Research On External Knowledge Integrated Reasoning For Commonsense Question Answering

Commonsense Question Answering(CQA)is an natural language processing task that involves answering questions through commonsense reasoning.Since the key knowledge is not given in question,thus,utilizing external knowledge to support the model’s reasoning is a common solution.However,previous works have faced several challenges,including noise interference,insufficient utilization of knowledge,and limited domain coverage of knowledge sources.Moreover,generative CQA requires further exploration.To address these challenges,this paper researches the following topics:(1)Multi-source knowledge fusion for discriminative CQAThe discriminative CQA task requires models to evaluate the plausibility of candidate answers to the given question.Previous works typically rely on acquiring "bridge-type"knowledge that connects the question and answer.However,these methods are impeded by noisy items in the knowledge source,and the utilization of multiple knowledge sources is inadequate.To address these issues,this paper proposes a multi-source knowledge fusion(MSKF)method that introduces multiple knowledge sources and multi-channel fused reasoning.Moreover,the greedy selection strategy is used to filter out noisy knowledge.(2)Knowledge enhancement and plausibility ranking for generative CQAThe generative CQA task requires models to produce a list of answers to the given question.In this task,Knowledge retrieval presents two challenges.Firstly,the candidate answers are not given,thus only the question can be used in retrieval.Secondly,the acquired knowledge should support the generation of diverse answers.Moreover,the sampled answers from models often contain many duplicates or errors,which negatively impact performance.To address these issues,this paper proposes a knowledge enhancement and plausibility ranking(KEPR)method.KEPR is based on a generate-then-rank architecture and incorporates"net-type" knowledge that contributes to the generation of various answers.(3)Mining knowledge within large-scale language models for CQAExternal knowledge retrieval is crucial for a CQA system.However,manually constructed knowledge sources have limited domains and sparse content,which makes them inadequate for supporting diverse commonsense questions.To address this issue,this paper proposes a unified fact obtaining(UFO)method that elicits intrinsic knowledge from language models.UFO transforms pre-trained language models into knowledge sources and instructs them to produce relevant facts through specific few-shot prompts.These prompts contain examples that cover various question expressions and different aspects of commonsense,making them highly generalizable.The knowledge produced by UFO can support commonsense reasoning in multiple domains.This paper aims to explore the retrieval and utilization of knowledge for both discriminative and generative CQA tasks.The proposed MSKF and KEPR methods fully utilize external knowledge to enhance discrimination accuracy and increase generation diversity.Additionally,the proposed UFO method leverages intrinsic knowledge from language models to address issues of limited domain and sparse content in manually constructed knowledge sources.The effectiveness of the proposed methods is demonstrated through detailed experimental results on multiple CQA datasets.