基于回答集语义的观点挖掘方法研究

Rule-based opinion mining approaches have been widely employed recently since rules are simple, easy to explain and efficient. There are three main problems in current researches on rule-based opinion mining methods: 1) the opinion mining rules they use are mainly monotonic first-order logic rules, which have representation limitations on complex non-monotonic extraction tasks; 2) the opinion mining rules they use are mainly generated based on manual work by domain experts, which is hard to guarantee the rule quality; 3) they rarely explore the problem of inconsistency in opinion mining rules. This project studies the approaches to opinion mining based on answer set semantics because of the following reasons. On the one hand, opinion mining rules based on answer set semantics are capable to break through the limitations of first-order logic rules, thus they are suitable to handle complex non-monotonic problems in opinion mining. On the other hand, rule learning and reasoning under answer set semantics are more efficient, and it is also easy to handle rule inconsistency in rule sets. This project will mainly focus on two issues: 1) rule learning approaches based on answer set semantics, and 2) rule inconsistency handling approaches based on answer set semantics. This study will not only be meaningful to opinion mining and rule learning, but also provide valuable techniques and theoretical supports to knowledge discovery in other domains.

基于规则的观点挖掘方法已被广泛应用。规则具有简单、可解释、高效等优势。当前基于规则的观点挖掘方法研究不足主要有三点：1）现有方法大多基于单调的一阶逻辑规则抽取观点，在复杂的非单调任务上具有局限性；2）现有方法使用的观点挖掘规则多是领域专家手工设计和筛选的，难以保证规则质量；3）探索观点挖掘规则不一致问题的研究较少。本项目研究基于回答集语义的观点挖掘方法。一方面，基于回答集语义的观点挖掘规则可以突破一阶逻辑规则不能进行非单调推理的局限，更加适合处理复杂的观点挖掘问题。另一方面，在回答集语义下可以高效地进行规则学习和推理，也可以方便地处理规则集中的不一致问题。本项目将着重探索以下两个问题：1）基于回答集语义的规则学习方法；2）基于回答集语义的规则不一致问题处理方法。本课题的研究不仅对观点挖掘、规则学习的理论创新具有重要意义，也为其它应用领域的知识发现提供理论支撑和技术储备。

本项目研究基于回答集语义的观点挖掘方法。一方面，基于回答集语义的观点挖掘规则可以突破一阶逻辑规则不能进行非单调推理的局限，更加适合处理复杂的观点挖掘问题。另一方面，在回答集语义下可以高效地进行规则学习和推理，也可以方便地处理规则集中的不一致问题。本项目的主要贡献如下：1）提出一种基于回答集编程（Answer Set Programming, ASP）的观点挖掘一般框架，该框架支持非单调推理；2）针对基于模板的规则学习方法提出两种算法：第一种采用贪心算法，第二种采用局部搜索算法，实验结果表明两种方法所选择的规则子集的性能均显著优于初始规则集的性能；针对基于数据的规则学习方法提出一种基于解释学习（Explanation-based Learning, EBL）的规则学习方法，实验结果表明该方法能有效利用有限的训练正例学习ASP抽取规则。3）提出两种处理ASP规则集不一致问题的方法。第一种基于马尔可夫逻辑网（Markov Logic Networks, MLNs），实验结果表明该方法能够有效处理多种类型领域知识之间的关系。第二种基于深度Q网络（Deep Q-Network, DQN），实验结果表明该方法在不需要领域知识和人工标注的前提下能够从试错经验中有效地学习选择策略。