Neural network-based dialogue semantic intention prediction method and learning training method

Abstract

The invention discloses a neural network-based chat robot dialogue semantic intention prediction method and a learning training method. The prediction method is used for a chat robot to predict a possibly continuous dialogue intention or point of interest (one of multiple dialogue intentions or points of interest) of a dialogist according to a sequence (said) statement eigenvector of the human dialogist; and by heuristically selecting an output dialogue of the chat robot and taking an answer of the dialogist as a verification feedback of a previous intention prediction result, online reinforcement learning and self-updating of an algorithm are realized, the accuracy of dialogue intention prediction is continuously improved, the round number of chat can be increased, and the interest of thedialogist can be enhanced.

Description

technical field [0001] The present disclosure relates to the technical field of semantic analysis, in particular to a neural network-based dialog semantic intent prediction method and a learning and training method. Background technique [0002] Through chatbots, people communicate with computers in natural language. In the process of chatting with the robot, a large amount of data containing rich user information is generated, which contains the user's potential needs and intentions. Intent recognition aims to judge whether the text or performance behavior posted by a user has a certain directional intention (such as consumption, leisure, knowledge, etc.), and accurately identifying user intentions can make chatbots more intelligent and enhance user experience; at the same time , and can also recommend consumer products in a targeted manner to better serve users. [0003] Conventional intent recognition uses template-matching-based methods and guided classification-based ...

AI Technical Summary

Problems solved by technology

[0011] The LSTM neural network model needs to rely on the original word vector input. For the semantic intent recognition task in chat robots, it is difficult to obtain large-scale corpus for training word vectors (especially human-machine dialogue). initialization vector representation

[0012] The LSTM neural network model only takes each word vector of a sentence (instead of each sentence text in several consecutive sentences) as input, and inputs it into the corresponding Cell unit in time series, lacking deep semantic information related to the task, so , the model does not fully characterize the chat text

[0013] The LSTM neural network model uses a supervised learning training algorithm to optimize the adjustment of network weights. Therefore, it is necessary to manually identify and label a large number of data samples, which will bring huge time and labor costs.

[0014] The LSTM neural network model adopts the offline learning and training method to realize the optimization and adjustment of the network weights. Therefore, it is impossible to use real-time experience to learn independently online. The network weights can only be updated after collecting a certain number of samples and going through the offline learning and training process again.

Images