User quality inspection demand classification method and system based on rule matching and deep learning

Abstract

The invention discloses a user quality inspection demand classification method and a system based on rule matching and deep learning, and the method comprises the steps: matching a to-be-classified text with a proper noun dictionary of each industry, calculating an industry probability value Pi matched with an industry label corresponding to each industry, and selecting a maximum industry probability value P; if the maximum industry probability value P is greater than or equal to a preset threshold G, outputting the industry label corresponding to the maximum industry probability value P as a classification result; otherwise, predicting an industry label of the to-be-classified text by using a deep learning model, and calculating a predicted industry probability value H; and for the maximum industry probability value P and the predicted industry probability value H, selecting the industry label corresponding to the maximum value as a final classification result and outputting the final classification result. According to the method, the quality inspection demand texts can be automatically and accurately classified according to the industries to which the quality inspection demand texts belong aiming at the user quality inspection demand texts.

Description

technical field [0001] The invention relates to the field of inspection and testing services and the field of machine learning, and in particular to a method and system for classifying user quality inspection requirements based on rule matching and deep learning. Background technique [0002] The domestic third-party inspection and testing website has inspection and testing agencies from various industries settled in to provide product quality inspection and testing services. Users can contact agencies by actively contacting or publishing quality inspection demand information. Faced with a large number of professional testing services, users tend to publish product testing demand information on the website, so that institutions that meet the testing conditions can contact. In order to achieve an accurate match between user needs and testing services, it is a key step to classify quality inspection needs by industry. At present, the classification of quality inspection deman...

AI Technical Summary

Problems solved by technology

[0003] At present, existing user quality inspection requirements classification methods include: (1) The current industry classification of quality inspection requirements texts mainly relies on manual classification by users. Users need to select specific labels from many industry categories, and the classification efficiency is low and limited. Due to the user's knowledge reserve, misclassification often occurs, which affects the quality of inspection and testing services

(2) Text classification for quality inspection requirements based on machine learning methods, such as support vector machine, K nearest neighbor, naive Bayesian and other algorithms are suitable for small-scale data sets, and there are shortcomings such as cumbersome feature extraction for text and weak model generalization performance ;(3) The text classification method based on the convolutional neural network predicts the text category by setting the parameters of the input layer, convolution layer, pooling layer and output layer. The training speed is fast, but the word order and semantic information of the text cannot be obtained. The effect is average; (4) The text classification method based on the long-term short-term memory network considers the context information of the text, and the classification performance is better, but the classification effect in small-scale data sets is average; (5) fasttext: it is a fast text Classification algorithm, but it does not fully consider the characteristics of short text and many professional vocabulary required by quality inspection. It is necessary to consider the characteristics of the text to design a suitable classification model

[0004] Compared with other Chinese texts such as news and user comments, the quality inspection demand texts published by users have the following characteristics: (1) There are many professional words, such as "protein / radio frequency antenna / xenon lamp / ballast, etc.", the frequency of occurrence of these words It is low, but it has a great impact on the classification results. The commonly used text classification methods do not fully consider these text features, and the classification effect is average.

(2) There are more than 20 classified industries, and the amount of data in each industry is unbalanced

Common deep learning models have been fully trained in large-scale data sets. The amount of data required for quality inspection in various industries is unbalanced (thousands at most and hundreds at least). The classification effect of using deep learning models is relatively low. Difference

(3) The text features are sparse, and the quality inspection requirements are published by different users. The description is relatively colloquial, and the text features are sparse. The text length of the quality inspection requirements is generally between 15 and 150 characters. It consists of fields of "test content" and "test location requirements". Different fields have different classification capabilities, which are not considered in general text classification methods.

[0005] To sum up, based on the classification of quality inspection needs based on user experience, multiple industry categories are preset, such as more than 20 industry categories such as food, agricultural products, environment, medicine, electronic appliances, and biology. Users can classify according to their own knowledge reserves , it is easy to cause misclassification for products with vague industry attributes

Text classification methods based on machine learning, this type of method has a strong dependence on the quality of text feature extraction, and the extraction process is cumbersome, and the generalization ability of the model is weak

Images