions and ontology, we conduct RNN and linear self-attention summarization in a similar way. As the slot-value pair input is not a sentence, we only keep the summarization vector $s^A \\in \\mathbb {R}^{d_{rnn}}$ and $s^O \\in \\mathbb {R}^{d_{rnn}}$ for each slot-value pair in system actions and ontology respectively." + ], + [ + "To determine whether the current user utterance refers to a slot-value pair $(s, v)$ in the ontology, the model employs inter-attention between user utterance, system action and ontology. Similar to the framework in BIBREF5, we employ two sources of interactions.", + "The first is the semantic similarity between the user utterance, represented by embedding $R^U$ and each slot-value pair from ontology $(s, v)$, represented by embedding $s^O$. We linearly combine vectors in $R^U$ via the normalized inner product with $s^O$, which is then employed to compute the similarity score $y_1$:", + "The second source involves the system actions. The reason is that if the system requested certain information in the previous round, it is very likely that the user will give answer in this round, and the answer may refer to the question, e.g. \u201cyes\u201d or \u201cno\u201d to the question. Thus, we first attend to system actions from user utterance and then combine with the ontology to get similarity score. Suppose there are $L$ slot-values pairs in the system actions from previous round, represented by $s_1^A, ..., s_L^A$:", + "The final similarity score between the user utterance and a slot-value pair $(s, v)$ from the ontology is a linear combination of $y_1$ and $y_2$ and normalized using sigmoid function.", + "where $\\beta $ is a learned coefficient. The loss function is the sum of binary cross entropy over all slot-value pairs in the ontology:", + "where $y_{(s, v)}\\in \\lbrace 0, 1\\rbrace $ is the ground truth. We illustrate the model structure of SIM in fig:model." + ], + [ + "We evaluated our model on Wizard of Oz (WoZ) BIBREF8 and the second Dialogue System Technology Challenges BIBREF11. Both tasks are for restaurant reservation and have slot-value pairs of both goal and request types. WoZ has 4 kinds of slots (area, food, price range, request) and 94 values in total. DSTC2 has an additional slot name and 220 values in total. WoZ has 800 dialogues in the training and development set and 400 dialogues in the test set, while DSTC2 dataset consists of 2118 dialogues in the training and development set, and 1117 dialogues in the test set." + ], + [ + "We use accuracy on the joint goal and turn request as the evaluation metrics. Both are sets of slot-value pairs, so the predicted set must exactly match the answer to be judged as correct. For joint goals, if a later turn generates a slot-value pair where the slot has been specified in previous rounds, we replace the value with the latest content." + ], + [ + "We fix GloVe BIBREF12 as the word embedding matrix. The models are trained using ADAM optimizer BIBREF13 with an initial learning rate of 1e-3. The dimension of POS and NER embeddings are 12 and 8, respectively. In character-CNN, each character is embedded into a vector of length 50. The CNN window size is 3 and hidden size is 50. We apply a dropout rate of 0.1 for the input to each module. The hidden size of RNN is 125.", + "During training, we pick the best model with highest joint goal score on development set and report the result on the test set.", + "For DSTC2, we adhere to the standard procedure to use the N-best list from the noisy ASR results for testing. The ASR results are very noisy. We experimented with several strategies and ended up using only the top result from the N-best list. The training and validation on DSTC2 are based on noise-free user utterance. The WoZ task does not have ASR results available, so we directly use noise-free user utterance." + ], + [ + "We compare our model SIM with a number of baseline systems: delexicalization model BIBREF8, BIBREF1, the neural belief tracker model (NBT) BIBREF0, global-locally self-attentive model GLAD BIBREF5, large-scale belief tracking model LSBT BIBREF7 and scalable multi-domain dialogue state tracking model SMDST BIBREF6.", + "Table TABREF17 shows that, on WoZ dataset, SIM achieves a new state-of-the-art joint goal accuracy of 89.5%, a significant improvement of 1.4% over GLAD, and turn request accuracy of 97.3%, 0.2% above GLAD. On DSTC2 dataset, where noisy ASR results are used as user utterance during test, SIM obtains comparable results with GLAD. Furthermore, the better representation in SIM makes it significantly outperform previous slot-independent models LSBT and SMDST.", + "Furthermore, as SIM has no slot-specific neural network structures, its model size is much smaller than previous models. Table TABREF20 shows that, on WoZ and DSTC2 datasets, SIM model has the same number of parameters, which is only 23% and 19% of that in GLAD model.", + "Ablation Study. We conduct an ablation study of SIM on WoZ dataset. As shown in Table TABREF21, the additional utterance word features, including character, POS, NER and exact matching embeddings, can boost the performance by 2.4% in joint goal accuracy. These features include POS, NER and exact match features. This indicates that for the dialogue state tracking task, syntactic information and text matching are very useful. Character-CNN captures sub-word level information and is effective in understanding spelling errors, hence it helps with 1.2% in joint goal accuracy. Variational dropout is also beneficial, contributing 0.9% to the joint goal accuracy, which shows the importance of uniform masking during dropout." + ], + [ + "In this paper, we propose a slot-independent neural model, SIM, to tackle the dialogue state tracking problem. Via incorporating better feature representations, SIM can effectively reduce the model complexity while still achieving superior or comparable results on various datasets, compared with previous models.", + "For future work, we plan to design general slot-free dialogue state tracking models which can be adapted to different domains during inference time, given domain-specific ontology information. This will make the model more agile in real applications." + ], + [ + "We thank the anonymous reviewers for the insightful comments. We thank William Hinthorn for proof-reading our paper." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1293/instruction.md b/qasper-1293/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..292bc785092f176d3582a6027b062c9de2d49d2c --- /dev/null +++ b/qasper-1293/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Czech Text Processing with Contextual Embeddings: POS Tagging, Lemmatization, Parsing and NER + +Question: What data is the Prague Dependency Treebank built on? \ No newline at end of file diff --git a/qasper-1431/instruction.md b/qasper-1431/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4f9b4783c609bbea9be9743e5f7e44391993eaf4 --- /dev/null +++ b/qasper-1431/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Benchmarking Zero-shot Text Classification: Datasets, Evaluation and Entailment Approach + +Question: Do they use pretrained models? \ No newline at end of file diff --git a/qasper-1455/instruction.md b/qasper-1455/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..86ab09cd5a334101a080e6b35ae0c128f0a192ec --- /dev/null +++ b/qasper-1455/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Weighed Domain-Invariant Representation Learning for Cross-domain Sentiment Analysis + +Question: How are different domains weighted in WDIRL? \ No newline at end of file diff --git a/qasper-1462/instruction.md b/qasper-1462/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..30292ac9117706ab8101655b3551df8a6c074984 --- /dev/null +++ b/qasper-1462/instruction.md @@ -0,0 +1,238 @@ +Name of Paper: The State of NLP Literature: A Diachronic Analysis of the ACL Anthology + +Question: What aspect of NLP research is examined? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Size", + "Demographics (focus of analysis: gender, age, and geographic diversity)", + "Demographics (focus of analysis: gender, age, and geographic diversity) ::: Gender", + "Demographics (focus of analysis: gender, age, and geographic diversity) ::: Academic Age", + "Demographics (focus of analysis: gender, age, and geographic diversity) ::: Location (Languages)", + "Areas of Research", + "Impact", + "Impact ::: #Citations and Most Cited Papers", + "Impact ::: Average Citations by Time Span", + "Impact ::: Aggregate Citation Statistics, by Paper Type and Venue", + "Impact ::: Citations to Papers by Areas of Research", + "Correlation of Age and Gender with Citations", + "Correlation of Age and Gender with Citations ::: Correlation of Academic Age with Citations", + "Conclusions" + ], + "paragraphs": [ + [ + "The ACL Anthology (AA) is a digital repository of tens of thousands of articles on Natural Language Processing (NLP) / Computational Linguistics (CL). It includes papers published in the family of ACL conferences as well as in other NLP conferences such as LREC and RANLP. AA is the largest single source of scientific literature on NLP.", + "This project, which we call NLP Scholar, examines the literature as a whole to identify broad trends in productivity, focus, and impact. We will present the analyses in a sequence of questions and answers. The questions range from fairly mundane to oh-that-will-be-good-to-know. Our broader goal here is simply to record the state of the AA literature: who and how many of us are publishing? what are we publishing on? where and in what form are we publishing? and what is the impact of our publications? The answers are usually in the form of numbers, graphs, and inter-connected visualizations.", + "We focus on the following aspects of NLP research: size, demographics, areas of research, impact, and correlation of citations with demographic attributes (age and gender).", + "", + "Target Audience: The analyses presented here are likely to be of interest to any NLP researcher. This might be particularly the case for those that are new to the field and wish to get a broad overview of the NLP publishing landscape. On the other hand, even seasoned NLP'ers have likely wondered about the questions raised here and might be interested in the empirical evidence.", + "", + "Data: The analyses presented below are based on information about the papers taken directly from AA (as of June 2019) and citation information extracted from Google Scholar (as of June 2019). Thus, all subsequent papers and citations are not included in the analysis. A fresh data collection is planned for January 2020.", + "", + "Interactive Visualizations: The visualizations we are developing for this work (using Tableau) are interactive\u2014so one can hover, click to select and filter, move sliders, etc. Since this work is high in the number of visualizations, the main visualizations are presented as figures in the paper and some sets of visualizations are pointed to online. The interactive visualizations and data will be made available through the first author's website after peer review.", + "", + "Related Work: This work builds on past research, including that on Google Scholar BIBREF0, BIBREF1, BIBREF2, BIBREF3, on the analysis of NLP papers BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, on citation intent BIBREF10, BIBREF11, BIBREF12, BIBREF13, BIBREF14, BIBREF15, and on measuring scholarly impact BIBREF16, BIBREF17, BIBREF18, BIBREF19, BIBREF20, BIBREF21.", + "", + "Caveats and Ethical Considerations: We list several caveats and limitations throughout the paper. A compilation of these is also available online in the About NLP Scholar page.", + "The analyses presented here are also available as a series of blog posts." + ], + [ + "Q. How big is the ACL Anthology (AA)? How is it changing with time?", + "A. As of June 2019, AA had $\\sim $50K entries, however, this includes some number of entries that are not truly research publications (for example, forewords, prefaces, table of contents, programs, schedules, indexes, calls for papers/participation, lists of reviewers, lists of tutorial abstracts, invited talks, appendices, session information, obituaries, book reviews, newsletters, lists of proceedings, lifetime achievement awards, erratum, and notes). We discard them for the analyses here. (Note: CL journal includes position papers like squibs, letter to editor, opinion, etc. We do not discard them.) We are then left with 44,896 articles. Figure FIGREF6 shows a graph of the number of papers published in each of the years from 1965 to 2018.", + "Discussion: Observe that there was a spurt in the 1990s, but things really took off since the year 2000, and the growth continues. Also, note that the number of publications is considerably higher in alternate years. This is due to biennial conferences. Since 1998 the largest of such conferences has been LREC (In 2018 alone LREC had over 700 main conferences papers and additional papers from its 29 workshops). COLING, another biennial conference (also occurring in the even years) has about 45% of the number of main conference papers as LREC.", + "Q. How many people publish in the ACL Anthology (NLP conferences)?", + "A. Figure FIGREF7 shows a graph of the number of authors (of AA papers) over the years:", + "Discussion: It is a good sign for the field to have a growing number of people join its ranks as researchers. A further interesting question would be:", + "Q. How many people are actively publishing in NLP?", + "A. It is hard to know the exact number, but we can determine the number of people who have published in AA in the last N years.", + "#people who published at least one paper in 2017 and 2018 (2 years): $\\sim $12k (11,957 to be precise)", + "#people who published at least one paper 2015 through 2018 (4 years):$\\sim $17.5k (17,457 to be precise)", + "Of course, some number of researchers published NLP papers in non-AA venues, and some number are active NLP researchers who may not have published papers in the last few years.", + "Q. How many journal papers exist in the AA? How many main conference papers? How many workshop papers?", + "A. See Figure FIGREF8.", + "Discussion: The number of journal papers is dwarfed by the number of conference and workshop papers. (This is common in computer science. Even though NLP is a broad interdisciplinary field, the influence of computer science practices on NLP is particularly strong.) Shared task and system demo papers are relatively new (introduced in the 2000s), but their numbers are already significant and growing.", + "Creating a separate class for \u201cTop-tier Conference\u201d is somewhat arbitrary, but it helps make certain comparisons more meaningful (for example, when comparing the average number of citations, etc.). For this work, we consider ACL, EMNLP, NAACL, COLING, and EACL as top-tier conferences, but certainly other groupings are also reasonable.", + "Q. How many papers have been published at ACL (main conference papers)? What are the other NLP venues and what is the distribution of the number of papers across various CL/NLP venues?", + "A. # ACL (main conference papers) as of June 2018: 4,839", + "The same workshop can co-occur with different conferences in different years, so we grouped all workshop papers in their own class. We did the same for tutorials, system demonstration papers (demos), and student research papers. Figure FIGREF9 shows the number of main conference papers for various venues and paper types (workshop papers, demos, etc.).", + "Discussion: Even though LREC is a relatively new conference that occurs only once in two years, it tends to have a high acceptance rate ($\\sim $60%), and enjoys substantial participation. Thus, LREC is already the largest single source of NLP conference papers. SemEval, which started as SenseEval in 1998 and occurred once in two or three years, has now morphed into an annual two-day workshop\u2014SemEval. It is the largest single source of NLP shared task papers." + ], + [ + "NLP, like most other areas of research, suffers from poor demographic diversity. There is very little to low representation from certain nationalities, race, gender, language, income, age, physical abilities, etc. This impacts the breadth of technologies we create, how useful they are, and whether they reach those that need it most. In this section, we analyze three specific attributes among many that deserve attention: gender (specifically, the number of women researchers in NLP), age (more precisely, the number of years of NLP paper publishing experience), and the amount of research in various languages (which loosely correlates with geographic diversity)." + ], + [ + "The ACL Anthology does not record demographic information about the paper authors. (Until recently, ACL and other NLP conferences did not record demographic information of the authors.) However, many first names have strong associations with a male or female gender. We will use these names to estimate the percentage of female first authors in NLP.", + "The US Social Security Administration publishes a database of names and genders of newborns. We use the dataset to identify 55,133 first names that are strongly associated with females (probability $\\ge $99%) and 29,873 first names that are strongly associated with males (probability $\\ge $99%). (As a side, it is interesting to note that there is markedly greater diversity in female names than in male names.) We identified 26,637 of the 44,896 AA papers ($\\sim $60%) where the first authors have one of these names and determine the percentage of female first author papers across the years. We will refer to this subset of AA papers as AA*.", + "", + "Note the following caveats associated with this analysis:", + "", + "The names dataset used has a lower representation of names from nationalities other than the US. However, there is a large expatriate population living in the US.", + "", + "Chinese names (especially in the romanized form) are not good indicators of gender. Thus the method presented here disregards most Chinese names, and the results of the analysis apply to the group of researchers excluding those with Chinese names.", + "", + "The dataset only records names associated with two genders.", + "", + "The approach presented here is meant to be an approximation in the absence of true gender information.", + "Q. What percent of the AA* papers have female first authors (FFA)? How has this percentage changed with time?", + "A. Overall FFA%: 30.3%. Figure FIGREF16 shows how FFA% has changed with time. Common paper title words and FFA% of papers that have those words are shown in the bottom half of the image. Note that the slider at the bottom has been set to 400, i.e., only those title words that occur in 400 or more papers are shown. The legend on the bottom right shows that low FFA scores are shown in shades of blue, whereas relatively higher FFA scores are shown in shades of green.", + "", + "Discussion: Observe that as a community, we are far from obtaining male-female parity in terms of first authors. A further striking (and concerning) observation is that the female first author percentage has not improved since the years 1999 and 2000 when the FFA percentages were highest (32.9% and 32.8%, respectively). In fact there seems to even be a slight downward trend in recent years. The calculations shown above are for the percentage of papers that have female first authors. The percentage of female first authors is about the same ($\\sim $31%). On average male authors had a slightly higher average number of publications than female authors.", + "To put these numbers in context, the percentage of female scientists world wide (considering all areas of research) has been estimated to be around 30%. The reported percentages for many computer science sub-fields are much lower. (See Women in Science (2015).) The percentages are much higher for certain other fields such as psychology and linguistics. (See this study for psychology and this study for linguistics.) If we can identify ways to move the needle on the FFA percentage and get it closer to 50% (or more), NLP can be a beacon to many other fields, especially in the sciences.", + "FFA percentages are particularly low for papers that have parsing, neural, and unsupervised in the title. There are some areas within NLP that enjoy a healthier female-male parity in terms of first authors of papers. Figure FIGREF20 shows FFA percentages for papers that have the word discourse in the title. There is burgeoning research on neural NLP in the last few years. Figure FIGREF21 shows FFA percentages for papers that have the word neural in the title.", + "Figure FIGREF22 shows lists of terms with the highest and lowest FFA percentages, respectively, when considering terms that occur in at least 50 paper titles (instead of 400 in the analysis above). Observe that FFA percentages are relatively higher in non-English European language research such as papers on Russian, Portuguese, French, and Italian. FFA percentages are also relatively higher for certain areas of NLP such as work on prosody, readability, discourse, dialogue, paraphrasing, and individual parts of speech such as adjectives and verbs. FFA percentages are particularly low for papers on theoretical aspects of statistical modelling, and areas such as machine translation, parsing, and logic. The full lists of terms and FFA percentages will be made available with the rest of the data." + ], + [ + "While the actual age of NLP researchers might be an interesting aspect to explore, we do not have that information. Thus, instead, we can explore a slightly different (and perhaps more useful) attribute: NLP academic age. We can define NLP academic age as the number of years one has been publishing in AA. So if this is the first year one has published in AA, then their NLP academic age is 1. If one published their first AA paper in 2001 and their latest AA paper in 2018, then their academic age is 18.", + "Q. How old are we? That is, what is the average NLP academic age of those who published papers in 2018? How has the average changed over the years? That is, have we been getting older or younger? What percentage of authors that published in 2018 were publishing their first AA paper?", + "A. Average NLP Academic Age of people that published in 2018: 5.41 years", + "Median NLP Academic Age of people that published in 2018: 2 years", + "Percentage of 2018 authors that published their first AA paper in 2018: 44.9%", + "Figure FIGREF24 shows how these numbers have changed over the years.", + "Discussion: Observe that the Average academic age has been steadily increasing over the years until 2016 and 2017, when the trend has shifted and the average academic age has started to decrease. The median age was 1 year for most of the 1965 to 1990 period, 2 years for most of the 1991 to 2006 period, 3 years for most of the 2007 to 2015 period, and back to 2 years since then. The first-time AA author percentage decreased until about 1988, after which it sort of stayed steady at around 48% until 2004 with occasional bursts to $\\sim $56%. Since 2005, the first-time author percentage has gone up and down every other year. It seems that the even years (which are also LREC years) have a higher first-time author percentage. Perhaps, this oscillation in first-time authors percentage is related to LREC\u2019s high acceptance rate.", + "Q. What is the distribution of authors in various academic age bins? For example, what percentage of authors that published in 2018 had an academic age of 2, 3, or 4? What percentage had an age between 5 and 9? And so on?", + "A. See Figure FIGREF25.", + "Discussion: Observe that about 65% of the authors that published in 2018 had an academic age of less than 5. This number has steadily reduced since 1965, was in the 60 to 70% range in 1990s, rose to the 70 to 72% range in early 2000s, then declined again until it reached the lowest value ($\\sim $60%) in 2010, and has again steadily risen until 2018 (65%). Thus, even though it may sometimes seem at recent conferences that there is a large influx of new people into NLP (and that is true), proportionally speaking, the average NLP academic age is higher (more experienced) than what it has been in much of its history." + ], + [ + "Automatic systems with natural language abilities are growing to be increasingly pervasive in our lives. Not only are they sources of mere convenience, but are crucial in making sure large sections of society and the world are not left behind by the information divide. Thus, the limits of what automatic systems can do in a language, limit the world for the speakers of that language.", + "We know that much of the research in NLP is on English or uses English datasets. Many reasons have been proffered, and we will not go into that here. Instead, we will focus on estimating how much research pertains to non-English languages.", + "We will make use of the idea that often when work is done focusing on a non-English language, then the language is mentioned in the title. We collected a list of 122 languages indexed by Wiktionary and looked for the presence of these words in the titles of AA papers. (Of course there are hundreds of other lesser known languages as well, but here we wanted to see the representation of these more prominent languages in NLP literature.)", + "Figure FIGREF27 is a treemap of the 122 languages arranged alphabetically and shaded such that languages that appear more often in AA paper titles have a darker shade of green.", + "Discussion: Even though the amount of work done on English is much larger than that on any other language, often the word English does not appear in the title, and this explains why English is not the first (but the second-most) common language name to appear in the titles. This is likely due to the fact that many papers fail to mention the language of study or the language of the datasets used if it is English. There is growing realization in the community that this is not quite right. However, the language of study can be named in other less prominent places than the title, for example the abstract, introduction, or when the datasets are introduced, depending on how central it is to the paper.", + "We can see from the treemap that the most widely spoken Asian and Western European languages enjoy good representation in AA. These include: Chinese, Arabic, Korean, Japanese, and Hindi (Asian) as well as French, German, Swedish, Spanish, Portuguese, and Italian (European). This is followed by the relatively less widely spoken European languages (such as Russian, Polish, Norwegian, Romanian, Dutch, and Czech) and Asian languages (such as Turkish, Thai, and Urdu). Most of the well-represented languages are from the Indo-European language family. Yet, even in the limited landscape of the most common 122 languages, vast swathes are barren with inattention. Notable among these is the extremely low representation of languages from Africa, languages from non-Indo-European language families, and Indigenous languages from around the world." + ], + [ + "Natural Language Processing addresses a wide range of research questions and tasks pertaining to language and computing. It encompasses many areas of research that have seen an ebb and flow of interest over the years. In this section, we examine the terms that have been used in the titles of ACL Anthology (AA) papers. The terms in a title are particularly informative because they are used to clearly and precisely convey what the paper is about. Some journals ask authors to separately include keywords in the paper or in the meta-information, but AA papers are largely devoid of this information. Thus titles are an especially useful source of keywords for papers\u2014keywords that are often indicative of the area of research.", + "Keywords could also be extracted from abstracts and papers; we leave that for future work. Further work is also planned on inferring areas of research using word embeddings, techniques from topic modelling, and clustering. There are clear benefits to performing analyses using that information. However, those approaches can be sensitive to the parameters used. Here, we keep things simple and explore counts of terms in paper titles. Thus the results are easily reproducible and verifiable.", + "Caveat: Even though there is an association between title terms and areas of research, the association can be less strong for some terms and areas. We use the association as one (imperfect) source of information about areas of research. This information may be combined with other sources of information to draw more robust conclusions.", + "Title Terms: The title has a privileged position in a paper. It serves many functions, and here are three key ones (from an article by Sneha Kulkarni): \"A good research paper title: 1. Condenses the paper's content in a few words 2. Captures the readers' attention 3. Differentiates the paper from other papers of the same subject area\".", + "If we examine the titles of papers in the ACL Anthology, we would expect that because of Function 1 many of the most common terms will be associated with the dominant areas of research. Function 2 (or attempting to have a catchy title) on the other hand, arguably leads to more unique and less frequent title terms. Function 3 seems crucial to the effectiveness of a title; and while at first glance it may seem like this will lead to unique title terms, often one needs to establish a connection with something familiar in order to convey how the work being presented is new or different.", + "It is also worth noting that a catchy term today, will likely not be catchy tomorrow. Similarly, a distinctive term today, may not be distinctive tomorrow. For example, early papers used neural in the title to distinguish themselves from non-nerual approaches, but these days neural is not particularly discriminative as far as NLP papers go.", + "Thus, competing and complex interactions are involved in the making of titles. Nonetheless, an arguable hypothesis is that: broad trends in interest towards an area of research will be reflected, to some degree, in the frequencies of title terms associated with that area over time. However, even if one does not believe in that hypothesis, it is worth examining the terms in the titles of tens of thousands of papers in the ACL Anthology\u2014spread across many decades.", + "Q. What terms are used most commonly in the titles of the AA papers? How has that changed with time?", + "A. Figure FIGREF28 shows the most common unigrams (single word) and bigrams (two-word sequences) in the titles of papers published from 1980 to 2019. (Ignoring function words.) The timeline graph at the bottom shows the percentage of occurrences of the unigrams over the years (the colors of the unigrams in the Timeline match those in the Title Unigram list). Note: For a given year, the timeline graph includes a point for a unigram if the sum of the frequency of the unigram in that year and the two years before it is at least ten. The period before 1980 is not included because of the small number of papers.", + "Discussion: Appropriately enough, the most common term in the titles of NLP papers is language. Presence of high-ranking terms pertaining to machine translation suggest that it is the area of research that has received considerable attention. Other areas associated with the high-frequency title terms include lexical semantics, named entity recognition, question answering, word sense disambiguation, and sentiment analysis. In fact, the common bigrams in the titles often correspond to names of NLP research areas. Some of the bigrams like shared task and large scale are not areas of research, but rather mechanisms or trends of research that apply broadly to many areas of research. The unigrams, also provide additional insights, such as the interest of the community in Chinese language, and in areas such as speech and parsing.", + "The Timeline graph is crowded in this view, but clicking on a term from the unigram list will filter out all other lines from the timeline. This is especially useful for determining whether the popularity of a term is growing or declining. (One can already see from above that neural has broken away from the pack in recent years.) Since there are many lines in the Timeline graph, Tableau labels only some (you can see neural and machine). However, hovering over a line, in the eventual interactive visualization, will display the corresponding term\u2014as shown in the figure.", + "Despite being busy, the graph sheds light on the relative dominance of the most frequent terms and how that has changed with time. The vocabulary of title words is smaller when considering papers from the 1980's than in recent years. (As would be expected since the number of papers then was also relatively fewer.) Further, dominant terms such as language and translation accounted for a higher percentage than in recent years where there is a much larger diversity of topics and the dominant research areas are not as dominant as they once were.", + "Q. What are the most frequent unigrams and bigrams in the titles of recent papers?", + "A. Figure FIGREF29 shows the most frequent unigrams and bigrams in the titles of papers published 2016 Jan to 2019 June (time of data collection).", + "Discussion: Some of the terms that have made notable gains in the top 20 unigrams and bigrams lists in recent years include: neural machine (presumably largely due to the phrase neural machine translation), neural network(s), word embeddings, recurrent neural, deep learning and the corresponding unigrams (neural, networks, etc.). We also see gains for terms related to shared tasks such as SemEval and task.", + "The sets of most frequent unigrams and bigrams in the titles of AA papers from various time spans are available online. Apart from clicking on terms, one can also enter the query (say parsing) in the search box at the bottom. Apart from filtering the timeline graph (bottom), this action also filters the unigram list (top left) to provide information only about the search term. This is useful because the query term may not be one of the visible top unigrams.", + "FigureFIGREF31 shows the timeline graph for parsing.", + "Discussion: Parsing seems to have enjoyed considerable attention in the 1980s, began a period of steep decline in the early 1990s, and a period of gradual decline ever since. One can enter multiple terms in the search box or shift/command click multiple terms to show graphs for more than one term.", + "FigureFIGREF32 shows the timelines for three bigrams statistical machine, neural machine, and machine translation:", + "Discussion: The graph indicates that there was a spike in machine translation papers in 1996, but the number of papers dropped substantially after that. Yet, its numbers have been comparatively much higher than other terms. One can also see the rise of statistical machine translation in the early 2000s followed by its decline with the rise of neural machine translation." + ], + [ + "Research articles can have impact in a number of ways\u2014pushing the state of the art, answering crucial questions, finding practical solutions that directly help people, making a new generation of potential-scientists excited about a field of study, and more. As scientists, it seems attractive to quantitatively measure scientific impact, and this is particularly appealing to governments and funding agencies; however, it should be noted that individual measures of research impact are limited in scope\u2014they measure only some kinds of contributions. Citations", + "The most commonly used metrics of research impact are derived from citations. A citation of a scholarly article is the explicit reference to that article. Citations serve many functions. However, a simplifying assumption is that regardless of the reason for citation, every citation counts as credit to the influence or impact of the cited work. Thus several citation-based metrics have emerged over the years including: number of citations, average citations, h-index, relative citation ratio, and impact factor.", + "It is not always clear why some papers get lots of citations and others do not. One can argue that highly cited papers have captured the imagination of the field: perhaps because they were particularly creative, opened up a new area of research, pushed the state of the art by a substantial degree, tested compelling hypotheses, or produced useful datasets, among other things.", + "Note however, that the number of citations is not always a reflection of the quality or importance of a piece of work. Note also that there are systematic biases that prevent certain kinds of papers from accruing citations, especially when the contributions of a piece of work are atypical, not easily quantified, or in an area where the number of scientific publications is low. Further, the citations process can be abused, for example, by egregious self-citations.", + "Nonetheless, given the immense volume of scientific literature, the relative ease with which one can track citations using services such as Google Scholar and Semantic Scholar, and given the lack of other easily applicable and effective metrics, citation analysis is an imperfect but useful window into research impact.", + "In this section, we examine citations of AA papers. We focus on two aspects:", + "Most cited papers: We begin by looking at the most cited papers overall and in various time spans. We will then look at most cited papers by paper-type (long, short, demo, etc) and venue (ACL, LREC, etc.). Perhaps these make interesting reading lists. Perhaps they also lead to a qualitative understanding of the kinds of AA papers that have received lots of citations.", + "Aggregate citation metrics by time span, paper type, and venue: Access to citation information allows us to calculate aggregate citation metrics such as average and median citations of papers published in different time periods, published in different venues, etc. These can help answer questions such as: on average, how well cited are papers published in the 1990s? on average, how many citations does a short paper get? how many citations does a long paper get? how many citations for a workshop paper? etc.", + "Data: The analyses presented below are based on information about the papers taken directly from AA (as of June 2019) and citation information extracted from Google Scholar (as of June 2019). We extracted citation information from Google Scholar profiles of authors who had a Google Scholar Profile page and had published at least three papers in the ACL Anthology. This yielded citation information for about 75% of the papers (33,051 out of the 44,896 papers). We will refer to this subset of the ACL Anthology papers as AA\u2019. All citation analysis below is on AA\u2019." + ], + [ + "Q. How many citations have the AA\u2019 papers received? How is that distributed among the papers published in various decades?", + "A. $\\sim $1.2 million citations (as of June 2019). Figure FIGREF36 shows a timeline graph where each year has a bar with height corresponding to the number of citations received by papers published in that year. Further, the bar has colored fragments corresponding to each of the papers and the height of a fragment (paper) is proportional to the number of citations it has received. Thus it is easy to spot the papers that received a large number of citations, and the years when the published papers received a large number of citations. Hovering over individual papers reveals an information box showing the paper title, authors, year of publication, publication venue, and #citations.", + "Discussion: With time, not only have the number of papers grown, but also the number of high-citation papers. We see a marked jump in the 1990s over the previous decades, but the 2000s are the most notable in terms of the high number of citations. The 2010s papers will likely surpass the 2000s papers in the years to come.", + "Q. What are the most cited papers in AA'?", + "A. Figure FIGREF37 shoes the most cited papers in the AA'.", + "Discussion: We see that the top-tier conference papers (green) are some of the most cited papers in AA\u2019. There are a notable number of journal papers (dark green) in the most cited list as well, but very few demo (purple) and workshop (orange) papers.", + "In the interactive visualizations (to be released later), one can click on the url to be to taken directly to the paper\u2019s landing page in the ACL Anthology website. That page includes links to meta information, the pdf, and associated files such as videos and appendices. There will also be functionality to download the lists. Alas, copying the lists from the screenshots shown here is not easy.", + "Q. What are the most cited AA' journal papers ? What are the most cited AA' workshop papers? What are the most cited AA' shared task papers? What are the most cited AA' demo papers? What are the most cited tutorials?", + "A. The most cited AA\u2019 journal papers, conference papers, workshop papers, system demo papers, shared task papers, and tutorials can be viewed online. The most cited papers from individual venues (ACL, CL journal, TACL, EMNLP, LREC, etc.) can also be viewed there.", + "Discussion: Machine translation papers are well-represented in many of these lists, but especially in the system demo papers list. Toolkits such as MT evaluation ones, NLTK, Stanford Core NLP, WordNet Similarity, and OpenNMT have highly cited demo or workshop papers.", + "The shared task papers list is dominated by task description papers (papers by task organizers describing the data and task), especially for sentiment analysis tasks. However, the list also includes papers by top-performing systems in these shared tasks, such as the NRC-Canada, HidelTime, and UKP papers.", + "Q. What are the most cited AA' papers in the last decade?", + "A. Figure FIGREF39 shows the most cited AA' papers in the 2010s. The most cited AA' papers from the earlier periods are available online.", + "Discussion: The early period (1965\u20131989) list includes papers focused on grammar and linguistic structure. The 1990s list has papers addressing many different NLP problems with statistical approaches. Papers on MT and sentiment analysis are frequent in the 2000s list. The 2010s are dominated by papers on word embeddings and neural representations." + ], + [ + "Q. How many citations did the papers published between 1990 and 1994 receive? What is the average number of citations that a paper published between 1990 and 1994 has received? What are the numbers for other time spans?", + "A. Total citations for papers published between 1990 and 1994: $\\sim $92k", + "Average citations for papers published between 1990 and 1994: 94.3", + "Figure FIGREF41 shows the numbers for various time spans.", + "Discussion: The early 1990s were an interesting period for NLP with the use of data from the World Wide Web and technologies from speech processing. This was the period with the highest average citations per paper, closely followed by the 1965\u20131969 and 1995\u20131999 periods. The 2000\u20132004 period is notable for: (1) a markedly larger number of citations than the previous decades; (2) third highest average number of citations. The drop off in the average citations for recent 5-year spans is largely because they have not had as much time to collect citations." + ], + [ + "Q. What are the average number of citations received by different types of papers: main conference papers, workshop papers, student research papers, shared task papers, and system demonstration papers?", + "A. In this analysis, we include only those AA\u2019 papers that were published in 2016 or earlier (to allow for at least 2.5 years to collect citations). There are 26,949 such papers. Figures FIGREF42 and FIGREF43 show the average citations by paper type when considering papers published 1965\u20132016 and 2010\u20132016, respectively. Figures FIGREF45 and FIGREF46 show the medians.", + "Discussion: Journal papers have much higher average and median citations than other papers, but the gap between them and top-tier conferences is markedly reduced when considering papers published since 2010.", + "System demo papers have the third highest average citations; however, shared task papers have the third highest median citations. The popularity of shared tasks and the general importance given to beating the state of the art (SOTA) seems to have grown in recent years\u2014something that has come under criticism.", + "It is interesting to note that in terms of citations, workshop papers are doing somewhat better than the conferences that are not top tier. Finally, the citation numbers for tutorials show that even though a small number of tutorials are well cited, a majority receive 1 or no citations. This is in contrast to system demo papers that have average and median citations that are higher or comparable to workshop papers.", + "Throughout the analyses in this article, we see that median citation numbers are markedly lower than average citation numbers. This is particularly telling. It shows that while there are some very highly cited papers, a majority of the papers obtain much lower number of citations\u2014and when considering papers other than journals and top-tier conferences, the number of citations is frequently lower than ten.", + "Q. What are the average number of citations received by the long and short ACL main conference papers, respectively?", + "A. Short papers were introduced at ACL in 2003. Since then ACL is by far the venue with the most number of short papers (compared to other venues). So we compare long and short papers published at ACL since 2003 to determine their average citations. Once again, we limit the papers to those published until 2016 to allow for the papers to have time to collect citations. Figure FIGREF47 shows the average and median citations for long and short papers.", + "Discussion: On average, long papers get almost three times as many citations as short papers. However, the median for long papers is two-and-half times that of short papers. This difference might be because some very heavily cited long papers push the average up for long papers.", + "Q. Which venue has publications with the highest average number of citations? What is the average number of citations for ACL and EMNLP papers? What is this average for other venues? What are the average citations for workshop papers, system demonstration papers, and shared task papers?", + "A. CL journal has the highest average citations per paper. Figure FIGREF49 shows the average citations for AA\u2019 papers published 1965\u20132016 and 2010\u20132016, respectively, grouped by venue and paper type. (Figure with median citations is available online.)", + "Discussion: In terms of citations, TACL papers have not been as successful as EMNLP and ACL; however, CL journal (the more traditional journal paper venue) has the highest average and median paper citations (by a large margin). This gap has reduced in papers published since 2010.", + "When considering papers published between 2010 and 2016, the system demonstration papers, the SemEval shared task papers, and non-SemEval shared task papers have notably high average (surpassing those of EACL and COLING); however their median citations are lower. This is likely because some heavily cited papers have pushed the average up. Nonetheless, it is interesting to note how, in terms of citations, demo and shared task papers have surpassed many conferences and even become competitive with some top-tier conferences such as EACL and COLING.", + "Q. What percent of the AA\u2019 papers that were published in 2016 or earlier are cited more than 1000 times? How many more than 10 times? How many papers are cited 0 times?", + "A. Google Scholar invented the i-10 index as another measure of author research impact. It stands for the number of papers by an author that received ten or more citations. (Ten here is somewhat arbitrary, but reasonable.) Similar to that, one can look at the impact of AA\u2019 as a whole and the impact of various subsets of AA\u2019 through the number of papers in various citation bins. Figure FIGREF50 shows the percentage of AA\u2019 papers in various citation bins. (The percentages of papers when considering papers from specific time spans are available online.)", + "Discussion: About 56% of the papers are cited ten or more times. 6.4% of the papers are never cited. Note also that some portion of the 1\u20139 bin likely includes papers that only received self-citations. It is interesting that the percentage of papers with 0 citations is rather steady (between 7.4% and 8.7%) for the 1965\u20131989, 1990\u20131999, and 2010\u20132016 periods. The majority of the papers lie in the 10 to 99 citations bin, for all except the recent periods (2010\u20132016 and 2016Jan\u20132016Dec). With time, the recent period should also have the majority of the papers in the 10 to 99 citations bin.", + "The numbers for the 2016Jan\u20132016Dec papers show that after 2.5 years, about 89% of the papers have at least one citation and about 33% of the papers have ten or more citations.", + "Q. What are the citation bin percentages for individual venues and paper types?", + "A. See Figure FIGREF51.", + "Discussion: Observe that 70 to 80% of the papers in journals and top-tier conferences have ten or more citations. The percentages are markedly lower (between 30 and 70%) for the other conferences shown above, and even lower for some other conferences (not shown above).", + "CL Journal is particularly notable for the largest percentage of papers with 100 or more citations. The somewhat high percentage of papers that are never cited (4.3%) are likely because some of the book reviews from earlier years are not explicitly marked in CL journal, and thus they were not removed from analysis. Also, letters to editors, which are more common in CL journal, tend to often obtain 0 citations.", + "CL, EMNLP, and ACL have the best track record for accepting papers that have gone on to receive 1000 or more citations. *Sem, the semantics conference, seems to have notably lower percentage of high-citation papers, even though it has fairly competitive acceptance rates.", + "Instead of percentage, if one considers raw numbers of papers that have at least ten citations (i-10 index), then LREC is particularly notable in terms of the large number of papers it accepts that have gone on to obtain ten or more citations ($\\sim $1600). Thus, by producing a large number of moderate-to-high citation papers, and introducing many first-time authors, LREC is one of the notable (yet perhaps undervalued) engines of impact on NLP.", + "About 50% of the SemEval shared task papers received 10 or more citations, and about 46% of the non-SemEval Shared Task Papers received 10 or more citations. About 47% of the workshop papers received ten or more citations. About 43% of the demo papers received 10 or more citations." + ], + [ + "Q. What is the average number of citations of AA' papers that have machine translation in the title? What about papers that have the term sentiment analysis or word representations?", + "A. Different areas of research within NLP enjoy varying amounts of attention. In Part II, we looked at the relative popularity of various areas over time\u2014estimated through the number of paper titles that had corresponding terms. (You may also want to see the discussion on the use of paper title terms to sample papers from various, possibly overlapping, areas.) Figure FIGREF53 shows the top 50 title bigrams ordered by decreasing number of total citations. Only those bigrams that occur in at least 30 AA' papers (published between 1965 and 2016) are considered. (The papers from 2017 and later are not included, to allow for at least 2.5 years for the papers to accumulate citations.)", + "Discussion: The graph shows that the bigram machine translation occurred in 1,659 papers that together accrued more than 93k citations. These papers have on average 68.8 citations and the median citations is 14. Not all machine translation (MT) papers have machine translation in the title. However, arguably, this set of 1,659 papers is a representative enough sample of machine translation papers; and thus, the average and median are estimates of MT in general. Second in the list are papers with statistical machine in the title\u2014most commonly from the phrase statistical machine translation. One expects considerable overlap in the papers across the sets of papers with machine translation and statistical machine, but machine translation likely covers a broader range of research including work before statistical MT was introduced, neural MT, and MT evaluation.", + "There are fewer papers with sentiment analysis in the title (356), but these have acquired citations at a higher average (104) than both machine translation and statistical machine. The bigram automatic evaluation jumps out because of its high average citations (337). Some of the neural-related bigrams have high median citations, for example, neural machine (49) and convolutional neural (40.5).", + "Figure FIGREF54 shows the lists of top 25 bigrams ordered by average citations.", + "Discussion: Observe the wide variety of topics covered by this list. In some ways that is reassuring for the health of the field as a whole; however, this list does not show which areas are not receiving sufficient attention. It is less clear to me how to highlight those, as simply showing the bottom 50 bigrams by average citations is not meaningful. Also note that this is not in any way an endorsement to write papers with these high-citation bigrams in the title. Doing so is of course no guarantee of receiving a large number of citations." + ], + [ + "In this section, we examine citations across two demographic dimensions: Academic age (number of years one has been publishing) and Gender. There are good reasons to study citations across each of these dimensions including, but not limited to, the following:", + "Areas of research: To better understand research contributions in the context of the area where the contribution is made.", + "Academic age: To better understand how the challenges faced by researchers at various stages of their career may impact the citations of their papers. For example, how well-cited are first-time NLP authors? On average, at what academic age do citations peak? etc.", + "Gender: To better understand the extent to which systematic biases (explicit and implicit) pervasive in society and scientific publishing impact author citations.", + "Some of these aspects of study may seem controversial. So it is worth addressing that first. The goal here is not to perpetuate stereotypes about age, gender, or even areas of research. The history of scientific discovery is awash with plenty of examples of bad science that has tried to erroneously show that one group of people is \u201cbetter\u201d than another, with devastating consequences.", + "People are far more alike than different. However, different demographic groups have faced (and continue to face) various socio-cultural inequities and biases. Gender and race studies look at how demographic differences shape our experiences. They examine the roles of social institutions in maintaining the inequities and biases.", + "This work is in support of those studies. Unless we measure differences in outcomes such as scientific productivity and impact across demographic groups, we will not fully know the extent to which these inequities and biases impact our scientific community; and we cannot track the effectiveness of measures to make our universities, research labs, and conferences more inclusive, equitable, and fair." + ], + [ + "We introduced NLP academic age earlier in the paper, where we defined NLP academic age as the number of years one has been publishing in AA. Here we examine whether NLP academic age impacts citations. The analyses are done in terms of the academic age of the first author; however, similar analyses can be done for the last author and all authors. (There are limitations to each of these analyses though as discussed further below.)", + "First author is a privileged position in the author list as it is usually reserved for the researcher that has done the most work and writing. The first author is also usually the main driver of the project; although, their mentor or advisor may also be a significant driver of the project. Sometimes multiple authors may be marked as first authors in the paper, but the current analysis simply takes the first author from the author list. In many academic communities, the last author position is reserved for the most senior or mentoring researcher. However, in non-university research labs and in large collaboration projects, the meaning of the last author position is less clear. (Personally, I prefer author names ordered by the amount of work done.)", + "Examining all authors is slightly more tricky as one has to decide how to credit the citations to the possibly multiple authors. It might also not be a clear indicator of differences across gender as a large number of the papers in AA have both male and female authors.", + "Q. How does the NLP academic age of the first author correlate with the amount of citations? Are first-year authors less cited than those with more experience?", + "A. Figure FIGREF59 shows various aggregate citation statistics corresponding to academic age. To produce the graph we put each paper in a bin corresponding to the academic age of the first author when the paper was published. For example, if the first author of a paper had an academic age of 3 when that paper was published, then the paper goes in bin 3. We then calculate #papers, #citations, median citations, and average citations for each bin. For the figure below, We further group the bins 10 to 14, 15 to 19, 20 to 34, and 35 to 50. These groupings are done to avoid clutter, and also because many of the higher age bins have a low number of papers.", + "Discussion: Observe that the number of papers where the first author has academic age 1 is much larger than the number of papers in any other bin. This is largely because a large number of authors in AA have written exactly one paper as first author. Also, about 60% of the authors in AA (17,874 out of the 29,941 authors) have written exactly one paper (regardless of author position).", + "The curves for the average and median citations have a slight upside down U shape. The relatively lower average and median citations in year 1 (37.26 and 10, respectively) indicate that being new to the field has some negative impact on citations. The average increases steadily from year 1 to year 4, but the median is already at the highest point by year 2. One might say, that year 2 to year 14 are the period of steady and high citations. Year 15 onwards, there is a steady decline in the citations. It is probably wise to not draw too many conclusions from the averages of the 35 to 50 bin, because of the small number of papers. There seems to be a peak in average citations at age 7. However, there is not a corresponding peak in the median. Thus the peak in average might be due to an increase in the number of very highly cited papers. Citations to Papers by First Author Gender", + "As noted in Part I, neither ACL nor the ACL Anthology have recorded demographic information for the vast majority of the authors. Thus we use the same setup discussed earlier in the section on demographics, to determine gender using the United States Social Security Administration database of names and genders of newborns to identify 55,133 first names that are strongly associated with females (probability $\\ge $99%) and 29,873 first names that are strongly associated with males (probability $\\ge $99%).", + "Q. On average, are women cited less than men?", + "A. Yes, on average, female first author papers have received markedly fewer citations than male first author papers (36.4 compared to 52.4). The difference in median is smaller (11 compared to 13). See Figure FIGREF60.", + "Discussion: The large difference in averages and smaller difference in medians suggests that there are markedly more very heavily cited male first-author papers than female first-author papers. The gender-unknown category, which here largely consist of authors with Chinese origin names and names that are less strongly associated with one gender have a slightly higher average, but the same median citations, as authors with female-associated first names.", + "The differences in citations, or citation gap, across genders may: (1) vary by period of time; (2) vary due to confounding factors such as academic age and areas of research. We explore these next.", + "Q. How has the citation gap across genders changed over the years?", + "A. Figure FIGREF61 (left side) shows the citation statistics across four time periods.", + "Discussion: Observe that female first authors have always been a minority in the history of ACL; however, on average, their papers from the early years (1965 to 1989) received a markedly higher number of citations than those of male first authors from the same period. We can see from the graph that this changed in the 1990s where male first-author papers obtained markedly more citations on average. The citation gap reduced considerably in the 2000s, and the 2010\u20132016 period saw a further slight reduction in the citation gap.", + "It is also interesting to note that the gender-unknown category has almost bridged the gap with the males in this most recent time period. Further, the proportion of the gender-unknown authors has increased over the years\u2014arguably, an indication of better representations of authors from around the world in recent years. (Nonetheless, as indicated in Part I, there is still plenty to be done to promote greater inclusion of authors from Africa and South America.)", + "Q. How have citations varied by gender and academic age? Are women less cited because of a greater proportion of new-to-NLP female first authors than new-to-NLP male first authors?", + "A. Figure FIGREF61 (right side) shows citation statistics broken down by gender and academic age. (This figure is similar to the academic age graph seen earlier, except that it shows separate average and median lines for female, male, and unknown gender first authors.)", + "Discussion: The graphs show that female first authors consistently receive fewer citations than male authors for the first fifteen years. The trend is inverted with a small citation gap in the 15th to 34th years period.", + "Q. Is the citation gap common across the vast majority of areas of research within NLP? Is the gap simply because more women work in areas that receive low numbers of citations (regardless of gender)?", + "A. Figure FIGREF64 shows the most cited areas of research along with citation statistics split by gender of the first authors of corresponding papers. (This figure is similar to the areas of research graph seen earlier, except that it shows separate citation statistics for the genders.) Note that the figure includes rows for only those bigram and gender pairs with at least 30 AA\u2019 papers (published between 1965 and 2016). Thus for some of the bigrams certain gender entries are not shown.", + "Discussion: Numbers for an additional 32 areas are available online. Observe that in only about 12% (7 of the top 59) of the most cited areas of research, women received higher average citations than men. These include: sentiment analysis, information extraction, document summarization, spoken dialogue, cross lingual (research), dialogue, systems, language generation. (Of course, note that some of the 59 areas, as estimated using title term bigrams, are overlapping. Also, we did not include large scale in the list above because the difference in averages is very small and it is not really an area of research.) Thus, the citation gap is common across a majority of the high-citations areas within NLP." + ], + [ + "This work examined the ACL Anthology to identify broad trends in productivity, focus, and impact. We examined several questions such as: who and how many of us are publishing? what are we publishing on? where and in what form are we publishing? and what is the impact of our publications? Particular attention was paid to the demographics and inclusiveness of the NLP community. Notably, we showed that only about 30% of first authors are female, and that this percentage has not improved since the year 2000. We also showed that, on average, female first authors are cited less than male first authors, even when controlling for academic age. We hope that recording citation and participation gaps across demographic groups will encourage our university, industry, and government research labs to be more inclusive and fair. Several additional aspects of the AA will be explored in future work (see the bottom of the blog posts).", + "Acknowledgments", + "This work was possible due to the helpful discussion and encouragement from a number of awesome people, including: Dan Jurafsky, Tara Small, Michael Strube, Cyril Goutte, Eric Joanis, Matt Post, Patrick Littell, Torsten Zesch, Ellen Riloff, Norm Vinson, Iryna Gurevych, Rebecca Knowles, Isar Nejadgholi, and Peter Turney. Also, a big thanks to the ACL Anthology team for creating and maintaining a wonderful resource." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1465/instruction.md b/qasper-1465/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..bbd7a7e063ccd92dbabdd10b1b6100b3d8d6d16f --- /dev/null +++ b/qasper-1465/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: BERTQA -- Attention on Steroids + +Question: What ensemble methods are used for best model? \ No newline at end of file diff --git a/qasper-1491/instruction.md b/qasper-1491/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..9caf973dcc055800cc121d324d81f652d30e2cf5 --- /dev/null +++ b/qasper-1491/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Grammatical Analysis of Pretrained Sentence Encoders with Acceptability Judgments + +Question: Do they report results only on English data? \ No newline at end of file diff --git a/qasper-1499/instruction.md b/qasper-1499/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..675408db5e0e21cb1c51c041d3b8df8562b72fa8 --- /dev/null +++ b/qasper-1499/instruction.md @@ -0,0 +1,118 @@ +Name of Paper: Question Answering by Reasoning Across Documents with Graph Convolutional Networks + +Question: How did they get relations between mentions? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Method", + "Dataset and task abstraction", + "Reasoning on an entity graph", + "Node annotations", + "Entity relational graph convolutional network", + "Experiments", + "Comparison", + "Ablation study", + "Error analysis", + "Related work", + "Conclusion", + "Acknowledgments", + "Architecture", + "Training details" + ], + "paragraphs": [ + [ + "The long-standing goal of natural language understanding is the development of systems which can acquire knowledge from text collections. Fresh interest in reading comprehension tasks was sparked by the availability of large-scale datasets, such as SQuAD BIBREF1 and CNN/Daily Mail BIBREF2 , enabling end-to-end training of neural models BIBREF3 , BIBREF4 , BIBREF5 . These systems, given a text and a question, need to answer the query relying on the given document. Recently, it has been observed that most questions in these datasets do not require reasoning across the document, but they can be answered relying on information contained in a single sentence BIBREF6 . The last generation of large-scale reading comprehension datasets, such as a NarrativeQA BIBREF7 , TriviaQA BIBREF8 , and RACE BIBREF9 , have been created in such a way as to address this shortcoming and to ensure that systems relying only on local information cannot achieve competitive performance.", + "Even though these new datasets are challenging and require reasoning within documents, many question answering and search applications require aggregation of information across multiple documents. The WikiHop dataset BIBREF0 was explicitly created to facilitate the development of systems dealing with these scenarios. Each example in WikiHop consists of a collection of documents, a query and a set of candidate answers (Figure 1 ). Though there is no guarantee that a question cannot be answered by relying just on a single sentence, the authors ensure that it is answerable using a chain of reasoning crossing document boundaries.", + "Though an important practical problem, the multi-hop setting has so far received little attention. The methods reported by BIBREF0 approach the task by merely concatenating all documents into a single long text and training a standard RNN-based reading comprehension model, namely, BiDAF BIBREF3 and FastQA BIBREF6 . Document concatenation in this setting is also used in Weaver BIBREF10 and MHPGM BIBREF11 . The only published paper which goes beyond concatenation is due to BIBREF12 , where they augment RNNs with jump-links corresponding to co-reference edges. Though these edges provide a structural bias, the RNN states are still tasked with passing the information across the document and performing multi-hop reasoning.", + "Instead, we frame question answering as an inference problem on a graph representing the document collection. Nodes in this graph correspond to named entities in a document whereas edges encode relations between them (e.g., cross- and within-document coreference links or simply co-occurrence in a document). We assume that reasoning chains can be captured by propagating local contextual information along edges in this graph using a graph convolutional network (GCN) BIBREF13 .", + "The multi-document setting imposes scalability challenges. In realistic scenarios, a system needs to learn to answer a query for a given collection (e.g., Wikipedia or a domain-specific set of documents). In such scenarios one cannot afford to run expensive document encoders (e.g., RNN or transformer-like self-attention BIBREF14 ), unless the computation can be preprocessed both at train and test time. Even if (similarly to WikiHop creators) one considers a coarse-to-fine approach, where a set of potentially relevant documents is provided, re-encoding them in a query-specific way remains the bottleneck. In contrast to other proposed methods (e.g., BIBREF12 , BIBREF10 , BIBREF3 ), we avoid training expensive document encoders.", + "In our approach, only a small query encoder, the GCN layers and a simple feed-forward answer selection component are learned. Instead of training RNN encoders, we use contextualized embeddings (ELMo) to obtain initial (local) representations of nodes. This implies that only a lightweight computation has to be performed online, both at train and test time, whereas the rest is preprocessed. Even in the somewhat contrived WikiHop setting, where fairly small sets of candidates are provided, the model is at least 5 times faster to train than BiDAF. Interestingly, when we substitute ELMo with simple pre-trained word embeddings, Entity-GCN still performs on par with many techniques that use expensive question-aware recurrent document encoders.", + "Despite not using recurrent document encoders, the full Entity-GCN model achieves over 2% improvement over the best previously-published results. As our model is efficient, we also reported results of an ensemble which brings further 3.6% of improvement and only 3% below the human performance reported by BIBREF0 . Our contributions can be summarized as follows:" + ], + [ + "In this section we explain our method. We first introduce the dataset we focus on, WikiHop by BIBREF0 , as well as the task abstraction. We then present the building blocks that make up our Entity-GCN model, namely, an entity graph used to relate mentions to entities within and across documents, a document encoder used to obtain representations of mentions in context, and a relational graph convolutional network that propagates information through the entity graph." + ], + [ + "The WikiHop dataset comprises of tuples $\\langle q, S_q, C_q, a^\\star \\rangle $ where: $q$ is a query/question, $S_q$ is a set of supporting documents, $C_q$ is a set of candidate answers (all of which are entities mentioned in $S_q$ ), and $a^\\star \\in C_q$ is the entity that correctly answers the question. WikiHop is assembled assuming that there exists a corpus and a knowledge base (KB) related to each other. The KB contains triples $\\langle s, r, o \\rangle $ where $s$ is a subject entity, $o$ an object entity, and $r$ a unidirectional relation between them. BIBREF0 used Wikipedia as corpus and Wikidata BIBREF15 as KB. The KB is only used for constructing WikiHop: BIBREF0 retrieved the supporting documents $q$0 from the corpus looking at mentions of subject and object entities in the text. Note that the set $q$1 (not the KB) is provided to the QA system, and not all of the supporting documents are relevant for the query but some of them act as distractors. Queries, on the other hand, are not expressed in natural language, but instead consist of tuples $q$2 where the object entity is unknown and it has to be inferred by reading the support documents. Therefore, answering a query corresponds to finding the entity $q$3 that is the object of a tuple in the KB with subject $q$4 and relation $q$5 among the provided set of candidate answers $q$6 .", + "The goal is to learn a model that can identify the correct answer $a^\\star $ from the set of supporting documents $S_q$ . To that end, we exploit the available supervision to train a neural network that computes scores for candidates in $C_q$ . We estimate the parameters of the architecture by maximizing the likelihood of observations. For prediction, we then output the candidate that achieves the highest probability. In the following, we present our model discussing the design decisions that enable multi-step reasoning and an efficient computation." + ], + [ + "In an offline step, we organize the content of each training instance in a graph connecting mentions of candidate answers within and across supporting documents. For a given query $q = \\langle s, r, ? \\rangle $ , we identify mentions in $S_q$ of the entities in $C_q \\cup \\lbrace s\\rbrace $ and create one node per mention. This process is based on the following heuristic:", + "we consider mentions spans in $S_q$ exactly matching an element of $C_q \\cup \\lbrace s\\rbrace $ . Admittedly, this is a rather simple strategy which may suffer from low recall.", + "we use predictions from a coreference resolution system to add mentions of elements in $C_q \\cup \\lbrace s\\rbrace $ beyond exact matching (including both noun phrases and anaphoric pronouns). In particular, we use the end-to-end coreference resolution by BIBREF16 .", + "we discard mentions which are ambiguously resolved to multiple coreference chains; this may sacrifice recall, but avoids propagating ambiguity.", + "To each node $v_i$ , we associate a continuous annotation $\\mathbf {x}_i \\in \\mathbb {R}^D$ which represents an entity in the context where it was mentioned (details in Section \"Node annotations\" ). We then proceed to connect these mentions i) if they co-occur within the same document (we will refer to this as DOC-BASED edges), ii) if the pair of named entity mentions is identical (MATCH edges\u2014these may connect nodes across and within documents), or iii) if they are in the same coreference chain, as predicted by the external coreference system (COREF edges). Note that MATCH edges when connecting mentions in the same document are mostly included in the set of edges predicted by the coreference system. Having the two types of edges lets us distinguish between less reliable edges provided by the coreference system and more reliable (but also more sparse) edges given by the exact-match heuristic. We treat these three types of connections as three different types of relations. See Figure 2 for an illustration. In addition to that, and to prevent having disconnected graphs, we add a fourth type of relation (COMPLEMENT edge) between any two nodes that are not connected with any of the other relations. We can think of these edges as those in the complement set of the entity graph with respect to a fully connected graph.", + "Our model then approaches multi-step reasoning by transforming node representations (Section \"Node annotations\" for details) with a differentiable message passing algorithm that propagates information through the entity graph. The algorithm is parameterized by a graph convolutional network (GCN) BIBREF13 , in particular, we employ relational-GCNs BIBREF17 , an extended version that accommodates edges of different types. In Section \"Entity relational graph convolutional network\" we describe the propagation rule.", + "Each step of the algorithm (also referred to as a hop) updates all node representations in parallel. In particular, a node is updated as a function of messages from its direct neighbours, and a message is possibly specific to a certain relation. At the end of the first step, every node is aware of every other node it connects directly to. Besides, the neighbourhood of a node may include mentions of the same entity as well as others (e.g., same-document relation), and these mentions may have occurred in different documents. Taking this idea recursively, each further step of the algorithm allows a node to indirectly interact with nodes already known to their neighbours. After $L$ layers of R-GCN, information has been propagated through paths connecting up to $L+1$ nodes.", + "We start with node representations $\\lbrace \\mathbf {h}_i^{(0)}\\rbrace _{i=1}^N$ , and transform them by applying $L$ layers of R-GCN obtaining $\\lbrace \\mathbf {h}_i^{(L)}\\rbrace _{i=1}^N$ . Together with a representation $\\mathbf {q}$ of the query, we define a distribution over candidate answers and we train maximizing the likelihood of observations. The probability of selecting a candidate $c \\in C_q$ as an answer is then ", + "$$ \nP(c|q, C_q, S_q) \\propto \\exp \\left(\\max _{i \\in \\mathcal {M}_c} f_o([\\mathbf {q}, \\mathbf {h}^{(L)}_i]) \\right)\\;,$$ (Eq. 16) ", + "where $f_o$ is a parameterized affine transformation, and $\\mathcal {M}_c$ is the set of node indices such that $i\\in \\mathcal {M}_c$ only if node $v_i$ is a mention of $c$ . The $\\max $ operator in Equation 16 is necessary to select the node with highest predicted probability since a candidate answer is realized in multiple locations via different nodes." + ], + [ + "Keeping in mind we want an efficient model, we encode words in supporting documents and in the query using only a pre-trained model for contextualized word representations rather than training our own encoder. Specifically, we use ELMo BIBREF20 , a pre-trained bi-directional language model that relies on character-based input representation. ELMo representations, differently from other pre-trained word-based models (e.g., word2vec BIBREF21 or GloVe BIBREF22 ), are contextualized since each token representation depends on the entire text excerpt (i.e., the whole sentence).", + "We choose not to fine tune nor propagate gradients through the ELMo architecture, as it would have defied the goal of not having specialized RNN encoders. In the experiments, we will also ablate the use of ELMo showing how our model behaves using non-contextualized word representations (we use GloVe).", + "ELMo encodings are used to produce a set of representations $\\lbrace \\mathbf {x}_i\\rbrace _{i=1}^N$ , where $\\mathbf {x}_i \\in \\mathbb {R}^D$ denotes the $i$ th candidate mention in context. Note that these representations do not depend on the query yet and no trainable model was used to process the documents so far, that is, we use ELMo as a fixed pre-trained encoder. Therefore, we can pre-compute representation of mentions once and store them for later use.", + "ELMo encodings are used to produce a query representation $\\mathbf {q} \\in \\mathbb {R}^K$ as well. Here, $\\mathbf {q}$ is a concatenation of the final outputs from a bidirectional RNN layer trained to re-encode ELMo representations of words in the query. The vector $\\mathbf {q}$ is used to compute a query-dependent representation of mentions $\\lbrace \\mathbf { \\hat{x}}_i\\rbrace _{i=1}^N$ as well as to compute a probability distribution over candidates (as in Equation 16 ). Query-dependent mention encodings $\\mathbf {\\hat{x}}_i = f_x(\\mathbf {q}, \\mathbf {x}_i)$ are generated by a trainable function $f_x$ which is parameterized by a feed-forward neural network." + ], + [ + "Our model uses a gated version of the original R-GCN propagation rule. At the first layer, all hidden node representation are initialized with the query-aware encodings $\\mathbf {h}_i^{(0)} = \\mathbf {\\hat{x}}_i$ . Then, at each layer $0\\le \\ell \\le L$ , the update message $\\mathbf {u}_i^{(\\ell )}$ to the $i$ th node is a sum of a transformation $f_s$ of the current node representation $\\mathbf {h}^{(\\ell )}_i$ and transformations of its neighbours: ", + "$$\\mathbf {u}^{(\\ell )}_i = f_s(\\mathbf {h}^{(\\ell )}_i) + \\frac{1}{|\\mathcal {N}_i|} \\sum _{j \\in \\mathcal {N}_i} \\sum _{r \\in \\mathcal {R}_{ij}} f_r(\\mathbf {h}_j^{(\\ell )})\\;,$$ (Eq. 22) ", + "where $\\mathcal {N}_i$ is the set of indices of nodes neighbouring the $i$ th node, $\\mathcal {R}_{ij}$ is the set of edge annotations between $i$ and $j$ , and $f_r$ is a parametrized function specific to an edge type $r\\in \\mathcal {R}$ . Recall the available relations from Section \"Ablation study\" , namely, $\\mathcal {R} =\\lbrace $ DOC-BASED, MATCH, COREF, COMPLEMENT $\\rbrace $ .", + "A gating mechanism regulates how much of the update message propagates to the next step. This provides the model a way to prevent completely overwriting past information. Indeed, if all necessary information to answer a question is present at a layer which is not the last, then the model should learn to stop using neighbouring information for the next steps. Gate levels are computed as ", + "$$\\mathbf {a}^{(\\ell )}_i = \\sigma \\left( f_a\\left([\\mathbf {u}^{(\\ell )}_i, \\mathbf {h}^{(\\ell )}_i ]\\right) \\right) \\;,$$ (Eq. 23) ", + "where $\\sigma (\\cdot )$ is the sigmoid function and $f_a$ a parametrized transformation. Ultimately, the updated representation is a gated combination of the previous representation and a non-linear transformation of the update message: ", + "$$\\mathbf {h}^{(\\ell + 1)}_i = \\phi (\\mathbf {u}^{(\\ell )}_i) \\odot \\mathbf {a}^{(\\ell )}_i + \\mathbf {h}^{(\\ell )}_i \\odot (1 - \\mathbf {a}^{(\\ell )}_i ) \\;,$$ (Eq. 24) ", + "where $\\phi (\\cdot )$ is any nonlinear function (we used $\\tanh $ ) and $\\odot $ stands for element-wise multiplication. All transformations $f_*$ are affine and they are not layer-dependent (since we would like to use as few parameters as possible to decrease model complexity promoting efficiency and scalability)." + ], + [ + "In this section, we compare our method against recent work as well as preforming an ablation study using the WikiHop dataset BIBREF0 . See Appendix \"Implementation and experiments details\" in the supplementary material for a description of the hyper-parameters of our model and training details." + ], + [ + "In this experiment, we compare our Enitity-GCN against recent prior work on the same task. We present test and development results (when present) for both versions of the dataset in Table 2 . From BIBREF0 , we list an oracle based on human performance as well as two standard reading comprehension models, namely BiDAF BIBREF3 and FastQA BIBREF6 . We also compare against Coref-GRU BIBREF12 , MHPGM BIBREF11 , and Weaver BIBREF10 . Additionally, we include results of MHQA-GRN BIBREF23 , from a recent arXiv preprint describing concurrent work. They jointly train graph neural networks and recurrent encoders. We report single runs of our two best single models and an ensemble one on the unmasked test set (recall that the test set is not publicly available and the task organizers only report unmasked results) as well as both versions of the validation set.", + "Entity-GCN (best single model without coreference edges) outperforms all previous work by over 2% points. We additionally re-ran BiDAF baseline to compare training time: when using a single Titan X GPU, BiDAF and Entity-GCN process 12.5 and 57.8 document sets per second, respectively. Note that BIBREF0 had to use BiDAF with very small state dimensionalities (20), and smaller batch size due to the scalability issues (both memory and computation costs). We compare applying the same reductions. Eventually, we also report an ensemble of 5 independently trained models. All models are trained on the same dataset splits with different weight initializations. The ensemble prediction is obtained as $\\arg \\max \\limits _c \\prod \\limits _{i=1}^5 P_i(c|q, C_q, S_q)$ from each model." + ], + [ + "To help determine the sources of improvements, we perform an ablation study using the publicly available validation set (see Table 3 ). We perform two groups of ablation, one on the embedding layer, to study the effect of ELMo, and one on the edges, to study how different relations affect the overall model performance.", + "We argue that ELMo is crucial, since we do not rely on any other context encoder. However, it is interesting to explore how our R-GCN performs without it. Therefore, in this experiment, we replace the deep contextualized embeddings of both the query and the nodes with GloVe BIBREF22 vectors (insensitive to context). Since we do not have any component in our model that processes the documents, we expect a drop in performance. In other words, in this ablation our model tries to answer questions without reading the context at all. For example, in Figure 1 , our model would be aware that \u201cStockholm\u201d and \u201cSweden\u201d appear in the same document but any context words, including the ones encoding relations (e.g., \u201cis the capital of\u201d) will be hidden. Besides, in the masked case all mentions become `unknown' tokens with GloVe and therefore the predictions are equivalent to a random guess. Once the strong pre-trained encoder is out of the way, we also ablate the use of our R-GCN component, thus completely depriving the model from inductive biases that aim at multi-hop reasoning.", + "The first important observation is that replacing ELMo by GloVe (GloVe with R-GCN in Table 3 ) still yields a competitive system that ranks far above baselines from BIBREF0 and even above the Coref-GRU of BIBREF12 , in terms of accuracy on (unmasked) validation set. The second important observation is that if we then remove R-GCN (GloVe w/o R-GCN in Table 3 ), we lose 8.0 points. That is, the R-GCN component pushes the model to perform above Coref-GRU still without accessing context, but rather by updating mention representations based on their relation to other ones. These results highlight the impact of our R-GCN component.", + "In this experiment we investigate the effect of the different relations available in the entity graph and processed by the R-GCN module. We start off by testing our stronger encoder (i.e., ELMo) in absence of edges connecting mentions in the supporting documents (i.e., using only self-loops \u2013 No R-GCN in Table 3 ). The results suggest that WikipHop genuinely requires multihop inference, as our best model is 6.1% and 8.4% more accurate than this local model, in unmasked and masked settings, respectively. However, it also shows that ELMo representations capture predictive context features, without being explicitly trained for the task. It confirms that our goal of getting away with training expensive document encoders is a realistic one.", + "We then inspect our model's effectiveness in making use of the structure encoded in the graph. We start naively by fully-connecting all nodes within and across documents without distinguishing edges by type (No relation types in Table 3 ). We observe only marginal improvements with respect to ELMo alone (No R-GCN in Table 3 ) in both the unmasked and masked setting suggesting that a GCN operating over a naive entity graph would not add much to this task and a more informative graph construction and/or a more sophisticated parameterization is indeed needed.", + "Next, we ablate each type of relations independently, that is, we either remove connections of mentions that co-occur in the same document (DOC-BASED), connections between mentions matching exactly (MATCH), or edges predicted by the coreference system (COREF). The first thing to note is that the model makes better use of DOC-BASED connections than MATCH or COREF connections. This is mostly because i) the majority of the connections are indeed between mentions in the same document, and ii) without connecting mentions within the same document we remove important information since the model is unaware they appear closely in the document. Secondly, we notice that coreference links and complement edges seem to play a more marginal role. Though it may be surprising for coreference edges, recall that the MATCH heuristic already captures the easiest coreference cases, and for the rest the out-of-domain coreference system may not be reliable. Still, modelling all these different relations together gives our Entity-GCN a clear advantage. This is our best system evaluating on the development. Since Entity-GCN seems to gain little advantage using the coreference system, we report test results both with and without using it. Surprisingly, with coreference, we observe performance degradation on the test set. It is likely that the test documents are harder for the coreference system.", + "We do perform one last ablation, namely, we replace our heuristic for assigning edges and their labels by a model component that predicts them. The last row of Table 3 (Induced edges) shows model performance when edges are not predetermined but predicted. For this experiment, we use a bilinear function $f_e(\\mathbf {\\hat{x}}_i, \\mathbf {\\hat{x}}_j) = \\sigma \\left( \\mathbf {\\hat{x}}^\\top _i \\mathbf {W}_e \\mathbf {\\hat{x}}_j \\right)$ that predicts the importance of a single edge connecting two nodes $i,j$ using the query-dependent representation of mentions (see Section \"Node annotations\" ). The performance drops below `No R-GCN' suggesting that it cannot learn these dependencies on its own.", + "Most results are stronger for the masked settings even though we do not apply the coreference resolution system in this setting due to masking. It is not surprising as coreferred mentions are labeled with the same identifier in the masked version, even if their original surface forms did not match ( BIBREF0 used Wikipedia links for masking). Indeed, in the masked version, an entity is always referred to via the same unique surface form (e.g., MASK1) within and across documents. In the unmasked setting, on the other hand, mentions to an entity may differ (e.g., \u201cUS\u201d vs \u201cUnited States\u201d) and they might not be retrieved by the coreference system we are employing, making the task harder for all models. Therefore, as we rely mostly on exact matching when constructing our graph for the masked case, we are more effective in recovering coreference links on the masked rather than unmasked version.", + "In Figure 3 , we show how the model performance goes when the input graph is large. In particular, how Entity-GCN performs as the number of candidate answers or the number of nodes increases." + ], + [ + "In this section we provide an error analysis for our best single model predictions. First of all, we look at which type of questions our model performs well or poorly. There are more than 150 query types in the validation set but we filtered the three with the best and with the worst accuracy that have at least 50 supporting documents and at least 5 candidates. We show results in Table 4 . We observe that questions regarding places (birth and death) are considered harder for Entity-GCN. We then inspect samples where our model fails while assigning highest likelihood and noticed two principal sources of failure i) a mismatch between what is written in Wikipedia and what is annotated in Wikidata, and ii) a different degree of granularity (e.g., born in \u201cLondon\u201d vs \u201cUK\u201d could be considered both correct by a human but not when measuring accuracy). See Table 6 in the supplement material for some reported samples.", + "Secondly, we study how the model performance degrades when the input graph is large. In particular, we observe a negative Pearson's correlation (-0.687) between accuracy and the number of candidate answers. However, the performance does not decrease steeply. The distribution of the number of candidates in the dataset peaks at 5 and has an average of approximately 20. Therefore, the model does not see many samples where there are a large number of candidate entities during training. Differently, we notice that as the number of nodes in the graph increases, the model performance drops but more gently (negative but closer to zero Pearson's correlation). This is important as document sets can be large in practical applications. See Figure 3 in the supplemental material for plots.", + "In Table 6 , we report three samples from WikiHop development set where out Entity-GCN fails. In particular, we show two instances where our model presents high confidence on the answer, and one where is not. We commented these samples explaining why our model might fail in these cases." + ], + [ + "In previous work, BiDAF BIBREF3 , FastQA BIBREF6 , Coref-GRU BIBREF12 , MHPGM BIBREF11 , and Weaver / Jenga BIBREF10 have been applied to multi-document question answering. The first two mainly focus on single document QA and BIBREF0 adapted both of them to work with WikiHop. They process each instance of the dataset by concatenating all $d \\in S_q$ in a random order adding document separator tokens. They trained using the first answer mention in the concatenated document and evaluating exact match at test time. Coref-GRU, similarly to us, encodes relations between entity mentions in the document. Instead of using graph neural network layers, as we do, they augment RNNs with jump links corresponding to pairs of corefereed mentions. MHPGM uses a multi-attention mechanism in combination with external commonsense relations to perform multiple hops of reasoning. Weaver is a deep co-encoding model that uses several alternating bi-LSTMs to process the concatenated documents and the query.", + "Graph neural networks have been shown successful on a number of NLP tasks BIBREF24 , BIBREF25 , BIBREF26 , including those involving document level modeling BIBREF27 . They have also been applied in the context of asking questions about knowledge contained in a knowledge base BIBREF28 . In schlichtkrull2017modeling, GCNs are used to capture reasoning chains in a knowledge base. Our work and unpublished concurrent work by BIBREF23 are the first to study graph neural networks in the context of multi-document QA. Besides differences in the architecture, BIBREF23 propose to train a combination of a graph recurrent network and an RNN encoder. We do not train any RNN document encoders in this work." + ], + [ + "We designed a graph neural network that operates over a compact graph representation of a set of documents where nodes are mentions to entities and edges signal relations such as within and cross-document coreference. The model learns to answer questions by gathering evidence from different documents via a differentiable message passing algorithm that updates node representations based on their neighbourhood. Our model outperforms published results where ablations show substantial evidence in favour of multi-step reasoning. Moreover, we make the model fast by using pre-trained (contextual) embeddings." + ], + [ + "We would like to thank Johannes Welbl for helping to test our system on WikiHop. This project is supported by SAP Innovation Center Network, ERC Starting Grant BroadSem (678254) and the Dutch Organization for Scientific Research (NWO) VIDI 639.022.518. Wilker Aziz is supported by the Dutch Organisation for Scientific Research (NWO) VICI Grant nr. 277-89-002." + ], + [ + "See table 5 for an outline of Entity-GCN architectural detail. Here the computational steps", + "ELMo embeddings are a concatenation of three 1024-dimensional vectors resulting in 3072-dimensional input vectors $\\lbrace \\mathbf {x}_i\\rbrace _{i=1}^N$ .", + "For the query representation $\\mathbf {q}$ , we apply 2 bi-LSTM layers of 256 and 128 hidden units to its ELMo vectors. The concatenation of the forward and backward states results in a 256-dimensional question representation.", + "ELMo embeddings of candidates are projected to 256-dimensional vectors, concatenated to the $\\mathbf {q}$ , and further transformed with a two layers MLP of 1024 and 512 hidden units in 512-dimensional query aware entity representations $\\lbrace \\mathbf {\\hat{x}}_i\\rbrace _{i=1}^N \\in \\mathbb {R}^{512}$ .", + "All transformations $f_*$ in R-GCN-layers are affine and they do maintain the input and output dimensionality of node representations the same (512-dimensional).", + "Eventually, a 2-layers MLP with [256, 128] hidden units takes the concatenation between $\\lbrace \\mathbf {h}_i^{(L)}\\rbrace _{i=1}^N$ and $\\mathbf {q}$ to predict the probability that a candidate node $v_i$ may be the answer to the query $q$ (see Equation 16 ).", + "During preliminary trials, we experimented with different numbers of R-GCN-layers (in the range 1-7). We observed that with WikiHop, for $L \\ge 3$ models reach essentially the same performance, but more layers increase the time required to train them. Besides, we observed that the gating mechanism learns to keep more and more information from the past at each layer making unnecessary to have more layers than required." + ], + [ + "We train our models with a batch size of 32 for at most 20 epochs using the Adam optimizer BIBREF29 with $\\beta _1=0.9$ , $\\beta _2=0.999$ and a learning rate of $10^{-4}$ . To help against overfitting, we employ dropout (drop rate $\\in {0, 0.1, 0.15, 0.2, 0.25}$ ) BIBREF30 and early-stopping on validation accuracy. We report the best results of each experiment based on accuracy on validation set." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1603/instruction.md b/qasper-1603/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..3a8d2ea3537c8d28765e5507562fbe09851670c4 --- /dev/null +++ b/qasper-1603/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Local Contextual Attention with Hierarchical Structure for Dialogue Act Recognition + +Question: What is dialogue act recognition? \ No newline at end of file diff --git a/qasper-1650/instruction.md b/qasper-1650/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..f5981ca44fa29b9544a55a7c2524fb8ba1048da8 --- /dev/null +++ b/qasper-1650/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Word Embeddings to Enhance Twitter Gang Member Profile Identification + +Question: How in YouTube content translated into a vector format? \ No newline at end of file diff --git a/qasper-1657/instruction.md b/qasper-1657/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0941d3410513e0d5d04b66585a9353f84c55e04c --- /dev/null +++ b/qasper-1657/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Synonym Discovery with Etymology-based Word Embeddings + +Question: Have the authors tried this approach on other languages? \ No newline at end of file diff --git a/qasper-1658/instruction.md b/qasper-1658/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..056b9afb48e523b3edeaa9d7a9650d2068898ca5 --- /dev/null +++ b/qasper-1658/instruction.md @@ -0,0 +1,69 @@ +Name of Paper: Luminoso at SemEval-2018 Task 10: Distinguishing Attributes Using Text Corpora and Relational Knowledge + +Question: What features did they train on? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Features", + "The Relational Inference Model", + "The Overfitting-Resistant Classifier", + "Other experiments", + "Results", + "Ablation Analysis", + "Reproducing These Results" + ], + "paragraphs": [ + [ + "Word embeddings are most effective when they learn from both unstructured text and a graph of general knowledge BIBREF0 . ConceptNet 5 BIBREF1 is an open-data knowledge graph that is well suited for this purpose. It is accompanied by a pre-built word embedding model known as ConceptNet Numberbatch, which combines skip-gram embeddings learned from unstructured text with the relational knowledge in ConceptNet.", + "A straightforward application of the ConceptNet Numberbatch embeddings took first place in SemEval 2017 task 2, on semantic word similarity. For SemEval 2018, we built a system with these embeddings as a major component for a slightly more complex task.", + "The Capturing Discriminative Attributes task BIBREF2 emphasizes the ability of a semantic model to recognize relevant differences between terms, not just their similarities. As the task description states, \u201cIf you can tell that americano is similar to capuccino and espresso but you can't tell the difference between them, you don't know what americano is.\u201d", + "The ConceptNet Numberbatch embeddings only measure the similarity of terms, and we hypothesized that we would need to represent more specific relationships. For example, the input triple \u201cfrog, snail, legs\u201d asks us to determine whether \u201clegs\u201d is an attribute that distinguishes \u201cfrog\u201d from \u201csnail\u201d. The answer is yes, because a frog has legs while a snail does not. The has relationship is one example of a specific relationship that is represented in ConceptNet.", + "To capture this kind of specific relationship, we built a model that infers relations between ConceptNet nodes, trained on the existing edges in ConceptNet and random negative examples. There are many models designed for this purpose; the one we decided on is based on Semantic Matching Energy (SME) BIBREF3 .", + "Our features consisted of direct similarity over ConceptNet Numberbatch embeddings, the relationships inferred over ConceptNet by SME, features that compose ConceptNet with other resources (WordNet and Wikipedia), and a purely corpus-based feature that looks up two-word phrases in the Google Books dataset.", + "We combined these features based on ConceptNet with features extracted from a few other resources in a LinearSVC classifier, using liblinear BIBREF4 via scikit-learn BIBREF5 . The classifier used only 15 features, of which 12 ended up with non-zero weights, from the five sources described. We aimed to avoid complexity in the classifier in order to prevent overfitting to the validation set; the power of the classifier should be in its features.", + "The classifier produced by this design (submitted late to the contest leaderboard) successfully avoided overfitting. It performed better on the test set than on the validation set, with a test INLINEFORM0 score of 0.7368, whose margin of error overlaps with the evaluation's reported high score of 0.75.", + "At evaluation time, we accidentally submitted our results on the validation data, instead of the test data, to the SemEval leaderboard. Our code had truncated the results to the length of the test data, causing us to not notice the mismatch. This erroneous submission got a very low score, of course. This paper presents the corrected test results, which we submitted to the post-evaluation CodaLab leaderboard immediately after the results appeared. We did not change the classifier or data; the change was a one-line change to our code for outputting the classifier's predictions on the test set instead on the validation set." + ], + [ + "In detail, these are the five sources of features we used:" + ], + [ + "To infer truth values for ConceptNet relations, we use a variant of the Semantic Matching Energy model BIBREF3 , adapted to work well on ConceptNet's vocabulary of relations. Instead of embedding relations in the same space as the terms, this model assigns new 10-dimensional embeddings to ConceptNet relations, yielding a compact model for ConceptNet's relatively small set of relations.", + "The model is trained to distinguish positive examples of ConceptNet edges from negative ones. The positive examples are edges directly contained in ConceptNet, or those that are entailed by changing the relation to a more general one or switching the directionality of a symmetric relation. The negative examples come from replacing one of the terms with a random other term, the relation with a random unentailed relation, or switching the directionality of an asymmetric relation.", + "We trained this model for approximately 3 million iterations (about 4 days of computation on an nVidia Titan Xp) using PyTorch BIBREF9 . The code of the model is available at https://github.com/LuminosoInsight/conceptnet-sme.", + "To extract features for the discriminative attribute task, we focus on a subset of ConceptNet relations that would plausibly be used as attributes: RelatedTo, IsA, HasA, PartOf, CapableOf, UsedFor, HasContext, HasProperty, and AtLocation.", + "For most of these relations, the first argument is the term, and the second argument is the attribute. We use two additional features for PartOf and AtLocation with their arguments swapped, so that the attribute is the first argument. The generic relation RelatedTo, unlike the others, is intended to be symmetric, so we add its value to the value of its swapped version and use it as a single feature." + ], + [ + "The classifier that we use to make a decision based on these features is scikit-learn's LinearSVC, using the default parameters in scikit-learn 0.19.1. (In Section SECREF4 , we discuss other models and parameters that we tried.) This classifier makes effective use of the features while being simple enough to avoid some amount of overfitting.", + "One aspect of the classifier that made a noticeable difference was the scaling of the features. We tried INLINEFORM0 and INLINEFORM1 -normalizing the columns of the input matrix, representing the values of each feature, and decided on INLINEFORM2 normalization.", + "We took advantage of the design of our features and the asymmetry of the task as a way to further mitigate overfitting. All of the features were designed to identify a property that INLINEFORM0 has and INLINEFORM1 does not, as is the case for the discriminative examples, so they should all make a non-negative contribution to a feature being discriminative. We can inspect the coefficients of the features in the SVC's decision boundary. If any feature gets a negative weight, it is likely a spurious result from overfitting to the training data. So, after training the classifier, we clip the coefficients of the decision boundary, setting all negative coefficients to zero.", + "If we were to remove these features and re-train, or require non-negative coefficients as a constraint on the classifier, then other features would inherently become responsible for overfitting. By neutralizing the features after training, we keep the features that are working well as they are, and remove a part of the model that appears to purely represent overfitting. Indeed, clipping the negative coefficients in this way increased our performance on the validation set.", + "Table TABREF8 shows the coeffcients assigned to each feature based on the training data." + ], + [ + "There are other features that we tried and later discarded. We experimented with a feature similar to the Google Books 2-grams feature, based on the AOL query logs dataset BIBREF10 . It did not add to the performance, most likely because any information it could provide was also provided by Google Books 2-grams. Similiarly, we tried extending the Google Books 2-grams data to include the first and third words of a selection of 3-grams, but this, too, appeared redundant with the 2-grams.", + "We also experimented with a feature based on bounding box annotations available in the OpenImages dataset BIBREF11 . We hoped it would help us capture attributes such as colors, materials, and shapes. While this feature did not improve the classifier's performance on the validation set, it did slightly improve the performance on the test set.", + "Before deciding on scikit-learn's LinearSVC, we experimented with a number of other classifiers. This included random forests, differentiable models made of multiple ReLU and sigmoid layers, and SVM with an RBF kernel or a polynomial kernel.", + "We also experimented with different parameters to LinearSVC, such as changing the default value of the penalty parameter INLINEFORM0 of the error term, changing the penalty from INLINEFORM1 to INLINEFORM2 , solving the primal optimization problem instead of the dual problem, and changing the loss from squared hinge to hinge. These changes either led to lower performance or had no significant effect, so in the end we used LinearSVC with the default parameters for scikit-learn version 0.19.1." + ], + [ + "When trained on the training set, the classifier we describe achieved an INLINEFORM0 score of 0.7617 on the training set, 0.7281 on the validation set, and 0.7368 on the test set. Table TABREF9 shows these scores along with their standard error of the mean, supposing that these data sets were randomly sampled from larger sets." + ], + [ + "We performed an ablation analysis to see what the contribution of each of our five sources of features was. We evaluated classifiers that used all non-empty subsets of these sources. Figure FIGREF11 plots the results of these 31 classifiers when evaluated on the validation set and the test set.", + "It is likely that the classifier with all five sources (ABCDE) performed the best overall. It is in a statistical tie ( INLINEFORM0 ) with ABDE, the classifier that omits Wikipedia as a source.", + "Most of the classifiers perfomed better on the test set than on the validation set, as shown by the dotted line. Some simple classifiers with very few features performed particularly well on the test set. One surprisingly high-performing classifier was A (ConceptNet vector similarity), which gets a test INLINEFORM0 score of 0.7355 INLINEFORM1 0.0091. This is simple enough to be called a heuristic instead of a classifier, and we can express it in closed form. It is equivalent to this expression over ConceptNet Numberbatch embeddings: INLINEFORM2 ", + "where INLINEFORM0 .", + "It is interesting to note that source A (ConceptNet vector similarity) appears to dominate source B (ConceptNet SME) on the test data. SME led to improvements on the validation set, but on the test set, any classifier containing AB performs equal to or worse than the same classifier with B removed. This may indicate that the SME features were the most prone to overfitting, or that the validation set generally required making more difficult distinctions than the test set." + ], + [ + "The code for our classifier is available on GitHub at https://github.com/LuminosoInsight/semeval-discriminatt, and its input data is downloadable from https://zenodo.org/record/1183358." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1667/instruction.md b/qasper-1667/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..9996b14932e8d61605e734f818311161f779efe6 --- /dev/null +++ b/qasper-1667/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Simultaneous Neural Machine Translation using Connectionist Temporal Classification + +Question: Do they compare simultaneous translation performance to regular machine translation? \ No newline at end of file diff --git a/qasper-1668/instruction.md b/qasper-1668/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..a09927fbc794be44674329eac4eadb48d4f894ab --- /dev/null +++ b/qasper-1668/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Simultaneous Neural Machine Translation using Connectionist Temporal Classification + +Question: Which metrics do they use to evaluate simultaneous translation? \ No newline at end of file diff --git a/qasper-1693/instruction.md b/qasper-1693/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..2f60250d00ce2b989aedba45aa26b74a19b5a25f --- /dev/null +++ b/qasper-1693/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Facet-Aware Evaluation for Extractive Text Summarization + +Question: What is the problem with existing metrics that they are trying to address? \ No newline at end of file diff --git a/qasper-1802/instruction.md b/qasper-1802/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..e39cd17fba3ab5b24c498a3b90efeebc7dda4c01 --- /dev/null +++ b/qasper-1802/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Vietnamese Open Information Extraction + +Question: Do they use an NER system in their pipeline? \ No newline at end of file diff --git a/qasper-1835/instruction.md b/qasper-1835/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..5d9127353f8abf5cfab3b9929de1da61c74b3f3e --- /dev/null +++ b/qasper-1835/instruction.md @@ -0,0 +1,57 @@ +Name of Paper: A Lightweight Front-end Tool for Interactive Entity Population + +Question: What programming language is the tool written in? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "LUWAK: A lightweight tool for interactive entity population", + "Implementation", + "LUWAK Dashboard", + "Entity Candidate Generation", + "Example: Housing Equipment Entity Population", + "Related Work and Discussion", + "Summary" + ], + "paragraphs": [ + [ + "Entity extraction is one of the most major NLP components. Most NLP tools (e.g., NLTK, Stanford CoreNLP, etc.), including commercial services (e.g., Google Cloud API, Alchemy API, etc.), provide entity extraction functions to recognize named entities (e.g., PERSON, LOCATION, ORGANIZATION, etc.) from texts. Some studies have defined fine-grained entity types and developed extraction methods BIBREF0 based on these types. However, these methods cannot comprehensively cover domain-specific entities. For instance, a real estate search engine needs housing equipment names to index these terms for providing fine-grained search conditions. There is a significant demand for constructing user-specific entity dictionaries, such as the case of cuisine and ingredient names for restaurant services. A straightforward solution is to prepare a set of these entity names as a domain-specific dictionary. Therefore, this paper focuses on the entity population task, which is a task of collecting entities that belong to an entity type required by a user.", + "We develop LUWAK, a lightweight tool for effective interactive entity population. The key features are four-fold:", + "We think these features are key components for effective interactive entity population.", + "We choose an interactive user feedback strategy for entity population for LUWAK. A major approach to entity population is bootstrapping, which uses several entities that have been prepared as a seed set for finding new entities. Then, these new entities are integrated into the initial seed set to create a new seed set. The bootstrapping approach usually repeats the procedure until it has collected a sufficient number of entities. The framework cannot prevent the incorporation of incorrect entities that do not belong to the entity type unless user interaction between iterations. The problem is commonly called semantic drift BIBREF1 . Therefore, we consider user interaction, in which feedback is given to expanded candidates, as essential to maintaining the quality of an entity set. LUWAK implements fundamental functions for entity population, including (a) importing an initial entity set, (b) generating entity candidates, (c) obtaining user feedback, and (d) publishing populated entity dictionary.", + "We aim to reduce the user\u2019s total workload as a key metric of an entity population tool. That is, an entity population tool should provide the easiest and fastest solution to collecting entities of a particular entity type. User interaction cost is a dominant factor in the entire workload of an interactive tool. Thus, we carefully design the user interface for users to give feedbacks to the tool intuitively. Furthermore, we also consider the end-to-end user cost reduction. We adhere to the concept of developing installation-free software to distribute the tool among a wide variety of users, including nontechnical clusters. This lightweight design of LUWAK might speed up the procedure of the whole interactive entity population workflow. Furthermore, this advantage might be beneficial to continuously improve the whole pipeline of interactive entity population system." + ], + [ + "Our framework adopts the interactive entity expansion approach. This approach organizes the collaboration of a human worker and entity expansion algorithms to generate a user-specific entity dictionary efficiently. We show the basic workflow of LUWAK in Figure 1 . (Step 1) LUWAK assumes that a user prepares an initial seed set manually. The seed set is shown in the Entity table. (Step 2) A user can send entities in the Entity table to an Expansion API for obtaining entity candidates. (Step 3) LUWAK shows the entity candidates in the Candidate table for user interaction. Then, the user checks accept/reject buttons to update the Entity table. After submitting the judgments, LUWAK shows the Entity table again. The user can directly add, edit, or delete entities in the table at any time. (Step 4) the user can also easily see how these entities stored in the Entity table appear in a document. (Step 5) After repeating the same procedure (Steps 2\u20134) for a sufficient time, the user can publish the Entity table as an output." + ], + [ + "LUWAK is implemented in pure JavaScript code, and it uses the LocalStorage of a web browser. A user does not have to install any packages except for a web browser. The only thing the user must do is to download the LUWAK software and put it in a local directory. We believe the cost of installing the tool will keep away a good amount of potential users. This philosophy follows our empirical feeling of the curse of installing required packages/libraries. Moreover, LUWAK does not need users to consider the usage and maintenance of these additional packages. That is why we are deeply committed to making LUWAK a pure client-side tool in the off-the-shelf style." + ], + [ + "LUWAK has a dashboard for quickly viewing an entity dictionary in progress. The dashboard consists of two tables: the Entity table and the Feedback table. The Entity table provides efficient ways to construct and modify an entity dictionary. Figure UID11 shows the screenshot of the Entity table. The table shows entities in the current entity set. Each row corresponds to an entity entry. Each entry has a label, which denotes whether the predefined entity type is a positive or a negative example, an original entity, which was used to find the entity, and the score, which denotes the confidence score. A user can directly edit the table by adding, renaming, and deleting entities. Moreover, the entity inactivation function allows a user to manually inactivate entities, so that entity expansion algorithms do not use the inactivated entities. The table implements a page switching function, a search function, and a sorting function to ensure visibility even when there is a large number of entities in the table." + ], + [ + "We design the entity candidate generation module as an external API (Expansion API). The Expansion API receives a set of entities with positive labels. The Expansion API returns top- $k$ entity candidates.", + "As an initial implementation, we used GloVe BIBREF2 as word embedding models for implementing an Expansion API. This API calculates the cosine similarity between a set of positive entities and entities candidates to generate a ranked list. We prepared models trained based on the CommonCrawl corpus and the Twitter corpus. Note that the specification of the expansion algorithm is not limited to the algorithm described in this paper, as LUWAK considers the Expansion API as an external function.", + "Moreover, we also utilize the category-based expansion module, in which we used is-a relationship between the ontological category and each entity and expanded seeds via category-level. For example, if most of the entities already inserted in the dictionary share the same category, such as Programming Languages, the system suggests that \"Programming Language\" entities should be inserted in the dictionary when we develop a job skill name dictionary. Category-based entity expansion is helpful to avoid the candidate entity one by one. We used Yago BIBREF3 as an existing knowledge base.", + "External API. In our design of LUWAK, Expansion APIs are placed as an external function outside LUWAK. There are three reasons why we adopt this design. First, we want LUWAK to remain a corpus-free tool. Users do not have to download any corpora or models to start using LUWAK, and it takes too much time to launch an Expansion API server. Second, LUWAK\u2019s design allows external contributors to build their own expansion APIs that are compatible with LUWAK\u2019s interface. We developed the initial version of the LUWAK package to contain an entity Expansion API so users can launch their expansion APIs internally. Third, the separation between LUWAK and the Expansion APIs enables Expansion APIs to use predetermined options for algorithms, including non-embedding-based methods (e.g., pattern-based methods). We can use more than one entity expansion model to find related entities. For instance, general embedding models, such as those built on Wikipedia, might be a good choice in early iterations, whereas more domain-specific models trained on domain-specific corpora might be helpful in later iterations. LUWAK is flexible to change and use more than one Expansion API. This design encourages us to continuously refine the entity expansion module easily." + ], + [ + "We show an example of populating house equipment entities using LUWAK for improving a real estate search engine. The preliminary step is to prepare seed entities that belong to the real estate house equipment entity type (e.g., kitchen, bath). In this case, a user is supposed to provide several entities ( $\\sim $ 10) as an initial set of the category. LUWAK first asks the user to upload an initial seed set. The user can add, rename, and delete entities on the Entity table as he or she wants. The user can also choose a set of entity expansion models at any time. Figure 2 shows the entity dashboard in this example.", + "When the user submits the current entity set by clicking the Expand Seed Set button (Figure UID11 ), LUWAK sends a request to the external Expansion APIs that are selected to obtain expanded entities. The returned values will be stored in the Feedback table, as Figure UID12 shows. The Feedback table provides a function to capture user feedback intuitively. The user can click the + or - buttons to assign positive or negative labels to the entity candidates. The score column stores the similarity score, which is calculated by the Expansion API as reference information for users. The user can also see how these entities are generated by looking at the original entities in the original column. The original entity information can be used to detect semantic drift. For instance, if the user finds the original entity of some entity candidates has negative labels, the user might consider inactivating the entity to prevent semantic drift.", + "In the next step, the user reflects the feedback by clicking the Submit Feedback button. Then, the user will see the entity dashboard with the newly added entities as shown in Figure UID13 . The user can inactivate the entity by clicking the inactivate button. The user can sort rows by column values to take a brief look at the current entity set. Also, the entity dashboard provides a search function to find an entity for action. The user can also check how entities appear in a test document. As shown in Figure UID14 , LUWAK highlights these entities in the current entity set. After the user is satisfied with the amount of the current entity set in the table, the Export button allows the user to download the entire table, including inactivated entities." + ], + [ + "Entity population is one of the important practical problems in NLP. Generated entity dictionaries can be used in various applications, including search engines, named entity extraction, and entity linking. Iterative seed expansion is known to be an efficient approach to construct user-specific entity dictionaries. Previous studies have aimed to construct a high-quality entity dictionary from a small number of seed entities BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 . As we stated in \"Entity Candidate Generation\" , LUWAK is flexible with the types of algorithms used for entity population. A user can select any combinations of different methods once the Expansion API of the methods are available.", + "Stanford Pattern-based Information Extraction and Diagnostics (SPIED) BIBREF8 is a pattern-based entity population system. SPIED requires not only an initial seed set but also document collection because it uses the pattern-based approach. After a user inputs initial seed entities, SPIED generates regular expression patterns to find entity candidates from a given document collection. This approach incurs a huge computational cost for calculating the scores of every regular expression pattern and every entity candidate in each iteration. Furthermore, SPIED adopts a bootstrapping approach, which does not involve user feedback for each iteration. This approach can easily result in semantic drift.", + "Interactive Knowledge Extraction BIBREF9 (IKE) is an interactive bootstrapping tool for collecting relation-extraction patterns. IKE also provides a search-based entity extraction function and an embedding-based entity expansion function for entity population. A user can interactively add entity candidates generated by an embedding-based algorithm to an entity dictionary. LUWAK is a more lightweight tool than IKE, which only focuses on the entity population task. LUWAK has numerous features, such as the multiple entity expansion model choices, that are not implemented in IKE. Moreover, LUWAK is a corpus-free tool that does not require a document collection for entity population. Thus, we differentiate LUWAK from IKE, considering it a more lightweight entity population tool." + ], + [ + "This paper has presented LUWAK, a lightweight front-end tool for interactive entity population. LUWAK provides a set of basic functions such as entity expansion and user feedback assignment. We have implemented LUWAK in pure JavaScript with LocalStorage to make it an installation-free tool. We believe that LUWAK plays an important role in delivering the values of existing entity expansion techniques to potential users including nontechnical people without supposing a large amount of human cost. Moreover, we believe that this design makes it easy to compare performances between interactive entity population pipelines and develop more sophisticated ones." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1856/instruction.md b/qasper-1856/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..0f46d666854173933b2fc7d269fe85b6c2b57962 --- /dev/null +++ b/qasper-1856/instruction.md @@ -0,0 +1,128 @@ +Name of Paper: DP-LSTM: Differential Privacy-inspired LSTM for Stock Prediction Using Financial News + +Question: How does the differential privacy mechanism work? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Problem Statement", + "Problem Statement ::: ARMA Model", + "Problem Statement ::: Sentiment Analysis", + "Problem Statement ::: Sentiment-ARMA Model and Loss Function", + "Problem Statement ::: Overview of LSTM", + "Problem Statement ::: Definition of Differential Privacy", + "Training DP-LSTM Neural Network", + "Training DP-LSTM Neural Network ::: Data Preprocessing and Normalization ::: Data Preprocessing", + "Training DP-LSTM Neural Network ::: Data Preprocessing and Normalization ::: Normalization", + "Training DP-LSTM Neural Network ::: Adding Noise", + "Training DP-LSTM Neural Network ::: Training Setting", + "Performance Evaluation", + "Conclusion", + "References" + ], + "paragraphs": [ + [ + "Stock prediction is crucial for quantitative analysts and investment companies. Stocks' trends, however, are affected by a lot of factors such as interest rates, inflation rates and financial news [12]. To predict stock prices accurately, one must use these variable information. In particular, in the banking industry and financial services, analysts' armies are dedicated to pouring over, analyzing, and attempting to quantify qualitative data from news. A large amount of stock trend information is extracted from the large amount of text and quantitative information that is involved in the analysis.", + "Investors may judge on the basis of technical analysis, such as charts of a company, market indices, and on textual information such as news blogs or newspapers. It is however difficult for investors to analyze and predict market trends according to all of these information [22]. A lot of artificial intelligence approaches have been investigated to automatically predict those trends [3]. For instance, investment simulation analysis with artificial markets or stock trend analysis with lexical cohesion based metric of financial news' sentiment polarity. Quantitative analysis today is heavily dependent on data. However, the majority of such data is unstructured text that comes from sources like financial news articles. The challenge is not only the amount of data that are involved, but also the kind of language that is used in them to express sentiments, which means emoticons. Sifting through huge volumes of this text data is difficult as well as time-consuming. It also requires a great deal of resources and expertise to analyze all of that [4].", + "To solve the above problem, in this paper we use sentiment analysis to extract information from textual information. Sentiment analysis is the automated process of understanding an opinion about a given subject from news articles [5]. The analyzed data quantifies reactions or sentiments of the general public toward people, ideas or certain products and reveal the information's contextual polarity. Sentiment analysis allows us to understand if newspapers are talking positively or negatively about the financial market, get key insights about the stock's future trend market.", + "We use valence aware dictionary and sentiment reasoner (VADER) to extract sentiment scores. VADER is a lexicon and rule-based sentiment analysis tool attuned to sentiments that are expressed in social media specifically [6]. VADER has been found to be quite successful when dealing with NY Times editorials and social media texts. This is because VADER not only tells about the negativity score and positively but also tells us about how positive or negative a sentiment is.", + "However, news reports are not all objective. We may increase bias because of some non-objective reports, if we rely on the information that is extracted from the news for prediction fully. Therefore, in order to enhance the prediction model's robustness, we will adopt differential privacy (DP) method. DP is a system for sharing information about a dataset publicly by describing groups' patterns within the dataset while withholding information about individuals in the dataset. DP can be achieved if the we are willing to add random noise to the result. For example, rather than simply reporting the sum, we can inject noise from a Laplace or gaussian distribution, producing a result that\u2019s not quite exact, that masks the contents of any given row.", + "In the last several years a promising approach to private data analysis has emerged, based on DP, which ensures that an analysis outcome is \"roughly as likely\" to occur independent of whether any individual opts in to, or to opts out of, the database. In consequence, any one individual's specific data can never greatly affect the results. General techniques for ensuring DP have now been proposed, and a lot of datamining tasks can be carried out in a DP method, frequently with very accurate results [21]. We proposed a DP-LSTM neural network, which increase the accuracy of prediction and robustness of model at the same time.", + "The remainder of the paper is organized as follows. In Section 2, we introduce stock price model, the sentiment analysis and differential privacy method. In Section 3, we develop the different privacy-inspired LSTM (DP-LSTM) deep neural network and present the training details. Prediction results are provided in Section 4. Section 5 concludes the paper." + ], + [ + "In this section, we first introduce the background of the stock price model, which is based on the autoregressive moving average (ARMA) model. Then, we present the sentiment analysis details of the financial news and introduce how to use them to improve prediction performance. At last, we introduce the differential privacy framework and the loss function." + ], + [ + "The ARMA model, which is one of the most widely used linear models in time series prediction [17], where the future value is assumed as a linear combination of the past errors and past values. ARMA is used to set the stock midterm prediction problem up. Let ${X}_t^\\text{A}$ be the variable based on ARMA at time $t$, then we have", + "where $X_{t-i}$ denotes the past value at time $t-i$; $\\epsilon _{t}$ denotes the random error at time $t$; $\\phi _i$ and $\\psi _j$ are the coefficients; $\\mu $ is a constant; $p$ and $q$ are integers that are often referred to as autoregressive and moving average polynomials, respectively." + ], + [ + "Another variable highly related to stock price is the textual information from news, whose changes may be a precursor to price changes. In our paper, news refers to a news article's title on a given trading day. It has been used to infer whether an event had informational content and whether investors' interpretations of the information were positive, negative or neutral. We hence use sentiment analysis to identify and extract opinions within a given text. Sentiment analysis aims at gauging the attitude, sentiments, evaluations and emotions of a speaker or writer based on subjectivity's computational treatment in a text [19]-[20].", + "Figure FIGREF3 shows an example of the sentiment analysis results obtained from financial news titles that were based on VADER. VADER uses a combination of a sentiment lexicon which are generally labelled according to their semantic orientation as either negative or positive. VADER has been found to be quite successful when dealing with news reviews. It is fully open-sourced under the MIT License. The result of VADER represent as sentiment scores, which include the positive, negative and neutral scores represent the proportion of text that falls in these categories. This means all these three scores should add up to 1. Besides, the Compound score is a metric that calculates the sum of all the lexicon ratings which have been normalized between -1(most extreme negative) and +1 (most extreme positive). Figure FIGREF5 shows the positive and negative wordcloud, which is an intuitive analysis of the number of words in the news titles." + ], + [ + "To take the sentiment analysis results of the financial news into account, we introduce the sentiment-ARMA model as follows", + "where $\\alpha $ and $\\lambda $ are weighting factors; $c$ is a constant; and $f_2(\\cdot )$ is similar to $f_1(\\cdot )$ in the ARMA model (DISPLAY_FORM2) and is used to describe the prediction problem.", + "In this paper, the LSTM neural network is used to predict the stock price, the input data is the previous stock price and the sentiment analysis results. Hence, the sentiment based LSTM neural network (named sentiment-LSTM) is aimed to minimize the following loss function:", + "where $T$ denotes the number of prediction time slots, i.e., $t = 1,...,p$ are the observations (training input data), $t = p+1,...,p+T$ are the predicts (training output data); and $\\hat{X}_t$ is given in (DISPLAY_FORM7)." + ], + [ + "Denote $\\mathcal {X}_t^{\\text{train}} = \\lbrace X_{t-i},S_{t-i}\\rbrace _{i=1}^p$ as the training input data. Figure FIGREF10 shows the LSTM's structure network, which comprises one or more hidden layers, an output layer and an input layer [16]. LSTM networks' main advantage is that the hidden layer comprises memory cells. Each memory cell recurrently has a core self-connected linear unit called \u201c Constant Error Carousel (CEC)\u201d [13], which provides short-term memory storage and has three gates:", + "Input gate, which controls the information from a new input to the memory cell, is given by", + "where $h_{t-1}$ is the hidden state at the time step $t-1$; $i_t$ is the output of the input gate layer at the time step $t$; $\\hat{c}_t$ is the candidate value to be added to the output at the time step $t$; $b_i$ and $b_c$ are biases of the input gate layer and the candidate value computation, respectively; $W_i$ and $W_c$ are weights of the input gate and the candidate value computation, respectively; and $\\sigma (x) = 1/(1+e^{-x})$ is the pointwise nonlinear activation function.", + "Forget gate, which controls the limit up to which a value is saved in the memory, is given by", + "where $f_t$ is the forget state at the time step $t$, $W_f$ is the weight of the forget gate; and $b_f$ is the bias of the forget gate.", + "Output gate, which controls the information output from the memory cell, is given by", + "where new cell states $c_t$ are calculated based on the results of the previous two steps; $o_t$ is the output at the time step $t$; $W_o$ is the weight of the output gate; and $b_o$ is the bias of the output gate [14]." + ], + [ + "Differential privacy is one of privacy's most popular definitions today, which is a system for publicly sharing information about a dataset by describing the patterns of groups within the dataset while withholding information about individuals in the dataset. It intuitively requires that the mechanism that outputs information about an underlying dataset is robust to one sample's any change, thus protecting privacy. A mechanism ${f}$ is a random function that takes a dataset $\\mathcal {N}$ as input, and outputs a random variable ${f}(\\mathcal {N})$. For example, suppose $\\mathcal {N}$ is a news articles dataset, then the function that outputs compound score of articles in $\\mathcal {N}$ plus noise from the standard normal distribution is a mechanism [7].", + "Although differential privacy was originally developed to facilitate secure analysis over sensitive data, it can also enhance the robustness of the data. Note that finance data, especially news data and stock data, is unstable with a lot of noise, with a more robust data the accuracy of prediction will be improved. Since we predict stock price by fusing news come from different sources, which might include fake news. Involving differential privacy in the training to improve the robustness of the finance news is meaningful." + ], + [ + "It is known that it is risky to predict stocks by considering news factors, because news can't guarantee full notarization and objectivity, many times extreme news will have a big impact on prediction models. To solve this problem, we consider entering the idea of the differential privacy when training. In this section, our DP-LSTM deep neural network training strategy is presented. The input data consists of three components: stock price, sentiment analysis compound score and noise." + ], + [ + "The data for this project are two parts, the first part is the historical S&P 500 component stocks, which are downloaded from the Yahoo Finance. We use the data over the period of from 12/07/2017 to 06/01/2018. The second part is the news article from financial domain are collected with the same time period as stock data. Since our paper illustrates the relationship between the sentiment of the news articles and stocks' price. Hence, only news article from financial domain are collected. The data is mainly taken from Webhose archived data, which consists of 306242 news articles present in JSON format, dating from December 2017 up to end of June 2018. The former 85% of the dataset is used as the training data and the remainder 15% is used as the testing data. The News publishers for this data are CNBC.com, Reuters.com, WSJ.com, Fortune.com. The Wall Street Journal is one of the largest newspapers in the United States, which coverage of breaking news and current headlines from the US and around the world include top stories, photos, videos, detailed analysis and in-depth thoughts; CNBC primarily carries business day coverage of U.S. and international financial markets, which following the end of the business day and on non-trading days; Fortune is an American multinational business magazine; Reuters is an international news organization. We preprocess the raw article body and use NLTK sentiment package alence Aware Dictionary and Sentiment Reasoner (VADER) to extract sentiment scores.", + "The stocks with missing data are deleted, and the dataset we used eventually contains 451 stocks and 4 news resources (CNBC.com, Reuters.com, WSJ.comFortune.com.). Each stock records the adjust close price and news compound scores of 121 trading days.", + "A rolling window with size 10 is used to separate data, that is, We predict the stock price of the next trading day based on historical data from the previous 10 days, hence resulting in a point-by-point prediction [15]. In particular, the training window is initialized with all real training data. Then we shift the window and add the next real point to the last point of training window to predict the next point and so forth. Then, according to the length of the window, the training data is divided into 92 sets of training input data (each set length 10) and training output data (each set length 1). The testing data is divided into input and output data of 9 windows (see Figure FIGREF20)." + ], + [ + "To detect stock price pattern, it is necessary to normalize the stock price data. Since the LSTM neural network requires the stock patterns during training, we use \u201cmin-max\u201d normalization method to reform dataset, which keeps the pattern of the data [11], as follow:", + "where $X_{t}^{n}$ denotes the data after normalization. Accordingly, de-normalization is required at the end of the prediction process to get the original price, which is given by", + "where $\\hat{X}_{t}^{n}$ denotes the predicted data and $\\hat{X}_{t}$ denotes the predicted data after de-normalization.", + "Note that compound score is not normalized, since the compound score range from -1 to 1, which means all the compound score data has the same scale, so it is not require the normalization processing." + ], + [ + "We consider the differential privacy as a method to improve the robustness of the LSTM predictions [8]. We explore the interplay between machine learning and differential privacy, and found that differential privacy has several properties that make it particularly useful in application such as robustness to extract textual information [9]. The robustness of textual information means that accuracy is guaranteed to be unaffected by certain false information [10].", + "The input data of the model has 5 dimensions, which are the stock price and four compound scores as $(X^t, S_1^t, S_2^t, S_3^t, S_4^t), t=1,...,T$, where $X^t$ represents the stock price and $S_i^t,~i=1,...,4$ respectively denote the mean compound score calculated from WSJ, CNBC, Fortune and Reuters. According to the process of differential privacy, we add Gaussian noise with different variances to the news according to the variance of the news, i.e., the news compound score after adding noise is given by", + "where $\\text{var}(\\cdot )$ is the variance operator, $\\lambda $ is a weighting factor and $\\mathcal {N}(\\cdot )$ denotes the random Gaussian process with zero mean and variance $\\lambda \\text{var}(S_i)$.", + "We used python to crawl the news from the four sources of each trading day, perform sentiment analysis on the title of the news, and get the compound score. After splitting the data into training sets and test sets, we separately add noise to each of four news sources of the training set, then, for $n$-th stock, four sets of noise-added data $(X^n_t, {\\widetilde{S}^t_1}, S^t_2, S^t_3, S^t_4)$, $(X^n_t, {S^t_1}, \\widetilde{S}^t_2, S^t_3, S^t_4)$, $(X^n_t, { S^t_1}, S^t_2, \\widetilde{S}^t_3, S^t_4)$, $(X^n_t, { S^t_1}, S^t_2, S^t_3, \\widetilde{S}^t_4)$ are combined into a new training data through a rolling window. The stock price is then combined with the new compound score training data as input data for our DP-LSTM neural network." + ], + [ + "The LSTM model in figure FIGREF10 has six layers, followed by an LSTM layer, a dropout layer, an LSTM layer, an LSTM layer, a dropout layer, a dense layer, respectively. The dropout layers (with dropout rate 0.2) prevent the network from overfitting. The dense layer is used to reshape the output. Since a network will be difficult to train if it contains a large number of LSTM layers [16], we use three LSTM layers here.", + "In each LSTM layer, the loss function is the mean square error (MSE), which is the sum of the squared distances between our target variable and the predicted value. In addition, the ADAM [17] is used as optimizer, since it is straightforward to implement, computationally efficient and well suited for problems with large data set and parameters.", + "There are many methods and algorithms to implement sentiment analysis systems. In this paper, we use rule-based systems that perform sentiment analysis based on a set of manually crafted rules. Usually, rule-based approaches define a set of rules in some kind of scripting language that identify subjectivity, polarity, or the subject of an opinion. We use VADER, a simple rule-based model for general sentiment analysis." + ], + [ + "In this section, we validate our DP-LSTM based on the S&P 500 stocks. We calculate the mean prediction accuracy (MPA) to evaluate the proposed methods, which is defined as", + "where $X_{t,\\ell }$ is the real stock price of the $\\ell $-th stock on the $t$-th day, $L$ is the number of stocks and $\\hat{X}_{t,\\ell }$ is the corresponding prediction result.", + "Figure FIGREF27 plots the average score for all news on the same day over the period. The compound score is fluctuating between -0.3 and 0.15, indicating an overall neutral to slightly negative sentiment. The Positive, Negative and Neutral scores represent the proportion of text that falls in these categories. The Compound score is a metric that calculates the sum of all the lexicon ratings which have been normalized between -1 (most extreme negative) and +1 (most extreme positive).", + "Figure FIGREF29 shows the $\\text{MPAs}$ of the proposed DP-LSTM and vanilla LSTM for comparison. In Table TABREF30, we give the mean MPA results for the prediction prices, which shows the accuracy performance of DP-LSTM is 0.32% higer than the LSTM with news. The result means the DP framework can make the prediction result more accuracy and robustness.", + "Note that the results are obtained by running many trials, since we train stocks separately and predict each price individually due to the different patterns and scales of stock prices. This in total adds up to 451 runs. The results shown in Table TABREF30 is the average of these 451 runs. Furthermore, we provide results for 9 duration over a period in Figure FIGREF29. The performance of our DP-LSTM is always better than the LSTM with news. Based on the sentiment-ARMA model and adding noise for training, the proposed DP-LSTM is more robust. The investment risk based on this prediction results is reduced.", + "In Figure FIGREF31, we can see the prediction results of DP-LSTM with is closer to the real S&P 500 index price line than other methods. The two lines (prediction results of LSTM with news and LSTM without news) almost coincide in Figure FIGREF31. We can tell the subtle differences from the Table TABREF32, that DP-LSTM is far ahead, and LSTM with news is slightly better than LSTM without news." + ], + [ + "In this paper, we integrated the deep neural network with the famous NLP models (VADER) to identify and extract opinions within a given text, combining the stock adjust close price and compound score to reduce the investment risk. We first proposed a sentiment-ARMA model to represent the stock price, which incorporates influential variables (price and news) based on the ARMA model. Then, a DP-LSTM deep neural network was proposed to predict stock price according to the sentiment-ARMA model, which combines the LSTM, compound score of news articles and differential privacy method. News are not all objective. If we rely on the information extracted from the news for prediction fully, we may increase bias because of some non-objective reports. Therefore, the DP-LSTM enhance robustness of the prediction model. Experiment results based on the S&P 500 stocks show that the proposed DP-LSTM network can predict the stock price accurately with robust performance, especially for S&P 500 index that reflects the general trend of the market. S&P 500 prediction results show that the differential privacy method can significantly improve the robustness and accuracy." + ], + [ + "[1] X. Li, Y. Li, X.-Y. Liu, D. Wang, \u201cRisk management via anomaly circumvent: mnemonic deep learning for midterm stock prediction.\u201d in Proceedings of 2nd KDD Workshop on Anomaly Detection in Finance (Anchorage \u201919), 2019.", + "[2] P. Chang, C. Fan, and C. Liu, \u201cIntegrating a piece-wise linear representation method and a neural network model for stock trading points prediction.\u201d IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews) 39, 1 (2009), 80\u201392.", + "[3] Akita, Ryo, et al. \u201cDeep learning for stock prediction using numerical and textual information.\u201d IEEE/ACIS 15th International Conference on Computer and Information Science (ICIS). IEEE, 2016.", + "[4] Li, Xiaodong, et al. \u201cDoes summarization help stock prediction? A news impact analysis.\u201d IEEE Intelligent Systems 30.3 (2015): 26-34.", + "[5] Ding, Xiao, et al. \u201cDeep learning for event-driven stock prediction.\u201d Twenty-fourth International Joint Conference on Artificial Intelligence. 2015.", + "[6] Hutto, Clayton J., and Eric Gilbert. \u201cVader: A parsimonious rule-based model for sentiment analysis of social media text.\u201d Eighth International AAAI Conference on Weblogs and Social Media, 2014.", + "[7] Ji, Zhanglong, Zachary C. Lipton, and Charles Elkan. \u201cDifferential privacy and machine learning: a survey and review.\u201d arXiv preprint arXiv:1412.7584 (2014).", + "[8] Abadi, Martin, et al. \u201cDeep learning with differential privacy.\u201d Proceedings of the 2016 ACM SIGSAC Conference on Computer and Communications Security, ACM, 2016.", + "[9] McMahan, H. Brendan, and Galen Andrew. \u201cA general approach to adding differential privacy to iterative training procedures.\u201d arXiv preprint arXiv:1812.06210 (2018).", + "[10] Lecuyer, Mathias, et al. \u201cCertified robustness to adversarial examples with differential privacy.\u201d arXiv preprint arXiv:1802.03471 (2018).", + "[11] Hafezi, Reza, Jamal Shahrabi, and Esmaeil Hadavandi. \u201cA bat-neural network multi-agent system (BNNMAS) for stock price prediction: Case study of DAX stock price.\u201d Applied Soft Computing, 29 (2015): 196-210.", + "[12] Chang, Pei-Chann, Chin-Yuan Fan, and Chen-Hao Liu. \u201cIntegrating a piecewise linear representation method and a neural network model for stock trading points prediction.\u201d IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews) 39.1 (2008): 80-92.", + "[13] Gers, Felix A., Nicol N. Schraudolph, and J\u00fcrgen Schmidhuber. \u201cLearning precise timing with LSTM recurrent networks.\u201d Journal of Machine Learning Research 3.Aug (2002): 115-143.", + "[14] Qin, Yao, et al. \u201cA dual-stage attention-based recurrent neural network for time series prediction.\u201d arXiv preprint arXiv:1704.02971 (2017).", + "[15] Malhotra, Pankaj, et al. \u201cLong short term memory networks for anomaly detection in time series.\u201d Proceedings. Presses universitaires de Louvain, 2015.", + "[16] Sak, Ha\u015fim, Andrew Senior, and Fran\u00e7oise Beaufays. \u201cLong short-term memory recurrent neural network architectures for large scale acoustic modeling.\u201d Fifteenth annual conference of the international speech communication association, 2014.", + "[17] Kingma, Diederik P., and Jimmy Ba. \u201cAdam: A method for stochastic optimization.\u201d arXiv preprint arXiv:1412.6980 (2014).", + "[18] Box, George EP, et al. Time series analysis: forecasting and control. John Wiley & Sons, 2015.", + "[19] Pang, Bo, and Lillian Lee. \u201cOpinion mining and sentiment analysis.\u201d Foundations and Trends in Information Retrieval 2.1\u20132 (2008): 1-135.", + "[20] Cambria, Erik. \u201cAffective computing and sentiment analysis.\u201d IEEE Intelligent Systems 31.2 (2016): 102-107.", + "[21] Dwork C, Lei J. Differential privacy and robust statistics//STOC. 2009, 9: 371-380.", + "[22] X. Li, Y. Li, Y. Zhan, and X.-Y. Liu. \u201cOptimistic bull or pessimistic bear: adaptive deep reinforcement learning for stock portfolio allocation.\u201d in Proceedings of the 36th International Conference on Machine Learning, 2019." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-1869/instruction.md b/qasper-1869/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..47ca101a789bbd9253ec604e1a15cbf0271b33af --- /dev/null +++ b/qasper-1869/instruction.md @@ -0,0 +1,114 @@ +Name of Paper: The Evolved Transformer + +Question: What is in the model search space? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Methods", + "Search Space", + "Seeding the Search Space with Transformer", + "Evolution with Progressive Dynamic Hurdles", + "Datasets", + "Training Details and Hyperparameters", + "Search Configurations", + "Results", + "Search Techniques", + "The Evolved Transformer: Performance and Analysis", + "Conclusion", + "Acknowledgements", + "Search Algorithms", + "Search Space Information", + "Ablation studies of the Evolved Transformer" + ], + "paragraphs": [ + [ + "Over the past few years, impressive advances have been made in the field of neural architecture search. Reinforcement learning and evolution have both proven their capacity to produce models that exceed the performance of those designed by humans BIBREF0 , BIBREF1 . These advances have mostly focused on improving image models, although some effort has also been invested in searching for sequence models BIBREF2 , BIBREF3 . In these cases, it has always been to find improved recurrent neural networks (RNNs), which were long established as the de facto neural model for sequence problems BIBREF4 , BIBREF5 .", + "However, recent works have shown that there are better alternatives to RNNs for solving sequence problems. Due to the success of convolution-based networks, such as Convolution Seq2Seq BIBREF6 , and full attention networks, such as the Transformer BIBREF7 , feed-forward networks are now a viable option for solving sequence-to-sequence (seq2seq) tasks. The main strength of feed-forward networks is that they are faster, and easier to train than RNNs.", + "The goal of this work is to examine the use of neural architecture search methods to design better feed-forward architectures for seq2seq tasks. Specifically, we apply tournament selection architecture search to evolve from the Transformer, considered to be the state-of-art and widely-used, into a better and more efficient architecture. To achieve this, we construct a search space that reflects the recent advances in feed-forward seq2seq models and develop a method called progressive dynamic hurdles (PDH) that allows us to perform our search directly on the computationally demanding WMT 2014 English-German (En-De) translation task. Our search produces a new architecture \u2013 called the Evolved Transformer (ET) \u2013 which demonstrates consistent improvement over the original Transformer on four well-established language tasks: WMT 2014 English-German, WMT 2014 English-French (En-Fr), WMT 2014 English-Czech (En-Cs) and the 1 Billion Word Language Model Benchmark (LM1B). In our experiments with big size models, the Evolved Transformer is twice as efficient as the Transformer in FLOPS without loss of quality. At a much smaller \u2013 mobile-friendly \u2013 model size of $\\sim $ 7M parameters, the Evolved Transformer outperforms the Transformer by 0.7 BLEU." + ], + [ + "RNNs have long been used as the default option for applying neural networks to sequence modeling BIBREF4 , BIBREF5 , with LSTM BIBREF8 and GRU BIBREF9 architectures being the most popular. However, recent work has shown that RNNs are not necessary to build state-of-the-art sequence models. For example, many high performance convolutional models have been designed, such as WaveNet BIBREF10 , Gated Convolution Networks BIBREF11 , Conv Seq2Seq BIBREF6 and Dynamic Lightweight Convolution model BIBREF12 . Perhaps the most promising architecture in this direction is the Transformer architecture BIBREF7 , which relies only on multi-head attention to convey spatial information. In this work, we use both convolutions and attention in our search space to leverage the strengths of both layer types.", + "The recent advances in sequential feed-forward networks are not limited to architecture design. Various methods, such as BERT BIBREF13 and Radford et. al's pre-training technique BIBREF14 , have demonstrated how models such as the Transformer can improve over RNN pre-training BIBREF15 , BIBREF16 . For translation specifically, work on scaling up batch size BIBREF17 , BIBREF12 , using relative position representations BIBREF18 , and weighting multi-head attention BIBREF19 have all pushed the state-of-the-art for WMT 2014 En-De and En-Fr. However, these methods are orthogonal to this work, as we are only concerned with improving the neural network architecture itself, and not the techniques used for improving overall performance.", + "The field of neural architecture search has also seen significant recent progress. The best performing architecture search methods are those that are computationally intensive BIBREF2 , BIBREF20 , BIBREF21 , BIBREF22 , BIBREF1 , BIBREF0 . Other methods have been developed with speed in mind, such as DARTS BIBREF23 , ENAS BIBREF3 , SMASH BIBREF24 , and SNAS BIBREF25 . These methods radically reduce the amount of time needed to run each search by approximating the performance of each candidate model, instead of investing resources to fully train and evaluate each candidate separately. Unfortunately, these faster methods produce slightly less competitive results. Zela et. al's zela18 utilization of Hyperband BIBREF26 and PNAS's BIBREF27 incorporation of a surrogate model are examples of approaches that try to both increase efficiency via candidate performance estimation and maximize search quality by training models to the end when necessary. The progressive dynamic hurdles method we introduce here is similar to these approaches in that we train our best models individually to the end, but optimize our procedure by discarding unpromising models early on." + ], + [ + "We employ evolution-based architecture search because it is simple and has been shown to be more efficient than reinforcement learning when resources are limited BIBREF0 . We use the same tournament selection BIBREF28 algorithm as Real et al. real19, with the aging regularization omitted, and so encourage the reader to view their in-depth description of the method. In the interest of saving space, we will only give a brief overview of the algorithm here.", + "Tournament selection evolutionary architecture search is conducted by first defining a gene encoding that describes a neural network architecture; we describe our encoding in the following Search Space subsection. An initial population is then created by randomly sampling from the space of gene encoding to create individuals. These individuals are assigned fitnesses based on training the the neural networks they describe on the target task and then evaluating their performance on the task's validation set. The population is then repeatedly sampled from to produce subpopulations, from which the individual with the highest fitness is selected as a parent. Selected parents have their gene encodings mutated \u2013 encoding fields randomly changed to different values \u2013 to produce child models. These child models are then assigned a fitness via training and evaluation on the target task, as the initial population was. When this fitness evaluation concludes, the population is sampled from once again, and the individual in the subpopulation with the lowest fitness is killed, meaning it is removed from the population. The newly evaluated child model is then added to the population, taking the killed individual's place. This process is repeated and results in a population with high fitness individuals, which in our case represent well-performing architectures." + ], + [ + "Our encoding search space is inspired by the NASNet search space BIBREF1 , but is altered to allow it to express architecture characteristics found in recent state-of-the-art feed-forward seq2seq networks. Crucially, we ensured that the search space can represent the Transformer, so that we can seed the search process with the Transformer itself.", + "Our search space consists of two stackable cells, one for the model encoder and one for the decoder (see Figure 1 ). Each cell contains NASNet-style blocks, which receive two hidden state inputs and produce new hidden states as outputs; the encoder contains six blocks and the decoder contains eight blocks, so that the Transformer can be represented exactly. The blocks perform separate transformations to each input and then combine the transformation outputs together to produce a single block output; we will refer to the transformations applied to each input as a branch. Our search space contains five branch-level search fields (input, normalization, layer, output dimension and activation), one block-level search field (combiner function) and one cell-level search field (number of cells).", + "In our search space, a child model's genetic encoding is expressed as: $[$ left input, left normalization, left layer, left relative output dimension, left activation, right input, right normalization, right layer, right relative output dimension, right activation, combiner function $]$ $\\times $ 14 + $[$ number of cells $]$ $\\times $ 2, with the first 6 blocks allocated to the encoder and the latter 8 allocated to the decoder. Given the vocabularies described in the Appendix, this yields a search space of $7.30 * 10^{115}$ models, although we do shrink this to some degree by introducing constraints (see the Appendix for more details)." + ], + [ + "While previous neural architecture search works rely on well-formed hand crafted search spaces BIBREF1 , we intentionally leave our search minimally tuned, in a effort to alleviate our manual burden and emphasize the role of the automated search method. To help navigate the large search space we create for ourselves, we find it easier to seed our initial population with a known strong model, in this case the Transformer. This anchors the search to a known good starting point and guarantees at least a single strong potential parent in the population as the generations progress. We offer empirical support for these claims in our Results section." + ], + [ + "The evolution algorithm we employ is adapted from the tournament selection evolutionary architecture search proposed by Real et al. real19, described above. Unlike Real et al. real19 who conducted their search on CIFAR-10, our search is conducted on a task that takes much longer to train and evaluate on. Specifically, to train a Transformer to peak performance on WMT'14 En-De requires $\\sim $ 300k training steps, or 10 hours, in the base size when using a single Google TPU V.2 chip, as we do in our search. In contrast, Real et al. real19 used the less resource-intensive CIFAR-10 task BIBREF29 , which takes about two hours to train on, to assess their models during their search, as it was a good proxy for ImageNet BIBREF30 performance BIBREF1 . However, in our preliminary experimentation we could not find a proxy task that gave adequate signal for how well each child model would perform on the full WMT'14 En-De task; we investigated using only a fraction of the data set and various forms of aggressive early stopping.", + "To address this problem we formulated a method to dynamically allocate resources to more promising architectures according to their fitness. This method, which we refer to as progressive dynamic hurdles (PDH), allows models that are consistently performing well to train for more steps. It begins as ordinary tournament selection evolutionary architecture search with early stopping, with each child model training for a relatively small $s_0$ number of steps before being evaluated for fitness. However, after a predetermined number of child models, $m$ , have been evaluated, a hurdle, $h_0$ , is created by calculating the the mean fitness of the current population. For the next $m$ child models produced, models that achieve a fitness greater than $h_0$ after $s_0$ train steps are granted an additional $s_1$ steps of training and then are evaluated again to determine their final fitness. Once another $m$ models have been considered this way, another hurdle, $h_1$ , is constructed by calculating the mean fitness of all members of the current population that were trained for the maximum number of steps. For the next $m$ child models, training and evaluation continues in the same fashion, except models with fitness greater than $m$0 after $m$1 steps of training are granted an additional $m$2 number of train steps, before being evaluated for their final fitness. This process is repeated until a satisfactory number of maximum training steps is reached. Algorithm 1 (Appendix) formalizes how the fitness of an individual model is calculated with hurdles and Algorithm 2 (Appendix) describes tournament selection augmented with progressive dynamic hurdles.", + "Although different child models may train for different numbers of steps before being assigned their final fitness, this does not make their fitnesses incomparable. Tournament selection evolution is only concerned with relative fitness rank when selecting which subpopulation members will be killed and which will become parents; the margin by which one candidate is better or worse than the other members of the subpopulation does not matter. Assuming no model overfits during its training and that its fitness monotonically increases with respect to the number of train steps it is allocated, a comparison between two child models can be viewed as a comparison between their fitnesses at the lower of the two's cumulative train steps. Since the model that was allocated more train steps performed, by definition, above the fitness hurdle for the lower number of steps and the model that was allocated less steps performed, by definition, at or below that hurdle at the lower number of steps, it is guaranteed that the model with more train steps was better when it was evaluated at the lower number of train steps.", + "The benefit of altering the fitness algorithm this way is that poor performing child models will not consume as many resources when their fitness is being computed. As soon as a candidate's fitness falls below a tolerable amount, its evaluation immediately ends. This may also result in good candidates being labeled as bad models if they are only strong towards the latter part of training. However, the resources saved as a result of discarding many bad models improves the overall quality of the search enough to justify potentially also discarding some good ones; this is supported empirically in our Results section." + ], + [ + "We use three different machine translation datasets to perform our experiments, all of which were taken from their Tensor2Tensor implementations. The first is WMT English-German, for which we mimic Vaswani et al.'s vaswani17 setup, using WMT'18 En-De training data without ParaCrawl BIBREF31 , yielding 4.5 million sentence pairs. In the same fashion, we use newstest2013 for development and test on newstest2014. The second translation dataset is WMT En-Fr, for which we also replicate Vaswani et.al's vaswani17 setup. We train on the 36 million sentence pairs of WMT'14 En-Fr, validate on newstest2013 and test on newstest2014. The final translation dataset is WMT English-Czech (En-Cs). We used the WMT'18 training dataset, again without ParaCrawl, and used newstest2013 and newstest2014 as validation and test sets. For all tasks, tokens were split using a shared source-target vocabulary of about 32k word-pieces BIBREF32 .", + "All datasets were generated using Tensor2Tensor's \u201cpacked\" scheme; sentences were shuffled and concatenated together with padding to form uniform 256 length inputs and targets, with examples longer than 256 being discarded. This yielded batch sizes of 4096 tokens per GPU or TPU chip; accordingly, 16 TPU chip configurations had $\\sim $ 66K tokens per batch and 8 GPU chip configurations had $\\sim $ 33K tokens per batch.", + "For language modeling we used the 1 Billion Word Language Model Benchmark (LM1B) BIBREF33 , also using its \u201cpacked\" Tensor2Tensor implementation. Again the tokens are split into a vocabulary of approximately 32k word-pieces and the sentences are shuffled." + ], + [ + "All of our experiments used Tensor2Tensor's Transformer TPU hyperparameter settings. These are nearly identical to those used by Vaswani et al. vaswani17, but modified to use the memory-efficient Adafactor BIBREF34 optimizer. Aside from using the optimizer itself, these hyperparameters set the warmup to a constant learning rate of $10^{-2}$ over 10k steps and then uses inverse-square-root learning-rate decay. For our experiments, we make only one change, which is to alter this decay so that it reaches 0 at the final step of training, which for our non-search experiments is uniformly 300k. We found that the our search candidate models, the Transformer, and the Evolved Transformer all benefited from this and so experimented with using linear decay, single-cycle cosine decay BIBREF35 and a modified inverse-square-root decay to 0 at 300k steps: $lr = step^{-0.00303926} - .962392$ ; every decay was paired with the same constant $10^{-2}$ warmup. We used WMT En-De validation perplexity to gauge model performance and found that the Transformer preferred the modified inverse-square-root decay. Therefore, this is what we used for both all our Transformer trainings and the architecture searches themselves. The Evolved Transformer performed best with cosine decay and so that is what we used for all of its trainings. Besides this one difference, the hyperparameter settings across models being compared are exactly the same. Because decaying to 0 resulted in only marginal weight changes towards the end of training, we did not use checkpoint averaging.", + "Per-task there is one additional hyperparameter difference, which is dropout rate. For ET and all search child models, dropout was applied uniformly after each layer, approximating the Transformer's more nuanced dropout scheme. For En-De and En-Cs, all \u201cbig\" sized models were given a higher dropout rate of 0.3, keeping in line with Vaswani et al. vaswani17, and all models with an input embedding size of 768 are given a dropout rate of 0.2. Aside from this, hyperparameters are identical across all translation tasks.", + "For decoding we used the same beam decoding configuration used by Vaswani et al. vaswani17. That is a beam size of 4, length penalty ( $\\alpha $ ) of 0.6, and maximum output length of input length + 50. All BLEU is calculated using case-sensitive tokenization and for WMT'14 En-De we also use the compound splitting that was used in Vaswani et al. vaswani17.", + "Our language model training setup is identical to our machine translation setup except we remove label smoothing and lower the intra-attention dropout rate to 0. This was taken from the Tensor2Tensor hyperparameters for LM1B[2]." + ], + [ + "All of the architecture searches we describe were run on WMT'14 En-De. They utilized the search space and tournament selection evolution algorithm described in our Methods section. Unless otherwise noted, each search used 200 workers, which were equipped with a single Google TPU V.2 chip for training and evaluation. We maintained a population of size 100 with subpopulation sizes for both killing and reproducing set to 30. Mutations were applied independently per encoding field at a rate of 2.5%. For fitness we used the negative log perplexity of the validation set instead of BLEU because, as demonstrated in our Results section, perplexity is more consistent and that reduced the noise of our fitness signal." + ], + [ + "In this section, we will first benchmark the performance of our search method, progressive dynamic hurdles, against other evolutionary search methods BIBREF21 , BIBREF0 . We will then benchmark the Evolved Transformer, the result of our search method, against the Transformer BIBREF7 ." + ], + [ + "We tested our evolution algorithm enhancements \u2013 using PDH and seeding the initial population with the Transformer \u2013 against control searches that did not use these techniques; without our enhancements, these controls function the same way as Real et. al's real19 searches, without aging regularization. Each search we describe was run 3 times and the top model from each run was retrained on a single TPU V.2 chip for 300k steps. The performance of the models after retraining is given in Table 1 .", + "Our proposed search (Table 1 row 1), which used both PDH and Transformer seeding, was run first, with hurdles created every 1k models ( $m = 1000$ ) and six 30k train step (1 hour) increments ( $s=<30, 30, 30, 30, 30, 30>$ ). To test the effectiveness of seeding with the Transformer, we ran an identical search that was instead seeded with random valid encodings (Table 1 row 2). To test the effectiveness of PDH, we ran three controls (Table 1 rows 3-5) that each used a fixed number of train steps for each child model instead of hurdles (Table 1 column 2). For these we used the step increments (30k), the maximum number of steps our proposed search ultimately reaches (180k), and the total number of steps each top model receives when fully trained to gauge its final performance (300k). To determine the number of child models each of these searches would be able to train, we selected the value that would make the total amount of resources used by each control search equal to the maximum amount of resources used for our proposed searches, which require various amounts of resources depending on how many models fail to overcome hurdles. In the three trials we ran, our proposed search's total number of train steps used was 422M $\\pm $ 21M, with a maximum of 446M. Thus the number of child models allotted for each non-PDH control search was set so that the total number of child model train steps used would be 446M.", + "As demonstrated in Table 1, the search we propose, with PDH and Transformer seeding, has the best performance on average. It also is the most consistent, having the lowest standard deviation. Of all the searches conducted, only a single control run \u2013 \u201c30K no hurdles\" (Table 1 row 3) \u2013 produced a model that was better than any of our proposed search's best models. At the same time, the \u201c30K no hurdles\" setup also produced models that were significantly worse, which explains its high standard deviation. This phenomenon was a chief motivator for our developing this method. Although aggressive early stopping has the potential to produce strong models for cheap, searches that utilize it can also venture into modalities in which top fitness child models are only strong early on. Without running models for longer, whether or not this is happening cannot be detected. The 180K and 300K no hurdles searches did have insight into long term performance, but in a resource-inefficient manner that hurt these searches by limiting the number of generations they produced; for the \u201c180k no hurdles\" run to train as many models as PDH would require 1.08B train steps, over double what PDH used in our worst case.", + "Searching with random seeding also proved to be ineffective, performing considerably worse than every other configuration. Of the five searches run, random seeding was the only one that had a top model perplexity higher than the Transformer, which is 4.75 $\\pm $ 0.01 in the same setup.", + "After confirming the effectiveness of our search procedure, we launched a larger scale version of our search using 270 workers. We trained 5k models per hurdle ( $m=5000$ ) and used larger step increments to get a closer approximation to 300k step performance: $s = <60, 60, 120>$ . The setup was the same as the Search Techniques experiments, except after 11k models we lowered the mutation rate to 0.01 and introduced the NONE value to the normalization mutation vocabulary.", + "The search ran for 15K child models, requiring a total of 979M train steps. Over 13K models did not make it past the first hurdle, drastically reducing the resources required to view the 240 thousandth train step for top models, which would have cost 3.6B train steps for the same number of models without hurdles. After the search concluded, we then selected the top 20 models and trained them for the full 300k steps, each on a single TPU V.2 chip. The model that ended with the best perplexity is what we refer to as the Evolved Transformer (ET). Figure 3 shows the ET architecture. The most notable aspect of the Evolved Transformer is the use of wide depth-wise separable convolutions in the lower layers of the encoder and decoder blocks. The use of depth-wise convolution and self-attention was previously described in QANet BIBREF37 , however the overall architectures of the Evolved Transformer and QANet are different in many ways: e.g., QANet has smaller kernel sizes and no branching structures. The performance and analysis of the Evolved Transformer will be shown in the next section." + ], + [ + "To test the effectiveness of the found architecture \u2013 the Evolved Transformer \u2013 we compared it to the Transformer in its Tensor2Tensor training regime on WMT'14 En-De. Table 2 shows the results of these experiments run on the same 8 NVIDIA P100 hardware setup that was used by Vaswani et al. vaswani17. Observing ET's improved performance at parameter-comparable \u201cbase\" and \u201cbig\" sizes, we were also interested in understanding how small ET could be shrunk while still achieving the same performance as the Transformer. To create a spectrum of model sizes for each architecture, we selected different input embedding sizes and shrank or grew the rest of the model embedding sizes with the same proportions. Aside from embedding depths, these models are identical at all sizes, except the \u201cbig\" 1024 input embedding size, for which all 8 head attention layers are upgraded to 16 head attention layers, as was done in Vaswani et al. vaswani17.", + "ET demonstrates stronger performance than the Transformer at all sizes, with the largest difference of 0.7 BLEU at the smallest, mobile-friendly, size of $\\sim $ 7M parameters. Performance on par with the \u201cbase\" Transformer was reached when ET used just 71.4% of its FLOPS and performance of the \u201cbig\" Transformer was exceeded by the ET model that used 44.8% less FLOPS. Figure 4 shows the FLOPS vs. BLEU performance of both architectures.", + "To test ET's generalizability, we also compared it to the Transformer on an additional three well-established language tasks: WMT'14 En-Fr, WMT'14 En-Cs, and LM1B. Upgrading to 16 TPU V.2 chips, we doubled the number of synchronous workers for these experiments, pushing both models to their higher potential BIBREF17 . We ran each configuration 3 times, except WMT En-De, which we ran 6 times; this was a matter of resource availability and we gave priority to the task we searched on. As shown in Table 3 , ET performs at least one standard deviation above the Transformer in each of these tasks. Note that the Transformer mean BLEU scores in both Tables 2 and 3 for WMT'14 En-Fr and WMT'14 En-De are higher than those originally reported by Vaswani et al. BIBREF7 .", + "As can be seen in Tables 2 and 3 , the Evolved Transformer is much more effective than the Transformer when its model size is small. When the model size becomes large, its BLEU performance saturates and the gap between the Evolved Transformer and the Transformer becomes smaller. One explanation for this behavior is that overfitting starts to occur at big model sizes, but we expect that data augmentation BIBREF17 or hyperparameter tuning could improve performance. The improvement gains in perplexity by ET, however, are still significant for big model sizes; it is worth emphasizing that perplexity is also a reliable metric for measuring machine translation quality BIBREF32 .", + "To understand what mutations contributed to ET's improved performance we conducted two rounds of ablation testing. In the first round, we began with the Transformer and applied each mutation to it individually to measure the performance change each mutation introduces in isolation. In the second round, we began with ET and removed each mutation individually to again measure the impact of each single mutation. In both cases, each model was trained 3 times on WMT En-De for 300k steps with identical hyperparameters, using the inverse-square-root decay to 0 that the Transformer prefers. Each training was conducted on a single TPU V.2 chip. The results of these experiments are presented in Table 4 of the Appendix; we use validation perplexity for comparison because it was our fitness metric.", + "In all cases, the augmented ET models outperformed the the augmented Transformer models, indicating that the gap in performance between ET and the Transformer cannot be attributed to any single mutation. The mutation with the seemingly strongest individual impact is the increase from 3 to 4 decoder blocks. However, even when this mutation is introduced to the Transformer and removed from ET, the resulting augmented ET model still has a higher fitness than the augmented Transformer model.", + "Some mutations seem to only hurt model performance such as converting the encoder's first multi-head attention into a Gated Linear Unit. However, given how we formulate the problem \u2013 finding an improved model with a comparable number of parameters to the Transformer \u2013 these mutations might have been necessary. For example, when the Transformer decoder block is repeated 4 times, the resulting model has 69.6M parameters, which is outside of our allowed parameter range. Thus, mutations that shrank ET's total number of parameters, even at a slight degradation of performance, were necessary so that other more impactful parameter-expensive mutations, such as adding an additional decoder block, could be used. If model size is not of concern, some of these mutations could potentially be reverted to further improve performance. Likewise, parameter efficient layers, such as the depthwise-separable convolutions, could potentially be swapped for their less efficient counterparts, such as standard convolution." + ], + [ + "We presented the first neural architecture search conducted to find improved feed-forward sequence models. We first constructed a large search space inspired by recent advances in seq2seq models and used it to search directly on the computationally intensive WMT En-De translation task. To mitigate the size of our space and the cost of training child models, we proposed using both our progressive dynamic hurdles method and seeding our initial population with a known strong model, the Transformer.", + "When run at scale, our search found the Evolved Transformer. In a side by side comparison against the Transformer in an identical training regime, the Evolved Transformer showed consistent stronger performance on both translation and language modeling. On WMT En-De, the Evolved Transformer was twice as efficient FLOPS-wise as the Transformer big model and at a mobile-friendly size, the Evolved Transformer demonstrated a 0.7 BLEU gain over the Transformer architecture." + ], + [ + "We would like to thank Ashish Vaswani, Jakob Uszkoreit, Niki Parmar, Noam Shazeer, Lukasz Kaiser and Ryan Sepassi for their help with Tensor2Tensor and for sharing their understanding of the Transformer. We are also grateful to David Dohan, Esteban Real, Yanping Huang, Alok Aggarwal, Vijay Vasudevan, and Chris Ying for lending their expertise in architecture search and evolution." + ], + [ + "In the following, we describe the algorithm that we use to calculate child model fitness with hurdles (Algorithm \"Search Algorithms\" ) and evolution architecture search with progressive dynamic hurdles (Algorithm UID28 ). [h!] Calculate Model Fitness with Hurdles inputs: $model$ : the child model $s$ : vector of train step increments $h$ : queue of hurdles append $-\\infty $ to $h$ TRAIN_N_STEPS( $model$ , $s_0$ ) $fitness \\leftarrow $ EVALUATE( $model$ ) $i \\leftarrow 0$ $s$0 $s$1 $s$2 TRAIN_N_STEPS( $s$3 , $s$4 ) $s$5 EVALUATE( $s$6 ) $s$7 return $s$8 " + ], + [ + "In our search space, a child model's genetic encoding is expressed as: $[$ left input, left normalization, left layer, left relative output dimension, left activation, right input, right normalization, right layer, right relative output dimension, right activation, combiner function $]$ $\\times $ 14 + $[$ number of cells $]$ $\\times $ 2, with the first 6 blocks allocated to the encoder and the latter 8 allocated to the decoder. In the following, we will describe each of the components." + ], + [ + "To understand what mutations contributed to ET's improved performance we conducted two rounds of ablation testing. In the first round, we began with the Transformer and applied each mutation to it individually to measure the performance change each mutation introduces in isolation. In the second round, we began with ET and removed each mutation individually to again measure the impact of each single mutation. In both cases, each model was trained 3 times on WMT En-De for 300k steps with identical hyperparameters, using the inverse-square-root decay to 0 that the Transformer prefers. Each training was conducted on a single TPU V.2 chip. The results of these experiments are presented in Table 4 ; we use validation perplexity for comparison because it was our fitness metric.", + "To highlight the impact of each augmented model's mutation, we present not only their perplexities but also the difference between their mean perplexity and their unaugmented base model mean perplexity in the \"Mean Diff\" columns:", + "base model mean perplexity - augmented mean perplexity", + "This delta estimates the change in performance each mutation creates in isolation. Red highlighted cells contain evidence that their corresponding mutation hurt overall performance. Green highlighted cells contain evidence that their corresponding mutation helped overall performance.", + "In half of the cases, both the augmented Transformer's and the augmented Evolved Transformer's performances indicate that the mutation was helpful. Changing the number of attention heads from 8 to 16 was doubly indicated to be neutral and changing from 8 head self attention to a GLU layer in the decoder was doubly indicated to have hurt performance. However, this and other mutations that seemingly hurt performance may have been necessary given how we formulate the problem: finding an improved model with a comparable number of parameters to the Transformer. For example, when the Transformer decoder block is repeated 4 times, the resulting model has 69.6M parameters, which is outside of our allowed parameter range. Thus, mutations that shrank ET's total number of parameters, even at a slight degradation of performance, were necessary so that other more impactful parameter-expensive mutations, such as adding an additional decoder block, could be used.", + "Other mutations have inconsistent evidence about how useful they are. This ablation study serves only to approximate what is useful, but how effective a mutation is also depends on the model it is being introduced to and how it interacts with other encoding field values." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2016/instruction.md b/qasper-2016/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..edb8218903a47e62f8c97d78a9ddeb7356d4345b --- /dev/null +++ b/qasper-2016/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Measuring the Reliability of Hate Speech Annotations: The Case of the European Refugee Crisis + +Question: Was the degree of offensiveness taken as how generally offensive the text was, or how personally offensive it was to the annotator? \ No newline at end of file diff --git a/qasper-2019/instruction.md b/qasper-2019/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..4a906ee1506cdf3529334c92f23b7bd22497324d --- /dev/null +++ b/qasper-2019/instruction.md @@ -0,0 +1,119 @@ +Name of Paper: Annotating and normalizing biomedical NEs with limited knowledge + +Question: What does their system consist of? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Resource building", + "Resource building ::: SNOMED CT", + "Resource building ::: Contextual regexp grammars", + "Development", + "Development ::: Annotation process", + "Results", + "Discussion", + "Discussion ::: Inconsistency in the ag", + "Discussion ::: Inconsistency in ha as regards ag", + "Discussion ::: Inconsistency in ha on the test set as regards t+d sets", + "Discussion ::: Asystematic/incomplete annotation", + "Discussion ::: Incorrect sctids", + "Conclusions", + "Acknowledgements" + ], + "paragraphs": [ + [ + "Named Entity Recognition (ner) is considered a necessary first step in the linguistic processing of any new domain, as it facilitates the development of applications showing co-occurrences of domain entities, cause-effect relations among them, and, eventually, it opens the (still to be reached) possibility of understanding full text content. On the other hand, Biomedical literature and, more specifically, clinical texts, show a number of features as regards ner that pose a challenge to NLP researchers BIBREF0: (1) the clinical discourse is characterized by being conceptually very dense; (2) the number of different classes for nes is greater than traditional classes used with, for instance, newswire text; (3) they show a high formal variability for nes (actually, it is rare to find entities in their \u201ccanonical form\u201d); and, (4) this text type contains a great number of ortho-typographic errors, due mainly to time constraints when drafted.", + "Many ways to approach ner for biomedical literature have been proposed, but they roughly fall into three main categories: rule-based, dictionary-based (sometimes called knowledge-based) and machine-learning based solutions. Traditionally, the first two approaches have been the choice before the availability of Human Annotated Datasets (had), albeit rule-based approaches require (usually hand-crafted) rules to identify terms in the text, while dictionary-based approaches tend to miss medical terms not mentioned in the system dictionary BIBREF1. Nonetheless, with the creation and distribution of had as well as the development and success of supervised machine learning methods, a plethora of data-driven approaches have emerged \u2014from Hidden Markov Models (HMMs) BIBREF2, Support Vector Machines (SVMs) BIBREF3 and Conditional Random Fields (CRFs) BIBREF4, to, more recently, those founded on neural networks BIBREF5. This fact has had an impact on knowledge-based methods, demoting them to a second plane. Besides, this situation has been favoured by claims on the uselessness of gazetteers for ner in, for example, Genomic Medicine (GM), as it was suggested by BIBREF0 [p. 26]CohenandDemner-Fushman:2014:", + "One of the findings of the first BioCreative shared task was the demonstration of the long-suspected fact that gazetteers are typically of little use in GM.", + "Although one might think that this view could strictly refer to the subdomain of GM and to the past \u2014BioCreative I was a shared task held back in 2004\u2014, we can still find similar claims today, not only referred to rule-based and dictionary-based methods, but also to stochastic ones BIBREF5.", + "In this paper, in spite of previous statements, we present a system that uses rule-based and dictionary-based methods combined (in a way we prefer to call resource-based). Our final goals in the paper are two-fold: on the one hand, to describe our system, developed for the PharmaCoNER shared task, dealing with the annotation of some of the nes in health records (namely, pharmacological, chemical and biomedical entities) using a revisited version of rule- and dictionary-based approaches; and, on the other hand, to give pause for thought about the quality of datasets (and, thus, the fairness) with which systems of this type are evaluated, and to highlight the key role of resource-based systems in the validation and consolidation of both the annotation guidelines and the human annotation practices.", + "In section SECREF2, we describe our initial resources and explain how they were built, and try to address the issues posed by features (1) and (2) above. Section SECREF3 depicts the core of our system and the methods we have devised to deal with text features (3) and (4). Results obtained in PharmaCoNER by our system are presented in section SECREF4. Section SECREF5 details some of our errors, but, most importantly, focusses on the errors and inconsistencies found in the evaluation dataset, given that they may shed doubts on the scores obtained by any system in the competition. Finally, we present some concluding remarks in section SECREF6." + ], + [ + "As it is common in resource-based system development, special effort has been devoted to the creation of the set of resources used by the system. These are mainly two \u2014a flat subset of the snomed ct medical ontology, and the library and a part of the contextual regexp grammars developed by BIBREF6 FSL:2018 for a previous competition on abbreviation resolution in clinical texts written in Spanish. The process of creation and/or adaptation of these resources is described in this section." + ], + [ + "Although the competition proposes two different scenarios, in fact, both are guided by the snomed ct ontology \u2014for subtask 1, entities must be identified with offsets and mapped to a predefined set of four classes (PROTEINAS, NORMALIZABLES, NO_NORMALIZABLES and UNCLEAR); for subtask 2, a list of all snomed ct ids (sctid) for entities occurring in the text must be given, which has been called concept indexing by the shared task organizers. Moreover, PharmaCoNER organizers decided to promote snomed ct substance ids over product, procedure or other possible interpretations also available in this medical ontology for a given entity. This selection must be done even if the context clearly refers to a different concept, according to the annotation guidelines (henceforth, AnnotGuide) and the praxis. Finally, PROTEINAS is ranked as the first choice for substances in this category.", + "These previous decisions alone on the part of the organizers greatly simplify the task at hand, making it possible to build (carefully compiled) subsets of the entities to be annotated. This is a great advantage over open domain ner, where (like in GM) the texts may contain an infinite (and very creative indeed) number of nes. For clinical cases, although the ne density is greater, there exist highly structured terminological resources for the domain. Moreover, the set of classes to use in the annotation exercise for subtask 1 has been dramatically cut down by the organizers.", + "With the above-mentioned initial constraints in mind, we have painstakingly collected, from the whole set of snomed ct terms, instances of entities as classified by the human annotators in the datasets released by the organizers and, when browsing the snomed ct web version, we have tried to use the ontological hierarchical relations to pull a complete class down from snomed ct. This way, we have gathered 80 classes \u2014from lipids to proteins to peptides or peptide hormones, from plasminogen activators to dyes to drugs or medicaments\u2014, that have been arranged in a ranked way so as to mimic human annotators choices. The number of entities so collected (henceforth, `primary entities') is 51,309." + ], + [ + "Some of the entities to be annotated, specially those in abbreviated form, are ambiguous without a context. This is the case, for instance, of PCR, whose expanded forms are (among other meanings; we use only English expanded forms) `reactive protein c', `polymerase chain reaction', `cardiorespiratory arrest'. In order to deal with these cases, we use a contextual regexp rule system with a lean and simple rule formalism previously developed BIBREF6. As an exemplification, we include one rule to deal with one of the cases of the preceeding ambiguity:", + "b:[il::bioqu\u00edmica|en sangre|hemoglobina|", + "hemograma|leucocit|par\u00e1sito|plaqueta|", + "prote.na|recuento|urea] - [PCR] - >", + "[m=prote\u00edna]", + "", + "A rule has a left hand side (LHS) and a right hand side (RHS). There is a focus in the LHS (PCR, within dashes) and a left and right context (that may be empty). When the left context includes a b: (like in this case), it indicates either left or right context. The words in the context can take other qualifiers \u2014in this case, the matching will be case insensitive (i to the left of bioqu\u00edmica) and local (l), which means the disjunction of words and/or stems can be found in a distance of 40 characters (this can be modifified by the user). Hence, the rule applies, selecting the prote\u00edna expansion (in RHS) of PCR if any of the words/stems specified as local context (40 chars maximum) is matched either to the left or right of the focus term (which is usually an abbreviation).", + "With no tweaking at all for the datasets in PharmaCoNER competition, the system annotates correctly 18 out of 20 occurrences of PCR in the test dataset (a precision of 0.9).", + "This component of the system is important because, only when the previous abbreviation is expanded as the first string (that of a protein name), it must be annotated, according to the AnnotGuide. The same ambiguity happens with Cr, which may mean `creatinine' or `chrome'. These expansions are both NORMALIZABLES, but, obviously, their sctid is different.", + "The system currently uses 104 context rules, only for abbreviations and acronyms in the clinical cases. These rules, contrary to what is commonly referred in the biomedical processing literature BIBREF5, do not require a special domain knowledge (none of the authors do have it) and can be written, most of the times, in a very straightforward way in the formalism briefly described above." + ], + [ + "In general, dictionary-based methods rely on strict string matching over a fixed set of lexical entries from the domain. This is clearly insufficient to deal with non-canonical linguistic forms of nes as used in clinical texts. For this reason, we have devised two different solutions to this shortcoming.", + "In the first place, we have munged a great number of our primary entities, in a way similar to that described in BIBREF7 FSL:2019a for gazetteers used for protected information anonymization in clinical texts. We basically transform canonical forms in other possible textual forms observed when working with biomedical texts. With such transformations, a system module converts a salt compound like clorhidrato de ciclopentolato into ciclopentolato clorhidrato, or simply the PP de potasio into its corresponding adjective pot\u00e1sico. Other, more complex conversions include the treatment of antibodies \u2014for instance anticuerpo contra especie de Leishmania becomes ac. Leishmania, among other variants\u2014, or pairs of antibiotics normally prescribed together \u2014which have a unique sctid and whose order we handle just as the `glueing' characters. Note, incidentally, that, while the input to this pre-processing step is always a string, the output can be a regular expression, that is linked to a sctid. Plural forms are also generated through this module, that uses 45 transformations (not all equally productive). Using these transformation rules, we produce 139,150 `secondary entities', many of them regexps. As a final (simple) example of this, consider the entity ant\u00edgeno CD13: after applying one of the previous string-to-regexp transformations, it is converted to:", + "(?:ant\u00edgeno )?CD[- ]?13", + "", + "With the previous regexp, the system is able to identify (and string-normalize) six different textual realizations of the same unique snomed ct term. There are more complex rules that, thus, produce many more potential strings. The important thing with this strategy is that through the generative power of these predictably-created regexps from snomed ct entities the system is able to improve its recall and overcome the limitations of traditional dictionary-based approaches.", + "Secondly, to tackle with careless drafting of clinical reports, a Levenshtein edit distance library is used on the whole background dataset. The process is run once, using our secondary entities as lexicon and a general vocabulary lexicon to rule out common words in the candidate search process. We have used distances in the range 1-3 (depending on string length) for sequences up to 3 words long. The output of this process, which links forms with spelling errors with canonical ones and, thus, to sctids, can be inspected prior to its inclusion in the system lexicon, if so desired." + ], + [ + "As such, the annotation process is very simple. The program reads the input byte stream trying to identify known entities by means of a huge regexp built through the pre-processing of the available resources. If the candidate entity is ambiguous and (at least) one contextual rule exists for it, it is applied. For the rest of the nes, the system assigns them the class and sctid found in our ranked in-memory lexicon. As already mentioned in passing, the system does not tokenize text prior to ner, a processing order that we consider the right choice for highly entity-dense texts. The data structures built during pre-processing are efficiently stored on disk for subsequent runs, so the pre-processing is redone only when resources are edited." + ], + [ + "According to the organizers, and taking into account the ha of the tiny subset from the background dataset released to the participants, the system obtained the scores presented in table TABREF16, ranking as second best system for subtask1 and best system for subtask2 BIBREF8..", + "Our results are consistent with our poor understanding of the classes for subtask 1. Having a null knowledge of Pharmacology, Biomedicine or even Chemistry, assigning classes (as requested for subtask 1) to entities is very hard, while providing a sctid (subtask 2) seems an easier goal. We will explain the point with an example entity \u2014\u00e1cido hialur\u00f3nico (`hyaluronic acid'). Using the ontological structure of snomed ct, one can find the following parent relations (just in English):", + "hyaluronic acid is-a mucopolysaccharide is-a protein", + "The authors have, in this case, promoted the PROTEINAS annotation for this entity, disregarding its interpretation as a replacement agent and overlooking a recommendation on polysaccharides in the AnnotGuide. Fortunately, all its interpretations share a unique sctid. The same may be true for", + "haemosiderin is-a protein", + "which is considered NORMALIZABLE in the test dataset. Similar cases are responsible for the lower performance on subtask 1 with respect to the more complex subtask 2.", + "In spite of these human classification errors, our system scores outperform those obtained by PharmacoNER Tagger BIBREF5, a simpler system using a binary classification and a very different organization of the dataset with a smaller fragment for test (10% of the data as opposed to 25% for the official competition). In fact, our system improves their F1-score (89.06) by 1.3 points when compared with our results for the more complex PharmaCoNER subtask 1." + ], + [ + "In this section, we perform error analysis for our system run on the test dataset. We will address both recall and precision errors, but mainly concentrate on the latter type, and on a thorough revision of mismatches between system and human annotations.", + "In general, error analysis is favoured by knowledge-based methods, since it is through the understanding of the underlying reasons for an error that the system could be improved. Moreover, and differently to what happens with the current wave of artificial neural network methods, the whole annotation process \u2014its guidelines for human annotators, the collection and appropriate structuring of resources, the adequate means to assign tags to certain entities but not to other, similar or even pertaining to the same class\u2014 must be clearly understood by the designer/developer/data architect of such systems. As a natural consequence of this attempt to mimick a task defined by humans to be performed, in the first place, also by humans, some inconsistencies, asystematic or missing assignments can be discovered, and this information is a valuable treasure not only for system developers but also for task organizers, guideline editors and future annotation campaigns, not to mention for the exactness of program evaluation results.", + "Most of the error types made by the system (i.e., by the authors) in class assignment for subtask 1 have already been discussed. In the same vein, as regards subtask 2, a great number of errors come from the selection of the `product containing substance' reading from snomed ct rather to the `substance' itself. This is due to inexperience of the authors on the domain and the wrong consideration of context when tagging entities \u2014the latter being clearly obviated in the AnnotGuide.", + "In the following paragraphs, some of the most relevant inconsistencies found when performing error analysis of our system are highlighted. The list is necessarily incomplete due to space constraints, and it is geared towards the explanation of our possible errors." + ], + [ + "Among some of the paradoxical examples in the AnnotGuide it stands out the double explicit consideration of gen (`gene'), when occurs alone in context, as both an entity to be tagged (positive rule P2 of the AnnotGuide) and a noun not to be tagged (negative rule N2). This inconsistency (and a bit of bad luck) has produced that none of the 6 occurrences as an independent noun \u2014not introducing an entity\u2014 is tagged in the train+dev (henceforth, t+d) while the only 2 in the same context in the test dataset have been tagged. This amounts for 2 true negatives (tns) for the evaluation script." + ], + [ + "The AnnotGuide proposal for the treatment of elliptical elements is somewhat confusing. For these cases, a longest match annotation is proposed, which is difficult to replicate automatically and not easy to remember for the human annotator. In many contexts, the annotator has made the right choice \u2014for instance, in receptores de estr\u00f3geno y de progesterona\u2014 whereas in others do not \u2014$|$anticuerpos anticardiolipina$|$ $|$IgG$|$ e $|$IgM$|$, with `$|$' marking the edges of the annotations. The last example occurs twice in the test dataset. Hence, the disagreement counts as 6 tns and 2 false positives (fps).", + "On the other hand, there is a clear reference to food materials and nutrition in the AnnotGuide, where they are included in the class of substances. However, none of the following entities is tagged in the test dataset: az\u00facar (which is mandatory according to AnnotGuide and was tagged in t+d; 1 fp); almid\u00f3n de ma\u00edz (also mandatory in AnnotGuide; 1 fp); and Loprof\u00edn, Aglutella, Aproten (hypoproteic nutrition products, 3 fps in total).", + "There is an explicit indication in the AnnotGuide to annotate salts, with the example iron salts. However, in the context sales de litio (`lithium salts'), only the chemical element has been tagged (1 fp).", + "There exist other differing-span mismatches between human and automatic annotation. These include anticuerpos anticitoplasma de neutr\u00f3filo, where the ha considers the first two words only (in one of the occurrences, 1 fp); in the text fragment b2 microglobulina, CEA y CA 19,9 normales, CA 19,9 is the correct span for the last entity (and not CA, 1 fp); A.S.T is the span selected (for A.S.T., 1 fp); finally, in the context lgM anticore only lgM has been tagged (1 fp).", + "Other prominent mismatch between had and AnnotGuide is that of DNA, which is explicitly included in the AnnotGuide (sects. P2 and O1). It accounts for 2 fps.", + "But perhaps one of the most common discrepancies between human and automatic annotation has to do with medicaments normally prescribed together, which have a unique sctid. Examples include amiloride/hidroclorotiazida (1 fp); and betametasona + calcipotriol (1 fp) in the test set. This situation was also observed in the t+d corpus fragment (tenofovir + emtricitabina, carbonato c\u00e1lcico /colecalciferol, lopinavir/ritonavir)." + ], + [ + "Some inconsistencies between dataset annotations have turned the authors crazy: NPT (acronym for `total parenteral nutrition, TPN') is tagged in the train+dev dataset 15 out of 21 times it occurs. The common sense of frequency in the ha of texts has led us to tag it in the background set. Unluckily, neither NPT nor its expansion have been tagged in the test dataset. This has also been the behaviour in ha for `parenteral nutrition' and `enteral nutrition' (and their corresponding acronyms) in test dataset, since these entities have not been tagged. We asked the organizers about this and other entities for which we had doubts, either because the AnnotGuide didn't cover their cases or because the ha didn't match the recommendations in the AnnotGuide. Woefully, communication with the organizers has not been very fluent on this respect. All in all, this bad decision on the part of the authors amounts for 6 fps (more than 7.5% of our fps according to evaluation script).", + "For other cases, decisions that may be clearly induced from the tagging of train+dev datasets, have not been applied in the test corpus fragment. These include cadenas ligeras (5 times in t+d, 1 fp in test); enzimas hep\u00e1ticas (tagged systematically in t+d, 1 fp); p53 (also tagged in t+d, 1 fp).", + "Another entity that stands out is hidratos de carbono (`carbohydrates'). It is tagged twice in the t+d dataset, occurring 4 times in the set (once as HC). However, although the form carbohidratos has been annotated twice in the test set, hidratos de carbono has been not (1 fp).", + "Moreover, suero (`Sodium chloride solution' or `serum') deserves its own comment. Both entity references are tagged in the train+dev datasets (although with the latter meaning it is tagged only 4 out of 12 occurrences). We decided to tag it due to its relevance. In the test dataset, it occurs 5 times with the blood material meaning, but it has only been tagged twice as such (one of them being an error, since it refers to the former meaning). Our system tagged all occurrences, but tagged also one of the instances with the former meaning as serum (3 fps).", + "Finally, there are some inconsistencies within the same dataset. For example, nutricional agent Kabiven is tagged as both NORMALIZABLES (with sctid) and NO_NORMALIZABLES in the very same text. The same happens with another nutritional complement, Cernebit, this time in two different files. The perfusion solution Isoplasmal G (with a typo in the datasets \u2014Isoplasmar G) is tagged as NORMALIZABLES and UNCLEAR. These examples reveal a vague understanding (or definition) of criteria as regards fluids and nutrition, as we pointed out at the beginning of this section." + ], + [ + "Some of the entities occurring in the test dataset have not always been tagged. This is the case for celulosa (annotated only once but used twice, 1 fp); vimentina (same situation as previous, 1 fp); LDH (tagged 20 times in t+d but not in one of the files, 1 fp); cimetidina (1 fp); reactantes de fase aguda (2 fps; 2 other occurrences were tagged); anticuerpos antinucleares (human annotators missed 1, considered fp)." + ], + [ + "On our refinement work with the system, some incorrect sctids have emerged. These errors impact on subtask 2 (some also on subtask 1). A large sample of them is enumerated below.", + "ARP (`actividad de renina plasm\u00e1tica', `plasma renin activity', PRA) cannot be linked to sctid for renina, which happens twice. In the context `perfil de antigenos [sic] extra\u00edbles del n\u00facleo (ENA)', ENA has been tagged with sctid of the antibody (1 fp). In one of the files, tioflavina is linked to sctid of tioflavina T, but it could be tioflavina S. Thus, it should be NO_NORMALIZABLE. Harvoni is ChEBI:85082 and not (1 fp). AcIgM contra CMV has a wrong sctid (1 fp). HBsAg has no sctid in the test set; it should be 22290004 (`Hepatitis B surface antigen') (1 fp).", + "There are other incorrect annotations, due to inadvertent human errors, like biotina tagged as PROTEINAS or VEB (`Epstein-Barr virus') being annotated when it is not a substance. Among these mismatches between ha and system annotation, the most remarkable is the case of synonyms in active principles. For instance, the brand name drug Dekapine has been linked to `\u00e1cido valproico' in the former case and to `valproato s\u00f3dico' in the latter. These terms are synonymous, but sadly they don't share sctid. Hence, this case also counts as a fp.", + "A gold standard dataset for any task is very hard to develop, so a continuous editing of it is a must. In this discussion, we have focused on false positives (fps) according to the script used for system evaluation, with the main purpose of understanding the domain knowledge encoded in the linguistic conventions (lexical/terminological items and constructions) used by health professionals, but also the decisions underlying both the AnnotGuide and the ha practice.", + "In this journey to system improvement and authors enlightenment, some inconsistencies, errors, omissions have come up, as it has been reflected in this section, so both the guidelines for and the practice of annotation can also be improved in future use scenarios of the clinical case corpus built and maintained by the shared task organizers.", + "Our conclusion on this state of affairs is that some of the inconsistencies spotted in this section show that there were not a rational approach to the annotation of certain entities contained in the datasets (apart from other errors and/or oversights), and, hence, the upper bound of any tagging system is far below the ideal 1.0 F1-score. To this respect, in very many cases, the authors have made the wrong choice, but in others they were guided by analogy or common sense. Maybe a selection founded on probability measures estimated on training material could have obtained better results with this specific test dataset. However, in the end, this cannot be considered as an indication of a better system performance, since, as it has been shown, the test dataset used still needs more refinement work to be used as the right dataset for automatic annotation evaluation." + ], + [ + "With this resource-based system developed for the PharmaCoNER shared task on ner of pharmacological, chemical and biomedical entities, we have demonstrated that, having a very limited knowledge of the domain, and, thus, making wrong choices many times in the creation of resources for the tasks at hand, but being more flexible with the matching mechanisms, a simple-design system can outperform a ner tagger for biomedical entities based on state-of-the-art artificial neural network technology. Thus, knowledge-based methods stand on their own merits in task resolution.", + "But, perhaps most importantly, the other key point brought to light in this contribution is that a resource-based approach also favours a more critical stance on the dataset(s) used to evaluate system performance. With these methods, system development can go hand in hand with dataset refinement in a virtuous circle that let us think that maybe next time we are planning to add a new gazetteer or word embedding to our system in order to try to improve system performance, we should first look at our data and, like King Midas, turn our Human Annotated Dataset into a true Gold Standard Dataset." + ], + [ + "We thank three anonymous reviewers of our manuscript for their careful reading and their many insightful comments and suggestions. We have made our best in providing a revised version of the manuscript that reflects their suggestions. Any remaining errors are our own responsability." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2026/instruction.md b/qasper-2026/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ba1f746683a83118f1fdbc9f896f5343a7b303d4 --- /dev/null +++ b/qasper-2026/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Dialectometric analysis of language variation in Twitter + +Question: What are the characteristics of the rural dialect? \ No newline at end of file diff --git a/qasper-2072/instruction.md b/qasper-2072/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ae2ed7c6650c9daa06fa48a068f615d147356d14 --- /dev/null +++ b/qasper-2072/instruction.md @@ -0,0 +1,126 @@ +Name of Paper: Dependency or Span, End-to-End Uniform Semantic Role Labeling + +Question: what were the baselines? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Background", + "Related Work", + "Overview", + "Token Representation", + "Deep Encoder", + "Predicate and Argument Representation", + "Scorers", + "Training Objective", + "Candidates Pruning", + "SRL Constraints", + "Experiments", + "Datasets", + "Setup", + "End-to-end Results", + "Results with Pre-identified Predicates", + "Ablation", + "Dependency or Span?", + "Conclusion" + ], + "paragraphs": [ + [ + "The purpose of semantic role labeling (SRL) is to derive the meaning representation for a sentence, which is beneficial to a wide range of natural language processing (NLP) tasks BIBREF0 , BIBREF1 . SRL can be formed as four subtasks, including predicate detection, predicate disambiguation, argument identification and argument classification. For argument annotation, there are two formulizations. One is based on text spans, namely span-based SRL. The other is dependency-based SRL, which annotates the syntactic head of argument rather than entire argument span. Figure FIGREF1 shows example annotations.", + "Great progress has been made in syntactic parsing BIBREF2 , BIBREF3 , BIBREF4 . Most traditional SRL methods rely heavily on syntactic features. To alleviate the inconvenience, recent works BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 propose end-to-end models for SRL, putting syntax aside and still achieving favorable results. However, these systems focus on either span or dependency SRL, which motivates us to explore a uniform approach.", + "Both span and dependency are effective formal representations for semantics, though for a long time it has been kept unknown which form, span or dependency, would be better for the convenience and effectiveness of semantic machine learning and later applications. Furthermore, researchers are interested in two forms of SRL models that may benefit from each other rather than their separated development. This topic has been roughly discussed in BIBREF19 , who concluded that the (best) dependency SRL system at then clearly outperformed the span-based (best) system through gold syntactic structure transformation. However, BIBREF19 johansson2008EMNLP like all other traditional SRL models themselves had to adopt rich syntactic features, and their comparison was done between two systems in quite different building styles. Instead, this work will develop full syntax-agnostic SRL systems with the same fashion for both span and dependency representation, so that we can revisit this issue under a more solid empirical basis.", + "In addition, most efforts focus on argument identification and classification since span and dependency SRL corpora have already marked predicate positions. Although no predicate identification is needed, it is not available in many downstream applications. Therefore, predicate identification should be carefully handled in a complete practical SRL system. To address this problem, BIBREF9 he2018jointly proposed an end-to-end approach for jointly predicting predicates and arguments for span SRL. Likewise, BIBREF11 cai2018full introduced an end-to-end model to naturally cover all predicate/argument identification and classification subtasks for dependency SRL.", + "To jointly predict predicates and arguments, we present an end-to-end framework for both span and dependency SRL. Our model extends the span SRL model of BIBREF9 he2018jointly, directly regarding all words in a sentence as possible predicates, considering all spans or words as potential arguments and learning distributions over possible predicates. However, we differ by (1) introducing unified argument representation to handle two different types of SRL tasks, and (2) employing biaffine scorer to make decisions for predicate-argument relationship.", + "The proposed models are evaluated on span SRL datasets: CoNLL 2005 and 2012 data, as well as the dependency SRL dataset of CoNLL 2008 and 2009 shared tasks. For span SRL, our single model outperforms the previous best results by 0.3% and 0.5% F INLINEFORM0 -score on CoNLL 2005 and 2012 test sets respectively. For dependency SRL, we achieve new state-of-the-art of 85.3% F INLINEFORM1 and 90.4% F INLINEFORM2 on CoNLL 2008 and 2009 benchmarks respectively." + ], + [ + "SRL is pioneered by BIBREF20 gildea2002, which uses the PropBank conventions BIBREF21 . Conventionally, span SRL consists of two subtasks, argument identification and classification. The former identifies the arguments of a predicate, and the latter assigns them semantic role labels, namely, determining the relation between arguments and predicates. The PropBank defines a set of semantic roles to label arguments, falling into two categories: core and non-core roles. The core roles (A0-A5 and AA) indicate different semantics in predicate-argument structure, while the non-core roles are modifiers (AM-adj) where adj specifies the adjunct type, such as temporal (AM-TMP) and locative (AM-LOC) adjuncts. For example shown in Figure FIGREF1 , A0 is a proto-agent, representing the borrower.", + "Slightly different from span SRL in argument annotation, dependency SRL labels the syntactic heads of arguments rather than phrasal arguments, which was popularized by CoNLL-2008 and CoNLL-2009 shared tasks BIBREF22 , BIBREF23 . Furthermore, when no predicate is given, two other indispensable subtasks of dependency SRL are predicate identification and disambiguation. One is to identify all predicates in a sentence, and the other is to determine the senses of predicates. As the example shown in Figure FIGREF1 , 01 indicates the first sense from the PropBank sense repository for predicate borrowed in the sentence." + ], + [ + "The traditional approaches on SRL were mostly about designing hand-crafted feature templates and then employ linear classifiers such as BIBREF24 , BIBREF25 , BIBREF12 . Even though neural models were introduced, early work still paid more attention on syntactic features. For example, BIBREF14 Fitzgerald2015 integrated syntactic information into neural networks with embedded lexicalized features, while BIBREF15 roth2016 embedded syntactic dependency paths between predicates and arguments. Similarly, BIBREF16 marcheggianiEMNLP2017 leveraged the graph convolutional network to encode syntax for dependency SRL. Recently, BIBREF17 Strubell2018 presented a multi-task neural model to incorporate auxiliary syntactic information for SRL, BIBREF18 li2018unified adopted several kinds of syntactic encoder for syntax encoding while BIBREF10 he:2018Syntax used syntactic tree for argument pruning.", + "However, using syntax may be quite inconvenient sometimes, recent studies thus have attempted to build SRL systems without or with little syntactic guideline. BIBREF5 zhou-xu2015 proposed the first syntax-agnostic model for span SRL using LSTM sequence labeling, while BIBREF7 he-acl2017 further enhanced their model using highway bidirectional LSTMs with constrained decoding. Later, BIBREF8 selfatt2018 presented a deep attentional neural network for applying self-attention to span SRL task. Likewise for dependency SRL, BIBREF6 marcheggiani2017 proposed a syntax-agnostic model with effective word representation and obtained favorable results. BIBREF11 cai2018full built a full end-to-end model with biaffine attention and outperformed the previous state-of-the-art.", + "More recently, joint predicting both predicates and arguments has attracted extensive interest on account of the importance of predicate identification, including BIBREF7 , BIBREF17 , BIBREF9 , BIBREF11 and this work. In our preliminary experiments, we tried to integrate the self-attention into our model, but it does not provide any significant performance gain on span or dependency SRL, which is not consistent with the conclusion in BIBREF8 and lets us exclude it from this work.", + "Generally, the above work is summarized in Table TABREF2 . Considering motivation, our work is most closely related to the work of BIBREF14 Fitzgerald2015, which also tackles span and dependency SRL in a uniform fashion. The essential difference is that their model employs the syntactic features and takes pre-identified predicates as inputs, while our model puts syntax aside and jointly learns and predicts predicates and arguments." + ], + [ + "Given a sentence INLINEFORM0 , we attempt to predict a set of predicate-argument-relation tuples INLINEFORM1 , where INLINEFORM2 is the set of all possible predicate tokens, INLINEFORM3 includes all the candidate argument spans or dependencies, and INLINEFORM6 is the set of the semantic roles. To simplify the task, we introduce a null label INLINEFORM7 to indicate no relation between arbitrary predicate-argument pair following BIBREF9 he2018jointly. As shown in Figure FIGREF5 , our uniform SRL model includes four main modules:", + " INLINEFORM0 token representation component to build token representation INLINEFORM1 from word INLINEFORM2 ,", + " INLINEFORM0 a BiHLSTM encoder that directly takes sequential inputs,", + " INLINEFORM0 predicate and argument representation module to learn candidate representations,", + " INLINEFORM0 a biaffine scorer which takes the candidate representations as input and predicts semantic roles." + ], + [ + "We follow the bi-directional LSTM-CNN architecture BIBREF26 , where convolutional neural networks (CNNs) encode characters inside a word INLINEFORM0 into character-level representation INLINEFORM1 then concatenated with its word-level INLINEFORM2 into context-independent representation. To further enhance the word representation, we leverage an external representation INLINEFORM3 from pretrained ELMo (Embeddings from Language Models) layers according to BIBREF27 ELMo. Eventually, the resulting token representation is concatenated as DISPLAYFORM0 " + ], + [ + "The encoder in our model adopts the bidirectional LSTM with highway connections (BiHLSTM) to contextualize the representation into task-specific representation: INLINEFORM0 , where the gated highway connections is used to alleviate the vanishing gradient problem when training very deep BiLSTMs." + ], + [ + "We employ contextualized representations for all candidate arguments and predicates. As referred in BIBREF2 , applying a multi-layer perceptron (MLP) to the recurrent output states before the classifier has the advantage of stripping away irrelevant information for the current decision. Therefore, to distinguish the currently considered predicate from its candidate arguments in SRL context, we add an MLP layer to contextualized representations for argument INLINEFORM0 and predicate INLINEFORM1 candidates specific representations respectively with ReLU BIBREF28 as its activation function: INLINEFORM2 INLINEFORM3 ", + "To perform uniform SRL, we introduce unified argument representation. For dependency SRL, we assume single word argument span by limiting the length of candidate argument to be 1, so our model uses the INLINEFORM0 as the final argument representation INLINEFORM1 directly. While for span SRL, we utilize the approach of span representation from BIBREF29 lee2017end. Each candidate span representation INLINEFORM2 is built by DISPLAYFORM0 ", + "where INLINEFORM0 and INLINEFORM1 are boundary representations, INLINEFORM2 indicates a span, INLINEFORM3 is a feature vector encoding the size of span, and INLINEFORM4 is the specific notion of headedness which is learned by attention mechanism BIBREF30 over words in each span (where INLINEFORM5 is the position inside span) as follows : INLINEFORM6 INLINEFORM7 " + ], + [ + "For predicate and arguments, we introduce two unary scores on their candidates: INLINEFORM0 INLINEFORM1 ", + "For semantic role, we adopt a relation scorer with biaffine attention BIBREF2 : DISPLAYFORM0 ", + " where INLINEFORM0 and INLINEFORM1 respectively denote the weight matrix of the bi-linear and the linear terms and INLINEFORM2 is the bias item.", + "The biaffine scorer differs from feed-forward networks scorer in bilinear transformation. Since SRL can be regarded as a classification task, the distribution of classes is uneven and the problem comes worse after the null labels are introduced. The output layer of the model normally includes a bias term designed to capture the prior probability of each class, with the rest of the model focusing on learning the likelihood of every classes occurring in data. The biaffine attention as Dozat and Manning (2017) in our model directly assigns a score for each specific semantic role and would be helpful for semantic role prediction. Actually, (He et al., 2018a) used a scorer as Equation (2), which is only a part of our scorer including both Equations ( EQREF14 ) and (). Therefore, our scorer would be more informative than previous models such as BIBREF9 ." + ], + [ + "The model is trained to optimize the probability INLINEFORM0 of the predicate-argument-relation tuples INLINEFORM1 given the sentence INLINEFORM2 , which can be factorized as: DISPLAYFORM0 ", + " where INLINEFORM0 represents the model parameters, and INLINEFORM1 , is the score for the predicate-argument-relation tuple, including predicate score INLINEFORM2 , argument score INLINEFORM3 and relation score INLINEFORM4 .", + "Our model adopts a biaffine scorer for semantic role label prediction, which is implemented as cross-entropy loss. Moreover, our model is trained to minimize the negative likehood of the golden structure INLINEFORM0 : INLINEFORM1 . The score of null labels are enforced into INLINEFORM2 . For predicates and arguments prediction, we train separated scorers ( INLINEFORM3 and INLINEFORM4 ) in parallel fed to the biaffine scorer for predicate and argument predication respectively, which helps to reduce the chance of error propagation." + ], + [ + "The number of candidate arguments for a sentence of length INLINEFORM0 is INLINEFORM1 for span SRL, and INLINEFORM2 for dependency. As the model deals with INLINEFORM3 possible predicates, the computational complexity is INLINEFORM4 for span, INLINEFORM5 for dependency, which is too computationally expensive. To address this issue, we attempt to prune candidates using two beams for storing the candidate arguments and predicates with size INLINEFORM6 and INLINEFORM7 inspired by BIBREF9 he2018jointly, where INLINEFORM8 and INLINEFORM9 are two manually setting thresholds. First, the predicate and argument candidates are ranked according to their predicted score ( INLINEFORM10 and INLINEFORM11 ) respectively, and then we reduce the predicate and argument candidates with defined beams. Finally, we take the candidates from the beams to participate the label prediction. Such pruning will reduce the overall number of candidate tuples to INLINEFORM12 for both types of tasks. Furthermore, for span SRL, we set the maximum length of candidate arguments to INLINEFORM13 , which may decrease the number of candidate arguments to INLINEFORM14 ." + ], + [ + "According to PropBank semantic convention, predicate-argument structure has to follow a few of global constraints BIBREF25 , BIBREF7 , we thus incorporate constraints on the output structure with a dynamic programing decoder during inference. These constraints are described as follows:", + " INLINEFORM0 Unique core roles (U): Each core role (A0-A5, AA) should appear at most once for each predicate.", + " INLINEFORM0 Continuation roles (C): A continuation role C-X can exist only when its base role X is realized before it.", + " INLINEFORM0 Reference roles (R): A reference role R-X can exist only when its base role X is realized (not necessarily before R-X).", + " INLINEFORM0 Non-overlapping (O): The semantic arguments for the same predicate do not overlap in span SRL.", + "As C and R constraints lead to worse performance in our models from our preliminary experiments, we only enforce U and O constraints on span SRL and U constraints on dependency SRL." + ], + [ + "Our models are evaluated on two PropBank-style SRL tasks: span and dependency. For span SRL, we test model on the common span SRL datasets from CoNLL-2005 BIBREF32 and CoNLL-2012 BIBREF31 shared tasks. For dependency SRL, we experiment on CoNLL 2008 BIBREF22 and 2009 BIBREF23 benchmarks. As for the predicate disambiguation in dependency SRL task, we follow the previous work BIBREF15 .", + "We consider two SRL setups: end-to-end and pre-identified predicates. For the former setup, our system jointly predicts all the predicates and their arguments in one shot, which turns into CoNLL-2008 setting for dependency SRL. In order to compare with previous models, we also report results with pre-identified predicates, where predicates have been beforehand identified in corpora. Therefore, the experimental results fall into two categories: end-to-end results and results with pre-identified predicates." + ], + [ + "CoNLL 2005 and 2012 The CoNLL-2005 shared task focused on verbal predicates only for English. The CoNLL-2005 dataset takes section 2-21 of Wall Street Journal (WSJ) data as training set, and section 24 as development set. The test set consists of section 23 of WSJ for in-domain evaluation together with 3 sections from Brown corpus for out-of-domain evaluation. The larger CoNLL-2012 dataset is extracted from OntoNotes v5.0 corpus, which contains both verbal and nominal predicates.", + "CoNLL 2008 and 2009 CoNLL-2008 and the English part of CoNLL-2009 shared tasks use the same English corpus, which merges two treebanks, PropBank and NomBank. NomBank is a complement to PropBank with similar semantic convention for nominal predicate-argument structure annotation. Besides, the training, development and test splits of English data are identical to that of CoNLL-2005." + ], + [ + "In our experiments, the word embeddings are 300-dimensional GloVe vectors BIBREF33 . The character representations with dimension 8 randomly initialized. In the character CNN, the convolutions have window sizes of 3, 4, and 5, each consisting of 50 filters. Moreover, we use 3 stacked bidirectional LSTMs with 200 dimensional hidden states. The outputs of BiLSTM employs two 300-dimensional MLP layers with the ReLU as activation function. Besides, we use two 150-dimensional hidden MLP layers with ReLU to score predicates and arguments respectively. For candidates pruning, we follow the settings of BIBREF9 he2018jointly, modeling spans up to length INLINEFORM0 for span SRL and INLINEFORM1 for dependency SRL, using INLINEFORM2 for pruning predicates and INLINEFORM3 for pruning arguments.", + "Training Details During training, we use the categorical cross-entropy as objective, with Adam optimizer BIBREF34 initial learning rate 0.001. We apply 0.5 dropout to the word embeddings and character CNN outputs and 0.2 dropout to all hidden layers and feature embeddings. In the LSTMs, we employ variational dropout masks that are shared across timesteps BIBREF35 , with 0.4 dropout rate. All models are trained for up to 600 epochs with batch size 40 on a single NVIDIA GeForce GTX 1080Ti GPU, which occupies 8 GB graphic memory and takes 12 to 36 hours." + ], + [ + "We present all results using the official evaluation script from the CoNLL-2005 and CoNLL-2009 shared tasks, and compare our model with previous state-of-the-art models.", + "Span SRL Table TABREF15 shows results on CoNLL-2005 in-domain (WSJ) and out-of-domain (Brown) test sets, as well as the CoNLL-2012 test set (OntoNotes). The upper part of table presents results from single models. Our model outperforms the previous models with absolute improvements in F INLINEFORM0 -score of 0.3% on CoNLL-2005 benchmark. Besides, our single model performs even much better than all previous ensemble systems.", + "Dependency SRL Table TABREF19 presents the results on CoNLL-2008. J & N (2008b) BIBREF36 was the highest ranked system in CoNLL-2008 shared task. We obtain comparable results with the recent state-of-the-art method BIBREF11 , and our model surpasses the model BIBREF10 by 2% in F INLINEFORM0 -score." + ], + [ + "To compare with to previous systems with pre-identified predicates, we report results from our models as well.", + "Span SRL Table TABREF22 shows that our model outperforms all published systems, even the ensemble model BIBREF8 , achieving the best results of 87.7%, 80.5% and 86.0% in F INLINEFORM0 -score respectively.", + "Dependency SRL Table TABREF29 compares the results of dependency SRL on CoNLL-2009 English data. Our single model gives a new state-of-the-art result of 90.4% F INLINEFORM0 on WSJ. For Brown data, the proposed syntax-agnostic model yields a performance gain of 1.7% F INLINEFORM1 over the syntax-aware model BIBREF18 ." + ], + [ + "To investigate the contributions of ELMo representations and biaffine scorer in our end-to-end model, we conduct a series of ablation studies on the CoNLL-2005 and CoNLL-2008 WSJ test sets, unless otherwise stated.", + "Table TABREF31 compares F INLINEFORM0 scores of BIBREF9 he2018jointly and our model without ELMo representations. We observe that effect of ELMo is somewhat surprising, where removal of the ELMo dramatically declines the performance by 3.3-3.5 F INLINEFORM1 on CoNLL-2005 WSJ. However, our model gives quite stable performance for dependency SRL regardless of whether ELMo is concatenated or not. The results indicate that ELMo is more beneficial to span SRL.", + "In order to better understand how the biaffine scorer influences our model performance, we train our model with different scoring functions. To ensure a fair comparison with the model BIBREF9 , we replace the biaffine scorer with their scoring functions implemented with feed-forward networks, and the results of removing biaffine scorer are also presented in Table TABREF31 . We can see 0.5% and 1.6% F INLINEFORM0 performance degradation on CoNLL 2005 and 2008 WSJ respectively. The comparison shows that the biaffine scorer is more effective for scoring the relations between predicates and arguments. Furthermore, these results show that biaffine attention mechanism is applicable to span SRL." + ], + [ + "It is very hard to say which style of semantic formal representation, dependency or span, would be more convenient for machine learning as they adopt incomparable evaluation metric. Recent researches BIBREF37 have proposed to learn semantic parsers from multiple datasets in Framenet style semantics, while our goal is to compare the quality of different models in the span and dependency SRL for Propbank style semantics. Following BIBREF19 johansson2008EMNLP, we choose to directly compare their performance in terms of dependency-style metric through a transformation way. Using the head-finding algorithm in BIBREF19 which used gold-standard syntax, we may determine a set of head nodes for each span. This process will output an upper bound performance measure about the span conversion due to the use of gold syntax.", + "We do not train new models for the conversion and the resulted comparison. Instead, we do the job on span-style CoNLL 2005 test set and dependency-style CoNLL 2009 test set (WSJ and Brown), considering these two test sets share the same text content. As the former only contains verbal predicate-argument structures, for the latter, we discard all nomial predicate-argument related results and predicate disambiguation results during performance statistics. Table TABREF33 shows the comparison.", + "On a more strict setting basis, the results from our same model for span and dependency SRL verify the same conclusion of BIBREF19 johansson2008EMNLP, namely, dependency form is in a favor of machine learning effectiveness for SRL even compared to the conversion upper bound of span form." + ], + [ + "This paper presents an end-to-end neural model for both span and dependency SRL, which may jointly learn and predict all predicates and arguments. We extend existing model and introduce unified argument representation with biaffine scorer to the uniform SRL for both span and dependency representation forms. Our model achieves new state-of-the-art results on the CoNLL 2005, 2012 and CoNLL 2008, 2009 benchmarks. Our results show that span and dependency SRL can be effectively handled in a uniform fashion, which for the first time enables us to conveniently explore the useful connection between two types of semantic representation forms." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2075/instruction.md b/qasper-2075/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..74041ab920ac00c54817346387394a24bd917df7 --- /dev/null +++ b/qasper-2075/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Learning to Detect Opinion Snippet for Aspect-Based Sentiment Analysis + +Question: Is the accuracy of the opinion snippet detection subtask reported? \ No newline at end of file diff --git a/qasper-2086/instruction.md b/qasper-2086/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..afe8bb131a73a0505035951165035933a284f623 --- /dev/null +++ b/qasper-2086/instruction.md @@ -0,0 +1,78 @@ +Name of Paper: Improving Slot Filling by Utilizing Contextual Information + +Question: How does their model utilize contextual information for each work in the given sentence in a multi-task setting? setting? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Model", + "Model ::: Slot Filling", + "Model ::: Consistency with Contextual Representation", + "Model ::: Prediction by Contextual Information", + "Model ::: Prediction by Contextual Information ::: Predicting Word Label", + "Model ::: Prediction by Contextual Information ::: Predicting Sentence Labels", + "Experiments", + "Conclusion & Future Work" + ], + "paragraphs": [ + [ + "Slot Filling (SF) is the task of identifying the semantic concept expressed in natural language utterance. For instance, consider a request to edit an image expressed in natural language: \u201cRemove the blue ball on the table and change the color of the wall to brown\u201d. Here, the user asks for an \"Action\" (i.e., removing) on one \u201cObject\u201d (blue ball on the table) in the image and changing an \u201cAttribute\u201d (i.e., color) of the image to new \u201cValue\u201d (i.e., brown). Our goal in SF is to provide a sequence of labels for the given sentence to identify the semantic concept expressed in the given sentence.", + "Prior work have shown that contextual information could be useful for SF. They utilize contextual information either in word level representation (i.e., via contextualize embedding e.g., BERT BIBREF0) or in the model computation graph (e.g., concatenating the context feature to the word feature BIBREF1). However, such methods fail to capture the explicit dependence between the context of the word and its label. Moreover, such limited use of contextual information (i.e., concatenation of the feature vector and context vector) in the model cannot model the interaction between the word representation and its context. In order to alleviate these issues, in this work, we propose a novel model to explicitly increase the predictability of the word label using its context and increasing the interactivity between word representations and its context. More specifically, in our model we use the context of the word to predict its label and by doing so our model learns label-aware context for each word in the sentence. In order to improve the interactivity between the word representation and its context, we increase the mutual information between the word representations and its context. In addition to these contributions, we also propose an auxiliary task to predict which labels are expressed in a given sentence. Our model is trained in a mutli-tasking framework. Our experiments on a SF dataset for identifying semantic concepts from natural language request to edit an image show the superiority of our model compared to previous baselines. Our model achieves the state-of-the-art results on the benchmark dataset by improving the F1 score by almost 2%, which corresponds to a 12.3% error rate reduction." + ], + [ + "The task of Slot Filling is formulated as a sequence labeling problem. Deep learning has been extensively employed for this task (BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11). The prior work has mainly utilized the recurrent neural network as the encoder to extract features per word and Conditional Random Field (CRF) BIBREF12 as the decoder to generate the labels per word. Recently the work BIBREF1 shows that the global context of the sentence could be useful to enhance the performance of neural sequence labeling. In their approach, they use a separate sequential model to extract word features. Afterwards, using max pooling over the representations of the words, they obtain the sentence representations and concatenate it to the word embedding as the input to the main task encoder (i.e. the RNN model to perform sequence labeling). The benefit of using the global context along the word representation is 2-fold: 1) it enhance the representations of the word by the semantics of the entire sentence thus the word representation are more contextualized 2) The global view of the sentence would increase the model performance as it contains information about the entire sentence and this information might not be encoded in word representations due to long decencies.", + "However, the simple concatenation of the global context and the word embeddings would not separately ensure these two benefits of the global context. In order to address this problem, we introduce a multi-task setting to separately ensure the aforementioned benefits of utilizing contextual information. In particular, to ensure the better contextualized representations of the words, the model is encourage to learn representations for the word which are consistent with its context. This is achieved via increasing the mutual information between the word representation and its context. To ensure the usefulness of the contextual information for the final task, we introduce two novel sub-tasks. The first one aims to employ the context of the word instead of the word representation to predict the label of the word. In the second sub-task, we use the global representation of the sentence to predict which labels exist in the given sentence in a multi-label classification setting. These two sub-tasks would encourage the contextual representations to be informative for both word level classification and sentence level classification." + ], + [ + "Our model is trained in a multi-task setting in which the main task is slot filling to identify the best possible sequence of labels for the given sentence. In the first auxiliary task we aim to increase consistency between the word representation and its context. The second auxiliary task is to enhance task specific information in contextual information. In this section, we explain each of these tasks in more details." + ], + [ + "The input to the model is a sequence of words $x_1,x_2,...,x_N$. The goal is to assign each word one of the labels action, object, attribute, value or other. Following other methods for sequence labelling, we use the BIO encoding schema. In addition to the sequence of words, the part-of-speech (POS) tags and the dependency parse tree of the input are given to the model.", + "The input word $x_i$ is represented by the concatenation of its pre-trained word embedding and its POS tag embedding, denoted by $e_i$. These representations are further abstracted using a 2-layer Bi-Directional Long Short-Term Memory (LSTM) to obtain feature vector $h_i$. We use the dependency tree of the sentence to utilize the syntactical information about the input text. This information could be useful to identify the important words and their dependents in the sentence. In order to model the syntactic tree, we utilize Graph Convolutional Network (GCN) BIBREF13 over the dependency tree. This model learns the contextualized representations of the words such that the representation of each word is contextualized by its neighbors. We employ 2-layer GCN with $h_i$ as the initial representation for the node (i.e., word) $i$th. The representations of the $i$th node is an aggregation of the representations of its neighbors. Formally the hidden representations of the $i$th word in $l$th layer of GCN is obtained by:", + "where $N(i)$ is the neighbors of the $i$th word in the dependency tree, $W_l$ is the weight matrix in $l$th layer and $deg(i)$ is the degree of the $i$th word in the dependency tree. The biases are omitted for brevity. The final representations of the GCN for $i$th word, $\\hat{h}_i$, represent the structural features for that word. Afterwards, we concatenate the structural features $\\hat{h}_i$ and sequential features $h_i$ to represent $i$th word by feature vector $h^{\\prime }_i$:", + "Finally in order to label each word in the sentence we employ a task specific 2-layer feed forward neural net followed by a logistic regression model to generate class scores $S_i$ for each word:", + "where $W_{LR}, W_1$ and $W_2$ are trainable parameters and $S_i$ is a vector of size number of classes in which each dimension of it is the score for the corresponding class. Since the main task is sequence labeling we exploit Conditional Random Field (CRF) as the final layer to predict the sequence of labels for the given sentence. More specifically, class scores $S_i$ are fed into the CRF layer as emission scores to obtain the final labeling score:", + "where $T$ is the trainable transition matrix and $\\theta $ is the parameters of the model to generate emission scores $S_i$. Viterbi loss $L_{VB}$ is used as the final loss function to be optimized during training. In the inference time, the Viterbi decoder is employed to find the sequence of labels with highest score." + ], + [ + "In this sub-task we aim to increase the consistency of the word representation and its context. To obtain the context of each word we perform max pooling over the all words of the sentence excluding the word itself:", + "where $h_i$ is the representation of the $i$th word from the Bi-LSTM. We aim to increase the consistency between vectors $h_i$ and $h^c_i$. One way to achieve this is by decreasing the distance between these two vectors. However, directly enforcing the word representation and its context to be close to each other would not be efficient as in long sentences the context might substantially differs from the word. So in order to make enough room for the model to represent the context of each word while it is consistent with the word representation, we employ an indirect method.", + "We propose to maximize the mutual information (MI) between the word representation and its context in the loss function. In information theory, MI evaluates how much information we know about one random variable if the value of another variable is revealed. Formally, the mutual information between two random variable $X_1$ and $X_2$ is obtained by:", + "Using this definition of MI, we can reformulate the MI equation as KL Divergence between the joint distribution $P_{X_1X_2}=P(X_1,X_2)$ and the product of marginal distributions $P_{X_1\\bigotimes X_2}=P(X_1)P(X_2)$:", + "Based on this understanding of MI, we can see that if the two random variables are dependent then the mutual information between them (i.e. the KL-Divergence in equation DISPLAY_FORM9) would be the highest. Consequently, if the representations $h_i$ and $h^c_i$ are encouraged to have large mutual information, we expect them to share more information. The mutual information would be introduced directly into the loss function for optimization.", + "One issue with this approach is that the computation of the MI for such high dimensional continuous vectors as $h_i$ and $h^c_i$ is prohibitively expensive. In this work, we propose to address this issue by employing the mutual information neural estimation (MINE) in BIBREF14 that seeks to estimate the lower bound of the mutual information between the high dimensional vectors via adversarial training. To this goal, MINE attempts to compute the lower bound of the KL divergence between the joint and marginal distributions of the given high dimensional vectors/variables. In particular, MINE computes the lower bound of the Donsker-Varadhan representation of KL-Divergence:", + "However, recently, it has been shown that other divergence metrics (i.e., the Jensen-Shannon divergence) could also be used for this purpose BIBREF15, BIBREF16, offering simpler methods to compute the lower bound for the MI. Consequently, following such methods, we apply the adversarial approach to obtain the MI lower bound via the binary cross entropy of a variable discriminator. This discriminator differentiates the variables that are sampled from the joint distribution from those that are sampled from product of the marginal distributions. In our case, the two variables are the word representation $h_i$ and context representation $h^c_i$. In order to sample from joint distributions, we simply concatenate $h_i$ and $h^c_i$ (i.e., the positive example). To sample from the product of the marginal distributions, we concatenate the representation $h_i$ with $h^c_j$ where $i\\ne j$ (i.e., the negative example). These samples are fed into a 2-layer feed forward neural network $D$ (i.e., the discriminator) to perform a binary classification (i.e., coming from the joint distribution or the product of the marginal distributions). Finally, we use the following binary cross entropy loss to estimate the mutual information between $h_i$ and $h^c_i$ to add into the overall loss function:", + "where $N$ is the length of the sentence and $[h,h^c_i]$ is the concatenation of the two vectors $h$ and $h^c_i$. This loss is added to the final loss function of the model." + ], + [ + "In addition to increasing consistency between the word representation and its context representation, we aim to increase the task specific information in contextual representations. This is desirable as the main task is utilizing the word representation to predict its label. Since our model enforce the consistency between the word representation and its context, increasing the task specific information in contextual representations would help the model's final performance.", + "In order to increase task-specific information in contextual representation, we train the model on two auxiliary tasks. The first one aims to use the context of each word to predict the label of that word and the goal of the second auxiliary task is to use the global context information to predict sentence level labels. We describe each of these tasks in more details in the following sections." + ], + [ + "In this sub-task we use the context representations of each word to predict its label. It will increase the information encoded in the context of the word about the label of the word. We use the same context vector $h^c_i$ for the $i$th word as described in the previous section. This vector is fed into a 2-layer feed forward neural network with a softmax layer at the end to output the probabilities for each class:", + "Where $W_2$ and $W_1$ are trainable parameters. Biases are omitted for brevity. Finally we use the following cross-entropy loss function to be optimized during training:", + "where $N$ is the length of the sentence and $l_i$ is the label of the $i$th word." + ], + [ + "The word label prediction enforces the context of each word to contain information about its label but it would not ensure the contextual information to capture the sentence level patterns for expressing intent. In other words, the word level prediction lacks a general view about the entire sentence. In order to increase the general information about the sentence in the representation of the words, we aim to predict the labels existing in a sentence from the representations of its words. More specifically, we introduce a new sub-task to predict which labels exit in the given sentence (Note that sentences might have only a subset of the labels; e.g. only action and object). We formulate this task as a multi-class classification problem. Formally, given the sentence $X=x_1,x_2,...,x_N$ and label set $S=\\lbrace action, attribute, object, value\\rbrace $ our goal is to predict the vector $L^s=l^s_1,l^s_2,...,l^s_{|S|}$ where $l^s_i$ is one if the sentence $X$ contains $i$th label from the label set $S$ otherwise it is zero.", + "First, we find representation of the sentence from the word representations. To this end, we use max pooling over all words of the sentence to obtain vector $H$:", + "Afterwards, the vector $H$ is further abstracted by a 2-layer feed forward neural net with a sigmoid function at the end:", + "where $W_2$ and $W_1$ are trainable parameters. Note that since this tasks is a multi-class classification the number of neurons at the final layer is equal to $|S|$. We optimize the following binary cross entropy loss function:", + "where $l_k$ is one if the sentence contains the $k$th label otherwise it is zero. Finally, to train the model we optimize the following loss function:", + "where $\\alpha $, $\\beta $ and $\\gamma $ are hyper parameters to be tuned using development set performance." + ], + [ + "In our experiments, we use Onsei Intent Slot dataset. Table TABREF21 shows the statics of this dataset. We use the following hyper parameters in our model: We set the word embedding and POS embedding to 768 and 30 respectively; The pre-trained BERT BIBREF17 embedding are used to initialize word embeddings; The hidden dimension of the Bi-LSTM, GCN and feed forward networks are 200; the hyper parameters $\\alpha $, $\\beta $ and $\\gamma $ are all set to 0.1; We use Adam optimizer with learning rate 0.003 to train the model. We use micro-averaged F1 score on all labels as the evaluation metric.", + "We compare our method with the models trained using Adobe internal NLU tool, Pytext BIBREF18 and Rasa BIBREF19 NLU tools. Table TABREF22 shows the results on Test set. Our model improves the F1 score by almost 2%, which corresponds to a 12.3% error rate reduction. This improvements proves the effectiveness of using contextual information for the task of slot filling.", + "In order to analyze the contribution of the proposed sub-tasks we also evaluate the model when we remove one of the sub-task and retrain the model. The results are reported in Table TABREF23. This table shows that all sub-tasks are required for the model to have its best performance. Among all sub-tasks the word level prediction using the contextual information has the major contribution to the model performance. This fact shows that contextual information trained to be informative about the final sub-task is necessary to obtain the representations which could boost the final model performance." + ], + [ + "In this work we introduce a new deep model for the task of Slot Filling. In a multi-task setting, our model increase the mutual information between word representations and its context, improve the label information in the context and predict which concepts are expressed in the given sentence. Our experiments on an image edit request corpus shows that our model achieves state-of-the-art results on this dataset." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2214/instruction.md b/qasper-2214/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..89ede7d94f056cc41a5817f7eb09b89789dfff75 --- /dev/null +++ b/qasper-2214/instruction.md @@ -0,0 +1,157 @@ +Name of Paper: Build it Break it Fix it for Dialogue Safety: Robustness from Adversarial Human Attack + +Question: What datasets are used? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Related Work", + "Baselines: Wikipedia Toxic Comments", + "Baselines: Wikipedia Toxic Comments ::: Wikipedia Toxic Comments", + "Baselines: Wikipedia Toxic Comments ::: Models", + "Baselines: Wikipedia Toxic Comments ::: Experiments", + "Build it Break it Fix it Method", + "Build it Break it Fix it Method ::: Break it Details ::: Definition of offensive", + "Build it Break it Fix it Method ::: Break it Details ::: Crowderworker Task", + "Build it Break it Fix it Method ::: Break it Details ::: Models to Break", + "Build it Break it Fix it Method ::: Fix it Details", + "Single-Turn Task", + "Single-Turn Task ::: Data Collection ::: Adversarial Collection", + "Single-Turn Task ::: Data Collection ::: Standard Collection", + "Single-Turn Task ::: Data Collection ::: Task Formulation Details", + "Single-Turn Task ::: Data Collection ::: Model Training Details", + "Single-Turn Task ::: Experimental Results", + "Single-Turn Task ::: Experimental Results ::: Break it Phase", + "Single-Turn Task ::: Experimental Results ::: Fix it Phase", + "Multi-Turn Task", + "Multi-Turn Task ::: Task Implementation", + "Multi-Turn Task ::: Models", + "Multi-Turn Task ::: Experimental Results ::: Break it Phase", + "Multi-Turn Task ::: Experimental Results ::: Fix it Phase", + "Conclusion", + "Additional Experimental Results ::: Additional Break It Phase Results", + "Additional Experimental Results ::: Additional Fix It Phase Results", + "Data Collection Interface Details" + ], + "paragraphs": [ + [ + "The detection of offensive language has become an important topic as the online community has grown, as so too have the number of bad actors BIBREF2. Such behavior includes, but is not limited to, trolling in public discussion forums BIBREF3 and via social media BIBREF4, BIBREF5, employing hate speech that expresses prejudice against a particular group, or offensive language specifically targeting an individual. Such actions can be motivated to cause harm from which the bad actor derives enjoyment, despite negative consequences to others BIBREF6. As such, some bad actors go to great lengths to both avoid detection and to achieve their goals BIBREF7. In that context, any attempt to automatically detect this behavior can be expected to be adversarially attacked by looking for weaknesses in the detection system, which currently can easily be exploited as shown in BIBREF8, BIBREF9. A further example, relevant to the natural langauge processing community, is the exploitation of weaknesses in machine learning models that generate text, to force them to emit offensive language. Adversarial attacks on the Tay chatbot led to the developers shutting down the system BIBREF1.", + "In this work, we study the detection of offensive language in dialogue with models that are robust to adversarial attack. We develop an automatic approach to the \u201cBuild it Break it Fix it\u201d strategy originally adopted for writing secure programs BIBREF10, and the \u201cBuild it Break it\u201d approach consequently adapting it for NLP BIBREF11. In the latter work, two teams of researchers, \u201cbuilders\u201d and \u201cbreakers\u201d were used to first create sentiment and semantic role-labeling systems and then construct examples that find their faults. In this work we instead fully automate such an approach using crowdworkers as the humans-in-the-loop, and also apply a fixing stage where models are retrained to improve them. Finally, we repeat the whole build, break, and fix sequence over a number of iterations.", + "We show that such an approach provides more and more robust systems over the fixing iterations. Analysis of the type of data collected in the iterations of the break it phase shows clear distribution changes, moving away from simple use of profanity and other obvious offensive words to utterances that require understanding of world knowledge, figurative language, and use of negation to detect if they are offensive or not. Further, data collected in the context of a dialogue rather than a sentence without context provides more sophisticated attacks. We show that model architectures that use the dialogue context efficiently perform much better than systems that do not, where the latter has been the main focus of existing research BIBREF12, BIBREF5, BIBREF13.", + "Code for our entire build it, break it, fix it algorithm will be made open source, complete with model training code and crowdsourcing interface for humans. Our data and trained models will also be made available for the community." + ], + [ + "The task of detecting offensive language has been studied across a variety of content classes. Perhaps the most commonly studied class is hate speech, but work has also covered bullying, aggression, and toxic comments BIBREF13.", + "To this end, various datasets have been created to benchmark progress in the field. In hate speech detection, recently BIBREF5 compiled and released a dataset of over 24,000 tweets labeled as containing hate speech, offensive language, or neither. The TRAC shared task on Aggression Identification, a dataset of over 15,000 Facebook comments labeled with varying levels of aggression, was released as part of a competition BIBREF14. In order to benchmark toxic comment detection, The Wikipedia Toxic Comments dataset (which we study in this work) was collected and extracted from Wikipedia Talk pages and featured in a Kaggle competition BIBREF12, BIBREF15. Each of these benchmarks examine only single-turn utterances, outside of the context in which the language appeared. In this work we recommend that future systems should move beyond classification of singular utterances and use contextual information to help identify offensive language.", + "Many approaches have been taken to solve these tasks \u2013 from linear regression and SVMs to deep learning BIBREF16. The best performing systems in each of the competitions mentioned above (for aggression and toxic comment classification) used deep learning approaches such as LSTMs and CNNs BIBREF14, BIBREF15. In this work we consider a large-pretrained transformer model which has been shown to perform well on many downstream NLP tasks BIBREF17.", + "The broad class of adversarial training is currently a hot topic in machine learning BIBREF18. Use cases include training image generators BIBREF19 as well as image classifiers to be robust to adversarial examples BIBREF20. These methods find the breaking examples algorithmically, rather than by using humans breakers as we do. Applying the same approaches to NLP tends to be more challenging because, unlike for images, even small changes to a sentence can cause a large change in the meaning of that sentence, which a human can detect but a lower quality model cannot. Nevertheless algorithmic approaches have been attempted, for example in text classification BIBREF21, machine translation BIBREF22, dialogue generation tasks BIBREF23 and reading comprehension BIBREF24. The latter was particularly effective at proposing a more difficult version of the popular SQuAD dataset.", + "As mentioned in the introduction, our approach takes inspiration from \u201cBuild it Break it\u201d approaches which have been successfully tried in other domains BIBREF10, BIBREF11. Those approaches advocate finding faults in systems by having humans look for insecurities (in software) or prediction failures (in models), but do not advocate an automated approach as we do here. Our work is also closely connected to the \u201cMechanical Turker Descent\u201d algorithm detailed in BIBREF25 where language to action pairs were collected from crowdworkers by incentivizing them with a game-with-a-purpose technique: a crowdworker receives a bonus if their contribution results in better models than another crowdworker. We did not gamify our approach in this way, but still our approach has commonalities in the round-based improvement of models through crowdworker interaction." + ], + [ + "In this section we describe the publicly available data that we have used to bootstrap our build it break it fix it approach. We also compare our model choices with existing work and clarify the metrics chosen to report our results." + ], + [ + "The Wikipedia Toxic Comments dataset (WTC) has been collected in a common effort from the Wikimedia Foundation and Jigsaw BIBREF12 to identify personal attacks online. The data has been extracted from the Wikipedia Talk pages, discussion pages where editors can discuss improvements to articles or other Wikipedia pages. We considered the version of the dataset that corresponds to the Kaggle competition: \u201cToxic Comment Classification Challenge\" BIBREF15 which features 7 classes of toxicity: toxic, severe toxic, obscene, threat, insult, identity hate and non-toxic. In the same way as in BIBREF26, every label except non-toxic is grouped into a class offensive while the non-toxic class is kept as the safe class. In order to compare our results to BIBREF26, we similarly split this dataset to dedicate 10% as a test set. 80% are dedicated to train set while the remaining 10% is used for validation. Statistics on the dataset are shown in Table TABREF4." + ], + [ + "We establish baselines using two models. The first one is a binary classifier built on top of a large pre-trained transformer model. We use the same architecture as in BERT BIBREF17. We add a linear layer to the output of the first token ([CLS]) to produce a final binary classification. We initialize the model using the weights provided by BIBREF17 corresponding to \u201cBERT-base\". The transformer is composed of 12 layers with hidden size of 768 and 12 attention heads. We fine-tune the whole network on the classification task. We also compare it the fastText classifier BIBREF27 for which a given sentence is encoded as the average of individual word vectors that are pre-trained on a large corpus issued from Wikipedia. A linear layer is then applied on top to yield a binary classification." + ], + [ + "We compare the two aforementioned models with BIBREF26 who conducted their experiments with a BiLSTM with GloVe pre-trained word vectors BIBREF28. Results are listed in Table TABREF5 and we compare them using the weighted-F1, i.e. the sum of F1 score of each class weighted by their frequency in the dataset. We also report the F1 of the offensive-class which is the metric we favor within this work, although we report both. (Note that throughout the paper, the notation F1 is always referring to offensive-class F1.) Indeed, in the case of an imbalanced dataset such as Wikipedia Toxic Comments where most samples are safe, the weighted-F1 is closer to the F1 score of the safe class while we focus on detecting offensive content. Our BERT-based model outperforms the method from BIBREF26; throughout the rest of the paper, we use the BERT-based architecture in our experiments. In particular, we used this baseline trained on WTC to bootstrap our approach, to be described subsequently." + ], + [ + "In order to train models that are robust to adversarial behavior, we posit that it is crucial collect and train on data that was collected in an adversarial manner. We propose the following automated build it, break it, fix it algorithm:", + "Build it: Build a model capable of detecting offensive messages. This is our best-performing BERT-based model trained on the Wikipedia Toxic Comments dataset described in the previous section. We refer to this model throughout as $A_0$.", + "Break it: Ask crowdworkers to try to \u201cbeat the system\" by submitting messages that our system ($A_0$) marks as safe but that the worker considers to be offensive.", + "Fix it: Train a new model on these collected examples in order to be more robust to these adversarial attacks.", + "Repeat: Repeat, deploying the newly trained model in the break it phase, then fix it again.", + "See Figure FIGREF6 for a visualization of this process." + ], + [ + "Throughout data collection, we characterize offensive messages for users as messages that would not be \u201cok to send in a friendly conversation with someone you just met online.\" We use this specific language in an attempt to capture various classes of content that would be considered unacceptable in a friendly conversation, without imposing our own definitions of what that means. The phrase \u201cwith someone you just met online\" was meant to mimic the setting of a public forum." + ], + [ + "We ask crowdworkers to try to \u201cbeat the system\" by submitting messages that our system marks as safe but that the worker considers to be offensive. For a given round, workers earn a \u201cgame\u201d point each time they are able to \u201cbeat the system,\" or in other words, trick the model by submitting offensive messages that the model marks as safe. Workers earn up to 5 points each round, and have two tries for each point: we allow multiple attempts per point so that workers can get feedback from the models and better understand their weaknesses. The points serve to indicate success to the crowdworker and motivate to achieve high scores, but have no other meaning (e.g. no monetary value as in BIBREF25). More details regarding the user interface and instructions can be found in Appendix SECREF9." + ], + [ + "During round 1, workers try to break the baseline model $A_0$, trained on Wikipedia Toxic Comments. For rounds $i$, $i > 1$, workers must break both the baseline model and the model from the previous \u201cfix it\" round, which we refer to as $A_{i-1}$. In that case, the worker must submit messages that both $A_0$ and $A_{i-1}$ mark as safe but which the worker considers to be offensive." + ], + [ + "During the \u201cfix it\" round, we update the models with the newly collected adversarial data from the \u201cbreak it\" round.", + "The training data consists of all previous rounds of data, so that model $A_i$ is trained on all rounds $n$ for $n \\le i$, as well as the Wikipedia Toxic Comments data. We split each round of data into train, validation, and test partitions. The validation set is used for hyperparameter selection. The test sets are used to measure how robust we are to new adversarial attacks. With increasing round $i$, $A_i$ should become more robust to increasingly complex human adversarial attacks." + ], + [ + "We first consider a single-turn set-up, i.e. detection of offensive language in one utterance, with no dialogue context or conversational history." + ], + [ + "We collected three rounds of data with the build it, break it, fix it algorithm described in the previous section. Each round of data consisted of 1000 examples, leading to 3000 single-turn adversarial examples in total. For the remainder of the paper, we refer to this method of data collection as the adversarial method." + ], + [ + "In addition to the adversarial method, we also collected data in a non-adversarial manner in order to directly compare the two set-ups. In this method \u2013 which we refer to as the standard method, we simply ask crowdworkers to submit messages that they consider to be offensive. There is no model to break. Instructions are otherwise the same.", + "In this set-up, there is no real notion of \u201crounds\", but for the sake of comparison we refer to each subsequent 1000 examples collected in this manner as a \u201cround\". We collect 3000 examples \u2013 or three rounds of data. We refer to a model trained on rounds $n \\le i$ of the standard data as $S_i$." + ], + [ + "Since all of the collected examples are labeled as offensive, to make this task a binary classification problem, we will also add safe examples to it.", + "The \u201csafe data\" is comprised of utterances from the ConvAI2 chit-chat task BIBREF29, BIBREF30 which consists of pairs of humans getting to know each other by discussing their interests. Each utterance we used was reviewed by two independent crowdworkers and labeled as safe, with the same characterization of safe as described before.", + "For each partition (train, validation, test), the final task has a ratio of 9:1 safe to offensive examples, mimicking the division of the Wikipedia Toxic Comments dataset used for training our baseline models. Dataset statistics for the final task can be found in Table TABREF21. We refer to these tasks \u2013 with both safe and offensive examples \u2013 as the adversarial and standard tasks." + ], + [ + "Using the BERT-based model architecture described in Section SECREF3, we trained models on each round of the standard and adversarial tasks, multi-tasking with the Wikipedia Toxic Comments task. We weight the multi-tasking with a mixing parameter which is also tuned on the validation set. Finally, after training weights with the cross entropy loss, we adjust the final bias also using the validation set. We optimize for the sensitive class (i.e. offensive-class) F1 metric on the standard and adversarial validation sets respectively.", + "For each task (standard and adversarial), on round $i$, we train on data from all rounds $n$ for $n \\le i$ and optimize for performance on the validation sets $n \\le i$." + ], + [ + "We conduct experiments comparing the adversarial and standard methods. We break down the results into \u201cbreak it\" results comparing the data collected and \u201cfix it\" results comparing the models obtained." + ], + [ + "Examples obtained from both the adversarial and standard collection methods were found to be clearly offensive, but we note several differences in the distribution of examples from each task, shown in Table TABREF21. First, examples from the standard task tend to contain more profanity. Using a list of common English obscenities and otherwise bad words, in Table TABREF21 we calculate the percentage of examples in each task containing such obscenities, and see that the standard examples contain at least seven times as many as each round of the adversarial task. Additionally, in previous works, authors have observed that classifiers struggle with negations BIBREF8. This is borne out by our data: examples from the single-turn adversarial task more often contain the token \u201cnot\" than examples from the standard task, indicating that users are easily able to fool the classifier with negations.", + "We also anecdotally see figurative language such as \u201csnakes hiding in the grass\u201d in the adversarial data, which contain no individually offensive words, the offensive nature is captured by reading the entire sentence. Other examples require sophisticated world knowledge such as that many cultures consider eating cats to be offensive. To quantify these differences, we performed a blind human annotation of a sample of the data, 100 examples of standard and 100 examples of adversarial round 1. Results are shown in Table TABREF16. Adversarial data was indeed found to contain less profanity, fewer non-profane but offending words (such as \u201cidiot\u201d), more figurative language, and to require more world knowledge.", + "We note that, as anticipated, the task becomes more challenging for the crowdworkers with each round, indicated by the decreasing average scores in Table TABREF27. In round 1, workers are able to get past $A_0$ most of the time \u2013 earning an average score of $4.56$ out of 5 points per round \u2013 showcasing how susceptible this baseline is to adversarial attack despite its relatively strong performance on the Wikipedia Toxic Comments task. By round 3, however, workers struggle to trick the system, earning an average score of only $1.6$ out of 5. A finer-grained assessment of the worker scores can be found in Table TABREF38 in the appendix." + ], + [ + "Results comparing the performance of models trained on the adversarial ($A_i$) and standard ($S_i$) tasks are summarized in Table TABREF22, with further results in Table TABREF41 in Appendix SECREF40. The adversarially trained models $A_i$ prove to be more robust to adversarial attack: on each round of adversarial testing they outperform standard models $S_i$.", + "Further, note that the adversarial task becomes harder with each subsequent round. In particular, the performance of the standard models $S_i$ rapidly deteriorates between round 1 and round 2 of the adversarial task. This is a clear indication that models need to train on adversarially-collected data to be robust to adversarial behavior.", + "Standard models ($S_i$), trained on the standard data, tend to perform similarly to the adversarial models ($A_i$) as measured on the standard test sets, with the exception of training round 3, in which $A_3$ fails to improve on this task, likely due to being too optimized for adversarial tasks. The standard models $S_i$, on the other hand, are improving with subsequent rounds as they have more training data of the same distribution as the evaluation set. Similarly, our baseline model performs best on its own test set, but other models are not far behind.", + "Finally, we remark that all scores of 0 in Table TABREF22 are by design, as for round $i$ of the adversarial task, both $A_0$ and $A_{i-1}$ classified each example as safe during the `break it' data collection phase." + ], + [ + "In most real-world applications, we find that adversarial behavior occurs in context \u2013 whether it is in the context of a one-on-one conversation, a comment thread, or even an image. In this work we focus on offensive utterances within the context of two-person dialogues. For dialogue safety we posit it is important to move beyond classifying single utterances, as it may be the case that an utterance is entirely innocuous on its own but extremely offensive in the context of the previous dialogue history. For instance, \u201cYes, you should definitely do it!\" is a rather inoffensive message by itself, but most would agree that it is a hurtful response to the question \u201cShould I hurt myself?\"" + ], + [ + "To this end, we collect data by asking crowdworkers to try to \u201cbeat\" our best single-turn classifier (using the model that performed best on rounds 1-3 of the adversarial task, i.e., $A_3$), in addition to our baseline classifier $A_0$. The workers are shown truncated pieces of a conversation from the ConvAI2 chit-chat task, and asked to continue the conversation with offensive responses that our classifier marks as safe. As before, workers have two attempts per conversation to try to get past the classifier and are shown five conversations per round. They are given a score (out of five) at the end of each round indicating the number of times they successfully fooled the classifier.", + "We collected 3000 offensive examples in this manner. As in the single-turn set up, we combine this data with safe examples with a ratio of 9:1 safe to offensive for classifier training. The safe examples are dialogue examples from ConvAI2 for which the responses were reviewed by two independent crowdworkers and labeled as safe, as in the s single-turn task set-up. We refer to this overall task as the multi-turn adversarial task. Dataset statistics are given in Table TABREF30." + ], + [ + "To measure the impact of the context, we train models on this dataset with and without the given context. We use the fastText and the BERT-based model described in Section SECREF3. In addition, we build a BERT-based model variant that splits the last utterance (to be classified) and the rest of the history into two dialogue segments. Each segment is assigned an embedding and the input provided to the transformer is the sum of word embedding and segment embedding, replicating the setup of the Next Sentence Prediction that is used in the training of BERT BIBREF17." + ], + [ + "During data collection, we observed that workers had an easier time bypassing the classifiers than in the single-turn set-up. See Table TABREF27. In the single-turn set-up, the task at hand gets harder with each round \u2013 the average score of the crowdworkers decreases from $4.56$ in round 1 to $1.6$ in round 3. Despite the fact that we are using our best single-turn classifier in the multi-turn set-up ($A_3$), the task becomes easier: the average score per round is $2.89$. This is because the workers are often able to use contextual information to suggest something offensive rather than say something offensive outright. See examples of submitted messages in Table TABREF29. Having context also allows one to express something offensive more efficiently: the messages supplied by workers in the multi-turn setting were significantly shorter on average, see Table TABREF21." + ], + [ + "During training, we multi-tasked the multi-turn adversarial task with the Wikipedia Toxic Comments task as well as the single-turn adversarial and standard tasks. We average the results of our best models from five different training runs. The results of these experiments are given in Table TABREF31.", + "As we observed during the training of our baselines in Section SECREF3, the fastText model architecture is ill-equipped for this task relative to our BERT-based architectures. The fastText model performs worse given the dialogue context (an average of 23.56 offensive-class F1 relative to 37.1) than without, likely because its bag-of-embeddings representation is too simple to take the context into account.", + "We see the opposite with our BERT-based models, indicating that more complex models are able to effectively use the contextual information to detect whether the response is safe or offensive. With the simple BERT-based architecture (that does not split the context and the utterance into separate segments), we observe an average of a 3.7 point increase in offensive-class F1 with the addition of context. When we use segments to separate the context from the utterance we are trying to classify, we observe an average of a 7.4 point increase in offensive-class F1. Thus, it appears that the use of contextual information to identify offensive language is critical to making these systems robust, and improving the model architecture to take account of this has large impact." + ], + [ + "We have presented an approach to build more robust offensive language detection systems in the context of a dialogue. We proposed a build it, break it, fix it, and then repeat strategy, whereby humans attempt to break the models we built, and we use the broken examples to fix the models. We show this results in far more nuanced language than in existing datasets. The adversarial data includes less profanity, which existing classifiers can pick up on, and is instead offensive due to figurative language, negation, and by requiring more world knowledge, which all make current classifiers fail. Similarly, offensive language in the context of a dialogue is also more nuanced than stand-alone offensive utterances. We show that classifiers that learn from these more complex examples are indeed more robust to attack, and that using the dialogue context gives improved performance if the model architecture takes it into account.", + "In this work we considered a binary problem (offensive or safe). Future work could consider classes of offensive language separately BIBREF13, or explore other dialogue tasks, e.g. from social media or forums. Another interesting direction is to explore how our build it, break it, fix it strategy would similarly apply to make neural generative models safe BIBREF31.", + "" + ], + [ + "Additional results regarding the crowdworkers' ability to \u201cbeat\" the classifiers are reported in Table TABREF38. In particular, we report the percent of messages sent by the crowdsource workers that were marked safe and offensive by both $A_0$ and $A_{i-1}$. We note that very infrequently ($<1\\%$ of the time) a message was marked offensive by $A_0$ but safe by $A_{i-1}$, showing that $A_0$ was relatively ineffective at catching adversarial behavior.", + "In Table TABREF39, we report the categorization of examples into classes of offensive language from the blind human annotation of round 1 of the single-turn adversarial and standard data. We observe that in the adversarial set-up, there were fewer examples of bullying language but more examples targeting a protected class." + ], + [ + "We report F1, precision, and recall for the offensive class, as well as weighted-F1 for models $S_i$ and $A_i$ on the single-turn standard and adversarial tasks in Table TABREF41." + ], + [ + "During the adversarial data collection, we asked users to generate a message that \u201c[the user believes] is not ok but that our system marks as ok,\" using the definition of \u201cok\" and \u201cnot ok\" described in the paper (i.e. \u201cok to send in a friendly conversation with someone you just met online\").", + "In order to generate a variety of responses, during the single-turn adversarial collection, we provided users with a topic to base their response on 50% of the time. The topics were pulled from a set of 1365 crowd-sourced open-domain dialogue topics. Example topics include diverse topics such as commuting, Gouda cheese, music festivals, podcasts, bowling, and Arnold Schwarzenegger.", + "Users were able to earn up to five points per round, with two tries for each point (to allow them to get a sense of the models' weaknesses). Users were informed of their score after each message, and provided with bonuses for good effort. The points did not affect the user's compensation, but rather, were provided as a way of gamifying the data collection, as this has been showed to increase data quality BIBREF25.", + "Please see an example image of the chat interface in Figure FIGREF42." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2223/instruction.md b/qasper-2223/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c8cb9c728c35c75ccfd8eaefd66bb33a4dfcdc2b --- /dev/null +++ b/qasper-2223/instruction.md @@ -0,0 +1,102 @@ +Name of Paper: Fully Automated Fact Checking Using External Sources + +Question: Do they report results only on English data? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "The Fact-Checking System", + "External Support Retrieval", + "Text Representation", + "Veracity Prediction", + "Dataset", + "Experimental Setup", + "Evaluation Metrics", + "Results", + "Application to cQA", + "Related Work", + "Conclusions and Future Work", + "Acknowledgments" + ], + "paragraphs": [ + [ + "Recent years have seen the proliferation of deceptive information online. With the increasing necessity to validate the information from the Internet, automatic fact checking has emerged as an important research topic. It is at the core of multiple applications, e.g., discovery of fake news, rumor detection in social media, information verification in question answering systems, detection of information manipulation agents, and assistive technologies for investigative journalism. At the same time, it touches many aspects, such as credibility of users and sources, information veracity, information verification, and linguistic aspects of deceptive language.", + "In this paper, we present an approach to fact-checking with the following design principles: (i) generality, (ii) robustness, (iii) simplicity, (iv) reusability, and (v) strong machine learning modeling. Indeed, the system makes very few assumptions about the task, and looks for supportive information directly on the Web. Our system works fully automatically. It does not use any heavy feature engineering and can be easily used in combination with task-specific approaches as well, as a core subsystem. Finally, it combines the representational strength of recurrent neural networks with kernel-based classification.", + "The system starts with a claim to verify. First, we automatically convert the claim into a query, which we execute against a search engine in order to obtain a list of potentially relevant documents. Then, we take both the snippets and the most relevant sentences in the full text of these Web documents, and we compare them to the claim. The features we use are dense representations of the claim, of the snippets and of related sentences from the Web pages, which we automatically train for the task using Long Short-Term Memory networks (LSTMs). We also use the final hidden layer of the neural network as a task-specific embedding of the claim, together with the Web evidence. We feed all these representations as features, together with pairwise similarities, into a Support Vector Machine (SVM) classifier using an RBF kernel to classify the claim as True or False.", + "Figure FIGREF1 presents a real example from one of the datasets we experiment with. The left-hand side of the figure contains a True example, while the right-hand side shows a False one. We show the original claims from snopes.com, the query generated by our system, and the information retrieved from the Web (most relevant snippet and text selection from the web page). The veracity of the claim can be inferred from the textual information.", + "Our contributions can be summarized as follows:", + "The remainder of this paper is organized as follows. Section SECREF2 introduces our method for fact checking claims using external sources. Section SECREF3 presents our experiments and discusses the results. Section SECREF4 describes an application of our approach to a different dataset and a slightly different task: fact checking in community question answering forums. Section SECREF5 presents related work. Finally, Section SECREF6 concludes and suggests some possible directions for future work." + ], + [ + "Given a claim, our system searches for support information on the Web in order to verify whether the claim is likely to be true. The three steps in this process are (i) external support retrieval, (ii) text representation, and (iii) veracity prediction." + ], + [ + "This step consists of generating a query out of the claim and querying a search engine (here, we experiment with Google and Bing) in order to retrieve supporting documents. Rather than querying the search engine with the full claim (as on average, a claim is two sentences long), we generate a shorter query following the lessons highlighted in BIBREF0 .", + "As we aim to develop a general-purpose fact checking system, we use an approach for query generation that does not incorporate any features that are specific to claim verification (e.g., no temporal indicators).", + "We rank the words by means of tf-idf. We compute the idf values on a 2015 Wikipedia dump and the English Gigaword. BIBREF0 suggested that a good way to perform high-quality search is to only consider the verbs, the nouns and the adjectives in the claim; thus, we exclude all words in the claim that belong to other parts of speech. Moreover, claims often contain named entities (e.g., names of persons, locations, and organizations); hence, we augment the initial query with all the named entities from the claim's text. We use IBM's AlchemyAPI to identify named entities. Ultimately, we generate queries of 5\u201310 tokens, which we execute against a search engine. We then collect the snippets and the URLs in the results, skipping any result that points to a domain that is considered unreliable. Finally, if our query has returned no results, we iteratively relax it by dropping the final tokens one at a time." + ], + [ + "Next, we build the representation of a claim and the corresponding snippets and Web pages. First, we calculate three similarities (a) between the claim and a snippet, or (b) between the claim and a Web page: (i) cosine with tf-idf, (ii) cosine over embeddings, and (iii) containment BIBREF1 . We calculate the embedding of a text as the average of the embeddings of its words; for this, we use pre-trained embeddings from GloVe BIBREF2 . Moreover, as a Web page can be long, we first split it into a set of rolling sentence triplets, then we calculate the similarities between the claim and each triplet, and we take the highest scoring triplet. Finally, as we have up to ten hits from the search engine, we take the maximum and also the average of the three similarities over the snippets and over the Web pages.", + "We further use as features the embeddings of the claim, of the best-scoring snippet, and of the best-scoring sentence triplet from a Web page. We calculate these embeddings (i) as the average of the embeddings of the words in the text, and also (ii) using LSTM encodings, which we train for the task as part of a deep neural network (NN). We also use a task-specific embedding of the claim together with all the above evidence about it, which comes from the last hidden layer of the NN." + ], + [ + "Next, we build classifiers: neural network (NN), support vector machines (SVM), and a combination thereof (SVM+NN).", + "The architecture of our NN is shown on top of Figure FIGREF7 . We have five LSTM sub-networks, one for each of the text sources from two search engines: Claim, Google Web page, Google snippet, Bing Web page, and Bing snippet. The claim is fed into the neural network as-is. As we can have multiple snippets, we only use the best-matching one as described above. Similarly, we only use a single best-matching triple of consecutive sentences from a Web page. We further feed the network with the similarity features described above.", + "All these vectors are concatenated and fully connected to a much more compact hidden layer that captures the task-specific embeddings. This layer is connected to a softmax output unit to classify the claim as true or false. The bottom of Figure FIGREF7 represents the generic architecture of each of the LSTM components. The input text is transformed into a sequence of word embeddings, which is then passed to the bidirectional LSTM layer to obtain a representation for the full sequence.", + "Our second classifier is an SVM with an RBF kernel. The input is the same as for the NN: word embeddings and similarities. However, the word embeddings this time are calculated by averaging rather than using a bi-LSTM.", + "Finally, we combine the SVM with the NN by augmenting the input to the SVM with the values of the units in the hidden layer. This represents a task-specific embedding of the input example, and in our experiments it turned out to be quite helpful. Unlike in the SVM only model, this time we use the bi-LSTM embeddings as an input to the SVM. Ultimately, this yields a combination of deep learning and task-specific embeddings with RBF kernels." + ], + [ + "We used part of the rumor detection dataset created by BIBREF3 . While they analyzed a claim based on a set of potentially related tweets, we focus on the claim itself and on the use of supporting information from the Web.", + "The dataset consists of 992 sets of tweets, 778 of which are generated starting from a claim on snopes.com, which ma2016detecting converted into a query. Another 214 sets of tweets are tweet clusters created by other researchers BIBREF4 , BIBREF5 with no claim behind them. ma2016detecting ignored the claim and did not release it as part of their dataset. We managed to find the original claim for 761 out of the 778 snopes.com-based clusters.", + "Our final dataset consists of 761 claims from snopes.com, which span various domains including politics, local news, and fun facts. Each of the claims is labeled as factually true (34%) or as a false rumor (66%). We further split the data into 509 for training, 132 for development, and 120 for testing. As the original split for the dataset was not publicly available, and as we only used a subset of their data, we had to make a new training and testing split. Note that we ignored the tweets, as we wanted to focus on a complementary source of information: the Web. Moreover, ma2016detecting used manual queries, while we use a fully automatic method. Finally, we augmented the dataset with Web-retrieved snippets, Web pages, and sentence triplets from Web pages." + ], + [ + "We tuned the architecture (i.e., the number of layers and their size) and the hyper-parameters of the neural network on the development dataset. The best configuration uses a bidirectional LSTM with 25 units. It further uses a RMSprop optimizer with 0.001 initial learning rate, L2 regularization with INLINEFORM0 =0.1, and 0.5 dropout after the LSTM layers. The size of the hidden layer is 60 with tanh activations. We use a batch of 32 and we train for 400 epochs.", + "For the SVM model, we merged the development and the training dataset, and we then ran a 5-fold cross-validation with grid-search, looking for the best kernel and its parameters. We ended up selecting an RBF kernel with INLINEFORM0 and INLINEFORM1 0.01." + ], + [ + "The evaluation metrics we use are P (precision), R (recall), and F INLINEFORM0 , which we calculate with respect to the false and to the true claims. We further report AvgR (macro-average recall), AvgF INLINEFORM1 (macro-average F INLINEFORM2 ), and Acc (accuracy)." + ], + [ + "Table TABREF14 shows the results on the test dataset. We can see that both the NN and the SVM models improve over the majority class baseline (all false rumors) by a sizable margin. Moreover, the NN consistently outperforms the SVM by a margin on all measures. Yet, adding the task-specific embeddings from the NN as features of the SVM yields overall improvements over both the SVM and the NN in terms of avgR, avgF INLINEFORM0 , and accuracy. We can see that both the SVM and the NN overpredict the majority class (false claims); however, the combined SVM+NN model is quite balanced between the two classes.", + "Table TABREF22 compares the performance of the SVM with and without task-specific embeddings from the NN, when training on Web pages vs. snippets, returned by Google vs. Bing vs. both. The NN embeddings consistently help the SVM in all cases. Moreover, while the baseline SVM using snippets is slightly better than when using Web pages, there is almost no difference between snippets vs. Web pages when NN embeddings are added to the basic SVM. Finally, gathering external support from either Google or Bing makes practically no difference, and using the results from both together does not yield much further improvement. Thus, (i) the search engines already do a good job at generating relevant snippets, and one does not need to go and download the full Web pages, and (ii) the choice of a given search engine is not an important factor. These are good news for the practicality of our approach.", + "Unfortunately, direct comparison with respect to BIBREF3 is not possible. First, we only use a subset of their examples: 761 out of 993 (see Section SECREF17 ), and we also have a different class distribution. More importantly, they have a very different formulation of the task: for them, the claim is not available as input (in fact, there has never been a claim for 21% of their examples); rather an example consists of a set of tweets retrieved using manually written queries.", + "In contrast, our system is fully automatic and does not use tweets at all. Furthermore, their most important information source is the change in tweets volume over time, which we cannot use. Still, our results are competitive to theirs when they do not use temporal features.", + "To put the results in perspective, we can further try to make an indirect comparison to the very recent paper by BIBREF6 . They also present a model to classify true vs. false claims extracted from snopes.com, by using information extracted from the Web. Their formulation of the task is the same as ours, but our corpora and label distributions are not the same, which makes a direct comparison impossible. Still, we can see that regarding overall classification accuracy they improve a baseline from 73.7% to 84.02% with their best model, i.e., a 39.2% relative error reduction. In our case, we go from 66.7% to 80.0%, i.e., an almost identical 39.9% error reduction. These results are very encouraging, especially given the fact that our model is much simpler than theirs regarding the sources of information used (they model the stance of the text, the reliability of the sources, the language style of the articles, and the temporal footprint)." + ], + [ + "Next, we tested the generality of our approach by applying it to a different setup: fact-checking the answers in community question answering (cQA) forums. As this is a new problem, for which no dataset exists, we created one. We augmented with factuality annotations the cQA dataset from SemEval-2016 Task 3 (CQA-QA-2016) BIBREF7 . Overall, we annotated 249 question\u2013answer, or INLINEFORM0 - INLINEFORM1 , pairs (from 71 threads): 128 factually true and 121 factually false answers.", + "Each question in CQA-QA-2016 has a subject, a body, and meta information: ID, category (e.g., Education, and Moving to Qatar), date and time of posting, user name and ID. We selected only the factual questions such as \u201cWhat is Ooredoo customer service number?\u201d, thus filtering out all (i) socializing, e.g., \u201cWhat was your first car?\u201d, (ii) requests for opinion/advice/guidance, e.g., \u201cWhich is the best bank around??\u201d, and (iii) questions containing multiple sub-questions, e.g., \u201cIs there a land route from Doha to Abudhabi. If yes; how is the road and how long is the journey?\u201d", + "Next, we annotated for veracity the answers to the retained questions. Note that in CQA-QA-2016, each answer has a subject, a body, meta information (answer ID, user name and ID), and a judgment about how well it addresses the question of its thread: Good vs. Potentially Useful vs. Bad . We only annotated the Good answers. We further discarded answers whose factuality was very time-sensitive (e.g., \u201cIt is Friday tomorrow.\u201d, \u201cIt was raining last week.\u201d), or for which the annotators were unsure.", + "We targeted very high quality, and thus we did not use crowdsourcing for the annotation, as pilot annotations showed that the task was very difficult and that it was not possible to guarantee that Turkers would do all the necessary verification, e.g., gathering evidence from trusted sources. Instead, all examples were first annotated independently by four annotators, and then each example was discussed in detail to come up with a final label. We ended up with 249 Good answers to 71 different questions, which we annotated for factuality: 128 Positive and 121 Negative examples. See Table TABREF26 for details.", + "We further split our dataset into 185 INLINEFORM0 \u2013 INLINEFORM1 pairs for training, 31 for development, and 32 for testing, preserving the general positive:negative ratio, and making sure that the questions for the INLINEFORM2 \u2013 INLINEFORM3 pairs did not overlap between the splits. Figure FIGREF23 presents an excerpt of an example from the dataset, with one question and three answers selected from a longer thread. Answer INLINEFORM4 contains false information, while INLINEFORM5 and INLINEFORM6 are true, as can be checked on an official governmental website.", + "We had to fit our system for this problem, as here we do not have claims, but a question and an answer. So, we constructed the query from the concatenation of INLINEFORM0 and INLINEFORM1 . Moreover, as Google and Bing performed similarly, we only report results using Google. We limited our run to snippets only, as we have found them rich enough above (see Section SECREF3 ). Also, we had a list of reputed and Qatar-related sources for the domain, and we limited our results to these sources only. This time, we had more options to calculate similarities compared to the rumors dataset: we can compare against INLINEFORM2 , INLINEFORM3 , and INLINEFORM4 \u2013 INLINEFORM5 ; we chose to go with the latter. For the LSTM representations, we use both the question and the answer.", + "Table TABREF27 shows the results on the cQA dataset. Once again, our models outperformed all baselines by a margin. This time, the predictions of all models are balanced between the two classes, which is probably due to the dataset being well balanced in general. The SVM model performs better than the NN by itself. This is due to the fact that the cQA dataset is significantly smaller than the rumor detection one. Thus, the neural network could not be trained effectively by itself. Nevertheless, the task-specific representations were useful and combining them with the SVM model yielded consistent improvements on all the measures once again." + ], + [ + "Journalists, online users, and researchers are well aware of the proliferation of false information on the Web, and topics such as information credibility and fact checking are becoming increasingly important as research directions. For example, there was a recent 2016 special issue of the ACM Transactions on Information Systems journal on Trust and Veracity of Information in Social Media BIBREF9 , there was a SemEval-2017 shared task on Rumor Detection BIBREF10 , and there is an upcoming lab at CLEF-2018 on Automatic Identification and Verification of Claims in Political Debates BIBREF11 .", + "The credibility of contents on the Web has been questioned by researches for a long time. While in the early days the main research focus was on online news portals BIBREF12 , BIBREF13 , BIBREF14 , the interest has eventually shifted towards social media BIBREF4 , BIBREF15 , BIBREF6 , BIBREF16 , which are abundant in sophisticated malicious users such as opinion manipulation trolls, paid BIBREF17 or just perceived BIBREF18 , BIBREF19 , sockpuppets BIBREF20 , Internet water army BIBREF21 , and seminar users BIBREF22 .", + "For instance, BIBREF23 studied the credibility of Twitter accounts (as opposed to tweet posts), and found that both the topical content of information sources and social network structure affect source credibility. Other work, closer to ours, aims at addressing credibility assessment of rumors on Twitter as a problem of finding false information about a newsworthy event BIBREF4 . This model considers user reputation, writing style, and various time-based features, among others.", + "Other efforts have focused on news communities. For example, several truth discovery algorithms are combined in an ensemble method for veracity estimation in the VERA system BIBREF24 . They proposed a platform for end-to-end truth discovery from the Web: extracting unstructured information from multiple sources, combining information about single claims, running an ensemble of algorithms, and visualizing and explaining the results. They also explore two different real-world application scenarios for their system: fact checking for crisis situations and evaluation of trustworthiness of a rumor. However, the input to their model is structured data, while here we are interested in unstructured text as input.", + "Similarly, the task defined by BIBREF25 combines three objectives: assessing the credibility of a set of posted articles, estimating the trustworthiness of sources, and predicting user's expertise. They considered a manifold of features characterizing language, topics and Web-specific statistics (e.g., review ratings) on top of a continuous conditional random fields model. In follow-up work, BIBREF26 proposed a model to support or refute claims from snopes.com and Wikipedia by considering supporting information gathered from the Web. They used the same task formulation for claims as we do, but different datasets. In yet another follow-up work, Popat:2017:TLE:3041021.3055133 proposed a complex model that considers stance, source reliability, language style, and temporal information.", + "Our approach to fact checking is related: we verify facts on the Web. However, we use a much simpler and feature-light system, and a different machine learning model. Yet, our model performs very similarly to this latter work (even though a direct comparison is not possible as the datasets differ), which is a remarkable achievement given the fact that we consider less knowledge sources, we have a conceptually simpler model, and we have six times less training data than Popat:2017:TLE:3041021.3055133.", + "Another important research direction is on using tweets and temporal information for checking the factuality of rumors. For example, BIBREF27 used temporal patterns of rumor dynamics to detect false rumors and to predict their frequency. BIBREF27 focused on detecting false rumors in Twitter using time series. They used the change of social context features over a rumor's life cycle in order to detect rumors at an early stage after they were broadcast.", + "A more general approach for detecting rumors is explored by BIBREF3 , who used recurrent neural networks to learn hidden representations that capture the variation of contextual information of relevant posts over time. Unlike this work, we do not use microblogs, but we query the Web directly in search for evidence. Again, while direct comparison to the work of BIBREF3 is not possible, due to differences in dataset and task formulation, we can say that our framework is competitive when temporal information is not used. More importantly, our approach is orthogonal to theirs in terms of information sources used, and thus, we believe there is potential in combining the two approaches.", + "In the context of question answering, there has been work on assessing the credibility of an answer, e.g., based on intrinsic information BIBREF28 , i.e., without any external resources. In this case, the reliability of an answer is measured by computing the divergence between language models of the question and of the answer. The spawn of community-based question answering Websites also allowed for the use of other kinds of information. Click counts, link analysis (e.g., PageRank), and user votes have been used to assess the quality of a posted answer BIBREF29 , BIBREF30 , BIBREF31 . Nevertheless, these studies address the answers' credibility level just marginally.", + "Efforts to determine the credibility of an answer in order to assess its overall quality required the inclusion of content-based information BIBREF32 , e.g., verbs and adjectives such as suppose and probably, which cast doubt on the answer. Similarly, BIBREF33 used source credibility (e.g., does the document come from a government Website?), sentiment analysis, and answer contradiction compared to other related answers.", + "Overall, credibility assessment for question answering has been mostly modeled at the feature level, with the goal of assessing the quality of the answers. A notable exception is the work of BIBREF34 , where credibility is treated as a task of its own right. Yet, note that credibility is different from factuality (our focus here) as the former is a subjective perception about whether a statement is credible, rather than verifying it as true or false as a matter of fact; still, these notions are often wrongly mixed in the literature. To the best of our knowledge, no previous work has targeted fact-checking of answers in the context of community Question Answering by gathering external support." + ], + [ + "We have presented and evaluated a general-purpose method for fact checking that relies on retrieving supporting information from the Web and comparing it to the claim using machine learning. Our method is lightweight in terms of features and can be very efficient because it shows good performance by only using the snippets provided by the search engines. The combination of the representational power of neural networks with the classification of kernel-based methods has proven to be crucial for making balanced predictions and obtaining good results. Overall, the strong performance of our model across two different fact-checking tasks confirms its generality and potential applicability for different domains and for different fact-checking task formulations.", + "In future work, we plan to test the generality of our approach by applying it to these and other datasets in combination with complementary methods, e.g., those focusing on microblogs and temporal information in Twitter to make predictions about rumors BIBREF27 , BIBREF3 . We also want to explore the possibility of providing justifications for our predictions, and we plan to integrate our method into a real-world application." + ], + [ + "This research was performed by the Arabic Language Technologies group at Qatar Computing Research Institute, HBKU, within the Interactive sYstems for Answer Search project (Iyas)." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2224/instruction.md b/qasper-2224/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..11876b72c664515a796a46d9c85d0d9933a6f197 --- /dev/null +++ b/qasper-2224/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Fully Automated Fact Checking Using External Sources + +Question: Does this system improve on the SOTA? \ No newline at end of file diff --git a/qasper-2240/instruction.md b/qasper-2240/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..92959d5e3912358a8fd72590e8b7dcebea7cd28c --- /dev/null +++ b/qasper-2240/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Assessing the Applicability of Authorship Verification Methods + +Question: Which is the best performing method? \ No newline at end of file diff --git a/qasper-2247/instruction.md b/qasper-2247/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..168f807ff30482079295d900de1b5050c2e87d27 --- /dev/null +++ b/qasper-2247/instruction.md @@ -0,0 +1,73 @@ +Name of Paper: Word frequency and sentiment analysis of twitter messages during Coronavirus pandemic + +Question: Which word frequencies reflect on the psychology of the twitter users, according to the authors? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Word frequency analysis and power law", + "Statistical analysis of tweets", + "Statistical analysis of tweets ::: Twitter id evolution", + "Statistical analysis of tweets ::: Unigram, Bigram an Trigram word frequency analysis", + "Evaluating the Goodness of fit of the model", + "Sentiment Analysis of Twitter Messages", + "Results of Sentiment Analysis of Twitter Messages", + "Conclusion" + ], + "paragraphs": [ + [ + "Catastrophic global circumstances have a pronounced effect on the lives of human beings across the world. The ramifications of such a scenario are experienced in diverse and multiplicative ways spanning routine tasks, media and news reports, detrimental physical and mental health, and also routine conversations. A similar footprint has been left by the global pandemic Coronavirus particularly since February 2020. The outbreak has not only created havoc in the economic conditions, physical health, working conditions, and manufacturing sector to name a few but has also created a niche in the minds of the people worldwide. It has had serious repercussions on the psychological state of the humans that is most evident now.", + "One of the best possible mechanisms of capturing human emotions is to analyze the content they post on social media websites like Twitter and Facebook. Not to be surprised, social media is ablaze with myriad content on Coronavirus reflecting facts, fears, numbers, and the overall thoughts dominating the people's minds at this time.", + "This paper is an effort towards analysis of the present textual content posted by people on social media from a statistical perspective. Two techniques have been deployed to undertake statistical interpretation of text messages posted on twitter; first being word frequency analysis and second sentiment analysis.", + "A well known and profoundly researched as well as used statistical tool for quantitative linguistics is word frequency analysis. Determining word frequencies in any document gives a strong idea about the patterns of word used and the sentimental content of the text. The analysis can be carried out in computational as well as statistical settings. An investigation of the probability distribution of word frequencies extracted from the Twitter text messages posted by different users during the coronavirus outbreak in 2020 has been presented. Power law has been shown to capture patterns in the text analysis BIBREF0, BIBREF1.", + "Sentiment analysis is a technique to gauge the sentimental content of a writing. It can help understand attitudes in a text related to a particular subject. Sentiment analysis is a highly intriguing field of research that can assist in inferring the emotional content of a text. Sentiment analysis has been performed on two datasets, one pertaining to tweets by World Health Organization (WHO) and the other tweets with 1000 retweets. The sentimental quotient from the tweets has been deduced by computing the positive and negative polarities from the messages.", + "The paper has been structured as follows. The next section presents a brief overview of some work in the area of word frequency analysis and emergence of power law. Section 3 details the analysis of the twitter data. Section 4 provides a discussion on the obtained results. Section 5 provides a discussion on the scope of sentiment analysis and outlines the methodology of sentiment analysis adopted in this paper. Section 6 presents the results of the sentiment analysis performed on the two datasets mentioned above. The final section concludes the paper." + ], + [ + "Several researchers have devised statistical and mathematical techniques to analyze literary artifacts. A substantially significant approach among these is inferring the pattern of frequency distributions of the words in the text BIBREF2. Zipf's law is mostly immanent in word frequency distributions BIBREF3, BIBREF4. The law essentially proclaims that for the words' vector $x$, the word frequency distribution $\\nu $ varies as an inverse power of $x$. Some other distributions that are prevalent include Zipf\u2013Mandelbrot BIBREF5, lognormal BIBREF6, BIBREF7, and Gauss\u2013Poisson BIBREF6. Such studies have been conducted for several languages such as Chinese BIBREF8, Japanese BIBREF9, Hindi BIBREF10 and many others BIBREF2. Not only single word frequencies, but also multi-word frequencies have been exhaustively explored. One of the examples is BIBREF11 wherein bigram and trigram frequencies and versatilities were analyzed and 577 different bigrams and 6,140 different trigrams were reported.", + "A well known distribution is the power law. This \u201cnon-normal\" distribution has been a subject of immense interest in the academic community due to its unique nature. The rightly skewed distribution is mathematically represented as follows:", + "where a is a constant and b is the scaling or exponential parameter.", + "Power law has been deployed in various studies. In BIBREF12, the authors explicitly focus on the presence of power law in social networks and use this property to create a degree threshold-based similarity measure that can help in link prediction. In an attempt to model the self similar computer network traffic, the authors claim that during fragmentation of the data into Ethernet frames leads, the power spectrum of the departure process mimics power law similar to that of observed traffic. They also state that the power law was independent of the input process BIBREF13. It was shown in BIBREF14 that internet topologies also can be modelled by power law. While investigating the presence of power laws in information retrieval data, the authors showed that query frequency and 5 out of 24 term frequency distributions could be best fit by a power law BIBREF15. Power law has found immense applications in different domains. In this paper, we intend to use it to model the word frequencies of twitter messages posted during this time." + ], + [ + "In this section, we present the details of the analysis performed on the data obtained pertaining to Twitter messages from January 2020 upto now, that is the time since the news of the Coronavirus outbreak in China was spread across nations. The word frequency data corresponding to the twitter messages has been taken from BIBREF16. The data source indicates that during March 11th to March 30th there were over 4 million tweets a day with the surge in the awareness. Also, the data prominently captures the tweets in English, Spanish, and French languages. A total of four datasets have been used to carry out the study.", + "The first is the data on Twitter Id evolution reflecting on number of tweets and the other three are unigram, bigram and trigram frequencies of words. In the following subsections analysis of each has been undertaken." + ], + [ + "First, we consider the data corresponding to the number of twitter ids tweeting about coronavirus at a particular time. Fig. FIGREF1 depicts the pattern of the twitter id evolution. A couple of peaks can be observed in its evolution in the months of February and March." + ], + [ + "Three forms of tokens of words have been considered for the study viz. unigram, bigram and trigram. These represent the frequencies of one word, two words together and finally three words coupled. The dataset provides the top 1000 unigrams, top 1000 bigrams and the top 1000 trigrams.", + "Fig. FIGREF3 gives the visualization of the word cloud for unigrams. It can be seen that Coronavirus was the most frequent word.", + "Fig. FIGREF6, FIGREF7, and FIGREF8 depict plots for the Rank or Index vs frequency distributions for unigram, bigram and trigram respectively. The graphical visualization well demonstrates that the nature or the pattern of the data follows power law. The power law distribution can be seen to closely fit the data.", + "The exponents in case of unigram and bigrams are -1.273 and -1.375 respectively while for trigram it gives a smaller value of -0.5266. The computed parameters are reported in Table TABREF9. Also, heavy tails are observed specifically in the case of unigrams and bigrams. A good fit by power law connotes a difference in the behaviour of tweet messages when compared to literary documents like novels and poems which are replicated by Zipf's law with exponents close to 1.", + "In the following section, we investigate the goodness of fit of our proposed model using measures: SSE, $R^2$ and RMSE." + ], + [ + "The quality of fit of the data using power law distribution has been evaluated using three goodness of fit metrics: SSE, $R^2$ and RMSE. The value obtained for the three datasets with the three forms of token of words has been shown in Table TABREF9.", + "We obtain a high value of $R^2$ in all the three cases: unigram (0.9172), bigram (0.8718) and trigram (0.9461). Also, the values of SSE and RMSE obtained in all the three cases is quite low. This confirms that power law provides a good model for the frequency data of the tweet messages." + ], + [ + "Sentiment analysis is a fast growing field due to its capacity to interpret emotional quotient of a text.It has often been defined as a computational study that relates to people's attitudes, emotions and opinions towards a subject BIBREF17. The key intent of sentiment analysis is to gauge opinions, identify hidden sentiments and finally to classify their polarity into positive, negative or neutral. Sentiment analysis has been immensely used in customer reviews BIBREF18, news and blogs BIBREF19, and stock market BIBREF20 to name a few. Several methods have been deployed for sentiment analysis including Support Vector Machines BIBREF21, Naive Bayes BIBREF22, and Artificial Neural Networks BIBREF23. There have also been several papers that have provided algorithms for sentiment analysis on twitter posts BIBREF24,BIBREF25, BIBREF26, BIBREF27.", + "In the present work, we use the Python built-in package TextBlob BIBREF28 to perform sentiment analysis of tweets pertaining to the coronavirus outbreak. The analysis has been conducted on two datasets: one corresponds to the tweets made by WHO and the other is the one that contains tweets that have been retweeted more than 1000 times. The polarity values of individual tweets have been computed. The interpretation of these values is as follows:", + "polarity $>$ 0 implies positive;", + "polarity $<$ 0 implies negative;", + "polarity $=$ 0 implies neutral.", + "The polarities range between -1 to 1.", + "These polarity values have then been plotted in a histogram to highlight the overall sentiment (i.e. more positivity or negativity). The plots have been given in Fig. FIGREF11, FIGREF12, FIGREF13, and FIGREF14. Table presents the results of the percentage of positive, negative and neutral tweets based on the polarities in the dataset. The following section outlines an analysis of the results obtained." + ], + [ + "In this section, we provide a detailed discussion of the results obtained from the sentiment analysis of the two datasets.", + "Fig. FIGREF11 corresponds to the histogram of sentiment polarities of tweets on Coronavirus by general public. It can be seen that majority of the tweets have a neutral sentiment followed by positive. The same can be inferred from Table TABREF10 that shows that around 54$\\%$ tweets are neutral, 29$\\%$ positive and a mere 15$\\%$ is negative.", + "Fig. FIGREF12 corresponds to the histogram of sentiment polarities of tweets on Coronavirus by WHO. It can be seen that majority of the tweets have a neutral and positive sentiment. Table TABREF10 that shows that around 60$\\%$ tweets are positive, 24$\\%$ neutral and a mere 15$\\%$ is negative. This shows how WHO is trying to retain the positive spirit through its social media accounts.", + "Fig. FIGREF13 and FIGREF14 represent the histograms produced by removing the neutral tweets. It readily reiterates that the positive emotions in the tweets are higher than negative ones. This is a great sign that indicates that the humans are still focusing more on the positive and sharing the same light through their messages." + ], + [ + "The paper is an effort towards deriving statistical characterization of tweet messages posted during the Coronavirus outbreak. The spread of the disease has created an environment of threat, risks and uncertainty among the population globally. This response can be gauged from the immensely high number of tweet messages corresponding to the outbreak in a short duration of 2-3 months. In the present work, data related to tweets made since January 2020 have been analyzed with two major perspectives: one understanding the word frequency pattern and the second sentiment analysis. The study has resulted in the following observations. The number of twitter ids tweeting about Coronavius has increased rapidly with several peaks during February and March. An empirical analysis of words in the messages was made by determining their frequencies of occurrence. The most frequent words were Coronavirus, Covid19 and Wuhan. Unigram, bigram and trigram frequencies (the top 1000) were modeled. It was seen that all of them followed the rightly skewed power law distribution with quite heavy tails in the first two cases. The exponential parameters for the same were determined to be -1.273 for unigram, -1.375 for bigram and -0.5266 for trigram. The plots for the same have been depicted. The goodness of fit for the model was determined using SSE, $R^2$ and RMSE. The results were found to be satisfactorily good. The model can be used to determine the pattern of the words used during this time. The sentiment analysis was performed on tweet messages by general public and WHO. The polarities of the individual messages were determined and a histogram of the same was plotted. It could be observed that most of the messages belonged to the neutral and positive categories. With WHO messages, $60\\%$ of the messages were found to be positive and with general messages, $29\\%$ were found to be positive and $54\\%$ neutral. In both the cases, just about $15\\%$ messages were of negative sentiment. The results obtained are a great reflection of the sentiments expressed worldwide during this pandemic." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2416/instruction.md b/qasper-2416/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..1dde7a84d0e7ac37cf8ae56b930543be4a8b8746 --- /dev/null +++ b/qasper-2416/instruction.md @@ -0,0 +1,116 @@ +Name of Paper: Combining Acoustics, Content and Interaction Features to Find Hot Spots in Meetings + +Question: Is this approach compared to some baseline? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction and Prior Work", + "Data", + "Data ::: Time Windowing", + "Data ::: Metric", + "Feature Description ::: Acoustic-Prosodic Features", + "Feature Description ::: Word-Based Features ::: Bag of words with TF-IDF", + "Feature Description ::: Word-Based Features ::: Embeddings", + "Feature Description ::: Speaker Activity Features", + "Feature Description ::: Laughter Count", + "Modeling ::: Non-Neural Models", + "Modeling ::: Feed-Forward Neural Networks ::: Pooling Techniques", + "Modeling ::: Feed-Forward Neural Networks ::: Hyperparameters", + "Modeling ::: Model Fusion", + "Experiments", + "Experiments ::: Speech Feature Results", + "Experiments ::: Word-Based Results", + "Experiments ::: Acoustic-Prosodic Feature Results", + "Experiments ::: Fusion Results and Discussion", + "Conclusion", + "Acknowledgments" + ], + "paragraphs": [ + [ + "A definition of the meeting \u201chot spots\u201d was first introduced in BIBREF2, where it was investigated whether human annotators could reliably identify regions in which participants are \u201chighly involved in the discussion\u201d. The motivation was that meetings generally have low information density and are tedious to review verbatim after the fact. An automatic system that could detect regions of high interest (as indicated by the involvement of the participants during the meeting) would thus be useful. Relatedly, automatic meeting summarization could also benefit from such information to give extra weight to hot spot regions in selecting or abstracting material for inclusion in the summary. Later work on the relationship between involvement and summarization BIBREF3 defined a different approach: hot spots are those regions chosen for inclusion in a summary by human annotators (\u201csummarization hot spots\u201d). In the present work we stick with the original \u201cinvolvement hot spot\u201d notion, and refer to such regions simply as \u201chot spots\u201d, regardless of their possible role in summarization. We note that high involvement may be triggered both by a meeting's content (\u201cwhat is being talked about\u201d, and \u201cwhat may be included in a textual summary\u201d), as well as behavioral and social factors, such as a desire to participate, to stake out a position, or to oppose another participant. As related notion in dialog system research is \u201clevel of interest\u201d BIBREF4.", + "The initial research on hot spots focused on the reliability of human annotators and correlations with certain low-level acoustic features, such as pitch BIBREF2. Also investigated were the correlation between hot spots and dialog acts BIBREF5 and hot spots and speaker overlap BIBREF6, without however conducting experiments in automatic hot spot prediction using machine learning techniques. Laskowski BIBREF7 redefined the hot spot annotations in terms of time-based windows over meetings, and investigated various classifier models to detect \u201chotness\u201d (i.e., elevated involvement). However, that work focused on only two types of speech features: presence of laughter and the temporal patterns of speech activity across the various participants, both of which were found to be predictive of involvement.", + "For the related problem of level-of-interest prediction in dialog systems BIBREF8, it was found that content-based classification can also be effective, using both a discriminative TF-IDF model and lexical affect scores, as well as prosodic features. In line with the earlier hot spot research on interaction patterns and speaker overlap, turn-taking features were shown to be helpful for spotting summarization hot spots, in BIBREF3, and even more so than the human involvement annotations. The latter result confirms our intuition that summarization-worthiness and involvement are different notions of \u201chotness\u201d.", + "In this paper, following Laskowski, we focus on the automatic prediction of the speakers' involvement in sliding-time windows/segments. We evaluate machine learning models based on a range of features that can be extracted automatically from audio recordings, either directly via signal processing or via the use of automatic transcriptions (ASR outputs). In particular, we investigate the relative contributions of three classes of information:", + "low-level acoustic-prosodic features, such as those commonly used in other paralinguistic tasks, such as sentiment analysis (extracted using openSMILE BIBREF0);", + "spoken word content, as encoded with a state-of-the-art lexical embedding approach such as BERT BIBREF1;", + "speaker interaction, based on speech activity over time and across different speakers.", + "We attach lower importance to laughter, even though it was found to be highly predictive of involvement in the ICSI corpus, partly because we believe it would not transfer well to more general types of (e.g., business) meetings, and partly because laughter detection is still a hard problem in itself BIBREF9. Generation of speaker-attributed meeting transcriptions, on the other hand, has seen remarkable progress BIBREF10 and could support the features we focus on here." + ], + [ + "The ICSI Meeting Corpus BIBREF11 is a collection of meeting recordings that has been thoroughly annotated, including annotations for involvement hot spots BIBREF12, linguistic utterance units, and word time boundaries based on forced alignment. The dataset is comprised of 75 meetings and about 70 hours of real-time audio duration, with 6 speakers per meeting on average. Most of the participants are well-acquainted and friendly with each other. Hot spots were originally annotated with 8 levels and degrees, ranging from `not hot' to `luke warm' to `hot +'. Every utterance was labeled with one of these discrete labels by a single annotator. Hightened involvement is rare, being marked on only 1% of utterances.", + "Due to the severe imbalance in the label distribution, Laskowski BIBREF13 proposed extending the involvement, or hotness, labels to sliding time windows. In our implementation (details below), this resulted in 21.7% of samples (windows) being labeled as \u201cinvolved\u201d.", + "We split the corpus into three subsets: training, development, and evaluation, keeping meetings intact. Table TABREF4 gives statistics of these partitions.", + "We were concerned with the relatively small number of meetings in the test sets, and repeated several of our experiments with a (jackknifing) cross-validation setup over the training set. The results obtained were very similar to those with the fixed train/test split results that we report here." + ], + [ + "As stated above, the corpus was originally labeled for hot spots at the utterance level, where involvement was marked by either a `b' or a `b+' label. Training and test samples for our experiments correspond to 60 s-long sliding windows, with a 15 s step size. If a certain window, e.g., a segment spanning the times 15 s ...75 s, overlaps with any involved speech utterance, then we label that whole window as `hot'. Fig. FIGREF6 gives a visual representation." + ], + [ + "In spite of the windowing approach, the class distribution is still skewed, and an accuracy metric would reflect the particular class distribution in our data set. Therefore, we adopt the unweighted average recall (UAR) metric commonly used in emotion classification research. UAR is a reweighted accuracy where the samples of both classes are weighted equally in aggregate. UAR thus simulates a uniform class distribution. To match the objective, our classifiers are trained on appropriately weighted training data. Note that chance performance for UAR is by definition 50%, making results more comparable across different data sets." + ], + [ + "Prosody encompasses pitch, energy, and durational features of speech. Prosody is thought to convey emphasis, sentiment, and emotion, all of which are presumably correlated with expressions of involvement. We used the openSMILE toolkit BIBREF0 to compute 988 features as defined by the emobase988 configuration file, operating on the close-talking meeting recordings. This feature set consists of low-level descriptors such as intensity, loudness, Mel-frequency cepstral coefficients, and pitch. For each low-level descriptor, functionals such as max/min value, mean, standard deviation, kurtosis, and skewness are computed. Finally, global mean and variance normalization are applied to each feature, using training set statistics. The feature vector thus captures acoustic-prosodic features aggregated over what are typically utterances. We tried extracting openSMILE features directly from 60 s windows, but found better results by extracting subwindows of 5 s, followed by pooling over the longer 60 s duration. We attribute this to the fact that emobase features are designed to operate on individual utterances, which have durations closer to 5 s than 60 s." + ], + [ + "Initially, we investigated a simple bag-of-words model including all unigrams, bigrams, and trigrams found in the training set. Occurrences of the top 10,000 n-grams were encoded to form a 10,000-dimensional vector, with values weighted according to TD-IDF. TF-IDF weights n-grams according to both their frequency (TF) and their salience (inverse document frequency, IDF) in the data, where each utterance was treated as a separate document. The resulting feature vectors are very sparse." + ], + [ + "The ICSI dataset is too small to train a neural embedding model from scratch. Therefore, it is convenient to use the pre-trained BERT embedding architecture BIBREF1 to create an utterance-level embedding vector for each region of interest. Having been trained on a large text corpus, the resulting embeddings encode semantic similarities among utterances, and would enable generalization from word patterns seen in the ICSI training data to those that have not been observed on that limited corpus.", + "We had previously also created an adapted version of the BERT model, tuned to to perform utterance-level sentiment classification, on a separate dataset BIBREF14. As proposed in BIBREF1, we fine-tuned all layers of the pre-trained BERT model by adding a single fully-connected layer and classifying using only the embedding corresponding to the classification ([CLS]) token prepended to each utterance. The difference in UAR between the hot spot classifiers using the pre-trained embeddings and those using the sentiment-adapted embeddings is small. Since the classifier using embeddings extracted by the sentiment-adapted model yielded slightly better performance, we report all results using these as input.", + "To obtain a single embedding for each 60 s window, we experimented with various approaches of pooling the token and utterance-level embeddings. For our first approach, we ignored the ground-truth utterance segmentation and speaker information. We merged all words spoken within a particular window into a single contiguous span. Following BIBREF1, we added the appropriate classification and separation tokens to the text and selected the embedding corresponding to the [CLS] token as the window-level embedding. Our second approach used the ground-truth segmentation of the dialogue. Each speaker turn was independently modeled, and utterance-level embeddings were extracted using the representation corresponding to the [CLS] token. Utterances that cross window boundaries are truncated using the word timestamps, so only words spoken within the given time window are considered. For all reported experiments, we use L2-norm pooling to form the window-level embeddings for the final classifier, as this performed better than either mean or max pooling." + ], + [ + "These features were a compilation of three different feature types:", + "Speaker overlap percentages: Based on the available word-level times, we computed a 6-dimensional feature vector, where the $i$th index indicates the fraction of time that $i$ or more speakers are talking within a given window. This can be expressed by $\\frac{t_i}{60}$ with $t_i$ indicating the time in seconds that $i$ or more people were speaking at the same time.", + "Unique speaker count: Counts the unique speakers within a window, as a useful metric to track the diversity of participation within a certain window.", + "Turn switch count: Counts the number of times a speaker begins talking within a window. This is a similar metric to the number of utterances. However, unlike utterance count, turn switches can be computed entirely from speech activity, without requiring a linguistic segmentation." + ], + [ + "Laskowski found that laughter is highly predictive of involvement in the ICSI data. Laughter is annotated on an utterance level and falls into two categories: laughter solely on its own (no words) or laughter contained within an utterance (i.e. during speech). The feature is a simple tally of the number of times people laughed within a window. We include it in some of our experiments for comparison purposes, though we do not trust it as general feature. (The participants in the ICSI meetings are far too familiar and at ease with each other to be representative with regards to laughter.)" + ], + [ + "In preliminary experiments, we compared several non-neural classifiers, including logistic regression (LR), random forests, linear support vector machines, and multinomial naive Bayes. Logistic regression gave the best results all around, and we used it exclusively for the results shown here, unless neural networks are used instead." + ], + [ + "For BERT and openSMILE vector classification, we designed two different feed-forward neural network architectures. The sentiment-adapted embeddings described in Section SECREF3 produce one 1024-dimensional vector per utterance. Since all classification operates on time windows, we had to pool over all utterances falling withing a given window, taking care to truncate words falling outside the window. We tested four pooling methods: L2-norm, mean, max, and min, with L2-norm giving the best results.", + "As for the prosodic model, each vector extracted from openSMILE represents a 5 s interval. Since there was both a channel/speaker-axis and a time-axis, we needed to pool over both dimensions in order to have a single vector representing the prosodic features of a 60 s window. The second to last layer is the pooling layer, max-pooling across all the channels, and then mean-pooling over time. The output of the pooling layer is directly fed into the classifier." + ], + [ + "The hyperparameters of the neural networks (hidden layer number and sizes) were also tuned in preliminary experiments. Details are given in Section SECREF5." + ], + [ + "Fig. FIGREF19 depicts the way features from multiple categories are combined. Speech activity and word features are fed directly into a final LR step. Acoustic-prosodic features are first combined in a feed-forward neural classifier, whose output log posteriors are in turn fed into the LR step for fusion. (When using only prosodic features, the ANN outputs are used directly.)" + ], + [ + "We group experiments by the type of feaures they are based on: acoustic-prosodic, word-based, and speech activity, evaluating each group first by itself, and then in combination with others." + ], + [ + "As discussed in Section SECREF3, a multitude of input features were investigated, with some being more discriminative. The most useful speech activity features were speaker overlap percentage, number of unique speakers, and number of turn switches, giving evaluation set UARs of 63.5%, 63.9%, and 66.6%, respectively. When combined the UAR improved to 68.0%, showing that these features are partly complementary." + ], + [ + "The TF-IDF model alone gave a UAR of 59.8%. A drastic increase in performance to 70.5% was found when using the BERT embeddings instead. Therefore we adopted embeddings for all further experiments based on word information.", + "Three different types of embeddings were investigated, i.e. sentiment-adapted embeddings at an utterance-level, unadapted embeddings at the utterance-level, and unadapted embeddings over time windows.", + "The adapted embeddings (on utterances) performed best, indicating that adaptation to sentiment task is useful for involvement classification. It is important to note, however, that the utterance-level embeddings are larger than the window-level embeddings. This is due to there being more utterances than windows in the meeting corpus.", + "The best neural architecture we found for these embeddings is a 5-layer neural network with sizes 1024-64-32-12-2. Other hyperparameters for this model are dropout rate = 0.4, learning rate = $10^{-7}$ and activation function \u201ctanh\u201d. The UAR on the evaluation set with just BERT embeddings as input is 65.2%.", + "Interestingly, the neural model was outperformed by a LR directly on the embedding vectors. Perhaps the neural network requires further fine-tuning, or the neural model is too prone to overfitting, given the small training corpus. In any case, we use LR on embeddings for all subsequent results." + ], + [ + "Our prosodic model is a 5-layer ANN, as described in Section SECREF15. The architecture is: 988-512-128-16-Pool-2. The hyperparameters are: dropout rate = 0.4, learning rate = $10^{-7}$, activation = \u201ctanh\". The UAR on the evaluation set with just openSMILE features is 62.0%." + ], + [ + "Table TABREF24 gives the UAR for each feature subset individually, for all features combined, and for a combination in which one feature subset in turn is left out. The one-feature-set-at-time results suggest that prosody, speech activity and words are of increasing importance in that order. The leave-one-out analysis agrees that the words are the most important (largest drop in accuracy when removed), but on that criterion the prosodic features are more important than speech-activity. The combination of all features is 0.4% absolute better than any other subset, showing that all feature subsets are partly complementary.", + "Fig. FIGREF25 shows the same results in histogram form, but also add those with laughter information. Laughter count by itself is the strongest cue to involvement, as Laskowski BIBREF7 had found. However, even given the strong individual laughter feature, the other features add information, pushing the UAR from from 75.1% to 77.5%." + ], + [ + "We studied detection of areas of high involvement, or \u201chot spots\u201d, within meetings using the ICSI corpus. The features that yielded the best results are in line with our intuitions. Word embeddings, speech activity features such a number of turn changes, and prosodic features are all plausible indicators of high involvement. Furthermore, the feature sets are partly complementary and yield best results when combined using a simple logistic regression model. The combined model achieves 72.6% UAR, or 77.5% with laughter feature.", + "For future work, we would want to see a validation on an independent meeting collection, such as business meetings. Some features, in particular laughter, are bound not be as useful in this case. More data could also enable the training of joint models that perform an early fusion of the different feature types. Also, the present study still relied on human transcripts, and it would be important to know how much UAR suffers with a realistic amount of speech recognition error. Transcription errors are expected to boost the importance of the features types that do not rely on words." + ], + [ + "We thank Britta Wrede, Elizabeth Shriberg and Kornel Laskowski for explanations concerning the details of the data." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2429/instruction.md b/qasper-2429/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..ee8dd01ad599b36da77691982eaacf6aefd9229d --- /dev/null +++ b/qasper-2429/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Limits of Detecting Text Generated by Large-Scale Language Models + +Question: What semantic features help in detecting whether a piece of text is genuine or generated? of \ No newline at end of file diff --git a/qasper-2442/instruction.md b/qasper-2442/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..747c2251061d9df348d70702eecac83d5b00a94b --- /dev/null +++ b/qasper-2442/instruction.md @@ -0,0 +1,244 @@ +Name of Paper: Casting a Wide Net: Robust Extraction of Potentially Idiomatic Expressions + +Question: How big PIE datasets are obtained from dictionaries? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "New Terminology: Potentially Idiomatic Expression (PIE)", + "Related Work", + "Related Work ::: Annotated Corpora and Annotation Schemes for Idioms", + "Related Work ::: Annotated Corpora and Annotation Schemes for Idioms ::: VNC-Tokens", + "Related Work ::: Annotated Corpora and Annotation Schemes for Idioms ::: Gigaword", + "Related Work ::: Annotated Corpora and Annotation Schemes for Idioms ::: IDIX", + "Related Work ::: Annotated Corpora and Annotation Schemes for Idioms ::: SemEval-2013 Task 5b", + "Related Work ::: Annotated Corpora and Annotation Schemes for Idioms ::: General Multiword Expression Corpora", + "Related Work ::: Annotated Corpora and Annotation Schemes for Idioms ::: Overview", + "Related Work ::: Extracting Idioms from Corpora", + "Coverage of Idiom Inventories ::: Background", + "Coverage of Idiom Inventories ::: Selected Idiom Resources (Data and Method)", + "Coverage of Idiom Inventories ::: Method", + "Coverage of Idiom Inventories ::: Results", + "Corpus Annotation", + "Corpus Annotation ::: Evaluating the Extraction Methods", + "Corpus Annotation ::: Base Corpus and Idiom Selection", + "Corpus Annotation ::: Extraction of PIE Candidates", + "Corpus Annotation ::: Annotation Procedure", + "Dictionary-based PIE Extraction", + "Dictionary-based PIE Extraction ::: String-based Extraction Methods ::: Exact String Match", + "Dictionary-based PIE Extraction ::: String-based Extraction Methods ::: Fuzzy String Match", + "Dictionary-based PIE Extraction ::: String-based Extraction Methods ::: Inflectional String Match", + "Dictionary-based PIE Extraction ::: String-based Extraction Methods ::: Additional Steps", + "Dictionary-based PIE Extraction ::: Parser-Based Extraction Methods", + "Dictionary-based PIE Extraction ::: Parser-Based Extraction Methods ::: In-Context Parsing", + "Dictionary-based PIE Extraction ::: Results", + "Dictionary-based PIE Extraction ::: Analysis", + "Conclusions and Outlook" + ], + "paragraphs": [ + [ + "Idiomatic expressions pose a major challenge for a wide range of applications in natural language processing BIBREF0. These include machine translation BIBREF1, BIBREF2, semantic parsing BIBREF3, sentiment analysis BIBREF4, and word sense disambiguation BIBREF5. Idioms show significant syntactic and morphological variability (e.g. beans being spilled for spill the beans), which makes them hard to find automatically. Moreover, their non-compositional nature makes idioms really hard to interpret, because their meaning is often very different from the meanings of the words that make them up. Hence, successful systems need not only be able to recognise idiomatic expressions in text or dialogue, but they also need to give a proper interpretation to them. As a matter of fact, current language technology performs badly on idiom understanding, a phenomenon that perhaps has not received enough attention.", + "Nearly all current language technology used in NLP applications is based on supervised machine learning. This requires large amounts of labelled data. In the case of idiom interpretation, however, only small datasets are available. These contain just a couple of thousand idiom instances, covering only about fifty different types of idiomatic expressions. In fact, existing annotated corpora tend to cover only a small set of idiom types, comprising just a few syntactic patterns (e.g., verb-object combinations), of which a limited number of instances are extracted from a large corpus.", + "This is not surprising as preparing and compiling such corpora involves a large amount of manual extraction work, especially if one wants to allow for form variation in the idiomatic expressions (for example, extracting cooking all the books for cook the books). This work involves both the crafting of syntactic patterns to match potential idiomatic expressions and the filtering of false extractions (non-instances of the target expression e.g. due to wrong parses), and increases with the amount of idiom types included in the corpus (which, in the worst case, means an exponential increase in false extractions). Thus, building a large corpus of idioms, especially one that covers many types in many syntactic constructions, is costly. If a high-precision, high-recall system can be developed for the task of extracting the annotation candidates, this cost will be greatly reduced, making the construction of a large corpus much more feasible.", + "The variability of idioms has been a significant topic of interest among researchers of idioms. For example, BIBREF6 investigates the internal and external modification of a set of idioms in a large English corpus, whereas BIBREF7, quantifies and classifies the variation of a set of idioms in a large corpus of Dutch, setting up a useful taxonomy of variation types. Both find that, although idiomatic expressions mainly occur in their dictionary form, there is a significant minority of idiom instances that occur in non-dictionary variants. Additionally, BIBREF8 show that idiom variants retain their idiomatic meaning more often and are processed more easily than previously assumed. This emphasises the need for corpora covering idiomatic expressions to include these variants, and for tools to be robust in dealing with them.", + "As such, the aim of this article is to describe methods and provide tools for constructing larger corpora annotated with a wider range of idiom types than currently in existence due to the reduced amount of manual labour required. In this way we hope to stimulate further research in this area. In contrast to previous approaches, we want to catch as many idiomatic expressions as possible, and we achieve this by casting a wide net, that is, we consider the widest range of possible idiom variants first and then filter out any bycatch in a way that requires the least manual effort.", + "We expect research will benefit from having larger corpora by improving evaluation quality, by allowing for the training of better supervised systems, and by providing additional linguistic insight into idiomatic expressions. A reliable method for extracting idiomatic expressions is not only needed for building an annotated corpus, but can also be used as part of an automatic idiom processing pipeline. In such a pipeline, extracting potentially idiomatic expressions can be seen as a first step before idiom disambiguation, and the combination of the two modules then functions as an complete idiom extraction system.", + "The main research question that we aim to answer in this article is whether dictionary-based extraction of potentially idiomatic expressions is robust and reliable enough to facilitate the creation of wide-coverage sense-annotated idiom corpora.", + "By answering this question we make several contributions to research on multiword expressions, in particular that of idiom extraction. Firstly, we provide an overview of existing research on annotating idiomatic expressions in corpora, showing that current corpora cover only small sets of idiomatic types (Section SECREF3). Secondly, we quantify the coverage and reliability of a set of idiom dictionaries, demonstrating that there is little overlap between resources (Section SECREF4). Thirdly, we develop and release an evaluation corpus for extracting potentially idiomatic expressions from text (Section SECREF5). Finally, various extraction systems and combinations thereof are implemented, made available to the research community, and evaluated empirically (Section SECREF6)." + ], + [ + "The ambiguity of phrases like wake up and smell the coffee poses a terminological problem. Usually, these phrases are called idiomatic expressions, which is suitable when they are used in an idiomatic sense, but not so much when they are used in a literal sense. Therefore, we propose a new term: potentially idiomatic expressions, or PIEs for short. The term potentially idiomatic expression refers to those expressions which can have an idiomatic meaning, regardless of whether they actually have that meaning in a given context. So, see the light is a PIE in both `After another explanation, I finally saw the light' and `I saw the light of the sun through the trees', while it is an idiomatic expression in the first context, and a literal phrase in the latter context.", + "The processing of PIEs involves three main challenges: the discovery of (new) PIE types, the extraction of instances of known PIE types in text, and the disambiguation of PIE instances in context. Here, we propose calling the discovery task simply PIE discovery, the extraction task simply PIE extraction, and the disambiguation task PIE disambiguation. Note that these terms contrast with the terms used in existing research. There, the discovery task is called type-based idiom detection and the disambiguation task is called token-based idiom detection (cf. BIBREF10, BIBREF11), although this usage is not always consistent. Because these terms are very similar, they are potentially confusing, and that is why we propose novel terminology.", + "Other terminology comes from literature on multiword expressions (MWEs) more generally, i.e. not specific to idioms. Here, the task of finding new MWE types is called MWE discovery and finding instances of known MWE types is called MWE identification BIBREF12. Note, however, that MWE identification generally consists of finding only the idiomatic usages of these types (e.g. BIBREF13). This means that MWE identification consists of both the extraction and disambiguation tasks, performed jointly. In this work, we propose to split this into two separate tasks, and we are concerned only with the PIE extraction part, leaving PIE disambiguation as a separate problem." + ], + [ + "This section is structured so as to reflect the dual contribution of the present work. First, we discuss existing resources annotated for idiomatic expressions. Second, we discuss existing approaches to the automatic extraction of idioms." + ], + [ + "There are four sizeable sense-annotated PIE corpora for English: the VNC-Tokens Dataset BIBREF9, the Gigaword dataset BIBREF14, the IDIX Corpus BIBREF10, and the SemEval-2013 Task 5 dataset BIBREF15. An overview of these corpora is presented in Table TABREF7." + ], + [ + "The VNC-Tokens dataset contains 53 different PIE types. BIBREF9 extract up to 100 instances from the British National Corpus for each type, for a total of 2,984 instances. These types are based on a pre-existing list of verb-noun combinations and were filtered for frequency and whether two idiom dictionaries both listed them. Instances were extracted automatically, by parsing the corpus and selecting all sentences with the right verb and noun in a direct-object relation. It is unclear whether the extracted sentences were manually checked, but no false extractions are mentioned in the paper or present in the dataset.", + "All extracted PIE instances were annotated for sense as either idiomatic, literal or unclear. This is a self-explanatory annotation scheme, but BIBREF9 note that senses are not binary, but can form a continuum. For example, the idiomaticity of have a word in `You have my word' is different from both the literal sense in `The French have a word for this' and the figurative sense in `My manager asked to have a word'. They instructed annotators to choose idiomatic or literal even in ambiguous middle-of-the-continuum cases, and restrict the unclear label only to cases where there is not enough context to disambiguate the meaning of the PIE." + ], + [ + "BIBREF14 present a corpus of 17 PIE types, for which they extracted all instances from the Gigaword corpus BIBREF18, yielding a total of 3,964 instances. BIBREF14 extracted these instances semi-automatically by manually defining all inflectional variants of the verb in the PIE and matching these in the corpus. They did not allow for inflectional variations in non-verb words, nor did they allow intervening words. They annotated these potential idioms as either literal or figurative, excluding ambiguous and unclear instances from the dataset." + ], + [ + "BIBREF10 build on the methodology of BIBREF14, but annotate a larger set of idioms (52 types) and extract all occurrences from the BNC rather than the Gigaword corpus, for a total of 4,022 instances including false extractions. BIBREF10 use a more complex semi-automatic extraction method, which involves parsing the corpus, manually defining the dependency patterns that match the PIE, and extracting all sentences containing those patterns from the corpus. This allows for larger form variations, including intervening words and inflectional variation of all words. In some cases, this yields many non-PIE extractions, as for recharge one's batteries in Example SECREF10. These were not filtered out before annotation, but rather filtered out as part of the annotation process, by having false extraction as an additional annotation label.", + "For sense annotation, they use an extensive tagset, distinguishing literal, non-literal, both, meta-linguistic, embedded, and undecided labels. Here, the both label (Example SECREF10) is used for cases where both senses are present, often as a form of deliberate word play. The meta-linguistic label (Example SECREF10) applies to cases where the PIE instance is used as a linguistic item to discuss, not as part of a sentence. The embedded label (Example SECREF10) applies to cases where the PIE is embedded in a larger figurative context, which makes it impossible to say whether a literal or figurative sense is more applicable. The undecided label is used for unclear and undecidable cases. They take into account the fact that a PIE can have multiple figurative senses, and enumerate these separately as part of the annotation.", + ". These high-performance, rugged tools are claimed to offer the best value for money on the market for the enthusiastic d-i-yer and tradesman, and for the first time offer the possibility of a battery recharging time of just a quarter of an hour. (from IDIX corpus, ID #314)", + ". Left holding the baby, single mothers find it hard to fend for themselves. (from BIBREF10, p.642)", + ". It has long been recognised that expressions such as to pull someone's leg, to have a bee in one's bonnet, to kick the bucket, to cook someone's goose, to be off one's rocker, round the bend, up the creek, etc. are semantically peculiar. (from BIBREF10, p.642)", + ". You're like a restless bird in a cage. When you get out of the cage, you'll fly very high. (from BIBREF10, p.642)", + "The both, meta-linguistic, and embedded labels are useful and linguistically interesting distinctions, although they occur very rarely (0.69%, 0.15%, and an unknown %, respectively). As such, we include these cases in our tagset (see Section SECREF5), but group them under a single label, other, to reduce annotation complexity. We also follow BIBREF10 in that we combine both the PIE/non-PIE annotation and the sense annotation in a single task." + ], + [ + "BIBREF15 created a dataset for SemEval-2013 Task 5b, a task on detecting semantic compositionality in context. They selected 65 PIE types from Wiktionary, and extracted instances from the ukWaC corpus BIBREF17, for a total of 4,350 instances. It is unclear how they extracted the instances, and how much variation was allowed for, although there is some inflectional variation in the dataset. An unspecified amount of manual filtering was done on the extracted instances.", + "The extracted PIE instances were labelled as literal, idiomatic, both, or undecidable. Interestingly, they crowdsourced the sense annotations using CrowdFlower, with high agreement (90%\u201394% pairwise). Undecidable cases and instances on which annotators disagreed were removed from the dataset." + ], + [ + "In addition to the aforementioned idiom corpora, there are also corpora focused on multiword expressions (MWEs) in a more general sense. As idioms are a subcategory of MWEs, these corpora also include some idioms. The most important of these are the PARSEME corpus BIBREF19 and the DiMSUM corpus BIBREF20.", + "DiMSUM provides annotations of over 5,000 MWEs in approximately 90K tokens of English text, consisting of reviews, tweets and TED talks. However, they do not categorise the MWEs into specific types, meaning we cannot easily quantify the number of idioms in the corpus. In contrast to the corpus-specific sense labels seen in other corpora, DiMSUM annotates MWEs with WordNet supersenses, which provide a broad category of meaning for each MWE.", + "Similarly, the PARSEME corpus consists of over 62K MWEs in almost 275K tokens of text across 18 different languages (with the notable exception of English). The main differences with DiMSUM, except for scale and multilingualism, are that it only includes verbal MWEs, and that subcategorisation is performed, including a specific category for idioms. Idioms make up almost a quarter of all verbal MWEs in the corpus, although the proportion varies wildly between languages. In both corpora, MWE annotation was done in an unrestricted manner, i.e. there was not a predefined set of expressions to which annotation was restricted." + ], + [ + "In sum, there is large variation in corpus creation methods, regarding PIE definition, extraction method, annotation schemes, base corpus, and PIE type inventory. Depending on the goal of the corpus, the amount of deviation that is allowed from the PIE's dictionary form to the instances can be very little BIBREF14, to quite a lot BIBREF10. The number of PIE types covered by each corpus is limited, ranging from 17 to 65 types, often limited to one or more syntactic patterns. The extraction of PIE instances is usually done in a semi-automatic manner, by manually defining patterns in a text or parse tree, and doing some manual filtering afterwards. This works well, but an extension to a large number of PIE types (e.g. several hundreds) would also require a large increase in the amount of manual effort involved. Considering the sense annotations done on the PIE corpora, there is significant variation, with BIBREF9 using only three tags, whereas BIBREF10 use six. Outside of PIE-specific corpora there are MWE corpora, which provide a different perspective. A major difference there is that annotation is not restricted to a pre-specified set of expressions, which has not been done for PIEs specifically." + ], + [ + "There are two main approaches to idiom extraction. The first approach aims to distinguish idioms from other multiword phrases, where the main purpose is to expand idiom inventories with rare or novel expressions BIBREF21, BIBREF22, BIBREF23, BIBREF24. The second approach aims to extract all occurrences of a known idiomatic expression in a text. In this paper, we focus on the latter approach. We rely on idiom dictionaries to provide a list of PIE types, and build a system that extracts all instances of those PIE types from a corpus. High-quality idiom dictionaries exist for most well-resourced languages, but their reliability and coverage is not known. As such, we quantify the coverage of dictionaries in Section SECREF4.", + "There is, to the best of our knowledge, no existing work that focuses on dictionary-based PIE extraction. However, there is closely-related work by BIBREF25, who present a system for the dictionary-based extraction of verb-noun combinations (VNCs) in English and Spanish. In their case, the VNCs can be any kind of multiword expression, which they subdivide into literal expressions, collocations, light verb constructions, metaphoric expressions, and idioms. They extract 173 English VNCs and 150 Spanish VNCs and annotate these with both their lexico-semantic MWE type and the amount of morphosyntactic variation they exhibit. BIBREF25 then compare a word sequence-based method, a chunking-based method, and a parse-based method for VNC extraction. Each method relies on the morpho-syntactic information in order to limit false extractions. Precision is evaluated manually on a sample of the extracted VNCs, and recall is estimated by calculating the overlap between the output of the three methods. Evaluation shows that the methods are highly complementary both in recall, since they extract different VNCs, and in precision, since combining the extractors yields fewer false extractions.", + "Whereas BIBREF25 focus on both idiomatic and literal uses of the set of expressions, like in this paper, BIBREF26 tackle only half of that task, namely extracting only literal uses of a given set of VMWEs in Polish. This complicates the task, since it combines extracting all occurrences of the VMWEs and then distinguishing literal from idiomatic uses. Interestingly, they also experiment with models of varying complexity, i.e. just words, part-of-speech tags, and syntactic structures. Their results are hard to put into perspective however, since the frequency of literal VMWEs in their corpus is very rare, whereas corpora containing PIEs tend to show a more balanced distribution.", + "Other similar work to ours also focuses on MWEs more generally, or on different subtypes of MWEs. In addition, these tend to combine both extraction and disambiguation in that they aim to extract only idiomatically used instances of the MWE, without extracting literally used instances or non-instances. Within this line of work, BIBREF27 focuses on verb-particle constructions, BIBREF28 on verbal MWEs (including idioms), and BIBREF29 on verbal MWEs (especially non-canonical variants).", + "Both BIBREF28 and BIBREF29 rely on a pre-defined set of expressions, whereas BIBREF27 also extracts unseen expressions, although based on a pre-defined set of particles and within the vary narrow syntactic frame of verb-particle constructions. The work of BIBREF27 is most similar to ours in that it builds an unsupervised system using existing NLP tools (PoS taggers, chunkers, parsers) and finds that a combination of systems using those tools performs best, as we find in Section SECREF69. BIBREF28 and BIBREF29, by contrast, use supervised classifiers which require training data, not just for the task in general, but specific to the set of expressions used in the task.", + "Although our approach is similar to that of BIBREF25, both in the range of methods used and in the goal of extracting certain multiword expressions regardless of morphosyntactic variation, there are two main differences. First, we use dictionaries, but extract entries automatically and do not manually annotate their type and variability. As a result, our methods rely only on the surface form of the expression taken from the dictionary. Second, we evaluate precision and recall in a more rigorous way, by using an evaluation corpus exhaustively annotated for PIEs. In addition, we do not put any restriction on the syntactic type of the expressions to be extracted, which BIBREF27, BIBREF28, BIBREF25, and BIBREF29 all do." + ], + [ + "Since our goal is developing a dictionary-based system for extracting potentially idiomatic expressions, we need to devise a proper method for evaluating such a system. This is not straightforward, even though the final goal of such a system is simple: it should extract all potentially idiomatic expressions from a corpus and nothing else, regardless of their sense and the form they are used in. The type of system proposed here hence has two aspects that can be evaluated: the dictionary that it is using as a resource for idiomatic expression, and the extractor component that finds idioms in a corpus.", + "The difficulty here is that there is no undisputed and unambiguous definition of what counts as an idiom BIBREF30, as is the case with multiword expressions in general BIBREF12. Of course, a complete set of idiomatic expressions for English (or any other language) is impossible to get due to the broad and ever-changing nature of language. This incompleteness is exacerbated by the ambiguity problem: if we had a clear definition of idiom we could make an attempt of evaluating idiom dictionaries on their accuracy, but it is practically impossible to come up with a definition of idiom that leaves no room for ambiguity. This ambiguity, among others, creates a large grey area between clearly non-idiomatic phrases on the one hand (e.g. buy a house), and clear potentially idiomatic phrases on the other hand (e.g. buy the farm). As a consequence, we cannot empirically evaluate the coverage of the dictionaries. Instead, in this work, we will quantify the divergence between various idiom dictionaries and corpora, with regard to their idiom inventories. If they show large discrepancies, we take that to mean that either there is little agreement on definitions of idiom or the category is so broad that a single resource can only cover a small proportion. Conversely, if there is large agreement, we assume that idiom resources are largely reliable, and that there is consensus around what is, and what is not, an idiomatic expression.", + "We use different idiom resources and assume that the combined set of resources yields an approximation of the true set of idioms in English. A large divergence between the idiom inventories of these resources would then suggest a low recall for a single resource, since many other idioms are present in the other resources. Conversely, if the idiom inventories largely overlap, that indicates that a single resource can already yield decent coverage of idioms in the English language. The results of the dictionary comparisons are in Section SECREF36." + ], + [ + "We evaluate the quality of three idiom dictionaries by comparing them to each other and to three idiom corpora. Before we report on the comparison we first describe why we select and how we prepare these resources. We investigate the following six idiom resources:", + "Wiktionary;", + "the Oxford Dictionary of English Idioms (ODEI, BIBREF31);", + "UsingEnglish.com (UE);", + "the Sporleder corpus BIBREF10;", + "the VNC dataset BIBREF9;", + "and the SemEval-2013 Task 5 dataset BIBREF15.", + "These dictionaries were selected because they are available in digital format. Wiktionary and UsingEnglish have the added benefit of being freely available. However, they are both crowdsourced, which means they lack professional editing. In contrast, ODEI is a traditional dictionary, created and edited by lexicographers, but it has the downside of not being freely available.", + "For Wiktionary, we extracted all idioms from the category `English Idioms' from the English version of Wiktionary. We took the titles of all pages containing a dictionary entry and considered these idioms. Since we focus on multiword idiomatic expressions, we filtered out all single-word entries in this category. More specifically, since Wiktionary is a constantly changing resource, we used the 8,482 idioms retrieved on 10-03-2017, 15:30. We used a similar extraction method for UE, a web page containing freely available resources for ESL learners, including a list of idioms. We extracted all idioms which have publicly available definitions, which numbered 3,727 on 10-03-2017, 15:30. Again, single-word entries and duplicates were filtered out. Concerning ODEI, all idioms from the e-book version were extracted, amounting to 5,911 idioms scraped on 13-03-2017, 10:34. Here we performed an extra processing step to expand idioms containing content in parentheses, such as a tough (or hard) nut (to crack). Using a set of simple expansion rules and some hand-crafted exceptions, we automatically generated all variants for this idiom, with good, but not perfect accuracy. For the example above, the generated variants are: {a tough nut, a tough nut to crack, a hard nut, a hard nut to crack}. The idioms in the VNC dataset are in the form verb_noun, e.g. blow_top, so they were manually expanded to a regular dictionary form, e.g. blow one's top before comparison." + ], + [ + "In many cases, using simple string-match to check overlap in idioms does not work, as exact comparison of idioms misses equivalent idioms that differ only slightly in dictionary form. Differences between resources are caused by, for example:", + "inflectional variation (crossing the Rubicon \u2014 cross the Rubicon);", + "variation in scope (as easy as ABC \u2014 easy as ABC);", + "determiner variation (put the damper on \u2014 put a damper on);", + "spelling variation (mind your p's and q's \u2014 mind your ps and qs);", + "order variation (call off the dogs \u2014 call the dogs off);", + "and different conventions for placeholder words (recharge your batteries \u2014 recharge one's batteries), where both your and one's can generalise to any possessive personal pronoun.", + "These minor variations do not fundamentally change the nature of the idiom, and we should count these types of variation as belonging to the same idiom (see also BIBREF32, who devise a measure to quantify different types of variation allowed by specific MWEs). So, to get a good estimate of the true overlap between idiom resources, these variations need to be accounted for, which we do in our flexible matching approach.", + "There is one other case of variation not listed above, namely lexical variation (e.g. rub someone up the wrong way - stroke someone the wrong way). We do not abstract over this, since we consider lexical variation to be a more fundamental change to the nature of the idiom. That is, a lexical variant is an indicator of the coverage of the dictionary, where the other variations are due to different stylistic conventions and do not indicate actual coverage. In addition, it is easy to abstract over the other types of variation in an NLP application, but this is not the case for lexical variation.", + "The overlap counts are estimated by abstracting over all variations except lexical variation in a semi-automatic manner, using heuristics and manual checking. Potentially overlapping idioms are selected using the following set of heuristics: whether an idiom from one resource is a substring (including gaps) of an idiom in the other resource, whether the words of an idiom form a subset of the words of an idiom in the other resource, and whether there is an idiom in the other resource which has a Levenshtein ratio of over 0.8. The Levenshtein ratio is an indicator of the Levenshtein distance between the two idioms relative to their length. These potential matches are then judged manually on whether they are really forms of the same idiom or not." + ], + [ + "The results of using exact string matching to quantify the overlap between the dictionaries is illustrated in Figure FIGREF37.", + "Overlap between the three dictionaries is low. A possible explanation for this lies with the different nature of the dictionaries. Oxford is a traditional dictionary, created and edited by professional lexicographers, whereas Wiktionary is a crowdsourced dictionary open to everyone, and UsingEnglish is similar, but focused on ESL-learners. It is likely that these different origins result in different idiom inventories. Similarly, we would expect that the overlap between a pair of traditional dictionaries, such as the ODEI and the Penguin Dictionary of English Idioms BIBREF33 would be significantly higher. It should also be noted, however, that comparisons between more similar dictionaries also found relatively little overlap (BIBREF34; BIBREF35). A counterpoint is provided by BIBREF36, who quantifies coverage of verb-particle constructions in three different dictionaries and finds large overlap \u2013 perhaps because verb-particle are a more restricted class.", + "As noted previously, using exact string matching is a very limited approach to calculating overlap. Therefore, we used heuristics and manual checking to get more precise numbers, as shown in Table TABREF39, which also includes the three corpora in addition to the three dictionaries. As the manual checking only involved judging similar idioms found in pairs of resources, we cannot calculate three-way overlap as in Figure FIGREF37. The counts of the pair-wise overlap between dictionaries differ significantly between the two methods, which serves to illustrate the limitations of using only exact string matching and the necessity of using more advanced methods and manual effort.", + "Several insights can be gained from the data in Table TABREF39. The relation between Wiktionary and the SemEval corpus is obvious (cf. Section SECREF12), given the 96.92% coverage. For the other dictionary-corpus pairs, the coverage increases proportionally with the size of the dictionary, except in the case of UsingEnglish and the Sporleder corpus. The proportional increase indicates no clear qualitative differences between the dictionaries, i.e. one does not have a significantly higher percentage of non-idioms than the other, when compared to the corpora.", + "Generally, overlap between dictionaries and corpora is low: the two biggest, ODEI and Wiktionary have only around 30% overlap, while the dictionaries also cover no more than approximately 70% of the idioms used in the various corpora. Overlap between the three corpora is also extremely low, at below 5%. This is unsurprising, since a new dataset is more interesting and useful when it covers a different set of idioms than used in an existing dataset, and thus is likely constructed with this goal in mind." + ], + [ + "In order to evaluate the PIE extraction methods developed in this work (Section SECREF6), we exhaustively annotate an evaluation corpus with all instances of a pre-defined set of PIEs. As part of this, we come up with a workable definition of PIEs, and measure the reliability of PIE annotation by inter-annotator agreement.", + "Assuming that we have a set of idioms, the main problem of defining what is and what is not a potentially idiomatic expression is caused by variation. In principle, potentially idiomatic expression is an instance of a phrase that, when seen without context, could have either an idiomatic or a literal meaning. This is clearest for the dictionary form of the idiom, as in Example SECREF5. Literal uses generally allow all kinds of variation, but not all of these variations allow a figurative interpretation, e.g. Example SECREF5. However, how much variation an idiom can undergo while retaining its figurative interpretation is different for each expression, and judgements of this might vary from one speaker to the other. An example of this is spill the bean, a variant of spill the beans, in Example SECREF5 judged by BIBREF21 as being highly questionable. However, even here a corpus example can be found containing the same variant used in a figurative sense (Example SECREF5).", + "As such, we assume that we cannot know a priori which variants of an expression allow a figurative reading, and are thus a potentially idiomatic expression. Therefore we consider every possible morpho-syntactic variation of an idiom a PIE, regardless of whether it actually allows a figurative reading. We believe the boundaries of this variation can only be determined based on corpus evidence, and even then they are likely variable.", + "Note that a similar question is tackled by BIBREF26, when they establish the boundary between a `literal reading of a VMWE' and a `coincidental co-occurrence'. BIBREF26's answer is similar to ours, in that they count something as a literal reading of a VMWE if it `the same or equivalent dependencies hold between [the expression]'s components as in its canonical form'.", + ". John kicked the bucket last night.", + ". * The bucket, John kicked last night.", + ". ?? Azin spilled the bean. (from BIBREF21)", + ". Alba reveals Fantastic Four 2 details The Invisible Woman actress spills the bean on super sequel (from ukWaC)" + ], + [ + "Evaluating the extraction methods is easier than evaluating dictionary coverage, since the goal of the extraction component is more clearly delimited: given a set of PIEs from one or more dictionaries, extract all occurrences of those PIEs from a corpus. Thus, rather than dealing with the undefined set of all PIEs, we can work with a clearly defined and finite set of PIEs from a dictionary.", + "Because we have a clearly defined set of PIEs, we can exhaustively annotate a corpus for PIEs, and use that annotated corpus for automatic evaluation of extraction methods using recall and precision. This allows us to facilitate and speed up annotation by pre-extracting sentences possibly containing a PIE. After the corpus is annotated, the precision and recall can be easily estimated by comparing the extracted PIE instances to those marked in the corpus. The details of the corpus selection, dictionary selection, extraction heuristic and annotation procedure are presented in Section SECREF46, and the details and results of the various extraction methods are presented in Section SECREF6." + ], + [ + "As a base corpus, we use the XML version of the British National Corpus BIBREF37, because of its size, variety, and wide availability. The BNC is pre-segmented into s-units, which we take to be sentences, w-units, which we take to be words, and c-units, punctuation. We then extract the text of all w-units and c-units. We keep the sentence segmentation, resulting in a set of plain text sentences. All sentences are included, except for sentences containing elements, which are filtered out. These elements indicate places where material from the original has been left out, e.g. for anonymisation purposes. Since this can result in incomplete sentences that cannot be parsed correctly, we filter out sentences containing these gaps.", + "We use only the written part of the BNC. From this, we extract a set of documents with the aim of having as much genre variation as possible. To achieve this, we select the first document in each genre, as defined by the classCode attribute (e.g. nonAc, commerce, letters). The resulting set of 46 documents makes up our base corpus. Note that these documents vary greatly in size, which means the resulting corpus is varied, but not balanced in terms of size (Table TABREF43). The documents are split across a development and test set, as specified at the end of Section SECREF46. We exclude documents with IDs starting with A0 from all annotation and evaluation procedures, as these were used during development of the extraction tool and annotation guidelines.", + "As for the set of potentially idiomatic expressions, we use the intersection of the three dictionaries, Wiktionary, Oxford, and UsingEnglish. Based on the assumption that, if all three resources include a certain idiom, it must unquestionably be an idiom, we choose the intersection (also see Figure FIGREF37). This serves to exclude questionable entries, like at all, which is in Wiktionary. The final set of idioms used for these experiments consists of 591 different multiword expressions. Although we aim for wide coverage, this is a necessary trade-off to ensure quality. At the same time, it leaves us with a set of idiom types that is approximately ten times larger than present in existing corpora. The set of 591 idioms includes idioms with a large variety of syntactic patterns, of which the most frequent ones are shown in Table TABREF44. The statistics show that the types most prevalent in existing corpora, verb-noun and preposition-noun combinations, are indeed the most frequent ones, but that there is a sizeable minority of types that do not fall into those categories, including coordinated adjectives, coordinated nouns, and nouns with prepositional phrases. This serves to emphasise the necessity of not restricting corpora to a small set of syntactic patterns." + ], + [ + "To annotate the corpus completely manually would require annotators to read the whole corpus, and cross-reference each sentence to a list of almost 600 PIEs, to check whether one of those PIEs occurs in a sentence. We do not consider this a feasible annotation settings, due to both the difficulty of recognising literal usages of idioms and the time cost needed to find enough PIEs, given their low overall frequency. As such, we use a pre-extraction step to present candidates for annotation to the human annotators.", + "Given the corpus and the set of PIEs, we heuristically extract the PIE candidates as follows: given an idiomatic expression, extract every sentence which contains all the defining words of the idiom, in any form. This ensures that all possibly matching sentences get extracted, while greatly pruning the amount of sentences for annotators to look at. In addition, it allows us to present the heuristically matched PIE type and corresponding words to the annotators, which makes it much easier to judge whether something is a PIE or not. This also means that annotators never have to go through the full list of PIEs during the annotation process.", + "Initially, the heuristic simply extracted any sentence containing all the required words, where a word is any of the inflectional variants of the words in the PIE, except for determiners and punctuation. This method produced large amounts of noise, that is, a set of PIE candidates with only a very low percentage of actual PIEs. This was caused by the presence of some highly frequent PIEs with very little defining lexical content, such as on the make, and in the running. For example, with the original method, every sentence containing the preposition on, and any inflectional form of the verb make was extracted, resulting in a huge number of non-PIE candidates.", + "To limit the amount of noise, two restrictions were imposed. The first restrictions disallows word order variation for PIEs which do not contain a verb. The rationale behind this is that word order variation is only possible with PIEs like spill the beans (e.g. the beans were spilled), and not with PIEs like in the running (*the running in??). The second restriction is that we limit the number of words that can be inserted between the words of a PIE, but only for PIEs like on the make, and in the running, i.e. PIEs which only contain prepositions, determiners and a single noun. The number of intervening words was limited to three tokens, allowing for some variation, as in Example SECREF45, but preventing sentences like Example SECREF45 from being extracted. This restriction could result in the loss of some PIE candidates with a large number of intervening words. However, the savings in annotation time clearly outweigh the small loss in recall in this situation.", + ". Either at New Year or before July you can anticipate a change in the everyday running of your life. (in the running - BNC - document CBC - sentence 458)", + ". [..] if [he] hung around near the goal or in the box for that matter instead of running all over the show [..] (in the running - BNC - document J1C - sentence 1341)" + ], + [ + "The manual annotation procedure consists of three different phases (pilot, double annotation, single annotation), followed by an adjudication step to resolve conflicting annotations. Two things are annotated: whether something is a PIE or not, and if it is a PIE, which sense the PIE is used in. In the first phase (0-100-*), we randomly select hundred of the 2,239 PIE candidates which are then annotated by three annotators. All annotators have a good command of English, are computational linguists, and familiar with the subject. The annotators include the first and last author of this paper.", + "The annotators were provided with a short set of guidelines, of which the main rule-of-thumb for labelling a phrase as a PIE is as follows: any phrase is a PIE when it contains all the words, with the same part-of-speech, and in the same grammatical relations as in the dictionary form of the PIE, ignoring determiners.", + "For sense annotation, annotators were to mark a PIE as idiomatic if it had a sense listed in one of the idiom dictionaries, and as literal if it had a meaning that is a regular composition of its component words. For cases which were undecidable due to lack of context, the ?-label was used. The other-label was used as a container label for all cases in which neither the literal or idiomatic sense was correct (e.g. meta-linguistic uses and embeddings in metaphorical frames, see also Section SECREF10).", + "The first phase of annotation serves to bring to light any inconsistencies between annotators and fill in any gaps in the annotation guidelines. The resulting annotations already show a reasonably high agreement of 0.74 Fleiss' Kappa. Table TABREF48 shows annotation details and agreement statistics for all three phases. The annotation tasks suffixed by -PIE indicate agreement on PIE/non-PIE annotation and the tasks suffixed by -sense indicate agreement on sense annotation for PIEs.", + "In the second phase of annotation (100-600-* & 600-1100-*), another 1000 of the 2239 PIE candidates are selected to be annotated by two pairs of annotators. This shows very high agreement, as shown in Table TABREF48. This is probably due to the improvement in guidelines and the discussion following the pilot round of annotation. The exception to this are the somewhat lower scores for the 600-1100-sense annotation task. Adjudication revealed that this is due almost exclusively because of a different interpretation of the literal and idiomatic senses of a single PIE type: on the ground. Excluding this PIE type, Fleiss' Kappa increases from 0.63 to 0.77.", + "Because of the high agreement on PIE annotation, we deem it sufficient for the remainder (1108 candidates) to be annotated by only the primary annotator in the third phase of annotation (1100-2239-*). The reliability of the single annotation can be checked by comparing the distribution of labels to the multi-annotated parts. This shows that it falls clearly within the ranges of the other parts, both in the proportion of PIEs and idiomatic senses (see Table TABREF49). The single-annotated part has 49.0% PIEs, which is only 4 percentage points above the 44.7% PIEs in the multi-annotated parts. The proportion of idioms is just 2 percentage points higher, with 55.9% versus 53.9.%.", + "Although inter-annotator agreement was high, there was still a significant number of cases in the triple and double annotated PIE candidate sets where not all annotators agreed. These cases were adjudicated through discussion by all annotators, until they were in agreement. In addition, all PIE candidates which initially received the ?-label (unclear or undecidable) for sense or PIE were resolved in the same manner. In the adjudication procedure, annotators were provided with additional context on each side of the idiom, in contrast to the single sentence provided during the initial annotation. The main reason to do adjudication, rather than simply discarding all candidates for which there was disagreement, was that we expected exactly those cases for which there are conflicting annotations to be the most interesting ones, since having non-standard properties would cause the annotations to diverge. Examples of such interesting non-standard cases are at sea as part of a larger satirical frame in Example SECREF46 and cut the mustard in Example SECREF46 where it is used in a headline as wordplay on a Cluedo character.", + ". The bovine heroine has connections with Cowpeace International, and deals with a huge treacle slick at sea. (at sea - BNC - document CBC - sentence 13550)", + ". Why not cut the Mustard? [..] WADDINGTON Games's proposal to axe Reverend Green from the board game Cluedo is a bad one. (cut the mustard - BNC - document CBC - sentence 14548)", + "We split the corpus at the document level. The corpus consists of 45 documents from the BNC, and we split it in such a way that the development set has 1,112 candidates across 22 documents and the test set has 1,127 candidates from 23 documents. Note that this means that the development and test set contain different genres. This ensures that we do not optimise our systems on genre-specific aspects of the data." + ], + [ + "We propose and implement four different extraction methods, of differing complexities: exact string match, fuzzy string match, inflectional string match, and parser-based extraction. Because of the absence of existing work on this task, we compare these methods to each other, where the more basic methods function as baselines. More complex methods serve to shine light on the difficulty of the PIE extraction task; if simple methods already work sufficiently well, the task is not as hard as expected, and vice versa. Below, each of the extraction methods is presented and discussed in detail." + ], + [ + "This is, very simply, extracting all instances of the exact dictionary form of the PIE, from the tokenized text of the corpus. Word boundaries are taken into account, so at sea does not match `that seawater'. As a result, all inflectional and other variants of the PIE are ignored." + ], + [ + "Fuzzy string match is a rough way of dealing with morphological inflection of the words in a PIE. We match all words in the PIE, taking into account word boundaries, and allow for up to 3 additional letters at the end of each word. These 3 additional characters serve to cover inflectional suffixes." + ], + [ + "In inflectional string match, we aim to match all inflected variations of a PIE. This is done by generating all morphological variants of the words in a PIE, generating all combinations of those words, and then using exact string match as described earlier.", + "Generating morphological variations consists of three steps: part-of-speech tagging, morphological analysis, and morphological reinflection. Since inflectional variation only applies to verbs and nouns, we use the Spacy part-of-speech tagger to detect the verbs and nouns. Then, we apply the morphological analyser morpha to get the base, uninflected form of the word, and then use the morphological generation tool morphg to get all possible inflections of the word. Both tools are part of the Morph morphological processing suite BIBREF38. Note that the Morph tools depend on the part-of-speech tag in the input, so that a wrong PoS may lead to an incorrect set of morphological variants.", + "For a PIE like spill the beans, this results in the following set of variants: $\\lbrace $spill the bean, spills the bean, spilled the bean, spilling the bean, spill the beans, spills the beans, spilled the beans, spilling the beans$\\rbrace $. Since we generate up to 2 variants for each noun, and up to 4 variants for each verb, the number of variants for PIEs containing multiple verbs and nouns can get quite large. On average, 8 additional variants are generated for each potentially idiomatic expression." + ], + [ + "For all string match-based methods, ways to improve performance are implemented, to make them as competitive as possible. Rather than doing exact string matching, we also allow words to be separated by something other than spaces, e.g. nuts-and-bolts for nuts and bolts. Additionally, there is an option to take into account case distinctions. With the case-sensitive option, case is preserved in the idiom lists, e.g. coals to Newcastle, and the string matching is done in a case-sensitive manner. This increases precision, e.g. by avoiding PIEs as part of proper names, but also comes at a cost of recall, e.g. for sentence-initial PIEs. Thirdly, there is the option to allow for a certain number of intervening words between each pair of words in the PIE. This should improve recall, at the cost of precision. For example, this would yield the true positive make a huge mountain out of a molehill for make a mountain out of a molehill, but also false positives like have a smoke and go for have a go.", + "A third shared property of the string-based methods is the processing of placeholders in PIEs. PIEs containing possessive pronoun placeholders, such as one's and someone's are expanded. That is, we remove the original PIE, and add copies of the PIE where the placeholder is replaced by one of the possessive personal pronouns. For example, a thorn in someone's side is replaced by a thorn in {my, your, his, ...} side. In the case of someone's, we also add a wildcard for any possessively used word, i.e. a thorn in \u2014's side, to match e.g. a thorn in Google's side. Similarly, we make sure that PIE entries containing \u2014, such as the mother of all \u2014, will match any word for \u2014 during extraction. We do the same for someone, for which we substitute objective pronouns. For one, this is not possible, since it is too hard to distinguish from the one used as a number." + ], + [ + "Parser-based extraction is potentially the widest-coverage extraction method, with the capacity to extract both morphological and syntactic variants of the PIE. This should be robust against the most common modifications of the PIE, e.g. through word insertions (spill all the beans), passivisation (the beans were spilled), and abstract over articles (spill beans).", + "In this method, PIEs are extracted using the assumption that any sentence which contains the lemmata of the words in the PIE, in the same dependency relations as in the PIE, contains an instance of the PIE type in question. More concretely, this means that the parse of the sentence should contain the parse tree of the PIE as a subtree. This is illustrated in Figure FIGREF57, which shows the parse tree for the PIE lose the plot, parsed without context. Note that this is a subtree of the parse tree for the sentence `you might just lose the plot completely', which is shown in Figure FIGREF58. Since the sentence parse contains the parse of the PIE, we can conclude that the sentence contains an instance of that PIE and extract the span of the PIE instance.", + "All PIEs are parsed in isolation, based on the assumption that all PIEs can be parsed, since they are almost always well-formed phrases. However, not all PIEs will be parsed correctly, especially since there is no context to resolve ambiguity. Errors tend to occur at the part-of-speech level, where, for example, verb-object combinations like jump ship and touch wood are erroneously tagged as noun-noun compounds. An analysis of the impact of parser error on PIE extraction performance is presented in Section SECREF73. Initially, we use the Spacy parser for parsing both the PIEs and the sentences.", + "Next, the sentence is parsed, and the lemma of the top node of the parsed PIE is matched against the lemmata of the sentence parse. If a match is found, the parse tree of the PIE is matched against the subtree of the matching sentence parse node. If the whole PIE parse tree matches, the span ranging from the first PIE token to the last is extracted. This span can thus include words that are not directly part of the PIE's dictionary form, in order to account for insertions like ships were jumped for jump ship, or have a big heart for have a heart.", + "During the matching, articles (a/an/the) are ignored, and passivisation is accounted for with a special rule. In addition, a number of special cases are dealt with. These are PIEs containing someone('s), something('s), one's, or \u2014. These words are used in PIEs as placeholders for a generic possessor (someone's/something's/one's), generic object (someone/something), or any word of the right PoS (\u2014).", + "For someone's, and something's, we match any possessive pronoun, or (proper) noun + possessive marker. For one's, only possessive pronouns are matched, since this is a placeholder for reflexive possessors. For someone and something, any non-possessive pronoun or (proper) noun is matched.", + "For \u2014 wildcards, any word can be matched, as long as it has the right relation to the right head. An additional challenge with these wildcards is that PIEs containing them cannot be parsed, e.g. too \u2014 for words is not parseable. This is dealt with by substituting the \u2014 by a PoS-ambiguous word, such as fine, or back.", + "Two optional features are added to the parser-based method with the goal of making it more robust to parser errors: generalising over dependency relation labels, and generalising over dependency relation direction. We expect this to increase recall at the cost of precision. In the first no labels setting, we match parts of the parse tree which have the same head lemma and the same dependent lemma, regardless of the relation label. An example of this is Figure FIGREF60, which has the wrong relation label between up and ante. If labels are ignored, however, we can still extract the PIE instance in Figure FIGREF61, which has the correct label. In the no directionality setting, relation labels are also ignored, and in addition the directionality of the relation is ignored, that is, we allow for the reversal of heads and dependents. This benefits performance in a case like Figure FIGREF62, which has stock as the head of laughing in a compound relation, whereas the parse of the PIE (Figure FIGREF63) has laughing as the head of stock in a dobj relation.", + "Note that similar settings were implemented by BIBREF26, who detect literal uses of VMWEs using a parser-based method with either full labelled dependencies, unlabelled dependencies, or directionless unlabelled dependencies (which they call BagOfDeps). They find that recall increases when less restrictions on the dependencies are used, but that this does not hurt precision, as we would expect. However, we cannot draw too many conclusions from these results due to the small size of their evaluation set, which consists of just 72 literal VMWEs in total." + ], + [ + "Since the parser-based method parses PIEs without any context, it often finds an incorrect parse, as for jump ship in Figure FIGREF65. As such, we add an option to the method that aims to increase the number of correct parses by parsing the PIE within context, that is, within a sentence. This can greatly help to disambiguate the parse, as in Figure FIGREF66. If the number of correct parses goes up, the recall of the extraction method should also increase. Naturally, it can also be the case that a PIE is parsed correctly without context, and incorrectly with context. However, we expect the gains to outweigh the losses.", + "The challenge here is thus to collect example sentences containing the PIE. Since the whole point of this work is to extract PIEs from raw text, this provides a catch-22-like situation: we need to extract a sentence containing a PIE in order to extract sentences containing a PIE.", + "The workaround for this problem is to use the exact string matching method with the dictionary form of the PIE and a very large plain text corpus to gather example sentences. By only considering the exact dictionary form we both simplify the finding of example sentences and the extraction of the PIE's parse from the sentence parse.", + "In case multiple example sentences are found, the shortest sentence is selected, since we assume it is easiest to parse. This is also the reason we make use of very large corpora, to increase the likelihood of finding a short, simple sentence. The example sentence extraction method is modified in such a way that sentences where the PIE is used meta-linguistically in quotes, e.g. \u201cthe well-known English idiom `to spill the beans' has no equivalents in other languages\u201d, are excluded, since they do not provide a natural context for parsing. When no example sentence can be found in the corpus, we back-off to parsing the PIE without context. After a parse has been found for each PIE (i.e. with or without context), the method proceeds identically to the regular parser-based method.", + "We make use of the combination of two large corpora for the extraction of example sentences: the English Wikipedia, and ukWaC BIBREF17. For the Wikipedia corpus, we use a dump (13-01-2016) of the English-language Wikipedia, and remove all Wikipedia markup. This is done using WikiExtractor. The resulting files still contain some mark-up, which is removed heuristically. The resulting corpus contains mostly clean, raw, untokenized text, numbering approximately 1.78 billion tokens.", + "As for ukWaC, all XML-markup was removed, and the corpus is converted to a one-sentence-per-line format. UkWaC is tokenized, which makes it difficult for a simple string match method to find PIEs containing punctuation, for example day in, day out. Therefore, all spaces before commas, apostrophes, and sentence-final punctuation are removed. The resulting corpus contains approximately 2.05 billion tokens, making for a total of 3.83 billion tokens in the combined ukWaC and Wikipedia corpus." + ], + [ + "In order to determine which of the methods described previously produces the highest quality extraction of potentially idiomatic expressions, we evaluate them, in various settings, on the corpus described in Section SECREF5.", + "For parser-based extraction, systems with and without in-context parsing, ignoring labels, and ignoring directionality are tested. For the three string-based extraction methods, varying numbers of intervening words and case sensitivity are evaluated. Evaluation is done using the development set, consisting of 22 documents and 1112 PIE candidates, and the test set, which consists of 23 documents and 1127 PIE candidates. For each method the best set of parameters and/or options is determined using the development set, after which the best variant by F1-score of each method is evaluated on the test set.", + "Since these documents in the corpus are exhaustively annotated for PIEs (see Section SECREF40), we can calculate true and false positives, and false negatives, and thus precision, recall and F1-score. The exact spans are ignored, because the spans annotated in the evaluation corpus are not completely reliable. These were automatically generated during candidate extraction, as described in Section SECREF45. Rather, we count an extraction as a true positive if it finds the correct PIE type in the correct sentence.", + "Note that we judge the system with the highest F1-score to be the best-performing system, since it is a clear and objective criterion. However, when using the system in practice, the best performance depends on the goal. When used as a preprocessing step for PIE disambiguation, the system with the highest F1-score is perhaps the most suitable, but as a corpus building tool, one might want to sacrifice some precision for an increase in recall. This helps to get the most comprehensive annotation of PIEs possible, without overloading the annotators with false extractions (i.e. non-PIEs), by maintaining high precision.", + "The results for each system on the development set are presented in Tables TABREF70 and TABREF71. Generally, results are in line with expectations: (the best) parse-based methods are better than (the best) string-based methods, and within string-based methods, inflectional matching works best. The same goes for the different settings: case-sensitivity increases precision at the cost of recall, allowing intervening words increases recall at the cost of precision, and the same goes for the no labels and no directionality options for parser-based extraction. Overall, in-context parser-based extraction works best, with an F1 of 88.54%, whereas fuzzy matching does very poorly.", + "Within string-based methods, exact matching has the highest precision, but low recall. Fuzzy matching increases recall at a disproportionately large precision cost, whereas inflectional matching combines the best of both worlds and has high recall at a small loss in precision. For the parser-based system, it is notable that parsing idioms within context yields a clear overall improvement by greatly improving recall at a small cost in precision.", + "We evaluate the best variant of each system, as determined by F1-score, on the test set. This gives us an indication of whether the system is robust enough, or was overfitted on the development data. Results on the test set are shown in Table TABREF72. On average, the results are lower than the results on the development set. The string-based methods perform clearly worse, with drops of about 4% F1-score for exact and inflectional match, and a large drop of almost 9% F1-score for fuzzy matching. The parser-based method, on the other hand, is more robust, with a small 0.59% increase in F1-score on the test set." + ], + [ + "Broadly speaking, the PIE extraction systems presented above perform in line with expectations. It is nevertheless useful to see where the best-performing system misses out, and where improvements like in-context parsing help performance.", + "We analyse the shortcomings of the in-context parser-based system by looking at the false positives and false negatives on the development set. We consider the output of the system with best overall performance, since it will provide the clearest picture. The system extracts 529 PIEs in total, of which 54 are false extractions (false positives), and it misses 69 annotated PIE instances (false negatives). Most false positives stem from the system's failure to capture nuances of PIE annotation. This includes cases where PIEs contain, or are part of, proper nouns (Example SECREF73), PIEs that are part of coordination constructions (Example SECREF73), and incorrect attachments (Example SECREF73). Among these errors, sentences containing proper nouns are an especially frequent problem.", + ". Drama series include [..] airline security thrills in Cleared For Takeoff and Head Over Heels [..] (in the clear - BNC - document CBC - sentence 5177)", + ". They prefer silk, satin or lace underwear in tasteful black or ivory. (in the black - BNC - document CBC - sentence 14673)", + ". [..] `I saw this chap make something out of an ordinary piece of wood \u2014 he fashioned it into an exquisite work of art.' (out of the woods - BNC - document ABV - sentence 1300)", + "The main cause of false negatives are errors made by the parser. In order to correctly extract a PIE from a sentence, both the PIE and the sentence have to be parsed correctly, or at least parsed in the same way. This means a missed extraction can be caused by a wrong parse for the PIE or a wrong parse for the sentence. These two error types form the largest class of false negatives. Since some PIE types are rather frequent, a wrong parse for a single PIE type can potentially lead to a large number of missed extractions.", + "It is not surprising that the parser makes many mistakes, since idioms often have unusual syntactic constructions (e.g. come a cropper) and contain words where default part-of-speech tags lead to the wrong interpretation (e.g. round is a preposition in round the bend, not a noun or adjective). This is especially true when idioms are parsed without context, and hence, where in-context parsing provides the largest benefit: the number of PIEs which are parsed incorrectly drops, which leads to F1-scores on those types going from 0% to almost 100% (e.g. in light of and ring a bell). Since parser errors are the main contributor to false negatives, hurting recall, we can observe that parsing idioms in context serves to benefit only recall, by 7 percentage points, at only a small loss in precision.", + "We find that adding context mainly helps for parsing expressions which are structurally relatively simple, but still ambiguous, such as rub shoulders, laughing stock, and round the bend. Compare, for example, the parse trees for laughing stock in isolation and within the extracted context sentence in Figures FIGREF74 and FIGREF75. When parsed in isolation, the relation between the two words is incorrectly labelled as a compound relation, whereas in context it is correctly labelled as a direct object relation. Note however, that for the most difficult PIEs, embedding them in a context does solve the parsing problem: a syntactically odd phrase is hard to phrase (e.g. for the time being), and a syntactically odd phrase in a sentence makes for a syntactically odd sentence that is still hard to parse (e.g. `London for the time being had been abandoned.'). Finding example sentences turned out not to be a problem, since appropriate sentences were found for 559 of 591 PIE types.", + "An alternative method for reducing parser error is to use a different, better parser. The Spacy parser was mainly chosen for implementation convenience and speed, and there are parsers which have better performance, as measured on established parsing benchmarks. To investigate the effectiveness of this method, we used the Stanford Neural Dependency Parser BIBREF39 to extract PIEs in the regular parsing, in-context parsing and the no labels settings. In all cases, using the Stanford parser yielded worse extraction performance than the Spacy parser. A possible explanation for why a supposedly better parser performs worse here is that parsers are optimised and trained to do well on established benchmarks, which consist of complete sentences, often from news texts. This does not necessarily correlate with parsing performance on short (sentences containing) idiomatic phrases. As such, we cannot assume that better overall parsing performance implies PIE extraction performance.", + "It should be noted that, when assessing the quality of PIE extraction performance, the parser-based methods are sensitive to specific PIE types. That is, if a single PIE type is parsed incorrectly, then it is highly probable that all instances of that type are missed. If this type is also highly frequent, this means that a small change in actual performance yields a large change in evaluation scores. Our goal is to have a PIE extraction system that is robust across all PIE types, and thus the current evaluation setting does not align exactly with our aim.", + "Splitting out performance per PIE type reveals whether there is indeed a large variance in performance across types. Table TABREF76 shows the 25 most frequent PIE types in the corpus, and the performance of the in-context-parsing-based system on each. Except two cases (in the black and round the bend), we see that the performance is in the 80\u2013100% range, even showing perfect performance on the majority of types.", + "For none of the types do we see low precision paired with high recall, which indicates that the parser never matches a highly frequent non-PIE phrase. For the system with the no labels and no-directionality options (per-type numbers not shown here), however, this does occur. For example, ignoring the labels for the parse of the PIE have a go leads to the erroneous matching of many sentences containing a form of have to go, which is highly frequent, thus leading to a large drop in precision.", + "Although performance is stable across the most frequent types, among the less frequent types it is more spotty. This hurts overall performance, and there are potential gains in mitigating the poor performance on these types, such as for the time being. At the same time, the string matching methods show much more stable performance across types, and some of them do so with very high precision. As such, a combination of two such methods could boost performance significantly. If we use a high-precision string match-based method, such as the exact string match variant with a precision of 97.35%, recall could be improved for the wrongly parsed PIE types, without a significant loss of precision.", + "We experiment with two such combinations, by simply taking the union of the sets of extracted idioms of both systems, and filtering out duplicates. Results are shown in Table TABREF77. Both combinations show the expected effect: a clear gain in recall at a minimal loss in precision. Compared to the in-context-parsing-based system, the combination with exact string matching yields a gain in recall of over 6%, and the combination with inflectional string matching yields an even bigger gain of almost 8%, at precision losses of 0.6% and 0.8%, respectively. This indicates that the systems are very much complementary in the PIEs they extract. It also means that, when used in practice, combining inflectional string matching and parse-based extraction is the most reliable configuration." + ], + [ + "We present an in-depth study on the automatic extraction of potentially idiomatic expressions based on dictionaries. The purpose of automatic dictionary-based extraction is, on the one hand, to function as a pre-extraction step in the building of a large idiom-annotated corpus. On the other hand, it can function as part of an idiom extraction system when combined with a disambiguation component. In both cases, the ultimate goal is to improve the processing of idiomatic expressions within NLP. This work consists of three parts: a comparative evaluation of the coverage of idiom dictionaries, the annotation of a PIE corpus, and the development and evaluation of several dictionary-based PIE extraction methods.", + "In the first part, we present a study of idiom dictionary coverage, which serves to answer the question of whether a single idiom dictionary, or a combination of dictionaries, can provide good coverage of the set of all English idioms. Based on the comparison of dictionaries to each other, we estimate that the overlap between them is limited, varying from 20% to 55%, which indicates a large divergence between the dictionaries. This can be explained by the fact that idioms vary widely by register, genre, language variety, and time period. In our case, it is also likely that the divergence is caused partly by the gap between crowdsourced dictionaries on the one hand, and a dictionary compiled by professional lexicographers on the other. Given these factors, we can conclude that a single dictionary cannot provide even close to complete coverage of English idioms, but that by combining dictionaries from various sources, significant gains can be made. Since `English idioms' are a diffuse and constantly changing set, we have no gold standard to compare to. As such, we conclude that multiple dictionaries should be used when possible, but that we cannot say any anything definitive on the coverage of dictionaries with regard to the complete set of English idioms (which can only be approximated in the first place). A more comprehensive of idiom resources could be made in the future by using more advanced automatic methods for matching, for example by using BIBREF32's (BIBREF32) method for measuring expression variability. This would make it easier to evaluate a larger number of dictionaries, since no manual effort would be required.", + "In the second part, we experiment with the exhaustive annotation of PIEs in a corpus of documents from the BNC. Using a set of 591 PIE types, much larger and more varied than in existing resources, we show that it is very much possible to establish a working definition of PIE that allows for a large amount of variation, while still being useful for reliable annotation. This resulted in high inter-annotator agreement, ranging from 0.74 to 0.91 Fleiss' Kappa. This means that we can build a resource to evaluate a wide-range idiom extraction system with relatively little effort. The final corpus of PIEs with sense annotations is publicly available consists of 2,239 PIE candidates, of which 1,050 actual PIEs instances, and contains 278 different PIE types.", + "Finally, several methods for the automatic extraction of PIE instances were developed and evaluated on the annotated PIE corpus. We tested methods of differing complexity, from simple string match to dependency parse-based extraction. Comparison of these methods revealed that the more computationally complex method, parser-based extraction, works best. Parser-based extraction is especially effective in capturing a larger amount of variation, but is less precise than string-based methods, mostly because of parser error. The best overall setting of this method, which parses idioms within context, yielded an F1-score of 89.13% on the test set. Parser error can be partly compensated by combining the parse-based method and the inflectional string match method, which yields an F1-score of 92.01% (on the development set). This aligns well with the findings by BIBREF27, who found that combining simpler and more complex methods improves over just using a simple method case for extracting verb-particle constructions. This level of performance means that we can use the tool in corpus building. This greatly reduces the amount of manual extraction effort involved, while still maintaining a high level of recall. We make the source code for the different systems publicly available. Note that, although used here in the context of PIE extraction, our methods are equally applicable to other phrase extraction tasks, for example the extraction of light-verb constructions, metaphoric constructions, collocations, or any other type of multiword expression (cf. BIBREF27, BIBREF25, BIBREF26). Similarly, our method can be conceived as a blueprint and extended to languages other than English. For this to be possible, for any given new language one would need a list of target expressions and, in the case of the parser-based method, a reliable syntactic parser. If this is not the case, the inflectional matching method can be used, which requires only a morphological analyser and generator. Obviously, for languages that are morphologically richer than English, one would need to develop strategies aimed at controlling non-exact matches, so as to enhance recall without sacrificing precision. Previous work on Italian, for example, has shown the feasibility of achieving such balance through controlled pattern matching BIBREF40. Languages that are typologically very different from English would obviously require a dedicated approach for the matching of PIEs in corpora, but the overall principles of extraction, using language-specific tools, could stay the same.", + "Currently, no corpora containing annotation of PIEs exist for languages other than English. However, the PARSEME corpus BIBREF19 already contains idioms (only idiomatic readings) for many languages and would only need annotation of literal usages of idioms to make up a set of PIEs. Paired with the Universal Dependencies project BIBREF41, which increasingly provides annotated data as well as processing tools for an ever growing number of languages, this seems an excellent starting point for creating PIE resources in multiple languages." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2472/instruction.md b/qasper-2472/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..c8f5c6dfef548bc68f056a7f234bacb01d6c7321 --- /dev/null +++ b/qasper-2472/instruction.md @@ -0,0 +1,78 @@ +Name of Paper: A General-Purpose Tagger with Convolutional Neural Networks + +Question: How many parameters does their CNN have? + +## Full Paper Text (JSON) + +```json +{ + "section_name": [ + "Introduction", + "Model", + "Character Composition Model", + "Context Encoding Model", + "Hyper-parameters", + "Data", + "Tasks", + "Setups", + "Results", + "Unnormalized Text", + "Conclusion" + ], + "paragraphs": [ + [ + "Recently, character composition models have shown great success in many NLP tasks, mainly because of their robustness in dealing with out-of-vocabulary (OOV) words by capturing sub-word informations. Among the character composition models, bidirectional long short-term memory (LSTM) models and convolutional neural networks (CNN) are widely applied in many tasks, e.g. part-of-speech (POS) tagging BIBREF0 , BIBREF1 , named entity recognition BIBREF2 , language modeling BIBREF3 , BIBREF4 , machine translation BIBREF5 and dependency parsing BIBREF6 , BIBREF7 .", + "In this paper, we present a state-of-the-art general-purpose tagger that uses CNNs both to compose word representations from characters and to encode context information for tagging. We show that the CNN model is more capable than the LSTM model for both functions, and more stable for unseen or unnormalized words, which is the main benefit of character composition models.", + "Yu:2017 compared the performance of CNN and LSTM as character composition model for dependency parsing, and concluded that CNN performs better than LSTM. In this paper, we show that this is also the case for POS tagging. Furthermore, we extend the scope to morphological tagging and supertagging, in which the tag set is much larger and long-distance dependencies between words are more important.", + "In these three tagging tasks, we compare our tagger with the bilstm-aux tagger BIBREF1 and the CRF-based morphological tagger MarMot BIBREF8 . The CNN tagger shows robust performance accross the three tasks, and achieves the highest average accuracy in all tasks. It (significantly) outperforms LSTM in morphological tagging, and outperforms both baselines in supertagging by a large margin.", + "To test the robustness of the taggers against the OOV problem, we also conduct experiments using artificially constructed unnormalized text by corrupting words in the normal dev set. Again, the CNN tagger outperforms the two baselines by a very large margin.", + "Therefore we conclude that our CNN tagger is a robust state-of-the-art general-purpose tagger that can effectively compose word representation from characters and encode context information." + ], + [ + "Our proposed CNN tagger has two main components: the character composition model and the context encoding model. Both components are essentially CNN models, capturing different levels of information: the first CNN captures morphological information from character n-grams, the second one captures contextual information from word n-grams. Figure FIGREF2 shows a diagram of both models of the tagger." + ], + [ + "The character composition model is similar to Yu:2017, where several convolution filters are used to capture character n-grams of different sizes. The outputs of each convolution filter are fed through a max pooling layer, and the pooling outputs are concatenated to represent the word." + ], + [ + "The context encoding model captures the context information of the target word by scanning through the word representations of its context window. The word representation could be only word embeddings ( INLINEFORM0 ), only composed vectors ( INLINEFORM1 ) or the concatenation of both ( INLINEFORM2 )", + "A context window consists of N words to both sides of the target word and the target word itself. To indicate the target word, we concatenate a binary feature to each of the word representations with 1 indicating the target and 0 otherwise, similar to Vu:2016. Additional to the binary feature, we also concatenate a position embedding to encode the relative position of each context word, similar to Gehring:2017." + ], + [ + "For the character composition model, we take a fixed input size of 32 characters for each word, with padding on both sides or cutting from the middle if needed. We apply four convolution filters with sizes of 3, 5, 7, and 9. Each filter has an output channel of 25 dimensions, thus the composed vector is 100-dimensional. We apply Gaussian noise with standard deviation of 0.1 is applied on the composed vector during training.", + "For the context encoding model, we take a context window of 15 (7 words to both sides of the target word) as input and predict the tag of the target word. We also apply four convolution filters with sizes of 2, 3, 4 and 5, each filter is stacked by another filter with the same size, and the output has 128 dimensions, thus the context representation is 512-dimensional. We apply one 512-dimensional hidden layer with ReLU non-linearity before the prediction layer. We apply dropout with probability of 0.1 after the hidden layer during training.", + "The model is trained with averaged stochastic gradient descent with a learning rate of 0.1, momentum of 0.9 and mini-batch size of 100. We apply L2 regularization with a rate of INLINEFORM0 on all the parameters of the network except the embeddings." + ], + [ + "We use treebanks from version 1.2 of Universal Dependencies (UD), and in the case of several treebanks for one language, we only use the canonical treebank. There are in total 22 treebanks, as in Plank:2016. Each treebank splits into train, dev, and test sets, we use the dev sets for early stop, and test on the test sets." + ], + [ + "We evaluate our method on three tagging tasks: POS tagging (Pos), morphological tagging (Morph) and supertagging (Stag).", + "For POS tagging we use Universal POS tags, which is an extension of Petrov:2012. The universal tag set tries to capture the \u201cuniversal\u201d properties of words and facilitate cross-lingual learning. Therefore the tag set is very coarse and leaves out most of the language-specific properties to morphological features.", + "Morphological tags encode the language-specific morphological features of the words, e.g., number, gender, case. They are represented in the UD treebanks as one string which contains several key-value pairs of morphological features.", + "Supertags BIBREF9 are tags that encode more syntactic information than standard POS tags, e.g. the head direction or the subcategorization frame. We use dependency-based supertags BIBREF10 which are extracted from the dependency treebanks. Adding such tags into feature models of statistical dependency parsers significantly improves their performance BIBREF11 , BIBREF12 . Supertags can be designed with different levels of granularity. We use the standard Model 1 from Ouchi:2014, where each tag consists of head direction, dependency label and dependent direction. Even with the basic supertag model, the Stag task is more difficult than Pos and Morph because it generally requires taking long-distance dependencies between words into consideration.", + "We select these tasks as examples for tagging applications because they differ strongly in tag set sizes. Generally, the Pos set sizes for all the languages are no more than 17 and Stag set sizes are around 200. When treating morphological features as a string (i.e. not splitting into key-value pairs), the sizes of the Morph tag sets range from about 100 up to 2000." + ], + [ + "As baselines to our models, we take the two state-of-the-art taggers MarMot (denoted as CRF) and bilstm-aux (denoted as LSTM). We train the taggers with the recommended hyper-parameters from the documentation.", + "To ensure a fair comparison (especially between LSTM and CNN), we generally treat the three tasks equally, and do not apply task-specific tuning on them, i.e., using the same features and same model hyper-parameters in each single task. Also, we do not use any pre-trained word embeddings.", + "For the LSTM tagger, we use the recommended hyper-parameters in the documentation including 64-dimensional word embeddings ( INLINEFORM0 ) and 100-dimensional composed vectors ( INLINEFORM1 ). We train the INLINEFORM2 , INLINEFORM3 and INLINEFORM4 models as in Plank:2016. We train the CNN taggers with the same dimensionalities for word representations.", + "For the CRF tagger, we predict Pos and Morph jointly as in the standard setting for MarMot, which performs much better than with separate predictions, as shown in Mueller:2013 and in our preliminary experiments. Also, it splits the morphological tags into key-value pairs, whereas the neural taggers treat the whole string as a tag. We predict Stag as a separate task." + ], + [ + "The test results for the three tasks are shown in Table TABREF17 in three groups. The first group of seven columns are the results for Pos, where both LSTM and CNN have three variations of input features: word only ( INLINEFORM0 ), character only ( INLINEFORM1 ) and both ( INLINEFORM2 ). For Morph and Stag, we only use the INLINEFORM3 setting for both LSTM and CNN.", + "On macro-average, three taggers perform close in the Pos task, with the CNN tagger being slightly better. In the Morph task, CNN is again slightly ahead of CRF, while LSTM is about 2 points behind. In the Stag task, CNN outperforms both taggers by a large margin: 2 points higher than LSTM and 8 points higher than CRF.", + "While considering the input features of the LSTM and CNN taggers, both taggers perform close with only INLINEFORM0 as input, which suggests that the two taggers are comparable in encoding context for tagging Pos. However, with only INLINEFORM1 , CNN performs much better than LSTM (95.54 vs. 92.61), and close to INLINEFORM2 (96.18). Also, INLINEFORM3 consistently outperforms INLINEFORM4 for all languages. This suggests that the CNN model alone is capable of learning most of the information that the word-level model can learn, while the LSTM model is not.", + "The more interesting cases are Morph and Stag, where CNN performs much higher than LSTM. We hypothesize three possible reasons to explain the considerably large difference. First, the LSTM tagger may be more sensitive to hyper-parameters and requires task specific tuning. We use the same setting which is tuned for the Pos task, thus it underperforms in the other tasks. Second, the LSTM tagger may not deal well with large tag sets. The tag set size for Morph are larger than Pos in orders of magnitudes, especially for Czech, Basque, Finnish and Slovene, all of which have more than 1000 distinct Morph tags in the training data, and the LSTM performs poorly on these languages. Third, the LSTM has theoretically unlimited access to all the tokens in the sentence, but in practice it might not learn the context as good as the CNN. In the LSTM model, the information of long-distance contexts will gradually fade away during the recurrence, whereas in the CNN model, all words are treated equally as long as they are in the context window. Therefore the LSTM underperforms in the Stag task, where the information from long-distance context is more important." + ], + [ + "It is a common scenario to use a model trained with news data to process text from social media, which could include intentional or unintentional misspellings. Unfortunately, we do not have social media data to test the taggers. However, we design an experiment to simulate unnormalized text, by systematically editing the words in the dev sets with three operations: insertion, deletion and substitution. For example, if we modify a word abcdef at position 2 (0-based), the modified words would be abxcdef, abdef, and abxdef, where x is a random character from the alphabet of the language.", + "For each operation, we create a group of modified dev sets, where all words longer than two characters are edited by the operation with a probability of 0.25, 0.5, 0.75, or 1. For each language, we use the models trained on the normal training sets and predict Pos for the three groups of modified dev set. The average accuracies are shown in Figure FIGREF19 .", + "Generally, all models suffer from the increasing degrees of unnormalized texts, but CNN always suffers the least. In the extreme case where almost all words are unnormalized, CNN performs 4 to 8 points higher than LSTM and 4 to 11 points higher than CRF. This suggests that the CNN is more robust to misspelt words. While looking into the specific cases of misspelling, CNN is more sensitive to insertion and deletion, while CRF and LSTM are more sensitive to substitution." + ], + [ + "In this paper, we propose a general-purpose tagger that uses two CNNs for both character composition and context encoding. On the universal dependency treebanks (v1.2), the tagger achieves state-of-the-art results for POS tagging and morphological tagging, and to the best of our knowledge, it also performs best for supertagging. The tagger works well across different tagging tasks without tuning the hyper-parameters, and it is also robust against unnormalized text." + ] + ] +} +``` \ No newline at end of file diff --git a/qasper-2486/instruction.md b/qasper-2486/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..32b9f8411e2276d17ab3f3fdcf9684b230ea1899 --- /dev/null +++ b/qasper-2486/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: Emerging Language Spaces Learned From Massively Multilingual Corpora + +Question: What neural machine translation models can learn in terms of transfer learning? \ No newline at end of file diff --git a/qasper-2489/instruction.md b/qasper-2489/instruction.md new file mode 100644 index 0000000000000000000000000000000000000000..6f2554526966e56831ca51df1f03c2b4a3a37d57 --- /dev/null +++ b/qasper-2489/instruction.md @@ -0,0 +1,3 @@ +Name of Paper: An Annotation Scheme of A Large-scale Multi-party Dialogues Dataset for Discourse Parsing and Machine Comprehension + +Question: How large is the proposed dataset? \ No newline at end of file