paper_id
stringlengths 10
10
⌀ | yes_no
bool 2
classes | paper_index
int64 0
519
⌀ | evidence
stringlengths 0
37.7k
| question
stringlengths 4
11.7k
| answer
stringlengths 1
26k
|
|---|---|---|---|---|---|
null | false
| 124
|
The task of speculation detection and scope resolution is critical in distinguishing factual information from speculative information. This has multiple use-cases, like systems that determine the veracity of information, and those that involve requirement analysis. This task is particularly important to the biomedical domain, where patient reports and medical articles often use this feature of natural language. This task is commonly broken down into two subtasks: the first subtask, speculation cue detection, is to identify the uncertainty cue in a sentence, while the second subtask: scope resolution, is to identify the scope of that cue. For instance, consider the example:
It might rain tomorrow.
The speculation cue in the sentence above is ‘might’ and the scope of the cue ‘might’ is ‘rain tomorrow’. Thus, the speculation cue is the word that expresses the speculation, while the words affected by the speculation are in the scope of that cue.
This task was the CoNLL-2010 Shared Task (BIBREF0), which had 3 different subtasks. Task 1B was speculation cue detection on the BioScope Corpus, Task 1W was weasel identification from Wikipedia articles, and Task 2 was speculation scope resolution from the BioScope Corpus. For each task, the participants were provided the train and test set, which is henceforth referred to as Task 1B CoNLL and Task 2 CoNLL throughout this paper.
For our experimentation, we use the sub corpora of the BioScope Corpus (BIBREF1), namely the BioScope Abstracts sub corpora, which is referred to as BA, and the BioScope Full Papers sub corpora, which is referred to as BF. We also use the SFU Review Corpus (BIBREF2), which is referred to as SFU.
This subtask of natural language processing, along with another similar subtask, negation detection and scope resolution, have been the subject of a body of work over the years. The approaches used to solve them have evolved from simple rule-based systems (BIBREF3) based on linguistic information extracted from the sentences, to modern deep-learning based methods. The Machine Learning techniques used varied from Maximum Entropy Classifiers (BIBREF4) to Support Vector Machines (BIBREF5,BIBREF6,BIBREF7,BIBREF8), while the deep learning approaches included Recursive Neural Networks (BIBREF9,BIBREF10), Convolutional Neural Networks (BIBREF11) and most recently transfer learning-based architectures like Bidirectional Encoder Representation from Transformers (BERT) (BIBREF12). Figures FIGREF1 and FIGREF1 contain a summary of the papers addressing speculation detection and scope resolution (BIBREF13, BIBREF5, BIBREF9, BIBREF3, BIBREF14, BIBREF15, BIBREF16, BIBREF17, BIBREF6, BIBREF11, BIBREF18, BIBREF10, BIBREF19, BIBREF7, BIBREF4, BIBREF8).
Inspired by the most recent approach of applying BERT to negation detection and scope resolution (BIBREF12), we take this approach one step further by performing a comparative analysis of three popular transformer-based architectures: BERT (BIBREF20), XLNet (BIBREF21) and RoBERTa (BIBREF22), applied to speculation detection and scope resolution. We evaluate the performance of each model across all datasets via the single dataset training approach, and report all scores including inter-dataset scores (i.e. train on one dataset, evaluate on another) to test the generalizability of the models. This approach outperforms all existing systems on the task of speculation detection and scope resolution. Further, we jointly train on multiple datasets and obtain improvements over the single dataset training approach on most datasets.
Contrary to results observed on benchmark GLUE tasks, we observe XLNet consistently outperforming RoBERTa. To confirm this observation, we apply these models to the negation detection and scope resolution task, and observe a continuity in this trend, reporting state-of-the-art results on three of four datasets on the negation scope resolution task.
The rest of the paper is organized as follows: In Section 2, we provide a detailed description of our methodology and elaborate on the experimentation details. In Section 3, we present our results and analysis on the speculation detection and scope resolution task, using the single dataset and the multiple dataset training approach. In Section 4, we show the results of applying XLNet and RoBERTa on negation detection and scope resolution and propose a few reasons to explain why XLNet performs better than RoBERTa. Finally, the future scope and conclusion is mentioned in Section 5.
we take this approach one step further by performing a comparative analysis of three popular transformer-based architectures: BERT, XLNet and RoBERTa, applied to speculation detection and scope resolution.
|
What analysis do the authors perform?
|
A comparative analysis.
|
null | false
| null |
We are expecting a baby. Can you describe your experience and provide some pointers as a first time parent?
|
* Baby sleeps all the time. You might end up hearing this a lot from others but in reality it might not be true. As everything in life, it depends. Some babies “sleep like a baby” but others might not. Our baby only sleeps if we hold him in our hands and he also had some reflex issues. Due to this reflex issue, as soon as we put him in the bed, he wakes up immediately and started to cry as the milk comes up from his stomach to his mouth. We are at-least a bit lucky here as he sleeps on our hands. Some babies may have colic and cries with no reason. That would have been much worse situation to be in.
* You might be wondering how to hold the baby properly and how to change diapers and so on and forth. Trust me, you will learn all these things in no time.
* Babies mainly cries for the following three reasons. 1. When they are hungry 2. When they are sleepy 3. When they need diaper change. Hence you should rule out all these three before you escalate the baby crying issue.
* Team work is very important. Work with your partner. Plan, schedule your time and execute it. Taking care of baby is really a two or more persons job. Your partner needs all the help that is available. Utilise your parental leave.
* Don’t get stressed out if the baby do not reach the milestones that you read on the internet. Every baby is unique and they reach milestones in their own timeline
* It is very important to enjoy the journey as you will never get those moments back and trust me, these will be very precious moment in your life.
|
|
null | false
| null |
Rhapsa scotosialis, the slender owlet moth, is a moth of the family Noctuidae. This species is endemic to New Zealand and is found throughout the country. It is regarded as one of the most common forest moths found in New Zealand. The larval host species for R. scotosialis is Piper excelsum.
|
What is a Rhapsa scotosialis
|
A Rhapsa scotosialis, the slender owlet moth, is a moth of the family Noctuidae. This species is endemic to New Zealand and is found throughout the country. It is regarded as one of the most common forest moths found in New Zealand. The larval host species for R. scotosialis is Piper excelsum.
|
null | false
| 69
|
Text summarization is a task to generate a shorter and concise version of a text while preserving the meaning of the original text. The task can be divided into two subtask based on the approach: extractive and abstractive summarization. Extractive summarization is a task to create summaries by pulling out snippets of text form the original text and combining them to form a summary. Abstractive summarization asks to generate summaries from scratch without the restriction to use the available words from the original text. Due to the limitations of extractive summarization on incoherent texts and unnatural methodology BIBREF0 , the research trend has shifted towards abstractive summarization.
Sequence-to-sequence models BIBREF1 with attention mechanism BIBREF2 have found great success in generating abstractive summaries, both from a single sentence BIBREF3 and from a long document with multiple sentences BIBREF4 . However, when generating summaries, it is necessary to determine the main topic and to sift out unnecessary information that can be omitted. Sequence-to-sequence models have the tendency to include all the information, relevant or not, that are found in the original text. This may result to unconcise summaries that concentrates wrongly on irrelevant topics. The problem is especially severe when summarizing longer texts.
In this paper, we propose to use entities found in the original text to infer the summary topic, mitigating the aforementioned problem. Specifically, we leverage on linked entities extracted by employing a readily available entity linking system. The importance of using linked entities in summarization is intuitive and can be explained by looking at Figure 1 as an example. First (O1 in the Figure), aside from auxiliary words to construct a sentence, a summary is mainly composed of linked entities extracted from the original text. Second (O2), we can depict the main topic of the summary as a probability distribution of relevant entities from the list of entities. Finally (O3), we can leverage on entity commonsense learned from a separate large knowledge base such as Wikipedia.
To this end, we present a method to effectively apply linked entities in sequence-to-sequence models, called Entity2Topic (E2T). E2T is a module that can be easily attached to any sequence-to-sequence based summarization model. The module encodes the entities extracted from the original text by an entity linking system (ELS), constructs a vector representing the topic of the summary to be generated, and informs the decoder about the constructed topic vector. Due to the imperfections of current ELS's, the extracted linked entities may be too ambiguous and coarse to be considered relevant to the summary. We solve this issue by using entity encoders with selective disambiguation and by constructing topic vectors using firm attention.
We experiment on two datasets, Gigaword and CNN, with varying lengths. We show that applying our module to a sequence-to-sequence model with attention mechanism significantly increases its performance on both datasets. Moreover, when compared with the state-of-the-art models for each dataset, the model obtains a comparable performance on the Gigaword dataset where the texts are short, and outperforms all competing models on the CNN dataset where the texts are longer. Furthermore, we provide analysis on how our model effectively uses the extracted linked entities to produce concise and better summaries.
We show that applying our module to a sequence-to-sequence model with attention mechanism significantly increases its performance on both datasets.
|
How does the module perform?
|
When applying the module to a sequence-to-sequence model with attention mechanism, it significantly increases its performance on both datasets.
|
null | false
| null |
Tennis is played on a rectangular, flat surface. The court is 78 feet (23.77 m) long, and 27 feet (8.2 m) wide for singles matches and 36 ft (11 m) for doubles matches. Additional clear space around the court is required in order for players to reach overrun balls. A net is stretched across the full width of the court, parallel with the baselines, dividing it into two equal ends. It is held up by either a cord or metal cable of diameter no greater than 0.8 cm (1⁄3 in). The net is 3 feet 6 inches (1.07 m) high at the posts and 3 feet (0.91 m) high in the centre. The net posts are 3 feet (0.91 m) outside the doubles court on each side or, for a singles net, 3 feet (0.91 m) outside the singles court on each side.
|
From the passage provided, extract the length and width of tennis court in metres.
|
Length : 23.77 m
Width: 8.2 m
|
null | false
| null |
Classify each of the following as cold or hot: snow, fire, refrigerator, lava, stove top, freezer
|
Snow: Cold
Fire: Hot
Refrigerator: Cold
Lava: Hot
Stove Top: Hot
Freezer: Cold
|
|
null | false
| 137
|
We build our bilingual LMs, named RAMEN, starting from BERT$_{\textsc {base}}$, BERT$_{\textsc {large}}$, RoBERTa$_{\textsc {base}}$, and RoBERTa$_{\textsc {large}}$ pre-trained models. Using BERT$_{\textsc {base}}$ allows us to compare the results with mBERT model. Using BERT$_{\textsc {large}}$ and RoBERTa allows us to investigate whether the performance of the target LM correlates with the performance of the source LM. We evaluate our models on two cross-lingual zero-shot tasks: (1) Cross-lingual Natural Language Inference (XNLI) and (2) dependency parsing.
We choose BERT and RoBERTa (Liu et al., 2019) as the source models for building our bilingual language models, named RAMEN3 for the ease of discussion.
|
What's the name of the authors' bilingual language models?
|
RAMEN.
|
1711.05345
| false
| null |
The procedure of transfer learning in this work is straightforward and includes two steps. The first step is to pre-train the model on one MCQA dataset referred to as the source task, which usually contains abundant training data. The second step is to fine-tune the same model on the other MCQA dataset, which is referred to as the target task, that we actually care about, but that usually contains much less training data. The effectiveness of transfer learning is evaluated by the model's performance on the target task.
The first step is to pre-train the model on one MCQA dataset referred to as the source task, which usually contains abundant training data. The second step is to fine-tune the same model on the other MCQA dataset, which is referred to as the target task, that we actually care about, but that usually contains much less training data.
|
How different is the dataset size of source and target?
|
the training dataset is large while the target dataset is usually much smaller
|
null | false
| null |
What is the best bidding system in duplicate bridge?
|
Bidding systems in bridge allow the partnership to communicate the strength and various characteristics of their hand to arrive at an appropriate contract. A good bidding system should be descriptive, accommodate many different hand types, and account for interference from the opponents. There is no clear best bidding system as they all have advantages and disadvantages but 2/1 (two-over-one) is accepted to be the standard that people use. The reason is that it can describe many hand types efficiently through natural bidding, it is easy to learn and understand, and handle interference from opponents.
|
|
null | false
| null |
Which episode did Daenerys Targaryen lose her dragon, Viserion, to a White Walker's spear?
|
The Night King, the leader of the White Walkers, threw a spear at Viserion, which killed him.
|
|
null | false
| null |
Make a list of different reasons that a kite might not be able to be flown
|
You might not be able to fly a kite if there isn't any wind, if the strings are tangled, if the kite is broken or if the kite gets caught up in a tree
|
|
null | false
| 369
|
Finally, we created a combined temporal-linguistic model. Since both the linguistic and the temporal models were built using the same framework, it was fairly simple to combine the two models. The combined model was created by merging the linguistic and temporal corpora and vocabularies. (Recall that we treated temporal features as words). We then experimented with the same four methods as in the last two sections to create our combined models.
Table 4 shows the performance of each of these models. Across the board, the combined models outperformed their corresponding linguistic and temporal models, though the difference with the linguistic models were not as great. These results suggest that at some level the temporal and the linguistic "styles" of users provide non-overlapping cues about the identity of said users. Also, note that as with the linguistic and temporal models, our combined confusion model outperformed the other combined models.
Another way to evaluate the performance of the different combined models is through the rank-statistics plot. This is shown in Figure 2 . The figure shows the distribution of the ranks of the $5,612$ users for different combined models. The x-axis is the rank percentile (divided into bins of $5\%$ ), the y-axis is the percentage of the users that fall in each bin. For example, for the confusion model, $69\%$ (3880) of the $5,612$ users were correctly linked between Twitter and Facebook when looking at the top $5\%$ (281) of the predictions by the model. From the figure, you can clearly see that the confusion model is superior to the other models, with TF-IDF a close second. You can also see from the figure that the rank plot for the random baseline is a horizontal line, with each rank percentile bin having $5\%$ of the users ( $5\%$ because the rank percentiles were divided into bins of $5\%$ ).
The combined model was created by merging the linguistic and temporal corpora and vocabularies.
|
How do they create the combined model?
|
By merging the linguistic and temporal corpora and vocabularies.
|
null | false
| null |
Pick a few months with 31 days.
|
January, March, May, July, August, October and December have 31 days.
|
|
null | false
| null |
The match details, including runs and dismissals, are recorded by two official scorers, one representing each team. The scorers are directed by the hand signals of an umpire (see image, right). For example, the umpire raises a forefinger to signal that the batter is out (has been dismissed); he raises both arms above his head if the batter has hit the ball for six runs. The scorers are required by the Laws to record all runs scored, wickets taken and overs bowled; in practice, they also note significant amounts of additional data relating to the game
|
How does an umpire signal that a batter has been dismissed in a cricket match ?
|
An umpire signals that a batter has been dismissed by raising a forefinger.
|
1909.09986
| false
| null |
At training time, we use the plan-to-DFS mapping to perform the correct sequence of traversals, and train a neural classifier to act as a controller, choosing which action to perform at each step. At test time, we use the controller to guide the truncated DFS process. This mechanism is inspired by transition based parsing BIBREF5. The action set at each stage is dynamic. During traversal, it includes the available children at each stage and pop. Before traversals, it includes a choose-i action for each available node $n_i$. We assign a score to each action, normalize with softmax, and train to choose the desired one using cross-entropy loss. At test time, we either greedily choose the best action, or we can sample plans by sampling actions according to their assigned probabilities.
At training time, we use the plan-to-DFS mapping to perform the correct sequence of traversals, and train a neural classifier to act as a controller, choosing which action to perform at each step.
|
How is neural planning component trained?
|
The answers are shown as follows:
* plan-to-DFS mapping to perform the correct sequence of traversals, and train a neural classifier to act as a controller
|
null | false
| null |
Which human rights organisation founded 1961 got Nobel 1977
|
Amnesty International
|
|
null | false
| null |
How many provinces are in Canada?
|
Canada has 10 provinces and 3 territories. The most populous provinces are Ontario, Quebec, and British Columbia.
|
|
null | false
| null |
Who are the best chess players of all time?
|
It's difficult to determine who the best chess players of all time are as there are a variety of factors to consider when determining chess strength. For example, you might consider things like chess rating, number of tournament wins, time spent as world chess champion, etc. It's especially difficult to compare players of different eras, as more modern chess players will have learned from past Grandmasters and have more advanced technology available to facilitate their understanding of the game.
Three chess players commonly discussed as possibly being the best of all time include: Bobby Fischer, Gary Kasparov, and Magnus Carlsen. Bobby Fischer is well-known for winning the "Game of the Century" at the age of 13 and, as an adult, won 20 consecutive games against some of the best chess players in the world. Gary Kasparov was the number one rated chess player for over 20 years and was World Chess Champion for 15 consecutive years. Magnus Carlsen achieved the highest chess rating ever in 2014 with a rating of 2882 and has been the number one rated chess player since 2011.
|
|
null | false
| null |
Who was the first Grand Prix driver to used a safety belt in 1967
|
Jackie Stuart
|
|
null | false
| null |
Please give me a bulleted list of men's Roland Garros champions of the last 5 years.
|
Sure! Here's the list of men's Roland Garros champions of the last 5 years:
* 2022 - Rafael Nadal
* 2021 - Novak Djokovic
* 2020 - Rafael Nadal
* 2019 - Rafael Nadal
* 2018 - Rafael Nadal
|
|
null | false
| 44
|
We aimed to mimic the presence of sparse/noisy content distribution, mandating the need to curate a novel dataset via specific lexicons. We scraped 500 random posts from recognized transport forum. A pool of 50 uni/bi-grams was created based on tf-idf representations, extracted from the posts, which was further pruned by annotators. Querying posts on Twitter with extracted lexicons led to a collection of $19,300$ tweets. In order to have lexical diversity, we added 2500 randomly sampled tweets to our dataset. In spite of the sparse nature of these posts, the lexical characteristics act as information cues.
Figure FIGREF4 pictorially represents our methodology. Our approach required an initial set of informative tweets for which we employed two human annotators annotating a random sub-sample of the original dataset. From the 1500 samples, 326 were marked as informative and 1174 as non informative ($\kappa =0.81$), discriminated on this criteria: Is the tweet addressing any complaint or raising grievances about modes of transport or services/ events associated with transportation such as traffic; public or private transport?. An example tweet marked as informative: No, metro fares will be reduced ???, but proper fare structure needs to presented right, it's bad !!!.
We utilized tf-idf for the identification of initial seed phrases from the curated set of informative tweets. 50 terms having the highest tf-idf scores were passed through the complete dataset and based on sub-string match, the transport relevant tweets were identified. The redundant tweets were filtered based on the cosine similarity score. Implicit information indicators were identified based on domain relevance score, a metric used to gauge the coverage of n-gram (1,2,3) when evaluated against a randomly created pool of posts.
We collected a pool of 5000 randomly sampled tweets different from the data collection period. The rationale behind having such a metric was to discard commonly occurring n-grams normalized by random noise and include ones which are of lexical importance. We used terms associated with high domain relevance score (threshold determined experimentally) as seed phrases for the next set of iterations. The growing dictionary augments the collection process. The process ran for 4 iterations providing us 7200 transport relevant tweets as no new lexicons were identified. In order to identify linguistic signals associated with the complaint posts, we randomly sampled a set of 2000 tweets which was used as training set, manually annotated into distinct labels: complaint relevant (702) and complaint non-relevant (1298) ($\kappa =0.79$). We employed these features on our dataset.
Linguistic markers. To capture linguistic aspects of complaints, we utilized Bag of Words, count of POS tags and Word2vec clusters.
Sentiment markers. We used quantified score based on the ratio of tokens mentioned in the following lexicons: MPQA, NRC, VADER and Stanford.
Information specific markers. These account for a set of handcrafted features associated with complaint, we used the stated markers (a) Text-Meta Data, this includes the count of URL's, hashtags, user mentions, special symbols and user mentions, used to enhance retweet impact; (b) Request Identification, we employed the model presented in BIBREF3 to identify if a specific tweet assertion is a request; (c) Intensifiers, we make use of feature set derived from the number of words starting with capital letters and the repetition of special symbols (exclamation, questions marks) within the same post; (d) Politeness Markers, we utilize the politeness score of the tweet extracted from the model presented in BIBREF3; (e) Pronoun Variation, these have the ability to reveal the personal involvement or intensify involvement. We utilize the frequency of pronoun types $\lbrace \textit {first, second, third, demonstrative and indefinite}$} using pre-defined dictionaries.
From the pool of 7200 transport relevant tweets, we sampled 3500 tweets which were used as the testing set. The results are reported in TableTABREF5 with 10 fold cross-validation. With increasing the number of iterations, the pool of seed phrases gets refined and augments the selection of transport relevant tweets. The proposed pipeline is tailored to identify complaint relevant tweets in a noisy scenario.
From the pool of 7200 transport relevant tweets, we sampled 3500 tweets which were used as the testing set.
|
How many tweets were used as a testing set?
|
3500 tweets.
|
null | false
| null |
Margot Elise Robbie (/ˈmɑːrɡoʊ ˈrɒbi/; born 2 July 1990) is an Australian actress and producer. Known for her work in both blockbuster and independent films, she has received various awards and nominations, including nominations for two Academy Awards, four Golden Globe Awards, and five British Academy Film Awards. Time magazine named her one of the 100 most influential people in the world in 2017 and she was ranked as one of the world's highest-paid actresses by Forbes in 2019.
|
Given a reference text about Margot Robbie, tell me when they were born and what they are famous for.
|
Margot Robbie was born on July 2nd, 1990. She is famous for being an Australian actress and one of Time Magazine's 100 most influential people in the world in 2017.
|
null | false
| null |
What is the most popular breed of dog in the US?
|
The Golden Retriever is consistently in the top 3 most popular dog breeds in the US in annual surveys.
|
|
null | false
| null |
Write an introductory email to a student I will be academically coaching through Services for Students with Disabilities office (SSD).
|
Hi Naomi,
My name is Chandler, and I am your Academic Coach through the Services for Students with Disabilities office (SSD).
I am looking forward to working with you during this time in your academic journey! As your academic coach, my role is to partner with you as we focus on your process of learning.
During our time together we will examine your:
learning experience
habits of working
current difficulties or barriers to success
While examining these areas we will work together to:
Develop individualized learning strategies and study habits
Navigate campus resources
Promote self-advocacy
We will also focus on building and sustaining Executive Functioning skills such as:
Time Management
Organization
Self-regulation
Self-efficacy
I look forward to hearing back from you.
|
|
null | false
| null |
Luis González de Ubieta y González del Campillo (1899 – 1950) was an admiral of the Spanish Republican Navy during the Spanish Civil War. He died in exile as the captain of the Panamanian merchant vessel Chiriqui, refusing to be rescued when the ship under his command sank in the Caribbean Sea not far from Barranquilla.
|
Who was Luis Gonzalez de Ubieta
|
Luis Gonzalez de Ubieta y Gonzalez del Campillo (1899 – 1950) was an admiral of the Spanish Republican Navy during the Spanish Civil War. He died in exile as the captain of the Panamanian merchant vessel Chiriqui, refusing to be rescued when the ship under his command sank in the Caribbean Sea not far from Barranquilla.
|
null | false
| 33
|
We obtain definitions for each of the two terms we wish to measure. Due to the sparsity and inconsistencies of the definitions in the UMLS, we not only use the definition of the term (CUI) but also include the definition of its related concepts. This follows the method proposed by PatwardhanP06 for general English and WordNet, and which was adapted for the UMLS and the medical domain by LiuMPMP12. In particular we add the definitions of any concepts connected via a parent (PAR), child (CHD), RB (broader than), RN (narrower than) or TERM (terms associated with CUI) relation. All of the definitions for a term are combined into a single super–gloss. At the end of this process we should have two super–glosses, one for each term to be measured for relatedness.
Next, we process each super–gloss as follows:
We extract a first–order co–occurrence vector for each term in the super–gloss from the co–occurrence matrix created previously.
We take the average of the first order co–occurrence vectors associated with the terms in a super–gloss and use that to represent the meaning of the term. This is a second–order co–occurrence vector.
After a second–order co–occurrence vector has been constructed for each term, then we calculate the cosine between these two vectors to measure the relatedness of the terms.
Due to the sparsity and inconsistencies of the definitions in the UMLS, we not only use the definition of the term (CUI) but also include the definition of its related concepts.
|
Why do the team use both the definition of the term and the definition of its related concepts?
|
Because the definitions in the UMLS are featured with sparsity and inconsistencies.
|
null | false
| 295
|
The AUD model described in BIBREF0 , BIBREF1 is a non-parametric Bayesian Hidden Markov Model (HMM). This model is topologically equivalent to a phone-loop model with two major differences:
In this work, we have used two variants of this original model. The first one (called HMM model in the remainder of this paper), following the analysis led in BIBREF8 , approximates the Dirichlet Process prior by a mere symmetric Dirichlet prior. This approximation, while retaining the sparsity constraint, avoids the complication of dealing with the variational treatment of the stick breaking process frequent in Bayesian non-parametric models. The second variant, which we shall denote Structured Variational AutoEncoder (SVAE) AUD, is based upon the work of BIBREF4 and embeds the HMM model into the Variational AutoEncoder framework BIBREF9 . A very similar version of the SVAE for AUD was developed independently and presented in BIBREF5 . The main noteworthy difference between BIBREF5 and our model is that we consider a fully Bayesian version of the HMM embedded in the VAE; and the posterior distribution and the VAE parameters are trained jointly using the Stochastic Variational Bayes BIBREF4 , BIBREF10 . For both variants, the prior over the HMM parameters were set to the conjugate of the likelihood density: Normal-Gamma prior for the mean and variance of the Gaussian components, symmetric Dirichlet prior over the HMM's state mixture's weights and symmetric Dirichlet prior over the acoustic units' weights. For the case of the uninformative prior, the prior was set to be vague prior with one pseudo-observation BIBREF11 .
In this work, we have used two variants of this original model. The first one (called HMM model in the remainder of this paper), following the analysis led in [9], approximates the Dirichlet Process prior by a mere symmetric Dirichlet prior. This approximation, while retaining the sparsity constraint, avoids the complication of dealing with the variational treatment of the stick breaking process frequent in Bayesian non-parametric models. The second variant, which we shall denote Structured Variational AutoEncoder (SVAE) AUD, is based upon the work of [5] and embeds the HMM model into the Variational AutoEncoder framework [10].
|
What are the two variants used in this work?
|
The first one is called HMM model. The second variant is Structured Variational AutoEncoder (SVAE) AUD.
|
null | false
| null |
Which of the following countries are in Asia: India, South Korea, Japan, United States of America, Morocco, New Zealand, Indonesia, Singapore, Mexico?
|
The Asian countries are - India, South Korea, Japan, Indonesia, Singapore.
|
|
null | false
| null |
Cricket is a bat-and-ball game played between two teams of eleven players on a field at the centre of which is a 22-yard (20-metre) pitch with a wicket at each end, each comprising two bails balanced on three stumps. The batting side scores runs by striking the ball bowled at one of the wickets with the bat and then running between the wickets, while the bowling and fielding side tries to prevent this (by preventing the ball from leaving the field, and getting the ball to either wicket) and dismiss each batter (so they are "out"). Means of dismissal include being bowled, when the ball hits the stumps and dislodges the bails, and by the fielding side either catching the ball after it is hit by the bat, but before it hits the ground, or hitting a wicket with the ball before a batter can cross the crease in front of the wicket. When ten batters have been dismissed, the innings ends and the teams swap roles. The game is adjudicated by two umpires, aided by a third umpire and match referee in international matches. They communicate with two off-field scorers who record the match's statistical information.
Forms of cricket range from Twenty20, with each team batting for a single innings of 20 overs (each "over" being a set of 6 fair opportunities for the batting team to score) and the game generally lasting three hours, to Test matches played over five days. Traditionally cricketers play in all-white kit, but in limited overs cricket they wear club or team colours. In addition to the basic kit, some players wear protective gear to prevent injury caused by the ball, which is a hard, solid spheroid made of compressed leather with a slightly raised sewn seam enclosing a cork core layered with tightly wound string.
The earliest reference to cricket is in South East England in the mid-16th century. It spread globally with the expansion of the British Empire, with the first international matches in the second half of the 19th century. The game's governing body is the International Cricket Council (ICC), which has over 100 members, twelve of which are full members who play Test matches. The game's rules, the Laws of Cricket, are maintained by Marylebone Cricket Club (MCC) in London. The sport is followed primarily in South Asia, Australasia, the United Kingdom, Southern Africa and the West Indies.
Women's cricket, which is organised and played separately, has also achieved international standard. The most successful side playing international cricket is Australia, which has won seven One Day International trophies, including five World Cups, more than any other country and has been the top-rated Test side more than any other country.
|
what are the rules of cricket ?
|
Cricket is a bat-and-ball game played between two teams of eleven players on a field at the centre of which is a 22-yard (20-metre) pitch with a wicket at each end, each comprising two bails balanced on three stumps.Forms of cricket range from Twenty20, with each team batting for a single innings of 20 overs (each "over" being a set of 6 fair opportunities for the batting team to score) and the game generally lasting three hours, to Test matches played over five days.The batting side scores runs by striking the ball bowled at one of the wickets with the bat and then running between the wickets, while the bowling and fielding side tries to prevent this (by preventing the ball from leaving the field, and getting the ball to either wicket) and dismiss each batter (so they are "out").
|
null | false
| null |
When talking about paint colours, which one of these colours is considered "cold", "warm" or "both". Both meaning the way the colour is perceived depends on the colours surrounding it.
Yellow, Yellow-Orange, Orange, Red-Orange, Red, Red-Violet, Violet, Blue-Violet, Blue, Blue-Green, Green, Yellow-Green
|
Warm: Yellow, Yellow-Orange, Orange, Red-Orange, Red
Cold: Violet, Blue-Violet, Blue, Blue-Green, Green
Both: Red-Violet, Yellow-Green
|
|
null | false
| 365
|
Words with multiple senses commonly exist in many languages. For example, the word bank can either mean a “financial establishment” or “the land alongside or sloping down to a river or lake”, based on different contexts. Such a word is called a “polyseme”. The task to identify the meaning of a polyseme in its surrounding context is called word sense disambiguation (WSD). Word sense disambiguation is a long-standing problem in natural language processing (NLP), and has broad applications in other NLP problems such as machine translation BIBREF0 . Lexical sample task and all-word task are the two main branches of WSD problem. The former focuses on only a pre-selected set of polysemes whereas the later intends to disambiguate every polyseme in the entire text. Numerous works have been devoted in WSD task, including supervised, unsupervised, semi-supervised and knowledge based learning BIBREF1 . Our work focuses on using supervised learning to solve all-word WSD problem.
Most supervised approaches focus on extracting features from words in the context. Early approaches mostly depend on hand-crafted features. For example, IMS by BIBREF2 uses POS tags, surrounding words and collections of local words as features. These approaches are later improved by combining with word embedding features BIBREF0 , which better represents the words' semantic information in a real-value space. However, these methods neglect the valuable positional information between the words in the sequence BIBREF3 . The bi-directional Long-Short-Term-Memory (LSTM) approach by BIBREF3 provides one way to leverage the order of words. Recently, BIBREF4 improved the performance by pre-training a LSTM language model with a large unlabelled corpus, and using this model to generate sense vectors for further WSD predictions. However, LSTM significantly increases the computational complexity during the training process.
The development of the so called “fixed-size ordinally forgetting encoding” (FOFE) has enabled us to consider more efficient method. As firstly proposed in BIBREF5 , FOFE provides a way to encode the entire sequence of words of variable length into an almost unique fixed-size representation, while also retain the positional information for words in the sequence. FOFE has been applied to several NLP problems in the past, such as language model BIBREF5 , named entity recognition BIBREF6 , and word embedding BIBREF7 . The promising results demonstrated by the FOFE approach in these areas inspired us to apply FOFE in solving the WSD problem. In this paper, we will first describe how FOFE is used to encode sequence of any length into a fixed-size representation. Next, we elaborate on how a pseudo language model is trained with the FOFE encoding from unlabelled data for the purpose of context abstraction, and how a classifier for each polyseme is built from context abstractions of its labelled training data. Lastly, we provide the experiment results of our method on several WSD data sets to justify the equivalent performance as the state-of-the-art approach.
As firstly proposed in BIBREF5 , FOFE provides a way to encode the entire sequence of words of variable length into an almost unique fixed-size representation, while also retain the positional information for words in the sequence.
|
In their method is the positional information for words in the sequence retained?
|
Yes.
|
null | false
| null |
In the series A Song of Ice and Fire, who is the founder of House Slynt?
|
Janos Slynt
|
|
null | false
| null |
Why do players smear black under their eyes?
|
Black eye smear reduces the blinding sunlight when looking up towards the sun, when catching/hitting the ball
|
|
null | false
| null |
What is the best age to drive a car?
|
In U.S legal age to drive is 16 years old. You will need to pass a written test and driving test in DMV to get your driver's license. However, I will not recommend driving at 16 because the insurance cost for young teen drivers is very high. It can be over $3000 per year and some states can even reach $5000. Also, most 16 years old is not prepared to handle the mental and physical demands of driving, their also easy to blend in their emotion when they are driving. So, my suggestion is to wait until they are 18 years old to drive a car.
|
|
null | false
| null |
what is Kubernetes?
|
Kubernetes is an open-source container orchestration system for automating software deployment, scaling, and management. Originally designed by Google, the project is now maintained by the Cloud Native Computing Foundation.
|
|
null | false
| 350
|
Machine reading comprehension and question answering has becomes a crucial application problem in evaluating the progress of AI system in the realm of natural language processing and understanding BIBREF0 . The computational linguistics communities have devoted significant attention to the general problem of machine reading comprehension and question answering.
However, most of existing reading comprehension tasks only focus on shallow QA tasks that can be tackled very effectively by existing retrieval-based techniques BIBREF1 . For example, recently we have seen increased interest in constructing extractive machine reading comprehension datasets such as SQuAD BIBREF2 and NewsQA BIBREF3 . Given a document and a question, the expected answer is a short span in the document. Question context usually contains sufficient information for identifying evidence sentences that entail question-answer pairs. For example, 90.2% questions in SQuAD reported by Min BIBREF4 are answerable from the content of a single sentence. Even in some multi-turn conversation tasks, the existing models BIBREF5 mostly focus on retrieval-based response matching.
In this paper, we focus on multiple-choice reading comprehension datasets such as RACE BIBREF6 in which each question comes with a set of answer options. The correct answer for most questions may not appear in the original passage which makes the task more challenging and allow a rich type of questions such as passage summarization and attitude analysis. This requires a more in-depth understanding of a single document and leverage external world knowledge to answer these questions. Besides, comparing to traditional reading comprehension problem, we need to fully consider passage-question-answer triplets instead of passage-question pairwise matching.
In this paper, we propose a new model, Dual Co-Matching Network, to match a question-answer pair to a given passage bidirectionally. Our network leverages the latest breakthrough in NLP: BERT BIBREF7 contextual embedding. In the origin BERT paper, the final hidden vector corresponding to first input token ([CLS]) is used as the aggregation representation and then a standard classification loss is computed with a classification layer. We think this method is too rough to handle the passage-question-answer triplet because it only roughly concatenates the passage and question as the first sequence and uses question as the second sequence, without considering the relationship between the question and the passage. So we propose a new method to model the relationship among the passage, the question and the candidate answer.
Firstly we use BERT as our encode layer to get the contextual representation of the passage, question, answer options respectively. Then a matching layer is constructed to get the passage-question-answer triplet matching representation which encodes the locational information of the question and the candidate answer matched to a specific context of the passage. Finally we apply a hierarchical aggregation method over the matching representation from word-level to sequence-level and then from sequence level to document-level. Our model improves the state-of-the-art model by 2.6 percentage on the RACE dataset with BERT base model and further improves the result by 3 percentage with BERT large model.
Firstly we use BERT as our encode layer to get the contextual representation of the passage, question, answer options respectively. Then a matching layer is constructed to get the passage-question-answer triplet matching representation which encodes the locational information of the question and the candidate answer matched to a specific context of the passage. Finally we apply a hierarchical aggregation method over the matching representation from word-level to sequence-level and then from sequence level to document-level.
|
What has been used as their encode layer?
|
BERT.
|
null | false
| null |
What years was the BMW 1M made?
|
The BMW 1M was made in 2011 only. It was only manufactured with a 6-speed manual transmission, none were made with an automatic transmission. The 1M was originally manufactured in 3 colors: Alpine White, Sapphire Black, and Valencia Orange.
|
|
null | false
| null |
What are the basic rules of baseball?
|
Baseball is a complex sport that can be confusing to newcomers. Baseball is a game played by two teams of 9 players on a large field with 4 bases. The team on defense has 9 players out in the field at once, which includes the pitcher and the catcher. The hitting team only has one player "at-bat." The pitcher throws the ball to the catcher, and the hitter attempts to hit the ball into the field and run to a base if they hit the ball into fair territory. If the player misses 3 times, they are out. If the pitcher throws a bad pitch 4 times, the hitter gets to go to first base automatically. The player can also get out if a fielder catches the ball in the air or tags first base before the runner gets there. The game itself is 9 innings of 3 outs for each team. Whoever has the most runs (when a player runs to all 4 bases) at the end of the game wins!
|
|
null | false
| null |
Identify which animal species is alive or extinct: Spinosaurus, Common Bottlenose Dolphin
|
Common Bottlenose Dolphin is alive, Spinosaurus is extinct.
|
|
null | false
| 70
|
Since, we there was no publicly available standard Nepali NER dataset and did not receive any dataset from the previous researchers, we had to create our own dataset. This dataset contains the sentences collected from daily newspaper of the year 2015-2016. This dataset has three major classes Person (PER), Location (LOC) and Organization (ORG). Pre-processing was performed on the text before creation of the dataset, for example all punctuations and numbers besides ',', '-', '|' and '.' were removed. Currently, the dataset is in standard CoNLL-2003 IO formatBIBREF25.
Since, this dataset is not lemmatized originally, we lemmatized only the post-positions like Ek, kO, l, mA, m, my, jF, sg, aEG which are just the few examples among 299 post positions in Nepali language. We obtained these post-positions from sanjaalcorps and added few more to match our dataset. We will be releasing this list in our github repository. We found out that lemmatizing the post-positions boosted the F1 score by almost 10%.
In order to label our dataset with POS-tags, we first created POS annotated dataset of 6946 sentences and 16225 unique words extracted from POS-tagged Nepali National Corpus and trained a BiLSTM model with 95.14% accuracy which was used to create POS-tags for our dataset.
The dataset released in our github repository contains each word in newline with space separated POS-tags and Entity-tags. The sentences are separated by empty newline. A sample sentence from the dataset is presented in table FIGREF13.
Since, we there was no publicly available standard Nepali NER dataset and did not receive any dataset from the previous researchers, we had to create our own dataset. This dataset contains the sentences collected from daily newspaper of the year 2015-2016. This dataset has three major classes Person (PER), Location (LOC) and Organization (ORG).
|
What dataset is the experiment conducted on?
|
They had to create their own dataset which contains the sentences collected from daily newspaper of the year 2015-2016.
|
null | false
| 26
|
One of the most fundamental topics in natural language processing is how best to derive high-level representations from constituent parts, as natural language meanings are a function of their constituent parts. How best to construct a sentence representation from distributed word embeddings is an example domain of this larger issue. Even though sequential neural models such as recurrent neural networks (RNN) BIBREF0 and their variants including Long Short-Term Memory (LSTM) BIBREF1 and Gated Recurrent Unit (GRU) BIBREF2 have become the de-facto standard for condensing sentence-level information from a sequence of words into a fixed vector, there have been many lines of research towards better sentence representation using other neural architectures, e.g. convolutional neural networks (CNN) BIBREF3 or self-attention based models BIBREF4 .
From a linguistic point of view, the underlying tree structure—as expressed by its constituency and dependency trees—of a sentence is an integral part of its meaning. Inspired by this fact, some recursive neural network (RvNN) models are designed to reflect the syntactic tree structure, achieving impressive results on several sentence-level tasks such as sentiment analysis BIBREF5 , BIBREF6 , machine translation BIBREF7 , natural language inference BIBREF8 , and discourse relation classification BIBREF9 .
However, some recent works have BIBREF10 , BIBREF11 proposed latent tree models, which learn to construct task-specific tree structures without explicit supervision, bringing into question the value of linguistically-motivated recursive neural models. Witnessing the surprising performance of the latent tree models on some sentence-level tasks, there arises a natural question: Are linguistic tree structures the optimal way of composing sentence representations for NLP tasks?
In this paper, we demonstrate that linguistic priors are in fact useful for devising effective neural models for sentence representations, showing that our novel architecture based on constituency trees and their tag information obtains superior performance on several sentence-level tasks, including sentiment analysis and natural language inference.
A chief novelty of our approach is that we introduce a small separate tag-level tree-LSTM to control the composition function of the existing word-level tree-LSTM, which is in charge of extracting helpful syntactic signals for meaningful semantic composition of constituents by considering both the structures and linguistic tags of constituency trees simultaneously. In addition, we demonstrate that applying a typical LSTM to preprocess the leaf nodes of a tree-LSTM greatly improves the performance of the tree models. Moreover, we propose a clustered tag set to replace the existing tags on the assumption that the original syntactic tags are too fined-grained to be useful in neural models.
In short, our contributions in this work are as follows:
A chief novelty of our approach is that we introduce a small separate tag-level tree-LSTM to control the composition function of the existing word-level tree-LSTM, which is in charge of extracting helpful syntactic signals for meaningful semantic composition of constituents by considering both the structures and linguistic tags of constituency trees simultaneously.
|
What's the function of the small separate tag-level tree-LSTM?
|
To control the composition function of the existing word-level tree-LSTM, which is in charge of extracting helpful syntactic signals for meaningful semantic composition of constituents by considering both the structures and linguistic tags simultaneously.
|
null | false
| null |
Why do people like sweet treats so much?
|
Sugar has been added heavily to the American diet. It's in foods you wouldn't even suspect, such as spaghetti sauce or juices, even though they are already sweet. Sugar is also highly addictive. It gives your brain a high, just like other addictive substances. Additionally, it does taste great, so the taste alone keeps someone coming back for more.
|
|
null | false
| 53
|
To show the effects of using the hybrid training data, we trained two models of the same architecture and hyperparameters. One model was trained on the hybrid data and will be referred to as hybrid NER model. The other model was trained on clinical notes only and will be referred to as i2b2 NER model. We evaluated the performance of the NER models by micro-F1 score on the test set of both the synthesized queries and the i2b2 data.
We used the BiLSTM-CRF implementation provided by the flair package BIBREF16. We set the hidden size value to be 256 in the LSTM structure and left everything else at default values for the SequenceTagger model on flair. For word embeddings, we used the ELMo embeddings fine-tuned on PubMed articles and flair embeddings BIBREF13 trained on $5\%$ of PubMed abstracts , respectively. We trained models for 10 epochs and experimented with different learning rate, mini batch size, and dropouts. We ran hyperparameter optimization tests to find the best combination. $S_c$ is set to be 0.6 in our experiment.
To show the effects of using the hybrid training data, we trained two models of the same architecture and hyperparameters. One model was trained on the hybrid data and will be referred to as hybrid NER model. The other model was trained on clinical notes only and will be referred to as i2b2 NER model. We evaluated the performance of the NER models by micro-F1 score on the test set of both the synthesized queries and the i2b2 data.
|
To demonstrate the effect of using mixed training data, what are the two models trained by the author to do respectively?
|
One model was trained on the hybrid data and will be referred to as hybrid NER model. The other model was trained on clinical notes only and will be referred to as i2b2 NER model. They evaluated the performance of the NER models by micro-F1 score on the test set of both the synthesized queries and the i2b2 data.
|
null | false
| null |
Domestically, Barcelona has won a record 76 trophies: 26 La Liga, 31 Copa del Rey, fourteen Supercopa de España, three Copa Eva Duarte, and two Copa de la Liga titles, as well as being the record holder for the latter four competitions. In international club football, the club has won 22 European and worldwide titles: five UEFA Champions League titles, a record four UEFA Cup Winners' Cups, a joint record five UEFA Super Cups, a record three Inter-Cities Fairs Cups, a joint record two Latin Cups and three FIFA Club World Cups.
|
From the passage provided extract how many European and worldwide tittles Barcelona has won.
|
Barcelona has won 22 European and worldwide titles
|
null | false
| 121
|
Until recently, efforts to bring out semantics from structured-data relied heavily on expert knowledge BIBREF22, BIBREF8. For example, in order to better transcribe numerical time series of weather data to a textual forecast, Reiter et al. BIBREF8 devise complex template schemes in collaboration with weather experts to build a consistent set of data-to-word rules.
Modern approaches to the wide range of tasks based on structured-data (e.g. table retrieval BIBREF2, BIBREF23, table classification BIBREF24, question answering BIBREF25) now propose to leverage progress in deep learning to represent these data into a semantic vector space (also called embedding space). In parallel, an emerging task, called “data-to-text", aims at describing structured data into a natural language description. This task stems from the neural machine translation (NMT) domain, and early work BIBREF26, BIBREF27, BIBREF10 represent the data records as a single sequence of facts to be entirely translated into natural language. Wiseman et al. BIBREF10 show the limits of traditional NMT systems on larger structured-data, where NMT systems fail to accurately extract salient elements.
To improve these models, a number of work BIBREF28, BIBREF12, BIBREF29 proposed innovating decoding modules based on planning and templates, to ensure factual and coherent mentions of records in generated descriptions. For example, Puduppully et al. BIBREF12 propose a two-step decoder which first targets specific records and then use them as a plan for the actual text generation. Similarly, Li et al. BIBREF28 proposed a delayed copy mechanism where their decoder also acts in two steps: 1) using a classical LSTM decoder to generate delexicalized text and 2) using a pointer network BIBREF30 to replace placeholders by records from the input data.
Closer to our work, very recent work BIBREF11, BIBREF19, BIBREF13 have proposed to take into account the data structure. More particularly, Puduppully et al. BIBREF13 follow entity-centric theories BIBREF31, BIBREF32 and propose a model based on dynamic entity representation at decoding time. It consists in conditioning the decoder on entity representations that are updated during inference at each decoding step. On the other hand, Liu et al. BIBREF11, BIBREF19 rather focus on introducing structure into the encoder. For instance, they propose a dual encoder BIBREF19 which encodes separately the sequence of element names and the sequence of element values. These approaches are however designed for single-entity data structures and do not account for delimitation between entities.
Our contribution differs from previous work in several aspects. First, instead of flatly concatenating elements from the data-structure and encoding them as a sequence BIBREF11, BIBREF12, BIBREF10, we constrain the encoding to the underlying structure of the input data, so that the delimitation between entities remains clear throughout the process. Second, unlike all works in the domain, we exploit the Transformer architecture BIBREF21 and leverage its particularity to directly compare elements with each others in order to avoid arbitrary assumptions on their ordering. Finally, in contrast to BIBREF33, BIBREF13 that use a complex updating mechanism to obtain a dynamic representation of the input data and its entities, we argue that explicit hierarchical encoding naturally guides the decoding process via hierarchical attention.
On the other hand, Liu et al. [18,17] rather focus on introducing structure into the encoder. For instance, they propose a dual encoder [17] which encodes separately the sequence of element names and the sequence of element values. These approaches are however designed for single-entity data structures and do not account for delimitation between entities.
|
Does the dual encoder proposed by Liu et al.account for delimitation between entities?
|
No.
|
null | false
| 225
|
Unsupervised learning of cross-lingual word embedding offers elegant matching of words across languages, but has fundamental limitations in translating sentences. In this paper, we propose simple yet effective methods to improve word-by-word translation of cross-lingual embeddings, using only monolingual corpora but without any back-translation. We integrate a language model for context-aware search, and use a novel denoising autoencoder to handle reordering. Our system surpasses state-of-the-art unsupervised neural translation systems without costly iterative training. We also analyze the effect of vocabulary size and denoising type on the translation performance, which provides better understanding of learning the cross-lingual word embedding and its usage in translation.
In this paper, we propose simple yet effective methods to improve word-by-word translation of crosslingual embeddings, using only monolingual corpora but without any back-translation.
|
Whether using only monolingual data that the proposed models can be efficiently trained?
|
Yes.
|
null | false
| 104
|
As discussed in section SECREF24 , our work is closely related to the crosslingual unsupervised SRL work of titovcrosslingual. The idea of using superlingual latent variables to capture cross-lingual information was proposed for POS tagging by naseem2009multilingual, which we use here for SRL. In a semi-supervised setting, pado2009cross used a graph based approach to transfer semantic role annotations from English to German. furstenau2009graph used a graph alignment method to measure the semantic and syntactic similarity between dependency tree arguments of known and unknown verbs.
For monolingual unsupervised SRL, swier2004unsupervised presented the first work on a domain-general corpus, the British National Corpus, using 54 verbs taken from VerbNet. garg2012unsupervised proposed a Bayesian model for this problem that we use here. titov2012bayesian also proposed a closely related Bayesian model. grenager2006unsupervised proposed a generative model but their parameter space consisted of all possible linkings of syntactic constituents and semantic roles, which made unsupervised learning difficult and a separate language-specific rule based method had to be used to constrain this space. Other proposed models include an iterative split-merge algorithm BIBREF18 and a graph-partitioning based approach BIBREF1 . marquez2008semantic provide a good overview of the supervised SRL systems.
The idea of using superlingual latent variables to capture crosslingual information was proposed for POS tagging by Naseem et al. (2009), which we use here for SRL.
|
For POS tagging, what idea has been proposed by Naseem et al.?
|
Using superlingual latent variables to capture crosslingual information.
|
1604.00125
| false
| null |
To evaluate the summarization performance of AttSum, we implement rich extractive summarization methods. Above all, we introduce two common baselines. The first one just selects the leading sentences to form a summary. It is often used as an official baseline of DUC, and we name it “LEAD”. The other system is called “QUERY_SIM”, which directly ranks sentences according to its TF-IDF cosine similarity to the query. In addition, we implement two popular extractive query-focused summarization methods, called MultiMR BIBREF2 and SVR BIBREF20 . MultiMR is a graph-based manifold ranking method which makes uniform use of the sentence-to-sentence relationships and the sentence-to-query relationships. SVR extracts both query-dependent and query-independent features and applies Support Vector Regression to learn feature weights. Note that MultiMR is unsupervised while SVR is supervised. Since our model is totally data-driven, we introduce a recent summarization system DocEmb BIBREF9 that also just use deep neural network features to rank sentences. It initially works for generic summarization and we supplement the query information to compute the document representation.
To verify the effectiveness of the joint model, we design a baseline called ISOLATION, which performs saliency ranking and relevance ranking in isolation. Specifically, it directly uses the sum pooling over sentence embeddings to represent the document cluster. Therefore, the embedding similarity between a sentence and the document cluster could only measure the sentence saliency. To include the query information, we supplement the common hand-crafted feature TF-IDF cosine similarity to the query. This query-dependent feature, together with the embedding similarity, are used in sentence ranking. ISOLATION removes the attention mechanism, and mixtures hand-crafted and automatically learned features. All these methods adopt the same sentence selection process illustrated in Section "Sentence Selection" for a fair comparison.
Above all, we introduce two common baselines. The first one just selects the leading sentences to form a summary. It is often used as an official baseline of DUC, and we name it “LEAD”. The other system is called “QUERY_SIM”, which directly ranks sentences according to its TF-IDF cosine similarity to the query.
Above all, we introduce two common baselines. The first one just selects the leading sentences to form a summary. It is often used as an official baseline of DUC, and we name it “LEAD”. The other system is called “QUERY_SIM”, which directly ranks sentences according to its TF-IDF cosine similarity to the query.
In addition, we implement two popular extractive query-focused summarization methods, called MultiMR BIBREF2 and SVR BIBREF20 .
Since our model is totally data-driven, we introduce a recent summarization system DocEmb BIBREF9 that also just use deep neural network features to rank sentences.
To verify the effectiveness of the joint model, we design a baseline called ISOLATION, which performs saliency ranking and relevance ranking in isolation.
|
What models do they compare to?
|
The answers are shown as follows:
* LEAD
* QUERY_SIM
* MultiMR
* SVR
* DocEmb
* ISOLATION
|
null | false
| null |
What should I do in Ho Chi Minh City, Vietnam for a day?
|
If you are visiting Ho Chi Minh City, Vietnam for a day, there are several must-visit attractions located in the heart of the city center.
One of the most famous tourist attractions is the Saigon Notre Dame Cathedral, which was constructed by French colonists in the 19th century and features a unique combination of Roman and Gothic styles that sets it apart from many other buildings in the city.
Another famous French-inspired architecture is Ho Chi Minh City Post Office. Every detail inside this tourist destination in Saigon surprises visitors with its delicate beauty. The unique design style from the roofs, arched doorways, tile floors… This is really a tourist destination in Saigon that you should not miss on your trip.
The Independence Palace, also known as the Reunification Convention Hall, is just a few minutes walk from the Notre Dame Cathedral. This iconic building is a symbol of the culture and history of the nation. Originally used as the home and workplace of the president of the Republic of Vietnam, it became a national monument when a North Vietnamese Army tank crashed through its gate during the Fall of Saigon on April 30, 1975, bringing an end to the 20-year war with the USA.
One of the most popular tourist attractions in Saigon is the Ben Thanh Market, which offers visitors hundreds of traditional and modern items to choose from. The market is not only a great place to find unique souvenirs, but it's also a food lover's paradise, where you can sample delicious traditional dishes at affordable prices.
For those who appreciate professional art performances, a visit to the Ho Chi Minh City Opera House is a must. This beautiful building, designed by multiple architects, boasts an impressive exterior that showcases its grandeur and majesty. Once inside, visitors are treated to a modern, sophisticated ambiance, complete with state-of-the-art lighting and sound technology. The Ho Chi Minh City Opera House is a must-see destination for anyone interested in the arts and modern architecture.
Another popular tourist destination in Saigon is the Vietnam History Museum, located on the former site of the Blanchard de la Brosse Museum. Built in 1929, the museum is home to tens of thousands of artifacts, 25,000 books, and many precious documents. Visitors can comfortably admire the historical value of these artifacts and gain knowledge about archeology, history, and ethnology. The Vietnam History Museum is an excellent place to explore and learn more about Vietnamese history and culture. With its impressive collection of artifacts and documents, visitors can gain a deeper understanding of the country's past and present. Whether you're a history buff or just looking for a unique cultural experience, a visit to the Vietnam History Museum is well worth your time.
If you're looking to immerse yourself in the vibrant and modern atmosphere of Ho Chi Minh City 1, a visit to Bui Vien Walking Street is a must. The street is famous for its bustling shopping and entertainment activities and is known for never sleeping. And there are many street artists, performing arts, and the vibrant sounds of the bars and pubs, all of which create a lively and modern city vibe. Tourists and locals alike flock to Bui Vien Street, not just for the energetic atmosphere, but also for the diverse and delicious selection of food on offer. Whether you're looking for a fun night out or a quick bite to eat, Bui Vien Street has something for everyone.
These attractions are conveniently located in the city center and offer a glimpse into Ho Chi Minh City's rich history and cultural heritage.
|
|
null | false
| 178
|
In this work, we use the ConceptNet BIBREF3 , WordNet BIBREF2 , and Microsoft Concept Graph (MCG) BIBREF11 , BIBREF12 knowledge bases for our ontology prediction experiments.
WordNet is a knowledge base (KB) of single words and relations between them such as hypernymy and meronymy. For our task, we use the hypernym relations only. ConceptNet is a KB of triples consisting of a left term $t_1$ , a relation $R$ , and a right term $t_2$ . The relations come from a fixed set of size 34. But unlike WordNet, terms in ConceptNet can be phrases. We focus on the Is-A relation in this work. MCG also consists of hierarchical relations between multi-word phrases, ranging from extremely general to specific. Examples from each dataset are shown in Table 1 .
For experiments involving unstructured text, we use the WaCkypedia corpus BIBREF13 .
In this work, we use the ConceptNet, WordNet , and Microsoft Concept Graph (MCG) knowledge bases for our ontology prediction experiments.
|
How many knowledge bases do they use for ontology prediction experiments?
|
Four.
|
null | false
| null |
Think up a bunch of different ways to exercise
|
Running, walking, jogging, weight lifting, swimming, yoga, biking, boxing, sit ups, pushups, planks, jumping jacks
|
|
null | false
| null |
Electric cars (or electric vehicles, EVs) have a smaller environmental footprint than conventional internal combustion engine vehicles (ICEVs). While aspects of their production can induce similar, less or alternative environmental impacts, they produce little or no tailpipe emissions, and reduce dependence on petroleum, greenhouse gas emissions, and health effects from air pollution. Electric motors are significantly more efficient than internal combustion engines and thus, even accounting for typical power plant efficiencies and distribution losses, less energy is required to operate an EV. Manufacturing batteries for electric cars requires additional resources and energy, so they may have a larger environmental footprint from the production phase. EVs also generate different impacts in their operation and maintenance. EVs are typically heavier and could produce more tire and road dust air pollution, but their regenerative braking could reduce such particulate pollution from brakes. EVs are mechanically simpler, which reduces the use and disposal of engine oil.
|
Does EV help address the energy crisis?
|
Although all cars have effects on other people, battery electric cars have major environmental benefits over conventional internal combustion engine vehicles (ICEVs), such as:
Elimination of harmful tailpipe pollutants such as various oxides of nitrogen, which kill thousands of people every year. EVs use 38 megajoules per 100 km in comparison to 142 megajoules per 100 km for ICE cars. Less CO2 emissions globally than fossil-fuelled cars, thus limiting climate change. Plug-in hybrids capture most of these benefits when they are operating in all-electric mode.
Electric cars have some disadvantages, such as:
Possible increased particulate matter emissions from tires compared to fossil-fueled cars. This is sometimes caused by the fact that most electric cars have a heavy battery, which means the car's tires are subjected to more wear. This is drastically reduced when EV-rated weight-specific tires are used on the EV which are built specifically for the extra weight.[citation needed] Devices to capture tyre particulates are being developed.
The brake pads, however, can be used less frequently than in non-electric cars, if regenerative braking is available and may thus sometimes produce less particulate pollution than brakes in non-electric cars. Also, some electric cars may have a combination of drum brakes and disc brakes, and drum brakes are known to cause less particulate emissions than disc brakes.
Reliance on rare-earth elements such as neodymium, lanthanum, terbium, and dysprosium, and other critical metals such as lithium and cobalt,though the quantity of rare metals used differs per car. Despite the name rare earth metals are plentiful.They make up a tiny share of the minerals used to make a car.
|
null | false
| null |
Classify each method of communication as "verbal" or "written": e-mail, phone call, slack chat, zoom meeting, text, post mail
|
E-mail: Written
Phone Call: Verbal
Slack Chat: Written
Zoom Meeting: Verbal
Text: Written
Post Mail: Written
|
|
null | false
| 41
|
Performance appraisal (PA) is an important HR process to periodically measure and evaluate every employee's performance vis-a-vis the goals established by the organization. A PA process involves purposeful multi-step multi-modal communication between employees, their supervisors and their peers, such as self-appraisal, supervisor assessment and peer feedback. Analysis of the structured data and text produced in PA is crucial for measuring the quality of appraisals and tracking actual improvements. In this paper, we apply text mining techniques to produce insights from PA text. First, we perform sentence classification to identify strengths, weaknesses and suggestions of improvements found in the supervisor assessments and then use clustering to discover broad categories among them. Next we use multi-class multi-label classification techniques to match supervisor assessments to predefined broad perspectives on performance. Finally, we propose a short-text summarization technique to produce a summary of peer feedback comments for a given employee and compare it with manual summaries. All techniques are illustrated using a real-life dataset of supervisor assessment and peer feedback text produced during the PA of 4528 employees in a large multi-national IT company.
In this paper, we apply text mining techniques to produce insights from PA text. First, we perform sentence classification to identify strengths, weaknesses and suggestions of improvements found in the supervisor assessments and then use clustering to discover broad categories among them. Next we use multi-class multi-label classification techniques to match supervisor assessments to predefined broad perspectives on performance. Finally, we propose a short-text summarization technique to produce a summary of peer feedback comments for a given employee and compare it with manual summaries.
|
How do they produce insights from PA text in the paper?
|
First, they perform sentence classification to identify strengths, weaknesses and suggestions of improvements found in the supervisor assessments and then use clustering to discover broad categories among them. Next they use multi-class multi-label classification techniques to match supervisor assessments to predefined broad perspectives on performance. Finally, they propose a short-text summarization technique to produce a summary of peer feedback comments for a given employee and compare it with manual summaries.
|
null | false
| null |
Which months of the year are considered summer, spring, winter and autumn.
|
Seasons depend on where you live. In the US, summer is usually June through August. Spring is March through May. Fall is September through November. Winter is December through February.
|
|
1702.06589
| true
| null |
Let $u,v \in \mathbb {R}^{d}$ be the sentence embeddings of question $q$ and of paraphrase $t$ . We experiment with the following similarity scores: i) DOTPRODUCT : $u^{T}v$ ; ii) BILIN : $u^{T}Sv$ , with $S\in \mathbb {R}^{d\times d}$ being a trainable matrix; iii) FC: u and v concatenated, followed by two sequential fully connected layers with ELU non-linearities; iv) FC-BILIN: weighted average of BILIN and FC. These models define parametrized similarity scoring functions $: Q\times T\rightarrow \mathbb {R}$ , where $Q$ is the set of natural language questions and $T$ is the set of paraphrases of logical forms.
These models define parametrized similarity scoring functions $: Q\times T\rightarrow \mathbb {R}$ , where $Q$ is the set of natural language questions and $T$ is the set of paraphrases of logical forms.
|
Does a neural scoring function take both the question and the logical form as inputs?
|
Yes.
|
null | false
| 85
|
In this paper, we have shown that using the designed RNN based network we could increase the performance of classification dramatically. We showed that keeping the sequential nature of the data can be hugely beneficial when working with textual data especially faced with the hard task of detecting more complex phenomena like emotions. We accomplished that by using a recurrent network in the process of generating our hidden representation. We have also used a max-pooling layer to capture the most relevant features and an average pooling layer to capture the text as a whole proving that we can achieve better performance by focusing on creating a more informative hidden representation. In future we can focus on improving these representations for example by using attention networks BIBREF37 , BIBREF38 to capture a more contextual representation or using language model based methods like BERT BIBREF39 that has been shown very successful in various NLP tasks.
In future we can focus on improving these representations for example by using attention networks to capture a more contextual representation or using language model based methods like BERT that has been shown very successful in various NLP tasks.
|
What is the author's plan for his future work?
|
They can focus on improving these representations for example by using attention networks to capture a more contextual representation or using language model based methods like BERT that has been shown very successful in various NLP tasks.
|
null | false
| null |
Macroglossum insipida, the hermit hummingbird hawkmoth, is a moth of the family Sphingidae. It was described by Arthur Gardiner Butler in 1875.
|
Was the hermit hummingbird hawkmoth described in the 20th or 21st century?
|
Neither. It was first described by Arthur Gardiner Butler in 1875, which is in the 19th century.
|
null | false
| null |
What are the healthiest cooking methods for preparing meals?
|
Steaming - cooking food by exposing retains nutrients and flavor.
Boiling - A healthy cooking method but boiling for too long causes the loss of some nutrients.
Grilling - cooking food over an open flame allows surplus fat to drip from the food.
Baking and roasting - involve cooking in an oven, which can retain nutrients and flavors.
Stir-frying - A healthy method of cooking food in high heat.
Poaching - involves cooking food in liquid at a low temperature.
The healthiest cooking methods involve minimal added fats, salt, and sugar.
|
|
null | false
| 20
|
As already stated, our main proposal is using neural embedding models to `project' documents in one language into the semantic space of another language. For this, we first trained a Continuous Bag-of-Words (CBOW) and a Continuous SkipGram model BIBREF0 for each of our monolingual subcorpora. The models were trained with identical hyperparameters: vector size of 300 components, symmetric window of 2 words, negative sampling with 10 samples, 5 iterations over the corpus, no down-sampling. The only language-dependent difference was that for the Ukrainian model we ignored words with the corpus frequency less than 10 and for the Russian model this threshold was set to 15 (as the Russian corpus is 3 times larger). All in all, the final Ukrainian model recognizes 429 215 words and the Russian one 271 720 words. Training was performed using CBOW and SkipGram implementation in Gensim library BIBREF7 .
After the models were trained, we followed the path outlined in BIBREF4 to learn a linear transformation matrix from Ukrainian to Russian. First, we extracted all noun pairs from Russian-Ukrainian bilingual dictionary BIBREF8 , with the constraint that their frequency in our corpora was above the already mentioned thresholds 15 and 10 for Russian and Ukrainian words correspondingly. That made it a list of about 5 thousand pairs of nouns being translations of each other.
For all these words, their vectors were found in the models corresponding to the words' languages. It provided us with a matrix of 5 thousand of 300-dimensional Ukrainian vectors and the matrix of corresponding 5 thousand of 300-dimensional Russian vectors. This data served as a training set to learn an optimal transformation matrix. The latter is actually a 300x301 matrix of coefficients, such that when the initial Ukrainian matrix is multiplied by this transformation matrix, the result is maximally close to the corresponding Russian matrix. This transformation matrix has 301 (not 300) columns, because we add one component equal to 1 to each vector, as a bias term.
Producing the transformation matrix is a linear regression problem: the input is 301 components of Ukrainian vectors (including the bias term) and the output is 300 components of Russian vectors. As we need 300 values as an output, there are actually 300 linear regression problems and that's why the resulting matrix size is 300x301 (301 weights for each of 300 components).
There are two main ways to solve a linear regression problem: one can either learn the optimal weights in an iterative way using some variant of gradient descent, or one can solve it numerically without iteration, using normal equation. For English and Spanish, BIBREF4 used stochastic gradient descent. However, normal equation is actually less error-prone and is guaranteed to find the global optimum. Its only disadvantage is that it becomes very computationally expensive when the number of features is large (thousands and more). However, in our case the number of features is only 301, so computational complexity is not an issue.
Thus, we use normal equation to find the optimal transformation matrix. The algebraic solution to each of 300 normal equations (one for each vector component $i$ ) is shown in the Equation 3 :
$$\beta _i = (\textbf {X}^\intercal * \textbf {X})^{-1} * \textbf {X}^\intercal * y_i$$ (Eq. 3)
where $\textbf {X}$ is the matrix of 5 thousand Ukrainian word vectors (input), $y_i$ is the vector of the $i$ th components of 5 thousand corresponding Russian words (correct predictions), and $\beta _i$ is our aim: the vector of 301 optimal coefficients which transform the Ukrainian vectors into the $i$ th component of the Russian vectors.
After solving such normal equations for all the 300 components $i$ , we have the 300x301 linear transformation matrix which fits the data best.
This matrix basically maps the Ukrainian vectors into the Russian ones. It is based on the assumption that the relations between semantic concepts in different languages are in fact very similar (students are close to teachers, while pirates are close to corsairs, and so on). In continuous distributional models which strive to represent these semantic spaces, mutual `geometrical' relations between vectors representing particular words are also similar across models (if they are trained on comparable corpora), but the exact vectors for words denoting one and the same notion are different. This is because the models themselves are stochastic and the particular values of vectors (unlike their positions in relation to each other) depend a lot on technical factors, including the random seed used to initialize vectors prior to training. In order to migrate from a model A to another model B, one has to `rotate and scale' A vectors in a uniform linear way. To learn the optimal transformation matrix means to find out the exact directions of rotating and scaling, which minimize prediction errors.
Linguistically speaking, once we learned the transformation matrix, we can predict what a Russian vector would most probably be, given a Ukrainian one. This essentially means we are able to `translate' Ukrainian words into Russian, by calculating the word in the Russian model with the vector closest to the predicted one.
We had to choose between CBOW or Continuous SkipGram models to use when learning the transformation matrix. Also, there was a question of whether to employ regularized or standard normal equations. Regularization is an attempt to avoid over-fitting by trying to somehow decrease the values of learned weights. The regularized normal equation is shown in 4 :
$$\beta _i = (\textbf {X}^\intercal * \textbf {X} + \lambda * L)^{-1} * \textbf {X}^\intercal * y_i$$ (Eq. 4)
Comparing to 3 , it adds the term $\lambda * L$ , where $L$ is the identity matrix of the size equal to the number of features, with 0 at the top left cell, and $\lambda $ is a real number used to tune the influence of regularization term (if $\lambda = 0$ , there is no regularization).
To test all the possible combinations of parameters, we divided the bilingual dictionary into 4500 noun pairs used as a training set and 500 noun pairs used as a test set. We then learned transformation matrices on the training set using both training algorithms (CBOW and SkipGram) and several values of regularization $\lambda $ from 0 to 5, with a step of 0.5. The resulting matrices were applied to the Ukrainian vectors from the test set and the corresponding Russian `translations' were calculated. The ratio of correct `translations' (matches) was used as an evaluation measure. It came out that regularization only worsened the results for both algorithms, so in the Table 1 we report the results without regularization.
For reference, we also report the accuracy of `quazi-translation' via Damerau-Levenshtein edit distance BIBREF9 , as a sort of a baseline. As already stated, the two languages share many cognates, and a lot of Ukrainian words can be orthographically transformed into their Russian translations (and vice versa) by one or two character replacements. Thus, we extracted 50,000 most frequent nouns from our Russian corpora; then for each Ukrainian noun in the bilingual dictionary we found the closest Russian noun (or 5 closest nouns for @5 metric) by edit distance and calculated how often it turned out to be the correct translation. As the Table 1 shows, notwithstanding the orthographic similarity of the two languages, CBOW consistently outperforms this approach even on the test set. On the training set, its superiority is even more obvious.
Thus, we use normal equation to find the optimal transformation matrix.
|
How to find the optimal transformation matrix?
|
Use normal equation.
|
null | false
| null |
The koala or, inaccurately, koala bear (Phascolarctos cinereus), is an arboreal herbivorous marsupial native to Australia. It is the only extant representative of the family Phascolarctidae and its closest living relatives are the wombats. The koala is found in coastal areas of the mainland's eastern and southern regions, inhabiting Queensland, New South Wales, Victoria, and South Australia. It is easily recognisable by its stout, tailless body and large head with round, fluffy ears and large, spoon-shaped nose. The koala has a body length of 60–85 cm (24–33 in) and weighs 4–15 kg (9–33 lb). Fur colour ranges from silver grey to chocolate brown. Koalas from the northern populations are typically smaller and lighter in colour than their counterparts further south. These populations possibly are separate subspecies, but this is disputed.
Koalas typically inhabit open Eucalyptus woodland, as the leaves of these trees make up most of their diet. This eucalypt diet has low nutritional and caloric content and contains toxins that deter most other mammals from feeding on it. Koalas are largely sedentary and sleep up to twenty hours a day. They are asocial animals, and bonding exists only between mothers and dependent offspring. Adult males communicate with loud bellows that intimidate rivals and attract mates. Males mark their presence with secretions from scent glands located on their chests. Being marsupials, koalas give birth to underdeveloped young that crawl into their mothers' pouches, where they stay for the first six to seven months of their lives. These young koalas, known as joeys, are fully weaned around a year old. Koalas have few natural predators and parasites, but are threatened by various pathogens, such as Chlamydiaceae bacteria and koala retrovirus.
|
Given this reference text about the Koala, tell me what makes up most of their diet, how long joeys spend in their mothers' pouches and what are some of the pathogens they are threatened by.
|
Eucalyptus makes up most of their diet. Joeys spend the first six to seven months of their lives in their mothers' pouches. They are threatened by pathogens like Chlamydiaceae and koala retrovirus.
|
null | false
| null |
The movement of electric charge is known as an electric current, the intensity of which is usually measured in amperes. Current can consist of any moving charged particles; most commonly these are electrons, but any charge in motion constitutes a current. Electric current can flow through some things, electrical conductors, but will not flow through an electrical insulator.
By historical convention, a positive current is defined as having the same direction of flow as any positive charge it contains, or to flow from the most positive part of a circuit to the most negative part. Current defined in this manner is called conventional current. The motion of negatively charged electrons around an electric circuit, one of the most familiar forms of current, is thus deemed positive in the opposite direction to that of the electrons. However, depending on the conditions, an electric current can consist of a flow of charged particles in either direction, or even in both directions at once. The positive-to-negative convention is widely used to simplify this situation.
The process by which electric current passes through a material is termed electrical conduction, and its nature varies with that of the charged particles and the material through which they are travelling. Examples of electric currents include metallic conduction, where electrons flow through a conductor such as metal, and electrolysis, where ions (charged atoms) flow through liquids, or through plasmas such as electrical sparks. While the particles themselves can move quite slowly, sometimes with an average drift velocity only fractions of a millimetre per second, the electric field that drives them itself propagates at close to the speed of light, enabling electrical signals to pass rapidly along wires.
In engineering or household applications, current is often described as being either direct current (DC) or alternating current (AC). These terms refer to how the current varies in time. Direct current, as produced by example from a battery and required by most electronic devices, is a unidirectional flow from the positive part of a circuit to the negative. If, as is most common, this flow is carried by electrons, they will be travelling in the opposite direction. Alternating current is any current that reverses direction repeatedly; almost always this takes the form of a sine wave. Alternating current thus pulses back and forth within a conductor without the charge moving any net distance over time. The time-averaged value of an alternating current is zero, but it delivers energy in first one direction, and then the reverse. Alternating current is affected by electrical properties that are not observed under steady state direct current, such as inductance and capacitance. These properties however can become important when circuitry is subjected to transients, such as when first energised.
|
What does current consist of?
|
Current can consist of any moving charged particles; most commonly these are electrons, but any charge in motion constitutes a current.
|
null | false
| null |
List five titles by George R.R. Martin.
|
A Game of Thrones, Dying of the Light, The Armageddon Rag, A Storm of Swords, A Song for Lya
|
|
null | false
| 338
|
A common task in natural language processing is to find the most frequently occurring word pairs in a text(s) in the expectation that these pairs will shed some light on the main ideas of the text, or offer insight into the structure of the language. One might be interested in pairings of adjacent words, but in some cases one is also interested in pairs of words in some small neighborhood. The neighborhood is usually refered to as a window, and to illustrate the concept consider the following text and bigram set:
Text: “I like kitties and doggies”
Window: 2
Bigrams: {(I like), (like kitties), (kitties and), (and doggies)} and this one:
Text: “I like kitties and doggies”
Window: 4
Bigrams: {(I like), (I kitties), (I and), (like kitties), (like and), (like doggies), (kitties and), (kitties doggies), (and doggies)}.
A common task in natural language processing is to find the most frequently occurring word pairs in a text(s) in the expectation that these pairs will shed some light on the main ideas of the text, or offer insight into the structure of the language.
|
Why is finding the most frequently occurring word pairs in a text a common task in NLP?
|
These pairs will shed some light on the main ideas of the text, or offer insight into the structure of the language
|
null | false
| 477
|
We use the filtered setting during evaluation on the standard training-validation-test split, randomly break ties for triplets with the same score (Sun et al., 2019b) and report Mean Rank (MR), Mean Reciprocal Rank (MRR), and Hit@K (K=1,3,10).
|
How are the ties resolved while ranking for knowledge graph completion task?
|
We randomly break ties for triplets with the same score (denoted as RANDOM in [1]). We have clarified this in Section 5.1.
[1] A Re-evaluation of Knowledge Graph Completion Methods. Sun et al., ACL 2020.
|
null | false
| null |
The Flash (Bartholomew Henry "Barry" Allen) is a superhero appearing in American comic books published by DC Comics. He is the second character known as the Flash, following Jay Garrick. The character first appeared in Showcase #4 (October 1956), created by writer Robert Kanigher and penciler Carmine Infantino.
Like other heroes who go by The Flash, Barry is a "speedster", with powers that derive mainly from his superhuman speed. He wears a distinct red and gold costume treated to resist friction and wind resistance, traditionally storing the costume compressed inside a ring. Originally created as a reimagining of the popular 1940s superhero The Flash (Jay Garrick), the success of the Barry Allen's Flash comic book helped to bring about the Silver Age of Comic Books, and contributed to a large growth in DC Comics' stable of science fiction comics and characters. During popular early volumes as the Flash, Barry established his own Rogues Gallery of colourful villains and sci-fi concepts like the Speed Force. Through crossovers with popular characters like Superman, Wonder Woman, and Batman, Barry Allen's Flash also helped establish DC's flagship Justice League title, whose success would define its publishing strategy for decades to come.
|
Name some of the characters from DC Comics.
|
Superman, Wonder Woman, Batman and Flash.
|
null | false
| 498
|
t-SNE. As shown in Figure 13(a), we visualize the memory state of ConvLSTM using t-SNE (Van der Maaten & Hinton, 2008). It is observed that the learned cell states by ConvLSTM are entangled among different action groups. The t-SNE visualization result matches the PhyDNet visualization on the RoboNet dataset shown in Figure 3(a). Thus, these results verify that the STMC also exists under the real-world human motion dataset. While in ModeRNN, we further visualize the learned features of the slot bus in Figure 13(b), which shows 2 clusters with clear boundaries, corresponding to two action groups in the KTH dataset. According to these t-SNE results, we can find that directly training the previous methods on the mixed dynamics will lead to severe STMC in representation learning, shown as the entanglement of hidden representations. These entangled representations will make the model provide a poor ambiguous prediction. In contrast, ModeRNN can effectively overcome the STMC by learning an accurate decoupling for mixed dynamics.****Figure 13: (a, c) Illustration of STMC on the existing ConvLSTM model on KTH and radar echo dataset of Guangzhou (GZ). (b, d) The slot bus of ModeRNN shows discriminative representations on different groups of video dynamics. The two groups in KTH respectively correspond to subtle hand motion (e.g., hand-waving, hand-clapping, and boxing) and more global body motion (e.g., running, walking, and jogging). The two groups in Radar are divided by different seasons.
|
The t-SNE visualization of Figures 3 and 5 for KTH and the Radar Echo dataset would be interesting. Are the modes as clearly separated?
|
We added these t-SNE results in Figure 13 in the Appendix. As shown: - STMC still exists for KTH and Radar Echo. - The slot bus of ModeRNN learns discriminative representations on different groups of video dynamics (KTH: body motion vs. hand motion; Radar Echo: winter vs. summer).
|
null | false
| null |
What are different ways of data collection?
|
Gathering data can be accomplished through a primary source (the researcher is the first person to obtain the data) or a secondary source (the researcher obtains the data that has already been collected by other sources, such as data disseminated in a scientific journal). Data analysis methodologies vary and include data triangulation and data percolation. The latter offers an articulate method of collecting, classifying, and analyzing data using five possible angles of analysis (at least three) to maximize the research's objectivity and permit an understanding of the phenomena under investigation as complete as possible: qualitative and quantitative methods, literature reviews (including scholarly articles), interviews with experts, and
|
|
null | false
| null |
Classify the following hikes in terms of difficulty: Half dome in Yosemite, Mt Everest in Nepal, Mission Peak in California
|
Mt Everest is the toughest among the list based on scaling height of 28k feet and requires multiple days of hike, acclimatization at different altitudes, requires carrying oxygen tanks and heavy gear.
Half-dome is also considered as a difficult hike but less tougher than Mt Everest. Half-dome can be done as a day-hike and is at 4k elevation gain from the Yosemite valley and with moderate fitness & stamina its possible to do.
Mission Peak is considered a medium difficulty hike. It takes between 2-3 hours to scale Mission peak, it can be strenuous for people with health difficulties and less stamina.
|
|
null | false
| null |
Does multitasking make someone more productive at work?
|
It is a commonly held belief that multitasking is a desirable attribute for an employee to possess. This is based on the idea that multitasking results in a higher volume of work product. However, the contrary view is that multitasking diminishes the quality of work. In other words, multitasking increases productivity if the goal is simply to produce more work product; but at the same time, it decreases the quality of each individual outcome.
|
|
null | false
| null |
Give me a list of 10 nematode species
|
Caenorhabditis elegans,
Ascaris lumbricoides,
Pristionchus pacificus,
Trichuris suis,
Necator americanus,
Ancylostoma duodenale,
Heligmosomoides polygyrus,
Nippostrongylus brasiliensis,
Lumbricus terrestris,
Ostertagia ostertagi
|
|
null | false
| null |
Give me a title of one of the "Harry Potter" novels.
|
Harry Potter and the Chamber of Secrets
|
|
null | false
| null |
Tell me if these are countries in the Americas or in Europe: Spain, Guyana, Peru, Netherlands, Belgium, Canada, French Guiana
|
Spain: Europe, Guyana: Americas, Peru: Americas, Netherlands: Europe, Belgium: Europe, Canada: Americas, French Guiana: Americas
|
|
null | false
| 104
|
Semantic Role Labeling (SRL) has emerged as an important task in Natural Language Processing (NLP) due to its applicability in information extraction, question answering, and other NLP tasks. SRL is the problem of finding predicate-argument structure in a sentence, as illustrated below:
INLINEFORM0
Here, the predicate WRITE has two arguments: `Mike' as A0 or the writer, and `a book' as A1 or the thing written. The labels A0 and A1 correspond to the PropBank annotations BIBREF0 .
As the need for SRL arises in different domains and languages, the existing manually annotated corpora become insufficient to build supervised systems. This has motivated work on unsupervised SRL BIBREF1 , BIBREF2 , BIBREF3 . Previous work has indicated that unsupervised systems could benefit from the word alignment information in parallel text in two or more languages BIBREF4 , BIBREF5 , BIBREF6 . For example, consider the German translation of sentence INLINEFORM0 :
INLINEFORM0
If sentences INLINEFORM0 and INLINEFORM1 have the word alignments: Mike-Mike, written-geschrieben, and book-Buch, the system might be able to predict A1 for Buch, even if there is insufficient information in the monolingual German data to learn this assignment. Thus, in languages where the resources are sparse or not good enough, or the distributions are not informative, SRL systems could be made more accurate by using parallel data with resource rich or more amenable languages.
In this paper, we propose a joint Bayesian model for unsupervised semantic role induction in multiple languages. The model consists of individual Bayesian models for each language BIBREF3 , and crosslingual latent variables to incorporate soft role agreement between aligned constituents. This latent variable approach has been demonstrated to increase the performance in a multilingual unsupervised part-of-speech tagging model based on HMMs BIBREF4 . We investigate the application of this approach to unsupervised SRL, presenting the performance improvements obtained in different settings involving labeled and unlabeled data, and analyzing the annotation effort required to obtain similar gains using labeled data.
We begin by briefly describing the unsupervised SRL pipeline and the monolingual semantic role induction model we use, and then describe our multilingual model.
The model consists of individual Bayesian models for each language (Garg and Henderson, 2012), and crosslingual latent variables to incorporate soft role agreement between aligned constituents.
|
What does the joint Bayesian model for unsupervised semantic role induction in multiple languages consist of?
|
Individual Bayesian models for each language (Garg and Henderson, 2012), and crosslingual latent variables.
|
null | false
| 112
|
Twitter, a micro-blogging and social networking site has emerged as a platform where people express themselves and react to events in real-time. It is estimated that nearly 500 million tweets are sent per day . Twitter data is particularly interesting because of its peculiar nature where people convey messages in short sentences using hashtags, emoticons, emojis etc. In addition, each tweet has meta data like location and language used by the sender. It's challenging to analyze this data because the tweets might not be grammatically correct and the users tend to use informal and slang words all the time. Hence, this poses an interesting problem for NLP researchers. Any advances in using this abundant and diverse data can help understand and analyze information about a person, an event, a product, an organization or a country as a whole. Many notable use cases of the twitter can be found here.
Along the similar lines, The Task 1 of WASSA-2017 BIBREF0 poses a problem of finding emotion intensity of four emotions namely anger, fear, joy, sadness from tweets. In this paper, we describe our approach and experiments to solve this problem. The rest of the paper is laid out as follows: Section 2 describes the system architecture, Section 3 reports results and inference from different experiments, while Section 4 points to ways that the problem can be further explored.
It’s challenging to analyze this data because the tweets might not be grammatically correct and the users tend to use informal and slang words all the time. Hence, this poses an interesting problem for NLP researchers. Any advances in using this abundant and diverse data can help understand and analyze information about a person, an event, a product, an organization or a country as a whole.
|
Why is it challenging to analyze the metadata of tweets?
|
It’s challenging to analyze this data because the tweets might not be grammatically correct and the users tend to use informal and slang words all the time.
|
null | false
| null |
What is the motto of the United States Marine Corps?
|
Adopted in 1883, Semper Fidelis: Latin for "Always Faithful" is the motto of the Marine Corps. Semper Fidelis symbolizes the lifelong commitment held by every Marine for the Corps and America, a promise reciprocated by the Corps to all Marines.
|
|
null | false
| null |
which side of the road you can drive in the UK?
|
In the UK, you are driving on the left side of the road, however, the driver is sitting on the right side of the car.
|
|
null | false
| 303
|
Online social networks have become one of the most disruptive communication platforms, as everyday billions of individuals use them to interact with each other. Their penetration in our everyday lives seems ever-growing and has in turn generated a massive volume of publicly available data open to analysis. The digital footprints left across these multiple media platforms provide us with a unique source to study and understand how the linguistic phenotype of a given user is related to social attributes such as socioeconomic status (SES).
The quantification and inference of SES of individuals is a long lasting question in the social sciences. It is a rather difficult problem as it may depend on a combination of individual characteristics and environmental variables BIBREF0 . Some of these features can be easier to assess like income, gender, or age whereas others, relying to some degree on self-definition and sometimes entangled with privacy issues, are harder to assign like ethnicity, occupation, education level or home location. Furthermore, individual SES correlates with other individual or network attributes, as users tend to build social links with others of similar SES, a phenomenon known as status homophily BIBREF1 , arguably driving the observed stratification of society BIBREF2 . At the same time, shared social environment, similar education level, and social influence have been shown to jointly lead socioeconomic groups to exhibit stereotypical behavioral patterns, such as shared political opinion BIBREF3 or similar linguistic patterns BIBREF4 . Although these features are entangled and causal relation between them is far from understood, they appear as correlations in the data.
Datasets recording multiple characteristics of human behaviour are more and more available due to recent developments in data collection technologies and increasingly popular online platforms and personal digital devices. The automatic tracking of online activities, commonly associated with profile data and meta-information; the precise recording of daily activities, interaction dynamics and mobility patterns collected through mobile personal devices; together with the detailed and expert annotated census data all provide new grounds for the inference of individual features or behavioral patterns BIBREF5 . The exploitation of these data sources has already been proven to be fruitful as cutting edge recommendation systems, advanced methods for health record analysis, or successful prediction tools for social behaviour heavily rely on them BIBREF6 . Nevertheless, despite the available data, some inference tasks, like individual SES prediction, remain an open challenge.
The precise inference of SES would contribute to overcome several scientific challenges and could potentially have several commercial applications BIBREF7 . Further, robust SES inference would provide unique opportunities to gain deeper insights on socioeconomic inequalities BIBREF8 , social stratification BIBREF2 , and on the driving mechanisms of network evolution, such as status homophily or social segregation.
In this work, we take a horizontal approach to this problem and explore various ways to infer the SES of a large sample of social media users. We propose different data collection and combination strategies using open, crawlable, or expert annotated socioeconomic data for the prediction task. Specifically, we use an extensive Twitter dataset of 1.3M users located in France, all associated with their tweets and profile information; 32,053 of them having inferred home locations. Individual SES is estimated by relying on three separate datasets, namely socioeconomic census data; crawled profession information and expert annotated Google Street View images of users' home locations. Each of these datasets is then used as ground-truth to infer the SES of Twitter users from profile and semantic features similar to BIBREF9 . We aim to explore and assess how the SES of social media users can be obtained and how much the inference problem depends on annotation and the user's individual and linguistic attributes.
We provide in Section SECREF2 an overview of the related literature to contextualize the novelty of our work. In Section SECREF3 we provide a detailed description of the data collection and combination methods. In Section SECREF4 we introduce the features extracted to solve the SES inference problem, with results summarized in Section SECREF5 . Finally, in Section SECREF6 and SECREF7 we conclude our paper with a brief discussion of the limitations and perspectives of our methods.
Individual SES is estimated by relying on three separate datasets, namely socioeconomic census data; crawled profession information and expert annotated Google Street View images of users' home locations.
|
How to estimate the Individual SES?
|
Individual SES is estimated by relying on three separate datasets, namely socioeconomic census data; crawled profession information and expert annotated Google Street View images of users' home locations.
|
null | false
| null |
Tamil Nadu (/ˌtæmɪl ˈnɑːduː/; Tamil: [ˈtamiɻ ˈnaːɽɯ] (listen), abbr. TN) is the southern-most state of India. The tenth largest Indian state by area and the sixth largest by population, Tamil Nadu is the home of the Tamil people, whose Tamil language—one of the longest surviving classical languages in the world—is widely spoken in the state and serves as its official language. The capital and largest city is Chennai.
|
What is the capital of Tamilnadu?
|
Chennai
|
null | false
| 36
|
We evaluate our model on two publicly available datasets. The statistics for both are shown in Table TABREF3 . The details of these datasets are as follows:
OntoNotes: OntoNotes 5.0 BIBREF16 includes texts from five different text genres: broadcast conversation (200k), broadcast news (200k), magazine (120k), newswire (625k), and web data (300k). This dataset is annotated with 18 categories.
Wiki(gold): The training data consists of Wikipedia sentences and was automatically generated using a distant supervision method, mapping hyperlinks in Wikipedia articles to Freebase, which we do not use in this study. The test data, mainly consisting of sentences from news reports, was manually annotated as described in BIBREF8 . The class hierarchy is shown in Figure FIGREF2 . This dataset is annotated with 7 main categories (bold text in Figure FIGREF2 ), which maps directly to OntoNotes. The miscellaneous category in Figure FIGREF2 does not have direct mappings, so future work may include redefining these categories so the mappings are more meaningful.
We evaluate our model on two publicly available datasets.
|
How do the authors evaluate the proposed scheme?
|
They evaluate their model on two publicly available datasets.
|
null | false
| 362
|
The proliferation of the online social media has lately resulted in the democratization of online content sharing. Among other media, Twitter is very popular for research and application purposes due to its scale, representativeness and ease of public access to its content. However, tweets, that are short messages of up to 140 characters, pose several challenges to traditional Natural Language Processing (NLP) systems due to the creative use of characters and punctuation symbols, abbreviations ans slung language.
Named Entity Recognition (NER) is a fundamental step for most of the information extraction pipelines. Importantly, the terse and difficult text style of tweets presents serious challenges to NER systems, which are usually trained using more formal text sources such as newswire articles or Wikipedia entries that follow particular morpho-syntactic rules. As a result, off-the-self tools trained on such data perform poorly BIBREF0 . The problem becomes more intense as the number of entities to be identified increases, moving from the traditional setting of very few entities (persons, organization, time, location) to problems with more. Furthermore, most of the resources (e.g., software tools) and benchmarks for NER are for text written in English. As the multilingual content online increases, and English may not be anymore the lingua franca of the Web. Therefore, having resources and benchmarks in other languages is crucial for enabling information access worldwide.
In this paper, we propose a new benchmark for the problem of NER for tweets written in French. The tweets were collected using the publicly available Twitter API and annotated with 13 types of entities. The annotators were native speakers of French and had previous experience in the task of NER. Overall, the generated datasets consists of INLINEFORM0 tweets, split in training and test parts.
The paper is organized in two parts. In the first, we discuss the data preparation steps (collection, annotation) and we describe the proposed dataset. The dataset was first released in the framework of the CAp 2017 challenge, where 8 systems participated. Following, the second part of the paper presents an overview of baseline systems and the approaches employed by the systems that participated. We conclude with a discussion of the performance of Twitter NER systems and remarks for future work.
The tweets were collected using the publicly available Twitter API and annotated with 13 types of entities.
|
How to collect datasets?
|
The dataset was collected using the publicly available Twitter API.
|
null | false
| null |
What fruit is wine produced from?
|
Wine is made from grapes
|
|
null | false
| null |
What are some weird ways I can cut my grass?
|
If you aren't using a lawnmower to cut your grass, here are some weird ideas to keep your grass trimmed:
- Buy a goat and leave it in your yard
- Spend all day cutting each blade with some craft scissors
- Flamethrower
- Dress up like the grim reaper and use a scythe
|
|
null | false
| null |
Why does sand get firmer when it is wet?
|
Sand is made up of tiny grain particles that are largely oval or flat. When they get wet, the particles can more easily move around and align, which packs them more tightly together.
|
|
null | false
| 61
|
The ability of semantic reasoning is essential for advanced natural language understanding (NLU) systems. Many NLU tasks that take sentence pairs as input, such as natural language inference (NLI) and machine reading comprehension (MRC), heavily rely on the ability of sophisticated semantic reasoning. For instance, the NLI task aims to determine whether the hypothesis sentence (e.g., a woman is sleeping) can be inferred from the premise sentence (e.g., a woman is talking on the phone). This requires the model to read and understand sentence pairs to make the specific semantic inference.
Bidirectional Encoder Representations from Transformer (BERT) BIBREF1 has shown strong ability in semantic reasoning. It was recently proposed and obtained impressive results on many tasks, ranging from text classification, natural language inference, and machine reading comprehension. BERT achieves this by employing two objectives in the pre-training, i.e., the masked language modeling (Masked LM) and the next sentence prediction (NSP). Intuitively, the Masked LM task concerns word-level knowledge, and the NSP task captures the global document-level information. The goal of NSP is to identify whether an input sentence is next to another input sentence. From the ablation study BIBREF1, the NSP task is quite useful for the downstream NLI and MRC tasks (e.g., +3.5% absolute gain on the Question NLI (QNLI) BIBREF2 task).
Despite its usefulness, we suggest that BERT has not made full use of the document-level knowledge. The sentences in the negative samples used in NSP are randomly drawn from other documents. Therefore, to discriminate against these sentences, BERT is prone to aggregating the shallow semantic, e.g., topic, neglecting context clues useful for detailed reasoning. In other words, the canonical NSP task would encourage the model to recognize the correlation between sentences, rather than obtaining the ability of semantic entailment. This setting weakens the BERT model from learning specific semantic for inference. Another issue that renders NSP less effective is that BERT is order-sensitive. Performance degradation was observed on typical NLI tasks when the order of two input sentences are reversed during the BERT fine-tuning phase. It is reasonable as the NSP task can be roughly analogy to the NLI task when the input comes as (premise, hypothesis), considering the causal order among sentences. However, this identity between NSP and NLI is compromised when the sentences are swapped.
Based on these considerations, we propose a simple yet effective method, i.e., introducing a IsPrev category to the classification task, which is a symmetric label of IsNext of NSP. The input of samples with IsPrev is the reverse of those with IsNext label. The advantages of using this previous sentence prediction (PSP) are three folds. (1) Learning the contrast between NSP and PSP forces the model to extract more detailed semantic, thereby the model is more capable of discriminating the correlation and entailment. (2) NSP and PSP are symmetric. This symmetric regularization alleviates the influence of the order of the input pair. (3) Empirical results indicate that our method is beneficial for all the semantic reasoning tasks that take sentence pair as input.
In addition, to further incorporating the document-level knowledge, NSP and PSP are extended with non-successive sentences, where the label smoothing technique is adopted. The proposed method yields a considerable improvement in our experiments. We evaluate the ability of semantic reasoning on standard NLI and MRC benchmarks, including the challenging HANS dataset BIBREF0. Analytical work on the HANS dataset provides a more comprehensible perspective towards the proposed method. Furthermore, the results on the Chinese benchmarks are provided to demonstrate its generality.
In summary, this work makes the following contributions:
The supervision signal from the original NSP task is weak for semantic inference. Therefore, a novel method is proposed to remedy the asymmetric issue and enhance the reasoning ability.
Both empirical and analytical evaluations are provided on the NLI and MRC datasets, which verifies the effectiveness of using more document-level knowledge.
We evaluate the ability of semantic reasoning on standard NLI and MRC benchmarks, including the challenging HANS dataset.
|
What is the dataset of this paper?
|
It is the standard natural language understanding (NLI) and machine reading comprehension (MRC) benchmarks, including the challenging HANS dataset.
|
null | false
| null |
Dunk Island, known as Coonanglebah in the Warrgamay and Dyirbal languages, is an island within the locality of Dunk in the Cassowary Coast Region, Queensland, Australia. It lies 4 kilometres (2.5 mi) off the Australian east coast, opposite the town of Mission Beach. The island forms part of the Family Islands National Park and is in the larger Great Barrier Reef World Heritage Area.
|
What is the indigenous name for Dunk Island?
|
Dunk Island is known as Coonanglebah in the Warrgamay and Dyirbal languages.
|
null | false
| null |
Categorize each of the following by the amount of noise they would make, either "a lot of noise", "some noise" or "no noise". A baby crying, a glass vase being dropped, a door being closed, a rabbit hopping across carpeted floors, a running microwave, reading a book
|
A baby crying: a lot of noise, a glass vase being dropped: a lot of noise, a door being closed: some noise, a rabbit hopping across carpeted floors: no noise, a running microwave: some noise, reading a book: no noise, slamming a door: a lot of noise
|
|
null | false
| null |
Who regulates banks in Australia?
|
The Australian Prudential Regulatory Authority (APRA) is the prudential regulator of the Australian financial services industry.
|
|
null | false
| null |
The plot is based on the story of the proof of Fermat's Last Theorem by Andrew Wiles, whose name is changed to "Daniel Keane" in the musical. After seven years of isolation in his attic, Keane believes he has found a proof of the theorem. The musical starts with a press conference, where Keane explains his proof to reporters and promises to return to normal life with his wife Anna and his family. After promising to Anna that he is now "done with Fermat", Keane is surprised in his study by none other than Fermat himself. Keane asks Fermat for the secret of his proof but is refused. Instead, Fermat introduces him to the "Aftermath", a "heavenly purgatory" where he meets the famous mathematicians Euclid, Pythagoras, Newton, and Gauss. They inform him that his proof contains a "big fat hole". In a second press conference, Keane is questioned by reporters about a flaw in the proof. Anna wishes for a corrected proof for her birthday. Fermat mocks Keane, and the other mathematicians inform him that "mathematics is a young man's game". Keane returns to his attic to try to fix his proof, while his "math widow" wife is frustrated. Fermat continues to taunt Keane, but he is invisible and inaudible to Anna, and the three dance a "bizarre tango à trois" while Anna is confused by Keane talking to Fermat. The other mathematicians from the Aftermath, after noticing that they can't keep up with the mathematics of the past century, decide to grant admission to Keane even if he is unable to prove the theorem. As Keane finally gives up and declares his attempts a failure, Anna suggests that "within your failure lie the seeds of your success", repeating a line earlier spoken by the mathematicians. This quickly leads to Keane realising how to close the gap in the argument, and the musical ends with another press conference, and Fermat congratulates Keane for his proof.
|
How many press conferences does Daniel Keane give in the musical "Fermat's Last Tango"?
|
In the musical "Fermat's Last Tango" the protagonist give three press conferences - in the first press conference, Daniel announces he has found proof of Fermat's Last Theorem; in the second press conference reporters question him about a flaw in the proof and in the third press conference, where he announces the conclusive proof of the theorem.
|
null | false
| 141
|
Automatic summarization, machine translation, question answering, and semantic parsing operations are useful for processing, analyzing, and extracting meaningful information from text. However, when applied to long texts, these tasks usually require some minimal syntactic structure to be identified, such as sentences BIBREF0 , BIBREF1 , BIBREF2 , which always end with a period (“.”) in English BIBREF3 .
However, written Thai does not use an explicit end-of-sentence marker to identify sentence boundaries BIBREF4 . Prior works have adapted traditional machine learning models to predict the beginning position of a sentence. The authors of BIBREF5 , BIBREF6 , BIBREF7 proposed traditional models to determine whether a considered space is a sentence boundary based on the words and their part of speech (POS) near the space. Meanwhile, Zhou N. et al. BIBREF8 considered Thai sentence segmentation as a sequence tagging problem and proposed a CRF-based model with n-gram embedding to predict which word is the sentence boundary. This method achieves the state-of-the-art result for Thai sentence segmentation and achieves greater accuracy than other models by approximately 10% on an Orchid dataset BIBREF9 .
Several deep learning approaches have been applied in various tasks of natural language processing (NLP), including the long short-term memory BIBREF10 , self-attention BIBREF11 , and other models. Huang Z. et al. BIBREF12 proposed a deep learning sequence tagging model called Bi-LSTM-CRF, which integrates a conditional random field (CRF) module to gain the benefit of both deep learning and traditional machine learning approaches. In their experiments, the Bi-LSTM-CRF model achieved an improved level of accuracy in many NLP sequence tagging tasks, such as named entity recognition, POS tagging and chunking.
The CRF module achieved the best result on the Thai sentence segmentation task BIBREF8 ; therefore, we adopt the Bi-LSTM-CRF model as our baseline. This paper makes the following three contributions to improve Bi-LSTM-CRF for sentence segmentation.
First, we propose adding n-gram embedding to Bi-LSTM-CRF due to its success in BIBREF8 and BIBREF12 . By including n-gram embedding, the model can capitalize on both approaches. First, the model gains the ability to extract past and future input features and sentence level tag information from Bi-LSTM-CRF; moreover, with the n-gram addition, it can also extract a local representation from n-gram embedding, which helps in capturing word groups that exist near sentence boundary. Although Jacovi A. et al. BIBREF13 reported that a convolutional neural network (CNN) can be used as an n-gram detector to capture local features, we chose n-gram embedding over a CNN due to its better accuracy, as will be shown in Section SECREF8 .
Second, we propose adding incorporative distant representation into the model via a self-attention mechanism, which can focus on the keywords of dependent clauses that are far from the considered word. Self-attention has been used in many recent state-of-the-art models, most notably the transformer BIBREF11 and BERT BIBREF14 . BERT has outperformed Bi-LSTM on numerous tasks, including question answering and language inference. Therefore, we choose to use self-attention modules to extract distant representations along with local representations to improve model accuracy.
Third, we also apply semi-supervised learning BIBREF15 , allowing us to employ unlimited amounts of unlabeled data, which is particularly important for low-resource languages such as Thai, for which annotation is costly and time-consuming. Many semi-supervised learning approaches have been proposed in the computer vision BIBREF16 , BIBREF17 and natural language processing BIBREF18 , BIBREF19 , BIBREF20 fields. Our choice for semi-supervised learning to enhance model representation is Cross-View Training (CVT) BIBREF20 . Clark K. et al. BIBREF20 claims that CVT can improve the representation layers of the model, which is our goal. However, CVT was not designed to be integrated with self-attention and CRF modules; consequently, we provide a modified version of CVT in this work.
Based on the above three contributions, we pursue two main experiments. The first experiment was conducted on two Thai datasets, Orchid and UGWC BIBREF21 , to evaluate our Thai sentence segmentation model. In this case, our model achieves F1 scores of 92.5% and 88.9% on Orchid and UGWC, respectively, and it outperforms all the baseline models. The second experiment was executed on the IWSLT dataset BIBREF22 and involves an English-language punctuation restoration task. This experiment demonstrates that our model is generalizable to different languages. Our model, which does not require pretrained word vectors, improved the overall F1 score by 0.9% compared to the baselines, including a model that uses pretrained word vectors.
There are five sections in the remainder of this paper. Section SECREF2 reviews the related works on Thai sentence segmentation, English punctuation restoration and introduces the original CVT. Section SECREF3 describes the proposed model architecture and the integration of cross-view training. The datasets, implementation process and evaluation metrics are explained in Section SECREF4 . The results of the experiments are discussed in Section SECREF5 . Finally, Section SECREF6 concludes the paper.
First, we propose adding n-gram embedding to Bi-LSTM-CRF due to its success in and .
|
What embedding does the paper propose adding to Bi-LSTM-CRF?
|
n-gram embedding.
|
null | false
| null |
Crystal Palace Football Club, commonly referred to as Palace, is a professional football club based in Selhurst in the Borough of Croydon, South London, England, which competes in the Premier League, the highest level of English football. The club's kit colours were claret and blue until 1973, when they changed to the red and blue vertical stripes worn today.
|
What professional football team play in red and blue and are based in south london?
|
Crystal Palace FC
|
null | false
| null |
Choose the word which is different from the rest: kiwi, eagle, emu, ostrich
|
all except eagle are flightless birds
|
|
null | false
| null |
define rainbow clours
|
VIBGYOR Violet–Indigo–Blue–Green–Yellow–Orange–Red
|
|
null | false
| null |
Who was the first president of the United States?
|
George Washington
|
|
null | false
| null |
How are rainbows formed
|
Rainbows are formed by a combination of various factors: the physics of light, it's components and interaction that combined with the dynamics of the rain droplets and the angle of sun hitting the water droplets.
Contrary to popular belief, the rain droplets are spherical and when light passes through this denser sphere, the light is scattered due to refraction into it's component wavelengths corresponding to the VIBGYOR pattern.
|
|
null | false
| null |
Tell me which of these are stars or planets: Mercury, Sirius, Betelgeuse, Jupiter, Aldebaran, Saturn, Mars, Rigel
|
Mercury - Planet
Sirius - Star
Betelgeuse - Star
Jupiter - Planet
Aldebaran - Star
Saturn - Planet
Mars - Planet
Rigel - Star
|
|
null | false
| null |
What is s3 equivalent in GCP?
|
If you're familiar with AWS, Google Storage is GCP's version of AWS Simple Storage Service (S3) and an S3 bucket would be equivalent to a Google Storage bucket across the two clouds.
|
|
null | false
| null |
if you have 2 ducks in front of a duck and 2 ducks behind a duck in a row, minimum how many ducks you have?
|
The answer would be 3 as you can have this setup with a minimum of 3 ducks.
|
|
null | false
| null |
Yu-Gi-Oh! tells the tale of Yugi Mutou, a timid young boy who loves all sorts of games, but is often bullied around. One day, he solves an ancient puzzle known as the Millennium Puzzle (千年パズル, Sennen Pazuru), causing his body to play host to a mysterious spirit with the personality of a gambler. From that moment onwards, whenever Yugi or one of his friends is threatened by those with darkness in their hearts, this other Yugi shows himself and challenges them to dangerous Shadow Games (闇のゲーム, Yami no Gēmu, lit. "Games of Darkness") which reveal the true nature of someone's heart, the losers of these contests often being subjected to a dark punishment called a Penalty Game (罰ゲーム, Batsu Gēmu). Whether it be cards, dice, or role-playing board games, he will take on challenges from anyone, anywhere.
As the series progresses, Yugi and his friends learn that this person inside of his puzzle is actually the spirit of a nameless Pharaoh from Ancient Egyptian times, who had lost his memories. As Yugi and his companions attempt to help the Pharaoh regain his memories, they find themselves going through many trials as they wager their lives facing off against gamers that wield the mysterious Millennium Items (千年アイテム, Sennen Aitemu) and the dark power of the Shadow Games.
|
What is the plot of Yu-Gi-Oh?
|
Yu-Gi-Oh! is about a boy named Yugi Mutou who plays a card game called Yu-Gi-Oh! He solves the Millennium Puzzle so his body now hosts the spirit of a gambler as well, and takes on various challenges against enemies. Most often, the challenge is the Yu-Gi-Oh! card game but sometimes it will be alternate games such as dice or board games. As the show moves forward, Yugi and his friends discover the spirit is actually an Ancient Egyptian Pharaoh with amnesia so they try to help him regain his memories by facing additional challenges from gamers with Millennium items.
|
null | false
| null |
Northam ran for Lieutenant Governor of Virginia in the 2013 election. Northam competed against U.S. Chief Technology Officer Aneesh Chopra for the Democratic nomination. On June 11, 2013, Northam won the Democratic primary over Chopra with 54% of the vote to Chopra's 46%.
On November 5, 2013, Northam was elected as Virginia's 40th Lieutenant Governor over Republican E. W. Jackson, receiving 55% of the vote to Jackson's 45%. Northam was the first Democrat since Tim Kaine in 2001 to be elected Lieutenant Governor of Virginia.
|
By what percentage of vote margin did Ralph Northam win over Aneesh Chopra in the 2013 election for Governor of Virginia?
|
8 percent
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.