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
| null |
On a visit to a movie theater, classify the likelihood of finding each of the following as either 'highly likely', 'likely', 'unlikely' or 'highly unlikely'. popcorn, M&Ms, sushi, salad, pasta.
|
It is highly likely that popcorn would be found at a movie theater.
It is highly likely that M&Ms would be found at a movie theater.
It is highly unlikely that sushi would be found at a movie theater.
It is highly unlikely that salad would be found at a movie theater.
It is highly unlikely that pasta would be found at a movie theater.
|
|
null | false
| 155
|
Cyberbullying has been defined by the National Crime Prevention Council as the use of the Internet, cell phones or other devices to send or post text or images intended to hurt or embarrass another person. Various studies have estimated that between to 10% to 40% of internet users are victims of cyberbullying BIBREF0 . Effects of cyberbullying can range from temporary anxiety to suicide BIBREF1 . Many high profile incidents have emphasized the prevalence of cyberbullying on social media. Most recently in October 2017, a Swedish model Arvida Byström was cyberbullied to the extent of receiving rape threats after she appeared in an advertisement with hairy legs.
Detection of cyberbullying in social media is a challenging task. Definition of what constitutes cyberbullying is quite subjective. For example, frequent use of swear words might be considered as bullying by the general population. However, for teen oriented social media platforms such as Formspring, this does not necessarily mean bullying (Table TABREF9 ). Across multiple SMPs, cyberbullies attack victims on different topics such as race, religion, and gender. Depending on the topic of cyberbullying, vocabulary and perceived meaning of words vary significantly across SMPs. For example, in our experiments we found that for word `fat', the most similar words as per Twitter dataset are `female' and `woman' (Table TABREF23 ). However, other two datasets do not show such particular bias against women. This platform specific semantic similarity between words is a key aspect of cyberbullying detection across SMPs. Style of communication varies significantly across SMPs. For example, Twitter posts are short and lack anonymity. Whereas posts on Q&A oriented SMPs are long and have option of anonymity (Table TABREF7 ). Fast evolving words and hashtags in social media make it difficult to detect cyberbullying using swear word list based simple filtering approaches. The option of anonymity in certain social networks also makes it harder to identify cyberbullying as profile and history of the bully might not be available.
Past works on cyberbullying detection have at least one of the following three bottlenecks. First (Bottleneck B1), they target only one particular social media platform. How these methods perform across other SMPs is unknown. Second (Bottleneck B2), they address only one topic of cyberbullying such as racism, and sexism. Depending on the topic, vocabulary and nature of cyberbullying changes. These models are not flexible in accommodating changes in the definition of cyberbullying. Third (Bottleneck B3), they rely on carefully handcrafted features such as swear word list and POS tagging. However, these handcrafted features are not robust against variations in writing style. In contrast to existing bottlenecks, this work targets three different types of social networks (Formspring: a Q&A forum, Twitter: microblogging, and Wikipedia: collaborative knowledge repository) for three topics of cyberbullying (personal attack, racism, and sexism) without doing any explicit feature engineering by developing deep learning based models along with transfer learning.
We experimented with diverse traditional machine learning models (logistic regression, support vector machine, random forest, naive Bayes) and deep neural network models (CNN, LSTM, BLSTM, BLSTM with Attention) using variety of representation methods for words (bag of character n-gram, bag of word unigram, GloVe embeddings, SSWE embeddings). Summary of our findings and research contributions is as follows.
First (Bottleneck B1), they target only one particular social media platform. How these methods perform across other SMPs is unknown. Second (Bottleneck B2), they address only one topic of cyberbullying such as racism, and sexism. Depending on the topic, vocabulary and nature of cyberbullying changes. These models are not flexible in accommodating changes in the definition of cyberbullying. Third (Bottleneck B3), they rely on carefully handcrafted features such as swear word list and POS tagging.
|
What are the bottlenecks in past efforts to detect cyberbullying?
|
First (Bottleneck B1), they target only one particular social media platform. Second (Bottleneck B2), they address only one topic of cyberbullying such as racism, and sexism.Third (Bottleneck B3), they rely on carefully handcrafted features such as swear word list and POS tagging.
|
null | false
| null |
Believe it or not, Jimi Hendrix only had one Top 40 hit. Which song was it?
|
"All Along the Watchtower"
|
|
null | false
| null |
What is a phenotype?
|
In genetics and genomics, the phenotype is the set of observable characteristics or traits of an organism. The term covers the organism's morphology (physical form and structure), its developmental processes, its biochemical and physiological properties, its behavior, and the products of behavior. An organism's phenotype results from two basic factors: the expression of an organism's genetic code (its genotype) and the influence of environmental factors. Both factors may interact, further affecting the phenotype. When two or more clearly different phenotypes exist in the same population of a species, the species is called polymorphic. A well-documented example of polymorphism is Labrador Retriever coloring; while the coat color depends on many genes, it is clearly seen in the environment as yellow, black, and brown. Richard Dawkins in 1978 and then again in his 1982 book The Extended Phenotype suggested that one can regard bird nests and other built structures such as caddisfly larva cases and beaver dams as "extended phenotypes".
Wilhelm Johannsen proposed the genotype–phenotype distinction in 1911 to make clear the difference between an organism's hereditary material and what that hereditary material produces. The distinction resembles that proposed by August Weismann (1834–1914), who distinguished between germ plasm (heredity) and somatic cells (the body). More recently, in The Selfish Gene (1976), Dawkins distinguished these concepts as replicators and vehicles.
The genotype–phenotype distinction should not be confused with Francis Crick's central dogma of molecular biology, a statement about the directionality of molecular sequential information flowing from DNA to protein, and not the reverse.
|
|
1911.03310
| false
| null |
Table TABREF15 shows that word-alignment based on mBERT representations surpasses the outputs of the standard FastAlign tool even if it was provided large parallel corpus. This suggests that word-level semantics are well captured by mBERT contextual embeddings. For this task, learning an explicit projection had a negligible effect on the performance.
Table TABREF15 shows that word-alignment based on mBERT representations surpasses the outputs of the standard FastAlign tool even if it was provided large parallel corpus. This suggests that word-level semantics are well captured by mBERT contextual embeddings. For this task, learning an explicit projection had a negligible effect on the performance.
|
How they demonstrate that language-neutral component is sufficiently general in terms of modeling semantics to allow high-accuracy word-alignment?
|
The answers are shown as follows:
* explicit projection had a negligible effect on the performance
|
null | false
| null |
Authoritarianism is a political system characterized by the rejection of political plurality, the use of strong central power to preserve the political status quo, and reductions in the rule of law, separation of powers, and democratic voting. Political scientists have created many typologies describing variations of authoritarian forms of government. Authoritarian regimes may be either autocratic or oligarchic and may be based upon the rule of a party or the military. States that have a blurred boundary between democracy and authoritarianism have some times been characterized as "hybrid democracies", "hybrid regimes" or "competitive authoritarian" states.
The political scientist Juan Linz, in an influential 1964 work, An Authoritarian Regime: Spain, defined authoritarianism as possessing four qualities:
Limited political pluralism, is realized with constraints on the legislature, political parties and interest groups.
Political legitimacy is based upon appeals to emotion and identification of the regime as a necessary evil to combat "easily recognizable societal problems, such as underdevelopment or insurgency."
Minimal political mobilization, and suppression of anti-regime activities.
Ill-defined executive powers, often vague and shifting extend the power of the executive.
Minimally defined, an authoritarian government lacks free and competitive direct elections to legislatures, free and competitive direct or indirect elections for executives, or both. Broadly defined, authoritarian states include countries that lack civil liberties such as freedom of religion, or countries in which the government and the opposition do not alternate in power at least once following free elections. Authoritarian states might contain nominally democratic institutions such as political parties, legislatures and elections which are managed to entrench authoritarian rule and can feature fraudulent, non-competitive elections. In contexts of democratic backsliding, scholars tend to identify authoritarian political leaders based on certain tactics, such as: politicizing independent institutions, spreading disinformation, aggrandizing executive power, quashing dissent, targeting vulnerable communities, stoking violence, and corrupting elections. Since 1946, the share of authoritarian states in the international political system increased until the mid-1970s but declined from then until the year 2000.
|
What are some defining qualities of authoritarianism?
|
1. Limited political pluralism or outright rejection of political pluralism.
2. Lack of civil liberties
3. Lack of free, fair and competitive elections
4. So-called democratic institutions corrupted by authoritarian rule
5. Use of election tactics such as spreading misinformation, politicizing independent institutions, quashing dissent, engaging in culture wars, targeting vulnerable communities, stoking violence, and aggrandizing executive power.
6. Use of a strong central power to preserve the political status quo at the cost of basic democratic ideals such as the rule of law, separation of powers, and voting rights.
|
null | false
| null |
Pedro Almodóvar Caballero (Spanish pronunciation: [ˈpeðɾo almoˈðoβaɾ kaβaˈʝeɾo]; born 25 September 1949) is a Spanish filmmaker. His films are marked by melodrama, irreverent humour, bold colour, glossy décor, quotations from popular culture, and complex narratives. Desire, passion, family, and identity are among Almodóvar's most prevalent subjects in his films. Acclaimed as one of the most internationally successful Spanish filmmakers, Almodóvar and his films have gained worldwide interest and developed a cult following.
Almodóvar's career came to during La Movida Madrileña, a cultural renaissance that followed after the end of Francoist Spain. His early films characterised the sense of sexual and political freedom of the period. In 1986, he established his own film production company, El Deseo, with his younger brother Agustín Almodóvar, who has been responsible for producing all of his films since Law of Desire (1987). His breakthrough film was Women on the Verge of a Nervous Breakdown (1988), which was nominated for the Academy Award for Best Foreign Language Film.
He achieved further success often collaborating with actors Antonio Banderas and Penélope Cruz. He directed Tie Me Up! Tie Me Down! (1989), High Heels (1991), and Live Flesh (1997). His next two films, All About My Mother (1999) and Talk to Her (2002), earned him an Academy Award each for Best International Feature Film and Best Original Screenplay, respectively. His later films include Volver (2006), Broken Embraces (2009), The Skin I Live In (2011), Julieta (2016), Pain and Glory (2019), and Parallel Mothers (2021).
Almodóvar has received numerous accolades including two Academy Awards, five BAFTA Awards, two Emmy Awards, two Golden Globe Awards, nine Goya Awards. He has also received the French Legion of Honour in 1997, the Gold Medal of Merit in the Fine Arts in 1999, and the European Film Academy Achievement in World Cinema Award in 2013 and was awarded the Golden Lion in 2019. He's also received an honorary doctoral degrees from Harvard University in 2009 and from University of Oxford in 2016.
|
From the passage, extract the list of movies that Pedro Almodovar produced. Include the year and the name. Place them in a bullet list.
|
1982 - Labyrinth of Passion
1988 - Women on the Verge of a Nervous Breakdown
2004 - Bad Education
2021 - Parallel Mothers
|
null | false
| null |
The Teiko Middle School basketball team dominated basketball teams within Japan, winning the middle school Nationals for three consecutive years. The star players of the team became known as the "Generation of Miracles". After graduating from middle school, the five star players went to different high schools with top basketball teams. However, there was a rumor that there was another player in the "Generation of Miracles": a phantom sixth man. This mysterious player is now a freshman at Seirin High, a new school with a powerful, if little-known, team. Now, Tetsuya Kuroko – the sixth member of the "Generation of Miracles", and Taiga Kagami – a naturally talented player who spent most of his youth in the US, aim to bring Seirin to the top of Japan by taking on Kuroko's former teammates one by one. The series chronicles Seirin's rise to become Japan's number one high school team. The Generation of Miracles include Ryota Kise, Shintaro Midorima, Daiki Aomine, Atsushi Murasakibara, and Seijuro Akashi.
|
What is the plot of Kuroko's Basketball?
|
Kuroko's Basketball is an anime about basketball. The story starts with Teiko Middle School where 6 talented players are known as the Generation of Miracles and they went undefeated for 3 years. However, all of the players split off for high school. The protagonist, Tetsuya, is the mysterious sixth man who comes to Seirin high school and meets Taiga Kagami, who is a naturally talented American player. The show focuses on this duo beating the rest of the Generation of Miracles to become the top Japanese high school basketball team.
|
1708.09157
| false
| null |
Our formulation of transfer learning builds on work in multi-task learning BIBREF15 , BIBREF9 . We treat each individual language as a task and train a joint model for all the tasks. We first discuss the current state of the art in morphological tagging: a character-level recurrent neural network. After that, we explore three augmentations to the architecture that allow for the transfer learning scenario. All of our proposals force the embedding of the characters for both the source and the target language to share the same vector space, but involve different mechanisms, by which the model may learn language-specific features.
Cross-lingual morphological tagging may be formulated as a multi-task learning problem. We seek to learn a set of shared character embeddings for taggers in both languages together through optimization of a joint loss function that combines the high-resource tagger and the low-resource one. The first loss function we consider is the following:
We treat each individual language as a task and train a joint model for all the tasks.
We seek to learn a set of shared character embeddings for taggers in both languages together through optimization of a joint loss function that combines the high-resource tagger and the low-resource one.
|
How are character representations from various languages joint?
|
The answers are shown as follows:
* shared character embeddings for taggers in both languages together through optimization of a joint loss function
|
null | false
| null |
Free climbing has been called "rock climbing's most commonly mistaken term", with problems including:
Incorrectly assuming that "free climbing" always means solo climbing, i.e. that you must always be alone and without any partner. Free climbing in traditional climbing and sport climbing uses a supporting belayer.
Incorrectly assuming that "free climbing" always means free soloing, i.e. that you must never use any climbing protection equipment. Free climbing in traditional climbing and sport climbing uses climbing protection (but not to aid progression).
Incorrectly assuming that "free climbing" always means onsighting or flashing, i.e. that you must always climb the route first try. Free climbing in traditional climbing and sport climbing also uses the redpoint as a definition of a first free ascent.
|
If it's called "free" climbing, why do climbers use gear to climb?
|
Free climbing refers to a form of rock climbing in which the climber ascends without the use of gear to aid progression. It is a common misunderstanding that free climbing means climbing with no protective gear. Both sport and traditional climbing make use of protective gear to prevent injury and death, and are both considered free climbing. This is in contrast to free solo climbing, another form of free climbing in which the climber uses no protective gear.
|
null | false
| null |
What is the largest or biggest land animal?
|
The largest or biggest land animal is the African bush elephant. It can reach weights of up to 10000 kgs (or 22000 lbs). However historically there have been dinosaurs that are greater in weight than elephants.
|
|
null | false
| null |
Can you sort the following letters alphabetically: d, h, a, f and y?
|
Yes, here is those letters in sorted in alphabetical order : a, d, f, h, y
|
|
null | false
| 164
|
Topic models (e.g., LDA BIBREF47 ) are usually unsupervised and therefore less biased towards human-defined categories. They are especially suited for insight-driven analysis, because they are constrained in ways that make their output interpretable. Although there is no guarantee that a “topic” will correspond to a recognizable theme or event or discourse, they often do so in ways that other methods do not. Their easy applicability without supervision and ready interpretability make topic models good for exploration. Topic models are less successful for many performance-driven applications. Raw word features are almost always better than topics for search and document classification. LSTMs and other neural network models are better as language models. Continuous word embeddings have more expressive power to represent fine-grained semantic similarities between words.
A topic model provides a different perspective on a collection. It creates a set of probability distributions over the vocabulary of the collection, which, when combined together in different proportions, best match the content of the collection. We can sort the words in each of these distributions in descending order by probability, take some arbitrary number of most-probable words, and get a sense of what (if anything) the topic is “about”. Each of the text segments also has its own distribution over the topics, and we can sort these segments by their probability within a given topic to get a sense of how that topic is used.
One of the most common questions about topic models is how many topics to use, usually with the implicit assumption that there is a “right” number that is inherent in the collection. We prefer to think of this parameter as more like the scale of a map or the magnification of a microscope. The “right” number is determined by the needs of the user, not by the collection. If the analyst is looking for a broad overview, a relatively small number of topics may be best. If the analyst is looking for fine-grained phenomena, a larger number is better.
After fitting the model, it may be necessary to circle back to an earlier phase. Topic models find consistent patterns. When authors repeatedly use a particular theme or discourse, that repetition creates a consistent pattern. But other factors can also create similar patterns, which look as good to the algorithm. We might notice a topic that has highest probability on French stopwords, indicating that we need to do a better job of filtering by language. We might notice a topic of word fragments, such as “ing”, “tion”, “inter”, indicating that we are not handling end-of-line hyphenation correctly. We may need to add to our stoplist or change how we curate multi-word terms.
Topic models (e.g., LDA [Blei et al., 2003]) are usually unsupervised and therefore less biased towards human-defined categories. They are especially suited for insight-driven analysis, because they are constrained in ways that make their output interpretable.
|
Why are unsupervised models suited for insight-driven analysis?
|
Because they are constrained in ways that make their output interpretable.
|
null | false
| null |
Give me a list of the top five most followed people on Instagram, along with their number of followers.
|
1. Cristiano Ronaldo, 567 millions
2. Lionel Messi, 448 millions
3. Selena Gomez, 407 millions
4. Kylie Jenner, 383 millions
5. Dwayne Johnson, 371 millions
|
|
1912.07025
| false
| null |
FLOAT SELECTED: TABLE IV: Class-wise average IoUs and per-pixel accuracies on the test set. Refer to Table I for full names of abbreviated region types listed at top of the table.
FLOAT SELECTED: TABLE I: Counts for various annotated region types in INDISCAPES dataset. The abbreviations used for region types are given below each region type.
FLOAT SELECTED: TABLE IV: Class-wise average IoUs and per-pixel accuracies on the test set. Refer to Table I for full names of abbreviated region types listed at top of the table.
FLOAT SELECTED: TABLE I: Counts for various annotated region types in INDISCAPES dataset. The abbreviations used for region types are given below each region type.
|
What accuracy does CNN model achieve?
|
Combined per-pixel accuracy for character line segments is 74.79
|
null | false
| null |
Why do people enjoy baseball?
|
Baseball requires skill, practice, hard work, hand eye coordination and grit. There are one hundred and fifty plus games per year in the major league, and most players will be on the field for 70% of those games. There are few thrills better than being in the stands watching your home team win a game in the last inning by hitting a game-winning homerun.
|
|
null | false
| null |
Nintendo was founded in 1889 as Nintendo Karuta by craftsman Fusajiro Yamauchi and originally produced handmade hanafuda playing cards. After venturing into various lines of business during the 1960s and acquiring a legal status as a public company, Nintendo distributed its first console, the Color TV-Game, in 1977. It gained international recognition with the release of Donkey Kong in 1981 and the Nintendo Entertainment System and Super Mario Bros. in 1985.
|
Given this reference text about a video game, when was Donkey Kong released?
|
1981
|
null | false
| null |
What is the main tax benefit of a Roth IRA and 401K?
|
The main tax benefit of a Roth IRA and 401K is that realized gains from selling investments within the accounts are not taxed.
|
|
null | false
| null |
What type of bike is suitable for children to learn: Balance Bike, Foldable bike, Mountain Bike, Road Bike
|
The bike suitable for children to start learning bicycle is Balance Bike
|
|
null | false
| null |
Who is the greatest Boston Red Sox player of all time?
|
The Boston Red Sox have had many amazing players over their long history, so it is hard to pick the single best player. Some of the most well-known and beloved players include David Ortiz, Carlton Fisk, Carl Yastrzemski, Ted Williams, Jim Rice, and Pedro Martinez.
|
|
null | false
| null |
Agriculture encompasses crop and livestock production, aquaculture, fisheries and forestry for food and non-food products. Agriculture was the key development in the rise of sedentary human civilization, whereby farming of domesticated species created food surpluses that enabled people to live in cities. While humans started gathering grains at least 105,000 years ago, nascent farmers only began planting them around 11,500 years ago. Sheep, goats, pigs and cattle were domesticated around 10,000 years ago. Plants were independently cultivated in at least 11 regions of the world. In the twentieth century, industrial agriculture based on large-scale monocultures came to dominate agricultural output.
|
Based on the information below, when might people have started baking bread?
|
There is not enough information in the passage to determine exactly when this would have happened, but because bread requires grain, it could not have been made prior to 105,000 years ago when humans first started gathering grain.
|
null | false
| null |
According to the World Register of Marine Species, there are only two genera in this family. Previously there were about 26, but most of these have been reallocated as sub-genera of the remaining two genera, Mycale and Phlyctaenopora:
|
If there are only two genera remaining, what happened to the other 24?
|
Most of the other 24 were reallocated as sub-genera of the remaining two genera.
|
2003.04032
| false
| null |
In this paper, we describe our scheme and process in annotating shallow discourse relations using PDTB-style. In view of the differences between English and Chinese, we made adaptations for the PDTB-3 scheme such as removing AltLexC and adding Progression into our sense hierarchy. To ensure the annotation quality, we formulated detailed annotation criteria and quality assurance strategies. After serious training, we annotated 3212 discourse relations, and we achieved a satisfactory consistency of labelling with a Kappa value of greater than 0.85 for most of the indicators. Finally, we display our annotation results in which the distribution of discourse relations and senses differ from that in other corpora which annotate news report or newspaper texts. Our corpus contains more Contingency, Temporal and Comparison relations instead of being governed by Expansion.
In view of the differences between English and Chinese, we made adaptations for the PDTB-3 scheme such as removing AltLexC and adding Progression into our sense hierarchy.
|
How are resources adapted to properties of Chinese text?
|
The answers are shown as follows:
* removing AltLexC and adding Progression into our sense hierarchy
|
null | false
| null |
The 2022 World Snooker Championship (officially the 2022 Betfred World Snooker Championship) was a professional snooker tournament that took place from 16 April to 2 May 2022 at the Crucible Theatre in Sheffield, England, the 46th consecutive year the World Snooker Championship was held at the venue. The 16th and final ranking event of the 2021–22 snooker season, the tournament was organised by the World Snooker Tour and sponsored by sports betting company Betfred. It was broadcast in the United Kingdom by the BBC, in Europe (including the UK) by Eurosport, and elsewhere in the world by Matchroom Sport and other broadcasters. The total prize fund was £2,395,000, of which the winner received £500,000.
Qualifying rounds for the tournament took place from 4 to 13 April 2022 at the English Institute of Sport, featuring 128 professional and invited amateur players. The main stage of the tournament featured 32 players: the top 16 players from the snooker world rankings and another 16 players from the qualifying rounds. Ashley Hugill, Jackson Page, and Hossein Vafaei were debutants at the Crucible, Vafaei being the first Iranian player to reach the main stage. Mark Selby was the defending champion, having won the 2021 final 18–15 against Shaun Murphy. He lost 10–13 to Yan Bingtao in a second-round match that produced the longest frame ever played at the Crucible, lasting 85 minutes.
Ronnie O'Sullivan equalled Steve Davis's record of 30 Crucible appearances. He reached a record 20th quarter-final and a record 13th semi-final before defeating Judd Trump 18–13 in the final to equal Stephen Hendry's record of seven world titles. It was O'Sullivan's 39th ranking title and 21st Triple Crown title. Aged 46 years and 148 days, he became the oldest world champion in the sport's history, surpassing Ray Reardon, who was aged 45 years and 203 days when he won his last world title in 1978. O'Sullivan also broke Hendry's record of 70 wins at the Crucible, setting a new record of 74.
Neil Robertson made a maximum break in his second-round match against Jack Lisowski, the fifth of his career and the 12th time a 147 had been achieved at the Crucible. Graeme Dott also made a maximum break in his third-round qualifying match against Pang Junxu, the second of his career and the fifth time that a maximum had been made in the World Championship qualifiers. The main stage produced a record 109 century breaks, surpassing the 108 centuries made the preceding year. Mark Williams made 16 centuries during the event, equalling the record set by Hendry in 2002.
|
Extract from the paragraph the winner of the 2022 World Snooker Championship.
|
Ronnie O'Sullivan
|
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.
In our current approach, subwords between source and target are not shared. A subword that is in both English and foreign vocabulary has two different embeddings.
|
Are the subwords between source and target shared in the authors' current approach?
|
No.
|
null | false
| null |
Give me a list of best Studio Ghibli Films
|
1. My Neighbour Totoro
2. Princess Mononoke
3. Howl's Moving Castle
4. Spirited Away
5. Kiki's Delivery Service
6. The tale of Princess Kaguya
7. Grave of the Fireflies
8. The Red Turtle
9. Ocean Waves
10. From Up on Poppy Hill
|
|
null | false
| null |
Classify each of the following as a French or Italian company: Campari, Dior, Hermes, Ferrari, Prada
|
Campari: Italian
Dior: French
Hermes: French
Ferrari: Italian
Prada: Italian
|
|
null | false
| 272
|
With the rapid growth of the internet, huge amounts of text data are generated in social networks, online shopping and news websites, etc. These data create demand for powerful and efficient text analysis techniques. Probabilistic topic models such as Latent Dirichlet Allocation (LDA) BIBREF0 are popular approaches for this task, by discovering latent topics from text collections. Many conventional topic models discover topics purely based on the word-occurrences, ignoring the meta information (a.k.a., side information) associated with the content. In contrast, when we humans read text it is natural to leverage meta information to improve our comprehension, which includes categories, authors, timestamps, the semantic meanings of the words, etc. Therefore, topic models capable of using meta information should yield improved modelling accuracy and topic quality.
In practice, various kinds of meta information are available at the document level and the word level in many corpora. At the document level, labels of documents can be used to guide topic learning so that more meaningful topics can be discovered. Moreover, it is highly likely that documents with common labels discuss similar topics, which could further result in similar topic distributions. For example, if we use authors as labels for scientific papers, the topics of the papers published by the same researcher can be closely related.
At the word level, different semantic/syntactic features are also accessible. For example, there are features regarding word relationships, such as synonyms obtained from WordNet BIBREF1 , word co-occurrence patterns obtained from a large corpus, and linked concepts from knowledge graphs. It is preferable that words having similar meaning but different morphological forms, like “dog” and “puppy”, are assigned to the same topic, even if they barely co-occur in the modelled corpus. Recently, word embeddings generated by GloVe BIBREF2 and word2vec BIBREF3 , have attracted a lot of attention in natural language processing and related fields. It has been shown that the word embeddings can capture both the semantic and syntactic features of words so that similar words are close to each other in the embedding space. It seems reasonable to expect that these word embedding will improve topic modelling BIBREF4 , BIBREF5 .
Conventional topic models can suffer from a large performance degradation over short texts (e.g., tweets and news headlines) because of insufficient word co-occurrence information. In such cases, meta information of documents and words can play an important role in analysing short texts by compensating the lost information in word co-occurrences. At the document level, for example, tweets are usually associated with hashtags, users, locations, and timestamps, which can be used to alleviate the data sparsity problem. At the word level, word semantic similarity and embeddings obtained or trained on large external corpus (e.g., Google News or Wikipedia) have been proven useful in learning meaningful topics from short texts BIBREF6 , BIBREF7 .
The benefit of using document and word meta information separately is shown in several models such as BIBREF8 , BIBREF9 , BIBREF5 . However, in existing models this is usually not efficient enough due to non-conjugacy and/or complex model structures. Moreover, only one kind of meta information (either at document level or at word level) is used in most existing models. In this paper, we propose MetaLDA, a topic model that can effectively and efficiently leverage arbitrary document and word meta information encoded in binary form. Specifically, the labels of a document in MetaLDA are incorporated in the prior of the per-document topic distributions. If two documents have similar labels, their topic distributions should be generated with similar Dirichlet priors. Analogously, at the word level, the features of a word are incorporated in the prior of the per-topic word distributions, which encourages words with similar features to have similar weights across topics. Therefore, both document and word meta information, if and when they are available, can be flexibly and simultaneously incorporated using MetaLDA. MetaLDA has the following key properties:
We conduct extensive experiments with several real datasets including regular and short texts in various domains. The experimental results demonstrate that MetaLDA achieves improved performance in terms of perplexity, topic coherence, and running time.
The experimental results demonstrate that MetaLDA achieves improved performance in terms of perplexity, topic coherence, and running time.
|
Does the MetaLDA achieve improved performance in terms of perplexity, topic coherence, and running time?
|
Yes, it does.
|
null | false
| null |
Which of the following venomous snakes lay eggs: copperhead, water moccasin, rattlesnake, coral snake.
|
The coral snake is the only venomous snake listed that is an egg-layer, or oviparous. Copperheads, water moccasin and rattlesnakes are ovoviviparous, and bear live young.
|
|
null | false
| null |
A tiger economy is the economy of a country which undergoes rapid economic growth, usually accompanied by an increase in the standard of living. The term was originally used for the Four Asian Tigers (South Korea, Taiwan, Hong Kong, and Singapore) as tigers are important in Asian symbolism, which also inspired the Tiger Cub Economies (Indonesia, Malaysia, Thailand, Vietnam and the Philippines). The Asian Tigers also inspired other economies later on; the Anatolian Tigers (certain cities in Turkey) in the 1980s, the Gulf Tiger (Dubai) in the 1990s, the Celtic Tiger (Republic of Ireland) in 1995–2000, the Baltic tigers (Baltic states) in 2000–2007, and the Tatra Tiger (Slovakia) in 2002–2007.
|
Which countries are referred to as tiger economy?
|
The term "Tiger Economy" originally used for the South Korea, Taiwan, Hong Kong and Singapore. It also inspired "Tiger Cub Economies" - Indonesia, Malaysia, Thailand, Vietnam and the Philippines.
|
null | false
| 127
|
Adversarial examples, a term introduced in BIBREF0, are inputs transformed by small perturbations that machine learning models consistently misclassify. The experiments are conducted in the context of computer vision (CV), and the core idea is encapsulated by an illustrative example: after imperceptible noises are added to a panda image, an image classifier predicts, with high confidence, that it is a gibbon. Interestingly, these adversarial examples can also be used to improve the classifier — either as additional training data BIBREF0 or as a regularisation objective BIBREF1 — thus providing motivation for generating effective adversarial examples.
The germ of this paper comes from our investigation of adversarial attack methods for natural language processing (NLP) tasks, e.g. sentiment classification, which drives us to quantify what is an “effective” or “good” adversarial example. In the context of images, a good adversarial example is typically defined according two criteria:
it has successfully fooled the classifier;
it is visually similar to the original example.
In NLP, defining a good adversarial example is a little more involving, because while criterion (b) can be measured with a comparable text similarity metric (e.g. BLEU or edit distance), an adversarial example should also:
be fluent or natural;
preserve its original label.
These two additional criteria are generally irrelevant for images, as adding minor perturbations to an image is unlikely to: (1) create an uninterpretable image (where else changing one word in a sentence can render a sentence incoherent), or (2) change how we perceive the image, say from seeing a panda to a gibbon (but a sentence's sentiment can be reversed by simply adding a negative adverb such as not). Without considering criterion (d), generating adversarial examples in NLP would be trivial, as the model can learn to simply replace a positive adjective (amazing) with a negative one (awful) to attack a sentiment classifier.
To the best of our knowledge, most studies on adversarial example generation in NLP have largely ignored these additional criteria BIBREF2, BIBREF3, BIBREF4, BIBREF5. We believe the lack of a rigorous evaluation framework partially explains why adversarial training for NLP models has not seen the same extent of improvement compared to CV models. As our experiments reveal, examples generated from most attacking methods are successful in fooling the classifier, but their language is often unnatural and the original label is not properly preserved.
The core contribution of our paper is to introduce a systematic, rigorous evaluation framework to assess the quality of adversarial examples for NLP. We focus on sentiment classification as the target task, as it is a popular application that highlights the importance of criteria discussed above. We test a number of attacking methods and also propose an alternative approach (based on an auto-encoder) for generating adversarial examples. We learn that a number of factors can influence the performance of adversarial attacks, including architecture of the classifier, sentence length and input domain.
The core contribution of our paper is to introduce a systematic, rigorous evaluation framework to assess the quality of adversarial examples for NLP.
|
What framework does the paper introduce?
|
A systematic, rigorous evaluation framework to assess the quality of adversarial examples for NLP.
|
null | false
| null |
Which of these are cities and which are countries: Paris, Argentina, Mountain View, Singapore, America, Cape Town
|
Paris, Mountain View and Cape Town are cities. Argentina and America are countries. Singapore is the name of both a city and a country.
|
|
null | false
| 326
|
Deep learning models have been widely used in many natural language processing (NLP) tasks. A major challenge is how to design and learn the semantic composition function while modeling a text sequence. The typical composition models involve sequential BIBREF0 , BIBREF1 , convolutional BIBREF2 , BIBREF3 , BIBREF4 and syntactic BIBREF5 , BIBREF6 , BIBREF7 compositional models.
In spite of their success, these models have two major limitations. First, they usually use a shared composition function for all kinds of semantic compositions, even though the compositions have different characteristics in nature. For example, the composition of the adjective and the noun differs significantly from the composition of the verb and the noun. Second, different composition functions are learned from scratch in different tasks. However, given a certain natural language, its composition functions should be the same (on meta-knowledge level at least), even if the tasks are different.
To address these problems, we need to design a dynamic composition function which can vary with different positions and contexts in a sequence, and share it across the different tasks. To share some meta-knowledge of composition function, we can adopt the multi-task learning BIBREF8 . However, the sharing scheme of most neural multi-task learning methods is feature-level sharing, where a subspace of the feature space is shared across all the tasks. Although these sharing schemes are successfully used in various NLP tasks BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 , they are not suitable to share the composition function.
In this paper, inspired by recent work on dynamic parameter generation BIBREF15 , BIBREF16 , BIBREF17 , we propose a function-level sharing scheme for multi-task learning, in which a shared meta-network is used to learn the meta-knowledge of semantic composition among the different tasks. The task-specific semantic composition function is generated by the meta-network. Then the task-specific composition function is used to obtain the task-specific representation of a text sequence. The difference between two sharing schemes is shown in Figure 1 . Specifically, we use two LSTMs as meta and basic (task-specific) network respectively. The meta LSTM is shared for all the tasks. The parameters of the basic LSTM are generated based on the current context by the meta LSTM, therefore the composition function is not only task-specific but also position-specific. The whole network is differentiable with respect to the model parameters and can be trained end-to-end.
We demonstrate the effectiveness of our architectures on two kinds of NLP tasks: text classification and sequence tagging. Experimental results show that jointly learning of multiple related tasks can improve the performance of each task relative to learning them independently.
Our contributions are of three-folds:
We demonstrate the effectiveness of our architectures on two kinds of NLP tasks: text classification and sequence tagging.
|
What do the authors demonstrate the effectiveness of their architectures on?
|
The authors demonstrate the effectiveness of the architectures on two kinds of NLP tasks: text classification and sequence tagging.
|
null | false
| 375
|
The use of open-source software has been steadily increasing for some time now, with the number of Java packages in Maven Central doubling in 2018. However, BIBREF0 states that there has been an 88% growth in the number of vulnerabilities reported over the last two years. In order to develop secure software, it is essential to analyze and understand security vulnerabilities that occur in software systems and address them in a timely manner. While there exist several approaches in the literature for identifying and managing security vulnerabilities, BIBREF1 show that an effective vulnerability management approach must be code-centric. Rather than relying on metadata, efforts must be based on analyzing vulnerabilities and their fixes at the code level.
Common Vulnerabilities and Exposures (CVE) is a list of publicly known cybersecurity vulnerabilities, each with an identification number. These entries are used in the National Vulnerability Database (NVD), the U.S. government repository of standards based vulnerability management data. The NVD suffers from poor coverage, as it contains only 10% of the open-source vulnerabilities that have received a CVE identifier BIBREF2. This could be due to the fact that a number of security vulnerabilities are discovered and fixed through informal communication between maintainers and their users in an issue tracker. To make things worse, these public databases are too slow to add vulnerabilities as they lag behind a private database such as Snyk's DB by an average of 92 days BIBREF0 All of the above pitfalls of public vulnerability management databases (such as NVD) call for a mechanism to automatically infer the presence of security threats in open-source projects, and their corresponding fixes, in a timely manner.
We propose a novel approach using deep learning in order to identify commits in open-source repositories that are security-relevant. We build regularized hierarchical deep learning models that encode features first at the file level, and then aggregate these file-level representations to perform the final classification. We also show that code2vec, a model that learns from path-based representations of code and claimed by BIBREF3 to be suitable for a wide range of source code classification tasks, performs worse than our logistic regression baseline.
In this study, we seek to answer the following research questions:
[leftmargin=*]
RQ1: Can we effectively identify security-relevant commits using only the commit diff? For this research question, we do not use any of the commit metadata such as the commit message or information about the author. We treat source code changes like unstructured text without using path-based representations from the abstract syntax tree.
RQ2: Does extracting class-level features before and after the change instead of using only the commit diff improve the identification of security-relevant commits? For this research question, we test the hypothesis that the source code of the entire Java class contains more information than just the commit diff and could potentially improve the performance of our model.
RQ3: Does exploiting path-based representations of Java source code before and after the change improve the identification of security-relevant commits? For this research question, we test whether code2vec, a state-of-the-art model that learns from path-based representations of code, performs better than our model that treats source code as unstructured text.
RQ4: Is mining commits using regular expression matching of commit messages an effective means of data augmentation for improving the identification of security-relevant commits? Since labelling commits manually is an expensive task, it is not easy to build a dataset large enough to train deep learning models. For this research question, we explore if collecting coarse data samples using a high-precision approach is an effective way to augment the ground-truth dataset.
The main contributions of this paper are:
[leftmargin=*]
Novel hierarchical deep learning models for the identification of security-relevant commits based on either the diff or the modified source code of the Java classes.
A comparative analysis of how various deep learning models perform across different input representations and how various regularization techniques help with the generalization of our models.
We envision that this work would ultimately allow for monitoring open-source repositories in real-time, in order to automatically detect security-relevant changes such as vulnerability fixes.
We also show that code2vec, a model that learns from path-based representations of code and claimed by Alon et al. [2018c] to be suitable for a wide range of source code classification tasks, performs worse than our logistic regression baseline.
|
Is a state-of-the-art model that learns from path-based representations of code and claimed by Common Vulnerabilities and Exposures (CVE) to be suitable for a wide range of source code classification tasks, performing better than the logistic regression baseline?
|
No.
|
null | false
| null |
In discrete mathematics, and more specifically in graph theory, a graph is a structure amounting to a set of objects in which some pairs of the objects are in some sense "related". The objects correspond to mathematical abstractions called vertices (also called nodes or points) and each of the related pairs of vertices is called an edge (also called link or line). Typically, a graph is depicted in diagrammatic form as a set of dots or circles for the vertices, joined by lines or curves for the edges. Graphs are one of the objects of study in discrete mathematics.
|
What is a graph in graph theory?
|
In graph theory, a graph is a structure amounting to a set of objects in which some pairs of the objects are in some sense "related".
|
null | false
| null |
In the US, we are fortunate to have a dizzying array of fruits that fill our grocery stores year-round. They come in all sizes, shapes, and colors, and we have all heard about the recommended five servings of fruits and vegetables a day. So, what are we eating them for? And how does the nutritional value vary between fruits? Is there any difference between whole fruits versus juice, fresh versus dried? Let’s take a look.
Differences between fruits
Just like other foods, different fruits have different nutrient values. Generally, whole fruits are good sources of fiber while fruit juices are not. And one cup of fruit juice, even 100% fruit juice, has a lot more sugar than one piece or one serving of whole fruit. In addition, whole fruits are more satiating than juices. When meeting the recommended fruit and vegetable intake, it is better to eat them (whole) than drink them (juice). However, one should not completely avoid drinking juice — if it is 100% juice — but you should limit consumption to no more than 4 to 8 ounces a day.
The freezer section of the grocery store is often stocked with quite a variety of frozen fruits. These are often peeled and cut already (like mango), which is convenient and often less expensive than fresh fruits. Frozen fruits are usually picked and quick-frozen near the point of harvest, therefore the nutrients are well preserved. Moreover, some seasonal fruits such as blueberries are readily available in frozen form. The key to selection is to choose plain frozen fruits without added sugar.
There are a number of fruits that are available in dried form, such as raisins, apricots, and pineapple — just to name a few. They also have good nutrient values, keep for a long time, are convenient to carry around, and are high in calories, making them a favorite for hikers and campers. However, some often have sugar added in the drying process, particularly mango and pineapple. Dried cranberries almost always have sugar added, as they are naturally very tart. Even for those without added sugar, the compact volume and sweetness make it quite easy to eat a lot in one sitting, and the calories can add up quickly.
Some dried fruits like raisins and apricots are also treated with sulfur dioxide to preserve freshness and color. For most people that is not a concern; however, some individuals are sensitive, especially those with asthma. Sulfur dioxide treatment is labeled on the package, so it is not difficult to avoid if necessary.
What about buying organic?
We have much choice when it comes to organic and conventionally grown fruits, be they fresh, frozen, or dried. Nutritionally, there is not enough difference to choose one over the other, though consumers might choose one over another based on farming practices and environmental impact. The US has regulations on pesticide use, but some fruits tend to have more residual pesticides than others, and it is always recommended that you wash fruits thoroughly before eating.
Differences in nutritional value
Different fruits are good sources of different nutrients. Citrus fruits are high in vitamin C, and other fruits are good sources of nutrients too. A few examples below:
Nutrient: Major functions - Good fruit sources
potassium: fluid and electrolyte balance, maintaining healthy blood pressure - oranges, raspberries, bananas, cherries, pomegranates, honeydew melons, avocados
iron: formation of red blood cells, brain growth in children - dried apricots, raisins
vitamin C: maintaining healthy immune system, wound healing, antioxidant - citrus fruits, strawberries, kiwi
folate: DNA synthesis, formation of red blood cells, early neural tube development in the fetus - oranges, mangos, avocados
vitamin A: night vision, cell growth, immune function - cantaloupe
Besides the above nutrients, certain fruits are also high in flavonoids. This is a diverse group of compounds, some of which are potent antioxidants that protect against oxidative damage, and may reduce the risk of certain diseases such as cardiovascular disease and diabetes. In particular, citrus fruits are high in the flavanones class of flavonoids, and blackberries, blueberries, cranberries, and cherries are high in the anthocyanidins class of flavonoids.
So, what fruits should we eat?
As you can see, there isn’t one fruit that has all the nutrients, so eating a variety is the key to good health. Try something new! Most adults should eat a variety of colors for about 2 cups a day. Prioritize whole fruits over juice. Eat what is in season, as it will be cheaper. And enjoy your fruits: eat mindfully to fully appreciate the smell, texture, and flavor. Bon appetit!
|
Based on the text, what are the best ways for people to eat fruits?
|
- Eat whole fruits instead of drinking 100% fruit juice, event 100% fruit juice, because whole fruits contain a lot more fiber and less super
- Eat fruits of the season. They are fresher and cheaper.
- It is ok to eat freshly frozen fruits. Pay attention to pick frozen fruits which do not have sugar added during the process
- There is no nutritional difference between organic and non-organic fruits. Always wash fruits before eating
- Dried fruits contain more sugar per unit weight. Don't overeat.
- Different fruits contain different nutrients. Eating a variety of different fruits is the key to good health
- A healthy adult should eat 2 cups of fruits a day with different varieties
|
null | false
| null |
what is machine learning meaning?
|
Machine learning is a branch of artificial intelligence (AI) and computer science that focuses on using data and algorithms to imitate how humans learn, gradually improving its accuracy.
|
|
null | false
| null |
Which of the following are countries? Pakistan, Azerbaijan, Hodor, Chicago, St. Louis, Michael, Jordan, Minecraft, Ninja, Teletubbies, Fortnite, Lesotho, Greece
|
Pakistan, Azerbaijan, Jordan, Lesotho, and Greece
|
|
null | false
| null |
What makes sourdough special?
|
Sourdough is made using sourdough starter, which is a fermented flour and water mixture that contains natural yeast. This gives the bread a distinct "sour" flavor which is different than breads made with commercial yeast.
|
|
null | false
| null |
List some capital cities that begin with the letter S.
|
Capital cities starting with the letter "S" include; San Marino, Santiago, Seoul, Singapore, Stockholm and Sydney.
|
|
null | false
| null |
How do you ride a bike?
|
You can learn to ride a bike by starting with training wheels. This will help you to develop your balance as muscle memory in body. After you ride for a couple weeks with training wheels you can remove the training wheels and use your feet to help you balance. With enough speed you will want to put your feet on the pedals and get enough velocity to keep your balance going in a straight line. Practice this for a couple days, and then you can practice turning and use your feet if you lose your balance. Do this for a couple more days and then you will be off to the races riding your own bike with perfect balance.
|
|
null | false
| null |
Are these surfing terms or not? Nutted, pitted, drop in, shwabang, glass, poot, crest
|
The surfing terms are the following: pitted, drop in, glass, crest
|
|
null | false
| null |
One of the best place to visit in Mexico
|
Cancun : You can visit Cancun for the tropical weather. You can visit Cancun all over the year. You can avoid September to October because of little bit of rain and cloudy weather.
Cancun has some splendid Resorts with All inclusive package . The beach is clean.
You can stay either in Cancun or Playa Del Carman or The Cancun . Both has similar resorts mostly. You can visit Tulum if you like calm and nature associated with Beach.
|
|
1910.03042
| false
| null |
Figure FIGREF3 provides an overview of Gunrock's architecture. We extend the Amazon Conversational Bot Toolkit (CoBot) BIBREF6 which is a flexible event-driven framework. CoBot provides ASR results and natural language processing pipelines through the Alexa Skills Kit (ASK) BIBREF7. Gunrock corrects ASR according to the context (asr) and creates a natural language understanding (NLU) (nlu) module where multiple components analyze the user utterances. A dialog manager (DM) (dm) uses features from NLU to select topic dialog modules and defines an individual dialog flow. Each dialog module leverages several knowledge bases (knowledge). Then a natural language generation (NLG) (nlg) module generates a corresponding response. Finally, we markup the synthesized responses and return to the users through text to speech (TTS) (tts). While we provide an overview of the system in the following sections, for detailed system implementation details, please see the technical report BIBREF1.
We extend the Amazon Conversational Bot Toolkit (CoBot) BIBREF6 which is a flexible event-driven framework. CoBot provides ASR results and natural language processing pipelines through the Alexa Skills Kit (ASK) BIBREF7. Gunrock corrects ASR according to the context (asr) and creates a natural language understanding (NLU) (nlu) module where multiple components analyze the user utterances. A dialog manager (DM) (dm) uses features from NLU to select topic dialog modules and defines an individual dialog flow. Each dialog module leverages several knowledge bases (knowledge). Then a natural language generation (NLG) (nlg) module generates a corresponding response. Finally, we markup the synthesized responses and return to the users through text to speech (TTS) (tts).
|
What the system designs introduced?
|
The answers are shown as follows:
* Amazon Conversational Bot Toolkit
* natural language understanding (NLU) (nlu) module
* dialog manager
* knowledge bases
* natural language generation (NLG) (nlg) module
* text to speech (TTS) (tts)
|
null | false
| null |
What states do not have a state gambling lottery?
|
Alabama, Alaska, Hawaii, Nevada, and Utah.
|
|
null | false
| null |
Why did the Soviet Union collapse?
|
There are a multitude of reasons for the collapse of the Soviet Union. But historians have generally narrowed it down to the following major causes:
Rise of ethnic nationalism in Soviet Republics: After decades of suppressing national identities, ethnic groups began asserting their nationality and demanded greater autonomy and political power. This resulted in numerous riots and protests throughout the Soviet Union with demands for independence and national self-determination.
Economic stagnation and fall of oil prices: The centrally planned economic model resulted in massive shortages of staple consumer products. Shops were continually empty and Soviet citizens needed to wait in long lines to purchase basic goods. After years of economic deprivation, Soviet citizens were deeply angry at their government. In addition, oil prices declined significantly in the late 1980s, which provided the Soviet government a key source of hard currency. This it made it difficult for the Soviet government to sustain high levels of military spending and keep the economy functioning.
Perestroika and Glasnost: Gorbachev's reforms of the 1980s were intended to reinvigorate the Soviet economy. However, they unintentionally triggered an opening of debate throughout the country. The increased freedoms allowed people to more openly criticize the government and demand changes.
Afghanistan and Chernobyl: In the late 1980s, two events symbolized the incompetence of the Soviet government. The failure of the Afghanistan war resulted in a loss of confidence in the Soviet military. The Chernobyl disaster revealed the incompetence and duplicity of the Soviet regime, resulting in a further erosion of trust.
|
|
null | false
| null |
Tell me which of these are shades of Blue that BWM paints their cars: Miami Blue, Laguna Seca Blue, Le Mans Blue, Interlagos Blue, Heritage Blue, Mexico Blue, Yas Marina Blue, Gulf Blue, Long Beach Blue.
|
Laguna Seca Blue, Le Mans Blue, Interlagos Blue, Yas Marina Blue, Long Beach Blue
|
|
null | false
| null |
On May 5, 2022, in a secret auction at the brand's museum in Germany, Mercedes-Benz sold one of just two 1955 300 SLR Uhlenhaut coupes from its extensive collection of historical automobiles—which dates back to the earliest days of the car in the late 19th century. The sale, to a private buyer, was for 135 million euros ($142,769,250). It handily outstripped the previous record-setting $48.4-million sale of a 1962 Ferrari 250 GTO at a 2018 auction to become the most expensive car ever sold at auction. Both of these high-dollar sales were brokered by RM Sotheby's.
|
What was the most expensive car ever sold at auction
|
A 1955 Mercedes-Benz 300 SLR was reportedly sold in a secret auction for 135 Million Euros
|
null | false
| null |
What do to if your cat starts drooling excessively?
|
Although excessively drooling could be a sign of your cat eating something distasteful, it could also be an underlying serious disease like teeth ache or poisoning, so be sure to seek out veterinary care as soon as possible.
|
|
null | false
| 55
|
We implement all the models with the same PyTorch framework. Below, we summarize the implementation details that are key for reproducing results for each model:
SSE: This model can converge very fast, for example, 2 or 3 epochs for the SNLI dataset. We control the convergence speed by updating the learning rate for each epoch: specifically, $lr=\frac{1}{2^{\frac{epoch\_i}{2}}}*{init\_lr}$ , where $init\_lr$ is the initial learning rate and $epoch\_i$ is the index of current epoch.
DecAtt: It is important to use gradient clipping for this model: for each gradient update, we check the L2 norm of all the gradient values, if it is greater than a threshold $b$ , we scale the gradient by a factor $\alpha = b/L2\_norm$ . Another useful procedure is to assemble batches of sentences with similar length.
ESIM: Similar but different from DecAtt, ESIM batches sentences with varied length and uses masks to filter out padding information. In order to batch the parse trees within Tree-LSTM recursion, we follow Bowman et al.'s bowman-EtAl:2016:P16-1 procedure that converts tree structures into the linear sequential structure of a shift reduce parser. Two additional masks are used for producing left and right children of a tree node.
PWIM: The cosine and Euclidean distances used in the word interaction layer have smaller values for similar vectors while dot products have larger values. The performance increases if we add a negative sign to make all the vector similarity measurements behave consistently.
We implement all the models with the same PyTorch framework. Below, we summarize the implementation details that are key for reproducing results for each model:
SSE: This model can converge very fast, for example, 2 or 3 epochs for the SNLI dataset. We control the convergence speed by updating the learning rate for each epoch: specifically, $lr=\frac{1}{2^{\frac{epoch\_i}{2}}}*{init\_lr}$ , where $init\_lr$ is the initial learning rate and $epoch\_i$ is the index of current epoch.
DecAtt: It is important to use gradient clipping for this model: for each gradient update, we check the L2 norm of all the gradient values, if it is greater than a threshold $b$ , we scale the gradient by a factor $\alpha = b/L2\_norm$ . Another useful procedure is to assemble batches of sentences with similar length.
ESIM: Similar but different from DecAtt, ESIM batches sentences with varied length and uses masks to filter out padding information. In order to batch the parse trees within Tree-LSTM recursion, we follow Bowman et al.'s bowman-EtAl:2016:P16-1 procedure that converts tree structures into the linear sequential structure of a shift reduce parser. Two additional masks are used for producing left and right children of a tree node.
PWIM: The cosine and Euclidean distances used in the word interaction layer have smaller values for similar vectors while dot products have larger values. The performance increases if we add a negative sign to make all the vector similarity measurements behave consistently.
We implement all the models with the same PyTorch framework.23 Below, we summarize the implementation details that are key for reproducing results for each model: ****•DecAtt: It is important to use gradient clipping for this model: for each gradient update, we check the L2 norm of all the gradient values, if it is greater than a threshold b, we scale the gradient by a factor a = b/L2-norm. Another useful procedure is to assemble batches of sentences with similar length.
|
What details are key for reproducing results related to DecAtt?
|
It is important to use gradient clipping for this model: for each gradient update, they check the L2 norm of all the gradient values, if it is greater than a threshold b, they scale the gradient by a factor a = b/L2-norm. Another useful procedure is to assemble batches of sentences with similar length.
|
null | false
| null |
Name the Chinese satellite launched for Mars which failed?
|
Yinghuo-1 was the Chinese satellite that failed the launch for Mars
|
|
null | false
| null |
The six main types of motorcycles are generally recognized as standard, cruiser, touring, sports, off-road, and dual-purpose. Sport touring is sometimes recognized as a seventh category or integrated with the touring category.
Although there are many names and systems for classifying types of motorcycles based on their characteristics and usage, there are generally six categories recognized by most motorcycle manufacturers and organizations.
Strong distinctions are usually made between the six main types of motorcycles and other motorcycles. Scooter, moped, underbone, miniature, pocket, electric, and three-wheeled motorcycles are generally excluded from the main categories but other classification schemes may include these as types of motorcycles.
There is no universal system for classifying all types of motorcycles. There are strict classification systems enforced by competitive motorcycle sport sanctioning bodies, or legal definitions of a motorcycle established by certain legal jurisdictions for motorcycle registration, emissions, road traffic safety rules or motorcyclist licensing. There are also informal classifications or nicknames used by manufacturers, riders, and the motorcycling media. Some experts do not recognize sub-types, like naked bike, that "purport to be classified" outside the six usual classes, because they fit within one of the main types and are recognizable only by cosmetic changes.
Street motorcycles are motorcycles designed for being ridden on paved roads. They have smooth tires with tread patterns and engines generally in the 125 cc (7.6 cu in) and over range. Typically, street motorcycles are capable of speeds up to 100 mph (160 km/h), and many of speeds in excess of 125 mph (201 km/h). Street motorcycles powered by electric motors are becoming more common, with firms like Harley-Davidson entering the market.
|
Without quoting directly from the text give me a summary of the different types of motorcycles.
|
Motorcycles are 2 wheeled vehicles powered by an internal combustion engine generally made for a single rider. Although there are many styles of motorcycles, they are loosely organized into 6 main categories. These categories most closely align to the intended use of the vehicle. There is your standard motorcycle, your more relaxed cruising motorcycle, the larger capacity and longer haul touring cycle, a racing more sportier sports cycle and then you have a off-road motorcycle often known as a dirtbike and lastly a dual-sport which is a motorcycle that can go both on and offroad.
|
2004.02083
| false
| null |
Morphological inflection has been thoroughly studied in monolingual high resource settings, especially through the recent SIGMORPHON challenges BIBREF8, BIBREF9, BIBREF10, with the latest iteration focusing more on low-resource settings, utilizing cross-lingual transfer BIBREF11. We use the guidelines of the state-of-the-art approach of BIBREF12 that achieved the highest inflection accuracy in the latest SIGMORPHON 2019 morphological inflection shared task. Our models are implemented in DyNet BIBREF13.
Morphological analysis is the task of creating a morphosyntactic description for a given word. It can be framed in a context-agnostic manner (as in our case) or within a given context, as for instance for the SIGMORPHON 2019 second shared task BIBREF11. We trained an encoder-decoder model that receives the form as character-level input, encodes it with a BiLSTM encoder, and then an attention enhanced decoder BIBREF14 outputs the corresponding sequence of morphological tags, implemented in DyNet. The baseline results are shown in Table . The exact-match accuracy of 67% is lower than the average accuracy that context-aware systems can achieve, and it highlights the challenge that the complexity of the tonal system of SJQ Chatino can pose.
Lemmatization is the task of retrieving the underlying lemma from which an inflected form was derived. Although in some languages the lemma is distinct from all forms, in SJQ Chatino the lemma is defined as the completive third-person singular form. As a computational task, lemmatization entails producing the lemma given an inflected form (and possibly, given a set of morphological tags describing the input form). Popular approaches tackle it as a character-level edit sequence generation task BIBREF15, or as a character-level sequence-to-sequence task BIBREF16. For our baseline lemmatization systems we follow the latter approach. We trained a character level encoder-decoder model, similar to the above-mentioned inflection system, implemented in DyNet.
We use the guidelines of the state-of-the-art approach of BIBREF12 that achieved the highest inflection accuracy in the latest SIGMORPHON 2019 morphological inflection shared task. Our models are implemented in DyNet BIBREF13.
We trained an encoder-decoder model that receives the form as character-level input, encodes it with a BiLSTM encoder, and then an attention enhanced decoder BIBREF14 outputs the corresponding sequence of morphological tags, implemented in DyNet.
We trained a character level encoder-decoder model, similar to the above-mentioned inflection system, implemented in DyNet.
|
What system is used as baseline?
|
The answers are shown as follows:
* DyNet
|
null | false
| 163
|
The task assigned to annotators was to read sentences one at a time and label them with with binary labels indicating the polarity (i.e., positive/negative). Note that, the participants were not instructed to annotate whether a sentence is sarcastic or not., to rule out the Priming Effect (i.e., if sarcasm is expected beforehand, processing incongruity becomes relatively easier BIBREF12 ). The setup ensures its “ecological validity” in two ways: (1) Readers are not given any clue that they have to treat sarcasm with special attention. This is done by setting the task to polarity annotation (instead of sarcasm detection). (2) Sarcastic sentences are mixed with non sarcastic text, which does not give prior knowledge about whether the forthcoming text will be sarcastic or not.
The eye-tracking experiment is conducted by following the standard norms in eye-movement research BIBREF13 . At a time, one sentence is displayed to the reader along with the “aspect” with respect to which the annotation has to be provided. While reading, an SR-Research Eyelink-1000 eye-tracker (monocular remote mode, sampling rate 500Hz) records several eye-movement parameters like fixations (a long stay of gaze) and saccade (quick jumping of gaze between two positions of rest) and pupil size.
The accuracy of polarity annotation varies between 72%-91% for sarcastic texts and 75%-91% for non-sarcastic text, showing the inherent difficulty of sentiment annotation, when sarcasm is present in the text under consideration. Annotation errors may be attributed to: (a) lack of patience/attention while reading, (b) issues related to text comprehension, and (c) confusion/indecisiveness caused due to lack of context.
For our analysis, we do not discard the incorrect annotations present in the database. Since our system eventually aims to involve online readers for sarcasm detection, it will be hard to segregate readers who misinterpret the text. We make a rational assumption that, for a particular text, most of the readers, from a fairly large population, will be able to identify sarcasm. Under this assumption, the eye-movement parameters, averaged across all readers in our setting, may not be significantly distorted by a few readers who would have failed to identify sarcasm. This assumption is applicable for both regular and multi-instance based classifiers explained in section SECREF6 .
Annotation errors may be attributed to: (a) lack of patience/attention while reading, (b) issues related to text comprehension, and (c) confusion indecisiveness caused due to lack of context.
|
What might be the reasons for annotation errors?
|
(a) lack of patience/attention while reading, (b) issues related to text comprehension, and (c) confusion indecisiveness caused due to lack of context.
|
null | false
| null |
What are the primary causes of obesity?
|
The primary causes are eating significantly more calories than you burn daily, lack of exercise/movement, lack of sleep and increased stress, your genetic makeup.
|
|
1912.02481
| false
| null |
The Spearman $\rho $ correlation for fastText models on the curated small dataset (clean), C1, improves the baselines by a large margin ($\rho =0.354$ for Twi and 0.322 for Yorùbá) even with a small dataset. The improvement could be justified just by the larger vocabulary in Twi, but in the case of Yorùbá the enhancement is there with almost half of the vocabulary size. We found out that adding some noisy texts (C2 dataset) slightly improves the correlation for Twi language but not for the Yorùbá language. The Twi language benefits from Wikipedia articles because its inclusion doubles the vocabulary and reduces the bias of the model towards religious texts. However, for Yorùbá, noisy texts often ignore diacritics or tonal marks which increases the vocabulary size at the cost of an increment in the ambiguity too. As a result, the correlation is slightly hurt. One would expect that training with more data would improve the quality of the embeddings, but we found out with the results obtained with the C3 dataset, that only high-quality data helps. The addition of JW300 boosts the vocabulary in both cases, but whereas for Twi the corpus mixes dialects and is noisy, for Yorùbá it is very clean and with full diacritics. Consequently, the best embeddings for Yorùbá are obtained when training with the C3 dataset, whereas for Twi, C2 is the best option. In both cases, the curated embeddings improve the correlation with human judgements on the similarity task a $\Delta \rho =+0.25$ or, equivalently, by an increment on $\rho $ of 170% (Twi) and 180% (Yorùbá).
One would expect that training with more data would improve the quality of the embeddings, but we found out with the results obtained with the C3 dataset, that only high-quality data helps.
|
What turn out to be more important high volume or high quality data?
|
The answers are shown as follows:
* high-quality
|
null | false
| null |
Article II, Section 2, Clause 2 of the United States Constitution, known as the Appointments Clause, empowers the president to nominate and, with the confirmation (advice and consent) of the United States Senate, to appoint public officials, including justices of the Supreme Court. This clause is one example of the system of checks and balances inherent in the Constitution. The president has the plenary power to nominate, while the Senate possesses the plenary power to reject or confirm the nominee. The Constitution sets no qualifications for service as a justice, thus a president may nominate anyone to serve, and the Senate may not set any qualifications or otherwise limit who the president can choose.
|
Explain how a Justice is appointed to the US Supreme Court
|
The US President nominates a Supreme Court judge who subsequently has to be confirmed by the US Senate.
|
null | false
| null |
What are the debt securities?
|
Debt securities are financial assets that define the terms of a loan between an issuer (the borrower) and an investor (the lender). The terms of a debt security typically include the principal amount to be returned upon maturity of the loan, interest rate payments, and the maturity date or renewal date.
|
|
null | false
| null |
Where does Delta Lake store the data?
|
When writing data, you can specify the location in your cloud storage. Delta Lake stores the data in that location in Parquet format.
|
|
null | false
| null |
What was the primary reason for the mutiny that occurred in 1857, in British India?
|
There were several reasons that can be attributed to the mutiny that occurred in 1857, but one of the major reasons was related to the Hindu religious beliefs. Most of the British army were comprised of Hindus who were extremely religious. The Britishers were insensitive to their religious beliefs, for e.g. they were asked to chew of the bullet before using it in their rifles that was made of animal fat. Most Hindus in the army were religiously vegetarian and were offended when asked to bite meat. This event started bringing together the Hindus in the army to plot a revolt against the Britishers. Mangal Panday, one of the Hindu soldiers has been credited with being the leader of this revolt and this revolt is often referred to as the first war of Indian independence.
|
|
null | false
| null |
How can I improve my finances?
|
Determine your short and long term financial goals.
Create a budget based on your total income, necessary, and discretionary expenses.
Pay down high interest debt first, such as credit card debt.
Pay off your credit cards in full each month.
Create an emergency fund that covers 3-6 months of expenses.
Max out your 401K contributions every year.
Use a high interest savings account instead of saving money with a traditional bank so you don't lose money due to inflation.
Create a diversified investment account.
Hire a financial planner to manage your assets.
Turn a hobby into a second job and earn passive income.
Reduce temptations to overspend by setting up an automatic transfer from your paycheck into a savings account.
|
|
null | false
| null |
What is the largest city in Washington state?
|
Seattle
|
|
null | false
| null |
How many U.S. President were lawyers?
|
25 U.S. Presidents were lawyers.
John Adams: Adams was the second president of the United States. A farmer, lawyer, writer and political philosopher; he played a major role in American history.
Thomas Jefferson: He was a lawyer, statesman, professor, writer, and the third president of the United States, during which time he wrote the Declaration of Independence with John Adams. He also played an important role in establishing the structure of the Library of Congress.
James Madison: He was one of the Founding Fathers and fourth president of the United States. A student of law and classical political theory, Madison was the leading author of the Bill of Rights.
James Monroe: He was an American statesman, lawyer, diplomat and Founding Father and fifth president of the United States during the time of the War of 1812.
John Q. Adams: He was an American statesman, journalist, diplomat, lawyer, and sixth president of the United States.
Martin Van Buren: He was the eighth President of the United States and became the youngest person ever to become President at that time. He studied law in New York, and was admitted to the New York bar in 1802.
John Tyler: He was the tenth President of the United States when President William Henry Harrison died in April 1841. He was the first Vice President to succeed in the Presidency after the death of his predecessor. He was a lawyer by trade and became Governor of Tennessee after his election in 1844.
James Polk: He was a lawyer, surveyor and railroad worker He was the eleventh President of the United States.
Millard Fillmore: He was a lawyer, political leader, minister of finance, diplomat and statesman. He was educated in the state schools, studied law, and was admitted to the bar in 1813.
Franklin Pierce: He was a lawyer, judge, and governor of the American state of New Hampshire. He was the fourteenth President of the United States.
James Buchanan: He was a statesman, lawyer, soldier, author, and diplomat who helped found the United States and signed the Declaration of Independence.
Abraham Lincoln: Lincoln was the president during the Civil War, and his administration was a time of success and failure. He was a former statesman, lawyer, journalist, soldier and president of the United States.
Rutherford B. Hayes: He was a statesman, lawyer, congressman, minister, and the nineteenth President of the United States from 1877-1881.
Chester Arthur: He was a statesman, lawyer, and professor, who had also been an admiral of the United States Navy and a delegate to
the Republican National Convention.
Grover Cleveland: He was a statesman, lawyer, and businessman. He became President when William McKinley was assassinated. He studied law in
St. Louis, Missouri and began his law practice in 1866.
Benjamin Harrison: He was a statesman, lawyer, soldier, and U.S. Representative for the state of Indiana. As U.S. President in 1891-93, Harrison achieved the repeal of the Sherman Silver Purchase Act and the creation of the National Cordage Company.
William McKinley: His important role was to establish relations between the two nations on a more cordial basis. He was a statesman, lawyer, a Union veteran of the Civil War and the twenty-fifth President of the United States.
William Taft: He was a lawyer, judge, and statesman, and a member of the commission which framed the constitution of Ohio. He graduated from the Cincinnati Law School and was admitted to the bar in 1839.
Woodrow Wilson: He was the twenty-eighth President of the United States. He was a statesman, lawyer, and diplomat, of the preeminent stature of his era. He was the first United States President to hold foreign office as a diplomat.
Calvin Coolidge: He was a statesman, lawyer, soldier, college president and the thirtieth President of the United States.
Franklin D. Roosevelt: Roosevelt was considered “The Golden-Gate President,” having been elected in 1932 in the midst of the Great Depression. He was a statesman, lawyer, a World War I veteran who served as Assistant Secretary of the Navy in the administration of President Theodore Roosevelt.
Richard Nixon: He was the thirty-seventh President of the United States, a statesman, lawyer, statesman, and a diplomat.
Gerald Ford: He was a statesman, lawyer and Christian man, who dedicated his life to God and our country. He was honored in the presence of our current leaders.
Bill Clinton: He was a statesman, lawyer, US Senator, governor, secretary of state, US president, and Rhodes Scholar. He focused his term on expanding education in the United States, with a particular focus on the poor and minorities.
Barack Obama: Obama was the 44th president of the United States and the nation’s first black president, was a statesman, a lawyer, a law and order man and a politician.
|
|
null | false
| 209
|
Language variety identification aims at labelling texts in a native language (e.g. Spanish, Portuguese, English) with their specific variation (e.g. Argentina, Chile, Mexico, Peru, Spain; Brazil, Portugal; UK, US). Although at first sight language variety identification may seem a classical text classification problem, cultural idiosyncrasies may influence the way users construct their discourse, the kind of sentences they build, the expressions they use or their particular choice of words. Due to that, we can consider language variety identification as a double problem of text classification and author profiling, where information about how language is shared by people may help to discriminate among classes of authors depending on their language variety.
This task is specially important in social media. Despite the vastness and accessibility of the Internet destroyed frontiers among regions or traits, companies are still very interested in author profiling segmentation. For example, when a new product is launched to the market, knowing the geographical distribution of opinions may help to improve marketing campaigns. Or given a security threat, knowing the possible cultural idiosyncrasies of the author may help to better understand who could have written the message.
Language variety identification is a popular research topic of natural language processing. In the last years, several tasks and workshops have been organized: the Workshop on Language Technology for Closely Related Languages and Language Variants @ EMNLP 2014; the VarDial Workshop @ COLING 2014 - Applying NLP Tools to Similar Languages, Varieties and Dialects; and the LT4VarDial - Joint Workshop on Language Technology for Closely Related Languages, Varieties and Dialect @ RANLP BIBREF0 BIBREF1 . We can find also several works focused on the task. In BIBREF2 the authors addressed the problem of identifying Arabic varieties in blogs and social fora. They used character $n$ -gram features to discriminate between six different varieties and obtained accuracies between 70%-80%. Similarly, BIBREF3 collected 1,000 news articles of two varieties of Portuguese. They applied different features such as word and character $n$ -grams and reported accuracies over 90%. With respect to the Spanish language, BIBREF4 focused on varieties from Argentina, Chile, Colombia, Mexico and Spain in Twitter. They used meta-learning and combined four types of features: i) character $n$ -gram frequency profiles, ii) character $n$ -gram language models, iii) Lempel-Ziv-Welch compression and iv) syllable-based language models. They obtained an interesting 60%-70% accuracy of classification.
We are interested in discovering which kind of features capture higher differences among varieties. Our hypothesis is that language varieties differ mainly in lexicographic clues. We show an example in Table 1 .
In this work we focus on the Spanish language variety identification. We differentiate from the previous works as follows: i) instead of $n$ -gram based representations, we propose a low dimensionality representation that is helpful when dealing with big data in social media; ii) in order to reduce the possible over-fitting, our training and test partitions do not share any author of instance between them; and iii) in contrast to the Twitter dataset of BIBREF4 , we will make available our dataset to the research community.
We differentiate from the previous works as follows: i) instead of $n$ -gram based representations, we propose a low dimensionality representation that is helpful when dealing with big data in social media; ii) in order to reduce the possible over-fitting, our training and test partitions do not share any author of instance between them; and iii) in contrast to the Twitter dataset of BIBREF4 , we will make available our dataset to the research community.
|
What is the use of the low dimensionality representation?
|
It is helpful when dealing with big data in social media.
|
null | false
| null |
What are the best vegetables for new gardeners to try growing?
|
Vegetables that have easy germination requirements and are low maintenance are the best for new gardeners to try to grow. Easy germination requirements include low to average soil temperature, low to average sun requirements, and faster germination timelines. Beans, peas, pumpkins, and squash are all some of the easiest vegetables to germinate from seeds. Low maintenance vegetables, once sprouted, require little care, including minimal weeding, pest control, and watering requirements. Carrots, radishes, and onions are great examples of low care vegetables to grow.
|
|
null | false
| null |
Provide a bulleted list of things to think about when buying a house
|
The following are things to think about when buying a house:
1. Location
2. Purchase price and mortgage rates
3. Floor Plan / Layout
4. Design / Finishings
5. Garage space
6. Storage space
7. How long one plans to live there
8. Property taxes
9. Insurance rates
10. Maintenance costs
|
|
null | false
| null |
Make me a workout playlist for the gym
|
- Till I Collapse by Eminem and Nate Dogg
- Sicko Mode by Travis Scott
- Humble by Kendrick Lamar
- Body by Loud Luxury
- Gecko by Oliver Heldens
- Mercy by Kanye West
- Badboy Luxury by thmpsn
- Headstrong by Trapt
- One Step Closer by Lilnkin Park
- Down with the Sickness by Disturbed
- Killing in the Name by Rage Against the Machine
- Lose Yourself by Eminem
- Duality by Slipknot
- Bangarang by Skrillex
- Stricken by Disturbed
- Pursuit of Happiness by Kid Cudi
- Butterfly Effect by Travis Scott
- Right Above It by Lil Wayne
- Trophies by Drake
- Super Gremlin by Kodak Black
- Super Hero by Metro Boomin
- Jimmy Cooks by Drake and 21 Savage
|
|
null | false
| null |
Cassandra introduced the Cassandra Query Language (CQL). CQL is a simple interface for accessing Cassandra, as an alternative to the traditional Structured Query Language (SQL). CQL adds an abstraction layer that hides implementation details of this structure and provides native syntaxes for collections and other common encodings. Language drivers are available for Java (JDBC), Python (DBAPI2), Node.JS (Datastax), Go (gocql) and C++.
The keyspace in Cassandra is a namespace that defines data replication across nodes. Therefore, replication is defined at the keyspace level. Below an example of keyspace creation, including a column family in CQL 3.0:
|
From the passage provided, extract list of language drivers for Cassandra. Separate them with a comma.
|
Java (JDBC), Python (DBAPI2), Node.JS (Datastax), Go (gocql), C++
|
null | false
| 24
|
Reading Comprehension (RC) has become a central task in natural language processing, with great practical value in various industries. In recent years, many large-scale RC datasets in English BIBREF0, BIBREF1, BIBREF2, BIBREF3, BIBREF4, BIBREF5, BIBREF6 have nourished the development of numerous powerful and diverse RC models BIBREF7, BIBREF8, BIBREF9, BIBREF10, BIBREF11. The state-of-the-art model BIBREF12 on SQuAD, one of the most widely used RC benchmarks, even surpasses human-level performance. Nonetheless, RC on languages other than English has been limited due to the absence of sufficient training data. Although some efforts have been made to create RC datasets for Chinese BIBREF13, BIBREF14 and Korean BIBREF15, it is not feasible to collect RC datasets for every language since annotation efforts to collect a new RC dataset are often far from trivial. Therefore, the setup of transfer learning, especially zero-shot learning, is of extraordinary importance.
Existing methods BIBREF16 of cross-lingual transfer learning on RC datasets often count on machine translation (MT) to translate data from source language into target language, or vice versa. These methods may not require a well-annotated RC dataset for the target language, whereas a high-quality MT model is needed as a trade-off, which might not be available when it comes to low-resource languages.
In this paper, we leverage pre-trained multilingual language representation, for example, BERT learned from multilingual un-annotated sentences (multi-BERT), in cross-lingual zero-shot RC. We fine-tune multi-BERT on the training set in source language, then test the model in target language, with a number of combinations of source-target language pair to explore the cross-lingual ability of multi-BERT. Surprisingly, we find that the models have the ability to transfer between low lexical similarity language pair, such as English and Chinese. Recent studies BIBREF17, BIBREF12, BIBREF18 show that cross-lingual language models have the ability to enable preliminary zero-shot transfer on simple natural language understanding tasks, but zero-shot transfer of RC has not been studied. To our knowledge, this is the first work systematically exploring the cross-lingual transferring ability of multi-BERT on RC tasks.
Surprisingly, we find that the models have the ability to transfer between low lexical similarity language pair, such as English and Chinese.
|
What kind of ability do they find that the models have?
|
The ability to transfer between low lexical similarity language pair, such as English and Chinese.
|
null | false
| null |
I which country is the best football played?
|
The best football is either played in Spain or Italy since they have the most skilled players and coaches. A reason for that is that investors have put a lot of money into these teams. However, football leagues like Germany, England, and Brazil are very good.
|
|
null | false
| null |
What is one of the main trigger for autism in current generation?
|
The research has found that exposing infants or kids less than Two years of age to 2 hours of screen time every day can trigger autism and other significant developmental delays. To avoid environmental factors which can trigger autism, it is suggested for parents or care givers to interact with kids and socialize them instead of exposing them to screen time. The research also shows the autism rates are high in nuclear family where the usage of gadgets or multimedia is high.
|
|
null | false
| 391
|
The overall architecture of DFN is depicted by Figure 3 . The input is a question $Q$ in length $l_q$ , a passage $P$ in length $l_p$ , and a list of $r$ answer candidates $\mathcal {A}=\lbrace A_1,...,A_r\rbrace $ in length $l_{a}^1,...,l_a^r$ . The model produces scores $c_1, c_2, ..., c_r$ for each answer candidate $A_1, A_2, ..., A_r$ respectively. The final prediction module selects the answer with the highest score.
The architecture consists of a standard Lexicon Encoding Layer and a Context Encoding Layer, on top of which are a Dynamic Fusion Layer and a Memory Generation Layer. The Dynamic Fusion Layer applies different attention strategies to different question types, and the Memory Generation Layer encodes question-related information in the passage for answer prediction. Multi-step reasoning is conducted over the output from the Dynamic Fusion and Memory Generation layers, in the Answer Scoring Module. The final output of the model is an answer choice $C\in \lbrace 1,2,...,r\rbrace $ from the Answer Prediction Module.
In the following subsections, we will describe the details of each component in DFN (bold letters represent trainable parameters).
Figure 3: Architecture of DFN. For simplicity, we only draw DFN for one answer candidate A. i) Passage, question and answer candidates are independently mapped through word and character encodings in the Lexicon Encoding Layer. ii) The independent encodings are then fed into a BiLSTM in the Context Encoding Layer. iii) The Dynamic Fusion Layer takes a customized attention strategy across the three representations of passage, question and answer candidates. iv) Memory Generation Layer generates a working memory. v) The Answer Scoring Module reads in the memory for a dynamic number of steps. vi) Answer prediction module generates the final output.
|
How does the DFN work?
|
i) Passage, question and answer candidates are independently mapped through word and character encodings in the Lexicon Encoding Layer. ii) The independent encodings are then fed into a BiLSTM in the Context Encoding Layer. iii) The Dynamic Fusion Layer takes a customized attention strategy across the three representations of passage, question and answer candidates. iv) Memory Generation Layer generates a working memory. v) The Answer Scoring Module reads in the memory for a dynamic number of steps. vi) Answer prediction module generates the final output.
|
null | false
| null |
Which is an imperial or a metric measurement: pint, millilitres
|
pint is imperial, millilitres is metric
|
|
null | false
| null |
What is intelligence?
|
The term intelligence is an open term and is often open to interpretation and relativity. Some people claim you can test for intelligence using an intelligence quotient (IQ) test. This is up for debate. Knowledge alone does not equal intelligence. The proper use of knowledge can be referred to as intelligence. A definition of intelligence is: the ability to acquire and apply knowledge and skills. According to this, nearly every has the ability of intelligence.
|
|
null | false
| null |
How to find the area of a rectangle?
|
A rectangle is a closed two dimensional quadrilateral structure with four right angles (90 degrees). Opposite sides of a rectangle are parallel and of equal length. Usually, the length of longer sides is called length of a rectangle and the length of shorter sides is called width/breadth of a rectangle.
Area of the rectangle is calculated by multiplying the length and width of the rectangle. For example, If the length of the rectangle is l and the width of the rectangle is w then area of the rectangle = (l x w).
If all sides of a rectangle are of equal length then it is called a square. Thus, square is a special case of a rectangle.
|
|
null | false
| null |
Visa Inc. (/ˈviːzə, ˈviːsə/; stylized as VISA) is an American multinational financial services corporation headquartered in San Francisco, California. It facilitates electronic funds transfers throughout the world, most commonly through Visa-branded credit cards, debit cards and prepaid cards. Visa is one of the world's most valuable companies.
Visa does not issue cards, extend credit or set rates and fees for consumers; rather, Visa provides financial institutions with Visa-branded payment products that they then use to offer credit, debit, prepaid and cash access programs to their customers. In 2015, the Nilson Report, a publication that tracks the credit card industry, found that Visa's global network (known as VisaNet) processed 100 billion transactions during 2014 with a total volume of US$6.8 trillion.
Visa was founded in 1958 by Bank of America (BofA) as the BankAmericard credit card program. In response to competitor Master Charge (now Mastercard), BofA began to license the BankAmericard program to other financial institutions in 1966. By 1970, BofA gave up direct control of the BankAmericard program, forming a cooperative with the other various BankAmericard issuer banks to take over its management. It was then renamed Visa in 1976.
Nearly all Visa transactions worldwide are processed through the company's directly operated VisaNet at one of four secure data centers, located in Ashburn, Virginia; Highlands Ranch, Colorado; London, England; and Singapore. These facilities are heavily secured against natural disasters, crime, and terrorism; can operate independently of each other and from external utilities if necessary; and can handle up to 30,000 simultaneous transactions and up to 100 billion computations every second.
Visa is the world's second-largest card payment organization (debit and credit cards combined), after being surpassed by China UnionPay in 2015, based on annual value of card payments transacted and number of issued cards. However, because UnionPay's size is based primarily on the size of its domestic market in China, Visa is still considered the dominant bankcard company in the rest of the world, where it commands a 50% market share of total card payments.
|
From the passage list down the location of Visa data centers. List the results in semi colon separated format.
|
Ashburn, Virginia; Highlands Ranch, Colorado; London, England; Singapore
|
null | false
| null |
What is backcountry skiing?
|
Backcountry skiing is skiing in the unmarked and ungroomed areas. In normal ski, people go to a ski resort and ski on the slopes marked and maintained by the ski resort. Backcountry ski may use ski lift, or other transportation methods to get to the top of the hill, but they don't follow the marked slope. Backcountry skiers usually bring special gear to avoid hazardous condition, such as avalanche or cliff.
|
|
null | false
| null |
Stanwell-Fletcher Lake is the largest lake on Somerset Island, the tenth-largest island of the Canadian Arctic Archipelago. The lake, along with most of Somerset Island, is located within the Qikiqtaaluk Region of the Canadian territory of Nunavut.
The harsh tundral climate and minimal soil cover of the highlands surrounding the lake allows only a few types of plants to grow there, mostly lichens, mosses, and arctic willow. Vegetation is more prominent in the thermokarst lowlands to the north of the lake, which are dominated by sedges during the summer. The low-lying areas to the southeast of the lake, and the nearby Creswell Bay, host a Canadian Important Bird Area. The area supports large bird populations during the summer, especially breeding populations of various shorebirds. 13 shorebird species have been observed in the area, while 11 are known to breed there. These species includes the white-rumped sandpiper, red phalarope, black-bellied plover, sanderling, American golden plover, ruddy turnstone, Baird's sandpiper, buff-breasted sandpiper, pectoral sandpiper, semipalmated sandpiper, and red knot. Other birds with populations in the area include the snow goose, king eider, long-tailed duck, northern fulmar, black-legged kittiwake, and peregrine falcon. Stanwell-Fletcher Lake is also a large and important habitat for anadromous populations of arctic char, which spend summer in coastal areas including Creswell Bay and surrounding waters, but migrate via the Union River to in the lake before it freezes over. Fourhorn sculpin are also present within the lake.
|
What are some examples of plants that grow around Stanwell-Fletcher Lake? Use a numbered list.
|
1. lichens
2. mosses
3. arctic willow
|
null | false
| null |
The Walt Disney World Resort, also called Walt Disney World or Disney World, is an entertainment resort complex in Bay Lake and Lake Buena Vista, Florida, United States, near the cities of Orlando and Kissimmee. Opened on October 1, 1971, the resort is operated by Disney Parks, Experiences and Products, a division of The Walt Disney Company. The property covers nearly 25,000 acres (39 sq mi; 101 km2), of which half has been used. The resort comprises four theme parks (Magic Kingdom, Epcot, Disney's Hollywood Studios, and Disney's Animal Kingdom), two water parks (Disney's Blizzard Beach and Disney's Typhoon Lagoon), 31 themed resort hotels, nine non-Disney hotels, several golf courses, a camping resort, and other entertainment venues, including the outdoor shopping center Disney Springs. On October 1, 2021, Walt Disney World started its celebration of its 50-year anniversary which lasted for 18 consecutive months ending on March 31, 2023.
|
Given this paragraph, what is the size in acres of Walt Disney World?
|
According to this text, Walt Disney World is 12,500 acres and sits on 25,000 acres of property owned by the company.
|
null | false
| 461
|
Consider the same problem setup as § 3, where we observe a small dataset of samples from the target channel distribution. Suppose we have adapted the MDN channel by optimizing over the parameters ψ, we can use the inverse-affine feature transformations (defined in Eq. ()) to transform the channel output y from a component of the target Gaussian mixture to the same component of the source Gaussian mixture. However, this transformation requires knowledge of both the channel input x and the mixture component i, which are not observed (latent) at the decoder. We address this by first determining the most-probable pair of channel input and mixture component for a given y (using the MAP rule), and applying the corresponding inverse-affine feature transformation as
The joint posterior over the channel input x and mixture component i, given the channel output y is based on the adapted (target) Gaussian mixture, given by
.
The adapted decoder based on the above affine feature transformation (see Fig.) is defined as In the previous method, an input transformation layer is introduced at the decoder only during adaptation, but not during training of the autoencoder. Alternatively, we propose an input transformation layer at the decoder that takes the channel output y and produces a best estimate of the encoded symbol x, which is then given as input to the decoder as shown in Fig.. This input transformation layer is included during the autoencoder training as a fixed non-linear transformation that does not have any trainable parameters. Since the decoder is trained to predict using x instead of y, it is inherently robust to changes in the distribution of y.
Given a generative model of the channel conditional density using Gaussian mixtures, we can estimate the plug-in Bayes posterior distribution of x given y, P θc (x | y) (ref.
Eq. ()). From this, we can find the MAP estimate of x given y as
The adapted decoder based on this input transformation, referred to as the MAP symbol estimation (SE) layer, is defined as
and illustrated in Fig.. Whenever the MDN model is adapted to changes in the channel distribution, resulting in a new MDN with parameters θ c , the MAP SE layer is also updated using θ c . This input transformation shields the decoder from changes to the distribution of the channel output y.
Since the MAP SE layer is also included in the autoencoder during training, the non-differentiable argmax function presents an obstacle to training the autoencoder using backpropagation. We address this by using a temperature-scaled softmax approximation to the argmax, which is differentiable and provides a close approximation for small temperature values. This approximation is used only during training, whereas the exact argmax is used during inference. Details on this approximation, and a modified autoencoder training algorithm with temperature annealing are discussed in Appendix C.5.
Comments. The proposed input transformation methods at the decoder have some similarities to equalization methods used in communication receivers. However, our problem setting considers a memoryless channel, and does not deal with intersymbol interference (ISI), which is the main focus of equalization methods. A key advantage of the proposed adaptation is that it is very computationally efficient to implement at the receiver of a communication system. A discussion of the computational complexity of the proposed methods is given in Appendix C.4.
Suppose we have adapted the MDN channel by optimizing over the parameters ψ, we can use the inverse-affine feature transformations (defined in Eq. (4)) to transform the channel output y from a component of the target Gaussian mixture to the same component of the source Gaussian mixture.
|
Is the adaptation method applicable to other e2e auto encoder models, e.g., GAN or VAE. In another words, is there any specific advantage of selecting MDN?
|
Just to clarify, the choice of a GAN, VAE, or MDN pertains only to the generative model of the channel, and not the end-to-end autoencoder. The proposed method is specific to a mixture density network channel model. Adaptation of other generative models such as GAN and VAE using a small number of samples is an interesting but challenging problem. The choice of MDN as the channel model for this problem comes with the following advantages:
The conditional density corresponding to the MDN is a mixture of Gaussians, whose parametric form is easy to adapt using symbol and component-conditional affine transformations. The number of affine transformation parameters is much smaller than the number of MDN parameters, making the adaptation less susceptible to overfitting on the small dataset from the target distribution.
The MDN choice makes it possible to derive a closed-form KL-divergence between the source and target Gaussian mixtures under some mild assumptions. This is used as a regularization term in the adaptation objective, which allows to directly constrain the change in distribution.
The affine transformation property between the mixture of Gaussians allows us to design an input transformation method at the decoder (Sec 4.1) that can compensate for changes in the distribution of the channel output. This same property also allows us to propose a validation metric that is used for selecting the hyper-parameter lambda (Appendix C.3).
|
null | false
| null |
How can I have a productive day?
|
Here are some ideas. Start with making a list of some things you would like to accomplish or do to make the day a meaningful success in your eyes. Prioritize those activities from most important to least important, or by time of day you would like to have them completed. Then open your planner or calendar and fill those activities into the appropriate time slot. Then check off those items as you complete them or participate in those activities to give yourself a sense of accomplishment. At the end of the day, you could journal to reflect on how the day went, what you could have done differently or approved, and plan the next day with those lessons in mind.
|
|
null | false
| null |
"Ghost Town" is a song recorded by American singer Adam Lambert for his third studio album, The Original High (2015). It was released as the album's lead single on April 21, 2015. The single garnered considerable commercial and critical success, becoming somewhat of a sleeper hit in select countries and receiving a substantial amount of radio play in the US.
|
Who recorded the song Ghost Town
|
Ghost Town is a song recorded by American singer Adam Lambert for his third studio album, The Original High (2015). It was released as the album's lead single on April 21, 2015. The single garnered considerable commercial and critical success, becoming somewhat of a sleeper hit in select countries and receiving a substantial amount of radio play in the US.
|
null | false
| null |
Why is Chicago so windy?
|
A lot of Chicago's wind comes from the air above nearby Lake Michigan. The sun heats up the land faster than it heats up the water in the lake. When the air in the city rises, the cooler air over the lake moves toward the land. The air movement makes a lot of wind.
|
|
null | false
| null |
Rosslyn Mountain Boys was a band that backed up a number of nationally known singers, and released two albums of their own.
|
If Rosslyn Mountain Boys had produced two more albums, how many albums would they have produced?
|
Rosslyn Mountain Boys produced two albums on their own, so if they produced two more albums then they would have produced four albums.
|
null | false
| null |
What are the benefits of living in Australia
|
Australia has many advantages as a country to live in. There are four major benefits, 1) the weather is warm and temperate and although there are seasons they are not extreme, 2) The country is politically stable, has had peaceful transfers of power and also stable finance sector with other countries willing to purcahse Australian government bonds, and 3) Significant growth has put the country in a strong position with GDP relative to other countries. The strength of the mining and agricultural sectors has driven the economy over the last decade and the consistency of the growth has lead Australia to being called the lucky country.
|
|
null | false
| 314
|
In information retrieval (IR), queries and documents are typically represented by term vectors where each term is a content word and weighted by tf-idf, i.e. the product of the term frequency and the inverse document frequency, or other weighting schemes BIBREF0 . The similarity of a query and a document is then determined as a dot product or cosine similarity. Although this works reasonably, the traditional IR scheme often fails to find relevant documents when synonymous or polysemous words are used in a dataset, e.g. a document including only “neoplasm" cannot be found when the word “cancer" is used in a query. One solution of this problem is to use query expansion BIBREF1 , BIBREF2 , BIBREF3 , BIBREF4 or dictionaries, but these alternatives still depend on the same philosophy, i.e. queries and documents should share exactly the same words.
While the term vector model computes similarities in a sparse and high-dimensional space, the semantic analysis methods such as latent semantic analysis (LSA) BIBREF5 , BIBREF6 and latent Dirichlet allocation (LDA) BIBREF7 learn dense vector representations in a low-dimensional space. These methods choose a vector embedding for each term and estimate a similarity between terms by taking an inner product of their corresponding embeddings BIBREF8 . Since the similarity is calculated in a latent (semantic) space based on context, the semantic analysis approaches do not require having common words between a query and documents. However, it has been shown that LSA and LDA methods do not produce superior results in various IR tasks BIBREF9 , BIBREF10 , BIBREF11 and the classic ranking method, BM25 BIBREF12 , usually outperforms those methods in document ranking BIBREF13 , BIBREF14 .
Neural word embedding BIBREF15 , BIBREF16 is similar to the semantic analysis methods described above. It learns low-dimensional word vectors from text, but while LSA and LDA utilize co-occurrences of words, neural word embedding learns word vectors to predict context words BIBREF10 . Moreover, training of semantic vectors is derived from neural networks. Both co-occurrence and neural word embedding approaches have been used for lexical semantic tasks such as semantic relatedness (e.g. king and queen), synonym detection (e.g. cancer and carcinoma) and concept categorization (e.g. banana and pineapple belong to fruits) BIBREF10 , BIBREF17 . But, Baroni et al. Baroni2014 showed that neural word embedding approaches generally performed better on such tasks with less effort required for parameter optimization. The neural word embedding models have also gained popularity in recent years due to their high performance in NLP tasks BIBREF18 .
Here we present a query-document similarity measure using a neural word embedding approach. This work is particularly motivated by the Word Mover's Distance BIBREF19 . Unlike the common similarity measure taking query/document centroids of word embeddings, the proposed method evaluates a distance between individual words from a query and a document. Our first experiment was performed on the TREC 2006 and 2007 Genomics benchmark sets BIBREF20 , BIBREF21 , and the experimental results showed that our approach was better than BM25 ranking. This was solely based on matching queries and documents by the semantic measure and no other feature was used for ranking documents.
In general, conventional ranking models (e.g. BM25) rely on a manually designed ranking function and require heuristic optimization for parameters BIBREF22 , BIBREF23 . In the age of information explosion, this one-size-fits-all solution is no longer adequate. For instance, it is well known that links to a web page are an important source of information in web document search BIBREF24 , hence using the link information as well as the relevance between a query and a document is crucial for better ranking. In this regard, learning to rank BIBREF22 has drawn much attention as a scheme to learn how to combine diverse features. Given feature vectors of documents and their relevance levels, a learning to rank approach learns an optimal way of weighting and combining multiple features.
We argue that the single scores (or features) produced by BM25 and our proposed semantic measure complement each other, thus merging these two has a synergistic effect. To confirm this, we measured the impact on document ranking by combining BM25 and semantic scores using the learning to rank approach, LamdaMART BIBREF25 , BIBREF26 . Trained on PubMed user queries and their click-through data, we evaluated the search performance based on the most highly ranked 20 documents. As a result, we found that using our semantic measure further improved the performance of BM25.
Taken together, we make the following important contributions in this work. First, to the best of our knowledge, this work represents the first investigation of query-document similarity for information retrieval using the recently proposed Word Mover's Distance. Second, we modify the original Word Mover's Distance algorithm so that it is computationally less expensive and thus more practical and scalable for real-world search scenarios (e.g. biomedical literature search). Third, we measure the actual impact of neural word embeddings in PubMed by utilizing user queries and relevance information derived from click-through data. Finally, on TREC and PubMed datasets, our proposed method achieves stronger performance than BM25.
Here we present a query-document similarity measure using a neural word embedding approach.
|
What method is used for evaluating the distance between individual words from a query and a document?
|
It is a neural word embedding approach.
|
null | false
| null |
Sort the following volumes in ascending order: 300 imperial gallons, 2000 imperial pints, 500 litres, 1 cubic meter, 10000 imperial ounces, 1500 imperial quarts.
|
10000 imperial ounce (~284 litres)
500 litres
1 cubic meter (1000 litres)
2000 imperial pints (~1136 litres)
300 imperial gallons (~1364 litres)
1500 imperial quarts (~1705 litres)
Conversions used:
An imperial gallon equals 4.54609 litres
An imperial pint equals 0.568 litres
A cubic meter equals 1000 litres
An imperial ounce equals 0.02841306 litres
An imperial quart equals 1.13652 litres
|
|
null | false
| null |
Tell me whether these are songs or video games: Madden, Fifa, Gimme Shelter, Running on Empty, New York New York, Call of Duty Modern Warfare, Fortnite, Sunday Morning, Mario Kart, Wonderwall
|
Video Games: Madden, Fifa, Call of Duty Modern Warfare, Fortnite, Mario Kart
Songs: Gimme Shelter, Running on Empty, New York New York, Sunday Morning, Wonderwall
|
|
null | false
| 101
|
This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/
Entity linking (EL), mapping entity mentions in texts to a given knowledge base (KB), serves as a fundamental role in many fields, such as question answering BIBREF0 , semantic search BIBREF1 , and information extraction BIBREF2 , BIBREF3 . However, this task is non-trivial because entity mentions are usually ambiguous. As shown in Figure FIGREF1 , the mention England refers to three entities in KB, and an entity linking system should be capable of identifying the correct entity as England cricket team rather than England and England national football team.
Entity linking is typically broken down into two main phases: (i) candidate generation obtains a set of referent entities in KB for each mention, and (ii) named entity disambiguation selects the possible candidate entity by solving a ranking problem. The key challenge lies in the ranking model that computes the relevance between candidates and the corresponding mentions based on the information both in texts and KBs BIBREF4 . In terms of the features used for ranking, we classify existing EL models into two groups: local models to resolve mentions independently relying on textual context information from the surrounding words BIBREF5 , BIBREF6 , BIBREF7 , BIBREF8 , and global (collective) models, which are the main focus of this paper, that encourage the target entities of all mentions in a document to be topically coherent BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 , BIBREF13 , BIBREF14 .
Global models usually build an entity graph based on KBs to capture coherent entities for all identified mentions in a document, where the nodes are entities, and edges denote their relations. The graph provides highly discriminative semantic signals (e.g., entity relatedness) that are unavailable to local model BIBREF15 . For example (Figure FIGREF1 ), an EL model seemly cannot find sufficient disambiguation clues for the mention England from its surrounding words, unless it utilizes the coherence information of consistent topic “cricket" among adjacent mentions England, Hussain, and Essex. Although the global model has achieved significant improvements, its limitation is threefold:
To mitigate the first limitation, recent EL studies introduce neural network (NN) models due to its amazing feature abstraction and generalization ability. In such models, words/entities are represented by low dimensional vectors in a continuous space, and features for mention as well as candidate entities are automatically learned from data BIBREF4 . However, existing NN-based methods for EL are either local models BIBREF16 , BIBREF17 or merely use word/entity embeddings for feature extraction and rely on another modules for collective disambiguation, which thus cannot fully utilize the power of NN models for collective EL BIBREF18 , BIBREF19 , BIBREF20 .
The second drawback of the global approach has been alleviated through approximate optimization techniques, such as PageRank/random walks BIBREF21 , graph pruning BIBREF22 , ranking SVMs BIBREF23 , or loopy belief propagation (LBP) BIBREF18 , BIBREF24 . However, these methods are not differentiable and thus difficult to be integrated into neural network models (the solution for the first limitation).
To overcome the third issue of inadequate training data, BIBREF17 has explored a massive amount of hyperlinks in Wikipedia, but these potential annotations for EL contain much noise, which may distract a naive disambiguation model BIBREF6 .
In this paper, we propose a novel Neural Collective Entity Linking model (NCEL), which performs global EL combining deep neural networks with Graph Convolutional Network (GCN) BIBREF25 , BIBREF26 that allows flexible encoding of entity graphs. It integrates both local contextual information and global interdependence of mentions in a document, and is efficiently trainable in an end-to-end fashion. Particularly, we introduce attention mechanism to robustly model local contextual information by selecting informative words and filtering out the noise. On the other hand, we apply GCNs to improve discriminative signals of candidate entities by exploiting the rich structure underlying the correct entities. To alleviate the global computations, we propose to convolute on the subgraph of adjacent mentions. Thus, the overall coherence shall be achieved in a chain-like way via a sliding window over the document. To the best of our knowledge, this is the first effort to develop a unified model for neural collective entity linking.
In experiments, we first verify the efficiency of NCEL via theoretically comparing its time complexity with other collective alternatives. Afterwards, we train our neural model using collected Wikipedia hyperlinks instead of dataset-specific annotations, and perform evaluations on five public available benchmarks. The results show that NCEL consistently outperforms various baselines with a favorable generalization ability. Finally, we further present the performance on a challenging dataset WW BIBREF19 as well as qualitative results, investigating the effectiveness of each key module.
However, existing NN-based methods for EL are either local models (Yamada et al., 2017; Gupta et al., 2017) or merely use word/entity embeddings for feature extraction and rely on another modules for collective disambiguation, which thus cannot fully utilize the power of NN models for collective EL (Globerson et al., 2016; Guo and Barbosa, 2017; Phan et al., 2018).
|
What is the limitation of existing NN-based methods？
|
They cannot fully utilize the power of NN models for collective EL.
|
null | false
| null |
What are the measurements needed to fit pants?
|
The main three measurements for pants are the waist, the length, and the inseam. There any many other measurements associated with pant fittings, including hip circumference, thigh circumference, knee circumference, and calf circumference.
|
|
null | false
| 159
|
The generator performance was assessed on the two evaluation metrics: the BLEU and the slot error rate ERR by adopting code from an open source benchmark toolkit for Natural Language Generation. We compared the proposed models against three strong baselines which have been recently published as state-of-the-art NLG benchmarks[]. https://github.com/shawnwun/RNNLG
HLSTM proposed by BIBREF3 thwsjy15 which used a heuristic gate to ensure that all of the slot-value information was accurately captured when generating.
SCLSTM proposed by BIBREF4 wensclstm15 which can jointly learn the gating signal and language model.
Enc-Dec proposed by BIBREF8 wentoward which applied the attention-based encoder-decoder architecture.
The generator performance was assessed on the two evaluation metrics: the BLEU and the slot error rate ERR by adopting code from an open source benchmark toolkit for Natural Language Generation.
|
What metrics are used to evaluate the generator performance?
|
The generator performance was assessed on the two evaluation metrics: the BLEU and the slot error rate ERR.
|
null | false
| 465
|
We design a simple non-learning-based goal set generator by exploiting the KG-based observation in the cooking theme. Algorithm 2 shows the pipeline for obtaining the goal set G_t. We first obtain the ingredient set I. For each ingredient i ∈ I, we first check whether it has been collected, then obtain its status set Si and requirement set R_i . We consider three types of goals: 1) “find” requires the agent to find and collect an uncollected ingredient, 2) “prepare” requires the agent to prepare an ingredient to satisfy a requirement, and 3) “eat”, that the agent is required to prepare and eat the final meal. Algorithm 3 shows the pipeline for assigning the goal-conditioned reward r^goal_t . We first obtain the type of a goal g, then check whether this goal has been accomplished given at and o^KG_(t+1). Some functions in Algorithm 2 can be reused here. r^goal_t is a binary reward that we will assign r_goal t = r_max if g is accomplished successfully, otherwise r_min (still not finished, or failed). Algorithm 2 and Algorithm 3 can also be implemented via learning-based methods. For example, the functions can be achieved by a QA model by answering specific questions.
|
How is the pseudo-reward constructed for each goal - in particular the agent must know when the goal has terminated and whether the termination was successful or a failure. How does the agent determine a successful criterion for each goal?
|
We provide more details in in Appendix B in our paper revision and the pipeline is as follows: For assigning the goal-conditioned reward r^goal at t, we first obtain the type of a goal g, then check whether this goal has been accomplished given at and o^KG at t+1. Some functions in Algorithm 2 can be reused here. r^goal at t is a binary reward that we will assign r^goal_t = r_max if g is accomplished successfully, otherwise r_min (failed).
|
null | false
| null |
What is counterintelligence?
|
Counterintelligence is an activity aimed at protecting an agency's intelligence program from an opposition's intelligence service. It includes gathering information and conducting activities to prevent espionage, sabotage, assassinations or other intelligence activities conducted by, for, or on behalf of foreign powers, organizations or persons.
|
|
null | false
| 219
|
We compared our full method (TextTiling with heuristic-max based on embeddings trained by virtual sentences) with several baselines:
Random. We randomly segmented conversation sessions. In this baseline, we were equipped with the prior probability of segmentation.
MMD. We applied the MinMax-Dotplotting (MMD) approach proposed by Ye et al. BIBREF24 . We ran the executable program provided by the authors.
TextTiling w/ tf INLINEFORM0 idf features. We implemented TextTiling ourselves according to BIBREF12 .
We tuned the hyperparameter INLINEFORM0 in Equation ()on the validation set to make the number of segmentation close to that of manual annotation, and reported precision, recall, and the F-score on the test set in Table TABREF18 . As seen, our approach significantly outperforms baselines by a large margin in terms of both precision and recall. Besides, we can see that MMD obtains low performance, which is mainly because the approach cannot be easily adapted to other datasets like short sentences of conversation utterances. In summary, we achieve an INLINEFORM1 -score higher than baseline methods by more than 20%, showing the effectiveness of enhancing TextTiling with modern word embeddings.
We further conducted in-depth analysis of different strategies of training word-embeddings and matching heuristics in Table TABREF21 . For word embeddings, we trained them on the 3M-sentence dataset with three strategies: (1) virtual-sentence context proposed in our paper; (2) within-sentence context, where all words (except the current one) within a sentence (either a query or reply) are regarded as the context; (3) window-based context, which is the original form of BIBREF25 : the context is the words in a window (previous 2 words and future 2 words in the sentence). We observe that our virtual-sentence strategy consistently outperforms the other two in all three matching heuristics. The results suggest that combining a query and a reply does provide more information in learning dialogue-specific word embeddings.
Regarding matching heuristics, we find that in the second and third strategies of training word embeddings, the complicated heuristic-max method yields higher INLINEFORM0 -scores than simple sum pooling by 2–3%. However, for the virtual-sentence strategy, heuristic-max is slightly worse than the sum pooling. (The degradation is only 0.1% and not significant.) This is probably because both heuristic-max and virtual sentences emphasize the rich interaction between a query and its corresponding reply; combining them does not result in further gain.
We also notice that heuristic-avg is worse than other similarity measures. As this method is mathematically equivalent to the average of word-by-word similarity, it may have an undesirable blurring effect.
To sum up, our experiments show that both the proposed embedding learning approach and the similarity heuristic are effective for session segmentation. The embedding-enhanced TextTiling approach largely outperforms baselines.
We conducted an external experiment to show the effect of session segmentation in dialogue systems. We integrated the segmentation mechanism into a state-of-the-practice retrieval-based system and evaluated the results by manual annotation, similar to our previous work BIBREF27 , BIBREF31 , BIBREF32 .
Concretely, we compared our session segmentation with fixed-length context, used in BIBREF11 . That is to say, the competing method always regards two previous utterances as context. We hired three workers to annotate the results with three integer scores (0–2 points, indicating bad, borderline, and good replies, respectively.) We sampled 30 queries from the test set of 100 sessions. For each query, we retrieved 10 candidates and computed p@1 and nDCG scores BIBREF33 (averaged over three annotators). Provided with previous utterances as context, each worker had up to 1000 sentences to read during annotation.
Table TABREF26 presents the results of the dialogue system with session segmentation. As demonstrated, our method outperforms the simple fixed-context approach in terms of both metrics. We computed the inner-annotator agreement: std INLINEFORM0 0.309; 3-discrete-class Fleiss' kappa score INLINEFORM1 0.411, indicating moderate agreement BIBREF34 .
Case Study. We present a case study on our website: https://sites.google.com/site/sessionsegmentation/. From the case study, we see that the proposed approach is able to segment the dialogue session appropriately, so as to better utilize background information from a conversation session.
In this paper, we addressed the problem of session segmentation for open-domain dialogue systems. We proposed an embedding-enhanced TextTiling approach, where we trained embeddings with the novel notion of virtual sentences; we also proposed several heuristics for similarity measure. Experimental results show that both our embedding learning and similarity measuring are effective in session segmentation, and that with our approach, we can improve the performance of a retrieval-based dialogue system.
We thank anonymous reviewers for useful comments and Jingbo Zhu for sharing the MMD executable program. This paper is partially supported by the National Natural Science Foundation of China (NSFC Grant Nos. 61272343 and 61472006), the Doctoral Program of Higher Education of China (Grant No. 20130001110032), and the National Basic Research Program (973 Program No. 2014CB340405).
Besides, we can see that MMD obtains low performance, which is mainly because the approach cannot be easily adapted to other datasets like short sentences of conversation utterances. In summary, we achieve an F-score higher than baseline methods by more than 20%, showing the effectiveness of enhancing TextTiling with modern word embeddings
|
How much is their method higher than baseline methods by on F-score?
|
More than 20%.
|
null | false
| null |
Identify all PII in this customer review "My name is Joe Smith and email address is joesmith9999@gmail.com. I'd like to leave a review about Langton's red wine. It was tasty for a red, but there were sediments at the bottom of my bottle"
|
I've found two pieces of personally identifiable information (PII) in the provided text:
- A person's name: "Joe Smith"
- An email address: "joesmith9999@gmail.com"
The Australian government defines PII as "Information that can be used on its own or with other information to identify, contact or locate a single person, or to identify an individual in context" (source: https://www.cyber.gov.au/acsc/view-all-content/glossary/personally-identifiable-information-pii)
|
|
null | false
| null |
All rivers in Serbia belong to the drainage basins of three seas: Black Sea, Adriatic Sea or Aegean Sea.
The largest in area, Black Sea drainage basin, covers an area of 81,261 km2 or 92% of the territory of Serbia. The entire basin is drained by only one river, the Danube, which flows into the Black Sea. All major rivers in Serbia, like Tisa, Sava, Velika Morava and Drina belong to it.
The Adriatic Sea drainage basin covers an area of 4,500 km2 or 5% of territory of Serbia. It comprises the western half of the Kosovo and Metohija and it is mostly drained by one river, the White Drin, which in Albania meets the Black Drin to create the Drin river, which flows into the Adriatic Sea. Smaller portion of it is drained by Crni Kamen-Radika river in the extreme southern region of Gora, which also drains into Black Drin in North Macedonia.
The Aegean Sea basin is the smallest in area, 2,650 km2 or 3% of the territory of Serbia, and covers the southern parts of Serbia, to the Macedonian and Bulgarian borders. It is drained by three rivers: Lepenac, Pčinja and Dragovištica. The first two flow into the Vardar river in North Macedonia, and the third flows into the Struma river in Bulgaria, and both of those rivers flow into the Aegean Sea.
All three basins meet at the Drmanska glava peak on the Crnoljeva mountain in central Kosovo, which represents the water divide of Serbia and the major one in the Balkans, as it divides three out of the Balkan's four drainage basins (the fourth being the Ionian Sea).
|
What are major rivers in Serbia
|
All rivers in Serbia belong to the drainage basins of three seas: Black Sea, Adriatic Sea or Aegean Sea.
The largest in area, Black Sea drainage basin, covers an area of 81,261 km2 or 92% of the territory of Serbia. The entire basin is drained by only one river, the Danube, which flows into the Black Sea. All major rivers in Serbia, like Tisa, Sava, Velika Morava and Drina belong to it.
The Adriatic Sea drainage basin covers an area of 4,500 km2 or 5% of territory of Serbia. It comprises the western half of the Kosovo and Metohija and it is mostly drained by one river, the White Drin, which in Albania meets the Black Drin to create the Drin river, which flows into the Adriatic Sea. Smaller portion of it is drained by Crni Kamen-Radika river in the extreme southern region of Gora, which also drains into Black Drin in North Macedonia.
The Aegean Sea basin is the smallest in area, 2,650 km2 or 3% of the territory of Serbia, and covers the southern parts of Serbia, to the Macedonian and Bulgarian borders. It is drained by three rivers: Lepenac, Pcinja and Dragovistica. The first two flow into the Vardar river in North Macedonia, and the third flows into the Struma river in Bulgaria, and both of those rivers flow into the Aegean Sea.
All three basins meet at the Drmanska glava peak on the Crnoljeva mountain in central Kosovo, which represents the water divide of Serbia and the major one in the Balkans, as it divides three out of the Balkan's four drainage basins (the fourth being the Ionian Sea).
|
null | false
| 486
|
Affordance-aware navigation is a major challenge in solving complex and long-horizon indoor tasks such as ALFRED with both navigation and object interactions. Specifically, given each object of interest in the scene, the agent is required to not only find and approach it but also end up at a pose (x, y, r, h), that is feasible for subsequent interactions with the object. For instance, to open a fridge, the robot should approach the fridge closely enough (so the door is within reach), look at it (so that the fridge is in the field of view), and leave enough room to open the door. To solve a long-horizon task involving multiple navigation and object interaction subgoals, it is natural to use an explicit semantic map, either 2D or 3D, of the environment (similar to Neural Active SLAM Chaplot et al. (2020a)), together with model-based planning (e.g. as in). This line of work tends to generalize better than models that directly learn mappings from human instructions to navigation & interaction actions (e.g., E.T.). With perfect knowledge of the environment, it is possible to achieve (nearly) perfect performance. In practice, however, the semantic map acquired at inference time is usually far from ideal, primarily due to Incompleteness (missing information due to insufficient exploration of the scene) and Inaccuracy (erroneous object location prediction on the map, especially for small objects).
To improve exploration performance, we propose a multimodal module that, at each step, predicts an exploration action a ∈ {MoveAhead, RotateLeft, RotateRight} by taking visual observations & actions in the past, step-by-step language instructions, and the explored area map which indicates where the agent has visited. We show that, compared to existing model-based approaches on ALFRED (e.g., which applies random exploration), our use of lowlevel language instructions leads to more efficient exploration. The proposed exploration module operates at the subgoal level and only predicts exploration actions (in contrast to E.T. which directly predicts actions for the entire task). The extra modality (the explored area) facilitates exploration by providing the agent with explicit spatial information. We illustrate the exploration module in Figure, elaborate its details in Section 4.3, and empirically demonstrate its advantages in Section 5.
The proposed exploration module operates at the subgoal level and only predicts exploration actions (in contrast to E.T. which directly predicts actions for the entire task).****The multimodal exploration module consists of several sub-modules, either learned or pretrained/fixed. At a high level, given a task, the exploration go through its navigation subgoals one by one, with the subgoal parser (see Section 4.4) predicting whether a subgoal is for navigation.
|
In sec 4, what is the meaning of module operating at the subgoal level compared to E.T.?
|
In contrast to E.T. which takes inputs for the entire task all at once and predicts the actions accordingly, our approach operates (like HLSM) by taking inputs for each subgoal and predicting the associated actions one by one.
|
null | false
| null |
What do you think about Lake Tahoe as a destination to rent a cabin either for summer or winter?
|
I have been to Lake Tahoe a couple of times, both in winter. It is a beautiful destination to rent a cabin, and during winter the snow, the mountains and the overall atmosphere is great as a destination to enjoy with friends and family. I have not been there during summer though, however I did hear that when the snow melts there is a whole other world of trails for hiking, biking and walking that are not visible during winter.
|
|
1805.00460
| false
| null |
We represent each instance of image, question, and user choice as a triplet consisting of image feature, question feature, and the label vector for the user's answer. In addition, collecting multiple choices from identical users enables us to represent any two instances by the same user as a pair of triplets, assuming source-target relation. With these pairs of triplets, we can train the system to predict a user's choice on a new image and a new question, given the same user's choice on the previous image and its associated question. User's choice $x_{ans_i}$ is represented as one-hot vector where the size of the vector is equal to the number of possible choices. We refer to the fused feature representation of this triplet consisting of image, question, and the user's choice as choice vector.
As discussed earlier, we attempt to reflect user's interest by asking questions that provide visual context. The foremost prerequisite for the interactive questions to perform that function is the possibility of various answers or interpretations. In other words, a question whose answer is so obvious that it can be answered in an identical way would not be valid as an interactive question. In order to make sure that each generated question allows for multiple possible answers, we internally utilize the VQA module. The question generated by the VQG module is passed on to VQA module, where the probability distribution $p_{ans}$ for all candidate answers $C$ is determined. If the most likely candidate $c_i=\max p_{ans}$ , where $c_i \in C$ , has a probability of being answer over a certain threshold $\alpha $ , then the question is considered to have a single obvious answer, and is thus considered ineligible. The next question generated by VQG is passed on to VQA to repeat the same process until the the following requirement is met:
In our experiments, we set $\alpha $ as 0.33. We also excluded the yes/no type of questions. Figure 4 illustrates an example of a question where the most likely answer had a probability distribution over the threshold (and is thus ineligible), and another question whose probability distribution over the candidate answers was more evenly distributed (and thus proceeds to narrative generation stage).
We represent each instance of image, question, and user choice as a triplet consisting of image feature, question feature, and the label vector for the user's answer. In addition, collecting multiple choices from identical users enables us to represent any two instances by the same user as a pair of triplets, assuming source-target relation. With these pairs of triplets, we can train the system to predict a user's choice on a new image and a new question, given the same user's choice on the previous image and its associated question.
As discussed earlier, we attempt to reflect user's interest by asking questions that provide visual context.
In other words, a question whose answer is so obvious that it can be answered in an identical way would not be valid as an interactive question.
If the most likely candidate $c_i=\max p_{ans}$ , where $c_i \in C$ , has a probability of being answer over a certain threshold $\alpha $ , then the question is considered to have a single obvious answer, and is thus considered ineligible.
In our experiments, we set $\alpha $ as 0.33. We also excluded the yes/no type of questions.
|
What are the features of used to customize target user interaction?
|
The answers are shown as follows:
* image feature
* question feature
* label vector for the user's answer
|
null | false
| null |
Classify the following into countries and cities:India, Mumbai, USA, Seattle, Germany, Berlin, United Kingdom, Australia, Tokyo, Sao Paulo
|
The following are countries - India, USA, Germany, United Kingdom, Australia
The following are cities - Mumbai, Seattle, Berlin, Tokyo, Sao Paulo
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.