Interpretable Natural Language Processing with Applications to Information Extraction

Abstract

Interpretability is very important for many NLP applications. Many such applications (e.g., information extraction, sentiment analysis) are applied to important decision-making areas like government policy, financial, law, and others. In these scenarios, machines must explain the information produced, if they are to be deployed in the real world. In this dissertation, we present some approaches for information extraction that mitigates the tension between generalization and explainability by jointly training for the two goals. First we investigate an approach uses an encoder-decoder architecture, which jointly trains a classifier for information extraction, and a rule decoder that generates syntactico-semantic rules that explain the decisions of the classifier. We evaluate the proposed approach on two different information extraction tasks and show that the decoder generates interpretable rules that serve as accurate explanations for the classifier's decisions, and, importantly, that the joint training generally improves the performance of the classifier. 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. Second, we investigate another 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 relation 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 relation extraction 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 great add-on to the manual rules and bring the rule-based system much closer to the neural models. Third, we also explore the usages of the model outputs in two ways: 1. Convert them to rules to bring global explanations to this approach; and 2. Use them for bootstrapping when we do not have enough data. Our globally-explainable models approach the performance of neural ones within a reasonable gap.