Coarse and Fine Granularity Graph Reasoning for Interpretable Multi-Hop Question Answering

Abstract

Interpretable multi-hop question answering requires step-by-step reasoning over multiple documents and finding scattered supporting facts to answer the question. Prior works have proposed the entity graph method to aggregate the entity information to improve the ability of reasoning. However, the entity graph loses some nonentity information that is also important to understand the semantics. Moreover, the entities distributed in the noisy sentences may mislead the reasoning process. In this paper, we propose the Coarse and Fine Granularity Graph Network (CFGGN), a novel interpretable model that combines both sentence information and entity information to answer the multi-hop questions. The CFGGN consists of a coarse-grain module to perform sentence-level reasoning and a fine-grain module to make entity-level reference. In sentence-level reasoning, the sentence graph is constructed to filter out the noisy sentences and capture the sentence features. In entity-level reference, a dynamic entity graph is used for the entity-level reasoning. We design a fusion module to integrate information of different granularity. To enhance the interpretability of the overall process, we calculate the reasoning score for each step and present the reasoning path from the multiple documents to the final answer. Evaluation on the HotpotQA dataset in the distractor setting shows that our method outperforms the published SOTA entity-based method in five out of six metrics. INDEX TERMS Question answering, multi-hop questions, graph network, interpretable model.

A Survey on Multi-hop Question Answering and Generation

“…There has been a significant research in the recent years to solve the task. A variety of methods model the task as performing inference over static or dynamic graphs to find the reasoning paths [14,15,30,34,40,61,113,128,142,179,181]. A number of works have also attempted to decompose the multi-hop questions into single hop questions or generate follow-up questions based on the retrieved information [14,95,102,138,181].…”

“…), where in the WikiHop dataset we have (𝑖𝑡𝑒𝑚 1 , 𝑝𝑟𝑜𝑝𝑒𝑟𝑡𝑦, ?). Therefore, automatic generation using KBs is argued to result in datasets limited by the incompleteness of entity relations and schema of the KB used [173,179].…”

“…• Yadav et al [171] use RoBERTa [92] for reranking trained to predict the F-1 score of evidence chain. • Zhang et al [179] and 20 concatenate paragraphs with the question and feed it to a BERT followed by a binary classifier and keep N(=3) paragraphs with the highest scores. Qiu et al [113] follows a similar approach but sets a threshold to retrieve variable number of paragraphs.…”

“…Updating embeddings of each token is necessary since the answer might not be an entity. • Zhang et al [179] create two different graphs: a) sentence graph with edges between sentences belonging to the same paragraph and between sentences that share an entity. b) entity graph with edges between entities appearing in the same sentence, between different mentions of same entities and between each entity in the paragraph to each entity in its title.…”

“…Finally, a three-way classifier layer is used to predict the answer type among 'yes', 'no' and the extracted span. Many methods focusing on the first two units use this baseline as the answer prediction unit [14,28,41,113,128,129,179]. • Qi et al [112] make two changes to this baseline: a) Concatenating passages before encoding makes the representations depend on the order while concatenating.…”