Summary: | 碩士 === 國立臺灣大學 === 資訊工程學研究所 === 93 === Named entity recognition is a fundamental task in biomedical text mining. Multiple-class entity annotation is more complicated and challenging than single-class entity annotation. In this thesis, we presented a single word classification approach to dealing with the multiple-class entity annotation problem using Support Vector Machines (SVMs). In other words, each token in a sentence is represented by a feature vector and classified as one of the given classes. Orthographical patterns, morphological patterns, results from existing gene/protein name taggers, context, part of speech (POS) tags, tags (class labels) of surrounding tokens, and other information are important features for named entity recognition. In addition, we employed a unique way of extracting and utilizing context information. Due to the huge number of non-entity instances (class ‘O’), we clustered the instances of this class into 5 subclasses to accelerate the SVM training process. We also applied a simple post-processing technique with the help of a dictionary and a post-processing technique via abbreviation extraction.
We presented the performance of our system using 13 different notions of correctness, showing the overall performance of our system is somewhere between 68.16% and 79.91% in terms of f-score, which is comparable to the performance of the top 3 systems in the JNLPBA shared task. Besides various notions of correctness used in evaluation, we defined 5 types of errors and showed how frequently our system made these types of mistakes. The error analysis also revealed the annotation discrepancies among the training and test corpora. Therefore, researchers approaching biomedical named entity recognition with machine learning algorithms should seek to improve their systems as well as be aware of the correctness of the underlying corpus.
|