We introduce SexTok, a multi-modal dataset composed of TikTok videos labeled as sexually suggestive (from the annotator’s point of view), sex-educational content, or neither. Such a dataset is necessary to address the challenge of distinguishing between sexually suggestive content and virtual sex education videos on TikTok. Children’s exposure to sexually suggestive videos has been shown to have adversarial effects on their development (Collins et al. 2017). Meanwhile, virtual sex education, especially on subjects that are more relevant to the LGBTQIA+ community, is very valuable (Mitchell et al. 2014). The platform’s current system removes/punishes some of both types of videos, even though they serve different purposes. Our dataset contains video URLs, and it is also audio transcribed. To validate its importance, we explore two transformer-based models for classifying the videos. Our preliminary results suggest that the task of distinguishing between these types of videos is learnable but challenging. These experiments suggest that this dataset is meaningful and invites further study on the subject.

To avoid detection by current NLP monitoring applications, progenitors of hate speech often replace one or more letters in offensive words with homoglyphs, visually similar Unicode characters. Harvesting real-world hate speech containing homoglyphs is challenging due to the vast replacement possibilities. We developed a character substitution scraping method and assembled the Offensive Tweets with Homoglyphs (OTH) Dataset (N=90,788) with more than 1.5 million occurrences of 1,281 non-Latin characters (emojis excluded). In an annotated sample (n=700), 40.14% of the tweets were found to contain hate speech. We assessed the performance of seven transformer-based hate speech detection models and found that they performed poorly in a zero-shot setting (F1 scores between 0.04 and 0.52) but normalizing the data dramatically improved detection (F1 scores between 0.59 and 0.71). Training the models using the annotated data further boosted performance (highest micro-averaged F1 score=0.88, using five-fold cross validation). This study indicates that a dataset containing homoglyphs known and unknown to the scraping script can be collected, and that neural models can be trained to recognize camouflaged real-world hate speech.

This work presents a manually annotated dataset for Information Extraction (IE) from legal wills, and relevant in-context learning experiments on the dataset. The dataset consists of entities, binary relations between the entities (e.g., relations between testator and beneficiary), and n-ary events (e.g., bequest) extracted from 45 legal wills from two US states. This dataset can serve as a foundation for downstream tasks in the legal domain. Another use case of this dataset is evaluating the performance of large language models (LLMs) on this IE task. We evaluated GPT-4 with our dataset to investigate its ability to extract information from legal wills. Our evaluation result demonstrates that the model is capable of handling the task reasonably well. When given instructions and examples as a prompt, GPT-4 shows decent performance for both entity extraction and relation extraction tasks. Nevertheless, the evaluation result also reveals that the model is not perfect. We observed inconsistent outputs (given a prompt) as well as prompt over-generalization.

In this paper, we present a dataset of subjective views (beliefs and attitudes) held by individuals or groups. We analyze the usefulness of the dataset by training a neural classifier that identifies belief-containing sentences that are relevant for our broader project of interest—scientific modeling of complex systems. We also explore and discuss difficulties related to annotation of subjective views and propose ways of addressing them.

We propose an explainable approach for relation extraction that mitigates the tension between generalization and explainability by jointly training for the two goals. Our approach uses a multi-task learning architecture, which jointly trains a classifier for relation extraction, and a sequence model that labels words in the context of the relations that explain the decisions of the relation classifier. We also convert the model outputs to rules to bring global explanations to this approach. This sequence model is trained using a hybrid strategy: supervised, when supervision from pre-existing patterns is available, and semi-supervised otherwise. In the latter situation, we treat the sequence model’s labels as latent variables, and learn the best assignment that maximizes the performance of the relation classifier. We evaluate the proposed approach on the two datasets and show that the sequence model provides labels that serve as accurate explanations for the relation classifier’s decisions, and, importantly, that the joint training generally improves the performance of the relation classifier. We also evaluate the performance of the generated rules and show that the new rules are a great add-on to the manual rules and bring the rule-based system much closer to the neural models.

Text generation often involves producing coherent and grammatically correct texts that also satisfy a given set of semantic constraints. While most approaches for conditional text generation have primarily focused on lexical constraints, they often struggle to effectively incorporate syntactic constraints, which provide a richer language for approximating semantic constraints. We address this gap by introducing NeuroStructural Decoding, a new decoding algorithm that incorporates syntactic constraints to further improve the quality of the generated text. We build NeuroStructural Decoding on the NeuroLogic Decoding (Lu etal. 2021) algorithm, which enables language generation models to produce fluent text while satisfying complex lexical constraints. Our algorithm is powerful and scalable. It tracks lexico-syntactic constraints (e.g., we need to observe dog as subject and ball as object)during decoding by parsing the partial generations at each step. To this end, we adapt a dependency parser to generate parses for incomplete sentences. Our approach is evaluated on three different language generation tasks, and the results show improved performance in both lexical and syntactic metrics compared to previous methods. The results suggest this is a promising solution for integrating fine-grained controllable generation into the conventional beam search decoding.

We introduce a method that self trains (or bootstraps) neural relation and explanation classifiers. Our work expands the supervised approach of CITATION, which jointly trains a relation classifier with an explanation classifier that identifies context words important for the relation at hand, to semi-supervised scenarios. In particular, our approach iteratively converts the explainable models’ outputs to rules and applies them to unlabeled text to produce new annotations. Our evaluation on the TACRED dataset shows that our method outperforms the rule-based model we started from by 15 F1 points, outperforms traditional self-training that relies just on the relation classifier by 5 F1 points, and performs comparatively with the prompt-based approach of CITATION (without requiring an additional natural language inference component).

This study investigates whether a legal natural language inference (NLI) model trained on the data from one US state can be transferred to another state. We fine-tuned a pre-trained model on the task of evaluating the validity of legal will statements, once with the dataset containing the Tennessee wills and once with the dataset containing the Idaho wills. Each model’s performance on the in-domain setting and the out-of-domain setting are compared to see if the models can across the states. We found that the model trained on one US state can be mostly transferred to another state. However, it is clear that the model’s performance drops in the out-of-domain setting. The F1 scores of the Tennessee model and the Idaho model are 96.41 and 92.03 when predicting the data from the same state, but they drop to 66.32 and 81.60 when predicting the data from another state. Subsequent error analysis revealed that there are two major sources of errors. First, the model fails to recognize equivalent laws across states when there are stylistic differences between laws. Second, difference in statutory section numbering system between the states makes it difficult for the model to locate laws relevant to the cases being predicted on. This analysis provides insights on how the future NLI system can be improved. Also, our findings offer empirical support to legal experts advocating the standardization of legal documents.

We propose a novel task-agnostic in-domain pre-training method that sits between generic pre-training and fine-tuning. Our approach selectively masks in-domain keywords, i.e., words that provide a compact representation of the target domain. We identify such keywords using KeyBERT (Grootendorst, 2020). We evaluate our approach using six different settings: three datasets combined with two distinct pre-trained language models (PLMs). Our results reveal that the fine-tuned PLMs adapted using our in-domain pre-training strategy outperform PLMs that used in-domain pre-training with random masking as well as those that followed the common pre-train-then-fine-tune paradigm. Further, the overhead of identifying in-domain keywords is reasonable, e.g., 7-15% of the pre-training time (for two epochs) for BERT Large (Devlin et al., 2019).

While fully supervised relation classification (RC) models perform well on large-scale datasets, their performance drops drastically in low-resource settings. As generating annotated examples are expensive, recent zero-shot methods have been proposed that reformulate RC into other NLP tasks for which supervision exists such as textual entailment. However, these methods rely on templates that are manually created which is costly and requires domain expertise. In this paper, we present a novel strategy for template generation for relation classification, which is based on adapting Harris’ distributional similarity principle to templates encoded using contextualized representations. Further, we perform empirical evaluation of different strategies for combining the automatically acquired templates with manual templates. The experimental results on TACRED show that our approach not only performs better than the zero-shot RC methods that only use manual templates, but also that it achieves state-of-the-art performance for zero-shot TACRED at 64.3 F1 score.

We introduce a synthetic dataset called Sentences Involving Complex Compositional Knowledge (SICCK) and a novel analysis that investigates the performance of Natural Language Inference (NLI) models to understand compositionality in logic. We produce 1,304 sentence pairs by modifying 15 examples from the SICK dataset (Marelli et al., 2014). To this end, we modify the original texts using a set of phrases modifiers that correspond to universal quantifiers, existential quantifiers, negation, and other concept modifiers in Natural Logic (NL) (MacCartney, 2009). We use these phrases to modify the subject, verb, and object parts of the premise and hypothesis. Lastly, we annotate these modified texts with the corresponding entailment labels following NL rules. We conduct a preliminary verification of how well the change in the structural and semantic composition is captured by neural NLI models, in both zero-shot and fine-tuned scenarios. We found that the performance of NLI models under the zero-shot setting is poor, especially for modified sentences with negation and existential quantifiers. After fine-tuning this dataset, we observe that models continue to perform poorly over negation, existential and universal modifiers.

In this paper, we introduce and justify a new task—causal link extraction based on beliefs—and do a qualitative analysis of the ability of a large language model—InstructGPT-3—to generate implicit consequences of beliefs. With the language model-generated consequences being promising, but not consistent, we propose directions of future work, including data collection, explicit consequence extraction using rule-based and language modeling-based approaches, and using explicitly stated consequences of beliefs to fine-tune or prompt the language model to produce outputs suitable for the task.

In this paper we revisit the direction of using lexico-syntactic patterns for relation extraction instead of today’s ubiquitous neural classifiers. We propose a semi-supervised graph-based algorithm for pattern acquisition that scores patterns and the relations they extract jointly, using a variant of PageRank. We insert light supervision in the form of seed patterns or relations, and model it with several custom teleportation probabilities that bias random-walk scores of patterns/relations based on their proximity to correct information. We evaluate our approach on Few-Shot TACRED, and show that our method outperforms (or performs competitively with) more expensive and opaque deep neural networks. Lastly, we thoroughly compare our proposed approach with the seminal RlogF pattern acquisition algorithm of, showing that it outperforms it for all the hyper parameters tested, in all settings.

With their Discovery of Inference Rules from Text (DIRT) algorithm, Lin and Pantel (2001) made a seminal contribution to the field of rule acquisition from text, by adapting the distributional hypothesis of Harris (1954) to rules that model binary relations such as X treat Y. DIRT’s relevance is renewed in today’s neural era given the recent focus on interpretability in the field of natural language processing. We propose a novel take on the DIRT algorithm, where we implement the distributional hypothesis using the contextualized embeddings provided by BERT, a transformer-network-based language model (Vaswani et al. 2017; Devlin et al. 2018). In particular, we change the similarity measure between pairs of slots (i.e., the set of words matched by a rule) from the original formula that relies on lexical items to a formula computed using contextualized embeddings. We empirically demonstrate that this new similarity method yields a better implementation of the distributional hypothesis, and this, in turn, yields rules that outperform the original algorithm in the question answering-based evaluation proposed by Lin and Pantel (2001).

We propose an unsupervised method for the identification of bridge phrases in multi-hop question answering (QA). Our method constructs a graph of noun phrases from the question and the available context, and applies the Steiner tree algorithm to identify the minimal sub-graph that connects all question phrases. Nodes in the sub-graph that bridge loosely-connected or disjoint subsets of question phrases due to low-strength semantic relations are extracted as bridge phrases. The identified bridge phrases are then used to expand the query based on the initial question, helping in increasing the relevance of evidence that has little lexical overlap or semantic relation with the question. Through an evaluation on HotpotQA, a popular dataset for multi-hop QA, we show that our method yields: (a) improved evidence retrieval, (b) improved QA performance when using the retrieved sentences; and (c) effective and faithful explanations when answers are provided.

While deep learning approaches to information extraction have had many successes, they can be difficult to augment or maintain as needs shift. Rule-based methods, on the other hand, can be more easily modified. However, crafting rules requires expertise in linguistics and the domain of interest, making it infeasible for most users. Here we attempt to combine the advantages of these two directions while mitigating their drawbacks. We adapt recent advances from the adjacent field of program synthesis to information extraction, synthesizing rules from provided examples. We use a transformer-based architecture to guide an enumerative search, and show that this reduces the number of steps that need to be explored before a rule is found. Further, we show that without training the synthesis algorithm on the specific domain, our synthesized rules achieve state-of-the-art performance on the 1-shot scenario of a task that focuses on few-shot learning for relation classification, and competitive performance in the 5-shot scenario.

Can language models read biomedical texts and explain the biomedical mechanisms discussed? In this work we introduce a biomedical mechanism summarization task. Biomedical studies often investigate the mechanisms behind how one entity (e.g., a protein or a chemical) affects another in a biological context. The abstracts of these publications often include a focused set of sentences that present relevant supporting statements regarding such relationships, associated experimental evidence, and a concluding sentence that summarizes the mechanism underlying the relationship. We leverage this structure and create a summarization task, where the input is a collection of sentences and the main entities in an abstract, and the output includes the relationship and a sentence that summarizes the mechanism. Using a small amount of manually labeled mechanism sentences, we train a mechanism sentence classifier to filter a large biomedical abstract collection and create a summarization dataset with 22k instances. We also introduce conclusion sentence generation as a pretraining task with 611k instances. We benchmark the performance of large bio-domain language models. We find that while the pretraining task help improves performance, the best model produces acceptable mechanism outputs in only 32% of the instances, which shows the task presents significant challenges in biomedical language understanding and summarization.

This paper presents the phonological, morphological, and syntactic distinctions between formal and informal Persian, showing that these two variants have fundamental differences that cannot be attributed solely to pronunciation discrepancies. Given that informal Persian exhibits particular characteristics, any computational model trained on formal Persian is unlikely to transfer well to informal Persian, necessitating the creation of dedicated treebanks for this variety. We thus detail the development of the open-source Informal Persian Universal Dependency Treebank, a new treebank annotated within the Universal Dependencies scheme. We then investigate the parsing of informal Persian by training two dependency parsers on existing formal treebanks and evaluating them on out-of-domain data, i.e. the development set of our informal treebank. Our results show that parsers experience a substantial performance drop when we move across the two domains, as they face more unknown tokens and structures and fail to generalize well. Furthermore, the dependency relations whose performance deteriorates the most represent the unique properties of the informal variant. The ultimate goal of this study that demonstrates a broader impact is to provide a stepping-stone to reveal the significance of informal variants of languages, which have been widely overlooked in natural language processing tools across languages.

Annotation inconsistencies between data sets can cause problems for low-resource NLP, where noisy or inconsistent data cannot be easily replaced. We propose a method for automatically detecting annotation mismatches between dependency parsing corpora, along with three related methods for automatically converting the mismatches. All three methods rely on comparing unseen examples in a new corpus with similar examples in an existing corpus. These three methods include a simple lexical replacement using the most frequent tag of the example in the existing corpus, a GloVe embedding-based replacement that considers related examples, and a BERT-based replacement that uses contextualized embeddings to provide examples fine-tuned to our data. We evaluate these conversions by retraining two dependency parsers—Stanza and Parsing as Tagging (PaT)—on the converted and unconverted data. We find that applying our conversions yields significantly better performance in many cases. Some differences observed between the two parsers are observed. Stanza has a more complex architecture with a quadratic algorithm, taking longer to train, but it can generalize from less data. The PaT parser has a simpler architecture with a linear algorithm, speeding up training but requiring more training data to reach comparable or better performance.

Natural language inference (NLI) is critical in many domains requiring complex decision-making, such as the biomedical domain. We introduce a novel semi-supervised procedure that bootstraps biomedical NLI datasets from positive entailment examples present in abstracts of biomedical publications. We focus on challenging texts where the hypothesis includes mechanistic information such as biochemical interactions between two entities. A key contribution of this work is automating the creation of negative examples that are informative without being simplistic. We generate a range of negative examples using nine strategies that manipulate the structure of the underlying mechanisms both with rules, e.g., flip the roles of the entities in the interaction, and, more importantly, by imposing the perturbed conditions as logical constraints in a neuro-logical decoding system (CITATION).We use this procedure to create a novel dataset for NLI in the biomedical domain, called . The accuracy of neural classifiers on this dataset is in the mid 70s F1, which indicates that this NLI task remains to be solved. Critically, we observe that the performance on the different classes of negative examples varies widely, from 97% F1 on the simple negative examples that change the role of the entities in the hypothesis, to barely better than chance on the negative examples generated using neuro-logic decoding.

This work introduces a natural language inference (NLI) dataset that focuses on the validity of statements in legal wills. This dataset is unique because: (a) each entailment decision requires three inputs: the statement from the will, the law, and the conditions that hold at the time of the testator’s death; and (b) the included texts are longer than the ones in current NLI datasets. We trained eight neural NLI models in this dataset. All the models achieve more than 80% macro F1 and accuracy, which indicates that neural approaches can handle this task reasonably well. However, group accuracy, a stricter evaluation measure that is calculated with a group of positive and negative examples generated from the same statement as a unit, is in mid 80s at best, which suggests that the models’ understanding of the task remains superficial. Further ablative analyses and explanation experiments indicate that all three text segments are used for prediction, but some decisions rely on semantically irrelevant tokens. This indicates that overfitting on these longer texts likely happens, and that additional research is required for this task to be solved.

We propose a method to teach an automated agent to learn how to search for multi-hop paths of relations between entities in an open domain. The method learns a policy for directing existing information retrieval and machine reading resources to focus on relevant regions of a corpus. The approach formulates the learning problem as a Markov decision process with a state representation that encodes the dynamics of the search process and a reward structure that minimizes the number of documents that must be processed while still finding multi-hop paths. We implement the method in an actor-critic reinforcement learning algorithm and evaluate it on a dataset of search problems derived from a subset of English Wikipedia. The algorithm finds a family of policies that succeeds in extracting the desired information while processing fewer documents compared to several baseline heuristic algorithms.

We propose a system that assists a user in constructing transparent information extraction models, consisting of patterns (or rules) written in a declarative language, through program synthesis. Users of our system can specify their requirements through the use of examples,which are collected with a search interface. The rule-synthesis system proposes rule candidates and the results of applying them on a textual corpus; the user has the option to accept the candidate, request another option, or adjust the examples provided to the system. Through an interactive evaluation, we show that our approach generates high-precision rules even in a 1-shot setting. On a second evaluation on a widely-used relation extraction dataset (TACRED), our method generates rules that outperform considerably manually written patterns. Our code, demo, and documentation is available at https://clulab.github.io/odinsynth.

An existing domain taxonomy for normalizing content is often assumed when discussing approaches to information extraction, yet often in real-world scenarios there is none. When one does exist, as the information needs shift, it must be continually extended. This is a slow and tedious task, and one which does not scale well. Here we propose an interactive tool that allows a taxonomy to be built or extended rapidly and with a human in the loop to control precision. We apply insights from text summarization and information extraction to reduce the search space dramatically, then leverage modern pretrained language models to perform contextualized clustering of the remaining concepts to yield candidate nodes for the user to review. We show this allows a user to consider as many as 200 taxonomy concept candidates an hour, to quickly build or extend a taxonomy to better fit information needs.

Recognizing causal precedence relations among the chemical interactions in biomedical literature is crucial to understanding the underlying biological mechanisms. However, detecting such causal relation can be hard because: (1) many times, such causal relations among events are not explicitly expressed by certain phrases but implicitly implied by very diverse expressions in the text, and (2) annotating such causal relation detection datasets requires considerable expert knowledge and effort. In this paper, we propose a strategy to address both challenges by training neural models with in-domain pre-training and knowledge distillation. We show that, by using very limited amount of labeled data, and sufficient amount of unlabeled data, the neural models outperform previous baselines on the causal precedence detection task, and are ten times faster at inference compared to the BERT base model.

Many state-of-the-art (SOTA) language models have achieved high accuracy on several multi-hop reasoning problems. However, these approaches tend to not be interpretable because they do not make the intermediate reasoning steps explicit. Moreover, models trained on simpler tasks tend to fail when directly tested on more complex problems. We propose the Explainable multi-hop Verbal Reasoner (EVR) to solve these limitations by (a) decomposing multi-hop reasoning problems into several simple ones, and (b) using natural language to guide the intermediate reasoning hops. We implement EVR by extending the classic reasoning paradigm General Problem Solver (GPS) with a SOTA generative language model to generate subgoals and perform inference in natural language at each reasoning step. Evaluation of EVR on the RuleTaker synthetic question answering (QA) dataset shows that EVR achieves SOTA performance while being able to generate all reasoning steps in natural language. Furthermore, EVR generalizes better than other strong methods when trained on simpler tasks or less training data (up to 35.7% and 7.7% absolute improvement respectively).

While neural networks produce state-of-the-art performance in several NLP tasks, they generally depend heavily on lexicalized information, which transfer poorly between domains. We present a combination of two strategies to mitigate this dependence on lexicalized information in fact verification tasks. We present a data distillation technique for delexicalization, which we then combine with a model distillation method to prevent aggressive data distillation. We show that by using our solution, not only does the performance of an existing state-of-the-art model remain at par with that of the model trained on a fully lexicalized data, but it also performs better than it when tested out of domain. We show that the technique we present encourages models to extract transferable facts from a given fact verification dataset.

Multi-hop reasoning requires aggregation and inference from multiple facts. To retrieve such facts, we propose a simple approach that retrieves and reranks set of evidence facts jointly. Our approach first generates unsupervised clusters of sentences as candidate evidence by accounting links between sentences and coverage with the given query. Then, a RoBERTa-based reranker is trained to bring the most representative evidence cluster to the top. We specifically emphasize on the importance of retrieving evidence jointly by showing several comparative analyses to other methods that retrieve and rerank evidence sentences individually. First, we introduce several attention- and embedding-based analyses, which indicate that jointly retrieving and reranking approaches can learn compositional knowledge required for multi-hop reasoning. Second, our experiments show that jointly retrieving candidate evidence leads to substantially higher evidence retrieval performance when fed to the same supervised reranker. In particular, our joint retrieval and then reranking approach achieves new state-of-the-art evidence retrieval performance on two multi-hop question answering (QA) datasets: 30.5 Recall@2 on QASC, and 67.6% F1 on MultiRC. When the evidence text from our joint retrieval approach is fed to a RoBERTa-based answer selection classifier, we achieve new state-of-the-art QA performance on MultiRC and second best result on QASC.

We introduce a method that transforms a rule-based relation extraction (RE) classifier into a neural one such that both interpretability and performance are achieved. Our approach jointly trains a RE classifier with a decoder that generates explanations for these extractions, using as sole supervision a set of rules that match these relations. Our evaluation on the TACRED dataset shows that our neural RE classifier outperforms the rule-based one we started from by 9 F1 points; our decoder generates explanations with a high BLEU score of over 90%; and, the joint learning improves the performance of both the classifier and decoder.

The general goal of text simplification (TS) is to reduce text complexity for human consumption. In this paper, we investigate another potential use of neural TS: assisting machines performing natural language processing (NLP) tasks. We evaluate the use of neural TS in two ways: simplifying input texts at prediction time and augmenting data to provide machines with additional information during training. We demonstrate that the latter scenario provides positive effects on machine performance on two separate datasets. In particular, the latter use of TS improves the performances of LSTM (1.82–1.98%) and SpanBERT (0.7–1.3%) extractors on TACRED, a complex, large-scale, real-world relation extraction task. Further, the same setting yields improvements of up to 0.65% matched and 0.62% mismatched accuracies for a BERT text classifier on MNLI, a practical natural language inference dataset.

While neural networks produce state-of-the- art performance in several NLP tasks, they generally depend heavily on lexicalized information, which transfer poorly between domains. Previous works have proposed delexicalization as a form of knowledge distillation to reduce the dependency on such lexical artifacts. However, a critical unsolved issue that remains is how much delexicalization to apply: a little helps reduce overfitting, but too much discards useful information. We propose Group Learning, a knowledge and model distillation approach for fact verification in which multiple student models have access to different delexicalized views of the data, but are encouraged to learn from each other through pair-wise consistency losses. In several cross-domain experiments between the FEVER and FNC fact verification datasets, we show that our approach learns the best delexicalization strategy for the given training dataset, and outperforms state-of-the-art classifiers that rely on the original data.

We propose a simple yet accurate method for dependency parsing that treats parsing as tagging (PaT). That is, our approach addresses the parsing of dependency trees with a sequence model implemented with a bidirectional LSTM over BERT embeddings, where the “tag” to be predicted at each token position is the relative position of the corresponding head. For example, for the sentence John eats cake, the tag to be predicted for the token cake is -1 because its head (eats) occurs one token to the left. Despite its simplicity, our approach performs well. For example, our approach outperforms the state-of-the-art method of (Fernández-González and Gómez-Rodríguez, 2019) on Universal Dependencies (UD) by 1.76% unlabeled attachment score (UAS) for English, 1.98% UAS for French, and 1.16% UAS for German. On average, on 12 UD languages, our method with minimal tuning performs comparably with this state-of-the-art approach: better by 0.11% UAS, and worse by 0.58% LAS.

Modeling natural language inference is a challenging task. With large annotated data sets available it has now become feasible to train complex neural network based inference methods which achieve state of the art performance. However, it has been shown that these models also learn from the subtle biases inherent in these datasets (CITATION). In this work we explore two techniques for delexicalization that modify the datasets in such a way that we can control the importance that neural-network based methods place on lexical entities. We demonstrate that the proposed methods not only maintain the performance in-domain but also improve performance in some out-of-domain settings. For example, when using the delexicalized version of the FEVER dataset, the in-domain performance of a state of the art neural network method dropped only by 1.12% while its out-of-domain performance on the FNC dataset improved by 4.63%. We release the delexicalized versions of three common datasets used in natural language inference. These datasets are delexicalized using two methods: one which replaces the lexical entities in an overlap-aware manner, and a second, which additionally incorporates semantic lifting of nouns and verbs to their WordNet hypernym synsets

Evidence retrieval is a critical stage of question answering (QA), necessary not only to improve performance, but also to explain the decisions of the QA method. We introduce a simple, fast, and unsupervised iterative evidence retrieval method, which relies on three ideas: (a) an unsupervised alignment approach to soft-align questions and answers with justification sentences using only GloVe embeddings, (b) an iterative process that reformulates queries focusing on terms that are not covered by existing justifications, which (c) stops when the terms in the given question and candidate answers are covered by the retrieved justifications. Despite its simplicity, our approach outperforms all the previous methods (including supervised methods) on the evidence selection task on two datasets: MultiRC and QASC. When these evidence sentences are fed into a RoBERTa answer classification component, we achieve state-of-the-art QA performance on these two datasets.

We propose an interpretable approach for event extraction that mitigates the tension between generalization and interpretability by jointly training for the two goals. Our approach uses an encoder-decoder architecture, which jointly trains a classifier for event extraction, and a rule decoder that generates syntactico-semantic rules that explain the decisions of the event classifier. We evaluate the proposed approach on three biomedical events and show that the decoder generates interpretable rules that serve as accurate explanations for the event classifier’s decisions, and, importantly, that the joint training generally improves the performance of the event classifier. Lastly, we show that our approach can be used for semi-supervised learning, and that its performance improves when trained on automatically-labeled data generated by a rule-based system.

This paper explores an unsupervised approach to learning a compositional representation function for multi-word expressions (MWEs), and evaluates it on the Tratz dataset, which associates two-word expressions with the semantic relation between the compound constituents (e.g. the label employer is associated with the noun compound government agency) (Tratz, 2011). The composition function is based on recurrent neural networks, and is trained using the Skip-Gram objective to predict the words in the context of MWEs. Thus our approach can naturally leverage large unlabeled text sources. Further, our method can make use of provided MWEs when available, but can also function as a completely unsupervised algorithm, using MWE boundaries predicted by a single, domain-agnostic part-of-speech pattern. With pre-defined MWE boundaries, our method outperforms the previous state-of-the-art performance on the coarse-grained evaluation of the Tratz dataset (Tratz, 2011), with an F1 score of 50.4%. The unsupervised version of our method approaches the performance of the supervised one, and even outperforms it in some configurations.

Neural networks are a common tool in NLP, but it is not always clear which architecture to use for a given task. Different tasks, different languages, and different training conditions can all affect how a neural network will perform. Capsule Networks (CapsNets) are a relatively new architecture in NLP. Due to their novelty, CapsNets are being used more and more in NLP tasks. However, their usefulness is still mostly untested. In this paper, we compare three neural network architectures—LSTM, CNN, and CapsNet—on a part of speech tagging task. We compare these architectures in both high- and low-resource training conditions and find that no architecture consistently performs the best. Our analysis shows that our CapsNet performs nearly as well as a more complex LSTM under certain training conditions, but not others, and that our CapsNet almost always outperforms our CNN. We also find that our CapsNet implementation shows faster prediction times than the LSTM for Scottish Gaelic but not for Spanish, highlighting the effect that the choice of languages can have on the models.

Large pretrained language models (LM) have been used successfully for multi-hop question answering. However, most of these directions are not interpretable, as they do not make the inference hops necessary to explain a candidate answer explicitly. In this work, we investigate the capability of a state-of-the-art transformer LM to generate explicit inference hops, i.e., to infer a new statement necessary to answer a question given some premise input statements. Our analysis shows that such LMs can generate new statements for some simple inference types, but performance remains poor for complex, real-world inference types such as those that require monotonicity, composition, and commonsense knowledge.

We propose a simple, fast, and mostly-unsupervised approach for non-factoid question answering (QA) called Alignment over Heterogeneous Embeddings (AHE). AHE simply aligns each word in the question and candidate answer with the most similar word in the retrieved supporting paragraph, and weighs each alignment score with the inverse document frequency of the corresponding question/answer term. AHE’s similarity function operates over embeddings that model the underlying text at different levels of abstraction: character (FLAIR), word (BERT and GloVe), and sentence (InferSent), where the latter is the only supervised component in the proposed approach. Despite its simplicity and lack of supervision, AHE obtains a new state-of-the-art performance on the “Easy” partition of the AI2 Reasoning Challenge (ARC) dataset (64.6% accuracy), top-two performance on the “Challenge” partition of ARC (34.1%), and top-three performance on the WikiQA dataset (74.08% MRR), outperforming many other complex, supervised approaches. Our error analysis indicates that alignments over character, word, and sentence embeddings capture substantially different semantic information. We exploit this with a simple meta-classifier that learns how much to trust the predictions over each representation, which further improves the performance of unsupervised AHE.

Many of the most pressing current research problems (e.g., public health, food security, or climate change) require multi-disciplinary collaborations. In order to facilitate this process, we propose a system that incorporates multi-domain extractions of causal interactions into a single searchable knowledge graph. Our system enables users to search iteratively over direct and indirect connections in this knowledge graph, and collaboratively build causal models in real time. To enable the aggregation of causal information from multiple languages, we extend an open-domain machine reader to Portuguese. The new Portuguese reader extracts over 600 thousand causal statements from 120 thousand Portuguese publications with a precision of 62%, which demonstrates the value of mining multilingual scientific information.

Building causal models of complicated phenomena such as food insecurity is currently a slow and labor-intensive manual process. In this paper, we introduce an approach that builds executable probabilistic models from raw, free text. The proposed approach is implemented through three systems: Eidos, INDRA, and Delphi. Eidos is an open-domain machine reading system designed to extract causal relations from natural language. It is rule-based, allowing for rapid domain transfer, customizability, and interpretability. INDRA aggregates multiple sources of causal information and performs assembly to create a coherent knowledge base and assess its reliability. This assembled knowledge serves as the starting point for modeling. Delphi is a modeling framework that assembles quantified causal fragments and their contexts into executable probabilistic models that respect the semantics of the original text, and can be used to support decision making.

We propose a lightly-supervised approach for information extraction, in particular named entity classification, which combines the benefits of traditional bootstrapping, i.e., use of limited annotations and interpretability of extraction patterns, with the robust learning approaches proposed in representation learning. Our algorithm iteratively learns custom embeddings for both the multi-word entities to be extracted and the patterns that match them from a few example entities per category. We demonstrate that this representation-based approach outperforms three other state-of-the-art bootstrapping approaches on two datasets: CoNLL-2003 and OntoNotes. Additionally, using these embeddings, our approach outputs a globally-interpretable model consisting of a decision list, by ranking patterns based on their proximity to the average entity embedding in a given class. We show that this interpretable model performs close to our complete bootstrapping model, proving that representation learning can be used to produce interpretable models with small loss in performance. This decision list can be edited by human experts to mitigate some of that loss and in some cases outperform the original model.

Generating a large amount of training data for information extraction (IE) is either costly (if annotations are created manually), or runs the risk of introducing noisy instances (if distant supervision is used). On the other hand, semi-supervised learning (SSL) is a cost-efficient solution to combat lack of training data. In this paper, we adapt Mean Teacher (Tarvainen and Valpola, 2017), a denoising SSL framework to extract semantic relations between pairs of entities. We explore the sweet spot of amount of supervision required for good performance on this binary relation extraction task. Additionally, different syntax representations are incorporated into our models to enhance the learned representation of sentences. We evaluate our approach on the Google-IISc Distant Supervision (GDS) dataset, which removes test data noise present in all previous distance supervision datasets, which makes it a reliable evaluation benchmark (Jat et al., 2017). Our results show that the SSL Mean Teacher approach nears the performance of fully-supervised approaches even with only 10% of the labeled corpus. Further, the syntax-aware model outperforms other syntax-free approaches across all levels of supervision.

An important task in the machine reading of biochemical events expressed in biomedical texts is correctly reading the polarity, i.e., attributing whether the biochemical event is a promotion or an inhibition. Here we present a novel dataset for studying polarity attribution accuracy. We use this dataset to train and evaluate several deep learning models for polarity identification, and compare these to a linguistically-informed model. The best performing deep learning architecture achieves 0.968 average F1 performance in a five-fold cross-validation study, a considerable improvement over the linguistically informed model average F1 of 0.862.

We propose an unsupervised strategy for the selection of justification sentences for multi-hop question answering (QA) that (a) maximizes the relevance of the selected sentences, (b) minimizes the overlap between the selected facts, and (c) maximizes the coverage of both question and answer. This unsupervised sentence selection can be coupled with any supervised QA model. We show that the sentences selected by our method improve the performance of a state-of-the-art supervised QA model on two multi-hop QA datasets: AI2’s Reasoning Challenge (ARC) and Multi-Sentence Reading Comprehension (MultiRC). We obtain new state-of-the-art performance on both datasets among systems that do not use external resources for training the QA system: 56.82% F1 on ARC (41.24% on Challenge and 64.49% on Easy) and 26.1% EM0 on MultiRC. Our justification sentences have higher quality than the justifications selected by a strong information retrieval baseline, e.g., by 5.4% F1 in MultiRC. We also show that our unsupervised selection of justification sentences is more stable across domains than a state-of-the-art supervised sentence selection method.

While neural networks produce state-of-the-art performance in many NLP tasks, they generally learn from lexical information, which may transfer poorly between domains. Here, we investigate the importance that a model assigns to various aspects of data while learning and making predictions, specifically, in a recognizing textual entailment (RTE) task. By inspecting the attention weights assigned by the model, we confirm that most of the weights are assigned to noun phrases. To mitigate this dependence on lexicalized information, we experiment with two strategies of masking. First, we replace named entities with their corresponding semantic tags along with a unique identifier to indicate lexical overlap between claim and evidence. Second, we similarly replace other word classes in the sentence (nouns, verbs, adjectives, and adverbs) with their super sense tags (Ciaramita and Johnson, 2003). Our results show that, while performance on the in-domain dataset remains on par with that of the model trained on fully lexicalized data, it improves considerably when tested out of domain. For example, the performance of a state-of-the-art RTE model trained on the masked Fake News Challenge (Pomerleau and Rao, 2017) data and evaluated on Fact Extraction and Verification (Thorne et al., 2018) data improved by over 10% in accuracy score compared to the fully lexicalized model.

In this work we investigate the signal contained in the language of food on social media. We experiment with a dataset of 24 million food-related tweets, and make several observations. First,thelanguageoffoodhaspredictive power. We are able to predict if states in the United States (US) are above the medianratesfortype2diabetesmellitus(T2DM), income, poverty, and education – outperforming previous work by 4–18%. Second, we investigate the effect of socioeconomic factors (income, poverty, and education) on predicting state-level T2DM rates. Socioeconomic factors do improve T2DM prediction, with the greatestimprovementcomingfrompovertyinformation(6%),but,importantly,thelanguage of food adds distinct information that is not captured by socioeconomics. Third, we analyze how the language of food has changed over a five-year period (2013 – 2017), which is indicative of the shift in eating habits in the US during that period. We find several food trends, and that the language of food is used differently by different groups such as differentgenders. Last,weprovideanonlinevisualization tool for real-time queries and semantic analysis.

Noise is inherent in real world datasets and modeling noise is critical during training as it is effective in regularization. Recently, novel semi-supervised deep learning techniques have demonstrated tremendous potential when learning with very limited labeled training data in image processing tasks. A critical aspect of these semi-supervised learning techniques is augmenting the input or the network with noise to be able to learn robust models. While modeling noise is relatively straightforward in continuous domains such as image classification, it is not immediately apparent how noise can be modeled in discrete domains such as language. Our work aims to address this gap by exploring different noise strategies for the semi-supervised named entity classification task, including statistical methods such as adding Gaussian noise to input embeddings, and linguistically-inspired ones such as dropping words and replacing words with their synonyms. We compare their performance on two benchmark datasets (OntoNotes and CoNLL) for named entity classification. Our results indicate that noise strategies that are linguistically informed perform at least as well as statistical approaches, while being simpler and requiring minimal tuning.

We present the Named Entity Recognition (NER) and disambiguation model used by the University of Arizona team (UArizona) for the SemEval 2019 task 12. We achieved fourth place on tasks 1 and 3. We implemented a deep-affix based LSTM-CRF NER model for task 1, which utilizes only character, word, pre- fix and suffix information for the identification of geolocation entities. Despite using just the training data provided by task organizers and not using any lexicon features, we achieved 78.85% strict micro F-score on task 1. We used the unsupervised population heuristics for task 3 and achieved 52.99% strict micro-F1 score in this task.

We introduce a machine learning approach for the identification of “white spaces” in scientific knowledge. Our approach addresses this task as link prediction over a graph that contains over 2M influence statements such as “CTCF activates FOXA1”, which were automatically extracted using open-domain machine reading. We model this prediction task using graph-based features extracted from the above influence graph, as well as from a citation graph that captures scientific communities. We evaluated the proposed approach through backtesting. Although the data is heavily unbalanced (50 times more negative examples than positives), our approach predicts which influence links will be discovered in the “near future” with a F1 score of 27 points, and a mean average precision of 68%.

Many approaches to automatically recognizing entailment relations have employed classifiers over hand engineered lexicalized features, or deep learning models that implicitly capture lexicalization through word embeddings. This reliance on lexicalization may complicate the adaptation of these tools between domains. For example, such a system trained in the news domain may learn that a sentence like “Palestinians recognize Texas as part of Mexico” tends to be unsupported, but this fact (and its corresponding lexicalized cues) have no value in, say, a scientific domain. To mitigate this dependence on lexicalized information, in this paper we propose a model that reads two sentences, from any given domain, to determine entailment without using lexicalized features. Instead our model relies on features that are either unlexicalized or are domain independent such as proportion of negated verbs, antonyms, or noun overlap. In its current implementation, this model does not perform well on the FEVER dataset, due to two reasons. First, for the information retrieval portion of the task we used the baseline system provided, since this was not the aim of our project. Second, this is work in progress and we still are in the process of identifying more features and gradually increasing the accuracy of our model. In the end, we hope to build a generic end-to-end classifier, which can be used in a domain outside the one in which it was trained, with no or minimal re-training.

This work explores the detection of individuals’ risk of type 2 diabetes mellitus (T2DM) directly from their social media (Twitter) activity. Our approach extends a deep learning architecture with several contributions: following previous observations that language use differs by gender, it captures and uses gender information through domain adaptation; it captures recency of posts under the hypothesis that more recent posts are more representative of an individual’s current risk status; and, lastly, it demonstrates that in this scenario where activity factors are sparsely represented in the data, a bag-of-word neural network model using custom dictionaries of food and activity words performs better than other neural sequence models. Our best model, which incorporates all these contributions, achieves a risk-detection F1 of 41.9, considerably higher than the baseline rate (36.9).

Several semi-supervised representation learning methods have been proposed recently that mitigate the drawbacks of traditional bootstrapping: they reduce the amount of semantic drift introduced by iterative approaches through one-shot learning; others address the sparsity of data through the learning of custom, dense representation for the information modeled. In this work, we are the first to adapt three of these methods, most of which have been originally proposed for image processing, to an information extraction task, specifically, named entity classification. Further, we perform a rigorous comparative analysis on two distinct datasets. Our analysis yields several important observations. First, all representation learning methods outperform state-of-the-art semi-supervised methods that do not rely on representation learning. To the best of our knowledge, we report the latest state-of-the-art results on the semi-supervised named entity classification task. Second, one-shot learning methods clearly outperform iterative representation learning approaches. Lastly, one of the best performers relies on the mean teacher framework (Tarvainen and Valpola, 2017), a simple teacher/student approach that is independent of the underlying task-specific model.

We challenge a common assumption in active learning, that a list-based interface populated by informative samples provides for efficient and effective data annotation. We show how a 2D scatterplot populated with diverse and representative samples can yield improved models given the same time budget. We consider this for bootstrapping-based information extraction, in particular named entity classification, where human and machine jointly label data. To enable effective data annotation in a scatterplot, we have developed an embedding-based bootstrapping model that learns the distributional similarity of entities through the patterns that match them in a large data corpus, while being discriminative with respect to human-labeled and machine-promoted entities. We conducted a user study to assess the effectiveness of these different interfaces, and analyze bootstrapping performance in terms of human labeling accuracy, label quantity, and labeling consensus across multiple users. Our results suggest that supervision acquired from the scatterplot interface, despite being noisier, yields improvements in classification performance compared with the list interface, due to a larger quantity of supervision acquired.

We propose a novel approach to semi-supervised learning for information extraction that uses ladder networks (Rasmus et al., 2015). In particular, we focus on the task of named entity classification, defined as identifying the correct label (e.g., person or organization name) of an entity mention in a given context. Our approach is simple, efficient and has the benefit of being robust to semantic drift, a dominant problem in most semi-supervised learning systems. We empirically demonstrate the superior performance of our system compared to the state-of-the-art on two standard datasets for named entity classification. We obtain between 62% and 200% improvement over the state-of-art baseline on these two datasets.

Recent efforts in bioinformatics have achieved tremendous progress in the machine reading of biomedical literature, and the assembly of the extracted biochemical interactions into large-scale models such as protein signaling pathways. However, batch machine reading of literature at today’s scale (PubMed alone indexes over 1 million papers per year) is unfeasible due to both cost and processing overhead. In this work, we introduce a focused reading approach to guide the machine reading of biomedical literature towards what literature should be read to answer a biomedical query as efficiently as possible. We introduce a family of algorithms for focused reading, including an intuitive, strong baseline, and a second approach which uses a reinforcement learning (RL) framework that learns when to explore (widen the search) or exploit (narrow it). We demonstrate that the RL approach is capable of answering more queries than the baseline, while being more efficient, i.e., reading fewer documents.

We propose a question answering (QA) approach for standardized science exams that both identifies correct answers and produces compelling human-readable justifications for why those answers are correct. Our method first identifies the actual information needed in a question using psycholinguistic concreteness norms, then uses this information need to construct answer justifications by aggregating multiple sentences from different knowledge bases using syntactic and lexical information. We then jointly rank answers and their justifications using a reranking perceptron that treats justification quality as a latent variable. We evaluate our method on 1,000 multiple-choice questions from elementary school science exams, and empirically demonstrate that it performs better than several strong baselines, including neural network approaches. Our best configuration answers 44% of the questions correctly, where the top justifications for 57% of these correct answers contain a compelling human-readable justification that explains the inference required to arrive at the correct answer. We include a detailed characterization of the justification quality for both our method and a strong baseline, and show that information aggregation is key to addressing the information need in complex questions.

For many applications of question answering (QA), being able to explain why a given model chose an answer is critical. However, the lack of labeled data for answer justifications makes learning this difficult and expensive. Here we propose an approach that uses answer ranking as distant supervision for learning how to select informative justifications, where justifications serve as inferential connections between the question and the correct answer while often containing little lexical overlap with either. We propose a neural network architecture for QA that reranks answer justifications as an intermediate (and human-interpretable) step in answer selection. Our approach is informed by a set of features designed to combine both learned representations and explicit features to capture the connection between questions, answers, and answer justifications. We show that with this end-to-end approach we are able to significantly improve upon a strong IR baseline in both justification ranking (+9% rated highly relevant) and answer selection (+6% P@1).

We describe challenges and advantages unique to coreference resolution in the biomedical domain, and a sieve-based architecture that leverages domain knowledge for both entity and event coreference resolution. Domain-general coreference resolution algorithms perform poorly on biomedical documents, because the cues they rely on such as gender are largely absent in this domain, and because they do not encode domain-specific knowledge such as the number and type of participants required in chemical reactions. Moreover, it is difficult to directly encode this knowledge into most coreference resolution algorithms because they are not rule-based. Our rule-based architecture uses sequentially applied hand-designed “sieves”, with the output of each sieve informing and constraining subsequent sieves. This architecture provides a 3.2% increase in throughput to our Reach event extraction system with precision parallel to that of the stricter system that relies solely on syntactic patterns for extraction.

Odin is an information extraction framework that applies cascades of finite state automata over both surface text and syntactic dependency graphs. Support for syntactic patterns allow us to concisely define relations that are otherwise difficult to express in languages such as Common Pattern Specification Language (CPSL), which are currently limited to shallow linguistic features. The interaction of lexical and syntactic automata provides robustness and flexibility when writing extraction rules. This paper describes Odin’s declarative language for writing these cascaded automata.

We describe a strategy for the acquisition of training data necessary to build a social-media-driven early detection system for individuals at risk for (preventable) type 2 diabetes mellitus (T2DM). The strategy uses a game-like quiz with data and questions acquired semi-automatically from Twitter. The questions are designed to inspire participant engagement and collect relevant data to train a public-health model applied to individuals. Prior systems designed to use social media such as Twitter to predict obesity (a risk factor for T2DM) operate on entire communities such as states, counties, or cities, based on statistics gathered by government agencies. Because there is considerable variation among individuals within these groups, training data on the individual level would be more effective, but this data is difficult to acquire. The approach proposed here aims to address this issue. Our strategy has two steps. First, we trained a random forest classifier on data gathered from (public) Twitter statuses and state-level statistics with state-of-the-art accuracy. We then converted this classifier into a 20-questions-style quiz and made it available online. In doing so, we achieved high engagement with individuals that took the quiz, while also building a training set of voluntarily supplied individual-level data for future classification.

QA systems have been making steady advances in the challenging elementary science exam domain. In this work, we develop an explanation-based analysis of knowledge and inference requirements, which supports a fine-grained characterization of the challenges. In particular, we model the requirements based on appropriate sources of evidence to be used for the QA task. We create requirements by first identifying suitable sentences in a knowledge base that support the correct answer, then use these to build explanations, filling in any necessary missing information. These explanations are used to create a fine-grained categorization of the requirements. Using these requirements, we compare a retrieval and an inference solver on 212 questions. The analysis validates the gains of the inference solver, demonstrating that it answers more questions requiring complex inference, while also providing insights into the relative strengths of the solvers and knowledge sources. We release the annotated questions and explanations as a resource with broad utility for science exam QA, including determining knowledge base construction targets, as well as supporting information aggregation in automated inference.

Lexical semantic models provide robust performance for question answering, but, in general, can only capitalize on direct evidence seen during training. For example, monolingual alignment models acquire term alignment probabilities from semi-structured data such as question-answer pairs; neural network language models learn term embeddings from unstructured text. All this knowledge is then used to estimate the semantic similarity between question and answer candidates. We introduce a higher-order formalism that allows all these lexical semantic models to chain direct evidence to construct indirect associations between question and answer texts, by casting the task as the traversal of graphs that encode direct term associations. Using a corpus of 10,000 questions from Yahoo! Answers, we experimentally demonstrate that higher-order methods are broadly applicable to alignment and language models, across both word and syntactic representations. We show that an important criterion for success is controlling for the semantic drift that accumulates during graph traversal. All in all, the proposed higher-order approach improves five out of the six lexical semantic models investigated, with relative gains of up to +13% over their first-order variants.

We investigate the importance of text analysis for stock price prediction. In particular, we introduce a system that forecasts companies’ stock price changes (UP, DOWN, STAY) in response to financial events reported in 8-K documents. Our results indicate that using text boosts prediction accuracy over 10% (relative) over a strong baseline that incorporates many financially-rooted features. This impact is most important in the short term (i.e., the next day after the financial event) but persists for up to five days.

Distant supervision is a successful paradigm that gathers training data for information extraction systems by automatically aligning vast databases of facts with text. Previous work has demonstrated its usefulness for the extraction of binary relations such as a person’s employer or a film’s director. Here, we extend the distant supervision approach to template-based event extraction, focusing on the extraction of passenger counts, aircraft types, and other facts concerning airplane crash events. We present a new publicly available dataset and event extraction task in the plane crash domain based on Wikipedia infoboxes and newswire text. Using this dataset, we conduct a preliminary evaluation of four distantly supervised extraction models which assign named entity mentions in text to entries in the event template. Our results indicate that joint inference over sequences of candidate entity mentions is beneficial. Furthermore, we demonstrate that the Searn algorithm outperforms a linear-chain CRF and strong baselines with local inference.