chunk_id
stringlengths
3
7
chunk
stringlengths
1
823
source_url
stringclasses
416 values
title
stringclasses
416 values
chunk_idx
int64
0
294
chunk_start_char
int64
0
139k
chunk_end_char
int64
303
139k
21_25
In this paper, we focus on text-based classification methods that do not aim to explicitly link general web pages to their knowledge-base counterparts. Therefore, we are interested in the ability of neural models to generalize beyond their training context. In addition to testing across time and topics, we also invest...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
25
14,306
14,941
21_26
Both neural models also outperform both language models on AUC significantly (p < 0.05). Precision and Recall are more mixed, with the CNN and SVM outperforming the HAN on Precision and the language model -again- performing best in terms of Recall. Together, the neural methods seem to work best on three out of the fou...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
26
14,941
15,487
21_27
In Table TABREF20 , we present three Spearman correlation metrics to express model congruence with human annotations. Mean annotation expresses the correlation of model error rates with the controversy values attributed to a web page by human annotators, with positive values expressing greater error rates on controver...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
27
15,487
16,054
21_28
a score between 0 and 2.5 that expresses how sure annotators are of document (non)controversy. Here, the HAN shows errors most strongly negatively correlated to the certainty of annotators. Finally, annotators disagree on the controversy of some documents, expressed as the standard deviation of their controversy annot...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
28
16,054
16,491
21_29
Overall, the neural methods have less biased performance in relation to (non)controversial documents, correlate more strongly with the certainty of human annotators and are susceptible to errors in similar conditions as when annotators disagree. Conclusion Controversy detection is a hard task, as it forms a latent c...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
29
16,491
16,985
21_30
First, we have demonstrated that neural methods perform as state-of-the-art tools in controversy detection on the ClueWeb09 BIBREF0 based testset, even beating matching models. Second, we investigated temporal stability, and demonstrated neural -and especially CNN- robustness in terms of Recall, F1 and AUC performanc...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
30
16,985
17,503
21_31
Fourthly, we show that neural methods are better able to generalize from Wikipedia pages to unseen general web pages in terms of Precision, F1 and AUC. Lastly, neural methods seem better in line with human annotators with regard to certainty and disagreement. Table 1:Wikipedia derived dataset statistics. Including th...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
31
17,503
18,012
21_32
Results obtained by evaluating on the Wikipedia derived dataset from 2018 by either: models trained on Wikipedia data from 2018 or 2009. Trained on data from the same time frame, the neural models show a slight advantage over the lexical models. Most noticeable however is the drop in performance by the lexical models ...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
32
18,012
18,511
21_33
Table 2: Comparison of results with previous work Table 4: Cross-topic stability experiment.Metrics are averaged across 10 leave-on-out topic folds. Table 6: Spearman’s correlations for estimated probability distance from true label.Mean controversy: Average annotator score, certainty: distance from controversy ann...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
33
18,511
19,157
22_0
Neural Language Modeling with Visual Features Multimodal language models attempt to incorporate non-linguistic features for the language modeling task. In this work, we extend a standard recurrent neural network (RNN) language model with features derived from videos. We train our models on data that is two orders-of-m...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
0
0
466
22_1
Our experiments on two corpora (YouCookII and 20bn-something-something-v2) show that the best performing architecture consists of middle fusion of visual and text features, yielding over 25% relative improvement in perplexity. We report analysis that provides insights into why our multimodal language model improves up...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
1
466
1,127
22_2
However, most language models are trained and applied in a manner that is oblivious to the environment in which human language operates BIBREF0 . These models are typically trained only on sequences of words, ignoring the physical context in which the symbolic representations are grounded, or ignoring the social conte...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
2
1,127
1,764
22_3
Despite the lack of big datasets, researchers have started investigating language grounding in images BIBREF8 , BIBREF9 , BIBREF10 and to lesser extent in videos BIBREF11 , BIBREF1 . However, language grounding has focused more on obtaining better word and sentence representations or other downstream tasks, and to les...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
3
1,764
2,397
22_4
Notably, this work did not use the recurrent neural model paradigm which has now become the de facto way of implementing neural LMs. The closest work to ours is that of BIBREF0 , who report perplexity gains of around 5–6% on three languages on the MS COCO dataset (with an English vocabulary of only 16K words). Our w...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
4
2,397
2,957
22_5
For a given video segment, we assume that there is a sequence of INLINEFORM0 video frames represented by features INLINEFORM1 , and the corresponding transcription INLINEFORM2 . In practice, we assume INLINEFORM3 since we can always assign a video frame to each word by replicating the video frames the requisite numbe...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
5
2,957
3,558
22_6
We opt for the simplest strategy, which concatenates the representations. For a word embedding INLINEFORM0 and corresponding visual representation INLINEFORM1 , the input to our RNNLM will be the concatenated vector INLINEFORM2 . For the examples where we were unable to compute visual features (see Section § SECREF3 )...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
6
3,558
3,919
22_7
In addition to concatenating the word and visual embedding, we explore two variants of our model that allow for a finer-grained integration of the two modalities: In this case, the RNNLM is given as input a vector INLINEFORM0 that is a weighted sum of the two embeddings: INLINEFORM1 where INLINEFORM0 are learned m...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
7
3,919
4,534
22_8
For other word histories, though, the visual context might not be needed or be even irrelevant for the next word prediction: if the previous word is “carpe", the next word is very likely to be “diem", regardless of visual context. We implement a simple weighting mechanism that learns a scalar weight for the visual emb...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
8
4,534
4,908
22_9
The input to the RNNLM is now INLINEFORM0 , where: INLINEFORM1 This approach does not add any new parameters to the model, but since the word representations INLINEFORM0 are learned, this mechanism has the potential to learn word embeddings that are also appropriate for weighting the visual context. Location of com...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
9
4,908
5,514
22_10
The Middle Fusion merges the visual features at the output of the 1st LSTM layer while the Late Fusion strategies merges the two features after the final LSTM layer. The idea behind the Middle and Late fusion is that we would like to minimize changes to the regular RNNLM architecture at the early stages and still be a...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
10
5,514
6,104
22_11
Thus, the input to the model is a sequence of wordpieces. Using wordpieces allows us to address out-of-vocabulary (OOV) word issues that would arise from having a fixed word vocabulary. In practice, a wordpiece RNNLM gives similar performance as a word-level model BIBREF16 . For about INLINEFORM1 of the segments, we w...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
11
6,104
6,663
22_12
For a INLINEFORM2 -second video and INLINEFORM3 wordpieces, each feature is uniformly allocated to INLINEFORM4 wordpieces. Our RNNLM models consist of 2 LSTM layers, each containing 2048 units which are linearly projected to 512 units BIBREF19 . The word-piece and video embeddings are of size 512 each. We do not use ...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
12
6,663
7,109
22_13
The INLINEFORM0 -norms of the gradients are clipped to a max norm of INLINEFORM1 for the LSTM weights and to 10,000 for all other weights. We train with Synchronous SGD with the Adafactor optimizer BIBREF20 until convergence on a development set, created by randomly selecting INLINEFORM2 of all utterances. Experiment...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
13
7,109
7,597
22_14
Note that no data from these datasets are present in the YouTube videos used for training. The perplexity of our models is shown in Table . Conclusion We present a simple strategy to augment a standard recurrent neural network language model with temporal visual features. Through an exploration of candidate architec...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
14
7,597
8,234
22_15
Our work is a first step towards creating and deploying large-scale multimodal systems that properly situate themselves into a given context, by taking full advantage of every available signal. Figure 1: Visualization of our different Language Models. Given word and visual embeddings, the input can be created by thre...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
15
8,234
8,825
22_16
Note: ⊕ denotes concatenation,⊗ denotes matrix multiplication, denotes dot product. Table 2: Withholding visual context from our best model leads to worse performance (similar to an RNNLM trained only on text). Table 1: Middle Fusion of text and frame-level visual features leads to significant reductions in perplexi...
https://arxiv.org/abs/1903.02930
Neural Language Modeling with Visual Features
16
8,825
9,284
23_0
Can Neural Networks Learn Symbolic Rewriting? This work investigates if the current neural architectures are adequate for learning symbolic rewriting. Two kinds of data sets are proposed for this research -- one based on automated proofs and the other being a synthetic set of polynomial terms. The experiments with use...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
0
0
702
23_1
One of the tasks in this domain is translation between natural languages – neural machine translation (NMT) systems established here the state-of-the-art performance. Recently, NMT produced first encouraging results in the autoformalization task BIBREF0, BIBREF1, BIBREF2, BIBREF3 where given an informal mathematical t...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
1
702
1,109
23_2
In particular, the NMT performance on a large synthetic -to-Mizar dataset produced by a relatively sophisticated toolchain developed for several decades BIBREF4 is surprisingly good BIBREF3, indicating that neural networks can learn quite complicated algorithms for symbolic data. This inspired us to pose a question: C...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
2
1,109
1,604
23_3
For example, neural models could compete with symbolic methods such as inductive logic programming BIBREF5 (ILP) that have been previously experimented with to learn simple rewrite tasks and theorem-proving heuristics from large formal corpora BIBREF6. Unlike (early) ILP, neural methods can however easily cope with la...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
3
1,604
2,083
23_4
Data To perform experiments answering our question we prepared two data sets – the first consists of examples extracted from proofs found by ATP (automated theorem prover) in a mathematical domain (AIM loops), whereas the second is a synthetic set of polynomial terms. Data ::: The AIM data set The data consists of...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
4
2,083
2,527
23_5
Many of the inferences in the proofs are paramodulations from an equation and have the form s = t u[(s)] = vu[(t)] = v where $s, t, u, v$ are terms and $\theta $ is a substitution. For the most common equations $s = t$, we gathered corresponding pairs of terms $\big (u[\theta (s)], u[\theta (t)]\big )$ which were re...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
5
2,527
2,888
23_6
We put the pairs to separate data sets (depending on the corresponding $s = t$): in total 8 data sets for ground rewrites (where $\theta $ is trivial) and 12 for nonground ones. The goal will be to learn rewriting for each of this 20 rules separately. Terms in the examples are treated as linear sequences of tokens wh...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
6
2,888
3,368
23_7
Lengths of the sequences of tokens vary between 1 and 343, with mean around 35. These 20 data sets were split into training, validation and test sets for our experiments ($60 \%, 10 \%, 30 \%$, respectively). In Table TABREF4 and Table TABREF5 there are presented examples of pairs of AIM terms in TPTP BIBREF9 format,...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
7
3,368
3,771
23_8
Data ::: The polynomial data set This is a synthetically created data set where the examples are pairs of equivalent polynomial terms. The first element of each pair is a polynomial in an arbitrary form and the second element is the same polynomial in a normalized form. The arbitrary polynomials are created randomly...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
8
3,771
4,410
23_9
The parameters of this process are set in such a way that it creates polynomial terms of average length around 25 symbols. Terms longer than 50 are filtered out. Several data sets of various difficulty were created by varying the number of available symbols. This were quite limited – at most 5 different variables and...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
9
4,410
5,030
23_10
Each data set consists of different 300 000 examples, see Table TABREF7 for examples. These data sets were split into training, validation and test sets for our experiments ($60 \%, 10 \%, 30 \%$, respectively). Experiments For experiments with both data sets we used an established NMT architecture BIBREF10 based o...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
10
5,030
5,430
23_11
After a small grid search we decided to inherit most of the hyperparameters of the model from the best results achieved in BIBREF3 where -to-Mizar translation is learned. We used relatively small LSTM cells consisting of 2 layers with 128 units. The “scaled Luong” version of the attention mechanism was used, as well ...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
11
5,430
5,888
23_12
Experiments ::: AIM data set First, NMT models were trained for each of the 20 rewrite rules in the AIM data set. It turned out that the models, as long as the number of examples was greater than 1000, were able to learn the rewriting task very well, reaching $90\%$ of accuracy on separated test sets. This means tha...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
12
5,888
6,329
23_13
We also run an experiment on the joint set of all rewrite rules (consisting of 41396 examples). Here the task was more difficult as a model needed not only to apply rewriting correctly, but also choose “the right” rewrite rule applicable for a given term. Nevertheless, the performance was also very good, reaching $83...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
13
6,329
6,816
23_14
Depending on the difficulty of the data, accuracy on the test sets achieved in our experiments varied between $70\%$ and $99\%$. The results in terms of accuracy are shown in Table TABREF13. This high performance of the model encouraged a closer inspection of the results. First, we checked if in the test sets there a...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
14
6,816
7,317
23_15
After filtering them out the measured accuracy drops – but only by $1 - 2 \%$. An examination of the examples wrongly rewritten by the model was done. It turns out that the wrong outputs almost always parse (in $97 - 99 \%$ of cases they are legal polynomial terms). Notably, depending on the difficulty of the data se...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
15
7,317
7,823
23_16
Below $1 \%$ of wrong outputs are correct modulo variable renaming. Conclusions and future work NMT is not typically applied to symbolic problems, but surprisingly, it performed very well for both described tasks. The first one was easier in terms of complexity of the rewriting (only one application of a rewrite rul...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
16
7,823
8,504
23_17
We see several interesting directions this work can be extended. Firstly, more interesting and difficult rewriting problems need to be provided for better delineation of the strength of the neural models. The described data are relatively simple and with no direct relevance to the real unsolved symbolic problems. But...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
17
8,504
9,117
23_18
Specifically, we are going to implement an approach based on the idea of TreeNN, which may be another effective approach for this kind of tasks BIBREF7, BIBREF12, BIBREF13. TreeNNs are built recursively from modules, where the modules corresponds to parts of symbolic expression (symbols) and the shape of the network r...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
18
9,117
9,641
23_19
Perhaps this way the model could learn more efficiently from examples (and achieve higher results even on the small AIM data sets). The authors have a positive experience of applying TreeNNs to learn remainders of arithmetical expressions modulo small natural numbers – TreeNNs outperformed here neural models based on ...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
19
9,641
10,267
23_20
This can be motivated by the interesting contrast between some simpler ILP systems suffering for combinatorial explosion in presence of a large number of examples and neural methods which definitely benefit form large data sets. We hope that this work will inspire and trigger a discussion on the above (and other) ide...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
20
10,267
10,821
23_21
649043 AI4REASON and by the Czech project AI&Reasoning CZ.02.1.01/0.0/0.0/15_003/0000466 and the European Regional Development Fund. Kaliszyk was supported by ERC Starting grant no. 714034 SMART. Table 4. Results of experiments with AIM data. (Names of the rules correspond to folder names in the Github repo.) Table ...
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
21
10,821
11,192
23_22
(Characteristic of formulas concerns the input polynomials. Labels of the data sets correspond to folder names in the Github repo.)
https://arxiv.org/abs/1911.04873
Can Neural Networks Learn Symbolic Rewriting?
22
11,192
11,324
24_0
Toward Interpretable Topic Discovery via Anchored Correlation Explanation Many predictive tasks, such as diagnosing a patient based on their medical chart, are ultimately defined by the decisions of human experts. Unfortunately, encoding experts' knowledge is often time consuming and expensive. We propose a simple way...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
0
0
727
24_1
We give a preliminary evaluation of Anchored CorEx, showing that it produces more coherent and interpretable topics on two distinct corpora. Introduction A clinician can look at a patient's electronic health record (EHR) and not only decide whether the patient has diabetes but also produce a succinct summary of the ...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
1
727
1,360
24_2
However, representing the intuitive judgments of human experts can be challenging, particularly when the formal system does not match the expert's knowledge. For example, many deterministic disease classifiers used in clinical informatics rely heavily upon administrative codes not available at time of diagnosis. Furth...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
2
1,360
1,928
24_3
For example, to discover patients with diabetes in a set of clinical notes, a doctor can begin by specifying disease-specific anchor terms BIBREF1 , BIBREF2 , such as “diabetes” or “insulin.” Our framework then uses these to help discover both latent topics associated with diabetes and records in which diabetes-relate...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
3
1,928
2,398
24_4
In this workshop paper, we introduce a simple approach to anchored information theoretic topic modeling using a novel combination of Correlation Explanation (CorEx) BIBREF3 and the information bottleneck BIBREF4 . This flexible framework enables the user to leverage domain knowledge to guide exploration of a collecti...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
4
2,398
2,988
24_5
What is more, we demonstrate the potential for this framework to perform weakly supervised learning in settings where labeling documents is prohibitively expensive BIBREF5 , BIBREF6 . With respect to interpretable machine learning, our contributions are twofold. First, our framework provides a way for human users to ...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
5
2,988
3,725
24_6
Methods Anchored Correlation Explanation can be understood as a combination of Total Correlation Explanation (CorEx) BIBREF3 , BIBREF7 and the multivariate information bottleneck BIBREF4 , BIBREF8 . We search for a set of probabilistic functions of the inputs INLINEFORM0 for INLINEFORM1 that optimize the following i...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
6
3,725
4,344
24_7
Total correlation, or multivariate mutual information BIBREF9 , is specified as INLINEFORM5 where INLINEFORM6 is the KL divergence. Maximizing INLINEFORM7 over latent factors INLINEFORM8 amounts to minimizing INLINEFORM9 , which measures how much dependence in INLINEFORM10 is explained by INLINEFORM11 . At the global ...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
7
4,344
4,877
24_8
The second term involves the mutual information between pairs of latent factors INLINEFORM0 ) and anchor variables INLINEFORM1 specified in the set INLINEFORM2 . This is inspired by the information bottleneck BIBREF4 , BIBREF8 , a supervised information-theoretic approach to discovering latent factors. The bottleneck...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
8
4,877
5,525
24_9
This framework is flexible: we can attach multiple anchors to one factor or one anchor to multiple factors. We have found empirically that INLINEFORM1 works well and does not need to be tuned. Anchors allow us to both seed CorEx and impose semantics on latent factors: when analyzing medical documents, for example, we...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
9
5,525
6,018
24_10
While there is not space here for a full description of the optimization, it is similar in principle to the approaches in BIBREF3 , BIBREF7 . Two points are worth noting: first, the TC objective is replaced by a lower bound to make optimization feasible BIBREF7 . Second, we impose a sparse connection constraint (each...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
10
6,018
6,465
24_11
Related Work There is a large body of work on integrating domain knowledge into topic models and other unsupervised latent variable models, often in the form of constraints BIBREF13 , prior distributions BIBREF14 , and token labels BIBREF15 . Like Anchored CorEx, seeded latent dirichlet allocation (SeededLDA) allows...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
11
6,465
6,986
24_12
BIBREF1 first proposed anchors in the context of topic modeling: words that are high precision indicators of underlying topics. In contrast to our approach, anchors are typically selected automatically, constrained to appear in only one topic, and used primarily to aid optimization BIBREF17 . In our information theo...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
12
6,986
7,628
24_13
In health informatics, “anchor” features chosen based on domain knowledge have been used to guide statistical learning BIBREF2 . In BIBREF6 , anchors are used as a source of distant supervision BIBREF19 , BIBREF20 for classifiers in the absence of ground truth labels. While Anchored CorEx can be used for discriminati...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
13
7,628
8,198
24_14
We utilize an information theoretic approach that makes no generative modeling assumptions. Results and Discussion To demonstrate the utility of Anchored CorEx, we run experiments on two document collections: 20 Newsgroups and the i2b2 2008 Obesity Challenge BIBREF22 data set. Both corpora provide ground truth label...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
14
8,198
8,686
24_15
The latent classes represent mutually exclusive categories, and each document is known to originate from a single category. We find that the correlation structure among the latent classes is less complex than in the Obesity Challenge data. Further, each category tends to exhibit some specialized vocabulary not used ex...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
15
8,686
9,339
24_16
All CorEx models used three layers of 40, 3, and 1 factors. fig:big shows an example hierarchical topic model extracted by Anchored CorEx. i2b2 Obesity Challenge 2008 The Obesity Challenge 2008 data set includes 1237 deidentified clinical discharge summaries from the Partners HealthCare Research Patient Data Reposit...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
16
9,339
9,663
24_17
All summaries have been labeled by clinical experts with obesity and 15 other conditions commonly comorbid with obesity, ranging from Coronary Artery Disease (663 positives) to Depression (247) to Hypertriglyceridemia (62). We preprocessed each document with a standard biomedical text pipeline that extracts common me...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
17
9,663
10,136
24_18
We converted each document to a binary bag-of-words with a vocabulary of 4114 (possibly negated) medical phrases. We used the 60/40 training/test split from the competition. We are primarily interested in the ability of Anchored CorEx to extract latent topics that are unambiguously associated with the 16 known condit...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
18
10,136
10,577
24_19
tab:obesity:topics shows the Obesity and Obstructive Sleep Apnea (OSA) topics for three iterations of Anchored CorEx with the ten most important terms (highest weighted connections to the latent factor) listed for each topic. Unsupervised CorEx (first row) does not discover any topics obviously related to obesity or O...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
19
10,577
11,101
24_20
In the second iteration (second row), we add the common name of each of the 16 diseases as an anchor to one factor (16 total). Adding obesity as an anchor produces a clear Obesity topic, including several medications known to cause weight gain (e.g., acebutolol, klonopin). The anchored OSA topic, however, is quite po...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
20
11,101
11,597
24_21
This is likely due to the fact that obesity is a major risk factor of OSA, and so OSA symptoms are highly correlated with obesity and its other symptoms. Thus, the total correlation objective will attempt to group obesity and OSA-related terms together under a single latent factor. The sparse connection constraint me...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
21
11,597
12,107
24_22
In the third iteration, we correct this by adding sleep apnea as a second anchor to the OSA topic, and the resulting topic is clearly associated with OSA, including terms related to respiratory problems and medications used to treat (or believed to increase risk for) OSA. There is no noticeable reduction in quality i...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
22
12,107
12,627
24_23
We interpret each anchored latent factor as a classifier for an associated class label and then compute test set F1 (using a threshold of 0.5) and area under the curve (AUC) scores (Obesity Challenge only). tab:class compares the classification performance of Unsupervised and Anchored CorEx on the soc.religion.christ...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
23
12,627
13,207
24_24
Unsupervised CorEx does a reasonable job of discovering a coherent religion topic that already contains the terms God, Christian, and Jesus. However, using the terms Jesus and Christian as anchors yields a topic that better predicts the actual soc.religion.christianity category. tab:obesity:class shows the Macro-AUC ...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
24
13,207
13,727
24_25
Surprisingly, Anchored CorEx outperforms Naive Bayes (NB) by a large margin. Of course, Anchored CorEx is not a replacement for supervised learning: NB beats Anchored CorEx on 20 Newsgroups and does not represent a “strong” baseline for Obesity 2008 (teams scored above 0.7 in Macro-F1 during the competition). It is no...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
25
13,727
14,160
24_26
Conclusion We have introduced a simple information theoretic approach to topic modeling that can leverage domain knowledge specified informally as anchors. Our framework uses a novel combination of CorEx and the information bottleneck. Preliminary results suggest it can extract more precise, interpretable topics thr...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
26
14,160
14,846
24_27
Figure 1. A hierarchical topic model learned by CorEx. Anchored latent factors are labeled in red with anchor words marked with a “*”. Table 1. Evolution of Obesity and Obstructive Sleep Apnea (OSA) topics as anchors are added. Colors and font weight indicate anchors, spurious terms, and intruder terms from other kn...
https://arxiv.org/abs/1606.07043
Toward Interpretable Topic Discovery via Anchored Correlation Explanation
27
14,846
15,355
25_0
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media We focus on named entity recognition (NER) for Chinese social media. With massive unlabeled text and quite limited labelled corpus, we propose a semi-supervised learning model based on B-LSTM neural network. To take advantage...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
0
0
554
25_1
When considering the instability of F-score driven method and meaningful information provided by label accuracy, we propose an integrated method to train on both F-score and label accuracy. Our integrated model yields 7.44\% improvement over previous state-of-the-art result. Introduction With the development of Inte...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
1
554
1,097
25_2
As the foundation of many downstream applications BIBREF4 , BIBREF5 , BIBREF6 such as information extraction, named entity recognition (NER) deserves more research in prevailing and challenging social media text. NER is a task to identify names in texts and to assign names with particular types BIBREF7 , BIBREF8 , BIB...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
2
1,097
1,652
25_3
It is caused by Chinese logographic characters which lack many clues to indicate whether a word is a name, such as capitalization. The scant labelled Chinese social media corpus makes the task more challenging BIBREF11 , BIBREF12 , BIBREF13 . To address the problem, one approach is to use the lexical embeddings learn...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
3
1,652
2,220
25_4
Considering the value of word segmentation in Chinese NER, another approach is to construct an integrated model to jointly train learned representations for both predicting word segmentations and NER BIBREF14 . However, the two above approaches are implemented within CRF model. We construct a semi-supervised model ba...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
4
2,220
2,819
25_5
In addition, we propose an integrated method to train on both F-Score and label accuracy. Specifically, we make contributions as follows: Model We construct a semi-supervised model which is based on B-LSTM neural network and combine transition probability to form structured output. We propose a method to train direc...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
5
2,819
3,257
25_6
Transition Probability B-LSTM neural network can learn from past input features and LSTM layer makes it more efficient BIBREF15 , BIBREF16 , BIBREF17 , BIBREF18 . However, B-LSTM cannot learn sentence level label information. Huang et al. huang2015bidirectional combine CRF to use sentence level label information. We...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
6
3,257
3,801
25_7
The prediction of label in position INLINEFORM0 is given as: DISPLAYFORM0 where INLINEFORM0 are the transformation parameters, INLINEFORM1 the hidden vector and INLINEFORM2 the bias parameter. For a input sentence INLINEFORM3 with a label sequence INLINEFORM4 , a sentence-level score is then given as: DISPLAYFORM0 ...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
7
3,801
4,315
25_8
In our model, INLINEFORM5 is computed as: DISPLAYFORM0 We define a structured margin loss INLINEFORM0 as Pei et al. pei-ge-chang:2014:P14-1: DISPLAYFORM0 where INLINEFORM0 is the length of setence INLINEFORM1 , INLINEFORM2 is a discount parameter, INLINEFORM3 a given correct label sequence and INLINEFORM4 a predic...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
8
4,315
4,654
25_9
For a given training instance INLINEFORM5 , our predicted label sequence is the label sequence with highest score: INLINEFORM6 The label sequence with the highest score can be obtained by carrying out viterbi algorithm. The regularized objective function is as follows: DISPLAYFORM0 INLINEFORM0 By minimizing the ob...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
9
4,654
5,326
25_10
In fact, the structured margin loss INLINEFORM3 reflect the loss in label accuracy. Considering the gap between label accuracy and F-Score in NER, we introduce a new training method to train directly on F-Score. To introduce F-Score driven training method, we need to take a look at the subgradient of equation ( EQREF9...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
10
5,326
5,826
25_11
The margin loss INLINEFORM2 serves as a trigger function to conduct the training process of B-LSTM based MMNN. We can introduce a new trigger function to guide the training process of neural network. F-Score Trigger Function The main criterion of NER task is F-score. However, high label accuracy does not mean high F-...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
11
5,826
6,304
25_12
We use the F-Score between corrected label sequence and predicted label sequence as trigger function, which can conduct the training process to optimize the F-Score of training examples. Our new structured margin loss can be described as: DISPLAYFORM0 where INLINEFORM0 is the F-Score between corrected label sequence...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
12
6,304
6,878
25_13
To take advantage of meaningful information provided by label accuracy, we introduce an integrated trigger function as follows: DISPLAYFORM0 where INLINEFORM0 is a factor to adjust the weight of label accuracy and F-Score. Because F-Score depends on the whole label sequence, we use beam search to find INLINEFORM0 l...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
13
6,878
7,451
25_14
Both Peng and Dredze peng-dredze:2015:EMNLP and Peng and Dredze peng-dredze:2016:P16-2 show the value of word segmentation to Chinese NER in social media. We present two methods to use word segmentation information in neural network model. Character and Position Embeddings To incorporate word segmentation information...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
14
7,451
8,007
25_15
Character Embeddings and Word Segmentation Features We can treat word segmentation as discrete features in neural network model. The discrete features can be easily incorporated into neural network model BIBREF20 . We use word embeddings from a LSTM pretrained on MSRA 2006 corpus to initialize the word segmentation f...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
15
8,007
8,506
25_16
We also use the same unlabelled text as Peng and Dredze peng-dredze:2016:P16-2 from Sina Weibo service in China and the text is word segmented by a Chinese word segmentation system Jieba as Peng and Dredze peng-dredze:2016:P16-2 so that our results are more comparable to theirs. Parameter Estimation We pre-trained e...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
16
8,506
8,974
25_17
mao2008chinese, we use bigram features as follow: INLINEFORM0 We use window approach BIBREF20 to extract higher level Features from word feature vectors. We treat bigram features as discrete features BIBREF20 for our neural network. Our models are trained using stochastic gradient descent with an L2 regularizer.
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
17
8,974
9,290
25_18
As for parameters in our models, window size for word embedding is 5, word embedding dimension, feature embedding dimension and hidden vector dimension are all 100, discount INLINEFORM0 in margin loss is INLINEFORM1 , and the hyper parameter for the INLINEFORM2 is INLINEFORM3 . As for learning rate, initial learning ...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
18
9,290
9,865
25_19
The results of two methods are shown as Table TABREF22 . We can see that positional character embeddings perform better in neural network. This is probably because positional character embeddings method can learn word segmentation information from unlabeled text while word segmentation can only use training corpus. W...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
19
9,865
10,436
25_20
We design a F-Score driven training method in our third model F-Score Driven Model I . We propose an integrated training method in our fourth model F-Score Driven Model II .The results of models are depicted as Figure UID11 . From the figure, we can know our models perfrom better with little loss in time. Table TABRE...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
20
10,436
10,902
25_21
Peng and Dredze peng-dredze:2016:P16-2 is the state-of-the-art NER system in Chinese Social media. By comparing the results of B-LSTM model and B-LSTM + MTNN model, we can know transition probability is significant for NER. Compared with B-LSTM + MMNN model, F-Score Driven Model I improves the result of named entity w...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
21
10,902
11,252
25_22
The integrated training model (F-Score Driven Model II) benefits from both label accuracy and F-Score, which achieves a new state-of-the-art NER system in Chinese social media. Our integrated model has better performance on named entity and nominal mention. To better understand the impact of the factor INLINEFORM0 , ...
https://arxiv.org/abs/1611.04234
F-Score Driven Max Margin Neural Network for Named Entity Recognition in Chinese Social Media
22
11,252
11,782