Retrieval-Augmented Generative Question Answering for Event Argument Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4649 - 4666 December 7-11, 2022 ©2022 Association for Computational Linguistics Retrieval-Augmented Generative Question Answering for Event Argument Extraction Xinya Du Department of Computer Science The University of Texas at Dallas xinya.du@utdallas.edu Heng Ji Department of Computer Science University of Illinois Urbana-Champaign hengji@illinois.edu Abstract Event argument extraction has long been stud- ied as a sequential prediction problem with extractive-based methods, tackling each ar- gument in isolation. Although recent work proposes generation-based methods to capture cross-argument dependency, they require gener- ating and post-processing a complicated target sequence (template). Motivated by these obser- vations and recent pretrained language models’ capabilities of learning from demonstrations. We propose a retrieval-augmented generative QA model (R-GQA) for event argument ex- traction. It retrieves the most similar QA pair and augments it as prompt to the current ex- ample’s context, then decodes the arguments as answers. Our approach outperforms sub- stantially prior methods across various settings (i.e. fully supervised, domain transfer, and few- shot learning). Finally, we propose a clustering- based sampling strategy (JointEnc) and conduct a thorough analysis of how different strategies influence the few-shot learning performance.1 1 Introduction Many documents report sequences of events corre- sponding to common situations in the real world. Arguments of different roles provide fine-grained understanding of the event (e.g. INDIVIDUALS, OR- GANIZATIONS, LOCATIONS) and also influence the determination of the event type (Grishman, 2019). As compared to detecting the trigger (usually verbs) of an event, extracting arguments involve recogniz- ing mention spans (consisting of multiple words) of various roles across sentences (Jurafsky and Martin, 2018). We list an example in Figure 1, given the context and the event type (nomination), all argu- ments for the three roles (i.e PERSON, POSITION, AGENT) should be extracted. 1The implementations will be released at https:// github.com/xinyadu/RGQA. Context: One of those difficult judges [John M.] is nominated (Type: nomination) by Adam to be [chief justice] in 2000…. ✅ Context: [Greg L.] was elected (Type: Elect) by Randy as [mayor of Columbus] in 1999…. Role Question Answers/ Extractions Person who is the person nominated? John M. Postion what position is the person nominated for? chief justice Agent who is the norminating agent? Adam Role Question Answers/ Extractions Person who is the person elected? Greg L. Postion what position is the person elected for? mayor of Columbus Agent who is the electing agent? Randy ❌ Context: [John N.] borrowed (Type: Transfer-Money) a large amount of cash to to buy shares in 2000 …. Role Question Answers/ Extractions Recipient Who is recipient agent? John N. Giver Who is the donating agent? N/A … … … Retrieved Demos (from Training) Current Example Figure 1: Current/test example’s context and question for each role have great similarities to the retrieved demonstrations (context and QA pairs). To overcome the error propagation of extractive models (Li et al., 2013; Du and Cardie, 2020b) and efficiently capture the cross-role dependencies, end-to-end template generation-based information extraction approaches (Li et al., 2021; Huang …

Content truncated to 500 words