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
|
|---|---|---|---|---|---|
1909.00124
| true
| null |
Baselines. We use one strong non-DNN baseline, NBSVM (with unigrams or bigrams features) BIBREF23 and six DNN baselines. The first DNN baseline is CNN BIBREF25, which does not handle noisy labels. The other five were designed to handle noisy labels.
The comparison results are shown in Table TABREF12. From the results, we can make the following observations. (1) Our NetAb model achieves the best ACC and F1 on all datasets except for F1 of negative class on Laptop. The results demonstrate the superiority of NetAb. (2) NetAb outperforms the baselines designed for learning with noisy labels. These baselines are inferior to ours as they were tailored for image classification. Note that we found no existing method to deal with noisy labels for SSC. Training Details. We use the publicly available pre-trained embedding GloVe.840B BIBREF48 to initialize the word vectors and the embedding dimension is 300.
Baselines. We use one strong non-DNN baseline, NBSVM (with unigrams or bigrams features) BIBREF23 and six DNN baselines. The first DNN baseline is CNN BIBREF25, which does not handle noisy labels. The other five were designed to handle noisy labels.
The comparison results are shown in Table TABREF12. From the results, we can make the following observations. (1) Our NetAb model achieves the best ACC and F1 on all datasets except for F1 of negative class on Laptop.
|
Is the model evaluated against a CNN baseline?
|
Yes.
|
null | false
| 121
|
We evaluate our model through two types of metrics. The BLEU score BIBREF34 aims at measuring to what extent the generated descriptions are literally closed to the ground truth. The second category designed by BIBREF10 is more qualitative.
The BLEU score BIBREF34 is commonly used as an evaluation metric in text generation tasks. It estimates the correspondence between a machine output and that of a human by computing the number of co-occurrences for ngrams ($n \in {1, 2, 3, 4}$) between the generated candidate and the ground truth. We use the implementation code released by BIBREF35.
These metrics estimate the ability of our model to integrate elements from the table in its descriptions. Particularly, they compare the gold and generated descriptions and measure to what extent the extracted relations are aligned or differ. To do so, we follow the protocol presented in BIBREF10. First, we apply an information extraction (IE) system trained on labeled relations from the gold descriptions of the RotoWire train dataset. Entity-value pairs are extracted from the descriptions. For example, in the sentence Isaiah Thomas led the team in scoring, totaling 23 points [...]., an IE tool will extract the pair (Isaiah Thomas, 23, PTS). Second, we compute three metrics on the extracted information:
$\bullet $ Relation Generation (RG) estimates how well the system is able to generate text containing factual (i.e., correct) records. We measure the precision and absolute number (denoted respectively RG-P% and RG-#) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that also appear in $s$.
$\bullet $ Content Selection (CS) measures how well the generated document matches the gold document in terms of mentioned records. We measure the precision and recall (denoted respectively CS-P% and CS-R%) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
$\bullet $ Content Ordering (CO) analyzes how well the system orders the records discussed in the description. We measure the normalized Damerau-Levenshtein distance BIBREF36 between the sequences of records extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
CS primarily targets the “what to say" aspect of evaluation, CO targets the “how to say it" aspect, and RG targets both. Note that for CS, CO, RG-% and BLEU metrics, higher is better; which is not true for RG-#. The IE system used in the experiments is able to extract an average of 17 factual records from gold descriptions. In order to mimic a human expert, a generative system should approach this number and not overload generation with brute facts.
We evaluate our model through two types of metrics. The BLEU score BIBREF34 aims at measuring to what extent the generated descriptions are literally closed to the ground truth. The second category designed by BIBREF10 is more qualitative.
The BLEU score BIBREF34 is commonly used as an evaluation metric in text generation tasks. It estimates the correspondence between a machine output and that of a human by computing the number of co-occurrences for ngrams ($n \in {1, 2, 3, 4}$) between the generated candidate and the ground truth. We use the implementation code released by BIBREF35.
These metrics estimate the ability of our model to integrate elements from the table in its descriptions. Particularly, they compare the gold and generated descriptions and measure to what extent the extracted relations are aligned or differ. To do so, we follow the protocol presented in BIBREF10. First, we apply an information extraction (IE) system trained on labeled relations from the gold descriptions of the RotoWire train dataset. Entity-value pairs are extracted from the descriptions. For example, in the sentence Isaiah Thomas led the team in scoring, totaling 23 points [...]., an IE tool will extract the pair (Isaiah Thomas, 23, PTS). Second, we compute three metrics on the extracted information:
$\bullet $ Relation Generation (RG) estimates how well the system is able to generate text containing factual (i.e., correct) records. We measure the precision and absolute number (denoted respectively RG-P% and RG-#) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that also appear in $s$.
$\bullet $ Content Selection (CS) measures how well the generated document matches the gold document in terms of mentioned records. We measure the precision and recall (denoted respectively CS-P% and CS-R%) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
$\bullet $ Content Ordering (CO) analyzes how well the system orders the records discussed in the description. We measure the normalized Damerau-Levenshtein distance BIBREF36 between the sequences of records extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
CS primarily targets the “what to say" aspect of evaluation, CO targets the “how to say it" aspect, and RG targets both. Note that for CS, CO, RG-% and BLEU metrics, higher is better; which is not true for RG-#. The IE system used in the experiments is able to extract an average of 17 factual records from gold descriptions. In order to mimic a human expert, a generative system should approach this number and not overload generation with brute facts.
We evaluate our model through two types of metrics. The BLEU score BIBREF34 aims at measuring to what extent the generated descriptions are literally closed to the ground truth. The second category designed by BIBREF10 is more qualitative.
The BLEU score BIBREF34 is commonly used as an evaluation metric in text generation tasks. It estimates the correspondence between a machine output and that of a human by computing the number of co-occurrences for ngrams ($n \in {1, 2, 3, 4}$) between the generated candidate and the ground truth. We use the implementation code released by BIBREF35.
These metrics estimate the ability of our model to integrate elements from the table in its descriptions. Particularly, they compare the gold and generated descriptions and measure to what extent the extracted relations are aligned or differ. To do so, we follow the protocol presented in BIBREF10. First, we apply an information extraction (IE) system trained on labeled relations from the gold descriptions of the RotoWire train dataset. Entity-value pairs are extracted from the descriptions. For example, in the sentence Isaiah Thomas led the team in scoring, totaling 23 points [...]., an IE tool will extract the pair (Isaiah Thomas, 23, PTS). Second, we compute three metrics on the extracted information:
$\bullet $ Relation Generation (RG) estimates how well the system is able to generate text containing factual (i.e., correct) records. We measure the precision and absolute number (denoted respectively RG-P% and RG-#) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that also appear in $s$.
$\bullet $ Content Selection (CS) measures how well the generated document matches the gold document in terms of mentioned records. We measure the precision and recall (denoted respectively CS-P% and CS-R%) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
$\bullet $ Content Ordering (CO) analyzes how well the system orders the records discussed in the description. We measure the normalized Damerau-Levenshtein distance BIBREF36 between the sequences of records extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
CS primarily targets the “what to say" aspect of evaluation, CO targets the “how to say it" aspect, and RG targets both. Note that for CS, CO, RG-% and BLEU metrics, higher is better; which is not true for RG-#. The IE system used in the experiments is able to extract an average of 17 factual records from gold descriptions. In order to mimic a human expert, a generative system should approach this number and not overload generation with brute facts.
We evaluate our model through two types of metrics. The BLEU score BIBREF34 aims at measuring to what extent the generated descriptions are literally closed to the ground truth. The second category designed by BIBREF10 is more qualitative.
The BLEU score BIBREF34 is commonly used as an evaluation metric in text generation tasks. It estimates the correspondence between a machine output and that of a human by computing the number of co-occurrences for ngrams ($n \in {1, 2, 3, 4}$) between the generated candidate and the ground truth. We use the implementation code released by BIBREF35.
These metrics estimate the ability of our model to integrate elements from the table in its descriptions. Particularly, they compare the gold and generated descriptions and measure to what extent the extracted relations are aligned or differ. To do so, we follow the protocol presented in BIBREF10. First, we apply an information extraction (IE) system trained on labeled relations from the gold descriptions of the RotoWire train dataset. Entity-value pairs are extracted from the descriptions. For example, in the sentence Isaiah Thomas led the team in scoring, totaling 23 points [...]., an IE tool will extract the pair (Isaiah Thomas, 23, PTS). Second, we compute three metrics on the extracted information:
$\bullet $ Relation Generation (RG) estimates how well the system is able to generate text containing factual (i.e., correct) records. We measure the precision and absolute number (denoted respectively RG-P% and RG-#) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that also appear in $s$.
$\bullet $ Content Selection (CS) measures how well the generated document matches the gold document in terms of mentioned records. We measure the precision and recall (denoted respectively CS-P% and CS-R%) of unique relations $r$ extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
$\bullet $ Content Ordering (CO) analyzes how well the system orders the records discussed in the description. We measure the normalized Damerau-Levenshtein distance BIBREF36 between the sequences of records extracted from $\hat{y}_{1:T}$ that are also extracted from $y_{1:T}$.
CS primarily targets the “what to say" aspect of evaluation, CO targets the “how to say it" aspect, and RG targets both. Note that for CS, CO, RG-% and BLEU metrics, higher is better; which is not true for RG-#. The IE system used in the experiments is able to extract an average of 17 factual records from gold descriptions. In order to mimic a human expert, a generative system should approach this number and not overload generation with brute facts.
We evaluate our model through two types of metrics. The BLEU score [23] aims at measuring to what extent the generated descriptions are literally closed to the ground truth. The second category designed by [39] is more qualitative.****Relation Generation (RG) estimates how well the system is able to generate text containing factual (i.e., correct) records.****Content Selection (CS) measures how well the generated document matches the gold document in terms of mentioned records.****Content Ordering (CO) analyzes how well the system orders the records discussed in the description.
|
What are the evaluation metrics?
|
BLEU, RG, CS, and CO.
|
null | false
| null |
The underlying mechanisms for developing cardiovascular diseases (CVD) vary depending on the disease. It is estimated that dietary risk factors are associated with 53% of CVD deaths. Coronary artery disease, stroke, and peripheral artery disease involve atherosclerosis. This may be caused by high blood pressure, smoking, diabetes mellitus, lack of exercise, obesity, high blood cholesterol, poor diet, excessive alcohol consumption,and poor sleep, among other things. High blood pressure is estimated to account for approximately 13% of CVD deaths, while tobacco accounts for 9%, diabetes 6%, lack of exercise 6%, and obesity 5%. Rheumatic heart disease may follow untreated strep throat.
|
Extract all risk factors for developing cardiovascular diseases for which numbers about their contribution are given and rank them by their relative importance.
|
1. Dietary Risk Factors
2. High Blood Pressure
3. Tobacco
4. Diabetes
5. Lack of exercise
6. Obesity
|
null | false
| 72
|
Most neural network models for document classification on social media focus on text infor-mation to the neglect of other information on these platforms. In this paper, we classify post stance on social media channels and develop UTCNN, a neural network model that incorporates user tastes, topic tastes, and user comments on posts. UTCNN not only works on social media texts, but also analyzes texts in forums and message boards. Experiments performed on Chinese Facebook data and English online debate forum data show that UTCNN achieves a 0.755 macro-average f-score for supportive, neutral, and unsupportive stance classes on Facebook data, which is significantly better than models in which either user, topic, or comment information is withheld. This model design greatly mitigates the lack of data for the minor class without the use of oversampling. In addition, UTCNN yields a 0.842 accuracy on English online debate forum data, which also significantly outperforms results from previous work as well as other deep learning models, showing that UTCNN performs well regardless of language or platform.
Most neural network models for document classification on social media focus on text information to the neglect of other information on these platforms. In this paper, we classify post stance on social media channels and develop UTCNN, a neural network model that incorporates user tastes, topic tastes, and user comments on posts.
|
What is the research field of the paper?
|
Document classification.
|
null | false
| 97
|
Our experiments aim to measure if the incorporation of subword information into LexVec results in similar improvements as observed in moving from Skip-gram to fastText, and whether unsupervised morphemes offer any advantage over n-grams. For IV words, we perform intrinsic evaluation via word similarity and word analogy tasks, as well as downstream tasks. OOV word representation is tested through qualitative nearest-neighbor analysis.
All models are trained using a 2015 dump of Wikipedia, lowercased and using only alphanumeric characters. Vocabulary is limited to words that appear at least 100 times for a total of 303517 words. Morfessor is trained on this vocabulary list.
We train the standard LexVec (LV), LexVec using n-grams (LV-N), and LexVec using unsupervised morphemes (LV-M) using the same hyper-parameters as BIBREF7 ( $\textnormal {window} = 2$ , $\textnormal {initial learning rate} = .025$ , $\textnormal {subsampling} = 10^{-5}$ , $\textnormal {negative samples} = 5$ , $\textnormal {context distribution smoothing} = .75$ , $\textnormal {positional contexts} = \textnormal {True}$ ).
Both Skip-gram (SG) and fastText (FT) are trained using the reference implementation of fastText with the hyper-parameters given by BIBREF6 ( $\textnormal {window} = 5$ , $\textnormal {initial learning rate} = .025$ , $\textnormal {subsampling} = 10^{-4}$ , $\textnormal {negative samples} = 5$ ).
All five models are run for 5 iterations over the training corpus and generate 300 dimensional word representations. LV-N, LV-M, and FT use 2000000 buckets when hashing subwords.
For word similarity evaluations, we use the WordSim-353 Similarity (WS-Sim) and Relatedness (WS-Rel) BIBREF28 and SimLex-999 (SimLex) BIBREF29 datasets, and the Rare Word (RW) BIBREF20 dataset to verify if subword information improves rare word representation. Relationships are measured using the Google semantic (GSem) and syntactic (GSyn) analogies BIBREF2 and the Microsoft syntactic analogies (MSR) dataset BIBREF30 .
We also evaluate all five models on downstream tasks from the VecEval suite BIBREF13 , using only the tasks for which training and evaluation data is freely available: chunking, sentiment and question classification, and natural language identification (NLI). The default settings from the suite are used, but we run only the fixed settings, where the embeddings themselves are not tunable parameters of the models, forcing the system to use only the information already in the embeddings.
Finally, we use LV-N, LV-M, and FT to generate OOV word representations for the following words: 1) “hellooo”: a greeting commonly used in instant messaging which emphasizes a syllable. 2) “marvelicious”: a made-up word obtained by merging “marvelous” and “delicious”. 3) “louisana”: a misspelling of the proper name “Louisiana”. 4) “rereread”: recursive use of prefix “re”. 5) “tuzread”: made-up prefix “tuz”.
Relationships are measured using the Google semantic (GSem) and syntactic (GSyn) analogies (Mikolov et al., 2013a) and the Microsoft syntactic analogies (MSR) dataset (Mikolov et al., 2013b).
|
What analogies are used to measure word relationships in the research?
|
Google semantic (GSem) and syntactic (GSyn) analogies and the Microsoft syntactic analogies (MSR) dataset.
|
null | false
| null |
Which of these cities are in Austria: Graz, Warsaw, Amsterdam, Vienna, New York, Jakarta
|
Out of the cities you listed, Graz and Vienna are in Austria.
|
|
null | false
| 136
|
Previous work on multimodal machine translation has shown that visual information is only needed in very specific cases, for example in the presence of ambiguous words where the textual context is not sufficient. As a consequence, models tend to learn to ignore this information. We propose a translate-and-refine approach to this problem where images are only used by a second stage decoder. This approach is trained jointly to generate a good first draft translation and to improve over this draft by (i) making better use of the target language textual context (both left and right-side contexts) and (ii) making use of visual context. This approach leads to the state of the art results. Additionally, we show that it has the ability to recover from erroneous or missing words in the source language.
Additionally, we show that it has the ability to recover from erroneous or missing words in the source language.
|
What ability does the proposed approach have in addition?
|
The ability to recover from erroneous or missing words in the source language.
|
null | false
| 36
|
We evaluate our model on two publicly available datasets. The statistics for both are shown in Table TABREF3 . The details of these datasets are as follows:
OntoNotes: OntoNotes 5.0 BIBREF16 includes texts from five different text genres: broadcast conversation (200k), broadcast news (200k), magazine (120k), newswire (625k), and web data (300k). This dataset is annotated with 18 categories.
Wiki(gold): The training data consists of Wikipedia sentences and was automatically generated using a distant supervision method, mapping hyperlinks in Wikipedia articles to Freebase, which we do not use in this study. The test data, mainly consisting of sentences from news reports, was manually annotated as described in BIBREF8 . The class hierarchy is shown in Figure FIGREF2 . This dataset is annotated with 7 main categories (bold text in Figure FIGREF2 ), which maps directly to OntoNotes. The miscellaneous category in Figure FIGREF2 does not have direct mappings, so future work may include redefining these categories so the mappings are more meaningful.
OntoNotes: OntoNotes 5.0 (Weischedel et al., 2013) includes texts from five different text genres: broadcast conversation (200k), broadcast news (200k), magazine (120k), newswire (625k), and web data (300k). This dataset is annotated with 18 categories.
|
How many categories of annotations does the dataset OntoNotes have?
|
18 categories.
|
null | false
| 439
|
As obtaining the full labels f (ϕ; x, y, z) for query points (x, y, z) from all the locations P I is nontrivial for complex geometry from real-world (Fig.) and due to additional engineering constraints on runtime loading and memory bottleneck for the sufficiently dense pre-computed query point occupancy labels, we propose to incorporate the loss w.r.t. the spatial gradients. For occupancy,
(1) where P 0+ I and P 0− I denotes inward and outward near-surface query points, and λ or represents the loss weight of the occupancy geometric regularization. The facing direction of the mesh surface (determining inward or outward) can be determined by normals (if available), or via rendering the surface in all views as in the RGBD captures (with the surface facing toward camera as the "outward" side). The three terms in Eq. 1 correspond to the three types of losses in Fig.(c) ("∇", "+", "−") respectively. For SDF, we set the loss similar to:
where P 0 I denotes the query points on the ground truth surface only, and λ sr and λ sn represents the loss weight for the signed distance geometric regularization term and normal term respectively. The three terms represents the Eikonal regularization, surface zero-SDF loss, and the surface normal loss respectively, with the last term optional. Note now the loss functions in Eq. 1, 2 no longer require the implicit ground truth label for query points far away from the ground truth surface, enabling training with the real-world imperfect scanned data.
In both cases, the availability of surface normals is optional.
Practically, we found only our proposed loss for the occupancy version (Fig.(c), Eq. 1) demonstrates capability to be applicable for the real-world single-view feed-forward scenario, and use Eq. 1 rather than Eq. 2 in all of our experiments. First, learning with the Eikonal term (Eq. 2) suffers severely from its sensitivity of model initialization. We found either the SDF or the Truncated SDF (TSDF) representation model cannot easily converge during training. Specifically, we found it poses numerical difficulties when the model is learned to predict a large distance value or a constant truncation value for the majority of the query points in the air that is far from any surfaces. This becomes a major problem when we extend from the single objects scenario or single scene fitting. Second, our loss function (Eq. 1) can significantly save memory footprint during training and enable large batch size training, which is crucial in our learning framework. Unlike in Eq. 2 where all the query points require spatial gradient computation, which leads to 3× higher of memory footprint, in our loss function (Eq. 1), only the non-surface query points do. We set the query point batch size of the non-surface points (512 in practice) to be much smaller than the critical near-surface points (4096 in practice), enabling learning with the image batch size of 32 in our real-scene training.
As obtaining the full labels ˜f(ϕ; x, y, z) for query points (x, y, z) from all the locations PI is nontrivial for complex geometry from real-world (Fig. 2) and due to additional engineering constraints
on runtime loading and memory bottleneck for the sufficiently dense pre-computed query point
occupancy labels, we propose to incorporate the loss w.r.t. the spatial gradients.
|
What is the engineering constraint?
|
We revised the caption of Fig 2 and the sentence before Eq. 1. The engineering constraint is referring to the difficulties of storing and run-time loading the sufficiently dense pre-computed occupancy labels for all the samples in the batch (batch size is 32) when training with the pre-computed full occupancy labels (our baseline of DISN + TSDF Voxelization), as we introduced in the introduction section. Hence for this baseline, we train with the relatively sparse labels (the voxels), whose resolution is lower than the pre-computed dense occupancy labels.
|
null | false
| null |
Why do some people prefer to drink hot water than cold water?
|
People live in the north tend to live drinking hot beverages, such as tea or hot coco due to the cold weather in the winter. However, for people who have low tolerance in caffeine, it is better for them to not drink the previously listed beverages. A warm cup of water could help heat up the body, while reduce the intake of caffeine.
Also, in development areas, cold water source may not be sanitary for people to intake directly. After boiling the water, it kills 99% of germs, and will make it safer for people to drink as well.
|
|
null | false
| null |
Where is Rihanna from?
|
The artist known as Rihanna was born on the island of Barbados.
|
|
null | false
| 138
|
Users of photo-sharing websites such as Flickr often provide short textual descriptions in the form of tags to help others find the images. With the availability of GPS systems in current electronic devices such as smartphones, latitude and longitude coordinates are nowadays commonly made available as well. The tags associated with such georeferenced photos often describe the location where these photos were taken, and Flickr can thus be regarded as a source of environmental information. The use of Flickr for modelling urban environments has already received considerable attention. For instance, various approaches have been proposed for modelling urban regions BIBREF0 , and for identifying points-of-interest BIBREF1 and itineraries BIBREF2 , BIBREF3 . However, the usefulness of Flickr for characterizing the natural environment, which is the focus of this paper, is less well-understood.
Many recent studies have highlighted that Flickr tags capture valuable ecological information, which can be used as a complementary source to more traditional sources. To date, however, ecologists have mostly used social media to conduct manual evaluations of image content with little automated exploitation of the associated tags BIBREF4 , BIBREF5 , BIBREF6 . One recent exception is BIBREF7 , where bag-of-words representations derived from Flickr tags were found to give promising result for predicting a range of different environemental phenomena.
Our main hypothesis in this paper is that by using vector space embeddings instead of bag-of-words representations, the ecological information which is implicitly captured by Flickr tags can be utilized in a more effective way. Vector space embeddings are representations in which the objects from a given domain are encoded using relatively low-dimensional vectors. They have proven useful in natural language processing, especially for encoding word meaning BIBREF8 , BIBREF9 , and in machine learning more generally. In this paper, we are interested in the use of such representations for modelling geographic locations. Our main motivation for using vector space embeddings is that they allow us to integrate the textual information we get from Flickr with available structured information in a very natural way. To this end, we rely on an adaptation of the GloVe word embedding model BIBREF9 , but rather than learning word vectors, we learn vectors representing locations. Similar to how the representation of a word in GloVe is determined by the context words surrounding it, the representation of a location in our model is determined by the tags of the photos that have been taken near that location. To incorporate numerical features from structured environmental datasets (e.g. average temperature), we associate with each such feature a linear mapping that can be used to predict that feature from a given location vector. This is inspired by the fact that salient properties of a given domain can often be modelled as directions in vector space embeddings BIBREF10 , BIBREF11 , BIBREF12 . Finally, evidence from categorical datasets (e.g. land cover types) is taken into account by requiring that locations belonging to the same category are represented using similar vectors, similar to how semantic types are sometimes modelled in the context of knowledge graph embedding BIBREF13 .
While our point-of-departure is a standard word embedding model, we found that the off-the-shelf GloVe model performed surprisingly poorly, meaning that a number of modifications are needed to achieve good results. Our main findings are as follows. First, given that the number of tags associated with a given location can be quite small, it is important to apply some kind of spatial smoothing, i.e. the importance of a given tag for a given location should not only depend on the occurrences of the tag at that location, but also on its occurrences at nearby locations. To this end, we use a formulation which is based on spatially smoothed version of pointwise mutual information. Second, given the wide diversity in the kind of information that is covered by Flickr tags, we find that term selection is in some cases critical to obtain vector spaces that capture the relevant aspects of geographic locations. For instance, many tags on Flickr refer to photography related terms, which we would normally not want to affect the vector representation of a given location. Finally, even with these modifications, vector space embeddings learned from Flickr tags alone are sometimes outperformed by bag-of-words representations. However, our vector space embeddings lead to substantially better predictions in cases where structured (scientific) information is also taken into account. In this sense, the main value of using vector space embeddings in this context is not so much about abstracting away from specific tag usages, but rather about the fact that such representations allow us to integrate numerical and categorical features in a much more natural way than is possible with bag-of-words representations.
The remainder of this paper is organized as follows. In the next section, we provide a discussion of existing work. Section SECREF3 then presents our model for embedding geographic locations from Flickr tags and structured data. Next, in Section SECREF4 we provide a detailed discussion about the experimental results. Finally, Section SECREF5 summarizes our conclusions.
Our main hypothesis in this paper is that by using vector space embeddings instead of bag-of-words representations, the ecological information which is implicitly captured by Flickr tags can be utilized in a more effective way. Vector space embeddings are representations in which the objects from a given domain are encoded using relatively low-dimensional vectors. They have proven useful in natural language processing, especially for encoding word meaning, and in machine learning more generally. In this paper, we are interested in the use of such representations for modelling geographic locations.
|
What does the paper explore?
|
The use of vector space embeddings for modelling geographic locations
|
null | false
| 64
|
Table 2 shows a comparison of the discussed settings on both SQuAD and TriviaQA. Without any fine-tuning (column 0) the performance is low, probably because the model never saw a real question, but we see significant gains with Cloze pretraining even with very little labeled data. The BiDAF+SA model, exceeds an F1 score of $50\%$ with only $1\%$ of the training data (454 questions for SQuAD, and 746 questions for TriviaQA), and approaches $90\%$ of the best performance with only $10\%$ labeled data. The gains over the SL setting, however, diminish as the size of the labeled set increases and are small when the full dataset is available.
Cloze pretraining outperforms the GDAN baseline from BIBREF0 using the same SQuAD dataset splits. Additionally, we show improvements in the $90\%$ data case unlike GDAN. Our approach is also applicable in the extremely low-resource setting of $1\%$ data, which we suspect GDAN might have trouble with since it uses the labeled data to do reinforcement learning. Furthermore, we are able to use the same cloze dataset to improve performance on both SQuAD and TriviaQA datasets. When we use the same unlabeled data to pre-train with a language modeling objective, the performance is worse, showing the bias we introduce by constructing clozes is important.
On the BioASQ dataset (Table 3 ) we again see a significant improvement when pretraining with the cloze questions over the supervised baseline. The improvement is smaller than what we observe with SQuAD and TriviaQA datasets – we believe this is because questions are generally more difficult in BioASQ. BIBREF14 showed that pretraining on SQuAD dataset improves the downstream performance on BioASQ. Here, we show a much larger improvement by pretraining on cloze questions constructed in an unsupervised manner from the same domain.
When we use the same unlabeled data to pre-train with a language modeling objective, the performance is worse[Since the GA Reader uses bidirectional RNN layers, when pretraining the LM we had to mask the inputs to the in_x001f_termediate layers partially to avoid the model being exposed to the labels it is predicting. This results in a only a subset of the parameters being pretrained, which is why we believe this baseline performs poorly.], showing the bias we introduce by constructing clozes is important.
|
Why are intermediate layers masked？
|
To avoid the model being exposed to the labels it is predicting
|
null | false
| null |
How many U.S. Presidents have served in the military?
|
Twenty-six U.S. Presidents have served in various branches of the military.
George Washington
Thomas Jefferson
James Madison
James Monroe
Andrew Jackson
William Henry Harrison
John Tyler
James K. Polk
Zachary Taylor
Millard Fillmore
Franklin Pierce
James Buchanan
Abraham Lincoln
Andrew Johnson
Ulysses S. Grant
Rutherford B. Hayes
James A. Garfield
Chester A. Arthur
Benjamin Harrison
William McKinley
Theodore Roosevelt
Harry S. Truman
Dwight D. Eisenhower
John F. Kennedy
Lyndon B. Johnson
Richard M. Nixon
Gerald R. Ford, Jr.
Jimmy Carter
Ronald Reagan
George Bush
George W. Bush
|
|
null | false
| null |
How do I make a perfect grilled cheese sandwich?
|
Step 1. Select a cheese that will melt well. Parmesan, for example, is not ideal for a grilled cheese sandwich. Cheddar, Havarti, and Colby Jack are all cheeses that taste great and will melt.
Step 2. Get two slices of bread, preferably not too thick; typical store-bought sliced bread like sourdough, potato, and white make adequate grilled bread.
Step 3. While you are assembling the sandwich, heat a shallow pan over medium-low heat on your stovetop.
Step 4. Spread some mayonnaise on the inside of each slice of bread. Sprinkle some garlic powder over the mayonnaise -- this gives a little extra flavor that one might not expect from a grilled cheese.
Step 5. Layer the cheese between the slices of bread. Don't put too much cheese on it, or it will leak out of the sandwich when you cut it!
Step 6. Assemble the sandwich and put a thin layer of mayo on the outside of each slice.
Step 7. Lay the sandwich inside your heated pan. It should sizzle immediately. Wait two minutes and check to see if the cheese is melting. If the bread is browning too quickly, turn the heat down. When it is adequately browned on one side, use a spatula to flip it over to grill the other side.
Step 8. Once your sandwich is grilled to your liking, turn off the stove and remove your sandwich from the pan. Slice diagonally and serve with a bowl of hot tomato soup. Enjoy!
|
|
1911.00133
| true
| null |
We submit 4,000 segments, sampled equally from each domain and uniformly within domains, to Mechanical Turk to be annotated by at least five Workers each and include in each batch one of 50 “check questions” which have been previously verified by two in-house annotators. After removing annotations which failed the check questions, and data points for which at least half of the annotators selected “Can't Tell”, we are left with 3,553 labeled data points from 2,929 different posts. We take the annotators' majority vote as the label for each segment and record the percentage of annotators who agreed. The resulting dataset is nearly balanced, with 52.3% of the data (1,857 instances) labeled stressful.
The resulting dataset is nearly balanced, with 52.3% of the data (1,857 instances) labeled stressful.
|
Is the dataset balanced across categories?
|
Yes.
|
null | false
| 171
|
Finally, we consider appropriateness per system. Following related work by BIBREF21, BIBREF24, we use Trueskill BIBREF25 to cluster systems into equivalently rated groups according to their partial relative rankings. The results in Table TABREF36 show that the highest rated systen is Alley, a purpose build bot for online language learning. Alley produces “polite refusal” (2b) - the top ranked strategy - 31% of the time. Comparatively, commercial systems politely refuse only between 17% (Cortana) and 2% (Alexa). Most of the time commercial systems tend to “play along” (3a), joke (3b) or don't know how to answer (1e) which tend to receive lower ratings, see Figure FIGREF38. Rule-based systems most often politely refuse to answer (2b), but also use medium ranked strategies, such as deflect (2c) or chastise (2d). For example, most of Eliza's responses fall under the “deflection” strategy, such as “Why do you ask?”. Data-driven systems rank low in general. Neuralconvo and Cleverbot are the only ones that ever politely refuse and we attribute their improved ratings to this. In turn, the “clean” seq2seq often produces responses which can be interpreted as flirtatious (44%), and ranks similarly to Annabelle Lee and Laurel Sweet, the only adult bots that politely refuses ( 16% of the time). Ritter:2010:UMT:1857999.1858019's IR approach is rated similarly to Capt Howdy and both produce a majority of retaliatory (2e) responses - 38% and 58% respectively - followed by flirtatious responses. Finally, Dr Love and Sophia69 produce almost exclusively flirtatious responses which are consistently ranked low by users.
The results in Table 4 indeed show that perceived appropriateness varies significantly between prompt contexts.
|
Does the perceived appropriateness vary significantly between prompt contexts?
|
Yes, it does.
|
null | false
| 90
|
BioASQ is a biomedical document classification, document retrieval, and question answering competition, currently in its seventh year. We provide an overview of our submissions to semantic question answering task (7b, Phase B) of BioASQ 7 (except for 'ideal answer' test, in which we did not participate this year). In this task systems are provided with biomedical questions and are required to submit ideal and exact answers to those questions. We have used BioBERT BIBREF0 based system , see also Bidirectional Encoder Representations from Transformers(BERT) BIBREF1, and we fine tuned it for the biomedical question answering task. Our system scored near the top for factoid questions for all the batches of the challenge. More specifially, in the third test batch set, our system achieved highest ‘MRR’ score for Factoid Question Answering task. Also, for List-type question answering task our system achieved highest recall score in the fourth test batch set. Along with our detailed approach, we present the results for our submissions and also highlight identified downsides for our current approach and ways to improve them in our future experiments. In last test batch results we placed 4th for List-type questions and 3rd for Factoid-type questions.)
The QA task is organized in two phases. Phase A deals with retrieval of the relevant document, snippets, concepts, and RDF triples, and phase B deals with exact and ideal answer generations (which is a paragraph size summary of snippets). Exact answer generation is required for factoid, list, and yes/no type question.
BioASQ organizers provide the training and testing data. The training data consists of questions, gold standard documents, snippets, concepts, and ideal answers (which we did not use in this paper, but we used last year BIBREF2). The test data is split between phases A and B. The Phase A dataset consists of the questions, unique ids, question types. The Phase B dataset consists of the questions, golden standard documents, snippets, unique ids and question types. Exact answers for factoid type questions are evaluated using strict accuracy (the top answer), lenient accuracy (the top 5 answers), and MRR (Mean Reciprocal Rank) which takes into account the ranks of returned answers. Answers for the list type question are evaluated using precision, recall, and F-measure.
BioASQ is a biomedical document classification, document retrieval, and question answering competition,
|
What is BioASQ?
|
A biomedical document classification, document retrieval, and question answering competition,
|
null | false
| null |
The giant panda is a bear species endemic to China. It is a Carnivora, the giant panda is a folivore, with bamboo shoots and leaves making up more than 99% of its diet.
|
What do giant Pandas eats?
|
the giant panda is a folivore, with bamboo shoots and leaves making up more than 99% of its diet.
|
null | false
| 32
|
Propaganda aims at influencing people's mindset with the purpose of advancing a specific agenda. In the Internet era, thanks to the mechanism of sharing in social networks, propaganda campaigns have the potential of reaching very large audiences BIBREF0, BIBREF1, BIBREF2.
Propagandist news articles use specific techniques to convey their message, such as whataboutism, red Herring, and name calling, among many others (cf. Section SECREF3). Whereas proving intent is not easy, we can analyse the language of a claim/article and look for the use of specific propaganda techniques. Going at this fine-grained level can yield more reliable systems and it also makes it possible to explain to the user why an article was judged as propagandist by an automatic system.
With this in mind, we organised the shared task on fine-grained propaganda detection at the NLP4IF@EMNLP-IJCNLP 2019 workshop. The task is based on a corpus of news articles annotated with an inventory of 18 propagandist techniques at the fragment level. We hope that the corpus would raise interest outside of the community of researchers studying propaganda. For example, the techniques related to fallacies and the ones relying on emotions might provide a novel setting for researchers interested in Argumentation and Sentiment Analysis.
With this in mind, we organised the shared task on fine-grained propaganda detection at the NLP4IF@EMNLP-IJCNLP 2019 workshop. The task is based on a corpus of news articles annotated with an inventory of 18 propagandist techniques at the fragment level.
|
What does the paper describe?
|
A task based on a corpus of news articles annotated with an inventory of 18 propagandist techniques at the fragment level.
|
null | false
| null |
Tell me whether these are mammals or reptiles: dog, crocodile, cat, lizard, snake, tiger, lion, turtle.
|
Dog: mammal
Crocodile: reptile
Cat: mammal
Lizard: reptile
Snake: reptile
Tiger: mammal
Lion: mammal
Turtle: reptile
|
|
null | false
| null |
Explain how "work from home" has an opportunity cost. Provide anecdotes to support your points.
|
These days, it is “cool” to be a techie. Jobs come with perks, including nap rooms and free gourmet lunches. Celebrity entrepreneurs, like Mark Cuban, talk about how they are learning to write code. Back in the late 1990s when I started getting paid to write code, this was not the case. I often found myself in windowless rooms, basements or chilly server rooms wrangling data or writing code. However, from early on in my career, my managers ensured that I was exposed to the real action: where the business value-add activities took place. I walked factory floors at Prestolite in Ann Arbor to see how the ERP system for which I was contributing programming code played a part in the convergence of IT with humans working with machines and parts to produce finished products. When I worked for Richard Helppie’s Superior Consultant Holdings Corporation, while between programming assignments, I shadowed an industrial engineer who was helping to redesign an Emergency Department’s (ED) physical layout; we watched the flow of doctors, patients and nurses. We asked questions like: “why are the doctors not using the stations intended for note-taking and instead are walking down two hallways to take their notes in empty offices?”; it turned out that the note-taking station in the heart of the ED was a place where doctors were exposed to all sorts of noise and other distractions.
Being a good programmer had a lot to do with understanding software architecture, APIs, etc… But being a better programmer meant understanding the world in which the software was actually put into practice.
Ford Motor Company’s F-150 has been America’s best selling truck for 46 consecutive years. Looking at the finished product with your eyes barely conveys the awesome complexity of the supply chain involved in bringing it to life. To get a better sense of the convergence of the F-150’s supply chain into its finished product, you can take a stroll on the catwalk that hovers above the assembly line as part of the Ford Rouge Factory Tour. The public can tour the plant and walk above a large portion of the assembly process. You can see Kanban in action as workers pull components from bins and perform their specialized task in a step of the ever-rolling line, while overhead displays help orchestrate replenishment. You can get a sense of the optimized flows of humans, robots and parts. And, maybe, if you look closely, you’ll spot opportunities for improvements in the process, in the safety measures, or in areas where automation has not yet taken hold. Consider whether or not you would see such opportunities by looking at charts and graphs…
Alan Mulally was figuratively and literally a gymnast (at least during his time at the University of Kansas). After taking over the reins as CEO of Ford in 2006, he famously transformed a culture with deep-rooted divisional fiefdoms that hindered communication, reporting, efficiency and agility into a new era of quality, efficiency and innovation. A disciplined engineer by training, he did not solely rely on charts, facts and figures to drive organizational change. He used his eyes and made sure that the leaders saw the firm’s operations in a certain, methodical way. For instance, he built trust through transparency by using simple visual tools (like color-coding for status).
Once Mulally brought divisional leaders together to a conference room to look at a pile of hood prop rods for the various cars and trucks that Ford manufactures. The prop rods all looked different. They were from different suppliers, made of different materials, had a variety of costs, and had different internal staff dedicated to their role in the supply chain and assembly. He did not send a spreadsheet in an email to make his point. He had the leadership team look at the rods on the table and aimed for them to understand that “this kind of variation is costly but doesn’t matter to consumers.”
Mulally performed organizational and operational gymnastics, starting with a junk-rated company, and ending up being called “The Savior of Ford”. Mulally understood the power of seeing and of showing.
In the 1970s, Japan rocked the automotive world by mass producing high-quality, fuel-efficient vehicles at highly competitive prices. The Toyota way eventually swept across the industry, and a new jargon (and mindset) made its way to Detroit, with terms like Kaizen, Kanban, and Genchi Genbutsu. The Gemba Walk made its way into automotive manufacturing and other industries, such as Overhead Door and at the COVID-19 vaccination center at Mount Morningside Hospital in New York City. “The literal translation for Gemba (現場) is ‘the real place’ and in business it refers to the real place where value is created, such as the factory floor.” These tools are not a magic bullet; a Harvard Business Review study found that MBWA (“management by walking around”) decreased performance, on average, in a hospital setting.
I used aspects of the Gemba Walk to help design a software system for lab technicians at a major national laboratory. When the CDC needed help to track the variants of the SARS-Cov-2 (“Coronavirus”) across the USA, I helped build a system that enabled lab technicians to efficiently select and route the appropriate samples for the genetic sequencing process, a step that comes after the COVID PCR test. I went to the lab, watched the technicians, observed the physical flow of humans and materials in the lab and talked with the people involved in the process. I was able to see (and measure) aspects of the process that I was tasked with automating. I made observations that I never could have made through video calls or emails.
Software practitioners (developers, designers, architects, UX engineers and product owners) are continuously building and refining systems for robotics, assembly lines, labs, logistics, warehouses, shipyards, and other industries. “Agile” software development is a common methodology (and has some roots that are tied to Toyota-inspired manufacturing practices). Agile facilitates frequent, structured communication, but the product owner often serves as a proxy for the actual business stakeholders. He or she may communicate with slides, diagrams and customer “user stories”. However, as the value chain is communicated through the various parties (product owner, architect, etc.), the real-world value chain (what does the assembly line actually look like when it is in full swing) can become distilled and distorted.
Executives and senior managers are keenly monitoring the Work from Home (WFH) phenomenon. The U.S. Census Bureau is tracking WFH trends. Academics are analyzing WFH. Celebrity executives are flip-flopping on WFH. Productivity is questioned. But, are we asking the right question? Can you WFH and walk?
It is clear that IT and non-IT professionals prefer WFH… So companies are in a bind: force in-office work or find some middle ground. A.P. Moller – Maersk, “an integrated transport and logistics company”, lists job vacancies on their website. When looking at their IT jobs, you will struggle to see a mention of “Remote” or “Work from Home.” And you will also see vacancies, such as this Lead Software Engineer that have been open since May of 2022.
A startup named “Gemba” is hoping to solve this vexing problem using Virtual Reality (VR). A Global Newswire article from January, 2023 describes Gemba’s origin story: “Gemba grew out of executive training company The Leadership Network, founded in 2013 by CEO Nathan Robinson and Chairman Victor Lewis. Through its collaboration with global giants like Toyota, Tesla, Google, BMW and Amazon, senior executives from non-competing companies have been able to visit their state-of-the-art factories and facilities to see best practice in action as part of an executive Gemba masterclass.”
We’ll see if technology will allow techies like myself to continue to work in our pajamas while getting the benefits of a Gemba Walk. In the meantime, executives and senior managers may want to consider scheduling on-site Gemba walks. Just don’t forget to supply a gourmet lunch, a foot massage, and some nice swag.
|
|
null | false
| null |
The Byzantine Empire, also referred to as the Eastern Roman Empire or Byzantium, was the continuation of the Roman Empire primarily in its eastern provinces during Late Antiquity and the Middle Ages, when its capital city was Constantinople. It survived the fragmentation and fall of the Western Roman Empire in the 5th century AD and continued to exist for an additional thousand years until the fall of Constantinople to the Ottoman Empire in 1453. During most of its existence, the empire remained the most powerful economic, cultural, and military force in Europe. The terms "Byzantine Empire" and "Eastern Roman Empire" were coined after the end of the realm; its citizens continued to refer to their empire as the Roman Empire and to themselves as Romans[note 2]—a term which Greeks continued to use for themselves into Ottoman times. Although the Roman state continued and its traditions were maintained, modern historians prefer to differentiate the Byzantine Empire from Ancient Rome as it was centered on Constantinople instead of Rome, oriented towards Greek rather than Latin culture, and was characterized by Eastern Orthodox Christianity.
|
What was the Byzantine Empire?
|
After the fall of the Roman Empire in the 5th century AD, it split into many nations. The eastern part of the Roman Empire became the Byzantine Empire with Constantinople as its capital city rather than Rome. The Byzantine citizens continued to refer to themselves as Romans, and believed themselves as the successor to the former empire, however their culture was more Greek than Latin. The Byzantine Empire continued to be the most powerful nation in Europe until the fall of Constantinople to the Ottoman Empire in 1453.
|
null | false
| null |
Born in Porto Calvo, Alagoas, Willian José started his career at local CRB's youth setup, before joining Grêmio Barueri in 2008, aged 17. On 1 August 2009 he made his first team – and Série A – debut, coming on as a late substitute in a 2–1 loss at Botafogo.
Willian José scored his first goal as a professional on 17 January of the following year, netting his side's only in a 1–1 draw at Sertãozinho for the Campeonato Paulista championship. He scored six further times during the campaign, which ended in relegation.
|
Where was William José born?
|
William José was born in Porto Calvo, in the State of Alagoas in Brazil. He started his career at the local Gremio Barueri soccer team.
|
null | false
| null |
Doña Bárbara is a novel written by the Venezuelan Rómulo Gallegos and published by Editorial Araluce, on February 15 , 1929 . 1 It has been republished more than forty times and translated into other languages. It consists of three parts and is set in Apure plains in Venezuela , on the Arauca river property .
|
What is Doña Barbara?
|
Doña Bárbara is a novel written by the Venezuelan Rómulo Gallegos and published by Editorial Araluce, on February 15 , 1929 . 1 It has been republished more than forty times and translated into other languages. It consists of three parts and is set in Apure plains in Venezuela , on the Arauca river property .
Doña Bárbara is the most popular Venezuelan novel: since its appearance in 1929, it has been read avidly perhaps because between the lines Gallegos expressed his rebellion against the dictatorial regime of Juan Vicente Gómez and the backwardness that the country was experiencing. The novel examines the sociological topic, rooted in positivism, civilization versus barbarism in rural Venezuelan life. Among other merits, the writer's mastery in terms of character creation stands out, as well as the description of the llanero landscape.
|
1803.03786
| false
| null |
The training dataset contains 2,815 examples, where 1,940 (i.e., 69%) are fake news and 1,968 (i.e., 70%) are click-baits; we further have 761 testing examples. However, there is 98% correlation between fake news and click-baits, i.e., a model trained on fake news would do well on click-baits and vice versa. Thus, below we focus on fake news detection only.
The training dataset contains 2,815 examples, where 1,940 (i.e., 69%) are fake news and 1,968 (i.e., 70%) are click-baits; we further have 761 testing examples.
|
what datasets were used?
|
The answers are shown as follows:
* training dataset contains 2,815 examples
* 761 testing examples
|
null | false
| null |
The Chesapeake Bay (/ˈtʃɛsəpiːk/ CHESS-ə-peek) is the largest estuary in the United States. The Bay is located in the Mid-Atlantic region and is primarily separated from the Atlantic Ocean by the Delmarva Peninsula, including parts of the Eastern Shore of Maryland, the Eastern Shore of Virginia, and the state of Delaware. The mouth of the Bay at its southern point is located between Cape Henry and Cape Charles. With its northern portion in Maryland and the southern part in Virginia, the Chesapeake Bay is a very important feature for the ecology and economy of those two states, as well as others surrounding within its watershed. More than 150 major rivers and streams flow into the Bay's 64,299-square-mile (166,534 km2) drainage basin, which covers parts of six states, New York, Pennsylvania, Delaware, Maryland, Virginia, and West Virginia, and all of Washington, D.C.
|
What is the largest estuary in the United States?
|
The Chesapeake Bay is the largest estuary in the United States.
|
null | false
| 214
|
Our datasets can be used for various hate speech tasks. In this paper, we focus on generative hate speech intervention.
The goal of this task is to generate a response to hate speech that can mitigate its use during a conversation. The objective can be formulated as the following equation:
where $c$ is the conversation, $r$ is the corresponding intervention response, and $D$ is the dataset. This task is closely related to the response generation and dialog generation, though several differences exist including dialog length, language cadence, and word imbalances. As a baseline, we chose the most common methods of these two tasks, such as Seq2Seq and VAE, to determine the initial feasibility of automatically generate intervention responses. More recent Reinforcement Learning method for dialog generation BIBREF21 can also be applied to this task with slight modification. Future work will explore more complex, and unique models.
Similar to BIBREF21, a generative model is considered as an agent. However, different from dialog generation, generative intervention does not have multiple turns of utterance, so the action of the agent is to select a token in the response. The state of the agent is given by the input posts and the previously generated tokens. Another result due to this difference is that the rewards with regard to ease of answering or information flow do not apply to this case, but the reward for semantic coherence does. Therefore, the reward of the agent is:
where $rw(c,r)$ is the reward with regard to the conversation $c$ and its reference response $r$ in the dataset. $p(r|c)$ denotes the probability of generating response $r$ given the conversation $c$, and $p_{back}(c|r)$ denotes the backward probability of generating the conversation based on the response, which is parameterized by another generation network. The reward is a weighted combination of these two parts, which are observed after the agent finishing generating the response. We refer the readers to BIBREF21 for details.
The encoder consists of 2 bidirectional GRU layers.
|
What does the encoder consist of?
|
The encoder consists of 2 bidirectional GRU layers.
|
null | false
| null |
Why do some scholars think that campaigns do not matter much for presidential election outcomes? Why do other scholars and practitioners think that campaigns do matter?
|
Scholars argue that there are three main reasons that campaigns do not matter in elections. One, that other factors have a greater influence than campaigns, such as if a candidate is an incumbent, party identification of the candidate, state of the economy, and finally, macro-trends like wars and presidential approval. Second, the public does not care about politics so they don't pay attention to campaign functions like speeches, candidate gaffes, or advertising. Third, in presidential elections, each side has about the same amount of experts and money which cancels out any advantage those resources might bring.
Other scholars argue that campaigns do matter, and they also have three main reasons for this. One, that after controlling for incumbency, campaign spending does impact a candidates performance. Secondly, Gerber and Green argue that campaigns personalized outreach, such as, door-to-door or digital contact, will influence the person contacted to go and vote for the candidate. Finally, for non-presidential elections, the campaign resources vary enormously between candidates and the campaign with the most resources is favored to win the election.
|
|
1603.01417
| false
| null |
We speculate that there are two main reasons for this performance disparity, all exacerbated by the removal of supporting facts. First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU.
First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU.
|
Why is supporting fact supervision necessary for DMN?
|
The answers are shown as follows:
* First, the GRU only allows sentences to have context from sentences before them, but not after them. This prevents information propagation from future sentences. Second, the supporting sentences may be too far away from each other on a word level to allow for these distant sentences to interact through the word level GRU.
|
null | false
| null |
Give me a list of things to enable a support team on their technical strengths
|
1) Setup a daily stand up that becomes a platform for people to discuss their challenges. This platform will be a significant source of technical brainstorming. It enables people with technical competencies to guide the person asking for help. At the same time it allows for every one else to get familiar with an issue and its solution even if they haven't experienced it yet. Culturally this should be a judgement free zone. It should eliminate inhibitions to raise challenges.
2) Setup a cadence (weekly, bi weekly, monthly) for members of team to discuss their most interesting customer issues. They should cover the problem statement, symptom, assessment/diagnosis of the issue, approaches to solve the problem. The last piece around approaches to solve the problem is key as it equips others to incorporate effective options in their troubleshooting toolbox
3) Collaborate with product managers to setup a broader feature level enablement cadence.
4) Create mentorship programs where junior members can shadow seniors on their customer calls and learn technical cops as well as communication cops by just being a fly on the wall. At the same time seniors should shadow juniors on their customer calls and give them direct feedback on what needs to be improved
|
|
null | false
| null |
Can I travel to Canada from the United States with my dog?
|
You can travel into Canada with a dog as long as you have proof of an unexpired rabies vaccination.
|
|
null | false
| 16
|
Data annotation is a major bottleneck for the application of supervised learning approaches to many problems. As a result, unsupervised methods that learn directly from unlabeled data are increasingly important. For tasks related to unsupervised syntactic analysis, discrete generative models have dominated in recent years – for example, for both part-of-speech (POS) induction BIBREF0 , BIBREF1 and unsupervised dependency parsing BIBREF2 , BIBREF3 , BIBREF4 . While similar models have had success on a range of unsupervised tasks, they have mostly ignored the apparent utility of continuous word representations evident from supervised NLP applications BIBREF5 , BIBREF6 . In this work, we focus on leveraging and explicitly representing continuous word embeddings within unsupervised models of syntactic structure.
Pre-trained word embeddings from massive unlabeled corpora offer a compact way of injecting a prior notion of word similarity into models that would otherwise treat words as discrete, isolated categories. However, the specific properties of language captured by any particular embedding scheme can be difficult to control, and, further, may not be ideally suited to the task at hand. For example, pre-trained skip-gram embeddings BIBREF7 with small context window size are found to capture the syntactic properties of language well BIBREF8 , BIBREF9 . However, if our goal is to separate syntactic categories, this embedding space is not ideal – POS categories correspond to overlapping interspersed regions in the embedding space, evident in Figure SECREF4 .
In our approach, we propose to learn a new latent embedding space as a projection of pre-trained embeddings (depicted in Figure SECREF5 ), while jointly learning latent syntactic structure – for example, POS categories or syntactic dependencies. To this end, we introduce a new generative model (shown in Figure FIGREF6 ) that first generates a latent syntactic representation (e.g. a dependency parse) from a discrete structured prior (which we also call the “syntax model”), then, conditioned on this representation, generates a sequence of latent embedding random variables corresponding to each word, and finally produces the observed (pre-trained) word embeddings by projecting these latent vectors through a parameterized non-linear function. The latent embeddings can be jointly learned with the structured syntax model in a completely unsupervised fashion.
By choosing an invertible neural network as our non-linear projector, and then parameterizing our model in terms of the projection's inverse, we are able to derive tractable exact inference and marginal likelihood computation procedures so long as inference is tractable in the underlying syntax model. In sec:learn-with-inv we show that this derivation corresponds to an alternate view of our approach whereby we jointly learn a mapping of observed word embeddings to a new embedding space that is more suitable for the syntax model, but include an additional Jacobian regularization term to prevent information loss.
Recent work has sought to take advantage of word embeddings in unsupervised generative models with alternate approaches BIBREF9 , BIBREF10 , BIBREF11 , BIBREF12 . BIBREF9 build an HMM with Gaussian emissions on observed word embeddings, but they do not attempt to learn new embeddings. BIBREF10 , BIBREF11 , and BIBREF12 extend HMM or dependency model with valence (DMV) BIBREF2 with multinomials that use word (or tag) embeddings in their parameterization. However, they do not represent the embeddings as latent variables.
In experiments, we instantiate our approach using both a Markov-structured syntax model and a tree-structured syntax model – specifically, the DMV. We evaluate on two tasks: part-of-speech (POS) induction and unsupervised dependency parsing without gold POS tags. Experimental results on the Penn Treebank BIBREF13 demonstrate that our approach improves the basic HMM and DMV by a large margin, leading to the state-of-the-art results on POS induction, and state-of-the-art results on unsupervised dependency parsing in the difficult training scenario where neither gold POS annotation nor punctuation-based constraints are available.
To this end, we introduce a new generative model (shown in Figure 2) that first generates a latent syntactic representation (e.g. a dependency parse) from a discrete structured prior (which we also call the “syntax model”), then, conditioned on this representation, generates a sequence of latent embedding random variables corresponding to each word, and finally produces the observed (pre-trained) word embeddings by projecting these latent vectors through a parameterized non-linear function.
|
How does their model produce observed (pre-trained) word embeddings?
|
By projecting these latent vectors through a parameterized non-linear function.
|
null | false
| null |
Scrub Daddy is a cleaning product company best known for an eponymous sponge it manufactures in the shape of a smiley face. The product is made of a polymer which changes texture – firm in cold water and soft in warm water. As of 2019, Scrub Daddy had the highest revenue of any product successfully pitched on the ABC reality show Shark Tank.
History
After damaging the exterior of a vehicle when cleaning the outside, detailer Aaron Krause invented a line of buffing and polishing pads. The company was acquired by 3M in August 2008. 3M did not purchase a line of sponges Krause had invented, leaving them to sit in his factory. Five years later, Krause used the remaining sponges to clean his dishes and lawn furniture. According to Krause, this is when he "realized his multi-million dollar idea."
According to the company's website, Scrub Daddy, Inc. was founded in 2012 with grassroots marketing. On the Shark Tank episode originally airing in October 2012, Lori Greiner made a 20% equity stake deal with Krause for $200,000. The following day, Greiner and Krause sold out of 42,000 sponges in under seven minutes on QVC. Greiner then helped Scrub Daddy to be sold in retail stores such as Bed, Bath & Beyond. In January 2017, Scrub Daddy's total revenues surpassed $100 million – the highest of any Shark Tank product. As of October 2019, the company's lifetime sales were $209 million. In Season 14 Episode 13, it was reported that Scrub Daddy in 10 years has 273 employees, 160 products, sold in 257,000 retail locations, top 5 grossing company in shark tank history, doing over $670m in retail sales since launch.
|
Given a reference text about Scrub Daddy, tell me the details on the deal the founder made on Shark Tank.
|
Aaron Krause, the founder of Scrub Daddy, made a deal with Lori Greiner on Shark Tank, giving her 20% equity in the company for $200,000.
|
null | false
| 160
|
One thread of related work is the exploration of different kinds of compositional function over tree structures. BIBREF11 [ BIBREF11 ] proposed the recursive neural network with standard compositional function. After that, some extensions are introduced to enhance the expressive power of compositional function, such as MV-RecNN BIBREF23 , SU-RNN BIBREF13 , RNTN BIBREF23 , while these models suffer from the problem of hard-coded compositional operations and overfitting.
Another thread of work is the idea of using one network to direct the learning of another network BIBREF16 . BIBREF33 [ BIBREF33 ] introduce a meta neural network to provide another network with a step size and a direction vector, which is helpful for parameter optimization. BIBREF16 [ BIBREF16 ] propose the dynamic filter network to implicitly learn a variety of filtering operations. BIBREF17 [ BIBREF17 ] introduce a learnet for one-shot learning, which can predict the parameters of a second network given a single exemplar. BIBREF18 [ BIBREF18 ] propose the model hypernetwork, which uses a small network to generate the weights for a larger network.
Different from these models, we employ the idea of parameter generation to address the limitation of weight-sharing or partially sharing paradigm of tree-based compositional models.
One thread of related work is the exploration of different kinds of compositional function over tree structures. Socher et al. proposed the recursive neural network with standard compositional function. After that, some extensions are introduced to enhance the expressive power of compositional function, such as MV-RecNN, SURNN, RNTN, while these models suffer from the problem of hard-coded compositional operations and overfitting.
|
What methods have other researchers proposed to explore the combinatorial functions of different kinds on tree structures?
|
Socher et al. proposed the recursive neural network with standard compositional function.
|
null | false
| null |
Tell me about the discovery of coffee bean?
|
Coffee was first discovered by the 9th century Ethiopian goat-herder Kaldi. According to a story written down in 1671. Kaldi noticed that his goats behaving erratically when eating the coffee bean in a region called Kaffa in Ethiopia. It made its way north, across the red sea into Yemen in the 15th Century. It then started to be grown across the world.
|
|
null | false
| null |
Why didn’t people like the end of Game of Thrones?
|
Note: this answer contains spoilers for the end of the HBO television series, “Game of Thrones.” The end of the Game of Thrones television series was a bit divisive among viewers for a number of reasons. First, some believed that one of the series’ principal and most popular characters, Daenerys Targaryen, had been – if flawed – a protagonist throughout most of the series until the last few episodes of the final season, and her switch to a villainous persona felt rushed. Similarly, some viewers believed that arcs for other main characters, including Jaime, Cersei, and Tyrion Lannister, Brienne of Tarth, Jon Snow, and others, were adversely impacted by the quick pacing of the final season and did away with much of those characters’ development over the course of many years. Others were confused and disappointed that Bran Stark became king of the Seven Kingdoms (and ultimate winner of the titular “game of thrones”), as there seemed to be little build-up or motivation for that narrative decision. Many suspect that George R.R. Martin, author of the “Song of Ice and Fire” series of novels on which the show was based, may have anticipated providing more narrative context and development for these choices in his then-unpublished final novels in the series, and that the television series simply needed to end well before Martin could complete his novels.
|
|
2002.02224
| false
| null |
In order to obtain the citation data of the Czech apex courts, it was necessary to recognize and extract the references from the CzCDC 1.0. Given that training data for both the reference recognition model BIBREF13, BIBREF34 and the text segmentation model BIBREF33 are publicly available, we were able to conduct extensive error analysis and put together a pipeline to arguably achieve the maximum efficiency in the task. The pipeline described in this part is graphically represented in Figure FIGREF10.
At this point, it is necessary to evaluate the performance of the above mentioned part of the pipeline before proceeding further. The evaluation of the performance is summarised in Table TABREF11. It shows that organising the two models into the pipeline boosted the performance of the reference recognition model, leading to a higher F1 measure in the initial recognition of the text spans and their classification.
In order to obtain the citation data of the Czech apex courts, it was necessary to recognize and extract the references from the CzCDC 1.0. Given that training data for both the reference recognition model BIBREF13, BIBREF34 and the text segmentation model BIBREF33 are publicly available, we were able to conduct extensive error analysis and put together a pipeline to arguably achieve the maximum efficiency in the task. The pipeline described in this part is graphically represented in Figure FIGREF10.
At this point, it is necessary to evaluate the performance of the above mentioned part of the pipeline before proceeding further. The evaluation of the performance is summarised in Table TABREF11. It shows that organising the two models into the pipeline boosted the performance of the reference recognition model, leading to a higher F1 measure in the initial recognition of the text spans and their classification.
|
How is quality of the citation measured?
|
The answers are shown as follows:
* it is necessary to evaluate the performance of the above mentioned part of the pipeline before proceeding further. The evaluation of the performance is summarised in Table TABREF11. It shows that organising the two models into the pipeline boosted the performance of the reference recognition model, leading to a higher F1 measure in the initial recognition of the text spans and their classification.
|
null | false
| null |
European-Americans first settled Nevada City in 1849, during the California Gold Rush, as Nevada (Spanish for "snow-covered", a reference to the snow-topped mountains in the area). The Gold Tunnel on the north side of Deer Creek was the city's first mine, built in 1850. The first sawmill in Nevada City was built on Deer Creek, just above town, in August 1850, by Lewis & Son, with a water wheel. In 1850–51, Nevada City was the state's most important mining town, and Nevada County the state's leading gold-mining county. In 1851, The Nevada Journal became the first newspaper published in the town and county. The first cemetery in town, the Pioneer Cemetery, was founded around 1851 behind the Nevada City United Methodist Church, Nevada County's first denominational church.
The town of Nevada was incorporated on April 19, 1856. In 1864, the word “City” was added to its name to relieve confusion with the nearby state of Nevada, and the town has legally been known as Nevada City ever since. The former town of Coyoteville later became Nevada City's northwestern section.
|
What is Nevada City known for?
|
Nevada City is a town in Northern California that was famous during the California Gold Rush that was first settled in 1849.
|
1909.02764
| false
| null |
Table TABREF16 shows the confusion matrices for facial and audio emotion recognition on our complete AMMER data set and Table TABREF17 shows the results per class for each method, including facial and audio data and micro and macro averages. The classification from facial expressions yields a macro-averaged $\text{F}_1$ score of 33 % across the three emotions joy, insecurity, and annoyance (P=0.31, R=0.35). While the classification results for joy are promising (R=43 %, P=57 %), the distinction of insecurity and annoyance from the other classes appears to be more challenging.
Table TABREF16 shows the confusion matrices for facial and audio emotion recognition on our complete AMMER data set and Table TABREF17 shows the results per class for each method, including facial and audio data and micro and macro averages. The classification from facial expressions yields a macro-averaged $\text{F}_1$ score of 33 % across the three emotions joy, insecurity, and annoyance (P=0.31, R=0.35).
|
How is face and audio data analysis evaluated?
|
The answers are shown as follows:
* confusion matrices
* $\text{F}_1$ score
|
null | false
| null |
The "Garbage collection" log of a JVM is formatted as follows:
1. Each line represents a single garbage collection operation starting at the timestamp.
2. The before and after sizes of different memory areas in the JVM are shown as "MemoryArea : BeforeSize->AfterSize(AllocatedSize)" where MemoryArea is one of PSYoungGen, ParOldGen, or Metaspace.
3. If the "MemoryArea:" is omitted, it represents the before and after size of the entire JVM's memory.
4. Each line has the time taken for the operation in seconds.
Summarize the following log and note any significant anomalies:
2023-03-30T07:00:19.800+0000: [GC (Allocation Failure) [PSYoungGen: 17197776K->2224032K(21782528K)] 64496630K->49524856K(79218176K), 3.2658630 secs] [Times: user=4.53 sys=0.00, real=3.27 secs]
2023-03-30T07:01:06.553+0000: [GC (Allocation Failure) [PSYoungGen: 17471392K->2195300K(22969344K)] 64772216K->49530782K(80404992K), 3.3074224 secs] [Times: user=4.63 sys=0.00, real=3.30 secs]
2023-03-30T07:01:56.129+0000: [GC (Allocation Failure) [PSYoungGen: 19045732K->2429792K(22598656K)] 66381214K->49767742K(80034304K), 3.5912859 secs] [Times: user=4.94 sys=0.00, real=3.59 secs]
2023-03-30T07:02:46.034+0000: [GC (Allocation Failure) [PSYoungGen: 19280224K->2428421K(23520768K)] 66618174K->49768148K(80956416K), 3.6520001 secs] [Times: user=5.07 sys=0.03, real=3.65 secs]
2023-03-30T07:03:39.130+0000: [GC (Allocation Failure) [PSYoungGen: 20488709K->2600800K(23257088K)] 67828436K->49943004K(80692736K), 3.8378192 secs] [Times: user=5.19 sys=0.00, real=3.84 secs]
2023-03-30T07:04:31.634+0000: [GC (Allocation Failure) [PSYoungGen: 20661088K->2550592K(23885312K)] 68003292K->49894476K(81320960K), 3.7886199 secs] [Times: user=5.15 sys=0.00, real=3.78 secs]
2023-03-30T07:05:28.784+0000: [GC (Allocation Failure) [PSYoungGen: 21416768K->2709510K(23698432K)] 68760652K->50055163K(81134080K), 3.9951697 secs] [Times: user=5.54 sys=0.00, real=3.99 secs]
2023-03-30T07:06:24.857+0000: [GC (Allocation Failure) [PSYoungGen: 21575686K->2709696K(24113664K)] 68921339K->50058933K(81549312K), 4.0210395 secs] [Times: user=5.47 sys=0.01, real=4.02 secs]
2023-03-30T07:07:21.991+0000: [GC (Allocation Failure) [PSYoungGen: 22106304K->2835749K(24000512K)] 69455541K->50186794K(81436160K), 4.0703042 secs] [Times: user=5.76 sys=0.00, real=4.06 secs]
2023-03-30T07:08:18.668+0000: [GC (Allocation Failure) [PSYoungGen: 22232357K->2785312K(24265216K)] 69583402K->50204626K(81700864K), 4.1296625 secs] [Times: user=5.77 sys=0.00, real=4.13 secs]
2023-03-30T07:09:16.891+0000: [GC (Allocation Failure) [PSYoungGen: 22510624K->2834405K(24177664K)] 69929938K->50255520K(81613312K), 4.2070487 secs] [Times: user=5.89 sys=0.01, real=4.21 secs]
2023-03-30T07:10:15.553+0000: [GC (Allocation Failure) [PSYoungGen: 22559717K->2842896K(24403456K)] 69980832K->50266688K(81839104K), 4.2489383 secs] [Times: user=5.83 sys=0.02, real=4.24 secs]
2023-03-30T07:11:15.412+0000: [GC (Allocation Failure) [PSYoungGen: 22863632K->2880069K(24334848K)] 70287424K->50306742K(81770496K), 4.2983311 secs] [Times: user=6.01 sys=0.00, real=4.29 secs]
2023-03-30T07:12:17.330+0000: [GC (Allocation Failure) [PSYoungGen: 22900805K->2670097K(24596992K)] 70327478K->50099432K(82032640K), 3.9450690 secs] [Times: user=5.44 sys=0.00, real=3.95 secs]
2023-03-30T07:13:15.713+0000: [GC (Allocation Failure) [PSYoungGen: 23009297K->2684375K(24459776K)] 70438632K->50115773K(81895424K), 3.9758416 secs] [Times: user=5.53 sys=0.00, real=3.97 secs]
2023-03-30T07:14:12.939+0000: [GC (Allocation Failure) [PSYoungGen: 23023575K->2678912K(24829952K)] 70454973K->50113093K(82265600K), 3.9702778 secs] [Times: user=5.52 sys=0.00, real=3.97 secs]
2023-03-30T07:15:12.343+0000: [GC (Allocation Failure) [PSYoungGen: 23508608K->2753575K(24717312K)] 70942789K->50189628K(82152960K), 4.0754481 secs] [Times: user=5.72 sys=0.00, real=4.08 secs]
2023-03-30T07:16:13.026+0000: [GC (Allocation Failure) [PSYoungGen: 23583271K->2762097K(24974336K)] 71019324K->50201762K(82409984K), 4.1128461 secs] [Times: user=5.66 sys=0.00, real=4.11 secs]
2023-03-30T07:17:14.129+0000: [GC (Allocation Failure) [PSYoungGen: 23924593K->2797957K(24905728K)] 71364258K->50239629K(82341376K), 4.1456776 secs] [Times: user=5.74 sys=0.01, real=4.15 secs]
2023-03-30T07:18:14.857+0000: [GC (Allocation Failure) [PSYoungGen: 23960453K->2804721K(25075712K)] 71402125K->50249103K(82511360K), 4.1905285 secs] [Times: user=5.73 sys=0.01, real=4.19 secs]
2023-03-30T07:19:15.979+0000: [GC (Allocation Failure) [PSYoungGen: 24189937K->3641846K(25027072K)] 71634319K->51171235K(82462720K), 3.6175882 secs] [Times: user=5.94 sys=0.00, real=3.62 secs]
2023-03-30T07:22:24.484+0000: [GC (Allocation Failure) [PSYoungGen: 25027062K->3360979K(24336896K)] 72556451K->52269877K(81772544K), 0.4407322 secs] [Times: user=5.66 sys=0.00, real=0.44 secs]
2023-03-30T07:22:38.974+0000: [GC (Allocation Failure) [PSYoungGen: 24007379K->4035567K(24681984K)] 72916277K->57145380K(82117632K), 0.8531910 secs] [Times: user=10.80 sys=0.23, real=0.85 secs]
2023-03-30T07:22:52.666+0000: [GC (Allocation Failure) [PSYoungGen: 24677029K->24677029K(24681984K)] 77786841K->82112670K(82117632K), 7.3509182 secs] [Times: user=22.60 sys=11.27, real=7.35 secs]
2023-03-30T07:23:00.017+0000: [Full GC (Ergonomics) [PSYoungGen: 24677029K->0K(24681984K)] [ParOldGen: 57435641K->57435322K(57435648K)] 82112670K->57435322K(82117632K), [Metaspace: 241941K->241941K(260096K)], 26.4487596 secs] [Times: user=313.82 sys=2.44, real=26.45 secs]
2023-03-30T07:24:07.186+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->12427037K(24681984K)] [ParOldGen: 57435322K->57435609K(57435648K)] 78081722K->69862646K(82117632K), [Metaspace: 241947K->241947K(260096K)], 28.8675082 secs] [Times: user=350.97 sys=1.74, real=28.87 secs]
2023-03-30T07:24:36.057+0000: [Full GC (System.gc()) [PSYoungGen: 12730000K->12427055K(24681984K)] [ParOldGen: 57435609K->57435556K(57435648K)] 70165609K->69862611K(82117632K), [Metaspace: 241947K->241947K(260096K)], 31.3736816 secs] [Times: user=379.38 sys=2.94, real=31.37 secs]
2023-03-30T07:25:18.096+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->16985330K(24681984K)] [ParOldGen: 57435556K->57435308K(57435648K)] 78081956K->74420638K(82117632K), [Metaspace: 241999K->241999K(260096K)], 31.4762980 secs] [Times: user=363.38 sys=3.10, real=31.48 secs]
2023-03-30T07:25:54.537+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->19116969K(24681984K)] [ParOldGen: 57435308K->57435152K(57435648K)] 78081708K->76552122K(82117632K), [Metaspace: 241999K->241999K(260096K)], 31.0418139 secs] [Times: user=377.34 sys=2.75, real=31.04 secs]
2023-03-30T07:26:27.487+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->19955901K(24681984K)] [ParOldGen: 57435152K->57435290K(57435648K)] 78081552K->77391191K(82117632K), [Metaspace: 241999K->241999K(260096K)], 22.9475977 secs] [Times: user=280.80 sys=1.57, real=22.95 secs]
2023-03-30T07:26:51.319+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20340549K(24681984K)] [ParOldGen: 57435290K->57435523K(57435648K)] 78081690K->77776072K(82117632K), [Metaspace: 242004K->242004K(260096K)], 37.2564843 secs] [Times: user=458.29 sys=3.35, real=37.26 secs]
2023-03-30T07:27:28.892+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20465997K(24681984K)] [ParOldGen: 57435523K->57435230K(57435648K)] 78081923K->77901227K(82117632K), [Metaspace: 242007K->242007K(260096K)], 31.4213545 secs] [Times: user=382.65 sys=2.74, real=31.42 secs]
2023-03-30T07:28:00.350+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20492397K(24681984K)] [ParOldGen: 57435230K->57435139K(57435648K)] 78081630K->77927536K(82117632K), [Metaspace: 242011K->242011K(260096K)], 17.3069966 secs] [Times: user=209.34 sys=0.45, real=17.31 secs]
2023-03-30T07:28:17.694+0000: [Full GC (Ergonomics) [PSYoungGen: 20639999K->20482496K(24681984K)] [ParOldGen: 57435139K->57435581K(57435648K)] 78075138K->77918078K(82117632K), [Metaspace: 242023K->242023K(260096K)], 39.0253664 secs] [Times: user=478.37 sys=3.87, real=39.02 secs]
2023-03-30T07:28:56.752+0000: [Full GC (Ergonomics) [PSYoungGen: 20629482K->20490559K(24681984K)] [ParOldGen: 57435581K->57435269K(57435648K)] 78065064K->77925828K(82117632K), [Metaspace: 242023K->242023K(260096K)], 32.7146380 secs] [Times: user=398.86 sys=2.93, real=32.71 secs]
2023-03-30T07:29:29.592+0000: [Full GC (Ergonomics) [PSYoungGen: 20627596K->20498740K(24681984K)] [ParOldGen: 57435269K->57435482K(57435648K)] 78062865K->77934223K(82117632K), [Metaspace: 242029K->242029K(260096K)], 39.9805382 secs] [Times: user=491.39 sys=4.10, real=39.98 secs]
2023-03-30T07:30:09.618+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20529006K(24681984K)] [ParOldGen: 57435482K->57435402K(57435648K)] 78081882K->77964408K(82117632K), [Metaspace: 242038K->242038K(260096K)], 31.3632706 secs] [Times: user=382.46 sys=2.74, real=31.36 secs]
2023-03-30T07:30:41.012+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20535638K(24681984K)] [ParOldGen: 57435402K->57435345K(57435648K)] 78081802K->77970983K(82117632K), [Metaspace: 242053K->242053K(260096K)], 31.0060106 secs] [Times: user=377.25 sys=2.72, real=31.00 secs]
2023-03-30T07:31:12.022+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20535719K(24681984K)] [ParOldGen: 57435345K->57435297K(57435648K)] 78081745K->77971016K(82117632K), [Metaspace: 242053K->242053K(260096K)], 31.1714473 secs] [Times: user=380.42 sys=2.74, real=31.18 secs]
2023-03-30T07:31:43.215+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20536490K(24681984K)] [ParOldGen: 57435297K->57435275K(57435648K)] 78081697K->77971766K(82117632K), [Metaspace: 242061K->242061K(260096K)], 30.9676462 secs] [Times: user=377.19 sys=2.88, real=30.96 secs]
2023-03-30T07:32:14.216+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20537679K(24681984K)] [ParOldGen: 57435275K->57435244K(57435648K)] 78081675K->77972923K(82117632K), [Metaspace: 242081K->242081K(260096K)], 31.2592798 secs] [Times: user=379.77 sys=3.04, real=31.26 secs]
2023-03-30T07:32:45.532+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20518698K(24681984K)] [ParOldGen: 57435244K->57435465K(57435648K)] 78081644K->77954164K(82117632K), [Metaspace: 242123K->242123K(260096K)], 38.9523351 secs] [Times: user=478.26 sys=4.04, real=38.95 secs]
2023-03-30T07:33:45.139+0000: [Full GC (Ergonomics) [PSYoungGen: 20627835K->20519977K(24681984K)] [ParOldGen: 57435465K->57435334K(57435648K)] 78063301K->77955311K(82117632K), [Metaspace: 242185K->242185K(260096K)], 31.8680480 secs] [Times: user=388.66 sys=2.88, real=31.87 secs]
2023-03-30T07:34:17.045+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20529987K(24681984K)] [ParOldGen: 57435334K->57435313K(57435648K)] 78081734K->77965301K(82117632K), [Metaspace: 242185K->242185K(260096K)], 30.6528436 secs] [Times: user=370.84 sys=3.05, real=30.65 secs]
2023-03-30T07:35:08.312+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20552578K(24681984K)] [ParOldGen: 57435313K->57435312K(57435648K)] 78081713K->77987890K(82117632K), [Metaspace: 242202K->242202K(260096K)], 31.3030289 secs] [Times: user=380.50 sys=2.90, real=31.30 secs]
2023-03-30T07:35:39.649+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20541558K(24681984K)] [ParOldGen: 57435312K->57435297K(57435648K)] 78081712K->77976855K(82117632K), [Metaspace: 242217K->242217K(260096K)], 31.4996226 secs] [Times: user=384.32 sys=3.00, real=31.50 secs]
2023-03-30T07:36:11.199+0000: [Full GC (Ergonomics) [PSYoungGen: 20646400K->20562436K(24681984K)] [ParOldGen: 57435297K->57435260K(57435648K)] 78081697K->77997697K(82117632K), [Metaspace: 242238K->242238K(260096K)], 31.1358920 secs] [Times: user=378.46 sys=3.00, real=31.14 secs]
|
1. The log has 47 GC operations over a 36-minute interval from 7:00 AM to 7:36 AM on 30th March 2023 UTC.
2. The GC operations start at roughly one per minute and take less than five seconds each, but after 7:23 AM, they abruptly take about 30 seconds each and occur twice a minute. In other words, initially, GC operations take up a small fraction of the JVM time, while later, they take up almost all of its time.
3. GC operations in this second period result in a very slight decrease in the total memory size and the PSYoungGen, ParOldGen, and Metaspace memory areas. This behavior is especially acute after 7:26:51 AM.
4. This indicates much time spent in GC operations with little result.
5. The total memory size initially slowly grows from 49.5GB to 52.3 GB, but in the second period maxes out at around 78 GB.
6. The allocated memory size is fixed at 82 GB during the second period of GC operations, indicating an outer limit.
7. I recommend increasing the maximum allocated memory beyond the 82 GB limit, doubling it if possible, and re-running the JVM.
|
|
null | false
| null |
Is it fun to play guitar? How do you start playing guitar?
|
Playing guitar is a worthwhile and super fun exercise for anyone, but especially those that are musically inclined. Even for a beginner, its not difficult to be able to start playing simple songs from great artists like the Rolling Stones, The Beatles, Jimi Hendrix, or even Led Zeppelin. The best way to get started is by understanding open chords, both major and minor, and simple string picking patterns.
|
|
1710.07395
| false
| null |
For logistic regression model implementation, we use l2 loss. We adopt the balanced class weight as described in Scikit learn. Logistic regression model with character-level n-gram features is presented as a strong baseline for comparison since it was shown very effective. BIBREF0 , BIBREF9
Logistic regression model with character-level n-gram features is presented as a strong baseline for comparison since it was shown very effective.
|
What is their baseline?
|
The answers are shown as follows:
* Logistic regression model with character-level n-gram features
|
null | false
| 170
|
We applied this method to discover gender-associated words in both domains. In Table TABREF9, we present a sample of the most gender-associated nouns from the celebrity domain. Several themes emerge: for example, female celebrities seem to be more associated with appearance (“gown,” “photo,” “hair,” “look”), while male celebrities are more associated with creating content (“movie,” “film,” “host,” “director”). This echoes real-world trends: for instance, on the red carpet, actresses tend to be asked more questions about their appearance –- what brands they are wearing, how long it took to get ready, etc. –- while actors are asked questions about their careers and creative processes (as an example, see BIBREF31).
Table TABREF9 also includes some of the most gender-associated verbs and adjectives from the professor domain. Female CS professors seem to be praised for being communicative and personal with students (“respond,” “communicate,” “kind,” “caring”), while male CS professors are recognized for being knowledgeable and challenging the students (“teach,”, “challenge,” “brilliant,” “practical”). These trends are well-supported by social science literature, which has found that female teachers are praised for “personalizing” instruction and interacting extensively with students, while male teachers are praised for using “teacher as expert” styles that showcase mastery of material BIBREF32.
These findings establish that there are clear differences in how people talk about women and men – even with Bonferroni correction, there are still over 500 significantly gender-associated nouns, verbs, and adjectives in the celebrity domain and over 200 in the professor domain. Furthermore, the results in both domains align with prior studies and real world trends, which validates that our methods can capture meaningful patterns and innovatively provide evidence at the large-scale. This analysis also hints that it can be helpful to abstract from words to topics to recognize higher-level patterns of gender associations, which motivates our next section on clustering.
Table TABREF11 displays a sample of our results – we find that the clusters are coherent in context and the labels seem reasonable. In the next section, we discuss human evaluations that we conducted to more rigorously evaluate the output, but first we discuss the value of these methods toward analysis.
At the word-level, we hypothesized that in the celebrity domain, women were more associated with appearance and men with creating content. Now, we can validate those hypotheses against labeled clusters – indeed, there is a cluster labeled clothing that is 100% female (i.e. 100% words are female-associated), and a 80% male cluster labeled movie. Likewise, in the professor domain, we had guessed that women are associated with communication and men with knowledge, and there is a 100% female cluster labeled communication and a 89% male cluster labeled cognition. Thus, cluster labeling proves to be very effective at pulling out the patterns that we believed we saw at the word-level, but could not formally validate.
The clusters we mentioned so far all lean heavily toward one gender association or the other, but some clusters are interesting precisely because they do not lean heavily – this allows us to see where semantic groupings do not align exactly with gender association. For example, in the celebrity domain, there is a cluster labeled lover that has a mix of female-associated words (“boyfriend,” “beau,” “hubby”) and male-associated words (“wife,” “girlfriend”). Jointly leveraging cluster labels and gender associations allows us to see that in the semantic context of having a lover, women are typically associated with male figures and men with female figures, which reflects heteronormativity in society.
These findings establish that there are clear differences in how people talk about women and men – even with Bonferroni correction, there are still over 500 significantly gender-associated nouns, verbs, and adjectives in the celebrity domain and over 200 in the professor domain.
|
Are there clear differences in how people talk about women and men?
|
Yes, there are.
|
null | false
| null |
Name some of the bravolebrities from Vanderpump Rules
|
James Kennedy, Katie Maloney, Raquel Leviss, Tom Sandoval, Ariana Madix, Tom Schwartz and Lala Kent
|
|
2002.08899
| true
| null |
Additionally, we provide evidence that the model learns knowledge of a separation between syntax and the lexicon that is similar to that of a human. Figure FIGREF6 displays the learned $\sigma (w)$ embeddings for some input words, across the domains. To avoid cherry-picking the results, we chose the input words arbitrarily, subject to the following constraint. We considered each word to typically have a different syntactic category than the other choices from that domain. This constraint was used to present a diverse selection of words. Table TABREF5 displays the output behavior of models that we damaged to resemble the damage that causes aphasia. To avoid cherry-picking the results, we arbitrarily chose an input for each domain, subject to the following constraint. The input is not in the train set and the undamaged LLA-LSTM model produces a translation that we judge to be correct. For all inputs that we chose, damage to the analog of Broca's area (the LSTMs) results in an output that describes content only if it is described by the input. However, the output does not show understanding of the input's syntax. In the naturalistic domains, damage to the analog of Wernicke's area (the Lexicon Unit) results in an output with incorrect content that would be acceptable if the input had different words but the same syntax. These knowledge distortions are precisely those that are expected in the respective human aphasics BIBREF0. We also provide corpus-level results from the damaged models by presenting mean precision on the test sets. Because the output languages in all of our domains use tokens to represent meanings in many cases, it is expected that the analog to Wernicke's area is responsible for maintaining a high precision.
TABREF5 displays the output behavior of models that we damaged to resemble the damage that causes aphasia.
Because the output languages in all of our domains use tokens to represent meanings in many cases, it is expected that the analog to Wernicke's area is responsible for maintaining a high precision.
|
Do they perform a quantitative analysis of their model displaying knowledge distortions?
|
Yes.
|
null | false
| null |
What's the best way to brew an espresso?
|
There is no general consensus on how to evaluate a cup of espresso, so there is no general "best way". Here are a few pointers that most of the coffee-lover community agrees will enhance the consistency and flavor of an espresso:
* Use freshly roasted beans! No amount of expensive espresso machines can save a stale batch of beans.
* Make sure your puck is evenly distributed. If needed, use a WDT tool to stir the puck gently from bottom to top, until you reach consistency.
* Observe the brewing process closely and make sure the fluid flows smoothly. If the stream is unbalanced or side-splitting, check your puck and your machine's pressurization.
|
|
null | false
| null |
what is typical enterprise sales person comp mix between base and commission?
|
50/50
|
|
null | false
| null |
Academic literature uses the abstract to succinctly communicate complex research. An abstract may act as a stand-alone entity instead of a full paper. As such, an abstract is used by many organizations as the basis for selecting research that is proposed for presentation in the form of a poster, platform/oral presentation or workshop presentation at an academic conference. Most bibliographic databases only index abstracts rather than providing the entire text of the paper. Full texts of scientific papers must often be purchased because of copyright and/or publisher fees and therefore the abstract is a significant selling point for the reprint or electronic form of the full text. The abstract can convey the main results and conclusions of a scientific article but the full text article must be consulted for details of the methodology, the full experimental results, and a critical discussion of the interpretations and conclusions. An abstract allows one to sift through copious numbers of papers for ones in which the researcher can have more confidence that they will be relevant to their research. Once papers are chosen based on the abstract, they must be read carefully to be evaluated for relevance. It is generally agreed that one must not base reference citations on the abstract alone, but the content of an entire paper.
|
Give me a summary of why an abstract is useful in academic literature?
|
An abstract is useful in academic research because it conveys the complex research in a concise way by acting as a stand-alone summary. Therefore, it is usually used by many organizations as a base for selecting research. Moreover, abstracts are usually provided free of charge by many bibliographic databases which makes the research works more accessible. In addition, researchers can understand if the work is relevant to their research just by reading the abstract.
|
null | false
| 11
|
Blogging gained momentum in 1999 and became especially popular after the launch of freely available, hosted platforms such as blogger.com or livejournal.com. Blogging has progressively been used by individuals to share news, ideas, and information, but it has also developed a mainstream role to the extent that it is being used by political consultants and news services as a tool for outreach and opinion forming as well as by businesses as a marketing tool to promote products and services BIBREF0 .
For this paper, we compiled a very large geolocated collection of blogs, written by individuals located in the U.S., with the purpose of creating insightful mappings of the blogging community. In particular, during May-July 2015, we gathered the profile information for all the users that have self-reported their location in the U.S., along with a number of posts for all their associated blogs. We utilize this blog collection to generate maps of the U.S. that reflect user demographics, language use, and distributions of psycholinguistic and semantic word classes. We believe that these maps can provide valuable insights and partial verification of previous claims in support of research in linguistic geography BIBREF1 , regional personality BIBREF2 , and language analysis BIBREF3 , BIBREF4 , as well as psychology and its relation to human geography BIBREF5 .
For this paper, we compiled a very large geolocated collection of blogs, written by individuals located in the U.S., with the purpose of creating insightful mappings of the blogging community.
|
What did the author collect in this article?
|
For this paper, the author compiled a very large geolocated collection of blogs, written by individuals located in the U.S., with the purpose of creating insightful mappings of the blogging community.
|
null | false
| null |
Please classify each of the following as European or Asian countries: England, Thailand, Greece, Japan.
|
England and Greece are European countries, while Thailand and Japan are Asian countries.
|
|
null | false
| null |
what is a computer?
|
A computer is a machine that can be programmed to carry out sequences of arithmetic or logical operations (computation) automatically.
|
|
1911.03597
| false
| null |
We adopt the mixture of two multilingual translation corpus as our training data: MultiUN BIBREF20 and OpenSubtitles BIBREF21. MultiUN consists of 463,406 official documents in six languages, containing around 300M words for each language. OpenSubtitles is a corpus consisting of movie and TV subtitles, which contains 2.6B sentences over 60 languages. We select four shared languages of the two corpora: English, Spanish, Russian and Chinese. Statistics of the training corpus are shown in Table TABREF14. Sentences are tokenized by Wordpiece as in BERT. A multilingual vocabulary of 50K tokens is used. For validation and testing, we randomly sample 10000 sentences respectively from each language pair. The rest data are used for training. For monolingual pre-training, we use English Wikipedia corpus, which contains 2,500M words.
We adopt the mixture of two multilingual translation corpus as our training data: MultiUN BIBREF20 and OpenSubtitles BIBREF21. MultiUN consists of 463,406 official documents in six languages, containing around 300M words for each language. OpenSubtitles is a corpus consisting of movie and TV subtitles, which contains 2.6B sentences over 60 languages. We select four shared languages of the two corpora: English, Spanish, Russian and Chinese. Statistics of the training corpus are shown in Table TABREF14.
|
What multilingual parallel data is used for training proposed model?
|
The answers are shown as follows:
* MultiUN BIBREF20
* OpenSubtitles BIBREF21
|
null | false
| 91
|
To make the conclusion convincing, we first choose two large-scale datasets that are publicly available, i.e., Chinese-English and English-French. Since English, French, and Chinese all belong to the subject-verb-object (SVO) family, we choose another very different subject-object-verb (SOV) language, Japanese, which might bring some interesting linguistic behaviors in English-Japanese translation.
For Chinese-English task, we use WMT17 Chinese-English dataset that consists of $20.6$M sentence pairs. For English-French task, we use WMT14 English-French dataset that comprises $35.5$M sentence pairs. For English-Japanese task, we follow BIBREF17 to use the first two sections of WAT17 English-Japanese dataset that consists of $1.9$M sentence pairs. Following the standard NMT procedure, we adopt the standard byte pair encoding (BPE) BIBREF18 with 32K merge operations for all language pairs. We believe that these datasets are large enough to confirm the rationality and validity of our experimental analyses.
We choose the state-of-the-art Transformer BIBREF1 model and the conventional RNN-Search model BIBREF0 as our test bed. We implement the Attribution method based on the Fairseq-py BIBREF19 framework for the above models. All models are trained on the training corpus for 100k steps under the standard settings, which achieve comparable translation results. All the following experiments are conducted on the test dataset, and we estimate the input word importance using the model generated hypotheses.
In the following experiments, we compare IG (Attribution) with several black-box methods (i.e., Content, Frequency, Attention) as introduced in Section SECREF8. In Section SECREF21, to ensure that the translation performance decrease attributes to the selected words instead of the perturbation operations, we randomly select the same number of words to perturb (Random), which serves as a baseline. Since there is no ranking for content words, we randomly select a set of content words as important words. To avoid the potential bias introduced by randomness (i.e., Random and Content), we repeat the experiments for 10 times and report the averaged results. We calculate the Attention importance in a similar manner as the Attribution, except that the attention scores use a max operation due to the better performance.
We evaluate the effectiveness of estimating word importance by the translation performance decrease. More specifically, unlike the usual way, we measure the decrease of translation performance when perturbing a set of important words that are of top-most word importance in a sentence. The more translation performance degrades, the more important the word is.
We use the standard BLEU score as the evaluation metric for translation performance. To make the conclusion more convincing, we conduct experiments on different types of synthetic perturbations (Section SECREF21), as well as different NMT architectures and language pairs (Section SECREF27). In addition, we compare with a supervised erasure method, which requires ground-truth translations for scoring word importance (Section SECREF30).
In this experiment, we investigate the effectiveness of word importance estimation methods under different synthetic perturbations. Since the perturbation on text is notoriously hard BIBREF20 due to the semantic shifting problem, in this experiment, we investigate three types of perturbations to avoid the potential bias :
Deletion perturbation removes the selected words from the input sentence, and it can be regarded as a specific instantiation of sentence compression BIBREF21.
Mask perturbation replaces embedding vectors of the selected words with all-zero vectors BIBREF22, which is similar to Deletion perturbation except that it retains the placeholder.
Grammatical Replacement perturbation replaces a word by another word of the same linguistic role (i.e., POS tags), yielding a sentence that is grammatically correct but semantically nonsensical BIBREF23, BIBREF24, such as “colorless green ideas sleep furiously”.
Figure FIGREF19 illustrates the experimental results on Chinese$\Rightarrow $English translation with Transformer. It shows that Attribution method consistently outperforms other methods against different perturbations on a various number of operations. Here the operation number denotes the number of perturbed words in a sentence. Specifically, we can make the following observations.
Under three different perturbations, perturbing words of top-most importance leads to lower BLEU scores than Random selected words. It confirms the existence of important words, which have greater impacts on translation performance. Furthermore, perturbing important words identified by Attribution outperforms the Random method by a large margin (more than 4.0 BLEU under 5 operations).
Figure FIGREF19 shows that two black-box methods (i.e., Content, Frequency) perform only slightly better than the Random method. Specifically, the Frequency method demonstrates even worse performances under the Mask perturbation. Therefore, linguistic properties (such as POS tags) and the word frequency can only partially help identify the important words, but it is not as accurate as we thought. In the meanwhile, it is intriguing to explore what exact linguistic characteristics these important words reveal, which will be introduced in Section SECREF5.
We also evaluate the Attention method, which bases on the encoder-decoder attention scores at the last layer of Transformer. Note that the Attention method is also used to simulate the best black-box method SOCRAT, and the results show that it is more effective than black-box methods and the Random baseline. Given the powerful Attention method, Attribution method still achieves best performances under all three perturbations. Furthermore, we find that the gap between Attribution and Attention is notably large (around $1.0+$ BLEU difference). Attention method does not provide as accurate word importance as the Attribution, which exhibits the superiority of gradient-based methods and consists with the conclusion reported in the previous study BIBREF8.
In addition, as shown in Figure FIGREF19, the perturbation effectiveness of Deletion, Mask, and Grammatical Replacement varies from strong to weak. In the following experiments, we choose Mask as the representative perturbation operation for its moderate perturbation performance, based on which we compare two most effective methods Attribution and Attention.
We validate the effectiveness of the proposed approach using a different NMT architecture RNN-Search on the Chinese$\Rightarrow $English translation task. The results are shown in Figure FIGREF20(a). We observe that the Attribution method still outperforms both Attention method and Random method by a decent margin. By comparing to Transformer, the results also reveal that the RNN-Search model is less robust to these perturbations. To be specific, under the setting of five operations and Attribution method, Transformer shows a relative decrease of $55\%$ on BLEU scores while the decline of RNN-Search model is $64\%$.
We further conduct experiments on another two language pairs (i.e., English$\Rightarrow $French, English$\Rightarrow $Japanese in Figures FIGREF20(b, c)) as well as the reverse directions (Figures FIGREF20(d, e, f)) using Transformer under the Mask perturbation. In all the cases, Attribution shows the best performance while Random achieves the worst result. More specifically, Attribution method shows similar translation quality degradation on all three language-pairs, which declines to around the half of the original BLEU score with five operations.
There exists another straightforward method, Erasure BIBREF7, BIBREF22, BIBREF25, which directly evaluates the word importance by measuring the translation performance degradation of each word. Specifically, it erases (i.e., Mask) one word from the input sentence each time and uses the BLEU score changes to denote the word importance (after normalization).
In Figure FIGREF31, we compare Erasure method with Attribution method under the Mask perturbation. The results show that Attribution method is less effective than Erasure method when only one word is perturbed. But it outperforms the Erasure method when perturbing 2 or more words. The results reveal that the importance calculated by erasing only one word cannot be generalized to multiple-words scenarios very well. Besides, the Erasure method is a supervised method which requires ground-truth references, and finding a better words combination is computation infeasible when erasing multiple words.
We close this section by pointing out that our gradient-based method consistently outperforms its black-box counterparts in various settings, demonstrating the effectiveness and universality of exploiting gradients for estimating word importance. In addition, our approach is on par with or even outperforms the supervised erasure method (on multiple-word perturbations). This is encouraging since our approach does not require any external resource and is fully unsupervised.
We use the standard BLEU score as the evaluation metric for translation performance.
|
What criteria does the author use to evaluate translation performance?
|
They use the standard BLEU score as the evaluation metric for translation performance.
|
null | false
| null |
Who are some of the most successful golfers in history?
|
Sam Snead, Tiger Woods, and Jack Nicklaus are the most successful golfers of all time with the top three most major championship wins.
|
|
null | false
| null |
Wilbur Snyder (September 15, 1929 – December 25, 1991) was an American football player and professional wrestler.
Wrestling career
Snyder's wrestling debut occurred during football's 1953 off-season. He was trained by Sandor Szabo and Warren Bockwinkel in Southern California. In 1954, Snyder retired from football completely to pursue a full-time career in wrestling. Part of his in-ring persona included the use of many football tactics, and it garnered Snyder a lot of national attention. He had already been a regional champion in Montreal, Quebec, Canada when he made a name for himself by defeating Verne Gagne and winning the United States Championship at Marigold Arena in Chicago on April 7, 1956.
|
Was Wilbur Snyder a good wrestler
|
Wilbur Snyder (September 15, 1929 – December 25, 1991) was an American football player and professional wrestler.
Snyder's wrestling debut occurred during football's 1953 off-season. He was trained by Sandor Szabo and Warren Bockwinkel in Southern California. In 1954, Snyder retired from football completely to pursue a full-time career in wrestling. Part of his in-ring persona included the use of many football tactics, and it garnered Snyder a lot of national attention.
He had already been a regional champion in Montreal, Quebec, Canada when he made a name for himself by defeating Verne Gagne and winning the United States Championship at Marigold Arena in Chicago on April 7, 1956.
|
null | false
| null |
Is there a fast way to put a duvet cover over a duvet?
|
Yes! First, lay the duvet out on top of a bed. Next, turn the duvet cover inside out, and place it on top of the duvet, with the bottom facing the top of the bed and duvet. Tie the corners of the duvet cover to the duvet. With a friend, reach inside the duvet cover to grab the furthest corners of the duvet cover with the hand facing the top of the bed, and grab the other corners with the opposite hand. Note: your arms should be crossed. Pull your arms so they are no longer crossed, and stretch them out as far as you can reach. You now have the duvet inside the duvet cover.
|
|
1909.03023
| false
| null |
Our annotation scheme introduces opportunities for the educational community to conduct further research on the relationship between features of student talk, student learning, and discussion quality. Although Chisholm and Godley Chisholm:11 and we found relations between our coding constructs and discussion quality, these were small-scale studies based on manual annotations. Once automated classifiers are developed, such relations between talk and learning can be examined at scale. Also, automatic labeling via a standard coding scheme can support the generalization of findings across studies, and potentially lead to automated tools for teachers and students.
The proposed annotation scheme also introduces NLP opportunities and challenges. Existing systems for classifying specificity and argumentation have largely been designed to analyze written text rather than spoken discussions. This is (at least in part) due to a lack of publicly available corpora and schemes for annotating argumentation and specificity in spoken discussions. The development of an annotation scheme explicitly designed for this problem is the first step towards collecting and annotating corpora that can be used by the NLP community to advance the field in this particular area. Furthermore, in text-based discussions, NLP methods need to tightly couple the discussion with contextual information (i.e., the text under discussion). For example, an argument move from one of the discussions mentioned in Section 4 stated “She's saying like free like, I don't have to be, I don't have to be this salesman's wife anymore, your know? I don't have to play this role anymore." The use of the term salesman shows the presence of specificity element (3) (see Section 3.2) because the text under discussion is indeed Death of a Salesman. If the students were discussing another text, the mention of the term salesman would not indicate one of the specificity elements, therefore lowering the specificity rating. Thus, using existing systems is unlikely to yield good performance. In fact, we previously BIBREF31 showed that while using an off-the-shelf system for predicting specificity in newspaper articles resulted in low performance when applied to classroom discussions, exploiting characteristics of our data could significantly improve performance. We have similarly evaluated the performance of two existing argument mining systems BIBREF18 , BIBREF33 on the transcripts described in Section SECREF4 . We noticed that since the two systems were trained to classify only claims and premises, they were never able to correctly predict warrants in our transcripts. Additionally, both systems classified the overwhelming majority of moves as premise, resulting in negative kappa in some cases. Using our scheme to create a corpus of classroom discussion data manually annotated for argumentation, specificity, and knowledge domain will support the development of more robust NLP prediction systems.
Our annotation scheme introduces opportunities for the educational community to conduct further research on the relationship between features of student talk, student learning, and discussion quality.
Once automated classifiers are developed, such relations between talk and learning can be examined at scale. Also, automatic labeling via a standard coding scheme can support the generalization of findings across studies, and potentially lead to automated tools for teachers and students.
The development of an annotation scheme explicitly designed for this problem is the first step towards collecting and annotating corpora that can be used by the NLP community to advance the field in this particular area.
|
what opportunities are highlighted?
|
The answers are shown as follows:
* Our annotation scheme introduces opportunities for the educational community to conduct further research
* Once automated classifiers are developed, such relations between talk and learning can be examined at scale
* automatic labeling via a standard coding scheme can support the generalization of findings across studies, and potentially lead to automated tools for teachers and students
* collecting and annotating corpora that can be used by the NLP community to advance the field in this particular area
|
null | false
| 14
|
Suppose a user wants to write a sentence “I will be 10 minutes late.” Ideally, she would type just a few keywords such as “10 minutes late” and an autocomplete system would be able to infer the intended sentence (Figure FIGREF1). Existing left-to-right autocomplete systems BIBREF0, BIBREF1 can often be inefficient, as the prefix of a sentence (e.g. “I will be”) fails to capture the core meaning of the sentence. Besides the practical goal of building a better autocomplete system, we are interested in exploring the tradeoffs inherent to such communication schemes between the efficiency of typing keywords, accuracy of reconstruction, and interpretability of keywords.
One approach to learn such schemes is to collect a supervised dataset of keywords-sentence pairs as a training set, but (i) it would be expensive to collect such data from users, and (ii) a static dataset would not capture a real user's natural predilection to adapt to the system BIBREF2. Another approach is to avoid supervision and jointly learn a user-system communication scheme to directly optimize the combination of efficiency and accuracy. However, learning in this way can lead to communication schemes that are uninterpretable to humans BIBREF3, BIBREF4 (see Appendix for additional related work).
In this work, we propose a simple, unsupervised approach to an autocomplete system that is efficient, accurate, and interpretable. For interpretability, we restrict keywords to be subsequences of their source sentences based on the intuition that humans can infer most of the original meaning from a few keywords. We then apply multi-objective optimization approaches to directly control and achieve desirable tradeoffs between efficiency and accuracy.
We observe that naively optimizing a linear combination of efficiency and accuracy terms is unstable and leads to suboptimal schemes. Thus, we propose a new objective which optimizes for communication efficiency under an accuracy constraint. We show this new objective is more stable and efficient than the linear objective at all accuracy levels.
As a proof-of-concept, we build an autocomplete system within this framework which allows a user to write sentences by specifying keywords. We empirically show that our framework produces communication schemes that are 52.16% more accurate than rule-based baselines when specifying 77.37% of sentences, and 11.73% more accurate than a naive, weighted optimization approach when specifying 53.38% of sentences. Finally, we demonstrate that humans can easily adapt to the keyword-based autocomplete system and save nearly 50% of time compared to typing a full sentence in our user study.
We empirically show that our framework produces communication schemes that are 52.16% more accurate than rule-based baselines when specifying 77.37% of sentences, and 11.73% more accurate than a naive, weighted optimization approach when specifying 53.38% of sentences.
|
Is the model tested against any baseline?
|
Yes.
|
null | false
| null |
Which one of the following objects is least like the others, and why?
Lemon, Car, Grass, Lime
|
A car is least like the others, because it is a man-made non-living object, while the rest are plants.
|
|
null | false
| null |
Population numbers have been recorded in Osaka since as early as 1873, in the early Meiji era. According to the census in 2005, there were 2,628,811 residents in Osaka, an increase of 30,037 or 1.2% from 2000. There were 1,280,325 households with approximately 2.1 persons per household. The population density was 11,836 persons per km2. The Great Kantō earthquake caused a mass migration to Osaka between 1920 and 1930, and the city became Japan's largest city in 1930 with 2,453,573 people, outnumbering even Tokyo, which had a population of 2,070,913. The population peaked at 3,252,340 in 1940, and had a post-war peak of 3,156,222 in 1965, but continued to declined since, as the residents moved out to the suburbs.
There were 144,123 registered foreigners, the two largest groups being Korean (60,110) and Chinese (39,551) 2021 years. Ikuno, with its Tsuruhashi district, is the home to one of the largest population of Korean residents in Japan, with 20,397 registered Zainichi Koreans.
|
Given this paragraph about the demographics of Osaka, what was the highest number of residents ever recorded to be living in Osaka at any one point in time?
|
The highest population ever recorded in Osaka was 3,252,340 in 1940.
|
null | false
| null |
Write a polite rejection response to a LinkedIn recruiter
|
Hello,
Thank you for your message. I’m currently happy where I am but hope to reconnect in the future. Best of luck to you on the search.
With gratitude,
|
|
null | false
| 414
|
Algorithm 1 GFCS: Gradient First, Coimage Second Sample w ∼ U (−1, 1) C 13:
q ← d ODS (x adv , s, w) ▷ See Eqn. 1 for definition.
14:
for α ∈ {ϵ, −ϵ} do
x adv ← Π xin,ν (x adv + α • q) 17:
S rem ← S ▷ Reset candidate surrogate set to input set; resume using loss gradients.
18:
break
The entirety of the proposed method is given in pseudocode as Algorithm 1. As indicated in Sec. 2.1, the method takes a victim classifier v, an input image x in , and a norm bound ν. The projection operator Π xin,ν confines its input to the ν-ball centred on x in : its inclusion in the algorithm represents a standard projected gradient ascent (PGA) implementation. Additionally, the method requires a loss function L, a set S of one or more surrogate models, and a step length ϵ representing the fixed length of the perturbations to be attempted at each iterate 1 along its candidate direction. The loss L can be any function of the iterate that serves as a suitable proxy for the adversarial objective, as in Sec. 2.1:
the only requirement here is that it be once differentiable. In our implementation, we make the popular and effective choice of the margin loss L f (x) = f ct (x) − f cs (x) where c s = argmax c v c (x) and c t = argmax c̸ =cs v c (x), i.e. the difference between the highest and second-highest (or, in the targeted case, the target) class scores. Note that the class IDs c t and c s are defined by the ranking according to v, but are evaluated on the net f parametrising the loss, which is either s or v depending on the line of the algorithm (9 and 15, respectively). The natural assumption of a surrogate method is that the surrogate will provide useful information about the victim, but there is no "hard" requirement on the surrogates s ∈ S other than being once-differentiable functions mapping X → Y. The definition of ODS direction d ODS for network f and input x is, as in,
where w is sampled from the uniform distribution over [−1, 1] C . By definition, it is the normalised gradient of a randomly weighted sum of all of the class scores. Equivalently, by linearity, it is a randomly weighted sum of all of the class-score gradients (i.e. rows of the Jacobian matrix), which are themselves a basis of the coimage of the linear approximation of f : the subspace which f exhibits any nonzero response to.
As indicated in Sec. 1, the logic of the method is simple. At any given iterate, the method tries to proceed in a SimBA-like manner by testing the change in adversarial loss at fixed-length steps along candidate directions, projected back into the feasible set where necessary. It does so exclusively using normalised loss gradients from the surrogates in the input set (drawn in random order, without replacement), unless and until it has exhausted them at that iterate without success. As we will demonstrate in Sec. 3.2, this intermediate failure state is seldom reached. If this state is reached, however, the method instead randomly samples a surrogate (with replacement) and an ODS direction from that surrogate, attempting a SimBA update each time, until an improvement in the loss is realised. Once such a successful update occurs, the method resets the candidate surrogate set to the input set and resumes using normalised loss gradients only. The method terminates on finding an adversarial example or on exceeding an upper bound on the query count if one has been specified.). The latter is omitted for LeBA , which does not admit non-singleton surrogate sets. All networks used are pretrained models available via PyTorch/torchvision. Parameter values of competitors are as they specify except where we note otherwise for reasons that will be discussed below. LeBA is run in "train" mode on a held-out set of 1000 images and then evaluated in "test" mode on the same set of 2000 used for all other methods. P-RGF always uses the adaptive coefficient mode. P-RGF and ODS-RGF are based on our own PyTorch port of the reference P-RGF code, which will be released along with this paper: no public implementation of ODS-RGF currently exists otherwise. We include the surrogate-free for comparison.
To delve deeper into the results of Sec. 3.1, we plot each attacked example as a 2D point whose x-coordinate is the number of queries expended by the surrogate loss gradient block of the algorithm, and whose y-coordinate is the analogous count for the coimage block. This gives the scatter plot of Fig., which is supplemented by marginal histograms corresponding to the axes opposite them. The figure shows results obtained using Inception-v3 as the victim: see Appendix A.3 for analogous figures for VGG-16 and ResNet-50. Note that the axes of the main scatter plots are log-log, while those of the marginal histograms are linear-log. Some phenomena are readily evident. For one, there is a large fraction of examples (represented by the dense horizontal rows of dots at the bottom of the plots) that succeed within a very low number of queries (on the order of 1-10), which are entirely or almost entirely due to the surrogate gradient transfer, with ODS used seldom or not at all. As these low-query clusters are extremely dense, the corresponding marginals should be consulted (best under zoom) in order to quantify them. For another, the number of examples outside of this regime falls considerably when the four-surrogate set is used instead of ResNet-152 on its own, as can be seen by comparing the left and right sides of the figure. It is clear that the examples that rely on the interplay between the gradient-and coimage-based direction generators are reduced to a nontrivial (i.e. sufficient to affect the failure rate if not handled) but nonetheless relatively small group. Overall, there is an order-of-magnitude difference between surrogate loss gradient queries and ODS queries in the points extending away from the dense low-query cluster at the bottom, i.e. the examples that rely on both submethods. That is, when the ODS block is required, it typically requires far more queries to progress the optimiser than in the much more common cases in which the gradient suffices.
its inclusion in the algorithm represents a standard projected gradient ascent (PGA) implementation.
|
In the page 9, what is PGA?
|
Projected Gradient Ascent. The abbreviation is defined on page 4. We have used ascent sign conventions throughout the paper, i.e. the attacker seeks to increase the loss. Thus, we refer to PGA rather than PGD.
|
null | false
| null |
The American Revolutionary War (April 19, 1775 – September 3, 1783), also known as the Revolutionary War or American War of Independence, was the military conflict of the American Revolution in which American Patriot forces under George Washington's command defeated the British, establishing and securing the independence of the United States.
|
what was the American Revolutionary War
|
The American Revolutionary War (April 19, 1775 – September 3, 1783), also known as the Revolutionary War or American War of Independence, was the military conflict of the American Revolution in which American Patriot forces under George Washington's command defeated the British, establishing and securing the independence of the United States. Fighting began on April 19, 1775 at the Battles of Lexington and Concord. The war was formalized and intensified following passage of the Lee Resolution, which asserted that the Thirteen Colonies were "free and independent states", and the Declaration of Independence, drafted by the Committee of Five and written primarily by Thomas Jefferson, by the Second Continental Congress in Philadelphia, on July 2, 1776 and July 4, 1776, respectively.
|
null | false
| null |
Do Ceanothus have shallow roots?
|
Ceanothus plants are super variable, but in general, they tend to have a deep taproot. Sometimes this is not the case if water sources are more reliable near the soil surface.
|
|
null | false
| null |
Write a short blog post about why you have chosen to delete your instagram account and why you think that instagram has an negative impact on society.
|
Hi friends, long time no talk! Today, I want to tell you all about why I decided to delete my instagram account. This is a decision I had been thinking about for a long time, and ultimately I decided it just made sense to get rid of the account. However, I didn't want to do this without sharing with you all why I made this decision, because I think it is important to explain. I've had an instagram account for years, but often felt as though I would get sucked into a scroll spiral (we've all been there, let's be real) where I would just sit on my phone for hours. I had already put in place screen time limits, but I felt like I wasn't really getting a lot out of the app. However, what ultimately made me decide to go full send and delete my account was reading about how social media causes low self esteem, and how it could be causing many of the mental health issues we are seeing with young people right now. After reading about this, I decided enough was enough, and I finally deleted my account. I would encourage you all to do the same - don't worry, I'm not giving up this blog, so you'll still be able to check in on what I'm up to. Let me know in the comments what you guys think, and talk next week!
|
|
null | false
| null |
In the late 18th century, French General Jean-Baptiste Vaquette de Gribeauval promoted standardized weapons in what became known as the Système Gribeauval after it was issued as a royal order in 1765. (Its focus at the time was artillery more than muskets or handguns.) One of the accomplishments of the system was that solid cast cannons were bored to precise tolerances, which allowed the walls to be thinner than cannons poured with hollow cores. However, because cores were often off center, the wall thickness determined the size of the bore. Standardized boring allowed cannons to be shorter without sacrificing accuracy and range because of the tighter fit of the shells. It also allowed standardization of the shells.
|
Extract names of all weapons mentioned in the paragraph below:
|
The following are names of weapons in the paragraph above: artillery, muskets, handguns, cannons, shells
|
null | false
| 151
|
The Humor Analysis based on Human Annotation (HAHA) 2019 BIBREF1 competition asked for analysis of two tasks in the Spanish language based on a corpus of publicly collected data described in Castro et al. BIBREF2 :
The HAHA dataset includes labeled data for 24,000 tweets and a test set of 6,000 tweets (80%/20% train/test split.) Each record includes the raw tweet text (including accents and emoticons), a binary humor label, the number of votes for each of five star ratings and a “Funniness Score” that is the average of the 1 to 5 star votes cast. Examples and data can be found on the CodaLab competition webpage.
The HAHA dataset includes labeled data for 24,000 tweets and a test set of 6,000 tweets (80%/20% train/test split.)
|
What data does the HAHA dataset include?
|
The HAHA dataset includes labeled data for 24,000 tweets and a test set of 6,000 tweets (80%/20% train/test split.)
|
1702.01517
| false
| null |
We show the final results for opinion recommendation, comparing our proposed model with the following state-of-the-art baseline systems:
RS-Average is the widely-adopted baseline (e.g., by Yelp.com), using the averaged review scores as the final score.
RS-Linear estimates the rating score that a user would give by INLINEFORM0 BIBREF49 , where INLINEFORM1 and INLINEFORM2 are the the training deviations of the user INLINEFORM3 and the product INLINEFORM4 , respectively.
RS-Item applies INLINEFORM0 NN to estimate the rating score BIBREF50 . We choose the cosine similarity between INLINEFORM1 to measure the distance between product.
RS-MF is a state-of-the-art recommendation model, which uses matrix factorisation to predict rating score BIBREF8 , BIBREF41 , BIBREF25 .
Sum-Opinosis uses a graph-based framework to generate abstractive summarisation given redundant opinions BIBREF51 .
Sum-LSTM-Att is a state-of-the-art neural abstractive summariser, which uses an attentional neural model to consolidate information from multiple text sources, generating summaries using LSTM decoding BIBREF44 , BIBREF3 .
All the baseline models are single-task models, without considering rating and summarisation prediction jointly. The results are shown in Table TABREF46 . Our model (“ Joint”) significantly outperforms both “RS-Average” and “RS-Linear” ( INLINEFORM0 using INLINEFORM1 -test), which demonstrates the strength of opinion recommendation, which leverages user characteristics for calculating a rating score for the user.
We show the final results for opinion recommendation, comparing our proposed model with the following state-of-the-art baseline systems:
RS-Average is the widely-adopted baseline (e.g., by Yelp.com), using the averaged review scores as the final score.
RS-Linear estimates the rating score that a user would give by INLINEFORM0 BIBREF49 , where INLINEFORM1 and INLINEFORM2 are the the training deviations of the user INLINEFORM3 and the product INLINEFORM4 , respectively.
RS-Item applies INLINEFORM0 NN to estimate the rating score BIBREF50 . We choose the cosine similarity between INLINEFORM1 to measure the distance between product.
RS-MF is a state-of-the-art recommendation model, which uses matrix factorisation to predict rating score BIBREF8 , BIBREF41 , BIBREF25 .
Sum-Opinosis uses a graph-based framework to generate abstractive summarisation given redundant opinions BIBREF51 .
Sum-LSTM-Att is a state-of-the-art neural abstractive summariser, which uses an attentional neural model to consolidate information from multiple text sources, generating summaries using LSTM decoding BIBREF44 , BIBREF3 .
All the baseline models are single-task models, without considering rating and summarisation prediction jointly. The results are shown in Table TABREF46 .
|
What are the baselines?
|
The answers are shown as follows:
* RS-Average
* RS-Linear
* RS-Item
* RS-MF
* Sum-Opinosis
* Sum-LSTM-Att
|
null | false
| null |
Classify each of the following as either a drink or a food: apple juice, cake, hamburger, carrot, orange juice, soda.
|
apple juice is a drink
cake is a food
hamburger is a food
carrot is a food
orange juice is a drink
soda is a drink
|
|
2002.05058
| false
| null |
The comparative evaluator is trained with maximum likelihood estimation (MLE) objective, as described in eq DISPLAY_FORM6
where $\mathcal {X}$ is the set of pairwise training examples contructed as described above, $Q(x_1, x_2) \in \lbrace >,<,\approx \rbrace $ is the true label for the pair ($x_1$, $x_2$), $D_\phi ^q(x_1, x_2)$ is the probability of the comparative discriminator's prediction being $q$ ($q \in \lbrace >,<,\approx \rbrace $) for the pair ($x_1$, $x_2$).
The comparative evaluator is trained with maximum likelihood estimation (MLE) objective, as described in eq DISPLAY_FORM6
where $\mathcal {X}$ is the set of pairwise training examples contructed as described above, $Q(x_1, x_2) \in \lbrace >,<,\approx \rbrace $ is the true label for the pair ($x_1$, $x_2$), $D_\phi ^q(x_1, x_2)$ is the probability of the comparative discriminator's prediction being $q$ ($q \in \lbrace >,<,\approx \rbrace $) for the pair ($x_1$, $x_2$).
|
How they add human prefference annotation to fine-tuning process?
|
The answers are shown as follows:
* human preference annotation is available
* $Q(x_1, x_2) \in \lbrace >,<,\approx \rbrace $ is the true label for the pair
|
null | false
| null |
Give me a list of the best movies I've never seen.
|
Top 10 movies I've ever seen, in no particular order. The Godfather, Dead Poet's Society, Rear Window, Scarface, Inception, Top Gun, Jaws, Skyfall, The Shawshank Redemption, A Few Good Men
|
|
null | false
| null |
What are the current f1 teams?
|
The current 10 Formula 1 teams are Red Bull Racing, Aston Martin Aramco, Mercedes-AMG PETRONAS, Scuderia Ferrari, McLaren, BWT Alpine, MoneyGram Haas, Alfa Romeo, Scuderia AlphaTauri, Williams
|
|
null | false
| null |
Milka Ivić (Serbian Cyrillic: Милка Ивић, 11 December 1923 – 7 March 2011) was a Serbian linguist.
|
Who is Milka Ivić?
|
Milka Ivić (Serbian Cyrillic: Милка Ивић, 11 December 1923 – 7 March 2011) was a Serbian linguist.
She was born in Belgrade. She took her doctorate in 1954 with the thesis Značenja srpskohrvatskoga instrumentala i njihov razvoj (The Meanings of Serbo-Croatian Instrumental and Their Development), and became a professor of Serbian and Croatian language at the University of Novi Sad. She was especially known for her book Pravci u lingvistici (Trends [Movements] in Linguistics). She was a member of the Norwegian Academy of Science and Letters from 1976, a corresponding member of the Slovenian Academy of Sciences and Arts from 1983,and a member of the Serbian Academy of Sciences and Arts.
|
1901.02534
| false
| null |
The release of the FEVER fact extraction and verification dataset BIBREF0 provides a large-scale challenge that tests a combination of retrieval and textual entailment capabilities. To verify a claim in the dataset as supported, refuted, or undecided, a system must retrieve relevant articles and sentences from Wikipedia. Then it must decide whether each of those sentences, or some combination of them, entails or refutes the claim, which is an entailment problem. Systems are evaluated on the accuracy of the claim predictions, with credit only given when correct evidence is submitted.
As entailment data, premises in FEVER data differ substantially from those in the image caption data used as the basis for the Stanford Natural Language Inference (SNLI) BIBREF1 dataset. Sentences are longer (31 compared to 14 words on average), vocabulary is more abstract, and the prevalence of named entities and out-of-vocabulary terms is higher.
The retrieval aspect of FEVER is not straightforward either. A claim may have small word overlap with the relevant evidence, especially if the claim is refuted by the evidence.
The release of the FEVER fact extraction and verification dataset BIBREF0 provides a large-scale challenge that tests a combination of retrieval and textual entailment capabilities. To verify a claim in the dataset as supported, refuted, or undecided, a system must retrieve relevant articles and sentences from Wikipedia. Then it must decide whether each of those sentences, or some combination of them, entails or refutes the claim, which is an entailment problem. Systems are evaluated on the accuracy of the claim predictions, with credit only given when correct evidence is submitted.
As entailment data, premises in FEVER data differ substantially from those in the image caption data used as the basis for the Stanford Natural Language Inference (SNLI) BIBREF1 dataset. Sentences are longer (31 compared to 14 words on average), vocabulary is more abstract, and the prevalence of named entities and out-of-vocabulary terms is higher.
The retrieval aspect of FEVER is not straightforward either. A claim may have small word overlap with the relevant evidence, especially if the claim is refuted by the evidence.
|
What is the FEVER task?
|
The answers are shown as follows:
* tests a combination of retrieval and textual entailment capabilities. To verify a claim in the dataset as supported, refuted, or undecided, a system must retrieve relevant articles and sentences from Wikipedia. Then it must decide whether each of those sentences, or some combination of them, entails or refutes the claim, which is an entailment problem
|
null | false
| 173
|
Encoding of words is perhaps the most important step towards a successful end-to-end natural language processing application. Although word embeddings have been shown to provide benefit to such models, they commonly treat words as the smallest meaning bearing unit and assume that each word type has its own vector representation. This assumption has two major shortcomings especially for languages with rich morphology: (1) inability to handle unseen or out-of-vocabulary (OOV) word-forms (2) inability to exploit the regularities among word parts. The limitations of word embeddings are particularly pronounced in sentence-level semantic tasks, especially in languages where word parts play a crucial role. Consider the Turkish sentences “Köy+lü-ler (villagers) şehr+e (to town) geldi (came)” and “Sendika+lı-lar (union members) meclis+e (to council) geldi (came)”. Here the stems köy (village) and sendika (union) function similarly in semantic terms with respect to the verb come (as the origin of the agents of the verb), where şehir (town) and meclis (council) both function as the end point. These semantic similarities are determined by the common word parts shown in bold. However ortographic similarity does not always correspond to semantic similarity. For instance the ortographically similar words knight and night have large semantic differences. Therefore, for a successful semantic application, the model should be able to capture both the regularities, i.e, morphological tags and the irregularities, i.e, lemmas of the word.
Morphological analysis already provides the aforementioned information about the words. However access to useful morphological features may be problematic due to software licensing issues, lack of robust morphological analyzers and high ambiguity among analyses. Character-level models (CLM), being a cheaper and accessible alternative to morphology, have been reported as performing competitively on various NLP tasks BIBREF0 , BIBREF1 , BIBREF2 . However the extent to which these tasks depend on morphology is small; and their relation to semantics is weak. Hence, little is known on their true ability to reveal the underlying morphological structure of a word and their semantic capabilities. Furthermore, their behaviour across languages from different families; and their limitations and strengths such as handling of long-range dependencies, reaction to model complexity or performance on out-of-domain data are unknown. Analyzing such issues is a key to fully understanding the character-level models.
To achieve this, we perform a case study on semantic role labeling (SRL), a sentence-level semantic analysis task that aims to identify predicate-argument structures and assign meaningful labels to them as follows:
$[$ Villagers $]$ comers came $[$ to town $]$ end point
We use a simple method based on bidirectional LSTMs to train three types of base semantic role labelers that employ (1) words (2) characters and character sequences and (3) gold morphological analysis. The gold morphology serves as the upper bound for us to compare and analyze the performances of character-level models on languages of varying morphological typologies. We carry out an exhaustive error analysis for each language type and analyze the strengths and limitations of character-level models compared to morphology. In regard to the diversity hypothesis which states that diversity of systems in ensembles lead to further improvement, we combine character and morphology-level models and measure the performance of the ensemble to better understand how similar they are.
We experiment with several languages with varying degrees of morphological richness and typology: Turkish, Finnish, Czech, German, Spanish, Catalan and English. Our experiments and analysis reveal insights such as:
We carry out an exhaustive error analysis for each language type and analyze the strengths and limitations of character-level models compared to morphology.
|
Do the authors analyze the strengths and limitations of character-level models compared to morphology?
|
Yes, they do.
|
null | false
| null |
Make me a menu for a dinner party which will use up potatoes, cherry tomatoes, salmon and heavy cream I have that is about to expire. The menu should have three courses and a dessert, and shouldn't include eggs due to the current egg shortage. Describe the order in which each food would be served.
|
The first course can be appetizers, for which you can make bruschetta to use up the cherry tomatoes. Next for the second course serve a potato and leek soup, which will use up the potatoes and heavy cream. For the main course you will serve salmon with a dill sauce that is also made with heavy cream. For dessert you can serve a sorbet, which is egg free. This is an egg free meal that will use up all your ingredients .
|
|
null | false
| null |
Give me a list of 10 airlines from around the world.
|
Delta, British Airways, Swiss Air, Singapore Airlines, Air India, Cathay Pacific, Virgin Atlantic, Easyjet, Southwest, Air France
|
|
null | false
| 349
|
Generating descriptions for multimedia contents such as images and videos, so called Image Captioning, is helpful for e-commerce companies or news agencies. For instance, in e-commerce field, people will no longer need to put much effort into understanding and describing products' images on their websites because image contents can be recognized and descriptions are automatically generated. Inspired by Horus BIBREF0 , Image Captioning system can also be integrated into a wearable device, which is able to capture surrounding images and generate descriptions as sound in real time to guide people with visually impaired.
Image Captioning has attracted attentions from researchers in recent years BIBREF1, BIBREF2, BIBREF3, and there has been promising attempts dealing with language barrier in this task by extending existed dataset captions into different languages BIBREF3, BIBREF4.
In this study, generating image captions in Vietnamese language is put into consideration. One straightforward approach for this task is to translate English captions into Vietnamese by human or by using machine translation tool, Google translation. With the method of translating directly from English to Vietnamese, we found that the descriptions are sometimes confusing and unnatural to native people. Moreover, image understandings are cultural dependent, as in Western, people usually have different ways to grasp images and different vocabulary choices for describing contexts. For instance, in Fig. FIGREF2, one MS-COCO English caption introduce about "a baseball player in motion of pitching", which makes sense and capture accurately the main activity in the image. Though it sounds sensible in English, the sentence becomes less meaningful when we try to translate it into Vietnamese. One attempt of translating the sentence is performed by Google Translation, and the result is not as expected.
Therefore, we come up with the approach of constructing a Vietnamese Image Captioning dataset with descriptions written manually by human. Composed by Vietnamese people, the sentences would be more natural and friendlier to Vietnamese users. The main resources we used from MS-COCO for our dataset are images. Besides, we consider having our dataset focus on sportball category due to several reasons:
By concentrating on a specific domain we are more likely to improve performance of the Image Captioning models. We expect our dataset can be used to confirm or reject this hypothesis.
Sportball Image Captioning can be used in certain sport applications, such as supportting journalists describing great amount of images for their articles.
Our primary contributions of this paper are as follows:
Firstly, we introduce UIT-ViIC, the first Vietnamese dataset extending MS-COCO with manually written captions for Image Captioning. UIT-ViIC is published for research purposes.
Secondly, we introduce our annotation tool for dataset construction, which is also published to help annotators conveniently create captions.
Finally, we conduct experiments to evaluate state-of-the-art models (evaluated on English dataset) on UIT-ViIC dataset, then we analyze the performance results to have insights into our corpus.
The structure of the paper is organized as follows. Related documents and studies are presented in Section SECREF2. UIT-ViIC dataset creation is described in Section SECREF3. Section SECREF4 describes the methods we implement. The experimental results and analysis are presented in Section SECREF5. Conclusion and future work are deduced in Section SECREF6.
Finally, we conduct experiments to evaluate state-of-the-art models (evaluated on English dataset) on UIT-ViIC dataset, then we analyze the performance results to have insights into our corpus.
|
What language of dataset do they use when evaluating state-of-the-art models?
|
English dataset.
|
null | false
| 99
|
In this work we formulate our classification problem as follows: given two classes of news articles, respectively $D$ (disinformation) and $M$ (mainstream), a set of news articles $A_i$ and associated class labels $C_i \in \lbrace D,M\rbrace $, and a set of tweets $\Pi _i=\lbrace T_i^1, T_i^2, ...\rbrace $ each of which contains an Uniform Resource Locator (URL) pointing explicitly to article $A_i$, predict the class $C_i$ of each article $A_i$. There is huge debate and controversy on a proper taxonomy of malicious and deceptive information BIBREF1BIBREF2BIBREF15BIBREF16BIBREF17BIBREF3BIBREF11. In this work we prefer the term disinformation to the more specific fake news to refer to a variety of misleading and harmful information. Therefore, we follow a source-based approach, a consolidated strategy also adopted by BIBREF6BIBREF16BIBREF2BIBREF1, in order to obtain relevant data for our analysis. We collected:
Disinformation articles, published by websites which are well-known for producing low-credibility content, false and misleading news reports as well as extreme propaganda and hoaxes and flagged as such by reputable journalists and fact-checkers;
Mainstream news, referring to traditional news outlets which deliver factual and credible information.
We believe that this is currently the most reliable classification approach, but it entails obvious limitations, as disinformation outlets may also publish true stories and likewise misinformation is sometimes reported on mainstream media. Also, given the choice of news sources, we cannot test whether our methodology is able to classify disinformation vs factual but not mainstream news which are published on niche, non-disinformation outlets.
We collected tweets associated to a dozen US mainstream news websites, i.e. most trusted sources described in BIBREF18, with the Streaming API, and we referred to Hoaxy API BIBREF16 for what concerns tweets containing links to 100+ US disinformation outlets. We filtered out articles associated to less than 50 tweets. The resulting dataset contains overall $\sim $1.7 million tweets for mainstream news, collected in a period of three weeks (February 25th, 2019-March 18th, 2019), which are associated to 6,978 news articles, and $\sim $1.6 million tweets for disinformation, collected in a period of three months (January 1st, 2019-March 18th, 2019) for sake of balance of the two classes, which hold 5,775 distinct articles. Diffusion censoring effects BIBREF14 were correctly taken into account in both collection procedures. We provide in Figure FIGREF4 the distribution of articles by source and political bias for both news domains.
As it is reported that conservatives and liberals exhibit different behaviors on online social platforms BIBREF19BIBREF20BIBREF21, we further assigned a political bias label to different US outlets (and therefore news articles) following the procedure described in BIBREF2. In order to assess the robustness of our method, we performed classification experiments by training only on left-biased (or right-biased) outlets of both disinformation and mainstream domains and testing on the entire set of sources, as well as excluding particular sources that outweigh the others in terms of samples to avoid over-fitting.
For what concerns the Italian scenario we first collected tweets with the Streaming API in a 3-week period (April 19th, 2019-May 5th, 2019), filtering those containing URLs pointing to Italian official newspapers websites as described in BIBREF22; these correspond to the list provided by the association for the verification of newspaper circulation in Italy (Accertamenti Diffusione Stampa). We instead referred to the dataset provided by BIBREF23 to obtain a set of tweets, collected continuously since January 2019 using the same Twitter endpoint, which contain URLs to 60+ Italian disinformation websites. In order to get balanced classes (April 5th, 2019-May 5th, 2019), we retained data collected in a longer period w.r.t to mainstream news. In both cases we filtered out articles with less than 50 tweets; overall this dataset contains $\sim $160k mainstream tweets, corresponding to 227 news articles, and $\sim $100k disinformation tweets, corresponding to 237 news articles. We provide in Figure FIGREF5 the distribution of articles according to distinct sources for both news domains. As in the US dataset, we took into account censoring effects BIBREF14 by excluding tweets published before (left-censoring) or after two weeks (right-censoring) from the beginning of the collection process.
The different volumes of news shared on Twitter in the two countries are due both to the different population size of US and Italy (320 vs 60 millions) but also to the different usage of Twitter platform (and social media in general) for news consumption BIBREF24. Both datasets analyzed in this work are available from the authors on request.
A crucial aspect in our approach is the capability to fully capturing sharing cascades on Twitter associated to news articles. It has been reported BIBREF25 that the Twitter streaming endpoint filters out tweets matching a given query if they exceed 1% of the global daily volume of shared tweets, which nowadays is approximately $5\cdot 10^8$; however, as we always collected less than $10^6$ tweets per day, we did not incur in this issue and we thus gathered 100% of tweets matching our query.
We built Twitter diffusion networks following an approach widely adopted in the literature BIBREF6BIBREF17BIBREF2. We remark that there is an unavoidable limitation in Twitter Streaming API, which does not allow to retrieve true re-tweeting cascades because re-tweets always point to the original source and not to intermediate re-tweeting users BIBREF8BIBREF14; thus we adopt the only viable approach based on Twitter's public availability of data. Besides, by disentangling different interactions with multiple layers we potentially reduce the impact of this limitation on the global network properties compared to the single-layer approach used in our baseline.
Using the notation described in BIBREF26. we employ a multi-layer representation for Twitter diffusion networks. Sociologists have indeed recognized decades ago that it is crucial to study social systems by constructing multiple social networks where different types of ties among same individuals are used BIBREF27. Therefore, for each news article we built a multi-layer diffusion network composed of four different layers, one for each type of social interaction on Twitter platform, namely retweet (RT), reply (R), quote (Q) and mention (M), as shown in Figure FIGREF11. These networks are not necessarily node-aligned, i.e. users might be missing in some layers. We do not insert "dummy" nodes to represent all users as it would have severe impact on the global network properties (e.g. number of weakly connected components). Alternatively one may look at each multi-layer diffusion network as an ensemble of individual graphs BIBREF26; since global network properties are computed separately for each layer, they are not affected by the presence of any inter-layer edges.
In our multi-layer representation, each layer is a directed graph where we add edges and nodes for each tweet of the layer type, e.g. for the RT layer: whenever user $a$ retweets account $b$ we first add nodes $a$ and $b$ if not already present in the RT layer, then we build an edge that goes from $b$ to $a$ if it does not exists or we increment the weight by 1. Similarly for the other layers: for the R layer edges go from user $a$ (who replies) to user $b$, for the Q layer edges go from user $b$ (who is quoted by) to user $a$ and for the M layer edges go from user $a$ (who mentions) to user $b$.
Note that, by construction, our layers do not include isolated nodes; they correspond to "pure tweets", i.e. tweets which have not originated any interactions with other users. However, they are present in our dataset, and their number is exploited for classification, as described below.
We used a set of global network indicators which allow us to encode each network layer by a tuple of features. Then we simply concatenated tuples as to represent each multi-layer network with a single feature vector. We used the following global network properties:
Number of Strongly Connected Components (SCC): a Strongly Connected Component of a directed graph is a maximal (sub)graph where for each pair of vertices $u,v$ there is a path in each direction ($u\rightarrow v$, $v\rightarrow u$).
Size of the Largest Strongly Connected Component (LSCC): the number of nodes in the largest strongly connected component of a given graph.
Number of Weakly Connected Components (WCC): a Weakly Connected Component of a directed graph is a maximal (sub)graph where for each pair of vertices $(u, v)$ there is a path $u \leftrightarrow v$ ignoring edge directions.
Size of the Largest Weakly Connected Component (LWCC): the number of nodes in the largest weakly connected component of a given graph.
Diameter of the Largest Weakly Connected Component (DWCC): the largest distance (length of the shortest path) between two nodes in the (undirected version of) largest weakly connected component of a graph.
Average Clustering Coefficient (CC): the average of the local clustering coefficients of all nodes in a graph; the local clustering coefficient of a node quantifies how close its neighbours are to being a complete graph (or a clique). It is computed according to BIBREF28.
Main K-core Number (KC): a K-core BIBREF13 of a graph is a maximal sub-graph that contains nodes of internal degree $k$ or more; the main K-core number is the highest value of $k$ (in directed graphs the total degree is considered).
Density (d): the density for directed graphs is $d=\frac{|E|}{|V||V-1|}$, where $|E|$ is the number of edges and $|N|$ is the number of vertices in the graph; the density equals 0 for a graph without edges and 1 for a complete graph.
Structural virality of the largest weakly connected component (SV): this measure is defined in BIBREF14 as the average distance between all pairs of nodes in a cascade tree or, equivalently, as the average depth of nodes, averaged over all nodes in turn acting as a root; for $|V| > 1$ vertices, $SV=\frac{1}{|V||V-1|}\sum _i\sum _j d_{ij}$ where $d_{ij}$ denotes the length of the shortest path between nodes $i$ and $j$. This is equivalent to compute the Wiener's index BIBREF29 of the graph and multiply it by a factor $\frac{1}{|V||V-1|}$. In our case we computed it for the undirected equivalent graph of the largest weakly connected component, setting it to 0 whenever $V=1$.
We used networkx Python package BIBREF30 to compute all features. Whenever a layer is empty. we simply set to 0 all its features. In addition to computing the above nine features for each layer, we added two indicators for encoding information about pure tweets, namely the number T of pure tweets (containing URLs to a given news article) and the number U of unique users authoring those tweets. Therefore, a single diffusion network is represented by a vector with $9\cdot 4+2=38$ entries.
Aforementioned network properties can be qualitatively explained in terms of social footprints as follows: SCC correlates with the size of the diffusion network, as the propagation of news occurs in a broadcast manner most of the time, i.e. re-tweets dominate on other interactions, while LSCC allows to distinguish cases where such mono-directionality is somehow broken. WCC equals (approximately) the number of distinct diffusion cascades pertaining to each news article, with exceptions corresponding to those cases where some cascades merge together via Twitter interactions such as mentions, quotes and replies, and accordingly LWCC and DWCC equals the size and the depth of the largest cascade. CC corresponds to the level of connectedness of neighboring users in a given diffusion network whereas KC identifies the set of most influential users in a network and describes the efficiency of information spreading BIBREF17. Finally, d describes the proportions of potential connections between users which are actually activated and SV indicates whether a news item has gained popularity with a single and large broadcast or in a more viral fashion through multiple generations.
For what concerns different Twitter actions, users primarily interact with each other using retweets and mentions BIBREF20.
The former are the main engagement activity and act as a form of endorsement, allowing users to rebroadcast content generated by other users BIBREF31. Besides, when node B retweets node A we have an implicit confirmation that information from A appeared in B's Twitter feed BIBREF12. Quotes are simply a special case of retweets with comments.
Mentions usually include personal conversations as they allow someone to address a specific user or to refer to an individual in the third person; in the first case they are located at the beginning of a tweet and they are known as replies, otherwise they are put in the body of a tweet BIBREF20. The network of mentions is usually seen as a stronger version of interactions between Twitter users, compared to the traditional graph of follower/following relationships BIBREF32.
In order to assess the robustness of our method, we performed classification experiments by training only on left-biased (or right-biased) outlets of both disinformation and mainstream domains and testing on the entire set of sources, as well as excluding particular sources that outweigh the others in terms of samples to avoid over-fitting.
|
To assess the robustness of the method, how do the authors perform classification experiments?
|
The authors performed classification experiments by training only on left-biased (or right-biased) outlets of both disinformation and mainstream domains and testing on the entire set of sources, as well as excluding particular sources that outweigh the others in terms of samples to avoid over-fitting.
|
null | false
| null |
What are some fun ways to keep a toddler busy?
|
1. If you have a spare box and some colored plastic balls, one idea is to cut holes in the box and color a ring around each hole to match a color of ball. Mount the box and show your little one how to drop each ball into its matching hole.
2. If you have a spare toothbrush and a yellow dry erase marker, you can work teach some teeth brushing skills in a fun way! Draw a big toothy smile and color in the lips and gums leaving the teeth bright white. Cover it up with clear tape or hard plastic wrap and cut it out. Color the teeth with the yellow marker and have your little one brush it clean.
3. Teach some valuable life skills and keep your little one entertained by getting a kid-safe knife and having him or her cut fruits.
|
|
null | false
| null |
Sex and the City is an American romantic comedy-drama television series created by Darren Star for HBO. An adaptation of Candace Bushnell's newspaper column and 1996 book anthology of the same name, the series premiered in the United States on June 6, 1998, and concluded on February 22, 2004, with 94 episodes broadcast over six seasons. Throughout its development, the series received contributions from various producers, screenwriters, and directors, principally Michael Patrick King.
|
What is "Sex and the City"?
|
Sex and the City is an American romantic comedy-drama television series created by Darren Star for HBO. An adaptation of Candace Bushnell's newspaper column and 1996 book anthology of the same name, the series premiered in the United States on June 6, 1998, and concluded on February 22, 2004, with 94 episodes broadcast over six seasons. Throughout its development, the series received contributions from various producers, screenwriters, and directors, principally Michael Patrick King.
Sex and the City has received both acclaim and criticism for its subjects and characters, and is credited with helping to increase HBO's popularity as a network. The series has won several accolades, including seven of its 54 Emmy Award nominations, eight of its 24 Golden Globe Award nominations, and three of its 11 Screen Actors Guild Award nominations. The series placed fifth on Entertainment Weekly's "New TV Classics" list, and has been listed as one of the best television series of all time by Time in 2007 and TV Guide in 2013. The series still airs in syndication worldwide. It spawned two feature films, Sex and the City (2008) and Sex and the City 2 (2010), and a prequel television series commissioned by The CW, The Carrie Diaries (2013–14).
A sequel series titled And Just Like That... premiered on HBO Max on December 9, 2021. The series features Parker, Davis and Nixon reprising their roles, with Cattrall choosing not to return partly due to her long-standing and ongoing feud with Parker.
|
null | false
| 207
|
Word Sense Disambiguation (WSD) is an important problem in Natural Language Processing (NLP), both in its own right and as a stepping stone to other advanced tasks in the NLP pipeline, applications such as machine translation BIBREF0 and question answering BIBREF1 . WSD specifically deals with identifying the correct sense of a word, among a set of given candidate senses for that word, when it is presented in a brief narrative (surrounding text) which is generally referred to as context. Consider the ambiguous word `cold'. In the sentence “He started to give me a cold shoulder after that experiment”, the possible senses for cold can be cold temperature (S1), a cold sensation (S2), common cold (S3), or a negative emotional reaction (S4). Therefore, the ambiguous word cold is specified along with the sense set {S1, S2, S3, S4} and our goal is to identify the correct sense S4 (as the closest meaning) for this specific occurrence of cold after considering - the semantic and the syntactic information of - its context.
In this effort, we develop our supervised WSD model that leverages a Bidirectional Long Short-Term Memory (BLSTM) network. This network works with neural sense vectors (i.e. sense embeddings), which are learned during model training, and employs neural word vectors (i.e. word embeddings), which are learned through an unsupervised deep learning approach called GloVe (Global Vectors for word representation) BIBREF2 for the context words. By evaluating our one-model-fits-all WSD network over the public gold standard dataset of SensEval-3 BIBREF3 , we demonstrate that the accuracy of our model in terms of F-measure is comparable with the state-of-the-art WSD algorithms'.
We outline the organization of the rest of the paper as follows. In Section 2, we briefly explore earlier efforts in WSD and discuss recent approaches that incorporate deep neural networks and word embeddings. Our main model that employs BLSTM with the sense and word embeddings is detailed in Section 3. We then present our experiments and results in Section 4 supported by a discussion on how to avoid some drawbacks of the current model in order to achieve higher accuracies and demand less number of training data which is desirable. Finally, in Section 5, we conclude with some future research directions for the construction of sense embeddings as well as applications of such model in other domains such as biomedicine.
By evaluating our one-model-fits-all WSD network over the public gold standard dataset of SensEval-3 , we demonstrate that the accuracy of our model in terms of F-measure is comparable with the state-of-the-art WSD algorithms'.
|
What dataset do they use?
|
The public gold standard dataset of SensEval-3.
|
null | false
| null |
Associazione Calcio Milan (Italian pronunciation: [assotʃatˈtsjoːne ˈkaltʃo ˈmiːlan]), commonly referred to as AC Milan (Italian pronunciation: [a ˈtʃi ˈmiːlan]) or simply Milan (Italian pronunciation: [miːlan]), is a professional football club in Milan, Italy, founded in 1899. The club has spent its entire history, with the exception of the 1980–81 and 1982–83 seasons, in the top flight of Italian football, known as Serie A since 1929–30.
AC Milan's 18 FIFA and UEFA trophies is the fourth highest out of any club (joint with Boca Juniors[nb 1]), and the most out of any Italian club. Milan has won a joint record three Intercontinental Cups and one FIFA Club World Cup, seven European Cup/Champions League titles (Italian record), the UEFA Super Cup a joint record five times and the Cup Winners' Cup twice. With 19 league titles, Milan is tied as the second most successful club in Serie A with local rivals Inter Milan (also with 19 league titles), behind Juventus (36 league titles). They have also won the Coppa Italia five times, and the Supercoppa Italiana seven.
Milan's home games are played at San Siro, also known as the Stadio Giuseppe Meazza. The stadium, which was built by Milan's second chairman Piero Pirelli in 1926 and is shared since 1947 with city rivals Internazionale, is the largest in Italian football, with a total capacity of 75,923. They have a long-standing rivalry with Inter, with whom they contest the Derby della Madonnina which is one of the most followed derbies in football.
The club is one of the wealthiest in Italian and world football. It was a founding member of the now-defunct G-14 group of Europe's leading football clubs as well as its replacement, the European Club Association.
|
Given the reference text about AC Milan, how many FIFA and UEFA trophies have they won?
|
AC Milan has won 18 FIFA and UEFA trophies, which is the fourth most out of any club.
|
null | false
| null |
What are three common alarm clock sounds?
|
People might use a variety of sounds to wake up in the morning. Three common alarm clock sounds are :
1. Bird songs
2. Beeping
3. Radio
|
|
null | false
| 20
|
This research addresses the problem of representing the semantics of text documents in multi-lingual comparable corpora. We present a new approach to this problem, based on neural embeddings, and test it on the task of clustering texts into meaningful classes depending on their topics. The setting is unsupervised, meaning that one either does not have enough annotated data to train a supervised classifier or does not want to be limited with a pre-defined set of classes. There is a lot of sufficiently good approaches to this problem in the case of mono-lingual text collections, but the presence of multiple languages introduces complications.
When a text collection contains documents in several languages, it becomes impractical to simply represent the documents as vectors of words occurring in them ("bag-of-words"), as the words surface forms are different, even in closely-related languages. Thus, one has to invent means to cross the inter-lingual gap and bring all documents to some sort of shared representation, without losing information about their topics or categories.
Of course, one obvious way to solve this problem is to translate all documents into one language, and then apply any clustering algorithm. However, this requires either buying human/machine translation services (which can be expensive if you deal with large text collection) or training own statistical machine translation model (which as a rule requires big parallel corpus). This is the reason to search for other solutions.
In this paper, a novel way of reducing the problem of cross-lingual document representation to a monolingual setting is proposed. Essentially, we train Continuous Bag-of-Words models BIBREF0 on large comparable monolingual corpora for two languages our dataset consists of. This provides us with vector representations of words, allowing to measure their semantic similarity. Then, a linear transformation matrix from vectors of language A to vectors of language B is learned, using a small bilingual dictionary as training data. This matrix is then employed to `project' word and document representations from semantic space of language A to semantic space of language B. It allows not only quite accurate `translation' of words, but also of document `semantic fingerprints' (dense representations of document semantics, calculated as an average of the trained distributional vectors for all the words in document).
This approach is evaluated in a setting, where the input is a collection of documents in several languages and some number of topics to which these documents belong (we also have large monolingual corpora to train distributional models on). For each document, we are given its language, but not its topic. The task is to cluster this collection so that documents belonging to one topic were clustered together, independent of their language. Note that we are interested in clustering the collection as a whole, not each language separately (which is trivial).
Our evaluation data consists of comparable corpora of Russian and Ukrainian academic texts. On this material, we show that the `translated semantic fingerprints' method represents documents in different languages precisely enough to allow almost exact clustering according to document topics, with only 5% of incorrect assignments. It significantly outperforms both naive bag-of-words baseline and the not-so-naive method of `orthographic translation' based on Damerau-Levenshtein distance, even enriched with dictionary mappings. At the same time, it does not require large parallel corpora or a ready-made statistical machine translation model.
The rest of the paper is structured as follows. In Section "Related Work" we describe the foundations of our approach and the related work. Section "Academic texts as Comparable Corpora" introduces the employed corpora and the story behind them. Section "Learning to Translate: Ukrainian-to-Russian transformations" is dedicated to learning the transformation matrix, and Section "Experiment Design and Evaluation" describes our experimental setting and evaluation results. We discuss the findings in Section "Discussion" and conclude in Section "Conclusion and Future Work" , also suggesting directions for future work.
It significantly outperforms both naive bag-of-words baseline and the not-so-naive method of `orthographic trans lation' based on Damerau-Levenshtein distance, even en riched with dictionary mappings.
|
What are the advantages of the `translated semantic fingerprints' method?
|
It significantly outperforms both naive bag-of-words baseline and the not-so-naive method of `orthographic trans lation' based on Damerau-Levenshtein distance, even en riched with dictionary mappings.
|
null | false
| null |
What is Asperger syndrome?
|
Asperger syndrome (AS), or Asperger's, is a neurodevelopmental condition characterized by significant difficulties in social interaction and nonverbal communication, along with restricted and repetitive patterns of behavior and interests. The syndrome is no longer recognized as a diagnosis, having been merged with other conditions into autism spectrum disorder (ASD). It was considered to differ from other diagnoses that were incorporated into ASD by relatively unimpaired spoken language and intelligence.
The syndrome was named after the Austrian pediatrician Hans Asperger, who, in 1944, described children in his care who struggled to form friendships, did not understand others’ gestures or feelings, engaged in one-sided conversations about their favorite interests, and were clumsy. In 1994, the diagnosis of Asperger's was included in the fourth edition (DSM-IV) of the American Diagnostic and Statistical Manual of Mental Disorders; however, with the publication of DSM-5 in 2013, the syndrome was removed, and the symptoms are now included within autism spectrum disorder along with classic autism and pervasive developmental disorder not otherwise specified (PDD-NOS). It was similarly merged into autism spectrum disorder in the International Classification of Diseases (ICD-11) as of 2021.
The exact cause of Asperger's is poorly understood. While it has high heritability, the underlying genetics have not been determined conclusively. Environmental factors are also believed to play a role. Brain imaging has not identified a common underlying condition. There is no single treatment, and the UK's National Health Service (NHS) guidelines suggest that ’treatment’ of any form of autism should not be a goal since autism is not an 'a disease that can be removed or cured’. According to the Royal College of Psychiatrists, while co-occurring conditions might require treatment, ’management of autism itself is chiefly about providing the education, training and social support/care required to improve the person's ability to function in the everyday world’. Only limited data support the effectiveness of particular interventions for autism. Interventions may include social skills training, cognitive behavioral therapy, physical therapy, speech therapy, parent training, and medications for associated problems, such as mood or anxiety. Autistic characteristics tend to become less evident in adulthood , but social and communication difficulties usually persist.
|
|
null | false
| null |
What's the best binding setup for ski touring?
|
It depends on the emphasis of ski touring. If you favor more downhill performance, the CAST bindings is the clear winner as it utilizes the battle tested Look Pivot Bindings for downhill skiing, however, its heavy weight and complex transition actions may deter people. If you favor uphill performance, a pair of tech bindings may be a good choice, as they are light in weight, although they may not be as reliable. A good compromise between uphill and downhill may be Atomic Shift Bindings, they offer close to alpine binding experience for the downhill while being relatively light.
|
|
null | false
| null |
Is a vegan diet health for the average person?
|
Vegan diets can be very healthy depending on food choices as compared to the common American diet. It helps to choose organic foods as part of your dietary choices. It is quite easy to include good proteins and low glycemic carbs to power one's activities and brain functions. As long as you try and stay away from sugar-filled choices, most people feel better and loose weight on a well-planned vegan diet. This also keeps people off of high fat and high sugar fast foods, not healthy for anyone.
|
|
2001.01589
| false
| null |
We will elaborate two popular word segmentation methods and our newly proposed segmentation strategies in this section. The two popular segmentation methods are morpheme segmentation BIBREF4 and Byte Pair Encoding (BPE) BIBREF5. After word segmentation, we additionally add an specific symbol behind each separated subword unit, which aims to assist the NMT model to identify the morpheme boundaries and capture the semantic information effectively. The sentence examples with different segmentation strategies for Turkish-English machine translation task are shown in Table 1.
We utilize the Zemberek with a morphological disambiguation tool to segment the Turkish words into morpheme units, and utilize the morphology analysis tool BIBREF12 to segment the Uyghur words into morpheme units. We employ the python toolkits of jieba for Chinese word segmentation. We apply BPE on the target-side words and we set the number of merge operations to 35K for Chinese and 30K for English and we set the maximum sentence length to 150 tokens. The training corpus statistics of Turkish-English and Uyghur-Chinese machine translation tasks are shown in Table 2 and Table 3 respectively.
The two popular segmentation methods are morpheme segmentation BIBREF4 and Byte Pair Encoding (BPE) BIBREF5. After word segmentation, we additionally add an specific symbol behind each separated subword unit, which aims to assist the NMT model to identify the morpheme boundaries and capture the semantic information effectively.
We utilize the Zemberek with a morphological disambiguation tool to segment the Turkish words into morpheme units, and utilize the morphology analysis tool BIBREF12 to segment the Uyghur words into morpheme units.
|
How does the word segmentation method work?
|
The answers are shown as follows:
* morpheme segmentation BIBREF4 and Byte Pair Encoding (BPE) BIBREF5
* Zemberek
* BIBREF12
|
null | false
| null |
Wilcox is a Canadian drama film, directed by Denis Côté and released in 2019. Told entirely without dialogue, the film tells the story of Wilcox (Guillaume Tremblay), a mysterious drifter travelling in an apparent search for adventure. Although he does interact with other characters in the film, the dialogue taking place in the story is not heard by the audience.
|
How can the dialogue be described in this film?
|
There is no dialogue.
|
null | false
| null |
Which of the following is useful for transportation: a glass of wine, a lamp, a train, an iced cube tray, a plane, a bicycle, an apple and a scooter.
|
The useful objects for transportation in the list are train, plane, bicyle and scooter.
|
|
null | false
| null |
What is buoyant force?
|
The upward force exerted on a body, partially or fully immersed in a fluid, is known as buoyant force. This upward force is also called Upthrust. This is related to the Archimedes principle. If an object is partially or fully submerged in any fluid, the upward force and the fluid displaced is equal to the upward force exerted by the fluid.
|
|
null | false
| null |
Tell a child a story about a dinosaur who plays baseball but has to take a break due to an injury
|
There once was a dinosaur named Danny. Danny loved baseball and was the star hitter on his school baseball team, the Diamondbacks. Every time he hit a homer, he imagined he was sending a comet into space. One day Danny stepped up to the plate and the pitcher slipped while pitching the ball... the ball was coming straight at Danny! He tried to move out of the way but when he did, HE tripped and his foot bent sideways. Oh no, it hurt so much! Danny's coach took him to the doctor and the doctor said he would need surgery to fix his foot and that he should try a different hobby than baseball, so his foot would have time to heal. Danny started thinking about how he could keep that feeling of hitting homers while he wasn't able to play baseball. He started thinking about those comets he was hitting into space, and he decided to start reading about astronomy. Astronomy, it turns out, is really interesting! Danny learned about planets and suns and moons. He loved that he could imagine the solar system as one big baseball game. His aunt bought him a telescope so he could see his homers in outer space. Danny thought it was so fun to learn about astronomy, and he decided he would become an astronomer baseball star when he grew up. The end.
|
|
null | false
| null |
what can we do when coffee spill on laptop to make it working
|
Firstly we need to immediately switch off the laptop. secondly take a cloth and clean it as much as possible. Also, if possible remove the battery and use hair dryer for drying as much as possible.
|
|
null | false
| null |
Where does College Avenue begin and end in Berkeley, CA?
|
College Avenue extends from the southern point of the University of California campus on Bancroft Avenue south until the border of Oakland just past Alcatraz Avenue. College Avenue extends from there to Broadway but that portion is in the city of Oakland, CA.
|
|
null | false
| 143
|
The explosion of available scientific articles in the Biomedical domain has led to the rise of Biomedical Information Extraction (BioIE). BioIE systems aim to extract information from a wide spectrum of articles including medical literature, biological literature, electronic health records, etc. that can be used by clinicians and researchers in the field. Often the outputs of BioIE systems are used to assist in the creation of databases, or to suggest new paths for research. For example, a ranked list of interacting proteins that are extracted from biomedical literature, but are not present in existing databases, can allow researchers to make informed decisions about which protein/gene to study further. Interactions between drugs are necessary for clinicians who simultaneously administer multiple drugs to their patients. A database of diseases, treatments and tests is beneficial for doctors consulting in complicated medical cases.
The main problems in BioIE are similar to those in Information Extraction:
This paper discusses, in each section, various methods that have been adopted to solve the listed problems. Each section also highlights the difficulty of Information Extraction tasks in the biomedical domain.
This paper is intended as a primer to Biomedical Information Extraction for current NLP researchers. It aims to highlight the diversity of the various techniques from Information Extraction that have been applied in the Biomedical domain. The state of biomedical text mining is reviewed regularly. For more extensive surveys, consult BIBREF0 , BIBREF1 , BIBREF2 .
BioIE systems aim to extract information from a wide spectrum of articles including medical literature, biological literature, electronic health records, etc. that can be used by clinicians and researchers in the field.
|
What are the functions of BioS systems?
|
BioIE systems aim to extract information from a wide spectrum of articles including medical literature, biological literature, electronic health records, etc. that can be used by clinicians and researchers in the field.
|
null | false
| 88
|
While the LSTM encoder has richer capacity than NBOW, it produces the same representation for the entire text description regardless of its contexts. However, the text description may present an entity from various aspects, and various relations only focus on fractional aspects of the description. This phenomenon also occurs in structure embedding for an entity BIBREF8 , BIBREF19 .
Given a relation for an entity, not all of words/phrases in its text description are useful to model a specific fact. Some of them may be important for the given relation, but may be useless for other relations. Therefore, we introduce an attention mechanism BIBREF20 to utilize an attention-based encoder that constructs contextual text encodings according to different relations.
For each position $i$ of the text description, the attention for a given relation $r$ is defined as $\alpha _i(r)$ , which is
$$e_i(r) &= \mathbf {v}_a^T \tanh (\mathbf {W}_a {\mathbf {z}}_i + \mathbf {U}_a \mathbf {r}), \\
\alpha _i(r)&=\operatorname{\mathbf {softmax}}(e_i(r))\nonumber \\
&=\frac{\exp (e_i(r))}{\sum ^{n}_{j=1} \exp (e_j(r))},$$ (Eq. 12)
where $\mathbf {r}\in \mathbb {R}^d$ is the relation embedding; ${\mathbf {z}}_i \in \mathbb {R}^d$ is the output of BLSTM at position $i$ ; $\mathbf {W}_a,\mathbf {U}_a \in \mathbb {R}^{d\times d}$ are parameters matrices; $\mathbf {v}_a \in \mathbb {R}^{d}$ is a parameter vector.
The attention $\alpha _i(r)$ is interpreted as the degree to which the network attends to partial representation $\mathbf {z}_{i}$ for given relation $r$ .
The contextual encoding of text description can be formed by a weighted sum of the encoding $\mathbf {z}_{i}$ with attention.
$$\mathbf {enc_3}(x_{1:n};r) &= \sum _{i=1}^{n} \alpha _i(r) * \mathbf {z}_i.$$ (Eq. 13)
Given a relation for an entity, not all of words/phrases in its text description are useful to model a specific fact. Some of them may be important for the given relation, but may be useless for other relations. Therefore, we introduce an attention mechanism [Bahdanau et al., 2014] to utilize an attentionbased encoder that constructs contextual text encodings according to different relations.
|
Why is the attention mechanism introduced？
|
As not all of words/phrases in its text description are useful to model a specific fact，attentionbased encoder can construct contextual text encodings according to different relations.
|
null | false
| null |
Tell me which of these things are fruits or vegetables: Carrot, Apple, Orange, Potato, Peas, Broccoli, Pears
|
From the list you provided: Carrot, Potato, Peas, and Broccoli are Vegetables.
The rest are fruits: Apples, Oranges and Pears.
|
|
null | false
| 154
|
Question answering (QA) has been a blooming research field for the last decade. Selection-based QA implies a family of tasks that find answer contexts from large data given questions in natural language. Three tasks have been proposed for selection-based QA. Given a document, answer extraction BIBREF0 , BIBREF1 finds answer phrases whereas answer selection BIBREF2 , BIBREF3 , BIBREF4 , BIBREF5 and answer triggering BIBREF6 , BIBREF7 find answer sentences instead, although the presence of the answer context is not assumed within the provided document for answer triggering but it is for the other two tasks. Recently, various QA tasks that are not selection-based have been proposed BIBREF8 , BIBREF9 , BIBREF10 , BIBREF11 ; however, selection-based QA remains still important because of its practical value to real applications (e.g., IBM Watson, MIT Start).
Several datasets have been released for selection-based QA. wang:07a created the QASent dataset consisting of 277 questions, which has been widely used for benchmarking the answer selection task. feng:15a presented InsuranceQA comprising 16K+ questions on insurance contexts. yang:15a introduced WikiQA for answer selection and triggering. jurczyk:16 created SelQA for large real-scale answer triggering. rajpurkar2016squad presented SQuAD for answer extraction and selection as well as for reading comprehension. Finally, morales-EtAl:2016:EMNLP2016 provided InfoboxQA for answer selection.
These corpora make it possible to evaluate the robustness of statistical question answering learning. Although all of these corpora target on selection-based QA, they are designed for different purposes such that it is important to understand the nature of these corpora so a better use of them can be made. In this paper, we make both intrinsic and extrinsic analyses of four latest corpora based on Wikipedia, WikiQA, SelQA, SQuAD, and InfoboxQA. We first give a thorough intrinsic analysis regarding contextual similarities, question types, and answer categories (Section SECREF2 ). We then map questions in all corpora to the current version of English Wikipedia and benchmark another selection-based QA task, answer retrieval (Section SECREF3 ). Finally, we present an extrinsic analysis through a set of experiments cross-testing these corpora using a convolutional neural network architecture (Section SECREF4 ).
We first give a thorough intrinsic analysis regarding contextual similarities, question types, and answer categories (Section 2 ). We then map questions in all corpora to the current version of English Wikipedia and benchmark another selection-based QA task, answer retrieval (Section 3 ). Finally, we present an extrinsic analysis through a set of experiments cross-testing these corpora using a convolutional neural network architecture (Section 4 ).
|
How to analyze the corpora in this paper?
|
The authors make intrinsic analysis and extrinsic analysis to these corpora.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.