Incorporating Item Response Theory (IRT) into NLP tasks can provide valuable information about model performance and behavior. Traditionally, IRT models are learned using human response pattern (RP) data, presenting a significant bottleneck for large data sets like those required for training deep neural networks (DNNs). In this work we propose learning IRT models using RPs generated from artificial crowds of DNN models. We demonstrate the effectiveness of learning IRT models using DNN-generated data through quantitative and qualitative analyses for two NLP tasks. Parameters learned from human and machine RPs for natural language inference and sentiment analysis exhibit medium to large positive correlations. We demonstrate a use-case for latent difficulty item parameters, namely training set filtering, and show that using difficulty to sample training data outperforms baseline methods. Finally, we highlight cases where human expectation about item difficulty does not match difficulty as estimated from the machine RPs.
Statistical natural language inference (NLI) models are susceptible to learning dataset bias: superficial cues that happen to associate with the label on a particular dataset, but are not useful in general, e.g., negation words indicate contradiction. As exposed by several recent challenge datasets, these models perform poorly when such association is absent, e.g., predicting that "I love dogs" contradicts "I don't love cats". Our goal is to design learning algorithms that guard against known dataset bias. We formalize the concept of dataset bias under the framework of distribution shift and present a simple debiasing algorithm based on residual fitting, which we call DRiFt. We first learn a biased model that only uses features that are known to relate to dataset bias. Then, we train a debiased model that fits to the residual of the biased model, focusing on examples that cannot be predicted well by biased features only. We use DRiFt to train three high-performing NLI models on two benchmark datasets, SNLI and MNLI. Our debiased models achieve significant gains over baseline models on two challenge test sets, while maintaining reasonable performance on the original test sets.
Our goal of patent claim generation is to realize "augmented inventing" for inventors by leveraging latest Deep Learning techniques. We envision the possibility of building an "auto-complete" function for inventors to conceive better inventions in the era of artificial intelligence. In order to generate patent claims with good quality, a fundamental question is how to measure it. We tackle the problem from a perspective of claim span relevancy. Patent claim language was rarely explored in the NLP field. It is unique in its own way and contains rich explicit and implicit human annotations. In this work, we propose a span-based approach and a generic framework to measure patent claim generation quantitatively. In order to study the effectiveness of patent claim generation, we define a metric to measure whether two consecutive spans in a generated patent claims are relevant. We treat such relevancy measurement as a span-pair classification problem, following the concept of natural language inference. Technically, the span-pair classifier is implemented by fine-tuning a pre-trained language model. The patent claim generation is implemented by fine-tuning the other pre-trained model. Specifically, we fine-tune a pre-trained Google BERT model to measure the patent claim spans generated by a fine-tuned OpenAI GPT-2 model. In this way, we re-use two of the state-of-the-art pre-trained models in the NLP field. Our result shows the effectiveness of the span-pair classifier after fine-tuning the pre-trained model. It further validates the quantitative metric of span relevancy in patent claim generation. Particularly, we found that the span relevancy ratio measured by BERT becomes lower when the diversity in GPT-2 text generation becomes higher.
Word embeddings carry stereotypical connotations from the text they are trained on, which can lead to invalid inferences in downstream models that rely on them. We use this observation to design a mechanism for measuring stereotypes using the task of natural language inference. We demonstrate a reduction in invalid inferences via bias mitigation strategies on static word embeddings (GloVe). Further, we show that for gender bias, these techniques extend to contextualized embeddings when applied selectively only to the static components of contextualized embeddings (ELMo, BERT).
Scenario-based question answering (SQA) has attracted increasing research attention. It typically requires retrieving and integrating knowledge from multiple sources, and applying general knowledge to a specific case described by a scenario. SQA widely exists in the medical, geography, and legal domains---both in practice and in the exams. In this paper, we introduce the GeoSQA dataset. It consists of 1,981 scenarios and 4,110 multiple-choice questions in the geography domain at high school level, where diagrams (e.g., maps, charts) have been manually annotated with natural language descriptions to benefit NLP research. Benchmark results on a variety of state-of-the-art methods for question answering, textual entailment, and reading comprehension demonstrate the unique challenges presented by SQA for future research.
The state-of-the-art pre-trained language representation models, such as Bidirectional Encoder Representations from Transformers (BERT), rarely incorporate commonsense knowledge or other knowledge explicitly. We propose a pre-training approach for incorporating commonsense knowledge into language representation models. We construct a commonsense-related multi-choice question answering dataset for pre-training a neural language representation model. The dataset is created automatically by our proposed "align, mask, and select" (AMS) method. We also investigate different pre-training tasks. Experimental results demonstrate that pre-training models using the proposed approach followed by fine-tuning achieve significant improvements over previous state-of-the-art models on two commonsense-related benchmarks, including CommonsenseQA and Winograd Schema Challenge. We also observe that fine-tuned models after the proposed pre-training approach maintain comparable performance on other NLP tasks, such as sentence classification and natural language inference tasks, compared to the original BERT models. These results verify that the proposed approach, while significantly improving commonsense-related NLP tasks, does not degrade the general language representation capabilities.
Abductive reasoning is inference to the most plausible explanation. For example, if Jenny finds her house in a mess when she returns from work, and remembers that she left a window open, she can hypothesize that a thief broke into her house and caused the mess, as the most plausible explanation. While abduction has long been considered to be at the core of how people interpret and read between the lines in natural language (Hobbs et al., 1988), there has been relatively little research in support of abductive natural language inference and generation. We present the first study that investigates the viability of language-based abductive reasoning. We introduce a challenge dataset, ART, that consists of over 20k commonsense narrative contexts and 200k explanations. Based on this dataset, we conceptualize two new tasks -- (i) Abductive NLI: a multiple-choice question answering task for choosing the more likely explanation, and (ii) Abductive NLG: a conditional generation task for explaining given observations in natural language. On Abductive NLI, the best model achieves 68.9% accuracy, well below human performance of 91.4%. On Abductive NLG, the current best language generators struggle even more, as they lack reasoning capabilities that are trivial for humans. Our analysis leads to new insights into the types of reasoning that deep pre-trained language models fail to perform--despite their strong performance on the related but more narrowly defined task of entailment NLI--pointing to interesting avenues for future research.
The ability to learn from large unlabeled corpora has allowed neural language models to advance the frontier in natural language understanding. However, existing self-supervision techniques operate at the word form level, which serves as a surrogate for the underlying semantic content. This paper proposes a method to employ weak-supervision directly at the word sense level. Our model, named SenseBERT, is pre-trained to predict not only the masked words but also their WordNet supersenses. Accordingly, we attain a lexical-semantic level language model, without the use of human annotation. SenseBERT achieves significantly improved lexical understanding, as we demonstrate by experimenting on SemEval Word Sense Disambiguation, and by attaining a state of the art result on the Word in Context task.
Natural language understanding (NLU) of text is a fundamental challenge in AI, and it has received significant attention throughout the history of NLP research. This primary goal has been studied under different tasks, such as Question Answering (QA) and Textual Entailment (TE). In this thesis, we investigate the NLU problem through the QA task and focus on the aspects that make it a challenge for the current state-of-the-art technology. This thesis is organized into three main parts: In the first part, we explore multiple formalisms to improve existing machine comprehension systems. We propose a formulation for abductive reasoning in natural language and show its effectiveness, especially in domains with limited training data. Additionally, to help reasoning systems cope with irrelevant or redundant information, we create a supervised approach to learn and detect the essential terms in questions. In the second part, we propose two new challenge datasets. In particular, we create two datasets of natural language questions where (i) the first one requires reasoning over multiple sentences; (ii) the second one requires temporal common sense reasoning. We hope that the two proposed datasets will motivate the field to address more complex problems. In the final part, we present the first formal framework for multi-step reasoning algorithms, in the presence of a few important properties of language use, such as incompleteness, ambiguity, etc. We apply this framework to prove fundamental limitations for reasoning algorithms. These theoretical results provide extra intuition into the existing empirical evidence in the field.
Recently, the pre-trained language model, BERT (and its robustly optimized version RoBERTa), has attracted a lot of attention in natural language understanding (NLU), and achieved state-of-the-art accuracy in various NLU tasks, such as sentiment classification, natural language inference, semantic textual similarity and question answering. Inspired by the linearization exploration work of Elman [8], we extend BERT to a new model, StructBERT, by incorporating language structures into pre-training. Specifically, we pre-train StructBERT with two auxiliary tasks to make the most of the sequential order of words and sentences, which leverage language structures at the word and sentence levels, respectively. As a result, the new model is adapted to different levels of language understanding required by downstream tasks. The StructBERT with structural pre-training gives surprisingly good empirical results on a variety of downstream tasks, including pushing the state-of-the-art on the GLUE benchmark to 89.0 (outperforming all published models), the F1 score on SQuAD v1.1 question answering to 93.0, the accuracy on SNLI to 91.7.
Humans understand language based on the rich background knowledge about how the physical world works, which in turn allows us to reason about the physical world through language. In addition to the properties of objects (e.g., boats require fuel) and their affordances, i.e., the actions that are applicable to them (e.g., boats can be driven), we can also reason about if-then inferences between what properties of objects imply the kind of actions that are applicable to them (e.g., that if we can drive something then it likely requires fuel). In this paper, we investigate the extent to which state-of-the-art neural language representations, trained on a vast amount of natural language text, demonstrate physical commonsense reasoning. While recent advancements of neural language models have demonstrated strong performance on various types of natural language inference tasks, our study based on a dataset of over 200k newly collected annotations suggests that neural language representations still only learn associations that are explicitly written down.
In this paper we present DELTA, a deep learning based language technology platform. DELTA is an end-to-end platform designed to solve industry level natural language and speech processing problems. It integrates most popular neural network models for training as well as comprehensive deployment tools for production. DELTA aims to provide easy and fast experiences for using, deploying, and developing natural language processing and speech models for both academia and industry use cases. We demonstrate the reliable performance with DELTA on several natural language processing and speech tasks, including text classification, named entity recognition, natural language inference, speech recognition, speaker verification, etc. DELTA has been used for developing several state-of-the-art algorithms for publications and delivering real production to serve millions of users.
In this paper, we present a fast and strong neural approach for general purpose text matching applications. We explore what is sufficient to build a fast and well-performed text matching model and propose to keep three key features available for inter-sequence alignment: original point-wise features, previous aligned features, and contextual features while simplifying all the remaining components. We conduct experiments on four well-studied benchmark datasets across tasks of natural language inference, paraphrase identification and answer selection. The performance of our model is on par with the state-of-the-art on all datasets with much fewer parameters and the inference speed is at least 6 times faster compared with similarly performed ones.
This paper proposes a hybrid neural network (HNN) model for commonsense reasoning. An HNN consists of two component models, a masked language model and a semantic similarity model, which share a BERT-based contextual encoder but use different model-specific input and output layers. HNN obtains new state-of-the-art results on three classic commonsense reasoning tasks, pushing the WNLI benchmark to 89%, the Winograd Schema Challenge (WSC) benchmark to 75.1%, and the PDP60 benchmark to 90.0%. An ablation study shows that language models and semantic similarity models are complementary approaches to commonsense reasoning, and HNN effectively combines the strengths of both. The code and pre-trained models will be publicly available at https://github.com/namisan/mt-dnn.
Machine learning algorithms are often vulnerable to adversarial examples that have imperceptible alterations from the original counterparts but can fool the state-of-the-art models. It is helpful to evaluate or even improve the robustness of these models by exposing the maliciously crafted adversarial examples. In this paper, we present TextFooler, a simple but strong baseline to generate natural adversarial text. By applying it to two fundamental natural language tasks, text classification and textual entailment, we successfully attacked three target models, including the powerful pre-trained BERT, and the widely used convolutional and recurrent neural networks. We demonstrate the advantages of this framework in three ways: (1) effective---it outperforms state-of-the-art attacks in terms of success rate and perturbation rate, (2) utility-preserving---it preserves semantic content and grammaticality, and remains correctly classified by humans, and (3) efficient---it generates adversarial text with computational complexity linear to the text length. *The code, pre-trained target models, and test examples are available at https://github.com/jind11/TextFooler.
Recently, pre-trained models have achieved state-of-the-art results in various language understanding tasks, which indicates that pre-training on large-scale corpora may play a crucial role in natural language processing. Current pre-training procedures usually focus on training the model with several simple tasks to grasp the co-occurrence of words or sentences. However, besides co-occurring, there exists other valuable lexical, syntactic and semantic information in training corpora, such as named entity, semantic closeness and discourse relations. In order to extract to the fullest extent, the lexical, syntactic and semantic information from training corpora, we propose a continual pre-training framework named ERNIE 2.0 which builds and learns incrementally pre-training tasks through constant multi-task learning. Experimental results demonstrate that ERNIE 2.0 outperforms BERT and XLNet on 16 tasks including English tasks on GLUE benchmarks and several common tasks in Chinese. The source codes and pre-trained models have been released at https://github.com/PaddlePaddle/ERNIE.
We present LINSPECTOR WEB, an open source multilingual inspector to analyze word representations. Our system provides researchers working in low-resource settings with an easily accessible web based probing tool to gain quick insights into their word embeddings especially outside of the English language. To do this we employ 16 simple linguistic probing tasks such as gender, case marking, and tense for a diverse set of 28 languages. We support probing of static word embeddings along with pretrained AllenNLP models that are commonly used for NLP downstream tasks such as named entity recognition, natural language inference and dependency parsing. The results are visualized in a polar chart and also provided as a table. LINSPECTOR WEB is available as an offline tool or at https://linspector.ukp.informatik.tu-darmstadt.de.
Language model pretraining has led to significant performance gains but careful comparison between different approaches is challenging. Training is computationally expensive, often done on private datasets of different sizes, and, as we will show, hyperparameter choices have significant impact on the final results. We present a replication study of BERT pretraining (Devlin et al., 2019) that carefully measures the impact of many key hyperparameters and training data size. We find that BERT was significantly undertrained, and can match or exceed the performance of every model published after it. Our best model achieves state-of-the-art results on GLUE, RACE and SQuAD. These results highlight the importance of previously overlooked design choices, and raise questions about the source of recently reported improvements. We release our models and code.
This paper presents the submissions by Team Dr.Quad to the ACL-BioNLP 2019 shared task on Textual Inference and Question Entailment in the Medical Domain. Our system is based on the prior work Liu et al. (2019) which uses a multi-task objective function for textual entailment. In this work, we explore different strategies for generalizing state-of-the-art language understanding models to the specialized medical domain. Our results on the shared task demonstrate that incorporating domain knowledge through data augmentation is a powerful strategy for addressing challenges posed by specialized domains such as medicine.
We present SpanBERT, a pre-training method that is designed to better represent and predict spans of text. Our approach extends BERT by (1) masking contiguous random spans, rather than random tokens, and (2) training the span boundary representations to predict the entire content of the masked span, without relying on the individual token representations within it. SpanBERT consistently outperforms BERT and our better-tuned baselines, with substantial gains on span selection tasks such as question answering and coreference resolution. In particular, with the same training data and model size as BERT-large, our single model obtains 94.6% and 88.7% F1 on SQuAD 1.1 and 2.0, respectively. We also achieve a new state of the art on the OntoNotes coreference resolution task (79.6\% F1), strong performance on the TACRED relation extraction benchmark, and even show gains on GLUE.