毕业生知识库

With the revolutionary changes of new media communication, online media platforms have become the main carrier of event diffusion, providing interactive and diverse multi-dimensional communication content with event-related emotions, topics and institutions. Mining these information facilitates the management department to perceive emotional situation, understand the distribution of topics, and track event-related institutions, which can provide decision support by assessing the impact of the event diffusion. This thesis aims to take advantage of the research progress in pre-trained language models, graph neural networks and reading comprehension based framework to analyze events from three aspects including emotion element extraction, topic mining and key institution recognition. The major works of this thesis are summarized as follows.

1. Firstly, a machine reading comprehension based method is proposed for emotion element extraction. The event contains rich emotion elements which reveal the deep emotional trends behind the event. To alleviate the problem of insufficient modeling granularity and lack of target information fusion in existing emotion element extraction models, this thesis proposes a machine reading comprehension based extraction method. The method first takes emotion element-related query as prior knowledge, and fuses target information with text content based on a pre-trained language model to obtain element-oriented text representation. Then, a hierarchical multi-task learning mechanism is designed to enhance the answer selection process and enables the model to extract multiple emotion elements at the same time. The experimental results on two public datasets demonstrate the efficacy of our proposed model with task-oriented query and multi-task learning structure.

2. Secondly, a heterogeneous graph based method is proposed for event topic mining. Topic serves as a generalized expression of the main content of an event. To tackle the document semantic sparsity and topic overlapping issue, we propose a heterogeneous text graph based event topic mining method, which constructs a heterogeneous text graph based on documents and words. During representation learning, a two-channel encoding module learns both structural and semantic information via multi-layer heterogeneous graph convolution network and auto-encodrer, seperately. We combine the hidden states learned by the auto-encoder and graph convolution network in each layer to obtain a more comprehensive representation of the document. Furthermore, a dual-supervised mechanism is used to uniformly guide the learning process of these two channels. Experiments on a real-world event topic dataset show that the proposed method that deeply combines pre-trained model and heterogeneous convolution neural graph significantly improves the topic mining performance.

3. Thirdly, a multi-turn question answering based method is proposed for topic and key institution recognition. Several institutions are involved in the course of event diffusion, and there is usually an implicit relationship between instutisions and event topics. To fully exploit event topic information to boost the performance of event-related key institution recognition, this thesis proposes a multi-turn question answering based method that jointly performs topic mining and key institution recognition. The method utilizes queries and pre-trained language model to obtain task-oriented contextualized semantic representation. The model peforms span selection to extract all the entities and answer multi-choice selection to mine topics in the first turn, then, it constructs queries based on previous answers and performs judgment to discriminate key institutions in the second turn. Experiments show that the proposed topic and key institution recognition method can effectively mine the deep relation among institutions, topics and key institusions to improve the performance of key institution recognition.