Hidetsugu Nanba

Text summarization challenge 2: text summarization evaluation at NTCIR workshop 3

We describe the outline of Text Summarization Challenge 2 (TSC2 hereafter), a sequel text summarization evaluation conducted as one of the tasks at the NTCIR Workshop 3. First, we describe briefly the previous evaluation, Text Summarization Challenge (TSC1) as introduction to TSC2. Then we explain TSC2 including the participants, the two tasks in TSC2, data used, evaluation methods for each task, and brief report on the results.

Automatic acquisition of script knowledge from a text collection

In this paper, we describe a method for automatic acquisition of script knowledge from a Japanese text collection. Script knowledge represents a typical sequence of actions that occur in a particular situation. We extracted sequences (pairs) of actions occurring in time order from a Japanese text collection and then chose those that were typical of certain situations by ranking these sequences (pairs) in terms of the frequency of their occurrence. To extract sequences of actions occurring in time order, we constructed a text collection in which texts describing facts relating to a similar situation were clustered together and arranged in time order.We also describe a preliminary experiment with our acquisition system and discuss the results.

Automatic detection of survey articles

We propose a method for detecting survey articles in a multilingual database. Generally, a survey article cites many important papers in a research domain. Using this feature, it is possible to detect survey articles. We applied HITS, which was devised to retrieve Web pages using the notions of authority and hub. We can consider that important papers and survey articles correspond to authorities and hubs, respectively. It is therefore possible to detect survey articles, by applying HITS to databases and by selecting papers with outstanding hub scores. However, HITS does not take into account the contents of each paper, so the algorithm may detect a paper citing many principal papers in mistake for survey articles. We therefore improve HITS by analysing the contents of each paper. We conducted an experiment and found that HITS was useful for the detection of survey articles, and that our method could improve HITS.

Alignment between a technical paper and presentation sheets using a hidden Markov model

We have been studying the automatic generation of presentation sheets from a technical paper. Our approach consists of obtaining a set of rules for generating presentation sheets by applying machine learning techniques to many pairs of technical papers and their presentation sheets collected from the World Wide Web. As a first step, in this paper, we propose a method for aligning technical papers and presentation sheets. Our method is based on Jings method, which uses a hidden Markov model (HMM). Although this method is useful to align short sentences in newspaper articles, it is inapplicable to align sentences in a paper including charts and long sentences. Therefore, we analyse features of papers and sheets, such as information from text appearance, and propose an alignment method that combines the use of these features and Jings method. The evaluation shows that our alignment method performed effectively.

Corpus and evaluation measures for multiple document summarization with multiple sources

In this paper, we introduce a large-scale test collection for multiple document summarization, the Text Summarization Challenge 3 (TSC3) corpus. We detail the corpus construction and evaluation measures. The significant feature of the corpus is that it annotates not only the important sentences in a document set, but also those among them that have the same content. Moreover, we define new evaluation metrics taking redundancy into account and discuss the effectiveness of redundancy minimization.

Producing more readable extracts by revising them

In this paper, we first experimentally investigated the factors that make extracts hard to read. We did this by having human subjects try to revise extracts to produce more readable ones. We then classified the factors into five, most of which are related to cohesion, after which we devised revision rules for each factor, and partially implemented a system that revises extracts.

Towards Multi-paper Summarization Using Reference Information

This paper presents a system to support writing a survey of a specific domain. The system utilizes reference information that consists of reference relationships between papers and the information derived from the description around citations. We think the following are inevitable for writing a survey : collecting papers of the specific domain, and understanding their essence and differences among them. Therefore, we firstly extract fragments of papers where the author describes the essence of a referred paper and the differences between his paper and it (we call them reference areas). Then with the information of reference areas, we identify the types of reference relationships that indicate the reasons for citations(we call them reference types). These types make it possible to collect papers in the same domain. The system can display the collection of the papers. It can also show abstracts and reference areas of the collected papers. With the system, we can understand the relationships between the collected papers.

Construction of a cooking ontology from cooking recipes and patents

A cooking ontology is an indispensable language resource for the language processing of cooking recipes. We have constructed a cooking ontology by means of pattern matching, statistical natural language processing techniques, and manual steps to identify hyponymy, synonymy, attributes, and meronymy.

Automatic extraction of citation information in Japanese patent applications

The need for academic researchers to retrieve patents and research papers is increasing, because applying for patents is now considered an important research activity. However, retrieving patents using keywords is a laborious task for researchers, because the terms used in patents for the purpose of enlarging the scope of the claims are generally more abstract than those used in research papers. Therefore, we have constructed a framework that facilitates patent retrieval for researchers, and have integrated research papers and patents by analysing the citation relationships between them. We obtained cited research papers in patents using two steps: (1) detection of sentences containing bibliographic information, and (2) extraction of bibliographic information from those sentences. To investigate the effectiveness of our method, we conducted two experiments. In the experiment involving Step 1, we prepared 42,073 sentences, among which a human subject manually identified 1,476 sentences containing citations of papers. For Step 2, we prepared 3,000 sentences, in which the titles, authors, and other bibliographic information were manually identified. We obtained a precision of 91.6%, and a recall of 86.9% in Step 1, and a precision of 86.2% and a recall of 85.1% in Step 2. Finally, we constructed an information retrieval system that provided two methods of retrieving research papers and patents. One method was retrieval by query, and another was from the citation relationships between research papers and patents.

Automatic Compilation of an Online Travel Portal from Automatically Extracted Travel Blog Entries

For travelers who plan to visit a particular tourist spot, information about it is required. In this paper, we propose a method for extracting and organizing appropriate information from weblogs (blogs). Recently, increased numbers of travelers have been writing of their travel experiences via blogs. We call these travel blog entries, and they contain much useful travel information. For example, some bloggers introduce useful web sites for a tourist spot, while others report on transportation between tourist spots. Here, we extract hyperlinks of web sites for tourist spots from travel blog entries and organize them via automatic classification. We also extract transportation information automatically from travel blog entries. To investigate the effectiveness of our method, we conducted experiments. For the extraction of transportation information, we obtained an 80.3% for Precision. For the classification of hyperlinks, we obtained a high Precision. Finally, we constructed a prototype system, which provides information about (1) transportation between tourist spots and (2) useful web sites for tourist spots.