Deep Neural Networks For Legal Question Answering Based On Knowledge Graph

The workload of legal practitioners constantly increases as people’s legal awareness is raised in the process of the rule of law.Legal question answering(QA)is of practical use to people since it provides advice and solutions to legal issues.Nevertheless,most people cannot afford to enjoy legal advisory services.Hence,it is significant to develop a legal question answering model to fulfill people’s needs in a labor-saving way.Despite the effectiveness of prior work,we argue that legal question answering is still challenging for several reasons.First,existing legal QA systems adopt keyword matching methods or deep learning methods without concerning the domain-specific knowledge contained in the laws and regulations.The introduction of legal knowledge graph in this thesis helps to convey domain knowledge and provide professional advice.Secondly,legal question answering data is in shortage compared with open-domain QA due to the scarcity of real users and the pricey expert annotation cost.Deep learning models are hard to train with such a few data,resulting in the poor performance of legal QA.Transfer learning techniques can help to prevent overfitting and generalize the model by extracting common knowledge from other tasks.This thesis utilizes a knowledge graph(KG)to leverage domain knowledge and improve model performance by transfer learning.The novel contributions are summarized as follows.1.Propose a legal answer selection model based on knowledge graph: this thesis constructs a legal knowledge graph based on legal disputes and corresponding case elements.A context-aware attention mechanism is designed to capture the contextual semantic question/answer representations as well as the knowledge representations guided by the corresponding context representations.The model further explores the features of different granularities from low to high level of feature abstractions so as to avoid information loss.2.Propose a legal answer selection model by transfer learning: to boost the performace of the proposed deep learning model on the legal domain,which contains limited labeled examples,this thesis proposes a legal answer selection model by transfer learning.A feature extractor is trained on datasets with rich training samples and adapted to the legal domain by fixing part of the network parameters and fine-tuned other parameters.3.Develope an online legal question answering system: an online legal QA system is developed based on the proposed deep model to fulfill people’s needs for legal consultation.To the best of our knowledge,it is the first Chinese legal QA system which integrates the domain knowledge from legal KG to comprehend the questions and answers and provides interpretable visual cues for deep QA model.