Extracting Troubles from Daily Reports based on Syntactic Pieces

Extracting Troubles from Daily Reports based on Syntactic Pieces Kakimoto Yoshifumi and Kazuhide Yamamoto Nagaoka University of Technology, 1603-1, Kamitomioka, Nagaoka, Niigata 940-2188 Japan {kakimoto, ykaz}@nlp.nagaokaut.ac.jp Abstract. It is expensive for companies to browse daily reports. Our aim is to create a system that extracts information about problems from reports. This system operates in two steps. First, it records expressions involving troubles in a dictionary from training data. Second, it expands the dictionary to include information not included in the training data. We experimentally tested this extraction system; in the tests, a two-values classifier attained an F-value of 0.772, and experimental extraction of troubles attained a precision of 0.400 and a recall of 0.827. Keywords: syntactic piece, extracting troubles, knowledge dictionary 1 Introduction In resent years, many companies request daily reports of text data from company members. The text data is Email and web form. However, daily reports are browsed by human, it is expensive for companies. Our aim is to cut back the workload of daily reports browsing. This paper discusses a type of information extraction. Ichimura et al. (2001) developed a system that reduces the cost of access to reports and assists decision-making for users. This system extracts ‘good’ or ‘bad’ practices using a human-created knowledge dictionary. However, the knowledge dictionary depends on the specific enterprise because it is constructed using reports from that enterprise. Saito and Watabe (2001) developed a system that extracts and visualizes information indicating troubles using extraction rules defined by humans. They chose the extraction categories ‘trouble’, ‘causality’ and ‘countermeasure’, which their system evaluated with a precision of 0.878, 0.701 and 0.703, respectively. This system depends on the specific domain, because they considered only printer problems. Both systems produce rules or a dictionary only from training data. This approach has two problems. First, because the dictionary and rules are made by humans, the cost is considerable. Second, this system cannot extract troubles that is not included in the training data. Our approach corrects these problems. Our system automatically creates a dictionary, and can extract troubles by expanding the dictionary to add data not included in the training data. 2 Definition of the Trouble In this paper, we define the trouble as ‘content regarding some problem in daily reports’. Troubles must take into account the context of the problem. A single word does not provide adequate troubles because it is too short. For example, when the word ‘壊れる(break)’ occurs in daily reports, we can identify a trouble word. But ‘椅子が壊れる(The chair breaks)’ and ‘サー バーが壊れる(The server breaks)’ have different meanings, calling for different responses and involving different degrees of risk. Aoki and Yamamoto (2007) investigate syntactic pieces that mine units of syntactic structure. Syntactic pieces are pairs consisting of a modifier and a modificand, as shown in Figure 1. 411 22nd Pacific Asia Conference on Language, Information and Computation, pages 411–417 入力文( input sentence ): 出力構文片( output syntactic pieces ): 古いパソコンのバッテリーがいきなり爆発した。 ( The battery of the old computer exploded suddenly. ) 古い( old ) ⇒ …

Content truncated to 500 words