Automatic Lexicon Acquisition and Precision-Recall Maximization for Untagged Text Corpora

• Main Authors: Chang, Jing-Shin, 張景新
• Other Authors: Keh-Yih Su
• Format: Others
• Language: zh-TW
• Published: 1997
• Online Access: http://ndltd.ncl.edu.tw/handle/62275166919316022114

博士 === 國立清華大學 === 電機工程學系 === 85 === Automatic lexicon acquisition from large text corpora is surveyed in thisdissertation, with special emphases on optimization techniques for maximizingthe joint precision- recall performance. Both English compound word extractionand Chinese unknown word identification tasks are studied in order to exploreprecision-recall optimization techniques in different languages of differentcomplexity using different available resources. In the English compound wordextraction task, the simplest system architecture, which assumes that thelexicon extraction task is conducted using a classifier (or a filter) based ona set of multiple association features, is studied. Under such circumstances,a two stage optimization scheme is proposed, in which the first stage aims atminimizing classification error and the second stage focuses on maximizingjoint precision-recall, starting from the minimum error status. To achieveminimum error rate, various approaches are used to improve the error rateperformance of the classifier. In addition, a non-linear learning algorithm isdeveloped for achieving maximum precision-recall performance in terms of userspecified objective function of precision and recall. In the Chinese unknownword extraction task, where contextual information as well as word associationmetrics are used, an iterative approach, which allows us to improve bothprecision and recall simultaneously, is proposed to iteratively improve theprecision and recall performance. For the English compound word extractiontask, the weighted precision and recall (WPR) using the proposed approach canachieve as high as about 88% for bigram compounds, and 88% for trigramcompounds for a training (testing) corpus of 20715 (2301) sentences sampledfrom technical manuals of cars. The F-measure performances are about 84% forbigrams and 86% for trigrams. By applying the proposed optimization method,the precision and recall profile is observed to follow the preferred criteriaof different lexicographers. For the Chinese unknown word identification task,experiment results show that both precision and recall rates are improvedalmost monotonically, in contrast to non-iterative segmentation-merging-filtering- and-disambiguation approaches, which often sacrifice precision forrecall or vice versa. With a corpus of 311,591 sentences, the performance is76% (bigram), 54% (trigram), and 70% (quadgram) in F-measure, which issignificantly better than using the non-iterative approach with F-measures of74% (bigram), 46% (trigram), and 58% (quadgram).