| With the development of contemporary publishing industry, the load of human-collation keeps increasing leading to the demand of automatic-collation. Automatic-collation is an important and attractive research field in natural language process(NLP) whose main goal is alleviating the load of people who work on collating books or texts.Due to the specific characteristics of interlanguage texts, they show more complexity than the common texts in original and presentation of text errors. The Chinese automatic-collation systems now existed are used to process common Chinese texts written by native Chinese speakers. Their training and test corpora are homogeneous. This paper try to develop a system used to process those texts written by people whose mother tongue is not Chinese. To compare the two kind of system, there existed more difficulties in the latter than the former in both research and application, such as training and test corpora are not homogeneous, more categories of errors, hard to label, etc. Our research use the inter-language corpus build by Department of Preparatory, Xinjiang University, abstracting 2063 sentences containing errors randomly. We constructed the AECIT (Automatic Error Checker for Interlanguage Texts) whose definition is picking up the word error in the sentence. We use a moving window, read a triple word string one time, make the mutual information to work as basic statistic method combining POS and reasonable word pair threshold, then we abstract the errors from sentences. Due to the technology of syntax analysis and semantic analysis is not good enough, we focus our research work on word-class. Finally, take the threshold is 3.0 for instance, AECIT gets: recall 73.3%, precious 50% and miss report 50%. If people change the threshold, they can get higher or lower value either in recall or previous which can meet the demand of different users. |
