The capabilities of today’s natural language processing systems are typically evaluated using large datasets of curated questions and answers. While these are critical benchmarks of progress, they also suffer from weakness due to artificial distributions and incomplete knowledge. Artifacts arising from artificial distributions can overstate language model performance, while incomplete knowledge limits fine-grained analysis. In this work, we introduce a complementary benchmarking approach based on SimPlified Language Activity Traces (SPLAT). SPLATs are corpora of language encodings of activity in some closed domain (we study traces from chess and baseball games in this work). SPLAT datasets use naturally-arising distributions, allow the generation of question-answer pairs at scale, and afford complete knowledge in their closed domains. We show that language models of three different architectures can answer questions about world states using only verb-like encodings of activity. Our approach is extensible to new language models and additional question-answering tasks.

Conversations aimed at determining good recommendations are iterative in nature. People often express their preferences in terms of a critique of the current recommendation (e.g., “It doesn’t look good for a date”), requiring some degree of common sense for a preference to be inferred. In this work, we present a method for transforming a user critique into a positive preference (e.g., “I prefer more romantic”) in order to retrieve reviews pertaining to potentially better recommendations (e.g., “Perfect for a romantic dinner”). We leverage a large neural language model (LM) in a few-shot setting to perform critique-to-preference transformation, and we test two methods for retrieving recommendations: one that matches embeddings, and another that fine-tunes an LM for the task. We instantiate this approach in the restaurant domain and evaluate it using a new dataset of restaurant critiques. In an ablation study, we show that utilizing critique-to-preference transformation improves recommendations, and that there are at least three general cases that explain this improved performance.

Recent advances in commonsense reasoning depend on large-scale human-annotated training sets to achieve peak performance. However, manual curation of training sets is expensive and has been shown to introduce annotation artifacts that neural models can readily exploit and overfit to. We propose a novel generative data augmentation technique, G-DAUGˆC, that aims to achieve more accurate and robust learning in a low-resource setting. Our approach generates synthetic examples using pretrained language models and selects the most informative and diverse set of examples for data augmentation. On experiments with multiple commonsense reasoning benchmarks, G-DAUGˆC consistently outperforms existing data augmentation methods based on back-translation, establishing a new state-of-the-art on WinoGrande, CODAH, and CommonsenseQA, as well as enhances out-of-distribution generalization, proving to be robust against adversaries or perturbations. Our analysis demonstrates that G-DAUGˆC produces a diverse set of fluent training examples, and that its selection and training approaches are important for performance.

Neural Network Language Models (NNLMs) generate probability distributions by applying a softmax function to a distance metric formed by taking the dot product of a prediction vector with all word vectors in a high-dimensional embedding space. The dot-product distance metric forms part of the inductive bias of NNLMs. Although NNLMs optimize well with this inductive bias, we show that this results in a sub-optimal ordering of the embedding space that structurally impoverishes some words at the expense of others when assigning probability. We present numerical, theoretical and empirical analyses which show that words on the interior of the convex hull in the embedding space have their probability bounded by the probabilities of the words on the hull.

Representation learning is a critical ingredient for natural language processing systems. Recent Transformer language models like BERT learn powerful textual representations, but these models are targeted towards token- and sentence-level training objectives and do not leverage information on inter-document relatedness, which limits their document-level representation power. For applications on scientific documents, such as classification and recommendation, accurate embeddings of documents are a necessity. We propose SPECTER, a new method to generate document-level embedding of scientific papers based on pretraining a Transformer language model on a powerful signal of document-level relatedness: the citation graph. Unlike existing pretrained language models, Specter can be easily applied to downstream applications without task-specific fine-tuning. Additionally, to encourage further research on document-level models, we introduce SciDocs, a new evaluation benchmark consisting of seven document-level tasks ranging from citation prediction, to document classification and recommendation. We show that Specter outperforms a variety of competitive baselines on the benchmark.

Language models pretrained on text from a wide variety of sources form the foundation of today’s NLP. In light of the success of these broad-coverage models, we investigate whether it is still helpful to tailor a pretrained model to the domain of a target task. We present a study across four domains (biomedical and computer science publications, news, and reviews) and eight classification tasks, showing that a second phase of pretraining in-domain (domain-adaptive pretraining) leads to performance gains, under both high- and low-resource settings. Moreover, adapting to the task’s unlabeled data (task-adaptive pretraining) improves performance even after domain-adaptive pretraining. Finally, we show that adapting to a task corpus augmented using simple data selection strategies is an effective alternative, especially when resources for domain-adaptive pretraining might be unavailable. Overall, we consistently find that multi-phase adaptive pretraining offers large gains in task performance.

Commonsense reasoning is a critical AI capability, but it is difficult to construct challenging datasets that test common sense. Recent neural question answering systems, based on large pre-trained models of language, have already achieved near-human-level performance on commonsense knowledge benchmarks. These systems do not possess human-level common sense, but are able to exploit limitations of the datasets to achieve human-level scores. We introduce the CODAH dataset, an adversarially-constructed evaluation dataset for testing common sense. CODAH forms a challenging extension to the recently-proposed SWAG dataset, which tests commonsense knowledge using sentence-completion questions that describe situations observed in video. To produce a more difficult dataset, we introduce a novel procedure for question acquisition in which workers author questions designed to target weaknesses of state-of-the-art neural question answering systems. Workers are rewarded for submissions that models fail to answer correctly both before and after fine-tuning (in cross-validation). We create 2.8k questions via this procedure and evaluate the performance of multiple state-of-the-art question answering systems on our dataset. We observe a significant gap between human performance, which is 95.3%, and the performance of the best baseline accuracy of 65.3% by the OpenAI GPT model.

Recurrent neural network language models (RNNLM) form a valuable foundation for many NLP systems, but training the models can be computationally expensive, and may take days to train on a large corpus. We explore a technique that uses large corpus n-gram statistics as a regularizer for training a neural network LM on a smaller corpus. In experiments with the Billion-Word and Wikitext corpora, we show that the technique is effective, and more time-efficient than simply training on a larger sequential corpus. We also introduce new strategies for selecting the most informative n-grams, and show that these boost efficiency.

We introduce a novel topic modeling approach based on constructing a semantic set cover for clusters of similar documents. Specifically, our approach first clusters documents using their Tf-Idf representation, and then covers each cluster with a set of topic words based on semantic similarity, defined in terms of a word embedding. Computing a topic cover amounts to solving a minimum set cover problem. Our evaluation compares our topic modeling approach to Latent Dirichlet Allocation (LDA) on three metrics: 1) qualitative topic match, measured using evaluations by Amazon Mechanical Turk (MTurk) workers, 2) performance on classification tasks using each topic model as a sparse feature representation, and 3) topic coherence. We find that qualitative judgments significantly favor our approach, the method outperforms LDA on topic coherence, and is comparable to LDA on document classification tasks.

Recurrent neural network language models (RNNLMs) are the current standard-bearer for statistical language modeling. However, RNNLMs only estimate probabilities for complete sequences of text, whereas some applications require context-independent phrase probabilities instead. In this paper, we study how to compute an RNNLM’s em marginal probability: the probability that the model assigns to a short sequence of text when the preceding context is not known. We introduce a simple method of altering the RNNLM training to make the model more accurate at marginal estimation. Our experiments demonstrate that the technique is effective compared to baselines including the traditional RNNLM probability and an importance sampling approach. Finally, we show how we can use the marginal estimation to improve an RNNLM by training the marginals to match n-gram probabilities from a larger corpus.

Intelligent systems require common sense, but automatically extracting this knowledge from text can be difficult. We propose and assess methods for extracting one type of commonsense knowledge, object-property comparisons, from pre-trained embeddings. In experiments, we show that our approach exceeds the accuracy of previous work but requires substantially less hand-annotated knowledge. Further, we show that an active learning approach that synthesizes common-sense queries can boost accuracy.

We propose an unsupervised importance sampling approach to selecting training data for recurrent neural network (RNNs) language models. To increase the information content of the training set, our approach preferentially samples high perplexity sentences, as determined by an easily queryable n-gram language model. We experimentally evaluate the heldout perplexity of models trained with our various importance sampling distributions. We show that language models trained on data sampled using our proposed approach outperform models trained over randomly sampled subsets of both the Billion Word (Chelba et al., 2014 Wikitext-103 benchmark corpora (Merity et al., 2016).

We describe a deployed scalable system for organizing published scientific literature into a heterogeneous graph to facilitate algorithmic manipulation and discovery. The resulting literature graph consists of more than 280M nodes, representing papers, authors, entities and various interactions between them (e.g., authorships, citations, entity mentions). We reduce literature graph construction into familiar NLP tasks (e.g., entity extraction and linking), point out research challenges due to differences from standard formulations of these tasks, and report empirical results for each task. The methods described in this paper are used to enable semantic features in www.semanticscholar.org.

Many Natural Language Processing (NLP) models rely on distributed vector representations of words. Because the process of training word vectors can require large amounts of data and computation, NLP researchers and practitioners often utilize pre-trained embeddings downloaded from the Web. However, finding the best embeddings for a given task is difficult, and can be computationally prohibitive. We present a framework, called VecShare, that makes it easy to share and retrieve word embeddings on the Web. The framework leverages a public data-sharing infrastructure to host embedding sets, and provides automated mechanisms for retrieving the embeddings most similar to a given corpus. We perform an experimental evaluation of VecShare’s similarity strategies, and show that they are effective at efficiently retrieving embeddings that boost accuracy in a document classification task. Finally, we provide an open-source Python library for using the VecShare framework.