Recent work finds modern natural language processing (NLP) models relying on spurious features for prediction. Mitigating such effects is thus important. Despite this need, there is no quantitative measure to evaluate or compare the effects of different forms of spurious features in NLP. We address this gap in the literature by quantifying model sensitivity to spurious features with a causal estimand, dubbed CENT, which draws on the concept of average treatment effect from the causality literature. By conducting simulations with four prominent NLP models -- TextRNN, BERT, RoBERTa and XLNet -- we rank the models against their sensitivity to artificial injections of eight spurious features. We further hypothesize and validate that models that are more sensitive to a spurious feature will be less robust against perturbations with this feature during inference. Conversely, data augmentation with this feature improves robustness to similar perturbations. We find statistically significant inverse correlations between sensitivity and robustness, providing empirical support for our hypothesis.
Questions of fairness, robustness, and transparency are paramount to address before deploying NLP systems. Central to these concerns is the question of reliability: Can NLP systems reliably treat different demographics fairly and function correctly in diverse and noisy environments? To address this, we argue for the need for reliability testing and contextualize it among existing work on improving accountability. We show how adversarial attacks can be reframed for this goal, via a framework for developing reliability tests. We argue that reliability testing -- with an emphasis on interdisciplinary collaboration -- will enable rigorous and targeted testing, and aid in the enactment and enforcement of industry standards.
Neural models for automated fact verification have achieved promising results thanks to the availability of large, human-annotated datasets. However, for each new domain that requires fact verification, creating a dataset by manually writing claims and linking them to their supporting evidence is expensive. We develop QACG, a framework for training a robust fact verification model by using automatically generated claims that can be supported, refuted, or unverifiable from evidence from Wikipedia. QACG generates question-answer pairs from the evidence and then converts them into different types of claims. Experiments on the FEVER dataset show that our QACG framework significantly reduces the demand for human-annotated training data. In a zero-shot scenario, QACG improves a RoBERTa model's F1 from 50% to 77%, equivalent in performance to 2K+ manually-curated examples. Our QACG code is publicly available.
Multiparty Dialogue Machine Reading Comprehension (MRC) differs from traditional MRC as models must handle the complex dialogue discourse structure, previously unconsidered in traditional MRC. To fully exploit such discourse structure in multiparty dialogue, we present a discourse-aware dialogue graph neural network, DADgraph, which explicitly constructs the dialogue graph using discourse dependency links and discourse relations. To validate our model, we perform experiments on the Molweni corpus, a large-scale MRC dataset built over multiparty dialogue annotated with discourse structure. Experiments on Molweni show that our discourse-aware model achieves statistically significant improvements compared against strong neural network MRC baselines.
The collection and annotation of task-oriented conversational data is a costly and time-consuming manner. Many augmentation techniques have been proposed to improve the performance of state-of-the-art (SOTA) systems in new domains that lack the necessary amount of data for training. However, these augmentation techniques (e.g. paraphrasing) also require some mediocre amount of data since they use learning-based approaches. This makes using SOTA systems in emerging low-resource domains infeasible. We, to tackle this problem, introduce a framework, that creates synthetic task-oriented dialogues in a fully automatic manner, which operates with input sizes of as small as a few dialogues. Our framework uses the simple idea that each turn-pair in a task-oriented dialogue has a certain function and exploits this idea to mix them creating new dialogues. We evaluate our framework within a low-resource setting by integrating it with a SOTA model TRADE in the dialogue state tracking task and observe significant improvements in the fine-tuning scenarios in several domains. We conclude that this end-to-end dialogue augmentation framework can be a crucial tool for natural language understanding performance in emerging task-oriented dialogue domains.
Recent question generation (QG) approaches often utilize the sequence-to-sequence framework (Seq2Seq) to optimize the log-likelihood of ground-truth questions using teacher forcing. However, this training objective is inconsistent with actual question quality, which is often reflected by certain global properties such as whether the question can be answered by the document. As such, we directly optimize for QG-specific objectives via reinforcement learning to improve question quality. We design three different rewards that target to improve the fluency, relevance, and answerability of generated questions. We conduct both automatic and human evaluations in addition to a thorough analysis to explore the effect of each QG-specific reward. We find that optimizing question-specific rewards generally leads to better performance in automatic evaluation metrics. However, only the rewards that correlate well with human judgement (e.g., relevance) lead to real improvement in question quality. Optimizing for the others, especially answerability, introduces incorrect bias to the model, resulting in poor question quality. Our code is publicly available at https://github.com/YuxiXie/RL-for-Question-Generation.
Obtaining training data for Multi-hop Question Answering (QA) is extremely time-consuming and resource-intensive. To address this, we propose the problem of \textit{unsupervised} multi-hop QA, assuming that no human-labeled multi-hop question-answer pairs are available. We propose MQA-QG, an unsupervised question answering framework that can generate human-like multi-hop training pairs from both homogeneous and heterogeneous data sources. Our model generates questions by first selecting or generating relevant information from each data source and then integrating the multiple information to form a multi-hop question. We find that we can train a competent multi-hop QA model with only generated data. The F1 gap between the unsupervised and fully-supervised models is less than 20 in both the HotpotQA and the HybridQA dataset. Further experiments reveal that an unsupervised pretraining with the QA data generated by our model would greatly reduce the demand for human-annotated training data for multi-hop QA.
Domain divergence plays a significant role in estimating the performance of a model when applied to new domains. While there is significant literature on divergence measures, choosing an appropriate divergence measures remains difficult for researchers. We address this shortcoming by both surveying the literature and through an empirical study. We contribute a taxonomy of divergence measures consisting of three groups -- Information-theoretic, Geometric, and Higher-order measures -- and identify the relationships between them. We then ground the use of divergence measures in three different application groups -- 1) Data Selection, 2) Learning Representation, and 3) Decisions in the Wild. From this, we identify that Information-theoretic measures are prevalent for 1) and 3), and higher-order measures are common for 2). To further help researchers, we validate these uses empirically through a correlation analysis of performance drops. We consider the current contextual word representations (CWR) to contrast with the older word distribution based representations for this analysis. We find that traditional measures over word distributions still serve as strong baselines, while higher-order measures with CWR are effective.