A Lexically-Intensive Algorithm for Domain-Specific Knowlegde Acquisition

/ / / | / / / / / L / / A Lexically-Intensive Algorithm for Domain-Specific Knowlegde Acquisition Rend Schneider * Text Understanding Systems Daimler-Benz Research and Technology Ulm, Germany rene. s chneider@dbag, ulm. DaimlerBenz. COM Abstract This paper is an outline of a statistical learn- ing algorithm for information extraction sys- tems. It is based on a lexicaUy intensive anal- ysis of a small number of texts that belong to one domain and provides a robust lemma- tisation of the word forms and the collection of the most important syntagmatic dependen- cies in weighted regular expressions. The lexi- cal and syntactical knowledge is collected in a very compact knowledge base that enables the analysis of correct and partly incorrect texts or messages, which due to transmission er- rors, spelling or grammatical mistakes other- wise would have been rejected by conventional systems. 1 Introduction The major tasks of information extraction systems (IE-Systems) are the unsupervised selection, fast analysis and efficient storage of relevant text pat- terns a person or a group of persons is interested in. It accomplishes this through the use of learned or handcrafted patterns. In the ideal case the re- sults lead to an appropriate reaction, executed by the computer itself (see Figure 1). The extracted information is stored in a template that usually is based on a slot-and-filler model. Whenever the tex- tual information does not fit automatically into the fillers, it has to be changed adequately to the form and content requirements of the template, otherwise the text is rejected. Thus, the templates architecture depends very much on the domain the IE--system was built for, i.e. before processing a text or a mes- sage and starting the linguistic analysis, the cate- gory that the text or message belong to is already This study is part of the project READ. The project READ is funded by the German Ministry for Ed- ucation and Research (BMBF) under grant 01IN503C. The author is responsible for the contents of the publi- cation. known or has been labeled automatically with the aid of a categorizer. In our investigation the system was built to process requests for business reports, ex- tracting the number, years and language of business reports a certain sender asked for. •. i~.~:._5:~:.~::..:,:~_.. WIISWl Im ... I v~$kl 1~ ~ly ha~/. il ycea cot jId .~a n~ iI*,e annt~ re~s ~ y~Jr com~iny .. ~: wan sew too I ! Tel : 1234~6"~0 d k | [,nfon'nationExwacUon-- ~ I ~ • . Unsm~ctumci Text - • cammn~ ~ Action Y~. ~.. 93, 94 t Figure 1: Overview of an Information Extraction System Although a lot of sophisticate investigation has been done in the area of information extrac- tion (Pazienza, 1997) (esp. since the start of the MUC-Conferences in 1987), only few works are con- cerned with the automatic acquisition of the knowl- edge bases that are needed for IE-tasks (Riloff, 1993), which makes the construction of a new sys- tem for a different extraction task still very …

Content truncated to 500 words