Learning Domain Feature from Text Corpora

For improving performance in automatically electronic documents processing, this paper proposes a concept of domain feature, which is defined as terms that can represent topics of a certain domain. Then it presents a non-lexicon-based approach automatically learning domain feature from text corpora. This approach combines the length first segment algorithm and domain feature possibility (DFP) algorithm. The former segments domain foreground corpora and extracts words and phrases in a satisfying recall rate, while the latter enhances the precision rate of learning by comparing different statistic properties that domain feature shows between foreground and background corpora. Experiments verify that given appropriate foreground and background corpora, this approach significantly improves efficiency in domain feature building and gets better result than manually building does. Algorithms combined in this approach can be widely used in other research domains of knowledge management.