The topic model of voice interactive dialogue based on the localized command and control system
Haiyan Geng, Baoliang Dong · 2021
The core problem of the human-machine dialogue system in the localized command and control system is to understand the information of the dialogue, and then give a reasonable answer that is satisfactory to the command system commander or assist the commander in completing a certain task. Facing huge interactive dialogue and other short text dialogue data, it is necessary to understand it efficiently and accurately. Daily interactive dialogues are carried out following certain themes, and mining the hidden subject information is a good way to grasp the inherent characteristics of the dialogue. The length of the dialogue text in the daily dialogue of the commander is small, and it has a high degree of randomness. The oral language is serious. Its themes are intertwined and the organizational structure is more chaotic than the news and other types of texts, which makes it difficult for the traditional topic model to capture the hidden words. The co-occurrence law of the contained words and documents cannot be directly applied to this type of text. Therefore, this article constructs a "pseudo-long document" method for short text problems. The topic information in a group of dialogues is relatively similar, and the word co-occurrence in the dialogue is more representative. Using this method to construct the training corpus of the LDA model can improve the topic model's capture of the document-word co-occurrence law to a certain extent. At the same time, this paper proposes an index based on the combination of perplexity and topic similarity to determine the optimal number of topics. Based on these two aspects, an interactive dialogue topic model is constructed.