An intention recognition algorithm based on an embedding method

Abstract

The invention discloses an intention recognition algorithm based on an embedding method. The algorithm comprises the steps of based on the data in the specific field, mapping the input text and intention of the user to the same vector space, enabling the input text and intention of the user to be subjected to similarity comparison in the same space, and converting a classification problem into a sorting problem; converting the intention classification problem into an intention sorting problem by mapping statements and intentions to the same vector space; modifying the loss function in the paper, so that the new loss function has better stability and robustness. The invention provides an intention recognition algorithm based on an embedding method. The word vectors pre-trained on the general data set need to occupy the memory space of more than 1 GB, and the model provided by the scheme of the invention only needs about 100 MB memory space due to the fact that only the concerned words and intentions are embedded, so that the occupied memory space is smaller.

Description

technical field [0001] The invention relates to machine learning algorithms, in particular to an intention recognition algorithm based on an embedding method. Background technique [0002] At present, the intention recognition algorithm in intelligent customer service generally selects a pre-trained word vector, converts and maps the user's input text into a word vector, uses the pre-trained word vector to represent the user's input sentence, and then uses the traditional machine learning algorithm Or deep learning algorithm for classification, converting intent recognition into a multi-classification problem. There are mainly four shortcomings in the existing technology: the final effect of intent recognition is greatly affected by the quality of word vectors, compared with the quality of word vectors, the classification algorithm selected has little influence on the final effect; due to the training of word vectors It is generally performed on a general data set, so words...

AI Technical Summary

Problems solved by technology

There are mainly four shortcomings in the existing technology: the final effect of intent recognition is greatly affected by the quality of word vectors, compared with the quality of word vectors, the classification algorithm selected has little influence on the final effect; due to the training of word vectors It is generally performed on a general data set, so words in a specific field may not appear in the word vector or the meaning of a word in a general data set may be different from the meaning of a word in a specific field, resulting in no learning in the pre-trained word vector Words in a specific field or inaccurate learning; word vectors pre-trained on general data sets will have a large number of words that are not used when dealing with problems in a specific field, resulting in waste of memory space; Existing techniques do not give good results when classifying intent sentences

Images