基于THUCTC的金融语料情感分析模型优化

近几年，情感分析技术引起人们的兴趣，在金融应用上，可以作为投资者投资前的参考.但是现有方法存在应用过于专一、数据偏差、结果过于笼统和不够精确的问题.因此本文优化一个通用的中文文本分类器，用于对在线评论数据和股票新闻数据进行情感分析.收集整理了2万条数据作为语料库，每条数据分别由3个人进行独立标注.之后对THUCTC进行优化，具体从3个方面对中文文本分类器进行优化，首先是词语切分，使用词干词典方法结合不同的分词法，实验比较后得到二分法为最好的结果；其次，为分类器选择最好的内核，发现Liblinear内核对即时性要求较高的投资人更好，另一方面Libsvm在提高准确率方面更有优势；最后在金融导向的情绪字典方面，它由Chi-square和TF-IDF方法构建，可用在普通文本分类器上.通过这种方式，本文的结果可以被推广且不会失去准确性.

Model Optimization of Financial Corpus Sentiment Analysis Based on THUCTC

Sentiment analysis has attracted interest recently. In financial applications, it can be a reference for investors. However, existing approaches are either so specific as to cause data drift or too general to be precise. Therefore, a general Chinese text classifier for online reviews and news on stocks is optimized. A corpus with 20000 items is first collected. Then, each item is labeled by three persons as ground truth. After that, the THUCTC is optimized, thus optimizing a general Chinese text classifier in three aspects. First, by tokenization, the THUCTC is modified to a 2-gram with a stemming dictionary method and got better results. Second, the best kernel is selected for classifier. The Liblinear kernel is found to be better for people pressed for time. On the other hand, the Libsvm kernel is good at promoting accuracy. Third, a finance-oriented sentiment dictionary is set based on Chi-square and TF-IDF approach. It can be used by on-the-shelf general text classifiers. In this way, the result can be generalized without the loss of preciseness.