Structural analysis methods (e.g., probing and feature attribution) are increasingly important tools for neural network analysis. We propose a new structural analysis method grounded in a formal theory of \textit{causal abstraction} that provides rich characterizations of model-internal representations and their roles in input/output behavior. In this method, neural representations are aligned with variables in interpretable causal models, and then \textit{interchange interventions} are used to experimentally verify that the neural representations have the causal properties of their aligned variables. We apply this method in a case study to analyze neural models trained on Multiply Quantified Natural Language Inference (MQNLI) corpus, a highly complex NLI dataset that was constructed with a tree-structured natural logic causal model. We discover that a BERT-based model with state-of-the-art performance successfully realizes the approximate causal structure of the natural logic causal model, whereas a simpler baseline model fails to show any such structure, demonstrating that neural representations encode the compositional structure of MQNLI examples.
We introduce Dynaboard, an evaluation-as-a-service framework for hosting benchmarks and conducting holistic model comparison, integrated with the Dynabench platform. Our platform evaluates NLP models directly instead of relying on self-reported metrics or predictions on a single dataset. Under this paradigm, models are submitted to be evaluated in the cloud, circumventing the issues of reproducibility, accessibility, and backwards compatibility that often hinder benchmarking in NLP. This allows users to interact with uploaded models in real time to assess their quality, and permits the collection of additional metrics such as memory use, throughput, and robustness, which -- despite their importance to practitioners -- have traditionally been absent from leaderboards. On each task, models are ranked according to the Dynascore, a novel utility-based aggregation of these statistics, which users can customize to better reflect their preferences, placing more/less weight on a particular axis of evaluation or dataset. As state-of-the-art NLP models push the limits of traditional benchmarks, Dynaboard offers a standardized solution for a more diverse and comprehensive evaluation of model quality.
Cryptic crosswords, the dominant English-language crossword variety in the United Kingdom, can be solved by expert humans using flexible, creative intelligence and knowledge of language. Cryptic clues read like fluent natural language, but they are adversarially composed of two parts: a definition and a wordplay cipher requiring sub-word or character-level manipulations. As such, they are a promising target for evaluating and advancing NLP systems that seek to process language in more creative, human-like ways. We present a dataset of cryptic crossword clues from a major newspaper that can be used as a benchmark and train a sequence-to-sequence model to solve them. We also develop related benchmarks that can guide development of approaches to this challenging task. We show that performance can be substantially improved using a novel curriculum learning approach in which the model is pre-trained on related tasks involving, e.g, unscrambling words, before it is trained to solve cryptics. However, even this curricular approach does not generalize to novel clue types in the way that humans can, and so cryptic crosswords remain a challenge for NLP systems and a potential source of future innovation.
There is growing evidence that pretrained language models improve task-specific fine-tuning not just for the languages seen in pretraining, but also for new languages and even non-linguistic data. What is the nature of this surprising cross-domain transfer? We offer a partial answer via a systematic exploration of how much transfer occurs when models are denied any information about word identity via random scrambling. In four classification tasks and two sequence labeling tasks, we evaluate baseline models, LSTMs using GloVe embeddings, and BERT. We find that only BERT shows high rates of transfer into our scrambled domains, and for classification but not sequence labeling tasks. Our analyses seek to explain why transfer succeeds for some tasks but not others, to isolate the separate contributions of pretraining versus fine-tuning, and to quantify the role of word frequency. These findings help explain where and why cross-domain transfer occurs, which can guide future studies and practical fine-tuning efforts.
Images have become an integral part of online media. This has enhanced self-expression and the dissemination of knowledge, but it poses serious accessibility challenges. Adequate textual descriptions are rare. Captions are more abundant, but they do not consistently provide the needed descriptive details, and systems trained on such texts inherit these shortcomings. To address this, we introduce the publicly available Wikipedia-based corpus Concadia, which consists of 96,918 images with corresponding English-language descriptions, captions, and surrounding context. We use Concadia to further characterize the commonalities and differences between descriptions and captions, and this leads us to the hypothesis that captions, while not substitutes for descriptions, can provide a useful signal for creating effective descriptions. We substantiate this hypothesis by showing that image captioning systems trained on Concadia benefit from having caption embeddings as part of their inputs. These experiments also begin to show how Concadia can be a powerful tool in addressing the underlying accessibility issues posed by image data.
We introduce Dynabench, an open-source platform for dynamic dataset creation and model benchmarking. Dynabench runs in a web browser and supports human-and-model-in-the-loop dataset creation: annotators seek to create examples that a target model will misclassify, but that another person will not. In this paper, we argue that Dynabench addresses a critical need in our community: contemporary models quickly achieve outstanding performance on benchmark tasks but nonetheless fail on simple challenge examples and falter in real-world scenarios. With Dynabench, dataset creation, model development, and model assessment can directly inform each other, leading to more robust and informative benchmarks. We report on four initial NLP tasks, illustrating these concepts and highlighting the promise of the platform, and address potential objections to dynamic benchmarking as a new standard for the field.
Multi-hop reasoning (i.e., reasoning across two or more documents) at scale is a key step toward NLP models that can exhibit broad world knowledge by leveraging large collections of documents. We propose Baleen, a system that improves the robustness and scalability of multi-hop reasoning over current approaches. Baleen introduces a per-hop condensed retrieval pipeline to mitigate the size of the search space, a focused late interaction retriever (FliBERT) that can model complex multi-hop queries, and a weak supervision strategy, latent hop ordering, to learn from limited signal about which documents to retrieve for a query. We evaluate Baleen on the new many-hop claim verification dataset HoVer, establishing state-of-the-art performance.
We introduce DynaSent ('Dynamic Sentiment'), a new English-language benchmark task for ternary (positive/negative/neutral) sentiment analysis. DynaSent combines naturally occurring sentences with sentences created using the open-source Dynabench Platform, which facilities human-and-model-in-the-loop dataset creation. DynaSent has a total of 121,634 sentences, each validated by five crowdworkers, and its development and test splits are designed to produce chance performance for even the best models we have been able to develop; when future models solve this task, we will use them to create DynaSent version 2, continuing the dynamic evolution of this benchmark. Here, we report on the dataset creation effort, focusing on the steps we took to increase quality and reduce artifacts. We also present evidence that DynaSent's Neutral category is more coherent than the comparable category in other benchmarks, and we motivate training models from scratch for each round over successive fine-tuning.
Systems for Open-Domain Question Answering (OpenQA) generally depend on a retriever for finding candidate passages in a large corpus and a reader for extracting answers from those passages. In much recent work, the retriever is a learned component that uses coarse-grained vector representations of questions and passages. We argue that this modeling choice is insufficiently expressive for dealing with the complexity of natural language questions. To address this, we define ColBERT-QA, which adapts the scalable neural retrieval model ColBERT to OpenQA. ColBERT creates fine-grained interactions between questions and passages. We propose a weak supervision strategy that iteratively uses ColBERT to create its own training data. This greatly improves OpenQA retrieval on both Natural Questions and TriviaQA, and the resulting end-to-end OpenQA system attains state-of-the-art performance on both of those datasets.
Crime reporting is a prevalent form of journalism with the power to shape public perceptions and social policies. How does the language of these reports act on readers? We seek to address this question with the SuspectGuilt Corpus of annotated crime stories from English-language newspapers in the U.S. For SuspectGuilt, annotators read short crime articles and provided text-level ratings concerning the guilt of the main suspect as well as span-level annotations indicating which parts of the story they felt most influenced their ratings. SuspectGuilt thus provides a rich picture of how linguistic choices affect subjective guilt judgments. In addition, we use SuspectGuilt to train and assess predictive models, and show that these models benefit from genre pretraining and joint supervision from the text-level ratings and span-level annotations. Such models might be used as tools for understanding the societal effects of crime reporting.