A text inference method and device based on rule embedding

Abstract

A text inference method based on rule embedding, including neural retrieval and inference of different components of logical rules based on pre-trained semantic logic network, and supports user demand changes or task migration; combined with the parallelism of semantic logic network and neural classification network The structure adopts the probability distribution distance function Jensen-Shannon divergence, and constrains the consistency of its inference results through network fine-tuning training. The semantic logic network proposed by the present invention encodes user rules into semantic vectors, can better preserve the semantic information of text while detecting logic rules, and supports language flexibility and text diversity. The present invention also proposes a method of integrating user rules into the neural classification network to improve the performance of text inference, that is, combining the parallel prediction structure inferred by the neural classification network and the semantic logic network, and adopting the consistency joint loss, which can make the semantic logic network and the neural network Classification networks benefit from each other and use rule detection results as evidence for text inference.

Description

technical field [0001] The invention discloses a text inference method and device based on rule embedding, belonging to the technical field of natural language processing. Background technique [0002] Public opinion subscription is an important application scenario in the new media era. It means that media organizations regularly push Internet public opinion or news texts that users pay attention to according to the needs of subscribers. User needs are usually expressed in the form of keyword logic rules, which describe The text content of user preferences. The task of text inference based on user needs refers to determining whether a text meets user needs, and this task has important application value in the above scenarios. [0003] Existing technologies for processing the above inference tasks are mainly divided into two categories. One is to infer based on keyword Boolean retrieval results, and to find the text that matches the logical expression by comparing the text ...

AI Technical Summary

Problems solved by technology

[0003] Existing technologies for processing the above inference tasks are mainly divided into two categories. One is to infer based on keyword Boolean retrieval results, and to find the text that matches the logical expression by comparing the text with the keyword logical expression defined by the user. However, this There are limitations in this keyword Boolean retrieval method, and the flexibility of natural language makes the text expression form with the same semantics have a great degree of freedom, which affects the matching results

The other is a classification method based on deep learning, inferring text types based on pre-trained word vectors and neural networks, and performing supervised learning on large-scale labeled datasets, so that the neural network can understand and infer whether the text meets user needs from the semantic level , such as the acquisition of text representation vectors based on convolutional neural networks is described in Chinese patent document CN113076488A: a method and system for recommending information based on user data, which uses preset keywords to perform feature modeling on specific sentences in text carrying user information , but its disadvantage is that it is difficult to deal with the diversity of topics involved in user needs, and it is difficult to adapt to changes in user needs

Images