2020/03/31 by Marco Valentino, Valentino, Marco, Mokanarangan Thayaparan +3
Computer Science · Decision Sciences · #Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Information Retrieval (cs.IR) #Scientific Computing and Data Management #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2004.00061
openalex publication_date 2020/03/31 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
This paper presents a novel framework for reconstructing multi-hop\nexplanations in science Question Answering (QA). While existing approaches for\nmulti-hop reasoning build explanations considering each question in isolation,\nwe propose a method to leverage explanatory patterns emerging in a corpus of\nscientific explanations. Specifically, the framework ranks a set of atomic\nfacts by integrating lexical relevance with the notion of unification power,\nestimated analysing explanations for similar questions in the corpus.\n An extensive evaluation is performed on the Worldtree corpus, integrating\nk-NN clustering and Information Retrieval (IR) techniques. We present the\nfollowing conclusions: (1) The proposed method achieves results competitive\nwith Transformers, yet being orders of magnitude faster, a feature that makes\nit scalable to large explanatory corpora (2) The unification-based mechanism\nhas a key role in reducing semantic drift, contributing to the reconstruction\nof many hops explanations (6 or more facts) and the ranking of complex\ninference facts (+12.0 Mean Average Precision) (3) Crucially, the constructed\nexplanations can support downstream QA models, improving the accuracy of BERT\nby up to 10% overall.\n