Research On Interpretable Approaches For Answering Multi-hop Questions Over Knowledge Graph

With the rapid development from sensation intelligence to cognitive intelligence in the field of AI,knowledge graph(KG),which serves as the foundation of cognitive intelligence,has attracted significant attention from the industry and academic research community.Despite the impressive progress of KG techniques,KG contains massive fact triples,and the topological structure between these triples is extremely complicated.Therefore,how to precisely and efficiently retrieve useful information from KGs is still an intractable challenge.In recent years,in order to connect the end-users with the intelligent information systems and retrieve information from KGs,knowledge graph question answering(KGQA)methods are proposed.KGQA targets at pinpointing the answer entities or properties by understanding the implicit question semantics.In this thesis,we focus on answering multi-hop questions over the KG.Multi-hop questions need more than one triples to find the answers.Existing multi-hop KGQA methods suffer from the following four issues.(1)Poor model interpretability.Existing methods could not provide intermediate reasoning steps due to the black-box nature of neural network.(2)Relying on strong supervision signals.Existing methods need to know the golden reasoning path in advance.(3)Spurious reasoning path.The path reasoning-based model incidentally arrives at the correct answer entity along a wrong reasoning path.(4)KG incompleteness.KGs suffer from serious incompleteness and insufficiency despite their large scales.This thesis devotes to alleviating above four challenges.Specially,enhancing the necessary interpretability is the first consideration for all the proposed methods in this thesis.To this end,the detailed research works are as follows:1.A reinforcement learning(RL)based method for answering multi-hop questions over the KG with weak supervision.This research focuses on enhancing the interpretability and training with weak supervision.It is impractical to exactly indicate how to answer multi-hop questions step by step,and only the final answer is labeled due to the cost of data annotations.In this research,multi-hop KGQA is formulated as a sequential decision task.We aim to design a policy-guided RL agent to sequentially extend the reasoning path by sampling the most promising action until it reaches a target entity.However,due to the training procedure with weak supervision,RL-based methods suffer from two major challenges.(i)Aimless exploration.(ii)Delayed and sparse rewards.In order to alleviate above two challenges,we inject a KG embedding about the potential target entity into the environment state of RL framework to avoid aimless searching paths.Moreover,inspired by reward shaping,we replace binary rewards with soft rewards,which aim to assign additional reward signals for target entities,thereby alleviating the problem of delayed and sparse rewards to some extent.Experimental results reveal that our proposed model improves the predicting performance,even if this research is trained in weak supervision manner.2.An adversarial reinforcement learning based method for answering multi-hop questions over the KG.This research focuses on inferring answer entities along correct reasoning paths rather than spurious ones.The reasons of spurious paths lie in that:(i)In large scale KGs,there are more spurious paths than correct ones.It is easier for the model to discover spurious ones first.(ii)The reward function only considers the predicted answers,but ignores the correctness of intermediate reasoning paths.In order to alleviate the issue of spurious paths,this research is inspired by adversarial learning,in which a generator and a discriminator are jointly trained via an adversarial process.In our solution,a path discriminator is responsible for distinguishing whether the reasoning chain is correct or not.Subsequently,the output from path discriminator is viewed as a reward to guide the optimization of the answer generator.At the same time,the path discriminator continually enhances the evaluation capacity to identify negative samples provided by the answer generator.These two components are improved via a mutual promotion process.Experimental results reveal that our proposed model achieves promising performance on answer Hits@1 and path accuracy compared with traditional RL-based solutions.3.A multi-agent collaborative learning method for answering questions over the incomplete KG.This research focuses on combating against the missing information in KGs.In order to alleviate the issue of KG incompleteness,text-enhanced methods and implicit reasoning approaches are proposed.Despite achieving comparable performance,above two technical branches either lack necessary interpretability or rely on heuristic shortest paths between topic entities and answers.In this research,we consider dynamically enriching the incomplete KG with auxiliary text-formed triples which are extracted from external text corpora,thereby enriching the graph connectivity.To filter out vast amounts of noisy and irrelevant information from text corpora,we expect to set up two collaborative agents with different purposes.The auxiliary agent,i.e.,evidence extractor,is employed to recommend high-confidence actions from the collection of text-formed triples,thereby distilling valuable text information,these selected actions will be added into the original KG action space.The main agent,i.e.,path generator,learns to sequentially extend the reasoning paths by executing promising actions from the joint action space.Moreover,to facilitate effective explorations,during training phase,the path generator adopts a beam search-based action selection strategy to maintain several partial reasoning paths in one episode.Simultaneously,we also design an adaptive sampling mechanism to increase the selecting priority of partial reasoning paths with text-formed actions.By interacting with the environment,above two policy-guided agents aim to collaboratively execute optimal actions and maximize the expected rewards.Experimental results reveal that our proposed model achieves comparable or superior performance under different KG settings.