Deep Learning Intent Recognition via Feature Migration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current intent recognition methods in man-machine dialogue face challenges in obtaining labeled data, especially for methods based on word and PINYIN sequences with large data volumes, leading to limited research and ineffective recognition models.
Innovation Solution
The proposed method involves training a deep learning network using feature vectors of words and PINYIN sequences, where historical voice data is converted into text, manually labeled, and inputted into a first deep learning network to train a language model, with parameter matrices initialized and updated in a second deep learning network to create an intent recognition model, utilizing a convolution layer for enhanced recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If rule-based methods based on dictionaries and templates are used for intent recognition, then the method is simple to implement, but the recognition accuracy is limited and cannot handle complex semantic understanding
Solution Approach 1:
The patent replaces rule-based mechanical systems (dictionaries and templates) with deep learning neural networks that automatically learn semantic patterns from data, enabling the system to perform complex intent recognition without manual rule configuration
Solution Approach 2:
The patent transforms the approach from fixed rule parameters to dynamic learned parameters through training deep learning models on large datasets, allowing the system to adapt to different intent recognition tasks and improve accuracy
2Measurement precision
If recognition models are trained using feature vectors of words and PINYIN sequences with massive data volumes, then the recognition accuracy can be improved, but obtaining labeled data becomes extremely difficult and time-consuming
Solution Approach 1:
The patent performs preliminary actions by pre-training language models on large unlabeled corpora to learn general language features, which are then transferred to the intent recognition task, eliminating the need to manually label all training data
Solution Approach 2:
The patent introduces an intermediary language model that bridges the gap between unlabeled data and the final intent recognition model, using the language model to process and transform raw text into meaningful representations that can be efficiently trained with less labeled data
3Measurement precision
If deep learning networks are trained with large volumes of labeled data, then the intent recognition accuracy improves, but the complexity of the system and computational resources required increase significantly
Solution Approach 1:
The patent segments the training process into distinct stages: pre-training on large unlabeled data, fine-tuning on smaller labeled datasets, and domain-specific adaptation, making the overall complex process more manageable and efficient
Solution Approach 2:
The patent performs preliminary pre-training of language models on large corpora before the actual intent recognition training, so that when the model is fine-tuned on labeled data, it requires fewer parameters and less computational complexity to achieve high accuracy
Data Source
AI summary
The present invention relates to the field of intelligent recognition, and discloses an intent recognition method based on a deep learning network, resolving a technical problem that accuracy of intent recognition is not high. A key point of the technical solutions is migrating features of a first deep learning network to a second deep learning network, mainly including: converting data sets of all fields into a word sequence WS and a corresponding PINYIN sequence PS; meanwhile, manually labeling the data set of a certain field and converting the data set into a word sequence WD, a PINYIN sequence PD, and a label; inputting the word sequence WS and the PINYIN sequence PS to the first deep learning network for training to obtain a language model, initializing and updating an encoding layer parameter matrix of the language model; and weighting and inputting the word sequence WD and the PINYIN sequence PD to the second deep learning network after the word sequence WD and the PINYIN sequence PD are inputted to the second deep learning network to be encoded, to train an intent recognition model. Accuracy of performing intent recognition by using the intent recognition model is higher.