Development of an Automatic Trend Exploration System using the MuST Data Collection

Proceedings of the Workshop on Information Extraction Beyond The Document, pages 1–11, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Development of an automatic trend exploration system using the MuST data collection Masaki Murata1 murata@nict.go.jp Qing Ma3,1 3qma@math.ryukoku.ac.jp Toshiyuki Kanamaru1,4 1kanamaru@nict.go.jp Hitoshi Isahara1 isahara@nict.go.jp 1National Institute of Information and Communications Technology 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0289, Japan 3Ryukoku University Otsu, Shiga, 520-2194, Japan Koji Ichii2 ichiikoji@hiroshima-u.ac.jp Tamotsu Shirado1 shirado@nict.go.jp Sachiyo Tsukawaki1 tsuka@nict.go.jp 2Hiroshima University 1-4-1 Kagamiyama, Higashi-hiroshima, Hiroshima 739-8527, Japan 4Kyoto University Yoshida-nihonmatsu-cho, Sakyo-ku, Kyoto, 606-8501, Japan Abstract The automatic extraction of trend informa- tion from text documents such as news- paper articles would be useful for explor- ing and examining trends. To enable this, we used data sets provided by a workshop on multimodal summarization for trend in- formation (the MuST Workshop) to con- struct an automatic trend exploration sys- tem. This system first extracts units, tem- porals, and item expressions from news- paper articles, then it extracts sets of ex- pressions as trend information, and finally it arranges the sets and displays them in graphs. For example, when documents concerning the politics are given, the sys- tem extracts “%” and “Cabinet approval rating” as a unit and an item expression in- cluding temporal expressions. It next ex- tracts values related to “%”. Finally, it makes a graph where temporal expressions are used for the horizontal axis and the value of percentage is shown on the ver- tical axis. This graph indicates the trend of Cabinet approval rating and is useful for investigating Cabinet approval rating. Graphs are obviously easy to recognize and useful for understanding information described in documents. In experiments, when we judged the extraction of a correct graph as the top output to be correct, the system accuracy was 0.2500 in evaluation A and 0.3334 in evaluation B. (In evalua- tion A, a graph where 75% or more of the points were correct was judged to be cor- rect; in evaluation B, a graph where 50% or more of the points were correct was judged to be correct.) When we judged the extraction of a correct graph in the top five outputs to be correct, accuracy rose to 0.4167 in evaluation A and 0.6250 in eval- uation B. Our system is convenient and ef- fective because it can output a graph that includes trend information at these levels of accuracy when given only a set of doc- uments as input. 1 Introduction We have studied ways to automatically extract trend information from text documents, such as newspaper articles, because such a capability will be useful for exploring and examining trends. In this work, we used data sets provided by a work- shop on multimodal summarization for trend in- formation (the MuST Workshop) to construct an automatic trend exploration system. This system firsts extract units, temporals, and item expres- sions from newspaper articles, then it extract sets of expressions as trend information, and finally it arranges the sets and displays them in graphs. For example, when …

Content truncated to 500 words