TAGPRIME: A Unified Framework for Relational Structure Extraction

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 12917–12932 July 9-14, 2023 ©2023 Association for Computational Linguistics TAGPRIME: A Unified Framework for Relational Structure Extraction I-Hung Hsu∗† Kuan-Hao Huang∗‡ Shuning Zhang⋄ Wenxin Cheng‡ Premkumar Natarajan† Kai-Wei Chang‡ Nanyun Peng‡ †Information Science Institute, University of Southern California ‡Computer Science Department, University of California, Los Angeles ⋄Computer Science and Technology Department, Tsinghua University {ihunghsu, pnataraj}@isi.edu, {khhuang, kwchang, violetpeng}@cs.ucla.edu zhang-sn19@mails.tsinghua.edu.cn, wenxin0319@ucla.edu Abstract Many tasks in natural language processing re- quire the extraction of relationship informa- tion for a given condition, such as event argu- ment extraction, relation extraction, and task- oriented semantic parsing. Recent works usu- ally propose sophisticated models for each task independently and pay less attention to the com- monality of these tasks and to have a unified framework for all the tasks. In this work, we propose to take a unified view of all these tasks and introduce TAGPRIME to address relational structure extraction problems. TAGPRIME is a sequence tagging model that appends prim- ing words about the information of the given condition (such as an event trigger) to the in- put text. With the self-attention mechanism in pre-trained language models, the priming words make the output contextualized repre- sentations contain more information about the given condition, and hence become more suit- able for extracting specific relationships for the condition. Extensive experiments and analyses on three different tasks that cover ten datasets across five different languages demonstrate the generality and effectiveness of TAGPRIME. 1 Introduction There are many tasks in natural language process- ing (NLP) that require extracting relational struc- tures from texts. For example, the event argument extraction task aims to identify event arguments and their corresponding roles for a given event trig- ger (Huang et al., 2022; Wang et al., 2019). In entity relation extraction, the model identifies the tail-entities and head-entities that forms specific relations (Wei et al., 2020; Yu et al., 2020). In task- oriented semantic parsing, the model predicts the slots and their semantic roles for a given intent in an ∗The authors contribute equally. utterance (Tür et al., 2010; Li et al., 2021). These tasks are beneficial to a wide range of applications, such as dialog systems (Liu et al., 2018), question answering (Yasunaga et al., 2021), and narrative generation (Chen et al., 2019a). Prior works usu- ally design models to specifically address each of the tasks (Sun et al., 2019; Miwa and Bansal, 2016; Han et al., 2019; Fu et al., 2019; Zhang et al., 2018). However, less attention is paid to the commonality among these tasks and having a unified framework to deal with them and provide a strong baseline for every task. In this work, we take a unified view of these NLP tasks. We call them relational structure extraction (RSE) tasks and formulate them as a unified task that identifies arguments to a given condition and classifies their relationships. The condition could be a textual span, such as an event trigger …

Content truncated to 500 words