SRLGRN: Semantic Role Labeling Graph Reasoning Network

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 8881–8891, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 8881 SRLGRN: Semantic Role Labeling Graph Reasoning Network Chen Zheng Michigan State University zhengc12@msu.edu Parisa Kordjamshidi Michigan State University kordjams@msu.edu Abstract This work deals with the challenge of learn- ing and reasoning over multi-hop question an- swering (QA). We propose a graph reasoning network based on the semantic structure of the sentences to learn cross paragraph reason- ing paths and find the supporting facts and the answer jointly. The proposed graph is a heterogeneous document-level graph that con- tains nodes of type sentence (question, title, and other sentences), and semantic role label- ing sub-graphs per sentence that contain argu- ments as nodes and predicates as edges. In- corporating the argument types, the argument phrases, and the semantics of the edges origi- nated from SRL predicates into the graph en- coder helps in finding and also the explainabil- ity of the reasoning paths. Our proposed ap- proach shows competitive performance on the HotpotQA distractor setting benchmark com- pared to the recent state-of-the-art models. 1 Introduction Understanding and reasoning over natural language plays a significant role in artificial intelligence tasks such as Machine Reading Comprehension (MRC) and Question Answering (QA). Several QA tasks have been proposed in recent years to evaluate the language understanding capabilities of machines (Rajpurkar et al., 2016; Joshi et al., 2017; Dunn et al., 2017). These tasks are single-hop QA tasks and consider answering a question given only one single paragraph. Many existing neural models rely on learning context and type-matching heuris- tics (Weissenborn et al., 2017). Those rarely build reasoning modules but achieve promising perfor- mance on single-hop QA tasks. The main reason is that these single-hop QA tasks are lacking a re- alistic evaluation of reasoning capabilities because they do not require complex reasoning. Recently multi-hop QA tasks, such as HotpotQA (Yang et al., 2018) and WikiHop (Welbl et al., Question 430: What team did the recipient of the 2007 Brownlow Medal play for? Paragraph 1: Title: "2007 Brownlow Medal" 0. “The 2007 Brownlow Medal was the 80th year the award … (AFL) home and away season." 1. “Jimmy Bartel won the medal by polling twenty-nine votes ..." Answer: Geelong Football Club Support fact: ["2007 Brownlow Medal", 1], ["Jimmy Bartel", 0] Paragraph 2: Title: "Jimmy Bartel" 0: “James Ross Bartel (born 4 December 1983) is a former Australian rules footballer who played for the Geelong Football Club in the …" 1: "A utility, 1.87 m tall and weighing 86 kg , Bartel is able …" Paragraph 10: Title: "2005 Brownlow Medal" 0: "The 2005 Brownlow Medal was the 78th year the award …" 1: "Ben Cousins of the West Coast Eagles won the medal …" � Figure 1: An example of HotpotQA data. 2018), have been proposed to assess multi-hop rea- soning ability. HotpotQA task provides annotations to evaluate document level question answering and finding supporting facts. Providing supervision for supporting facts improves explainabilty …

Content truncated to 500 words