摘要
【目的/意义】文本表示是自然语言处理的基础工作,是信息检索、文本分类、问答系统的关键问题。【方法/过程】论文介绍了传统的文本表示方法,按照文本不同的粒度,回顾了近五年国内外基于神经网络模型的词表示、句子表示、篇章(段落)表示的方法,并提出了未来的研究方向。【结果/结论】实验发现,通过在神经网络模型中融入更多的特征能得到更优的词向量,但词向量还缺乏统一的评价标准,句子向量表示通常根据具体NLP任务建模,不同结构的模型在特征表示、运算速度上各有优劣势,篇章表示通常使用层次组合模型。
【Purpose/significance】Text representation is the basic work of natural language processing,the key task of information retrieval,text classification and question answering system.【Method/process】The paper introduces the traditional method of text representation,according to the different granularity of the text,the paper reviews the method of representation of word,sentence and document(paragraph)based on neural network in recent five years,and puts forward the future research direction.【Result/conclusion】The experiment found that incorporating more features into the neural network,a better word embedding can be obtained,but word embedding lacks the unified evaluation standard,the sentence vector representation usually according to the specific NLP task,the different structure model each has different advantage and disadvantage in the feature representation and the computing speed.And the document representation usually uses hierarchical model.
作者
李枫林
柯佳
LI Feng-lin;KE Jia(School of Information Management,Wuhan University,Wuhan 430072,China)
出处
《情报科学》
CSSCI
北大核心
2019年第1期156-164,共9页
Information Science
关键词
神经网络
文本表示
词向量
neural network
text representation
word embedding
作者简介
李枫林,(1962),男,武汉人,教授,博士生导师,主要从事电子商务研究.
