Evaluation aspect determination method and device

Abstract

The invention discloses an evaluation aspect determination method and device, and the method comprises the steps: generating a word vector of an evaluation text according to a character feature vectorof the evaluation text; then, inputting the word vector subjected to part-of-speech tagging into a first bidirectional long-short-term memory network model configured with a dependency transition structure for operation to obtain a first prediction tag sequence corresponding to the evaluation text; and further, based on the first prediction tag sequence, determining an evaluation aspect of the evaluation text. The capability of a bidirectional long-short-term memory network model for learning long-term sequence dependence and syntactic structure priori knowledge provided by dependence transition information are combined, so that the accuracy of determination in the aspect of evaluation is improved, and the technical problems that extra feature engineering is needed and a rule template needs to be manually made in advance are avoided.

Description

technical field [0001] The present invention generally relates to the field of information technology, and specifically relates to a method and device for determining evaluation aspects. Background technique [0002] In modern society, the scope of Internet applications is very wide, which brings great fun and convenience to daily life. On the Internet, we can chat, play games and consult information, etc. More importantly, people can also express their attitudes and views on goods and services, social hotspot events and parties through Internet products such as blogs, microblogs, and post bars. [0003] In recent years, the publication of subjective attitudes and opinions has shown a blowout growth trend. How to determine the evaluation aspects with emotional color has important research value in application fields such as enterprise product competitive intelligence, social public opinion monitoring and market trend forecasting. [0004] In related technologies, the adop...

AI Technical Summary

Problems solved by technology

[0004] In related technologies, the adopted schemes are as follows: first, the determination method of evaluation based on frequency and association rules. Although this method is easy to implement, it requires experts to select templates and formulate matching rules based on experience, and is easily affected by frequent noise. , the determination accuracy is low; second, on the basis of the evaluation method based on frequency and association rules, the point-wise mutual information (Point-wise Mutual Information, PMI) calculation method is introduced to improve the determination accuracy, but this method needs to be Collecting the category indication discriminator of the product, and at the same time, the calculation of point mutual information requires the help of a search engine, which will consume a lot of time; third, the method of determining the evaluation aspect based on deep learning, through the Recurrent Neural Network (RNN) and pre-trained word embeddings for model training, but this method ignores the grammatical constraints in the language that can convey beneficial information; fourth, use the recurrent neural network to learn the distributed representation of the dependent syntactic path, and use the distributed representation as Conditional Random Field (CRF) features for training, but this method requires the help of numerous feature engineering to work, cannot effectively use the advantages of neural networks with strong modeling capabilities, and there is also a distributed representation of dependent information The separation of the training process and the determination process of the evaluation aspect of the

Images