Improving Zero-shot Relation Extraction via Rule-based and Prompt-based Methods

Abstract

Relation extraction (RE) is a foundational task in information extraction that aims to identify semantic relations between entities from unstructured text. While supervised relation extraction models have considerably advanced the state-of-the-art, they often perform poorly in low-resource settings. Zero-shot RE is vital when annotations are not available either due to costs or time constraints. As a result, zero-shot RE has garnered interest in the research community. In this dissertation, we propose rule-based and prompt-based methods to advance zero-shot RE. First, we introduce a new rule learning method that given a seed rule, learns many rules with a synonymous meaning. Rule-based approaches have the advantage of interpretability. Furthermore, the interpretability provided by rules is actionable. On the other hand, rule-based approaches lack the generalization power of deep learning systems. In this work, we aim to marry the advantages of the two directions. To that end, we propose an extension to Harris (1954)’s distributional hypothesis for rules. In particular, we propose to measure the similarity between pairs of slots (i.e., the set of concepts matched by a rule) using contextualized embeddings instead of lexical overlap. Empirical results demonstrate that this new similarity method yields a better implementation of the distributional hypothesis. Next, we develop a zero-shot RE method that formulates RE as a textual entailment task. Our method automatically generates templates using our extension of the distributional hypothesis to rules. These templates verbalize relation types, and are fed as hypotheses to an off-the-shelf entailment engine for classification. Our method achieves state-of-the-art performance for zero-shot TACRED, a popular RE benchmark. Finally, we introduce an effective prompt-based method for RE. With the arrival of large language models, many approaches have been proposed for RE, but they are often ineffective or require an accompanying masked language model or complex post-prompt processing. In this work, we propose a high-performing prompt-based method for RE that does not require any additional resources. Our experiments on four main RE datasets showed that our method outperforms previous state-of-the-art by a large margin.