Emotion session generation method based on an emotion dictionary and word probability distribution

Abstract

The invention discloses an emotion session generation method based on an emotion dictionary and word probability distribution. The method comprises the following steps: carrying out emotion word embedding on words in a generated sentence; embedding emotional words into the encoder-decoder framework, calculating the generation probability of the next word corresponding to the sentiment word and thegeneral word in the generated sentence by utilizing the state of a decoder;connecting the generation probabilities of the sentiment words and the general words according to a preset rule, and establishing a session model to obtain a next word in the generated sentence; calculating the emotion similarity of the generated part in the input sentence and the generated sentence, and training the session model by adopting an emotion loss function, so that the emotion deviation between the generated sentence and the input sentence is smaller than a set deviation threshold value. According to the method for considering the emotion when the sentence is generated, grammar smoothness and emotion expression when the sentence is generated are balanced, and efficiency and satisfaction of session generation are improved.

Description

technical field [0001] The invention relates to the technical field of data mining, in particular to an emotional conversation generation method based on an emotional dictionary and word probability distribution. Background technique [0002] Due to the popularity of smartphones and the development of broadband wireless technology, we are now in the age of social media, and more people are connected to each other through data. It is natural for machine conversation generation to develop into a social method. Earlier interactive systems, such as Eliza (Weizenbaum, 1966), Parry (Colby, 1975) and Alice (Wallace, 2009), were all designed to mimic human behavior for text conversation generation, and passed the Turing test within the control range ( Turing, 1950; Shieber, 1994). Despite their impressive success, these predecessors to current session generation operated primarily based on handcrafted rules. Therefore, they can only perform well in limited environments. [0003] ...

AI Technical Summary

Problems solved by technology

[0005] The first problem arises because sentiment annotation is a relatively subjective task, and sentiment classification is also challenging

In large corpora, high-quality sentiment labels are difficult to obtain

[0006] The second problem is that it is difficult to consider emotion in a natural and coherent way due to the need to balance grammatical fluency and emotional expression of the generated sentences

Simply embedding emotions in existing neural models will only produce incomprehensible expressions, which hardly produce satisfactory answers

Images