Syntactic Multi-view Learning for Open Information Extraction

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4072 - 4083 December 7-11, 2022 ©2022 Association for Computational Linguistics Syntactic Multi-view Learning for Open Information Extraction Kuicai Dong1,2, Aixin Sun1, Jung-Jae Kim2, Xiaoli Li1,2,3 1 School of Computer Science and Engineering, Nanyang Technological University, Singapore kuicai001@e.ntu.edu.sg, axsun@ntu.edu.sg 2 Institute for Infocomm Research, A*STAR, Singapore 3 A*STAR Centre for Frontier AI Research, Singapore {jjkim, xlli}@i2r.a-star.edu.sg Abstract Open Information Extraction (OpenIE) aims to extract relational tuples from open-domain sentences. Traditional rule-based or statistical models have been developed based on syntac- tic structures of sentences, identified by syn- tactic parsers. However, previous neural Ope- nIE models under-explore the useful syntac- tic information. In this paper, we model both constituency and dependency trees into word- level graphs, and enable neural OpenIE to learn from the syntactic structures. To better fuse het- erogeneous information from both graphs, we adopt multi-view learning to capture multiple relationships from them. Finally, the finetuned constituency and dependency representations are aggregated with sentential semantic repre- sentations for tuple generation. Experiments show that both constituency and dependency information, and the multi-view learning are effective. Our model is publicly available.1 1 Introduction toOpen Information Extraction (OpenIE) aims generate structured tuples from unstruc- tured open-domain text (Yates et al., 2007). The extracted tuples are in the form of ⟨Subject, Relation, Object⟩ for binary relation, and ⟨ARG0, Relation, ARG1, . . . , ARGn⟩ for n- ary relation. It has been a critical NLP task as it is domain-independent and does not rely on pre- defined ontology schema. The structured relational tuples are beneficial to many downstream tasks such as question answering (Khot et al., 2017), knowledge base population (Martínez-Rodríguez et al., 2018; Gashteovski et al., 2020) and word embedding generation (Stanovsky et al., 2015). In general, traditional OpenIE systems are either statistical or rule-based. They extract relational tuples mainly based on certain sentence patterns heuristically defined on syntactic structures. The 1https://github.com/daviddongkc/smile_oie Mary plush room NP PP NP VP VP S NP NP S 'scat likes playing toys in the case poss nsubj xcomp prep dobj amod det pobj Constituency edge Dependency edge root Mary plush room NP PP NP VP VP S NP NP S 'scat likes playing toys in the Mary plush room 'scat likes playing toys in the det nsubj case poss amod dobj xcomp prep pobj root Mary plush room 'scat likes playing toys in the NP NP VP NP PP NP VP VP S S cat nsubj 's case Mary poss likes root playing xcomp plush amod toys dobj room pobj the det in prep Constituency tree arcs Constituency tree Dependency tree arcs Constituency and dependency tree (to be used in Introduction) Constructed constituency graph from arcs Constructed dependency graph from arcs room S‐VP‐S‐VP‐PP‐NP the S‐VP‐S‐VP‐PP‐NP in S‐VP‐S‐VP‐PP toys S‐VP‐S‐VP‐NP plush S‐VP‐S‐VP‐NP playing S‐VP‐S‐VP likes S‐VP cat S‐NP 's S‐NP‐NP Mary S‐NP‐NP Mary plush room 'scat likes playing toys in the NP NP VP NP PP NP VP VP S Constituency tree arcs Constructed …

Content truncated to 500 words