A Semi-supervised Approach for Generating a Table-of-Contents

International Conference RANLP 2009 - Borovets, Bulgaria, pages 312–317 A Semi-supervised Approach for Generating a Table-of-Contents Viet Cuong Nguyen, Le Minh Nguyen, and Akira Shimazu School of Information Science Japan Advanced Institute of Science and Technology 1-1 Asahidai, Nomi, Ishikawa 923-1211, Japan {cuongnv,nguyenml,shimazu}@jaist.ac.jp Abstract This paper presents a semi-supervised model for generating a table-of-contents as an in- dicative summarization. We mainly focus on using word cluster-based information derived from a large amount of unannotated data by an unsupervised algorithm. We integrate word cluster-based features into a discrimina- tive structured learning model, and show that our approach not only increases the quality of the resulting table-of-contents, but also re- duces the number of iterations in the train- ing process. In the experiments, our model shows better results than the baseline model in generating a table-of-contents, about 6.5% improvement in terms of averaged ROUGE-L score. Keywords text generation, text summarization, semi-supervised learning 1 Introduction In our research, to help people quickly having an overview of a topic or an event, we automatically col- lect articles from online newspapers related to that topic or event, and summarize them. We should be able to make a summarized document from those ar- ticles by extracting important sentences. However, there is a lot of articles with many useful information which makes the document too long as well as does not contain the deep comments of the authors of those ar- ticles about the topic or event. Our aim is to make a concise summary in the form of table-of-contents, automatically. A table-of-contents is a hierarchical structure of titles and locations of segments in a very long text such as books or a set of texts such as multiple docu- ments which describe the same topic [2]. It is a type of indicative summarization that is especially suited for locating and accessing information. With a table- of-contents as a navigation tool, a reader can quickly get not only an overview of the content of a very long text or multiple documents via the titles of segments, but also the location of needed information via the locations of segments. The task of automatically generating a table-of- contents involves two subtasks: (1) separating every article into a hierarchical structure of segments (a tree of segments) and merge them to make a unique tree of segments [6, 11, 12, 15] and (2) generating a title for each segment in that tree to make a table-of-contents [1, 2]. In this paper, we focus on subtask two with the assumption that the tree of segments are easily got from existing methods such as C99 [6] or TextTiling [11] with a text structuring method [5]. The subtask two is previously mentioned in [1, 2]. In [1], for each segment, they used the most important noun phrase based on its frequency to make the title of that segment. This method made the title too short and having very low quality. In [2], they made a better table-of-contents with a supervised learning method which accounts …

Content truncated to 500 words