Topic word-based language generation method

Abstract

The invention provides a topic word-based language generation method; a traditional Seq2Seq model is used to train context and descriptors, intermediate level information ci corresponding to implicit level information of input Xi is calculated at a model encoding portion, word vectors of the descriptors are used to generate intermediate level information oi through an attention mechanism, and the information ci and oi are subjected to joint influencing through a structure combined with the attention mechanism to finally generate a sequence regarding both the context and the descriptors. The topic word-based language generation method provided herein has the advantages that generated text is smoother and highly diversified, conveniences are brought to users, and potential users are guided in purchase deciding.

Description

technical field [0001] The invention relates to a language generation method, in particular to a method for language generation based on subject words using deep learning and an attention mechanism. Background technique [0002] With the development of Internet technology, online user reviews have a great impact on e-commerce and consumers. Studies have shown that most consumers will collect product and service-related information online before making a purchase decision, and will share consumption experience and purchase evaluation online after purchase. In addition, a large number of Internet users read user reviews before purchasing products or services, and are influenced by the content of reviews. Therefore, massive online user reviews are an important source of information to help consumers discover product quality and make corresponding purchase decisions. However, due to the cumbersome review process at this stage, users are unwilling to spend more time evaluating ...

AI Technical Summary

Problems solved by technology

[0010] (1) Although template generation is simple and feasible, the generated text is incomplete and of low quality

[0011] (2) Although the generated text structure is hierarchical through pattern generation, this method is only suitable for articles with a fixed structure and lacks flexibility

However, due to the complex semantic and grammatical relationships between sentences, it is not easy to construct a text rule base

[0013] (4) The concept of the method based on attribute features is simple, and the generated text is relatively flexible, but the content relationship between attributes is more complicated, and the workload is heavy, that is, it is difficult to determine what kind of text changes can be added to the collection as attribute features

[0029] Although the Seq2Seq model based on the attention mechanism can generate better text, because the model usually uses the generation method of imitating the language in the training set, there are a large number of general words such as "very good" and "unclear" in the training text, making the generated Online reviews tend to be "conservative" and lack diversity, that is, they produce a panacea-style language with little information, which leads to the language generated by the model not being able to reflect any information

Images