基于LSTM的金融新闻倾向性

Financial news tendency based on LSTM

为尽早发现负面新闻以降低对公司的影响,提出采用判断新闻文本关键句群倾向性的方法判断新闻的倾向性。对于公司名识别,在综合原有方法优势的基础上增加百度百科查询,向公司名基础词典加入公司名和公司代码映射;在关键句群抽取环节中,使用doc2vec模型计算句子和新闻标题相似度,综合句子位置信息、句子中领域动词信息、句子中公司名信息;使用Word2vec模型并结合TFIDF的句子表示方法,使句子的表示更加准确、更有侧重。使用LSTM模型对关键句群进行分类,实验结果表明,该模型分类效果优于传统机器学习分类模型和CNN。

To find negative news to reduce the impact on companies,the method that the tendency of a news text was decided by judging key sentences of the text was proposed.As to company name recognition,Baidu Baike query based on the method existed was added,and mapping between company name and company code was introduced to fundamental company name dictionary.In the course of key sentences extraction,doc2vec model was used to calculate the similarity between sentence and news title,and position of sentence,information of field verbs in sentence and information of company name in sentence were also taken into consideration.Considering the analysis of key sentences tendency,a combination of word2vec and TFIDF was used to represent sentences,which made sentence representation more accurate and more focused.LSTM model was trained to classify the key sentences.Experimental results show that the method proposed is better than traditional machine learning classification model and CNN in classification performance.