Loss minimization based keyword distillation

Keywords provide rich semantic information for documents. It benefits many applications such as topic retrieval, document clustering, etc. However, there still exist a large amount documents without keywords. Manually assigning keywords to existing documents is very laborious. Therefore it is highly desirable to automate the process. Traditional methods are mainly based on a predefined controlled-vocabulary, which is limited by unknown words. This paper presents a new approach based on Bayesian decision theory. The approach casts keyword distillation to a problem of loss minimization. To determine which word can be assigned as keywords becomes a problem to estimate the loss. Feature selection is one of the most important issues in machine learning. Several plausible attributes are always be assigned as the learning features, but they are all based on the assumption of words' independence. Machine learning based on them dose not produce satisfactory results. In this paper, taking the words' context and linkages between words into account, we extend the work of feature selection. Experiments show that our approach significantly improves the quality of extracted keywords.