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 |
Jens Stub (3 March 1764 – 24 October 1819) was a Norwegian priest and politician. He was vicar on the island of Veøya and served as a representative at the Norwegian Constitutional Assembly at Eidsvoll in 1814.
|
In what month did Jens Stub die?
|
Jens Stub died in October.
|
null | false
| null |
What year did World War 2 end?
|
1945
|
|
1807.09000
| true
| null |
Experiment 1 directly tested the hypothesis that speakers increase their specificity in contexts with asymmetry in visual access. We found that speakers are not only context-sensitive in choosing referring expressions that distinguish target from distractors in a shared context, but are occlusion-sensitive, adaptively compensating for uncertainty. Critically, this resulted in systematic differences in behavior across the occlusion conditions that are difficult to explain under an egocentric theory: in the presence of occlusions, speakers were spontaneously willing to spend additional time and keystrokes to give further information beyond what they produce in the corresponding unoccluded contexts, even though that information is equally redundant given the visible objects in their display.
These results strongly suggest that the speaker's informativity influences listener accuracy. In support of this hypothesis, we found a strong negative correlation between informativity and error rates across items and conditions: listeners make fewer errors when utterances are a better fit for the target relative to the distractor ( $\rho = -0.81$ , bootstrapped 95% CI $= [-0.9, -0.7]$ ; Fig. 6 B). This result suggests that listener behavior is driven by an expectation of speaker informativity: listeners interpret utterances proportionally to how well they fit objects in context.
Are human adults expert mind-readers, or fundamentally egocentric? The longstanding debate over the role of theory of mind in communication has largely centered around whether listeners (or speakers) with private information consider their partner's perspective BIBREF30 , BIBREF16 . Our work presents a more nuanced picture of how a speaker and a listener use theory of mind to modulate their pragmatic expectations. The Gricean cooperative principle emphasizes a natural division of labor in how the joint effort of being cooperative is shared BIBREF4 , BIBREF60 . It can be asymmetric when one partner is expected to, and able to, take on more complex reasoning than the other, in the form of visual perspective-taking, pragmatic inference, or avoiding further exchanges of clarification and repair. One such case is when the speaker has uncertainty over what the listener can see, as in the director-matcher task. Our Rational Speech Act (RSA) formalization of cooperative reasoning in this context predicts that speakers (directors) naturally increase the informativity of their referring expressions to hedge against the increased risk of misunderstanding; Exp. 1 presents direct evidence in support of this hypothesis.
Importantly, when the director (speaker) is expected to be appropriately informative, communication can be successful even when the matcher (listener) does not reciprocate the effort. If visual perspective-taking is effortful and cognitively demanding BIBREF39 , the matcher will actually minimize joint effort by not taking the director's visual perspective. This suggests a less egocentric explanation of when and why listeners neglect the speaker's visual perspective; they do so when they expect the speaker to disambiguate referents sufficiently. While adaptive in most natural communicative contexts, such neglect might backfire and lead to errors when the speaker (inexplicably) violates this expectation. From this point of view, the “failure” of listener theory of mind in these tasks is not really a failure; instead, it suggests that both speakers and listeners may use theory of mind to know when (and how much) they should expect others to be cooperative and informative, and subsequently allocate their resources accordingly BIBREF36 . Exp. 2 is consistent with this hypothesis; when directors used underinformative scripted instructions (taken from prior work), listeners made significantly more errors than when speakers were allowed to provide referring expressions at their natural level of informativity, and speaker informativeness strongly modulated listener error rates.
Experiment 1 directly tested the hypothesis that speakers increase their specificity in contexts with asymmetry in visual access. We found that speakers are not only context-sensitive in choosing referring expressions that distinguish target from distractors in a shared context, but are occlusion-sensitive, adaptively compensating for uncertainty.
These results strongly suggest that the speaker's informativity influences listener accuracy. In support of this hypothesis, we found a strong negative correlation between informativity and error rates across items and conditions: listeners make fewer errors when utterances are a better fit for the target relative to the distractor ( $\rho = -0.81$ , bootstrapped 95% CI $= [-0.9, -0.7]$ ; Fig. 6 B). This result suggests that listener behavior is driven by an expectation of speaker informativity: listeners interpret utterances proportionally to how well they fit objects in context.
Our Rational Speech Act (RSA) formalization of cooperative reasoning in this context predicts that speakers (directors) naturally increase the informativity of their referring expressions to hedge against the increased risk of misunderstanding; Exp. 1 presents direct evidence in support of this hypothesis.
Exp. 2 is consistent with this hypothesis; when directors used underinformative scripted instructions (taken from prior work), listeners made significantly more errors than when speakers were allowed to provide referring expressions at their natural level of informativity, and speaker informativeness strongly modulated listener error rates.
|
Did participants behave unexpectedly?
|
No.
|
null | false
| 30
|
Recently, neural machine translation (NMT) has gained popularity in the field of machine translation. The conventional encoder-decoder NMT proposed by Cho2014 uses two recurrent neural networks (RNN): one is an encoder, which encodes a source sequence into a fixed-length vector, and the other is a decoder, which decodes the vector into a target sequence. A newly proposed attention-based NMT by DzmitryBahdana2014 can predict output words using the weights of each hidden state of the encoder by the attention mechanism, improving the adequacy of translation.
Even with the success of attention-based models, a number of open questions remain in NMT. Tu2016 argued two of the common problems are over-translation: some words are repeatedly translated unnecessary and under-translation: some words are mistakenly untranslated. This is due to the fact that NMT can not completely convert the information from the source sentence to the target sentence. Mi2016a and Feng2016 pointed out that NMT lacks the notion of coverage vector in phrase-based statistical machine translation (PBSMT), so unless otherwise specified, there is no way to prevent missing translations.
Another problem in NMT is an objective function. NMT is optimized by cross-entropy; therefore, it does not directly maximize the translation accuracy. Shen2016 pointed out that optimization by cross-entropy is not appropriate and proposed a method of optimization based on a translation accuracy score, such as expected BLEU, which led to improvement of translation accuracy. However, BLEU is an evaluation metric based on n-gram precision; therefore, repetition of some words may be present in the translation even though the BLEU score is improved.
To address to problem of repeating and missing words in the translation, tu2016neural introduce an encoder-decoder-reconstructor framework that optimizes NMT by back-translation from the output sentences into the original source sentences. In their method, after training the forward translation in a manner similar to the conventional attention-based NMT, they train a back-translation model from the hidden state of the decoder into the source sequence by a new decoder to enforce agreement between source and target sentences.
In order to confirm the language independence of the framework, we experiment on two parallel corpora of English-Japanese and Japanese-English translation tasks using encode-decoder-reconstructor. Our experiments show that their method offers significant improvement in BLEU scores and alleviates the problem of repeating and missing words in the translation on English-Japanese translation task, though the difference is not significant on Japanese-English translation task.
In addition, we jointly train a model of forward translation and back-translation without pre-training, and then evaluate this model. As a result, the encoder-decoder-reconstructor can not be trained well without pre-training, so it proves that we have to train the forward translation model in a manner similar to the conventional attention-based NMT as pre-training.
The main contributions of this paper are as follows:
In order to confirm the language independence of the framework, we experiment on two parallel corpora of English-Japanese and Japanese-English translation tasks using encode-decoderreconstructor.
|
What experiments did the authors do to verify the framework proposed by Tu et al.?
|
They experiment on two parallel corpora of English-Japanese and Japanese-English translation tasks using encode-decoderreconstructor.
|
null | false
| 398
|
In the life cycle of many production machine-learning applications, maintaining and improving deployed models is the dominant factor in their total cost and effectiveness–much greater than the cost of de novo model construction. Yet, there is little tooling for model life-cycle support. For such applications, a key task for supporting engineers is to improve and maintain the quality in the face of changes to the input distribution and new production features. This work describes a new style of data management system called Overton that provides abstractions to support the model life cycle by helping build models, manage supervision, and monitor application quality.
Overton is used in both near-real-time and backend production applications. However, for concreteness, our running example is a product that answers factoid queries, such as “how tall is the president of the united states?” In our experience, the engineers who maintain such machine learning products face several challenges on which they spend the bulk of their time.
Fine-grained Quality Monitoring While overall improvements to quality scores are important, often the week-to-week battle is improving fine-grained quality for important subsets of the input data. An individual subset may be rare but are nonetheless important, e.g., 0.1% of queries may correspond to a product feature that appears in an advertisement and so has an outsized importance. Traditional machine learning approaches effectively optimize for aggregate quality. As hundreds of such subsets are common in production applications, this presents data management and modeling challenges. An ideal system would monitor these subsets and provide tools to improve these subsets while maintaining overall quality.
Support for Multi-component Pipelines Even simple machine learning products comprise myriad individual tasks. Answering even a simple factoid query, such as “how tall is the president of the united states?” requires tackling many tasks including (1) find the named entities (`united states', and `president'), (2) find the database ids for named entities, (3) find the intent of the question, e.g., the height of the topic entity, (4) determine the topic entity, e.g., neither president nor united states, but the person Donald J. Trump, who is not explicitly mentioned, and (5) decide the appropriate UI to render it on a particular device. Any of these tasks can go wrong. Traditionally, systems are constructed as pipelines, and so determining which task is the culprit is challenging.
Updating Supervision When new features are created or quality bugs are identified, engineers provide additional supervision. Traditionally, supervision is provided by annotators (of varying skill levels), but increasingly programmatic supervision is the dominant form of supervision BIBREF0, BIBREF1, which includes labeling, data augmentation, and creating synthetic data. For both privacy and cost reasons, many applications are constructed using programmatic supervision as a primary source. An ideal system can accept supervision at multiple granularities and resolve conflicting supervision for those tasks.
There are other desiderata for such a system, but the commodity machine learning stack has evolved to support them: building deployment models, hyperparameter tuning, and simple model search are now well supported by commodity packages including TensorFlow, containers, and (private or public) cloud infrastructure. By combining these new systems, Overton is able to automate many of the traditional modeling choices, including deep learning architecture, its hyperparameters, and even which embeddings are used.
Overton provides the engineer with abstractions that allow them to build, maintain, and monitor their application by manipulating data files–not custom code. Inspired by relational systems, supervision (data) is managed separately from the model (schema). Akin to traditional logical independence, Overton's schema provides model independence: serving code does not change even when inputs, parameters, or resources of the model change. The schema changes very infrequently–many production services have not updated their schema in over a year.
Overton takes as input a schema whose design goal is to support rich applications from modeling to automatic deployment. In more detail, the schema has two elements: (1) data payloads similar to a relational schema, which describe the input data, and (2) model tasks, which describe the tasks that need to be accomplished. The schema defines the input, output, and coarse-grained data flow of a deep learning model. Informally, the schema defines what the model computes but not how the model computes it: Overton does not prescribe architectural details of the underlying model (e.g., Overton is free to embed sentences using an LSTM or a Transformer) or hyperparameters, like hidden state size. Additionally, sources of supervision are described as data–not in the schema–so they are free to rapidly evolve.
As shown in Figure FIGREF7, given a schema and a data file, Overton is responsible to instantiate and train a model, combine supervision, select the model's hyperparameters, and produce a production-ready binary. Overton compiles the schema into a (parameterized) TensorFlow or PyTorch program, and performs an architecture and hyperparameter search. A benefit of this compilation approach is that Overton can use standard toolkits to monitor training (TensorBoard equivalents) and to meet service-level agreements (Profilers). The models and metadata are written to an S3-like data store that is accessible from the production infrastructure. This has enabled model retraining and deployment to be nearly automatic, allowing teams to ship products more quickly.
In retrospect, the following three choices of Overton were the most important in meeting the above challenges.
(1) Code-free Deep Learning In Overton-based systems, engineers focus exclusively on fine-grained monitoring of their application quality and improving supervision–not tweaking deep learning models. An Overton engineer does not write any deep learning code in frameworks like TensorFlow. To support application quality improvement, we use a technique, called model slicing BIBREF3. The main idea is to allow the developer to identify fine-grained subsets of the input that are important to the product, e.g., queries about nutrition or queries that require sophisticated disambiguation. The system uses developer-defined slices as a guide to increase representation capacity. Using this recently developed technique led to state-of-the-art results on natural language benchmarks including GLUE and SuperGLUE BIBREF4.
(2) Multitask Learning Overton was built to natively support multitask learning BIBREF5, BIBREF6, BIBREF7 so that all model tasks are concurrently predicted. A key benefit is that Overton can accept supervision at whatever granularity (for whatever task) is available. Overton models often perform ancillary tasks like part-of-speech tagging or typing. Intuitively, if a representation has captured the semantics of a query, then it should reliably perform these ancillary tasks. Typically, ancillary tasks are also chosen either to be inexpensive to supervise. Ancillary task also allow developers to gain confidence in the model's predictions and have proved to be helpful for aids for debugging errors.
(3) Weak Supervision Applications have access to supervision of varying quality and combining this contradictory and incomplete supervision is a major challenge. Overton uses techniques from Snorkel BIBREF1 and Google's Snorkel DryBell BIBREF0, which have studied how to combine supervision in theory and in software. Here, we describe two novel observations from building production applications: (1) we describe the shift to applications which are constructed almost entirely with weakly supervised data due to cost, privacy, and cold-start issues, and (2) we observe that weak supervision may obviate the need for popular methods like transfer learning from massive pretrained models, e.g., BERT BIBREF8–on some production workloads, which suggests that a deeper trade-off study may be illuminating.
In summary, Overton represents a first-of-its kind machine-learning lifecycle management system that has a focus on monitoring and improving application quality. A key idea is to separate the model and data, which is enabled by a code-free approach to deep learning. Overton repurposes ideas from the database community and the machine learning community to help engineers in supporting the lifecycle of machine learning toolkits. This design is informed and refined from use in production systems for over a year in multiple machine-learned products.
Overton is used in both near-real-time and backend production applications.
|
What applications are Overton used in?
|
Near-real-time and backend production applications.
|
null | false
| 377
|
In the process, we have gather a total of 449 expert judgments (one expert failed to answer a given problem), 146 missing hypotheses and 47 explanations for negative judgments.
Despite being marked as “yes”, the problems with a reported missing hypothesis should really be classified as “no”, if one does not assume external knowledge. (See below for further discussion on the reported missing hypotheses.) Thus crunching the numbers, we see that more that half of the responses express some doubt about entailment. Remember that all problems were marked as “yes” by the creators of the RTE3 testsuite — we find here that one average, one expert in two is likely to cast a doubt over this “yes”.
However, each problem was classified by three experts. The histogram below shows the distribution of number of experts casting doubt on entailment, over all problems.
[ybar interval, ymax=0.5,ymin=0, minor y tick num = 0] coordinates (0, 0.26666666666666666) (1, 0.17333333333333334) (2, 0.3466666666666667) (3, 0.21333333333333335) (4,0) ;
Unfortunately we can only draw preliminary conclusions, due to the limited number of respondents for each problem. However, we can make the following observations:
We find this level of agreement indicative of a good level of reliability. Additionally, with three experts per problem, we are very likely to discover most missing hypotheses and incorrect entailments.
In our compilation of answers, we have marked 42 problems as straight “No”, 64 as “Yes” with missing implicit hypotheses and “44” as plain “Yes”. This means that, we expect, in our opinion, 28% of problems to be incorrectly labeled in RTE3 even assuming reasonable world knowledge. An additional 42% of problems require additional (yet reasonable to assume) hypotheses for entailment to hold formally, as prescribed by RTE3. This leaves only 30% of problems to acceptable as such. The reason that the amount of doubt is larger than in the average numbers quoted above is that, for many problems, certain missing hypotheses and/or error were not detected by a majority experts, but, after careful inspection, we judge that the minority report is justified.
We have additionally tagged each missing hypothesis according to the following classification:
We have additionally tagged each missing hypothesis according to the following classification: 1. Linguistic subtleties (Labeled “Language”; Example: “ownership in the past is enough to justify the possessive in the present”; 9 occurrences in our sample) 2. Lexical meaning, sometimes specific to the context of the problem; (Labeled “Lexicon”; Example: “buying entails selling”; 15 occurrences in our sample) 3. World knowledge (Labeled “World”; Example: “Increased amounts of CO2 and other greenhouse gases cause Greenhouse effect.”; 13 occurrences in our sample) 4. Other missing hypothesis, see below for further details.
|
How to tag each missing hypothesis?
|
The authors have additionally tagged each missing hypothesis according to the following classification: 1. Linguistic subtleties; 2. Lexical meaning; 3. World knowledge; 4. Other missing hypothesis.
|
null | false
| null |
Jill, a 60-year-old personal training client, wants to improve her hip flexibility. Identify two exercises commonly done in a Pilates class that would address core strength and flexibility, thus improving Jill's lower back health.
|
One exercise is the criss-cross, which helps target the participant’s core strength and flexibility. The participant alternates bringing one shoulder towards the opposing knee, while keeping their hands on the back of their head, lying face-up, and curling their head up. Another exercise would be plank leg lift, which starts as a regular high plank. The participant alternates lifting one leg off the floor, keeping their core engaged to stabilize the movement.
|
|
null | false
| null |
What is "woke" and why is it so controversial?
|
Woke is a general political term used to group a set of similar ideologies in an attempt to characterize the political motivations typically ascribed to the far left. Although many definitions of woke exist, these definitions run along a few common threads. Woke is typically used to describe the ideals underpinning what is known as equality of outcome. This is to be contrasted with the opposing political forces who support equality of opportunity.
Those who engage in wokeness more often than not cite disparities as evidence of systemic biases or expressions of racism, sexism, classism, genderism, agism and other prejudices. A common example would be that females are disproportionately represented in elementary school teaching whereas males are disproportionately represented in railroad construction. Woke politics would lead us to believe that this is evidence of sexism: women and men are socialized differently so they are pushed into different positions by institutions. Woke opponents would lead us to believe this is evidence of inate preferences different between the sexes: males are generally stronger than females, so men are more suited to performing heavy labor.
Wokeness is controversial because its opponents do not believe disparity is evidence of bias. Disparities can exist due to natural or cultural differences, neither of which can be governed out of existence. The opponents of woke ideology also proclaim that it ignores the root cause of problems which plague society. If one's analysis stops at disparity and from disparity an answer is assumed, one can miss the nuances present in a situation. A common unavoidable disparity cited is that 100% of births are to females, but this is not sexism. Analysis past disparity is required to prove or disprove systemic or individual bias.
|
|
null | false
| null |
Name some famous hindi film actresses
|
Here is a list of some famous hindi actresses - Priyanka Chopra, Aishwarya Rai, Deepika Padukone, Alia Bhatt, Anushka Sharma and Katrina Kaif
|
|
null | false
| null |
Shohei Ohtani, nicknamed "Shotime", is a Japanese professional baseball pitcher, designated hitter and outfielder for the Los Angeles Angels of Major League Baseball (MLB). He previously played for the Hokkaido Nippon-Ham Fighters of Nippon Professional Baseball's (NPB) Pacific League.
Considered early on as an elite two-way player, Ohtani was the first pick of the Fighters in the 2012 draft. He played in NPB for the Fighters from 2013 through 2017 as a pitcher and an outfielder, and won the 2016 Japan Series with them. The Fighters posted Ohtani to MLB after the 2017 season, and he signed with the Angels, soon winning the 2018 American League (AL) Rookie of the Year Award.
Following an injury-plagued 2019 and 2020, Ohtani would go on to have a 2021 season widely considered to be historic, as he became the first in the history of MLB with 10+ home runs and 20+ stolen bases as a hitter and 100+ strikeouts and 10+ pitching appearances as a pitcher in the same season while also holding at least a share of the major league lead in home runs in fourteen starts. Because of his elite contributions both offensively and as a pitcher, a rarity for two-way players, Ohtani's peak is widely considered among the greatest in baseball history, with some comparing it favorably to the early career of Babe Ruth. For his efforts, he was awarded the 2021 American League Most Valuable Player Award. He followed this in 2022 by becoming the first player in the modern era to qualify for both the hitting and pitching leaderboards in one season, reaching the limits of 3.1 plate appearances and one inning pitched per game with 586 at bats against 166 innings pitched.
Internationally, Ohtani also represented Japan in the 2023 World Baseball Classic, winning the MVP Award for the tournament following Team Japan's victory over Team USA.
|
Please give me a short bulleted list of the key arguments the Shohei Ohtani in his career.
|
Shohei Otani is a baseball player in the history of MLB. He did:
・Won the 2018 American League (AL) Rookie of the Year Award.
・10+ home runs and 20+ stolen bases as a hitter
・100+ strikeouts and 10+ pitching appearances as a pitcher in the same season.
・Awarded the 2021 American League Most Valuable Player Award.
・Awarded the 2023 World Baseball Classic Most Valuable Player Award
|
null | false
| null |
There are ten different eye layouts—indeed every technological method of capturing an optical image commonly used by human beings, with the exceptions of zoom and Fresnel lenses, occur in nature. Eye types can be categorised into "simple eyes", with one concave photoreceptive surface, and "compound eyes", which comprise a number of individual lenses laid out on a convex surface. Note that "simple" does not imply a reduced level of complexity or acuity. Indeed, any eye type can be adapted for almost any behaviour or environment. The only limitations specific to eye types are that of resolution—the physics of compound eyes prevents them from achieving a resolution better than 1°. Also, superposition eyes can achieve greater sensitivity than apposition eyes, so are better suited to dark-dwelling creatures. Eyes also fall into two groups on the basis of their photoreceptor's cellular construction, with the photoreceptor cells either being cilliated (as in the vertebrates) or rhabdomeric. These two groups are not monophyletic; the cnidaria also possess cilliated cells, and some gastropods, as well as some annelids possess both.
Some organisms have photosensitive cells that do nothing but detect whether the surroundings are light or dark, which is sufficient for the entrainment of circadian rhythms. These are not considered eyes because they lack enough structure to be considered an organ, and do not produce an image.
|
Given this paragraph about different types of eyes, how are eye types categorized and why?
|
Eyes are categorized into two types, "simple eyes" and "compound eyes". In this case, "simple eyes" do not mean that they have reduced complexity or acuity, but rather that they have one concave photoreceptive surface. This is in contrast to "compound eyes" which are comprised of a number of individual lenses on a convex surface. In other words, eyes fall into two groups based on their photoreceptor's cellular construction either being cilliated (as in the vertebrates) or rhabdomeric.
|
null | false
| null |
S26 is a line on the Berlin S-Bahn. It runs from Teltow in the district (Kreis) of Potsdam-Mittelmark via Südkreuz, through the Berlin Nord-Süd Tunnel with a stop at Friedrichstraße, and ends in Waidmannslust in the borough of Reinickendorf. On weekends the S26 runs from Teltow to Potsdamer Platz.
Trains on this line run every 20 minutes. This interval is in sections shortened to a 10-minute interval by the S25. There is no night service on this line.
|
Given a reference text about S26, tell me where it begins and ends.
|
S26 begins in Tetlow and ends in Waidmannslust.
|
1911.12579
| false
| null |
Generally, closer words are considered more important to a word’s meaning. The word embeddings models have the ability to capture the lexical relations between words. Identifying such relationship that connects words is important in NLP applications. We measure that semantic relationship by calculating the dot product of two vectors using Eq. DISPLAY_FORM48. The high cosine similarity score denotes the closer words in the embedding matrix, while less cosine similarity score means the higher distance between word pairs. We present the cosine similarity score of different semantically or syntactically related word pairs taken from the vocabulary in Table TABREF77 along with English translation, which shows the average similarity of 0.632, 0.650, 0.591 yields by CBoW, SG and GloVe respectively. The SG model achieved a high average similarity score of 0.650 followed by CBoW with a 0.632 average similarity score. The GloVe also achieved a considerable average score of 0.591 respectively. However, the average similarity score of SdfastText is 0.388 and the word pair Microsoft-Bill Gates is not available in the vocabulary of SdfastText. This shows that along with performance, the vocabulary in SdfastText is also limited as compared to our proposed word embeddings.
Moreover, the average semantic relatedness similarity score between countries and their capitals is shown in Table TABREF78 with English translation, where SG also yields the best average score of 0.663 followed by CBoW with 0.611 similarity score. The GloVe also yields better semantic relatedness of 0.576 and the SdfastText yield an average score of 0.391. The first query word China-Beijing is not available the vocabulary of SdfastText. However, the similarity score between Afghanistan-Kabul is lower in our proposed CBoW, SG, GloVe models because the word Kabul is the name of the capital of Afghanistan as well as it frequently appears as an adjective in Sindhi text which means able.
The SG model achieved a high average similarity score of 0.650 followed by CBoW with a 0.632 average similarity score. The GloVe also achieved a considerable average score of 0.591 respectively. However, the average similarity score of SdfastText is 0.388 and the word pair Microsoft-Bill Gates is not available in the vocabulary of SdfastText.
Moreover, the average semantic relatedness similarity score between countries and their capitals is shown in Table TABREF78 with English translation, where SG also yields the best average score of 0.663 followed by CBoW with 0.611 similarity score. The GloVe also yields better semantic relatedness of 0.576 and the SdfastText yield an average score of 0.391.
|
How does proposed word embeddings compare to Sindhi fastText word representations?
|
Proposed SG model vs SINDHI FASTTEXT:
Average cosine similarity score: 0.650 vs 0.388
Average semantic relatedness similarity score between countries and their capitals: 0.663 vs 0.391
|
null | false
| null |
What is rationing?
|
Rationing is the restriction of goods.
It is mostly done when there are insufficient goods/products to meet the demand or needs of the people.
It is used to save resources.
|
|
null | false
| null |
Who is raiden shogun?
|
Raiden Shogun is used to refer to Janpanese General. Nowadays she is more referred to a character in Genshin Impact, a popular action role-playing game developed by miHoYo. She is a playable Electro character and also the current Electro Archon of Inazuma1. The Raiden Shogun is unique as she is comprised of two beings in one body: Ei, the current Electro Archon of Inazuma; and the Shogun, the puppet created by Ei to act as the ruler of Inazuma in her stead, which also serves as her vessel. With the might of lightning at her disposal, she commits herself to the solitary pursuit of eternity.
|
|
null | false
| null |
Give me 10 reasons why people dance
|
Passion, love of art, exercise, money, personal expression, communication, celebrations, cultural heritage, to connect and form bonds
|
|
null | false
| null |
Give me a bulleted list of colleges or universities located Massachusetts
|
- Harvard
- Boston College
- Boston University
- Northeastern
- Babson College
- Bentley
- Stonehill
- University of Massachusetts Amherst
- University of Massachusetts Dartmouth
- University of Massachusetts Boston
- Bridgewater State University
- Bunker Hill Community College
- Massachusetts Institute of Technology
- Berklee College of Music
- Tufts University
|
|
null | false
| 212
|
The automatic correction of errors in text [In a such situaction INLINEFORM0 In such a situation] is receiving more and more attention from the natural language processing community. A series of competitions has been devoted to grammatical error correction (GEC): the CoNLL-2013 shared task BIBREF0 , the CoNLL-2014 shared task BIBREF1 , and finally the BEA 2019 shared task BIBREF2 . This paper presents the contributions from the Cambridge University Engineering Department to the latest GEC competition at the BEA 2019 workshop.
We submitted systems to two different tracks. The low-resource track did not permit the use of parallel training data except a small development set with around 4K sentence pairs. For our low-resource system we extended our prior work on finite state transducer based GEC BIBREF3 to handle new error types such as punctuation errors as well as insertions and deletions of a small number of frequent words. For the restricted track, the organizers provided 1.2M pairs (560K without identity mappings) of corrected and uncorrected sentences. Our goal on the restricted track was to explore the potential of purely neural models for grammatical error correction. We confirm the results of BIBREF4 and report substantial gains by applying back-translation BIBREF5 to GEC – a data augmentation technique common in machine translation. Furthermore, we noticed that large parts of the training data do not match the target domain. We mitigated the domain gap by over-sampling the in-domain training corpus, and by fine-tuning through continued training. Our final model is an ensemble of four neural machine translation (NMT) models and two neural language models (LMs) with Transformer architecture BIBREF6 . Our purely neural system was also part of the joint submission with the Cambridge University Computer Lab described by BIBREF7 .
We mitigated the domain gap by over-sampling the in-domain training corpus, and by fine-tuning through continued training.
|
How did they mitigate the domain gap
|
By over-sampling the in-domain training corpus, and fine-tuning through continued training.
|
null | false
| 383
|
Parkinson's disease (PD) is a neurodegenerative disorder characterized by the progressive loss of dopaminergic neurons in the mid-brain producing several motor and non-motor impairments in the patients BIBREF0. Motor symptoms include among others, bradykinesia, rigidity, resting tremor, micrographia, and different speech impairments. The speech impairments observed in PD patients are typically grouped as hypokinetic dysarthria, and include symptoms such as vocal folds rigidity, bradykinesia, and reduced control of muscles and limbs involved in the speech production. The effects of dysarthria in the speech of PD patients include increased acoustic noise, reduced intensity, harsh and breathy voice quality, increased voice nasality, monopitch, monoludness, speech rate disturbances, imprecise articulation of consonants BIBREF1, and involuntary introduction of pauses BIBREF2. Clinical observations in the speech of patients can be objectively and automatically measured by using computer aided methods supported in signal processing and pattern recognition with the aim to address two main aspects: (1) to support the diagnosis of the disease by classifying healthy control (HC) subjects and patients, and (2) to predict the level of degradation of the speech of the patients according to a specific clinical scale.
Most of the studies in the literature to classify PD from speech are based on computing hand-crafted features and using classifiers such as support vector machines (SVMs) or K-nearest neighbors (KNN). For instance, in BIBREF3, the authors computed features related to perturbations of the fundamental frequency and amplitude of the speech signal to classify utterances from 20 PD patients and 20 HC subjects, Turkish speakers. Classifiers based on KNN and SVMs were considered, and accuracies of up to 75% were reported. Later, in BIBREF4 the authors proposed a phonation analysis based on several time frequency representations to assess tremor in the speech of PD patients. The extracted features were based on energy and entropy computed from time frequency representations. Several classifiers were used, including Gaussian mixture models (GMMs) and SVMs. Accuracies of up to 77% were reported in utterances of the PC-GITA database BIBREF5, formed with utterances from 50 PD patients and 50 HC subjects, Colombian Spanish native speakers. The authors from BIBREF6 computed features to model different articulation deficits in PD such as vowel quality, coordination of laryngeal and supra-laryngeal activity, precision of consonant articulation, tongue movement, occlusion weakening, and speech timing. The authors studied the rapid repetition of the syllables /pa-ta-ka/ pronounced by 24 Czech native speakers, and reported an accuracy of 88% discriminating between PD patients and HC speakers, using an SVM classifier. Additional articulation features were proposed in BIBREF7, where the authors modeled the difficulty of PD patients to start/stop the vocal fold vibration in continuous speech. The model was based on the energy content in the transitions between unvoiced and voiced segments. The authors classified PD patients and HC speakers with speech recordings in three different languages (Spanish, German, and Czech), and reported accuracies ranging from 80% to 94% depending on the language; however, the results were optimistic, since the hyper-parameters of the classifier were optimized based on the accuracy on the test set. Another articulation model was proposed in BIBREF8. The authors considered a forced alignment strategy to segment the different phonetic units in the speech utterances. The phonemes were segmented and grouped to train different GMMs. The classification was performed based on a threshold of the difference between the posterior probabilities from the models created for HC subjects and PD patients. The model was tested with Colombian Spanish utterances from the PC-GITA database BIBREF5 and with the Czech data from BIBREF9. The authors reported accuracies of up to 81% for the Spanish data, and of up to 94% for the Czech data.
In addition to the hand-crafted feature extraction models, there is a growing interest in the research community to consider deep learning models in the assessment of the speech of PD patients BIBREF10, BIBREF11, BIBREF12. Deep learning methods have the potential to extract more abstract and robust features than those manually computed. These features could help to improve the accuracy of different models to classify pathological speech, such as PD BIBREF13. A deep learning based articulation model was proposed in BIBREF11 to model the difficulties of the patients to stop/start the vibration of the vocal folds. Transitions between voiced and unvoiced segments were modeled with time-frequency representations and convolutional neural networks (CNNs). The authors considered speech recordings of PD patients and HC speakers in three languages: Spanish, German, and Czech, and reported accuracies ranging from 70% to 89%, depending on the language. However, in a language independent scenario, i.e., training the CNN with utterances from one language and testing with the remaining two, the results were not satisfactory (accuracy$<60\%$).
The classification of PD from speech in different languages has to be carefully conducted to avoid bias towards the linguistic content present in each language. For instance, Czech and German languages are richer than Spanish language in terms of consonant production, which may cause that it is easier to produce consonant sounds by Czech PD patients than by Spanish PD patients. Despite these language dependent issues, the results in the classification of PD in different languages could be improved using a transfer learning strategy among languages, i.e., to train a base model with utterances from one language, and then, to perform a fine-tuning of the weights with utterances from the target language BIBREF14. Similar approaches based on transfer learning have been recently considered to classify PD using handwriting BIBREF15. In the present study, we propose a methodology to classify PD via a transfer learning strategy with the aim to improve the accuracy in different languages. CNNs trained with utterances from one language are used to initialize a model to classify speech utterances from PD patients in a different language. The models are evaluated with speech utterances in Spanish, German, and Czech languages. The results suggest that the use of a transfer learning strategy improved the accuracy of the models over 8% with respect to those obtained when the model is trained only with utterance from the target language.
The results suggest that the use of a transfer learning strategy improved the accuracy of the models over 8% with respect to those obtained when the model is trained only with utterance from the target language.
|
How much was the accuracy improved by the transfer learning strategy?
|
over 8%.
|
null | false
| null |
I've heard astronauts talk about the normalization of deviance and the risks therein. What is the normalization of deviance, and how do I avoid it?
|
The normalization of deviance was a term coined by the sociologist Diane Vaughan who participated in the review of the Challenger disaster. It was later the the subject of a number of public talks by astronaut Mike Mullane. The theory is essentially that when we accept deviations from an expected standard, the deviation becomes the standard. In the Challenger disaster, engineers kept noticing rubber o-rings that came back from tests in an out of tolerance state. Because there were no dramatic failures during the tests, the engineers adjusted the standard for the o-rings to absorb these abnormalities. Unfortunately, by normalizing deviance from the expected standard, it led to an o-ring failing during the Challenger launch, and the Challenger shuttle disintegrated in flight. The normalization of deviance is an important pitfall to avoid when running any organization, but especially those which put its members in high risk situations, such as space travel. To avoid it, you must remember, or learn, how and why certain standards were developed and adhere to them, even when pressures are high.
|
|
null | false
| 102
|
Intuitively, when a human writes a description from a table she keeps track of information at two levels. At the macro level, it is important to decide which is the appropriate field to attend to next and at a micro level (i.e., within a field) it is important to know which values to attend to next. To capture this behavior, we use a bifocal attention mechanism as described below.
Macro Attention: Consider the INLINEFORM0 -th field INLINEFORM1 which has values INLINEFORM2 . Let INLINEFORM3 be the representation of this field in the infobox. This representation can either be (i) the word embedding of the field name or (ii) some function INLINEFORM4 of the values in the field or (iii) a concatenation of (i) and (ii). The function INLINEFORM5 could simply be the sum or average of the embeddings of the values in the field. Alternately, this function could be a GRU (or LSTM) which treats these values within a field as a sequence and computes the field representation as the final representation of this sequence (i.e., the representation of the last time-step). We found that bidirectional GRU is a better choice for INLINEFORM6 and concatenating the embedding of the field name with this GRU representation works best. Further, using a bidirectional GRU cell to take contextual information from neighboring fields also helps (these are the orange colored cells in the top-left block in Figure FIGREF3 with macro attention). Given these representations INLINEFORM7 for all the INLINEFORM8 fields we compute an attention over the fields (macro level). DISPLAYFORM0
where INLINEFORM0 is the state of the decoder at time step INLINEFORM1 . INLINEFORM2 and INLINEFORM3 are parameters, INLINEFORM4 is the total number of fields in the input, INLINEFORM5 is the macro (field level) context vector at the INLINEFORM6 -th time step of the decoder.
Micro Attention: Let INLINEFORM0 be the representation of the INLINEFORM1 -th value in a given field. This representation could again either be (i) simply the embedding of this value (ii) or a contextual representation computed using a function INLINEFORM2 which also considers the other values in the field. For example, if INLINEFORM3 are the values in a field then these values can be treated as a sequence and the representation of the INLINEFORM4 -th value can be computed using a bidirectional GRU over this sequence. Once again, we found that using a bi-GRU works better then simply using the embedding of the value. Once we have such a representation computed for all values across all the fields, we compute the attention over these values (micro level) as shown below : DISPLAYFORM0
where INLINEFORM0 is the state of the decoder at time step INLINEFORM1 . INLINEFORM2 and INLINEFORM3 are parameters, INLINEFORM4 is the total number of values across all the fields.
Fused Attention: Intuitively, the attention weights assigned to a field should have an influence on all the values belonging to the particular field. To ensure this, we reweigh the micro level attention weights based on the corresponding macro level attention weights. In other words, we fuse the attention weights at the two levels as: DISPLAYFORM0
where INLINEFORM0 is the field corresponding to the INLINEFORM1 -th value, INLINEFORM2 is the macro level context vector.
Intuitively, when a human writes a description from a table she keeps track of information at two levels. At the macro level, it is important to decide which is the appropriate field to attend to next and at a micro level (i.e., within a field) it is important to know which values to attend to next. To capture this behavior, we use a bifocal attention mechanism as described below.
|
What is "Fused Bifocal Attention Mechanism" for?
|
To capture and imitate a behavior when human completes writing a description of paired information.
|
null | false
| null |
What is Azure HDInsight?
|
Azure HDInsight is a cloud-based service from Microsoft for big data analytics that helps organizations process large amounts of streaming or historical data. It runs popular open-source frameworks—including Apache Hadoop, Spark, Hive, Kafka, and more.
|
|
null | false
| null |
Which of the following are fruits and which are vegetables, from this list: Tomatoes, Carrots, Onions, Broccoli, Mushrooms, Cauliflower, Cabbage, Chives, Grapes, Sausages?
|
Most authorities place Tomatoes, and Grapes in the fruit category, while Carrots, Onions, Broccoli, Cauliflower, Cabbage and Chives are definitely in the vegetables category.
And be aware of the trick item on the list: Sausages belong in neither category? That is a type of meat. And as for Mushrooms, they are actually a fungi but most frequently categorised in the "vegetables" section.
|
|
null | false
| 95
|
People are increasingly using social networking platforms such as Twitter, Facebook, YouTube, etc. to communicate their opinions and share information. Although the interactions among users on these platforms can lead to constructive conversations, they have been increasingly exploited for the propagation of abusive language and the organization of hate-based activities BIBREF0, BIBREF1, especially due to the mobility and anonymous environment of these online platforms. Violence attributed to online hate speech has increased worldwide. For example, in the UK, there has been a significant increase in hate speech towards the immigrant and Muslim communities following the UK's leaving the EU and the Manchester and London attacks. The US also has been a marked increase in hate speech and related crime following the Trump election. Therefore, governments and social network platforms confronting the trend must have tools to detect aggressive behavior in general, and hate speech in particular, as these forms of online aggression not only poison the social climate of the online communities that experience it, but can also provoke physical violence and serious harm BIBREF1.
Recently, the problem of online abusive detection has attracted scientific attention. Proof of this is the creation of the third Workshop on Abusive Language Online or Kaggle’s Toxic Comment Classification Challenge that gathered 4,551 teams in 2018 to detect different types of toxicities (threats, obscenity, etc.). In the scope of this work, we mainly focus on the term hate speech as abusive content in social media, since it can be considered a broad umbrella term for numerous kinds of insulting user-generated content. Hate speech is commonly defined as any communication criticizing a person or a group based on some characteristics such as gender, sexual orientation, nationality, religion, race, etc. Hate speech detection is not a stable or simple target because misclassification of regular conversation as hate speech can severely affect users’ freedom of expression and reputation, while misclassification of hateful conversations as unproblematic would maintain the status of online communities as unsafe environments BIBREF2.
To detect online hate speech, a large number of scientific studies have been dedicated by using Natural Language Processing (NLP) in combination with Machine Learning (ML) and Deep Learning (DL) methods BIBREF3, BIBREF4, BIBREF5, BIBREF6, BIBREF7, BIBREF0. Although supervised machine learning-based approaches have used different text mining-based features such as surface features, sentiment analysis, lexical resources, linguistic features, knowledge-based features or user-based and platform-based metadata BIBREF8, BIBREF9, BIBREF10, they necessitate a well-defined feature extraction approach. The trend now seems to be changing direction, with deep learning models being used for both feature extraction and the training of classifiers. These newer models are applying deep learning approaches such as Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), etc.BIBREF6, BIBREF0 to enhance the performance of hate speech detection models, however, they still suffer from lack of labelled data or inability to improve generalization property.
Here, we propose a transfer learning approach for hate speech understanding using a combination of the unsupervised pre-trained model BERT BIBREF11 and some new supervised fine-tuning strategies. As far as we know, it is the first time that such exhaustive fine-tuning strategies are proposed along with a generative pre-trained language model to transfer learning to low-resource hate speech languages and improve performance of the task. In summary:
We propose a transfer learning approach using the pre-trained language model BERT learned on English Wikipedia and BookCorpus to enhance hate speech detection on publicly available benchmark datasets. Toward that end, for the first time, we introduce new fine-tuning strategies to examine the effect of different embedding layers of BERT in hate speech detection.
Our experiment results show that using the pre-trained BERT model and fine-tuning it on the downstream task by leveraging syntactical and contextual information of all BERT's transformers outperforms previous works in terms of precision, recall, and F1-score. Furthermore, examining the results shows the ability of our model to detect some biases in the process of collecting or annotating datasets. It can be a valuable clue in using pre-trained BERT model for debiasing hate speech datasets in future studies.
Furthermore, examining the results shows the ability of our model to detect some biases in the process of collecting or annotating datasets.
|
What is the ability of the generative pre-trained language model?
|
Examining the results shows the ability of the model to detect some biases in the process of collecting or annotating datasets
|
null | false
| null |
Otto, Prince of Bismarck, Count of Bismarck-Schönhausen, Duke of Lauenburg (German: Otto Fürst von Bismarck, Graf von Bismarck-Schönhausen, Herzog zu Lauenburg, pronounced [ˈɔtoː fɔn ˈbɪsmaʁk] (listen); 1 April 1815 – 30 July 1898), born Otto Eduard Leopold von Bismarck, was a Prussian and later German statesman and diplomat. From his origins in the upper class of Junker landowners, Bismarck rose rapidly in Prussian politics, and from 1862 to 1890 he was the minister president and foreign minister of Prussia. Before his rise to the executive, he was the Prussian ambassador to Russia and France and served in both houses of the Prussian parliament. He masterminded the unification of Germany in 1871 and served as the first chancellor of the German Empire until 1890, in which capacity he dominated European affairs. He had served as chancellor of the North German Confederation from 1867 to 1871, alongside his responsibilities in the Kingdom of Prussia. He cooperated with King Wilhelm I of Prussia to unify the various German states, a partnership that would last for the rest of Wilhelm's life. The King granted Bismarck the titles of Count of Bismarck-Schönhausen in 1865 and Prince of Bismarck in 1871. Bismarck provoked three short, decisive wars against Denmark, Austria, and France. Following the victory against Austria, he abolished the supranational German Confederation and instead formed the North German Confederation as the first German national state, aligning the smaller North German states behind Prussia, while excluding Austria. Receiving the support of the independent South German states in the Confederation's defeat of France, he formed the German Empire – which also excluded Austria – and united Germany.
|
Who was Otto von Bismarck?
|
Otto von Bismarck was a Prussian and German soldier, statesman and diplomat who was largely responsible for Prussian unification with German as well as securing and ensuring its position through diplomacy.
|
null | false
| null |
What are things to look out for when buying an old house in New England?
|
The most important piece in buying an old New England home (100+ years old) is to make sure the foundation is in good condition. Field stones (if applicable) and support beams need to be checked to insure safety and stability for the years to come. Next, you'll want to make sure it is insulated and the HVAC system is in good condition. Along that, review what type of electricity the house had wired as some older homes may be using older setups that may be dangerous or hazardous.
|
|
null | false
| 242
|
Automatic Text Summarization deals with the task of condensing documents into a summary, whose level is similar to a human-generated summary. It is mostly distributed into two distinct domains, i.e., Abstractive Summarization and Extractive Summarization. Abstractive summarization( Dejong et al. ,1978) involves models to deduce the crux of the document. It then presents a summary consisting of words and phrases that were not there in the actual document, sometimes even paraphrasing BIBREF1 . A state of art method proposed by Wenyuan Zeng BIBREF2 produces such summaries with length restricted to 75. There have been many recent developments that produce optimal results, but it is still in a developing phase. It highly relies on natural language processing techniques, which is still evolving to match human standards. These shortcomings make abstractive summarization highly domain selective. As a result, their application is skewed to the areas where NLP techniques have been superlative. Extractive Summarization, on the other hand, uses different methods to identify the most informative/dominant sentences through the text, and then present the results, ranking them accordingly. In this paper, we have proposed two novel stand-alone summarization methods.The first method is based on Glove Model BIBREF3 ,and other is based on Facebook's InferSent BIBREF4 . We have also discussed how we can effectively subdue shortcomings of one model by using it in coalition with models which capture the view that other faintly held.
On the basis of pre-trained BERT, we add an additional pooling module, design some pooling strategies for integrating the multi-layer representations of the classification token.
|
What module do they add on the basis of pre-trained BERT?
|
Pooling module.
|
null | false
| null |
Which is a species of fish? Seahorse or Arabian
|
Seahorse
|
|
null | false
| 307
|
Question answering (QA) is a high-level natural language processing task that requires automatically providing answers to natural language questions. The approaches used to construct QA solvers vary depending on the questions and domain, from inference methods that attempt to construct answers from semantic, syntactic, or logical decompositions, to retrieval methods that work to identify passages of text likely to contain the answer in large corpora using statistical methods. Because of the difficulty of this task, overall QA task performance tends to be low, with generally between 20% and 80% of natural (non-artificially generated) questions answered correctly, depending on the questions, the domain, and the knowledge and inference requirements.
Standardized science exams have recently been proposed as a challenge task for question answering BIBREF0 , as these questions have very challenging knowledge and inference requirements BIBREF1 , BIBREF2 , but are expressed in simple-enough language that the linguistic challenges are likely surmountable in the near-term. They also provide a standardized comparison of modern inference techniques against human performance, with individual QA solvers generally answering between 40% to 50% of multiple choice science questions correctly BIBREF3 , BIBREF4 , BIBREF5 , BIBREF6 , BIBREF7 , and top-performing ensemble models nearly reaching a passing grade of 60% on middle school (8th grade) science exams during a recent worldwide competition of 780 teams sponsored by the Allen Institute for AI BIBREF8 .
One of the central shortcomings of question answering models is that while solvers are steadily increasing the proportion of questions they answer correctly, most solvers generally lack the capacity to provide human-readable explanations or justifications for why those answers are correct. This “explainable inference” task is seen as a limitation of current machine learning models in general (e.g. Ribeiro et al., Ribeiro2016), but is critical for domains such as science or medicine where user trust and detecting potentially costly errors are important. More than this, evidence from the cognitive and pedagogy literature suggests that explanations (when tutoring others) and self-explanations (when engaged in self-directed learning) are an important aspect of learning, helping humans better generalize the knowledge they have learned BIBREF9 , BIBREF10 , BIBREF11 . This suggests that explainable methods of inference may not only be desirable for users, but may be a requirement for automated systems to have human-like generalization and inference capabilities.
Building QA solvers that generate explanations for their answers is a challenging task, requiring a number of inference capacities. Central among these is the idea of information aggregation, or the idea that explanations for a given question are rarely found in a contiguous passage of text, and as such inference methods must generally assemble many separate pieces of knowledge from different sources in order to arrive at a correct answer. Previous estimates BIBREF2 suggest elementary science questions require an average of 4 pieces of knowledge to answer and explain those answers (here our analysis suggests this is closer to 6), but inference methods tend to have difficulty aggregating more than 2 pieces of knowledge from free-text together due to the semantic or contextual “drift” associated with this aggregation BIBREF12 . Because of the difficulty in assembling training data for the information aggregation task, some have approached explanation generation as a distant supervision problem, with explanation quality modelled as a latent variable BIBREF7 , BIBREF13 . While these techniques have had some success in constructing short explanations, semantic drift likely limits the viability of this technique for explanations requiring more than two pieces of information to be aggregated.
To address this, here we construct a large corpus of explanation graphs (see Figure 1 ) to serve as training data for explainable inference tasks. The contributions of this work are are:
While these techniques have had some success in constructing short explanations, semantic drift likely limits the viability of this technique for explanations requiring more than two pieces of information to be aggregated. To address this, here we construct a large corpus of explanation graphs (see Figure 1 ) to serve as training data for explainable inference tasks.
|
What task was their corpus created for?
|
Explainable inference tasks.
|
null | false
| null |
What are the common formats of competitive debate in the US?
|
- Public Forum
- Lincoln-Douglas
- Policy
- Congressional Debate
- British Parliamentary
- World Schools Debate
- International Public Policy Forum
|
|
null | false
| null |
Apart from fission induced by a neutron, harnessed and exploited by humans, a natural form of spontaneous radioactive decay (not requiring a neutron) is also referred to as fission, and occurs especially in very high-mass-number isotopes. Spontaneous fission was discovered in 1940 by Flyorov, Petrzhak, and Kurchatov in Moscow, in an experiment intended to confirm that, without bombardment by neutrons, the fission rate of uranium was negligible, as predicted by Niels Bohr; it was not negligible.
The unpredictable composition of the products (which vary in a broad probabilistic and somewhat chaotic manner) distinguishes fission from purely quantum tunneling processes such as proton emission, alpha decay, and cluster decay, which give the same products each time. Nuclear fission produces energy for nuclear power and drives the explosion of nuclear weapons. Both uses are possible because certain substances called nuclear fuels undergo fission when struck by fission neutrons, and in turn emit neutrons when they break apart. This makes a self-sustaining nuclear chain reaction possible, releasing energy at a controlled rate in a nuclear reactor or at a very rapid, uncontrolled rate in a nuclear weapon.
The amount of free energy contained in nuclear fuel is millions of times the amount of free energy contained in a similar mass of chemical fuel such as gasoline, making nuclear fission a very dense source of energy. The products of nuclear fission, however, are on average far more radioactive than the heavy elements which are normally fissioned as fuel, and remain so for significant amounts of time, giving rise to a nuclear waste problem. However, the seven long-lived fission products make up only a small fraction of fission products. Neutron absorption which does not lead to fission produces Plutonium (from 238
U) and minor actinides (from both 235
U and 238
U) whose radiotoxicity is far higher than that of the long lived fission products. Concerns over nuclear waste accumulation and the destructive potential of nuclear weapons are a counterbalance to the peaceful desire to use fission as an energy source. The thorium fuel cycle produces virtually no plutonium and much less minor actinides, but 232
U - or rather its decay products - are a major gamma ray emitter. All actinides are fertile or fissile and fast breeder reactors can fission them all albeit only in certain configurations. Nuclear reprocessing aims to recover usable material from spent nuclear fuel to both enable uranium (and thorium) supplies to last longer and to reduce the amount of "waste". The industry term for a process that fissions all or nearly all actinides is a "closed fuel cycle".
|
What are the examples of quantum tunneling processes?
|
proton emission, alpha decay, and cluster decay.
|
null | false
| null |
WABG-TV's first broadcast was on October 20, 1959, on VHF channel 6. Originally it was a CBS affiliate, with a secondary affiliation with ABC owing to WABG radio (960 AM)'s affiliation with ABC Radio. In 1966, the station built a 1,200-foot (366 m) tower near Inverness and increased its effective radiated power to 100,000 watts. Following its completion in October 1966, WABG dropped CBS to become a full-time ABC affiliate in November 1966. Until then, the only areas of the state to receive a sole ABC affiliate were the northwest (from Memphis' WHBQ-TV) and the Gulf Coast (from WVUE in New Orleans). Until future sister station WXVT signed on in 1980, WJTV in Jackson served as the default CBS affiliate for the southern counties in the Delta area, while WREG-TV in Memphis served the northern half of the market, Mississippi cities like Kosciusko received WABG-TV on their analog television set as their default ABC station from 1970 to 2009, as WAPT in Jackson did not have a strong signal on analog television in the Kosciusko area. On March 12, 1986, the studios of WABG were damaged by a fire triggered by an electrical short circuit after the station signed off the air for the night. With no injuries being reported, the station was forced to relocate its transmission signal.[citation needed]
|
List all of the call signs mentioned in the paragraph
|
Sure, here is a full list of all the call signs that were mentioned in the paragraph:
- WABG-TV
- WABG
- WHBQ-TV
- WVUE
- WXVT
- WJTV
- WREG-TV
- WAPT
|
null | false
| null |
Is the Tiger population Increasing in India?
|
India is one of the very few countries in which the Tiger population is flourishing. Thanks to the reserve forests and dedicated support by the public and NGOs.There has been a steady increase in the Tiger population year over year and Karnataka state has been the front-runner in protecting and growing the Tiger population
|
|
null | false
| null |
An isobar is a line of equal or constant pressure on a graph, plot, or map; an isopleth or contour line of pressure. More accurately, isobars are lines drawn on a map joining places of equal average atmospheric pressure reduced to sea level for a specified period of time. In meteorology, the barometric pressures shown are reduced to sea level, not the surface pressures at the map locations. The distribution of isobars is closely related to the magnitude and direction of the wind field, and can be used to predict future weather patterns. Isobars are commonly used in television weather reporting.
|
What does the distribution of isobars indicate?
|
The distribution of isobars is closely related to the magnitude and direction of the wind field, and can be used to predict future weather patterns.
|
null | false
| null |
In 1872 the Boundary Commission travelled through this area while surveying the border between Canada and the United States. During the 1874 March West trek of the North-West Mounted Police, the force established its first camp (called Short Creek Camp) at this location. The community began being served by the Soo Line in 1893, permitting the coal mines in the region to become operational. Roche Percée incorporated as a village on January 12, 1909.
|
Using this passage, extract all the years associated with the area.
|
1872, 1874, 1893, 1909
|
null | false
| null |
Please write a haiku
|
A Siamese cat
Trotting through the sticks and leaves
Follows his next meal
|
|
null | false
| 446
|
Self-supervised Learning In computer vision, self-supervised learning (SSL) aims to learn visual representations without any labels. Discriminative SSL methods facilitate learning with proxy tasks. Among them, contrastive learning has become a promising principle. It uses Siamese networks to minimize the similarity of two augmented views (positive pairs) and maximize the similarity of two different images (negative pairs). These methods are either contrastive or non-contrastive ones: contrastive SSL methods require negative pairs to prevent training collapse; non-contrastive SSL methods are generally more efficient as they maintain remarkable performances using only positive pairs. However, these methods do not perform well on decentralized non-IID data. We analyze their similarities and variances and propose a generalized FedSSL framework.
Federated Learning Federated learning (FL) is a distributed training technique for learning from decentralized parties without transmitting raw data to a central server. Among many studies that address the non-IID data challenge, Personalized FL (PFL) aims to learn personalized models for clients. Although some PFL methods interpolate global and local models, our proposed FedEMA differ in the motivation, application scenario, and measurement of the decay rate. Besides, the majority of existing works only consider supervised learning where clients have fully labeled data. Although recent works propose federated semi-supervised learning or federated domain adaptation, they still need labels in either the server or clients. This paper focuses on purely unlabeled decentralized data.
Federated Unsupervised Learning Learning representations from unlabeled decentralized data while preserving data privacy is still a nascent field. Federated unsupervised representation learning is first proposed by van Berlo et al. (2020) based on autoencoder, but it neglects the non-IID data challenge. Zhang et al. (2020a) address the non-IID issue with potential privacy risk for sharing features. Although Zhuang et al. (2020) address the issue based on BYOL as our FedEMA, they do not shed light on why BYOL works best. Since SSL methods are evolving rapidly and new methods are emerging, we introduce a generalized FedSSL framework and deeply investigate the fundamental components to build up practical guidelines for the generic FedSSL framework.
|
I cannot see significant novelty in both BYOL and FedAvg-like algorithm. Will you fully discuss the relationship to ICCV 2021 paper FedBYOL?
|
Thank you for raising the concern. We would like to clarify that our methods are not a simple combination of BYOL and FedAvg. Although ICCV 2021 paper [1] proposes a new method (FedU) based on BYOL, they do not shed light on why BYOL works best. Since SSL methods are evolving rapidly and new methods (like SimSiam [2], MoCoV2, SimCLR) are emerging, it is crucial to understand the fundamental components of FedSSL. In this paper, we first introduce a generalized FedSSL framework that embraces existing SSL methods and has flexibility catering to future methods. Then, we took an investigative approach to deeply investigate the fundamental components of FedSSL to build up intuitions and practical guidelines for the generic FedSSL framework. Based on these insights, we propose a new model update approach, FedEMA, which improves the performance substantially. We have revised the Related Work section with these discussions. Besides, we would like to point out that FedEMA outperforms [1], as shown in Table 3, 4, 5, and 7 in the manuscript.
[1] Collaborative unsupervised visual representation learning from decentralized data. ICCV 2021.
|
null | false
| null |
Classify each of the following as fast or slow: Toyota Supra, Toyota Corolla, Toyota Highlander
|
Toyota Supra is fast, Toyota Corolla is slow, Toyota Highlander is slow
|
|
null | false
| 506
|
We design an effective alternating optimization framework to minimize the overall objective in equation 2. Specifically, the upper-level problem can be decomposed into N separate sub-problems with fixed generalized models and to be solved on local devices in parallel. Next, we can further settle the inner-level problem to derive the generalized models with fixed personalized models. Since the solution to the sub-problems of the upper-level objective has been well-explored in recent PFL methods, we hereby mainly focus on the inner-level problem. We alternately update the generalized models Ω K and the cluster indicator C K to obtain the optimal generalized models. We view the personalized models, i.e., Θ I = [θ i , ..., θ N ], as private data, and distributionally update the generalized models Ω K on clients with fixed cluster indicator C K . During each server round, the server conducts k-Means clustering on uploaded local parameters Ω t I,R to cluster each client into K disjoint groups, and the clustering results C K are rearranged to the matrix form as
and empty set, respectively), the element (P t ) i,j is defined as 1 |Cj | , or set 0 otherwise. In this way, the elements of every column in P t amount to 1, i.e. N i=1 (P t ) i,j = 1, ∀j, t. When considering the relationship between the consecutive P t , we can formulate the iterate as P t+1 = P t Q t , where Q t ∈ R K×K is a square matrix. We can find that to maintain the above property of P t (∀t), the matrix Q t must satisfies that:
It is noticed that the clustering is based on the latest model parameters Ω t+1 I that depends on Ω t I , and the latest gradient updates given by clients. Hence, P t+1 is determined by and only by P t and Algorithm 1 CGPFL: Personalized Federated Learning with Clustered Generalization
Server sends Ω t K to clients according to P t . 3:
for local device i = 1 to N in parallel do 4:
Initialization:
Local update for the sub-problem of G(ΘI , ΩK ): 6:
for r = 0 to R − 1 do 7:
for s = 0 to S − 1 do 8:
Update the personalized model:
end for 10:
Local update:
end for 12:
end for 13:
Clients send back ω t i,R and server conducts (k-means++) clustering on models Ω t I,R to obtain P t+1 . 14:
Global aggregation:
). 15: end for 16: return The personalized models Θ T I .
Q t . Then we can consider this global iteration as a discrete-time Markov chain and Q t corresponds the transition probability matrix.
During each local round, the clients need to first utilize local datasets to solve the regularized optimization objective, i.e., the upper-level objective in equation 2 with fixed ω t i,r
to obtain a δ-approximate solution θi (ω t i,r ). Then, each client is required to calculate the gradients ∇ ωi G( θi (ω t i,r ), ω t i,r ) with fixed θi (ω t i,r ) and update the model using
) , where β is the learning rate and
), ω t i,r . To reduce the communication overhead, our CGPFL allows the clients to process several iterations before uploading the latest model parameters to the server. The details of CGPFL is given in algorithm 1, from which we can summarize the parameters update process as:
where P t+1 = P t Q t and J t P t = I K (J t ∈ R K×N and I K is an identity matrix), ∀t.
Then we can consider this global iteration as a discrete-time Markov chain and Qt corresponds the transition probability matrix
|
The global iteration is considered a discrete-time Markov chain. However, the transition matrix Qt is not a static matrix. How does the Markov chain handle this dynamic transition matrix? Has these dynamics been analysed in any part of the paper?
|
Sorry that we don't analyze this dynamic transition matrix in this paper. We just notice that the global iteration is a discrete-time Markov chain, but we don't solve the optimization problem from the perspective of Markov chain. We believe that this fact may lead to some other approaches to the personalized federated learning framework which can be a future direction.
|
null | false
| 36
|
The results for each class type are shown in Table TABREF19 , with some specific examples shown in Figure FIGREF18 . For the Wiki(gold) we quote the micro-averaged F-1 scores for the entire top level entity category. The total F-1 score on the OntoNotes dataset is 88%, and the total F-1 cross-validation score on the 112 class Wiki(gold) dataset is 53%. It is worth noting that one could improve Wiki(gold) results by training directly using this dataset. However, the aim is not to tune our model specifically on this class hierarchy. We instead aim to present a framework which can be modified easily to any domain hierarchy and has acceptable out-of-the-box performances to any fine-grained dataset. The results in Table TABREF19 (OntoNotes) only show the main 7 categories in OntoNotes which map to Wiki(gold) for clarity. The other categories (date, time, norp, language, ordinal, cardinal, quantity, percent, money, law) have F-1 scores between 80-90%, with the exception of time (65%)
The total F-1 score on the OntoNotes dataset is 88%, and the total F-1 cross-validation score on the 112 class Wiki(GOLD) dataset is 53%.
|
How many scores do the OntoNotes dataset and Wiki(GOLD) dataset get?
|
The total F-1 score on the OntoNotes dataset is 88%, and the total F-1 cross-validation score on the 112 class Wiki(GOLD) dataset is 53%.
|
null | false
| null |
What is it like to own a dog that sheds everywhere?
|
A dog that sheds everywhere is quite a way to live. You wake up with fur in your mouth. Then when you make the bed, fur flies everywhere. You go to get changed and there is fur all over your clothing. When you brush your teeth, you notice fur in your toothbrush. When you're finally ready for the day, you spend a few extra minutes brushing fur off your clothing and vacuuming up behind you before you start your day. Finally, everything you do in the house likely has fur in it so you just get used to it and learn to live with your fur baby.
|
|
null | false
| null |
In 1839, Russell designed St James Old Cathedral, the oldest building in Melbourne (though moved just outside the city grid in 1914) and one of the very few pre-Gold Rush buildings in the city.
Russell later practised as an architect in Melbourne until he was forced to retire by old age. He kept his mind to the last and died at Richmond, Melbourne, on 10 April 1900, aged 92. He married and was survived by two sons and two daughters.
In addition to being an architect and surveyor Russell did 'extensive work as a sketcher, amateur photographer, etcher, lithographer (and) carver'. He did valuable work as an amateur artist by preserving many original sketches of Melbourne in its early years, in both water-colour and pencil and in 1952 an exhibition of his 'lost paintings' was held at the Tye gallery in Melbourne Russell's artwork is in the collections of the State Library of Victoria, National Gallery of Victoria, the National Library of Australia and the State Library of New South Wales. There is a portrait of him in old age painted c1890 by Frederick McCubbin in the collection of the National Library of Australia.
In 1969, the Victorian chapter of the Australian Institute of Architects named their St. Kilda Road property 'Robert Russell House' in recognition of his work.
|
Given this paragraph about Robert Russell, what is he known for?
|
Rober Russell was an architect, surveyor and amateur artist who designed the oldest building in Melbourne, Australia, i.e., St James Old Cathedral in 1839. His art work including numerous sketches of Melbourne in the 1800s is on display in various galleries and museums in Victoria and New South Wales, Australia, such as the State Library of Victoria, National Gallery of Victoria, the National Library of Australia and the State Library of New South Wales.
|
null | false
| null |
Eudaronia is a genus of sea snails, marine gastropod mollusks in the family Eudaroniidae within the superfamily Seguenzioidea.
|
Extract the family that Eudaronia is part of from the text.
|
Eudaroniidae
|
null | false
| null |
Rowing, sometimes called crew in the United States, is the sport of racing boats using oars. It differs from paddling sports in that rowing oars are attached to the boat using oarlocks, while paddles are not connected to the boat. Rowing is divided into two disciplines: sculling and sweep rowing. In sculling, each rower holds two oars—one in each hand, while in sweep rowing each rower holds one oar with both hands. There are several boat classes in which athletes may compete, ranging from single sculls, occupied by one person, to shells with eight rowers and a coxswain, called eights. There are a wide variety of course types and formats of racing, but most elite and championship level racing is conducted on calm water courses 2 kilometres (1.2 mi) long with several lanes marked using buoys.
|
How many disciplines of rowing are there?
|
There are two primary rowing disciplines: sweep rowing and sculling.
|
null | false
| null |
Guion Stewart Bluford Jr. (born November 22, 1942) is an American aerospace engineer, retired United States Air Force (USAF) officer and fighter pilot, and former NASA astronaut, in which capacity he became the first African American to go to space. While assigned to NASA, he remained a USAF officer rising to the rank of colonel. He participated in four Space Shuttle flights between 1983 and 1992. In 1983, as a member of the crew of the Orbiter Challenger on the mission STS-8, he became the first African American in space as well as the second person of African descent in space, after Cuban cosmonaut Arnaldo Tamayo Méndez.
|
Who was the first African American to go to space?
|
Guion Stewart Bluford Jr. was the first African American to go to space. He was an American aerospace engineer, retired United States Air force officer and fighter pilot, and former NASA astronaut. He participated in for Space Shuttle flights between 1983 and 1992.
|
null | false
| null |
The Crucifixion with the Virgin and St John by Hendrick ter Brugghen is an oil painting, now in the Metropolitan Museum of Art in New York City. It was probably painted c. 1625 as an altarpiece for a Catholic schuilkerk, a "hidden church" or "church in the attic", in the Calvinist Dutch United Provinces, probably Utrecht. When discovered in a bombed out church in South Hackney, London in 1956, it was unknown, but by the time it appeared in Sotheby's salesroom in November of that year it was recognized as an important example of Utrecht Caravaggism. It was acquired by the museum in the sale.
|
Where is the Crucifixion with the Virgin and St John painting located now?
|
The Crucifixion with the Virgin and St John oil painting is located in the Metropolitan Museum of Art in New York City.
|
null | false
| null |
What are some pro tips when cold calling?
|
Cold calling is more art than science. The primary mistake people make when cold calling is in defining what success looks like. For most sales roles, the goal of a cold call is to establish rapport, not to book an order. Think about it - when was the last time you purchased something by phone the first time you were contacted? Odds are that several follow ups were required before you were ready to buy, no matter how qualified a buyer you appeared to be.
Establishing rapport means getting your prospect to open up, because he/she has determined that you are worth the time to do so. So the keys to establishing rapport are simple: be courteous and respectful, provide appropriate context, and be OPEN. Let's consider each of these individually. First, courtesy and respect- after briefly introducing yourself, ask the prospect if now is a good time to talk for a few minutes? We've all experienced the annoyance of a caller who dives right into the conversation immediately after hearing a "hello". Be respectful of your prospect's schedule and give him/her an "out." If it's not a good time, you're not going to have a successful call no matter what you try.
Next, provide appropriate context. Don't assume he/she knows nothing about what you are trying to sell. If that were the case, this prospect wouldn't be a very good target for you! Prepare a short introductory statement that can be delivered in a genuine way with a friendly tone. For example, "I'm with ACME Mobile Window Repair and wanted to share some spring promotions with you - have you seen our trucks around your neighborhood?" Setting context will help build rapport by providing a clear understanding of why you are calling, and what you hope to get out of the conversation. Successful sales people are very comfortable with the fact that they are trying to sell something. Own it, but do so in a low key and friendly way, and this will help lower the defenses of your prospect if he/she sounds guarded
Lastly, be open. This means open to going off script, open to hearing about competition or open to negative impressions of what you're selling. An open dialogue allows room for questions, both from you and from your prospect. Think about an enjoyable conversation you've had lately. Was it a natural back and forth, with questions posed and answers provided? Did the questions take the conversation into a deeper dialogue? A successful cold call is one that feels like a natural conversation about a topic of shared interest. If your prospect goes off on a tangent, don't redirect too soon, or better yet - ask a question - "I'm curious - What made you think of carpets when I mentioned windows?" If he/she goes negative about your offering, ask another question - "What in your experience has caused you to feel this way?"
Cold calling is a skill that can improve with practice. The more comfortable you are being courteous, setting context, and remaining open, the better your sales outcomes will be. And it never hurts to smile - even over the phone!
|
|
null | false
| null |
What would be a first date option in Boston?
|
There are many options for first dates in Boston. You could go for a nice walk in the Boston Common and grab a coffee/tea somewhere nearby. You could go to the Boston Aquarium and walk along the water afterwards. You could go to Seaport and spend time in the indoor mini golf spot. You could also wander around Quincy market and grab a bite to eat and then visit the Sam Adams brewery. Lastly, you could go for a walk along the esplanade and grab a drink and Night Shift brewery if the weather is nice.
|
|
null | false
| 364
|
Automatic dialog/conversation systems have served humans for a long time in various fields, ranging from train routing nbcitetrain to museum guiding nbcitemuseum. In the above scenarios, the dialogs are domain-specific, and a typical approach to such in-domain systems is by human engineering, for example, using manually constructed ontologies nbciteyoungsigdial, natural language templates nbcitetemplate, and even predefined dialog states nbcitestatetracking.
Recently, researchers have paid increasing attention to open-domain, chatbot-style human-computer conversation, because of its important commercial applications, and because it tackles the real challenges of natural language understanding and generation nbciteretrieval1,acl,aaai. For open-domain dialogs, rules and temples would probably fail as we can hardly handle the great diversity of dialog topics and natural language sentences. With the increasing number of human-human conversation utterances available on the Internet, previous studies have developed data-oriented approaches in the open domain, which can be roughly categorized into two groups: retrieval systems and generative systems.
When a user issues an utterance (called a query), retrieval systems search for a most similar query in a massive database (which consists of large numbers of query-reply pairs), and respond to the user with the corresponding reply nbciteretrieval1,retrieval2. Through information retrieval, however, we cannot obtain new utterances, that is, all replies have to appear in the database. Also, the ranking of candidate replies is usually judged by surface forms (e.g., word overlaps, tf $\cdot $ idf features) and hardly addresses the real semantics of natural languages.
Generative dialog systems, on the other hand, can synthesize a new sentence as the reply by language models nbciteBoWdialog,acl,aaai. Typically, a recurrent neural network (RNN) captures the query's semantics with one or a few distributed, real-valued vectors (also known as embeddings); another RNN decodes the query embeddings to a reply. Deep neural networks allow complicated interaction by multiple non-linear transformations; RNNs are further suitable for modeling time-series data (e.g., a sequence of words) especially when enhanced with long short term memory (LSTM) or gated recurrent units (GRUs). Despite these, RNN also has its own weakness when applied to dialog systems: the generated sentence tends to be short, universal, and meaningless, for example, “I don't know” nbcitenaacl or “something” nbciteaaai. This is probably because chatbot-like dialogs are highly diversified and a query may not convey sufficient information for the reply. Even though such universal utterances may be suited in certain dialog context, they make users feel boring and lose interest, and thus are not desirable in real applications.
In this paper, we are curious if we can combine the above two streams of approaches for open-domain conversation. To this end, we propose an ensemble of retrieval and generative dialog systems. Given a user-issued query, we first obtain a candidate reply by information retrieval from a large database. The query, along with the candidate reply, is then fed to an utterance generator based on the “bi-sequence to sequence” (biseq2seq) model nbcitemultiseq2seq. Such sequence generator takes into consideration the information contained in not only the query but also the retrieved reply; hence, it alleviates the low-substance problem and can synthesize replies that are more meaningful. After that we use the scorer in the retrieval system again for post-reranking. This step can filter out less relevant retrieved replies or meaningless generated ones. The higher ranked candidate (either retrieved or generated) is returned to the user as the reply.
From the above process, we see that the retrieval and generative systems are integrated by two mechanisms: (1) The retrieved candidate is fed to the sequence generator to mitigate the “low-substance” problem; (2) The post-reranker can make better use of both the retrieved candidate and the generated utterance. In this sense, we call our overall approach an ensemble in this paper. To the best of our knowledge, we are the first to combine retrieval and generative models for open-domain conversation.
Experimental results show that our ensemble model consistently outperforms each single component in terms of several subjective and objective metrics, and that both retrieval and generative methods contribute an important portion to the overall approach. This also verifies the rationale for building model ensembles for dialog systems.
In this paper, we are curious if we can combine the above two streams of approaches for open-domain conversation. To this end, we propose an ensemble of retrieval and generative dialog systems. Given a userissued query, we first obtain a candidate reply by information retrieval from a large database. The query, along with the candidate reply, is then fed to an utterance generator based on the “bi-sequence to sequence” (biseq2seq) model [30].
|
What model is the utterance generator they use based on?
|
The "bi-sequence to sequence" (biseq2seq) model.
|
null | false
| null |
Tell me if the following foods would be safe or unsafe for someone with a nut allergy to eat? Peanut butter, eggs, apples, pb&j, carrots, pad thai
|
Peanut butter: unsafe, eggs: safe, apples: safe, pb&j: unsafe, carrots: safe, pad thai: unsafe
|
|
null | false
| 73
|
Building effective machine learning models for text requires data and different resources such as pre-trained word embeddings and reusable lexicons. Unfortunately, most of these resources are not entirely transferable to different domains, tasks or languages. Sentiment analysis is one such task that requires additional effort to transfer knowledge between languages.
In this paper, we studied the research question: Can we build reusable sentiment analysis models that can be utilized for making inferences in different languages without requiring separate models and resources for each language? To that end, we built a recurrent neural network model in the language that had largest data available. We took a general-to-specific model building strategy where the larger corpus that had reviews from different domains was first used to train the RNN model and a smaller single-domain corpus of sentiment reviews was used to specialize the model on the given domain. During scoring time, we used corpora for the given domain in different languages and translated them to English to be able to classify sentiments with the trained model. Experimental results showed that the proposed multilingual approach outperforms both the majority baseline and the lexicon-based baseline.
In this paper we made the sentiment analysis model specific to a single domain. For future work, we would like to investigate the effectiveness of our model on different review domains including hotel reviews and on different problems such as detecting stance.
For future work, we would like to investigate the effectiveness of our model on different review domains including hotel reviews and on different problems such as detecting stance.
|
What does the author hope to do about his future work?
|
For future work, they would like to investigate the effectiveness of their model on different review domains including hotel reviews and on different problems such as detecting stance.
|
null | false
| null |
When building a pre-stained shelf from Ikea, classify each of the following as useful or not useful for the task: a wooden dowel, a watercolor paint kit, a windowpane, a rubber mallet, a multitool, a screwdriver.
|
Wooden dowel: useful, a watercolor paint kit: not useful, a windowpane: not useful, a rubber mallet: useful, a multitool: useful, a screwdriver: useful.
|
|
null | false
| null |
Why did my cookies turn out flat?
|
Many factors contribute to a flat cookie. The biggest one tends to be the temperature of the butter when the ingredients are mixed together. If you want a taller, more dense cookie, it is recommended you use cold butter. The best way to do this is to take your sticks of butter and cut it into cubes or slices. Then your mixer will have an easier time incorporating the butter into the sugars and other ingredients. Melted butter will definitely result in a thinner more flat cookie.
|
|
null | false
| null |
Did David Petrarca direct any episodes in the second season of Game of Thrones?
|
Yes, David Petrarca directed two episodes of the second season of Game of Thrones: Garden of Bones and The Ghost of Harrenhal.
|
|
null | false
| null |
What are five totally distinct, creative ways to describe the usefulness of a broken stapler?
|
While a broken stapler may not have many obvious uses, there are a few ways to find utility from this broken piece of office equipment. First, and as long as the stapler is broken, Stephen Root, otherwise known as Milton from Office Space, would no longer have a reason to take revenge on Initech. Secondly, you can use the broken stapler to store additional staples for staplers that are in fact working. A third use, depending on the nature of the break, would be to use the stapler to get a refund from Office Max. Fourth, it could be used as a prank in the office to watch people attempt to staple pieces of paper after using the copy machine (as long as the copy machine, another notoriously broken piece of office equipment, actually works). Lastly, and probably most obvious, a broken stapler could be used as a paper weight.
|
|
null | false
| null |
name some popular songs from the 1970's
|
Stayin' Alive, Y.M.C.A, Let's Get it On, I'll be there and Let it be
|
|
1910.03042
| false
| null |
We assessed the degree of conversational depth by measuring users' mean word count. Prior work has found that an increase in word count has been linked to improved user engagement (e.g., in a social dialog system BIBREF13). For each user conversation, we extracted the overall rating, the number of turns of the interaction, and the user's per-utterance word count (averaged across all utterances). We modeled the relationship between word count and the two metrics of user engagement (overall rating, mean number of turns) in separate linear regressions.
We modeled the relationship between word count and the two metrics of user engagement (overall rating, mean number of turns) in separate linear regressions.
|
What are all the metrics to measure user engagement?
|
The answers are shown as follows:
* overall rating
* mean number of turns
|
null | false
| null |
Swift signed a record deal with Big Machine Records in 2005 and released her eponymous debut album the following year. With 157 weeks on the Billboard 200 by December 2009, the album was the longest-charting album of the 2000s decade. Swift's second studio album, Fearless (2008), topped the Billboard 200 for 11 weeks and was the only album from the 2000s decade to spend one year in the top 10. The album was certified Diamond by the RIAA. It also topped charts in Australia and Canada, and has sold 12 million copies worldwide. Her third studio album, the self-written Speak Now (2010), spent six weeks atop the Billboard 200 and topped charts in Australia, Canada, and New Zealand.
|
From the passage provided, extract the second studio album that Taylor Swift released.
|
Fearless
|
null | false
| null |
What is a dwarf planet ?
|
In August 2006, the International Astronomical Union (IAU) defined a dwarf planet as a celestial body that :
- orbits the Sun,
- has enough mass to assume a hydrostatic equilibrium (nearly round) shape,
- has not "cleared the neighbourhood" around its orbit,
- is not a satellite.
|
|
null | false
| null |
Tell me whether these are names of soft drinks or coffee: Starbucks, Peet's, Blue Bottle, Coca-Cola, Sprite, Canada Dry, Pepsi
|
Starbucks is a brand of coffee
Peet's is a brand of coffee
Blue Bottle is a brand of coffee
Coca-Cola is a soft drink
Sprite is a soft drink
Canada Dry is a soft drink
Pepsi is a soft drink
|
|
null | false
| 0
|
We evaluate our cross-lingual pre-training based transfer approach against several strong baselines on two public datatsets, Europarl BIBREF31 and MultiUN BIBREF32, which contain multi-parallel evaluation data to assess the zero-shot performance. In all experiments, we use BLEU as the automatic metric for translation evaluation.
The statistics of Europarl and MultiUN corpora are summarized in Table TABREF18. For Europarl corpus, we evaluate on French-English-Spanish (Fr-En-Es), German-English-French (De-En-Fr) and Romanian-English-German (Ro-En-De), where English acts as the pivot language, its left side is the source language, and its right side is the target language. We remove the multi-parallel sentences between different training corpora to ensure zero-shot settings. We use the devtest2006 as the validation set and the test2006 as the test set for Fr$\rightarrow $Es and De$\rightarrow $Fr. For distant language pair Ro$\rightarrow $De, we extract 1,000 overlapping sentences from newstest2016 as the test set and the 2,000 overlapping sentences split from the training set as the validation set since there is no official validation and test sets. For vocabulary, we use 60K sub-word tokens based on Byte Pair Encoding (BPE) BIBREF33.
For MultiUN corpus, we use four languages: English (En) is set as the pivot language, which has parallel data with other three languages which do not have parallel data between each other. The three languages are Arabic (Ar), Spanish (Es), and Russian (Ru), and mutual translation between themselves constitutes six zero-shot translation direction for evaluation. We use 80K BPE splits as the vocabulary. Note that all sentences are tokenized by the tokenize.perl script, and we lowercase all data to avoid a large vocabulary for the MultiUN corpus.
We use traditional transfer learning, pivot-based method and multilingual NMT as our baselines. For the fair comparison, the Transformer-big model with 1024 embedding/hidden units, 4096 feed-forward filter size, 6 layers and 8 heads per layer is adopted for all translation models in our experiments. We set the batch size to 2400 per batch and limit sentence length to 100 BPE tokens. We set the $\text{attn}\_\text{drop}=0$ (a dropout rate on each attention head), which is favorable to the zero-shot translation and has no effect on supervised translation directions BIBREF22. For the model initialization, we use Facebook's cross-lingual pretrained models released by XLM to initialize the encoder part, and the rest parameters are initialized with xavier uniform. We employ the Adam optimizer with $\text{lr}=0.0001$, $t_{\text{warm}\_\text{up}}=4000$ and $\text{dropout}=0.1$. At decoding time, we generate greedily with length penalty $\alpha =1.0$.
Regarding MLM, TLM and BRLM, as mentioned in the pre-training phase of transfer protocol, we first pre-train MLM on monolingual data of both source and pivot languages, then leverage the parameters of MLM to initialize TLM and the proposed BRLM, which are continued to be optimized with source-pivot bilingual data. In our experiments, we use MLM+TLM, MLM+BRLM to represent this training process. For the masking strategy during training, following BIBREF29, $15\%$ of BPE tokens are selected to be masked. Among the selected tokens, $80\%$ of them are replaced with [MASK] token, $10\%$ are replaced with a random BPE token, and $10\%$ unchanged. The prediction accuracy of masked words is used as a stopping criterion in the pre-training stage. Besides, we use fastalign tool BIBREF34 to extract word alignments for BRLM-HA.
Table TABREF19 and TABREF26 report zero-shot results on Europarl and Multi-UN evaluation sets, respectively. We compare our approaches with related approaches of pivoting, multilingual NMT (MNMT) BIBREF19, and cross-lingual transfer without pretraining BIBREF16. The results show that our approaches consistently outperform other approaches across languages and datasets, especially surpass pivoting, which is a strong baseline in the zero-shot scenario that multilingual NMT systems often fail to beat BIBREF19, BIBREF20, BIBREF23. Pivoting translates source to pivot then to target in two steps, causing inefficient translation process. Our approaches use one encoder-decoder model to translate between any zero-shot directions, which is more efficient than pivoting. Regarding the comparison between transfer approaches, our cross-lingual pretraining based transfer outperforms transfer method that does not use pretraining by a large margin.
Regarding comparison between the baselines in table TABREF19, we find that pivoting is the strongest baseline that has significant advantage over other two baselines. Cross-lingual transfer for languages without shared vocabularies BIBREF16 manifests the worst performance because of not using source$\leftrightarrow $pivot parallel data, which is utilized as beneficial supervised signal for the other two baselines.
Our best approach of MLM+BRLM-SA achieves the significant superior performance to all baselines in the zero-shot directions, improving by 0.9-4.8 BLEU points over the strong pivoting. Meanwhile, in the supervised direction of pivot$\rightarrow $target, our approaches performs even better than the original supervised Transformer thanks to the shared encoder trained on both large-scale monolingual data and parallel data between multiple languages.
MLM alone that does not use source$\leftrightarrow $pivot parallel data performs much better than the cross-lingual transfer, and achieves comparable results to pivoting. When MLM is combined with TLM or the proposed BRLM, the performance is further improved. MLM+BRLM-SA performs the best, and is better than MLM+BRLM-HA indicating that soft alignment is helpful than hard alignment for the cross-lingual pretraining.
Like experimental results on Europarl, MLM+BRLM-SA performs the best among all proposed cross-lingual pretraining based transfer approaches as shown in Table TABREF26. When comparing systems consisting of one encoder-decoder model for all zero-shot translation, our approaches performs significantly better than MNMT BIBREF19.
Although it is challenging for one model to translate all zero-shot directions between multiple distant language pairs of MultiUN, MLM+BRLM-SA still achieves better performances on Es $\rightarrow $ Ar and Es $\rightarrow $ Ru than strong pivoting$_{\rm m}$, which uses MNMT to translate source to pivot then to target in two separate steps with each step receiving supervised signal of parallel corpora. Our approaches surpass pivoting$_{\rm m}$ in all zero-shot directions by adding back translation BIBREF33 to generate pseudo parallel sentences for all zero-shot directions based on our pretrained models such as MLM+BRLM-SA, and further training our universal encoder-decoder model with these pseudo data. BIBREF22 gu2019improved introduces back translation into MNMT, while we adopt it in our transfer approaches. Finally, our best MLM+BRLM-SA with back translation outperforms pivoting$_{\rm m}$ by 2.4 BLEU points averagely, and outperforms MNMT BIBREF22 by 4.6 BLEU points averagely. Again, in supervised translation directions, MLM+BRLM-SA with back translation also achieves better performance than the original supervised Transformer.
We first evaluate the representational invariance across languages for all cross-lingual pre-training methods. Following BIBREF23, we adopt max-pooling operation to collect the sentence representation of each encoder layer for all source-pivot sentence pairs in the Europarl validation sets. Then we calculate the cosine similarity for each sentence pair and average all cosine scores. As shown in Figure FIGREF27, we can observe that, MLM+BRLM-SA has the most stable and similar cross-lingual representations of sentence pairs on all layers, while it achieves the best performance in zero-shot translation. This demonstrates that better cross-lingual representations can benefit for the process of transfer learning. Besides, MLM+BRLM-HA is not as superior as MLM+BRLM-SA and even worse than MLM+TLM on Fr-En, since MLM+BRLM-HA may suffer from the wrong alignment knowledge from an external aligner tool. We also find an interesting phenomenon that as the number of layers increases, the cosine similarity decreases.
We further sample an English-Russian sentence pair from the MultiUN validation sets and visualize the cosine similarity between hidden states of the top encoder layer to further investigate the difference of all cross-lingual pre-training methods. As shown in Figure FIGREF38, the hidden states generated by MLM+BRLM-SA have higher similarity for two aligned words. It indicates that MLM+BRLM-SA can gain better word-level representation alignment between source and pivot languages, which better relieves the burden of the domain shift problem.
To freeze parameters is a common strategy to avoid catastrophic forgetting in transfer learning BIBREF27. Table TABREF43 shows the performance of transfer learning with freezing different layers on MultiUN test set, in which En$\rightarrow $Ru denotes the parent model, Ar$\rightarrow $Ru and Es$\rightarrow $Ru are two child models, and all models are based on MLM+BRLM-SA. We can find that updating all parameters during training will cause a notable drop on the zero-shot direction due to the catastrophic forgetting. On the contrary, freezing all the parameters leads to the decline on supervised direction because the language features extracted during pre-training is not sufficient for MT task. Freezing the first four layers of the transformer shows the best performance and keeps the balance between pre-training and fine-tuning.
We evaluate our cross-lingual pre-training based transfer approach against several strong baselines on two public datatsets, Europarl (Koehn 2005) and MultiUN (Eisele and Chen 2010), which contain multi-parallel evaluation data to assess the zero-shot performance.
|
Which two public datasets were used in the paper?
|
Europarl (Koehn 2005) and MultiUN (Eisele and Chen 2010).
|
null | false
| null |
What are some fun activities to do in Maui, Hawaii?
|
Maui, Hawaii is a beautiful place to visit and offers activities that appeal to adventurers, families, and those who want to relax.
Maui's amazing beaches offer surfing, outrigger paddling, stand-up paddling, fishing, kite surfing and other activities for novice or experienced water lovers. More of a land-lubber than a mermaid? Take a hike in the Haleakala volcano crater or one of many other hikes around the island.
Relaxing by the pool more your speed, check out the great hotels or many private house rentals that offer vacationers a chance to unplug and get away from it all. Most of the resorts are closely located to shops and restaurants for a no worries kid of trip.
Whatever your passions are, Maui delivers with aloha!
|
|
null | false
| null |
why wax snowboard?
|
Waxing snowboard improves the speed of the skiing by reducing friction
|
|
null | false
| null |
Explain the black and white smoke during the Pope election
|
Beginning in the early 1800s, the ballots used by cardinals were burned after each ballot to indicate a failed election. The lack of smoke instead signalled a successful election. Since 1914, black smoke (fumata nera) emerging from a temporary chimney installed on the roof of the Sistine Chapel indicates that the ballot did not result in an election, while white smoke (fumata bianca) announces that a new pope has been chosen.
Prior to 1945 (when Pius XII changed the form of ballot to use unsigned ballots, first carried out in 1958), the sealing wax on the complex type ballots illustrated above had the effect of making the smoke from burning the ballots either black or white, depending on whether or not damp straw was added. Until the 20th century, sealing wax customarily had beeswax mixed into its composition. The use of wax made solely from animal fat does not give as much white colored smoke, as does wax that includes beeswax. In the 1939 conclave there was some confusion over the smoke color, which was even more apparent in the 1958 conclave. The lack of sealing wax on the ballots explains the confusion over the color of the smoke in the 1958 conclave. The Siri thesis was based on the confusion over the smoke color on the first day of that conclave.
Since 1963, chemicals have been added to the burning process to augment the smoke's black or white color. Beginning in 2005, a successful election is also accentuated by bells ringing at the appearance of the white smoke.
During the 2013 conclave, the Vatican disclosed the chemicals used to color the smoke—the black smoke comprises potassium perchlorate, anthracene, and sulfur; the white smoke is composed of potassium chlorate, lactose, and pine rosin.
|
|
null | false
| 104
|
Since the CoNLL annotations have 21 semantic roles in total, we use 21 roles in our model as well as the baseline. Following garg2012unsupervised, we set the number of PRs to 2 (excluding INLINEFORM0 , INLINEFORM1 and INLINEFORM2 ), and SRs to 21-2=19. Table TABREF27 shows the results.
In the first setting (Line 1), we train and test the monolingual model on the CoNLL data. We observe significant improvements in F1 score over the Baseline (Line 0) in both languages. Using the CoNLL 2009 dataset alone, titovcrosslingual report an F1 score of 80.9% (PU=86.8%, CO=75.7%) for German. Thus, our monolingual model outperforms their monolingual model in German. For English, they report an F1 score of 83.6% (PU=87.5%, CO=80.1%), but note that our English results are not directly comparable to theirs due to differences argument identification, as discussed in section SECREF25 . As their argument identification score is lower, perhaps their system is discarding “difficult” arguments which leads to a higher clustering score.
In the second setting (Line 2), we use the additional monolingual Europarl (EP) data for training. We get equivalent results in English and a significant improvement in German compared to our previous setting (Line 1). The German dataset in CoNLL is quite small and benefits from the additional EP training data. In contrast, the English model is already quite good due to a relatively big dataset from CoNLL, and good accuracy syntactic parsers. Unfortunately, titovcrosslingual do not report results with this setting.
The third setting (Line 3) gives the results of our multilingual model, which adds the word alignments in the EP data. Comparing with Line 2, we get non-significant improvements in both languages. titovcrosslingual obtain an F1 score of 82.7% (PU=85.0%, CO=80.6%) for German, and 83.7% (PU=86.8%, CO=80.7%) for English. Thus, for German, our multilingual Bayesian model is able to capture the cross-lingual patterns at least as well as the external penalty term in BIBREF6 . We cannot compare the English results unfortunately due to differences in argument identification.
We also compared monolingual and bilingual training data using a setting that emulates the standard supervised setup of separate training and test data sets. We train only on the EP dataset and test on the CoNLL dataset. Lines 4 and 5 of Table TABREF27 give the results. The multilingual model obtains small improvements in both languages, which confirms the results from the standard unsupervised setup, comparing lines 2 to 3.
These results indicate that little information can be learned about semantic roles from this parallel data setup. One possible explanation for this result is that the setup itself is inadequate. Given the definition of aligned arguments, only 8% of English arguments and 17% of German arguments are aligned. This plus our experiments suggest that improving the alignment model is a necessary step to making effective use of parallel data in multilingual SRI, for example by joint modeling with SRI. We leave this exploration to future work.
Since the CoNLL annotations have 21 semantic roles in total, we use 21 roles in our model as well as the baseline.
|
How many semantic roles do the CoNLL annotations have?
|
21.
|
null | false
| null |
Give me a list of some of the best ski areas spread across the North America. Organize the list by the mountain region they belong to. Give a very short synopsis if desired about why that resort was selected.
|
Any list of top Ski resorts is one that will be very contested among skiers, but here are a few that stand out:
Pacific Northwest:
Whistler Blackcomb - One of the largest ski resorts in the world, peak to peak gondola, and a vibrant base area village
Mount Baker - Large amount of snow, one of the few remaining locally owned resorts
Crystal Mountain - The largest ski area in Washington state, impressive views of Mt Rainier
Mt Bachelor - Ski lift to the summit and can ski circumnavigate the base of the mountain
Canadian Rockies:
Revelstoke - Highest vertical lift in Canada
Red Mountain - Remote and short lift lines
US Rockies:
Alta - Famous Skiers only mountain with expansive bowls
Park City - Large open terrain with a vibrant village
Crested Butte - Some of the steepest most challenging terrain in the world
Big Sky - Peak gondola to access some of the most challenging terrain in the US
Jackson Hole - Some of the best and most abundant powder and home of famous Corbetts Couloir
Sierra range:
Tahoe/Palisades: Large resort with expansive views of Lake Tahoe
East Coast:
Stowe Mountain - One of the oldest ski resorts in the US, home of the 10th mountain division
Burke Mountain - Home of the US Ski Race team
In total, there are 826 ski resorts in North America - all are loved for different reasons and have passionate groups of skiers and snowboarders that consider them home, and might argue why their mountain is the best mountain.
|
|
null | false
| 102
|
We compare with the following models:
1. BIBREF0 : This is a conditional language model which uses a feed-forward neural network to predict the next word in the description conditioned on local characteristics (i.e., words within a field) and global characteristics (i.e., overall structure of the infobox).
2. BIBREF9 : This model was proposed in the context of the WeatherGov and RoboCup datasets which have a much smaller vocabulary. They use an improved attention model with additional regularizer terms which influence the weights assigned to the fields.
3. Basic Seq2Seq: This is the vanilla encode-attend-decode model BIBREF1 . Further, to deal with the large vocabulary ( INLINEFORM0 400K words) we use a copying mechanism as a post-processing step. Specifically, we identify the time steps at which the decoder produces unknown words (denoted by the special symbol UNK). For each such time step, we look at the attention weights on the input words and replace the UNK word by that input word which has received maximum attention at this timestep. This process is similar to the one described in BIBREF30 . Even lebret2016neural have a copying mechanism tightly integrated with their model.
We compare with the following models: 1. (Lebret et al., 2016): This is a conditional language model which uses a feed-forward neural network to predict the next word in the description words within a field) and global characteristics (i.e., overall structure of the infobox). 2.(Mei et al., 2016): This model was proposed in the context of the WeatherGov and RoboCup datasets which have a much smaller vocabulary. They use an improved attention model with additional regularizer terms which influence the weights assigned to the fields. 3.Basic Seq2Seq: This is the vanilla encode- attend-decode model (Bahdanau et al., 2014). Fur_x001f_ther, to deal with the large vocabulary (~400K words) we use a copying mechanism as a post processing step. Specifically, we identify the time steps at which the decoder produces unknown words (denoted by the special symbol UNK). For each such time step, we look at the attention weights on the input words and replace the UNK word by that input word which has received max imum attention at this timestep. This process is similar to the one described in (Luong et al., 2015). Even Lebret et al. (2016) have a copying mechanism tightly integrated with their model.
|
What baseline model is the proposed model compared to?
|
They compared their models to (Lebret et al., 2016), (Mei et al., 2016), andBasic Seq2Seq.
|
null | false
| 164
|
Preparing the data can be a complex and time-consuming process, often involving working with partially or wholly unstructured data. The pre-processing steps have a big impact on the operationalizations, subsequent analyses and reproducibility efforts BIBREF30 , and they are usually tightly linked to what we intend to measure. Unfortunately, these steps tend to be underreported, but documenting the pre-processing choices made is essential and is analogous to recording the decisions taken during the production of a scholarly edition or protocols in biomedical research. Data may also vary enormously in quality, depending on how it has been generated. Many historians, for example, work with text produced from an analogue original using Optical Character Recognition (OCR). Often, there will be limited information available regarding the accuracy of the OCR, and the degree of accuracy may even vary within a single corpus (e.g. where digitized text has been produced over a period of years, and the software has gradually improved). The first step, then, is to try to correct for common OCR errors. These will vary depending on the type of text, the date at which the `original' was produced, and the nature of the font and typesetting.
One step that almost everyone takes is to tokenize the original character sequence into the words and word-like units. Tokenization is a more subtle and more powerful process than people expect. It is often done using regular expressions or scripts that have been circulating within the NLP community. Tokenization heuristics, however, can be badly confused by emoticons, creative orthography (e.g., U$A, sh!t), and missing whitespace. Multi-word terms are also challenging. Treating them as a single unit can dramatically alter the patterns in text. Many words that are individually ambiguous have clear, unmistakable meanings as terms, like “black hole" or “European Union". However, deciding what constitutes a multi-word term is a difficult problem. In writing systems like Chinese, tokenization is a research problem in its own right.
Beyond tokenization, common steps include lowercasing, removing punctuation, stemming (removing suffixes), lemmatization (converting inflections to a base lemma), and normalization, which has never been clearly defined, but often includes grouping abbreviations like “U.S.A." and “USA", ordinals like “1st" and “first", and variant spellings like “noooooo". The main goal of these steps is to improve the ratio of tokens (individual occurrences) to types (the distinct things in a corpus). Each step requires making additional assumptions about which distinctions are relevant: is “apple” different from “Apple”? Is “burnt” different from “burned”? Is “cool" different from “coooool"? Sometimes these steps can actively hide useful patterns, like social meaning BIBREF32 . Some of us therefore try do as little modification as possible.
From a multilingual perspective, English and Chinese have an unusually simple inflectional system, and so it is statistically reasonable to treat each inflection as a unique word type. Romance languages have considerably more inflections than English; many indigenous North American languages have still more. For these languages, unseen data is far more likely to include previously-unseen inflections, and therefore, dealing with inflections is more important. On the other hand, the resources for handling inflections vary greatly by language, with European languages dominating the attention of the computational linguistics community thus far.
We sometimes also remove words that are not relevant to our goals, for example by calculating vocabulary frequencies. We construct a “stoplist” of words that we are not interested in. If we are looking for semantic themes we might remove function words like determiners and prepositions. If we are looking for author-specific styles, we might remove all words except function words. Some words are generally meaningful but too frequent to be useful within a specific collection. We sometimes also remove very infrequent words. Their occurrences are too low for robust patterns and removing them helps reducing the vocabulary size.
The choice of processing steps can be guided by theory or knowledge about the domain as well as experimental investigation. When we have labels, predictive accuracy of a model is a way to assess the effect of the processing steps. In unsupervised settings, it is more challenging to understand the effects of different steps. Inferences drawn from unsupervised settings can be sensitive to pre-processing choices BIBREF33 . Stemming has been found to provide little measurable benefits for topic modeling and can sometimes even be harmful BIBREF34 . All in all, this again highlights the need to document these steps.
Finally, we can also mark up the data, e.g., by identifying entities (people, places, organizations, etc.) or parts of speech. Although many NLP tools are available for such tasks, they are often challenged by linguistic variation, such as orthographic variation in historical texts BIBREF35 and social media BIBREF32 . Moreover, the performance of NLP tools often drops when applying them outside the training domain, such as applying tools developed on newswire texts to texts written by younger authors BIBREF36 . Problems (e.g., disambiguation in named entity recognition) are sometimes resolved using considerable manual intervention. This combination of the automated and the manual, however, becomes more difficult as the scale of the data increases, and the `certainty' brought by the latter may have to be abandoned.
Preparing the data can be a complex and time consuming process, often involving working with partially or wholly unstructured data.
|
Could data preparing be a complex and time consuming process?
|
Yes.
|
null | false
| null |
What are some items that you might see in a fridge?
|
Some typical items that you might find in a refrigerator include milk, butter, cheese, yoghurt, juice, eggs and cold water.
|
|
null | false
| null |
Britain's income tax has changed over the years. Originally it taxed a person's income regardless of who was beneficially entitled to that income, but now tax is paid on income to which the taxpayer is beneficially entitled. Most companies were taken out of the income tax net in 1965 when corporation tax was introduced. These changes were consolidated by the Income and Corporation Taxes Act 1970. Also the schedules under which tax is levied have changed. Schedule B was abolished in 1988, Schedule C in 1996 and Schedule E in 2003. For income tax purposes, the remaining schedules were superseded by the Income Tax (Trading and Other Income) Act 2005, which also repealed Schedule F. For corporation tax purposes, the Schedular system was repealed and superseded by the Corporation Tax Acts of 2009 and 2010. The highest rate of income tax peaked in the Second World War at 99.25%. This was slightly reduced after the war and was around 97.5 percent (nineteen shillings and sixpence in the pound) through the 1950s and 60s.
HM Revenue and Customs has published online a comprehensive set of manuals about the UK tax system.
Tax revenues as a percentage of GDP for the UK in comparison to the OECD and the EU 15
In 1971, the top rate of income tax on earned income was cut to 75%. A surcharge of 15% on investment income kept the overall top rate on that income at 90%. In 1974 the top tax rate on earned income was again raised, to 83%. With the investment income surcharge this raised the overall top rate on investment income to 98%, the highest permanent rate since the war. This applied to incomes over £20,000 (equivalent to £221,741 in 2021 terms),. In 1974, as many as 750,000 people were liable to pay the top rate of income tax. Margaret Thatcher, who favoured indirect taxation, reduced personal income tax rates during the 1980s. In the first budget after her election victory in 1979, the top rate was reduced from 83% to 60% and the basic rate from 33% to 30%. The basic rate was further cut in three subsequent budgets, to 29% in 1986 budget, 27% in 1987 and 25% in 1988. The top rate of income tax was cut to 40% in the 1988 budget. The investment income surcharge was abolished in 1985.
Subsequent governments reduced the basic rate further, to the present level of 20% in 2007. Since 1976 (when it stood at 35%), the basic rate has been reduced by 15%, but this reduction has been largely offset by increases in national insurance contributions and value added tax.
In 2010 a new top rate of 50% was introduced on income over £150,000. Revenue to the Exchequer subsequently went down as top rate earners found methods to avoid taxation. In the 2012 budget this rate was cut to 45% for 2013–14; this was followed by an increase in the tax paid by additional rate taxpayers from £38 billion to £46 billion. Chancellor George Osborne claimed that the lower, more competitive tax rate had caused the increase.
In September 2022 the government announced that from April 2023 the top rate of tax would be reduced from 45% to 40% and the basic rate reduced from 20% to 19%. The abolition of the 45% additional rate of tax was subsequently cancelled through measures set out in the 2022's Autumn statement.
Business rates were introduced in England and Wales in 1990 and are a modernised version of a system of rating that dates back to the Elizabethan Poor Law of 1601. As such, business rates retain many previous features from, and follow some case law of, older forms of rating. The Finance Act 2004 introduced an income tax regime known as "pre-owned asset tax" which aims to reduce the use of common methods of inheritance tax avoidance.
|
what is the latest top rate of income tax
|
45% above 150,000 as of Aug 2022
|
null | false
| null |
Give me a list of Serie A winners in the last 5 years.
|
2021-2022: AC Milan
2020-2021: Inter
2019-2020: Juventus
2018-2019: Juventus
2017-2018: Juventus
|
|
null | false
| null |
Are rugby and American football similar?
|
In many regards, yes. In many regards, no. The main similarities are that ball-carrying players are tackled by the opposing team to keep them from scoring the ball in the end zone. The ball is a similar shape, but a different size. The main differences are in the number of players on the field at any given time; 15 players per side in rugby. 11 players per side in football. Also, in football, a player can advance the ball by throwing it forward to a teammate. In rugby, the ball must travel laterally or backward to a teammate by hand. The ball can only be advanced by running or kicking the ball.
|
|
null | false
| 213
|
According to The National Institute on Deafness, one in thousand infants is born deaf. An additional one to six per thousand are born with hearing loss at different levels BIBREF0. Sign language is commonly used by Deaf and Hard-of-Hearing (DHH) persons to communicate via hand gestures. An automatic sign language recognizer enables an ASL user to translate the sign language to written text or speech, allowing them to communicate with people who are not familiar with ASL. There is a tremendous rise in the popularity of personal digital assistants; available on user's personal and wearable devices (Google Now, Amazon Alexa and Apple Siri, etc.) and also in the form of standalone devices (Amazon Echo and Google Home smart speakers). These devices are primarily controlled through voice, and hence, their functionality is not readily available to DHH users. An automatic sign recognizer can also enable the interaction between a DHH user and a digital assistant.
Most current systems have capability of ASL recognition with RGB video data BIBREF1, BIBREF2, BIBREF3. An ASL sign is performed by a combination of hand gestures, facial expressions and postures of the body. Sequential motion of specific body locations (such as hand-tip, neck and arm) provide informative cues about a sign. Using video data, it is difficult to extract different body locations and associated motion sequences from a series of RGB frames. Microsoft Kinect is a 3D camera sensor which can use the depth information of a person to capture 3D coordinates of his/her body location across a video. This sequence of 3D body location is referred by skeletal data BIBREF4. To the best of our knowledge, there is no publicly available skeletal dataset in literature for ASL recognition.
With skeletal data, an ASL sign can be seen as a sequence of 3D coordinates or a 3D time series BIBREF5. Recurrent neural networks (RNN) have shown strong performance for sequential modeling BIBREF6. In this work, we investigate the impact of RGB video data in recognition accuracy when combined with skeletal data. We also propose a combined RNN network with a simple spatial data augmentation technique. In summary, the contributions of this work are:
We propose an RNN architecture with a novel spatial data augmentation technique.
We propose an architecture which uses both RGB and skeletal data to improve recognition accuracy.
We introduce and publicly release a multi–modal dataset for ASL called GMU-ASL51.
We also propose a combined RNN network with a simple spatial data augmentation technique.
|
What kind of technique do they propose?
|
A combined RNN network with a simple spatial data augmentation technique.
|
null | false
| null |
Identify which animal species is alive or extinct: Ganges River Dolphin, Helicoprion
|
Helicoprion is extinct, Ganges River Dolphin is alive.
|
|
null | false
| null |
which weighs more, a pound of bricks or a pound of feathers?
|
Both things weigh the same amount, one pound.
|
|
1807.07961
| false
| null |
In order to obtain insights about why the more fine-grained bi-sense emoji embedding helps in understanding the complexed sentiments behind tweets, we visualize the attention weights for ATT-E-LSTM and MATT-BiE-LSTM for comparison. The example tweets with corresponding attention weights calculated by word-emoji embedding and senti-emoji embedding are shown in Figure FIGREF27 , where the contexts are presented in the captions. The emojis used are , , and , respectively.
Therefore, we construct the new input INLINEFORM0 to each LSTM unit by concatenating the original word embedding and the attention vector in Equation EQREF21 to distribute the senti-emoji information to each step. This model is called Multi-level Attention-based LSTM with Bi-sense Emoji Embedding (MATT-BiE-LSTM). We choose the same binary cross-entropy as the loss function with the same network configuration with WATT-BiE-LSTM. DISPLAYFORM0
Attention-based LSTM with emojis:We also use the word-emoji embedding to calculate the emoji-word attention following Equation EQREF20 and EQREF21 , and the only difference is that we replace the attention-derived senti-emoji embedding with the pre-trained word-emoji embedding by fasttext, denoted as ATT-E-LSTM.
In Figure FIGREF27 (a), the ATT-E-LSTM model (baseline) assigns relatively more weights on the word “no” and “pressure”, while MATT-BiE-LSTM attends mostly on the word “happy” and “lovely”. The different attention distributions suggest that the proposed senti-emoji embedding is capable of recognizing words with strong sentiments that are closely related to the true sentiment even with the presence of words with conflicting sentiments, such as “pressure” and “happy”. while ATT-E-LSTM tends to pick up all sentimental words which could raise confusions. The senti-emoji embedding is capable of extracting representations of complexed semantics and sentiments which help guide the attentions even in cases when the word sentiment and emoji sentiment are somewhat contradictory to each other. From Figure FIGREF27 (b) and (c) we can observe that the ATT-E-LSTM assigns more weights on the sentiment-irrelevant words than the MATT-BiE-LSTM such as “hoodies”, “wait” and “after”, indicating that the proposed model is more robust to irrelevant words and concentrates better on important words. Because of the senti-emoji embedding obtained through bi-sense emoji embedding and the sentence-level LSTM encoding on the text input (described in Section SECREF13 ), we are able to construct a more robust embedding based on the semantic and sentiment information from the whole context compared to the word-emoji embedding used in ATT-E-LSTM which takes only word-level information into account.
In order to obtain insights about why the more fine-grained bi-sense emoji embedding helps in understanding the complexed sentiments behind tweets, we visualize the attention weights for ATT-E-LSTM and MATT-BiE-LSTM for comparison.
Therefore, we construct the new input INLINEFORM0 to each LSTM unit by concatenating the original word embedding and the attention vector in Equation EQREF21 to distribute the senti-emoji information to each step. This model is called Multi-level Attention-based LSTM with Bi-sense Emoji Embedding (MATT-BiE-LSTM)
Attention-based LSTM with emojis:We also use the word-emoji embedding to calculate the emoji-word attention following Equation EQREF20 and EQREF21 , and the only difference is that we replace the attention-derived senti-emoji embedding with the pre-trained word-emoji embedding by fasttext, denoted as ATT-E-LSTM.
In Figure FIGREF27 (a), the ATT-E-LSTM model (baseline) assigns relatively more weights on the word “no” and “pressure”, while MATT-BiE-LSTM attends mostly on the word “happy” and “lovely”. The different attention distributions suggest that the proposed senti-emoji embedding is capable of recognizing words with strong sentiments that are closely related to the true sentiment even with the presence of words with conflicting sentiments, such as “pressure” and “happy”. while ATT-E-LSTM tends to pick up all sentimental words which could raise confusions. The senti-emoji embedding is capable of extracting representations of complexed semantics and sentiments which help guide the attentions even in cases when the word sentiment and emoji sentiment are somewhat contradictory to each other. From Figure FIGREF27 (b) and (c) we can observe that the ATT-E-LSTM assigns more weights on the sentiment-irrelevant words than the MATT-BiE-LSTM such as “hoodies”, “wait” and “after”, indicating that the proposed model is more robust to irrelevant words and concentrates better on important words. Because of the senti-emoji embedding obtained through bi-sense emoji embedding and the sentence-level LSTM encoding on the text input (described in Section SECREF13 ), we are able to construct a more robust embedding based on the semantic and sentiment information from the whole context compared to the word-emoji embedding used in ATT-E-LSTM which takes only word-level information into account.
|
What evidence does visualizing the attention give to show that it helps to obtain a more robust understanding of semantics and sentiments?
|
The answers are shown as follows:
* The different attention distributions suggest that the proposed senti-emoji embedding is capable of recognizing words with strong sentiments that are closely related to the true sentiment even with the presence of words with conflicting sentiments
|
null | false
| null |
Christopher Crosby Farley was born on February 15, 1964, in Madison, Wisconsin, and grew up in Maple Bluff. His father, Thomas John Farley Sr. (1936–99), owned an oil company, and his mother, Mary Anne (née Crosby), was a homemaker. He had four siblings: Tom Jr., Kevin, John, and Barbara. His cousin, Jim, is chief executive officer at Ford Motor Company.
The Farleys are Irish Catholics. Chris attended parochial schools in Madison, including Edgewood High School of the Sacred Heart. According to Joel Murray, a fellow Second City cast member, Chris would "always make it to Mass."
Many of his summers were spent as a camper and counselor at Red Arrow Camp, near Minocqua, Wisconsin. He graduated from Marquette University in 1986, with a double major in communications and theater. At Marquette, he played rugby union and discovered a love of comedy.
After college, he worked with his father at the Scotch Oil Company in Madison. He first learned the art of improvisational comedy at the Ark Improv Theatre in Madison.
Farley made his way to Chicago, performing first at Improv Olympic. He then attended Chicago's Second City Theatre, starting the same day as Stephen Colbert, initially as part of Second City's touring group. He was eventually promoted to their main stage in 1989, and was a cast member of three revues, The Gods Must Be Lazy, It Was Thirty Years Ago Today, and Flag Smoking Permitted in Lobby Only or Censorama.
|
Please extract the names of each of the Second City Theatre venues in which Chris Farley was a cast member, separated by commas.
|
The Gods Must Be Lazy, It Was Thirty Years Ago Today, Flag Smoking Permitted in Lobby Only or Censorama
|
null | false
| null |
The system was proclaimed complete in 1992, but two of the original Interstates—I-95 and I-70—were not continuous: both of these discontinuities were due to local opposition, which blocked efforts to build the necessary connections to fully complete the system. I-95 was made a continuous freeway in 2018, and thus I-70 remains the only original Interstate with a discontinuity.
I-95 was discontinuous in New Jersey because of the cancellation of the Somerset Freeway. This situation was remedied when the construction of the Pennsylvania Turnpike/Interstate 95 Interchange Project started in 2010 and partially opened on September 22, 2018, which was already enough to fill the gap.
However, I-70 remains discontinuous in Pennsylvania, because of the lack of a direct interchange with the Pennsylvania Turnpike at the eastern end of the concurrency near Breezewood. Traveling in either direction, I-70 traffic must exit the freeway and use a short stretch of US-30 (which includes a number of roadside services) to rejoin I-70. The interchange was not originally built because of a legacy federal funding rule, since relaxed, which restricted the use of federal funds to improve roads financed with tolls. Solutions have been proposed to eliminate the discontinuity, but they have been blocked by local opposition, fearing a loss of business.
|
Given this paragraph, which two US interstates were not originally continuous? Have they now been made continuous?
|
The two discontinuous interstates were the I-95 and the I-70. The I-95 was discontinuous up until September 22, 2018 when the construction of the Pennsylvania Turnpike/Interstate was completed. The I-70, however, remains discontinuous.
|
null | false
| null |
Can you recommend best taco places in Austin?
|
Guero's on S Congress offers great breakfast tacos with reasonable prices and diverse choices. Taco 620 can be hard to spot as it is a food truck located at the corner of FM 620 and FM 2222. But you can’t miss its delicious breakfast tacos. It offers only the classics: egg & potato and egg & cheese. The tacos are fresh, soft and tasty. For more formal choices for lunch or dinner, Cuantos on 12th St and Suerte on 6th St are Mexican restaurants with great seating area and lovely vibes for dating or friends/family gatherings.
|
|
null | false
| null |
"I'm So Excited" is a song by Australian singer Anja Nissen.
|
Who is the singer of the song I'm So Excited?
|
"I'm So Excited" is a song by Australian singer Anja Nissen. It was written and produced by Redfoo and will.i.am. The song features vocals from will.i.am and Cody Wise. The song was released on 31 October 2014 and debuted at number 42 on the ARIA Singles Chart, selling 3,023 copies.
|
null | false
| null |
Which of the following are ice cream toppings and which are salad dressings: thousand island, chocolate sauce, hot fudge, balsamic vinaigrette, whipped cream, and Caesar.
|
The ice cream toppings are chocolate sauce, hot fudge, and whipped cream. The salad dressings are thousand island, balsamic vinaigrette, and Caesar.
|
|
null | false
| null |
Which of the following books were written by JK Rowling: Harry Potter and the Goblet of Fire, Harry Potter and the Half Blood Prince, Cursed Child, Lord of the Rings trilogy, Percy Jackson, A Song of Ice and Fire.
|
The books written by her are: Harry Potter and the Goblet of Fire, Harry Potter and the Half Blood Prince, Cursed Child.
|
|
1811.09786
| false
| null |
Across all 26 datasets, RCRN outperforms not only standard BiLSTMs but also 3L-BiLSTMs which have approximately equal parameterization. 3L-BiLSTMs were overall better than BiLSTMs but lose out on a minority of datasets. RCRN outperforms a wide range of competitive baselines such as DiSAN, Bi-SRUs, BCN and LSTM-CNN, etc. We achieve (close to) state-of-the-art performance on SST, TREC question classification and 16 Amazon review datasets.
Across all 26 datasets, RCRN outperforms not only standard BiLSTMs but also 3L-BiLSTMs which have approximately equal parameterization.
|
Does their model have more parameters than other models?
|
The answers are shown as follows:
* approximately equal parameterization
|
null | false
| 376
|
There has been a surge of recent interest in detecting human actions in videos. Work in this space has mainly focused on learning actions from articulated human pose BIBREF7 , BIBREF8 , BIBREF9 or mining spatial and temporal information from videos BIBREF10 , BIBREF11 . A number of resources have been produced, including Action Bank BIBREF12 , NTU RGB+D BIBREF13 , SBU Kinect Interaction BIBREF14 , and PKU-MMD BIBREF15 .
Most research on video action detection has gathered video information for a set of pre-defined actions BIBREF2 , BIBREF16 , BIBREF1 , an approach known as explicit data gathering BIBREF0 . For instance, given an action such as “open door,” a system would identify videos that include a visual depiction of this action. While this approach is able to detect a specific set of actions, whose choice may be guided by downstream applications, it achieves high precision at the cost of low recall. In many cases, the set of predefined actions is small (e.g., 203 activity classes in BIBREF2 ), and for some actions, the number of visual depictions is very small.
An alternative approach is to start with a set of videos, and identify all the actions present in these videos BIBREF17 , BIBREF18 . This approach has been referred to as implicit data gathering, and it typically leads to the identification of a larger number of actions, possibly with a small number of examples per action.
In this paper, we use an implicit data gathering approach to label human activities in videos. To the best of our knowledge, we are the first to explore video action recognition using both transcribed audio and video information. We focus on the popular genre of lifestyle vlogs, which consist of videos of people demonstrating routine actions while verbally describing them. We use these videos to develop methods to identify if actions are visually present.
The paper makes three main contributions. First, we introduce a novel dataset consisting of 1,268 short video clips paired with sets of actions mentioned in the video transcripts, as well as manual annotations of whether the actions are visible or not. The dataset includes a total of 14,769 actions, 4,340 of which are visible. Second, we propose a set of strong baselines to determine whether an action is visible or not. Third, we introduce a multimodal neural architecture that combines information drawn from visual and linguistic clues, and show that it improves over models that rely on one modality at a time.
By making progress towards automatic action recognition, in addition to contributing to video understanding, this work has a number of important and exciting applications, including sports analytics BIBREF19 , human-computer interaction BIBREF20 , and automatic analysis of surveillance video footage BIBREF21 .
The paper is organized as follows. We begin by discussing related work, then describe our data collection and annotation process. We next overview our experimental set-up and introduce a multimodal method for identifying visible actions in videos. Finally, we discuss our results and conclude with general directions for future work.
Second, we propose a set of strong baselines to determine whether an action is visible or not.
|
What is used by them to determine whether an action is visible or not?
|
A set of strong baselines
|
null | false
| null |
The first lightweight events were added to the World Championships in 1974 for men and 1985 for women. Lightweight rowing was added to the Olympics in 1996 but this came under threat in 2002 when the Programme Commission of the IOC recommended that, outside combat sports and weightlifting, there should not be weight category events. The executive board overturned this recommendation and lightweight rowing continues at the Olympics, though only in the men's and women's double sculls.
|
From the passage list all the years and the sporting events mentioned
|
Years mentioned are 1974, 1985, 1996 and 2002. The sporting events are World Championship and Olympics.
|
2002.00876
| false
| null |
Optimizations ::: a) Parallel Scan Inference
Optimizations ::: b) Vectorized Parsing
Optimizations ::: c) Semiring Matrix Operations
Torch-Struct aims for computational and memory efficiency. Implemented naively, dynamic programming algorithms in Python are prohibitively slow. As such Torch-Struct provides key primitives to help batch and vectorize these algorithms to take advantage of GPU computation and to minimize the overhead of backpropagating through chart-based dynamic programmming. Figure FIGREF17 shows the impact of these optimizations on the core algorithms.
a) Parallel Scan Inference
b) Vectorized Parsing
c) Semiring Matrix Operations
Torch-Struct aims for computational and memory efficiency. Implemented naively, dynamic programming algorithms in Python are prohibitively slow. As such Torch-Struct provides key primitives to help batch and vectorize these algorithms to take advantage of GPU computation and to minimize the overhead of backpropagating through chart-based dynamic programmming.
|
What general-purpose optimizations are included?
|
The answers are shown as follows:
* Parallel Scan Inference
* Vectorized Parsing
* Semiring Matrix Operations
|
null | false
| null |
Robert Toru Kiyosaki (born April 8, 1947) is an American entrepreneur, businessman and author. Kiyosaki is the founder of Rich Global LLC and the Rich Dad Company, a private financial education company that provides personal finance and business education to people through books and videos. The company's main revenues come from franchisees of the Rich Dad seminars that are conducted by independent individuals using Kiyosaki's brand name. He is also the creator of the Cashflow board and software games to educate adults and children about business and financial concepts.
Kiyosaki is the author of more than 26 books, including the international self-published personal finance Rich Dad Poor Dad series of books which has been translated into 51 languages and sold over 41 million copies worldwide. Kiyosaki is the subject of a class action suit filed by people who attended his seminars and has been the subject of two investigative documentaries by CBC Canada and WTAE USA. Kiyosaki's company, Rich Global LLC, filed for bankruptcy in 2012.
|
Given this paragraph, explain who Robert Kiyosaki is?
|
Robert Kiyosaki is an American entrepreneur and author who became famous primarily due to his book Rich Dad Poor Dad which has sold over 41 million copies. He is also the founder of Rich Global LLC and the Rich Dad Company, which provide financial education through books and videos. He is undergoing lawsuits related to these companies which in turn are gaining him more press due to documentaries produced on the subject of these lawsuits.
|
null | false
| null |
Modern rowing as a competitive sport can be traced to the early 17th century when professional watermen held races (regattas) on the River Thames in London, England. Often prizes were offered by the London Guilds and Livery Companies. Amateur competition began towards the end of the 18th century with the arrival of "boat clubs" at British public schools. Similarly, clubs were formed at colleges within Oxford and Cambridge in the early nineteenth century. Public rowing clubs were beginning at the same time in England, Germany, and the United States. The first American college rowing club was formed in 1843 at Yale College.
|
What was the first American college rowing club?
|
The first American college rowing club was founded at Yale College in 1843/
|
null | false
| null |
Divide these numbers into prime, composite or neither. 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12.
|
2, 3, 5, 7 and 11 are all prime numbers. 4, 6, 8, 9, 10 and 12 are composite numbers. 0 and 1 are neither prime nor composite numbers.
|
|
null | false
| null |
Tell me whether these teams have ever won the Champions League: Real Madrid, Manchester United, Manchester City, Liverpool, Paris Saint-Germain
|
Winners;
Real Madrid, Liverpool, Manchester United
Despite the recent success in their national leagues, Manchester City and Paris Saint-Germain are still competing to win their first UEFA Champions League title.
|
|
null | false
| null |
Taylor Alison Swift (born December 13, 1989) is an American singer-songwriter. Her genre-spanning discography, songwriting and artistic reinventions have received critical praise and wide media coverage. Born in West Reading, Pennsylvania, Swift moved to Nashville at age 14 to become a country artist. She signed a songwriting deal with Sony/ATV Music Publishing in 2004 and a recording contract with Big Machine Records in 2005. Her 2006 self-titled debut album made her the first female country artist to write a U.S. platinum-certified album.
Swift's next albums, Fearless (2008) and Speak Now (2010), explored country pop. The former's "Love Story" and "You Belong with Me" were the first country songs to top the U.S. pop and all-genre airplay charts, respectively. She experimented with rock and electronic styles on Red (2012), which featured her first Billboard Hot 100 number-one song, "We Are Never Ever Getting Back Together", and eschewed her country image in her synth-pop album, 1989 (2014), supported by chart-topping songs "Shake It Off", "Blank Space", and "Bad Blood". Media scrutiny inspired the urban-flavored Reputation (2017) and its number-one single "Look What You Made Me Do".
Exiting Big Machine, Swift signed with Republic Records in 2018 and released her seventh studio album, Lover (2019), followed by the autobiographical documentary Miss Americana (2020). She ventured into indie folk and alternative rock in her 2020 albums Folklore and Evermore, whose singles "Cardigan" and "Willow" topped the Hot 100. Swift began re-recording her first six albums after a dispute over their masters, re-releasing two in 2021—Fearless (Taylor's Version) and Red (Taylor's Version). The latter's "All Too Well (10 Minute Version)" became the longest song to top the Hot 100. Her 2022 album Midnights and single "Anti-Hero" broke all-time streaming records. Swift has directed music videos and films, such as All Too Well: The Short Film (2021), and played supporting roles in others.
Having sold over 200 million records globally, Swift is one of the best-selling musicians in history. She is the most streamed woman on Spotify, and the only act to have five albums open with over one million copies sold in the US. Among her accolades are 12 Grammy Awards, including three Album of the Year wins; an Emmy Award; 40 American Music Awards; 29 Billboard Music Awards; three IFPI Global Recording Artist of the Year awards; and 92 Guinness World Records. Swift has been featured in rankings such as Rolling Stone's 100 Greatest Songwriters of All Time, Billboard's Greatest of All Time Artists, the Time 100 and Forbes Celebrity 100. Honored with titles such as Artist of the Decade and Woman of the Decade, Swift is an advocate for artists' rights and women's empowerment. Her music is credited with influencing a generation of singers and songwriters.
|
Please give me a short bulleted list of the major achievements Taylor Swift has achieved.
|
Taylor Swift moved to Nashville at age 14 where She signed a songwriting deal with Sony/ATV Music that Published in 2004 and a recording contract with Big Machine Records in 2005. Her 2006 self-titled debut album made her the first female country artist to write a U.S. platinum-certified album. She has sold over 200 million records globally and is considered to be one of the best selling musicians in history. She is the most streamed woman on Spotify. She has 12 grammy awards, 29 billboard music awards, 3 IFPI Global Recording Artist of the Year awards, and 92 Guinness World Records.
|
null | false
| null |
What are some weight loss tips?
|
Diet and exercise are two key factors of weight loss. To lose weight, you need to consume fewer calories than you're burning in a day, which is called a calorie deficit. It's generally recommended to eat plenty of fruits and vegetables in order to meet weight loss goals, as these food groups are both filling and low calorie. Additionally, foods high in protein and fiber will help to keep you feeling full and lead to less binging and snacking on high calorie foods. In terms of exercise, ensure that you are staying active each day. Many people find that going for a long walk every day is a manageable way to meet their weight loss goals. Lastly, always ensure to drink plenty of water to encourage calorie burning within your body.
|
|
null | false
| null |
341 George Street, Sydney is a heritage-listed bank building located at 341 George Street, in the Sydney central business district, in the City of Sydney local government area of New South Wales, Australia. It was built from 1927 to 1932 and housed the headquarters of the Bank of New South Wales, and later Westpac. It is also known as Westpac Bank building and Bank of NSW building. Westpac sold the building in 2002, but continues to lease the lower floors for use as banking chambers. The upper floors are leased by other tenants.
It was added to the New South Wales State Heritage Register on 2 April 1999. The building served as the Head Office for the Bank of New South Wales until relocated to 60 Martin Place in 1970.
|
Given a reference text about 341 George Street, Sydney, tell me when what companies had headquarters there.
|
Westpac and Bank of New South Wales had their headquarters at 341 George Street, Sydney.
|
null | false
| null |
Think of different things I could do at the park
|
Ride a bike, use roller skates, go on the ferris wheel, take a hike, visit the duck pond, Sit on the lawn, have a picnic, take some photos
|
|
null | false
| null |
Which of the following items is the odd one out? Pacific, Atlantic, Baltic and Indian
|
Pacific, Atlantic and Indian are all oceans. Therefore, Baltic is the odd one out as it is a sea.
|
|
null | false
| null |
The Missouri and North Arkansas (reporting mark M&NA) was a railroad in Missouri and Arkansas from 1906 to 1946.
The railroad began as the Eureka Springs Railway in 1883 as a line from the St. Louis–San Francisco Railway in Seligman, MO, reaching the resort town of Eureka Springs, AR in 1883.
|
In which country did the M&NA railroad exist.
|
The M&NA was a railroad in Missouri and Arkansas, which made it part of the United States.
|
null | false
| 57
|
The vast amounts of data collected by healthcare providers in conjunction with modern data analytics techniques present a unique opportunity to improve health service provision and the quality and safety of medical care for patient benefit BIBREF0 . Much of the recent research in this area has been on personalised medicine and its aim to deliver better diagnostics aided by the integration of diverse datasets providing complementary information. Another large source of healthcare data is organisational. In the United Kingdom, the National Health Service (NHS) has a long history of documenting extensively the different aspects of healthcare provision. The NHS is currently in the process of increasing the availability of several databases, properly anonymised, with the aim of leveraging advanced analytics to identify areas of improvement in NHS services.
One such database is the National Reporting and Learning System (NRLS), a central repository of patient safety incident reports from the NHS in England and Wales. Set up in 2003, the NRLS now contains more than 13 million detailed records. The incidents are reported using a set of standardised categories and contain a wealth of organisational and spatio-temporal information (structured data), as well as, crucially, a substantial component of free text (unstructured data) where incidents are described in the `voice' of the person reporting. The incidents are wide ranging: from patient accidents to lost forms or referrals; from delays in admission and discharge to serious untoward incidents, such as retained foreign objects after operations. The review and analysis of such data provides critical insight into the complex functioning of different processes and procedures in healthcare towards service improvement for safer carer.
Although statistical analyses are routinely performed on the structured component of the data (dates, locations, assigned categories, etc), the free text remains largely unused in systematic processes. Free text is usually read manually but this is time-consuming, meaning that it is often ignored in practice, unless a detailed review of a case is undertaken because of the severity of harm that resulted. There is a lack of methodologies that can summarise content and provide content-based groupings across the large volume of reports submitted nationally for organisational learning. Methods that could provide automatic categorisation of incidents from the free text would sidestep problems such as difficulties in assigning an incident category by virtue of a priori pre-defined lists in the reporting system or human error, as well as offering a unique insight into the root cause analysis of incidents that could improve the safety and quality of care and efficiency of healthcare services.
Our goal in this work is to showcase an algorithmic methodology that detects content-based groups of records in a given dataset in an unsupervised manner, based only on the free and unstructured textual description of the incidents. To do so, we combine recently developed deep neural-network high-dimensional text-embedding algorithms with network-theoretical methods. In particular, we apply multiscale Markov Stability (MS) community detection to a sparsified geometric similarity graph of documents obtained from text vector similarities. Our method departs from traditional natural language processing tools, which have generally used bag-of-words (BoW) representation of documents and statistical methods based on Latent Dirichlet Allocation (LDA) to cluster documents BIBREF1 . More recent approaches have used deep neural network based language models clustered with k-means, without a full multiscale graph analysis BIBREF2 . There have been some previous applications of network theory to text analysis. For example, Lanchichinetti and co-workers BIBREF3 used a probabilistic graph construction analysed with the InfoMap algorithm BIBREF4 ; however, their community detection was carried out at a single-scale and the representation of text as BoW arrays lacks the power of neural network text embeddings. The application of multiscale community detection allows us to find groups of records with consistent content at different levels of resolution; hence the content categories emerge from the textual data, rather than fitting with pre-designed classifications. The obtained results could thus help mitigate possible human error or effort in finding the right category in complex category classification trees.
We showcase the methodology through the analysis of a dataset of patient incidents reported to the NRLS. First, we use the 13 million records collected by the NRLS since 2004 to train our text embedding (although a much smaller corpus can be used). We then analyse a subset of 3229 records reported from St Mary's Hospital, London (Imperial College Healthcare NHS Trust) over three months in 2014 to extract clusters of incidents at different levels of resolution in terms of content. Our method reveals multiple levels of intrinsic structure in the topics of the dataset, as shown by the extraction of relevant word descriptors from the grouped records and a high level of topic coherence. Originally, the records had been manually coded by the operator upon reporting with up to 170 features per case, including a two-level manual classification of the incidents. Therefore, we also carried out an a posteriori comparison against the hand-coded categories assigned by the reporter (healthcare personnel) at the time of the report submission. Our results show good overall correspondence with the hand-coded categories across resolutions and, specifically, at the medium level of granularity. Several of our clusters of content correspond strongly to well-defined categories, yet our results also reveal complementary categories of incidents not defined in the external classification. In addition, the tuning of the granularity afforded by the method can be used to provide a distinct level of resolution in certain areas corresponding to specialise or particular sub-themes.
In particular, we apply multiscale Markov Stability (MS) community detection to a sparsified geometric similarity graph of documents obtained from text vector similarities.
|
What community detection is applied in particular?
|
Multiscale Markov Stability (MS) community detection.
|
null | false
| null |
What is the Antartic?
|
The Antarctic is a polar region around Earth's South Pole, opposite the Arctic region around the North Pole. The Antarctic comprises the continent of Antarctica, the Kerguelen Plateau and other island territories located on the Antarctic Plate or south of the Antarctic Convergence. The Antarctic region includes the ice shelves, waters, and all the island territories in the Southern Ocean situated south of the Antarctic Convergence, a zone approximately 32 to 48 km (20 to 30 mi) wide varying in latitude seasonally. The region covers some 20 percent of the Southern Hemisphere, of which 5.5 percent (14 million km2) is the surface area of the Antarctica continent itself. All of the land and ice shelves south of 60°S latitude are administered under the Antarctic Treaty System. Biogeographically, the Antarctic realm is one of eight biogeographic realms of Earth's land surface.
|
|
1909.09491
| false
| null |
Evaluation ::: Plackett-Luce Model for SMT Reranking
After being de-duplicated, the N-best list has an average size of around 300, and with 7491 features. Refer to Formula DISPLAY_FORM9, this is ideal to use the Plackett-Luce model. Results are shown in Figure FIGREF12. We observe some interesting phenomena.
This experiment displays, in large-scale features, the Plackett-Luce model correlates with BLEU score very well, and alleviates overfitting in some degree.
Plackett-Luce Model for SMT Reranking
After being de-duplicated, the N-best list has an average size of around 300, and with 7491 features.
This experiment displays, in large-scale features, the Plackett-Luce model correlates with BLEU score very well, and alleviates overfitting in some degree.
|
What experiments with large-scale features are performed?
|
The answers are shown as follows:
* Plackett-Luce Model for SMT Reranking
|
null | false
| null |
Ratan Tata was the interim chairman of Tata Sons. He continues to head the main two Tata trusts Sir Dorabji Tata and Allied Trusts and Sir Ratan Tata Trust and their allied trusts, with a combined stake of 66% in Tata Sons, Tata group's holding company.He has served in various capacities in organizations in India and abroad. He is a member of Prime Minister's 'Council on Trade and Industry' and the 'National Manufacturing Competitiveness Council'. He is on the jury panel of Pritzker Architecture Prize – considered to be one of the world's premier architecture prizes.Over the years Tata has served on the Cornell University Board of Trustees, personally advising the school's administration in matters of international involvement, particularly regarding projects connected to India. More broadly, Tata has served on the board's Academic Affairs, Student Life, and Development Committees. In 2013 he was named Cornell Entrepreneur of the Year.He is a director on the boards of Alcoa Inc., Mondelez International and Board of Governors of the East–West Center. He is also a member of the board of trustees of University of Southern California, Harvard Business School Board of Dean's Advisors, X Prize and Cornell University. He is a member on the board of International Advisory Council at Bocconi University.He is also a member of the Harvard Business School India Advisory Board (IAB) since 2006 and previously a member of the Harvard Business School Asia-Pacific Advisory Board (APAB) 2001–2006.[citation needed]In 2013, he was appointed to the board of trustees of the Carnegie Endowment for International Peace.In February 2015, Ratan took an advisory role at Kalari Capital, a venture capital firm founded by Vani Kola.In October 2016, Tata Sons removed Cyrus Mistry as its chairman, nearly 4 years after he took over the reins of the over $100 billion conglomerate, Ratan Tata made a comeback, taking over the company's interim boss for 4 months. On 12 January 2017, Natarajan Chandrasekaran was named as the chairman of Tata Sons, a role he assumed in February 2017.
|
What are the names of Tata trusts which Ratan Tata heads?
|
Sir Dorabji Tata and Allied Trusts and Sir Ratan Tata Trust
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.