PeerDA: Data Augmentation via Modeling Peer Relation for Span Identification Tasks

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8681–8699 July 9-14, 2023 ©2023 Association for Computational Linguistics PeerDA: Data Augmentation via Modeling Peer Relation for Span Identification Tasks∗ Weiwen Xu12,† Xin Li2,‡ Yang Deng1 Wai Lam1 Lidong Bing2 1The Chinese University of Hong Kong 2DAMO Academy, Alibaba Group {wwxu,wlam}@se.cuhk.edu.hk ydeng@nus.edu.sg {xinting.lx,l.bing}@alibaba-inc.com Abstract Span identification aims at identifying specific text spans from text input and classifying them into pre-defined categories. Different from pre- vious works that merely leverage the Subordi- nate (SUB) relation (i.e. if a span is an instance of a certain category) to train models, this pa- per for the first time explores the Peer (PR) relation, which indicates that two spans are instances of the same category and share sim- ilar features. Specifically, a novel Peer Data Augmentation (PeerDA) approach is proposed which employs span pairs with the PR relation as the augmentation data for training. PeerDA has two unique advantages: (1) There are a large number of PR span pairs for augment- ing the training data. (2) The augmented data can prevent the trained model from over-fitting the superficial span-category mapping by push- ing the model to leverage the span semantics. Experimental results on ten datasets over four diverse tasks across seven domains demonstrate the effectiveness of PeerDA. Notably, PeerDA achieves state-of-the-art results on six of them.1 1 Introduction Span Identification (SpanID) is a family of Nat- ural Language Processing (NLP) tasks with the goal of detecting specific spans from the input text and further classifying them into pre-defined cat- egories (Papay et al., 2020). It serves as the ini- tial step for complex text analysis by narrowing down the search scopes of important spans, which holds a pivotal position in the field of NLP (Ding et al., 2021; Xu et al., 2021). Recently, different ∗ This work was supported by Alibaba Group through Alibaba Research Intern Program. It was also partially sup- ported by a grant from the Research Grant Council of the Hong Kong Special Administrative Region, China (Project Code: 14200620). † This work was done when Weiwen Xu was an intern at Alibaba DAMO Academy. ‡ Xin Li is the corresponding author. 1Our code and data are available at https://github. com/DAMO-NLP-SG/PeerDA Task Category Span NER PER ORG LOC Hawaii London Hangzhou … … CCE ABSA … SUB PR PR PR Context: Gotta dress up for London fashion week and party in style! Original data SUB Query: Answer: Highlight the parts (if any) related to “LOC”. Details: the name of politically or geographically defined locations such as cities, provinces, etc. London Augmented data PR Query-1: Answer: Highlight the parts (if any) similar to “Hawaii”. London PR Query-2: Answer: Highlight the parts (if any) similar to “Hangzhou”. London (a) Relations in SpanID (b) SpanID in MRC Paradigm SUB SUB Figure 1: (a) Illustrations of Subordinate (SUB) and Peer (PR) relations in SpanID tasks. (b) The constructions of augmented data with PR relations in MRC paradigm. We use …

Content truncated to 500 words