Efficient Statement Identification for Automatic Market Forecasting

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1128–1136, Beijing, August 2010 Efficient Statement Identification for Automatic Market Forecasting Henning Wachsmuth Universit¨at Paderborn Software Quality Lab hwachsmuth@slab.upb.de Peter Prettenhofer and Benno Stein Bauhaus-Universit¨at Weimar Web Technology & Information Systems benno.stein@uni-weimar.de Abstract Strategic business decision making in- volves the analysis of market forecasts. Today, the identification and aggregation of relevant market statements is done by human experts, often by analyzing doc- uments from the World Wide Web. We present an efficient information extrac- tion chain to automate this complex nat- ural language processing task and show results for the identification part. Based on time and money extraction, we iden- tify sentences that represent statements on revenue using support vector classifica- tion. We provide a corpus with German online news articles, in which more than 2,000 such sentences are annotated by do- main experts from the industry. On the test data, our statement identification al- gorithm achieves an overall precision and recall of 0.86 and 0.87 respectively. 1 Introduction Touch screen market to hit $9B by 2015. 50 sup- pliers provide multi-touch screens, and that num- ber is likely to rise.1 Strategic business decision making is a highly complex process that requires experience as well as an overall view of economics, politics, and technological developments. Clearly, for the time being this process cannot be done by a computer at the level of a human expert. However, important tasks may be automated such as market forecast- ing, which relies on identifying and aggregating relevant information from the World Wide Web (Berekoven et. al., 2001). An analyst who inter- prets the respective data can get a reasonable idea about the future market volume, for example. The 1Adapted from http://industry.bnet.com. problem is that a manually conducted Web search is time-consuming and usually far from being ex- haustive. With our research we seek to develop an efficient system that finds and analyzes market forecast information with retrieval, extraction and natural language processing (NLP) techniques. We contribute to the following situation. For a given product, technology, or industry sector we identify and aggregate statements on its market development found on relevant websites. In par- ticular, we extract time information (“by 2015”) and money information (“$9B”) and use support vector classification to identify sentences that rep- resent market statements. The statements’ sub- jects (“touch screen”) are found by relating recog- nized named entities to the time and money infor- mation, which we then normalize and aggregate. In this paper we report on results for the statement identification. To the best of our knowledge no data for the investigation of such market analysis tasks has been made publicly available until now. We provide such a corpus with statements on rev- enue annotated in news articles from the Web; the corpus was created in close collaboration with our industry partner Resolto Informatik GmbH. We pursue two objectives, namely, to support human experts with respect to the effectiveness and completeness of their analysis, and to estab- lish …

Content truncated to 500 words