Đang chuẩn bị nút TẢI XUỐNG, xin hãy chờ
Tải xuống
The EPISTLE system is being developed in a research project for exploring the feasibility of a variety of intelligent applications for the processing of business and office text (!'Z; the authors of are the project workers). Although ultimately intended functions include text generation (e.g., 4), present efforts focus on text analysis: developing the capability to take in essentially unconstrained business text and to output grammar and style critiques, on a sentence by sentence basis. Briefly, we use a large on-line dictionary and a bottom-up parser in connection with an Augmented Phrase Structure Grammar (5) to obtain an approximately correct. | NATURAL LANGUAGE TEXTS ARE NOT NECESSARILY GRAMMATICAL AND UNAMBIGUOUS OR EVEN COMPLETE. Lance A. Miller Behavioral Sciences and Linguistics Group IBM Watson Research Center p. 0. Box 218 Yorktown Heights NY 10598 The EPISTLE system is being developed in a research project for exploring the feasibility of a variety of intelligent applications for the processing of business and office text 1-3 the authors of 3 are the project workers . Although ultimately intended functions include text generation e.g. 4 present efforts focus on text analysis developing the capability to take in essentially unconstrained business text and to output grammar and style critiques on a sentence by sentence basis. Briefly we use a large on-line dictionary and a bottom-up parser in connection with an Augmented Phrase Structure Grammar 5 to obtain an approximately correct structural description of the surface text e.g. we posit no transformations or recovery of deleted material to infer underlying deep structures . In this process we always try to force a single parse output even in the presence of true ambiguity. Grammatical critiques are provided by having very strong grammar restrictions in an initial processing of the sentence should the application of grammar rules fail to lead to the identification of a complete syntactically correct sentence we then process the material a second time adding other rules which essentially relax certain constraints such as subject-verb number agreement thereby permitting us to recognize a wide variety of true grammatical errors. The stylistic critiques are based on measurements of the detailed hierarchical structure descriptions produced by the parser letting us detect a variety of stylistic characteristics judged by experts to be undesirable too great a distance between subject and verb too much embedding unbalanced subject predicate size excessive negation or quantification etc. The text corpus used for system construction and testing is a set of .
