融入事件知识的主题表示方法

发布时间：2018-07-11 15:20

本文选题：事件 + 主题模型　；参考：《计算机学报》2017年04期

【摘要】：主题模型已被广泛用于发现文档潜在主题.已有方法多采用词或短语来表示主题,然而这些方法生成的主题缺乏深层次的语义信息,可解释性比较差.文中提出使用结构化的事件来表示主题.一方面,事件包含比词或短语更丰富的语义;另一方面,一组相关的事件能更合理地解释并区分不同的主题.为解决事件作为基本单元所带来的稀疏性问题,该文在Biterm Topic Model(BTM)的基础上提出两种主题模型,采用两种不同的方式将事件的语义知识融入到主题生成过程中.其中,第1种模型利用Generalized Pólya Urn(GPU)模型天然的聚类效果加大语义相近的事件分配到同一主题的概率,而第2种模型则通过为每个biterm引入指示变量,合理地利用语义知识有效地解决同一个biterm中两个事件的主题分配问题.该文不仅从主题凝聚度和KL散度两个指标直接对主题模型进行评估,还通过将主题表示结果引入到文本分类任务中对模型进行了外部评估.实验结果表明文中提出的模型从共现和语义两个层面有效地解决了事件稀疏性问题.与基于词或短语的主题表示相比,事件结构所包含的语义信息提高了主题生成质量,使主题表示具有更强的可读性和主题判别性.
[Abstract]:Topic models have been widely used to discover potential topics in documents. Most of the existing methods use words or phrases to express the topic, but the topics generated by these methods lack deep semantic information and can be interpreted poorly. In this paper, a structured event is proposed to represent the topic. On the one hand, events contain more semantics than words or phrases; on the other hand, a set of related events can explain and distinguish different topics more reasonably. In order to solve the problem of sparsity caused by event as the basic unit, this paper proposes two topic models based on the Biterm topic Model (BTM), which integrates the semantic knowledge of the event into the topic generation process in two different ways. The first model uses the natural clustering effect of Generalized P 贸 lya Urn (GPU) model to increase the probability of semantically similar events being assigned to the same topic, while the second model introduces indicator variables for each biterm. The rational use of semantic knowledge can effectively solve the topic assignment problem of two events in the same biterm. This paper not only evaluates the topic model directly from the two indexes of topic coherence and KL divergence, but also evaluates the model externally by introducing the result of topic representation into the task of text classification. The experimental results show that the proposed model can effectively solve the event sparsity problem from the co-occurrence and semantic levels. Compared with the topic representation based on words or phrases, the semantic information contained in the event structure improves the topic generation quality and makes the topic representation more readable and subject discriminant.
【作者单位】：武汉大学计算机学院;乐山师范学院计算机科学学院;武汉大学软件工程国家重点实验室;
【基金】：国家自然科学基金(61133012,61373108,61373056) 国家社科重大招标计划项目(11&ZD189)资助~~
【分类号】：TP391.1

【相似文献】