A BERT-based Approach For Counter Argument Retrieval

Dialog systems have been receiving more and more attentions recently.Generally,these systems can be divided into two categories: task-complete and chitchat dialog systems.The former helps humans in everyday tasks such as hotel and flight booking,hospital registration,and so on.The latter can converse with humans freely in any topic.The objective of this thesis is to build a chitchat dialog system that can debate with humans in controversial topics such as banning guns.Such a dialog agent can benefit humans in multiple ways.For example,during the court trial,lawyers need to find counter-evidence based on the opponent's point of view,and the chatbot can effectively retrieve legal files based on the counter viewpoint.In areas such as financial advisory,corporate decision-making,and public event decision-making,such chatbot can better help human understand the pros and cons of events,allowing human to view things dialectically and make information decisions.Developing a valuable chatbot which retrieves counter argument poses many challenges: First of all,the agent needs to correctly understand the context of the conversation.For example,if a user is discussing religious matters,the chatbot's responses should be appropriate and related to religion.Secondly,the chatbot should be able to response with new perspective and make the discussion more open.For instance,if the user's comments are against gun control,the argumentative chatbot should introduce the reason for supporting banning guns.Finally,the chatbot needs to have a strong generalization ability in order to discuss diverse topics with users.In other words,when a new topic is added to the argumentative chatbot,the system should maintain the performance of each module at a small cost.This thesis proposes THE Debater,a novel chatbot system for counter argument based on information retrieval approach.More specifically,given a user comment,the system selects an appropriate answer from a dataset of question and answer(QA)pairs.To address the first problem,i.e.the context-appropriate matching,THE Debaters adds a layer of semantic matching based on BERT,a state-of-the-art language model recently proposed by Google.Second,a stance detection module is integrated to judge whether a user is in favor or against a specific topic so that the bot knows to debate from other perspective.Last but not least,in order to improve the generalization ability of the system,THE Debater follows different strategies such as making use of large and available data resources,and applying transfer learning methods.Extensive experiments show that our system can achieve good results in both semantic matching and stance detection.In addition,our system also shows promising results towards building a good argumentative chatbot.All in all,the main contributions of this thesis are as follows: 1)Applying BERTbased model for semantic matching: although BERT has been used for other tasks,this is the first time that it is considered for building a chatbot for counter argument;2)Integrating stance detection: previous studies in argumentative conversation systems have not considered this issue.and 3)Exploiting different deep transfer learning technologies to improve the system generality: using network-based transfer learning to improve model's effectiveness in a semi-supervised learning manner;using instancebased transfer learning to accommodate to the shift in sample distribution and expand the amount of data,thus improve the system performance.