A questionnaire survey result analysis method based on AI

Abstract

The present invention relates to an AI-based questionnaire survey result analysis method, belonging to the field of data processing technology. The method includes the following steps: merging the filling results in the semi-open questions into the questions to generate text, and constructing the questionnaire answer text by combining the generated text with the answer text of the open questions; segmenting the text to generate the theme classification code for each short sentence; performing sentiment polarity analysis to obtain the sentiment tendency of the answer text of the open questions; predicting the text intention through the extracted subject words, sentiment polarity and answer consistency results; and generating an analysis report. The present invention combines all the features of the survey results to mine the potential correlation between the answer results, and predicts the intention category based on the theme classification, sentiment polarity and answer consistency, thereby expanding the text information features and improving the accuracy of the prediction; at the same time, the generated analysis report clarifies the single targeted answer text.

Description

Technical Field

[0001] The present invention belongs to the field of data processing technology, and specifically relates to a questionnaire survey result analysis method based on AI. Background Art

[0002] In questionnaire surveys, in addition to semi-open questions, some open-ended questions will also be designed in order to understand the respondents' true thoughts or opinions more comprehensively and accurately.

[0003] However, due to technical limitations, the current analysis of such open-ended question text data collected is mainly based on manual coding, which is time-consuming, costly, and difficult to standardize. At the same time, in the design of questionnaires, there is often a correlation between semi-open questions and open questions, so as to facilitate the collection of divergent answers to structured questions. However, in the analysis process, it is generally only based on the text data of open-ended questions, and the correlation between semi-open questions and open questions is not explored to conduct in-depth analysis of the answer results. The generated analysis report is not perfect, thus ignoring some single but very effective suggestions. Summary of the invention

[0004] In order to solve the above-mentioned problems existing in the prior art, the present invention provides an AI-based questionnaire survey result analysis method, which combines all the features of the survey results to mine the potential correlation between the answer results, and predicts the intention category based on topic classification, sentiment polarity and answer consistency, thereby expanding the text information features and improving the accuracy of the prediction; at the same time, the generated analysis report clarifies the unified and targeted answer text.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] The present disclosure provides an AI-based questionnaire survey result analysis method, comprising the following steps:

[0007] Collect data: merge the results of the semi-open questions into the questions to generate text, and construct the questionnaire answer text by combining the generated text with the answer text of the open questions;

[0008] Extraction of keywords: The topic model is used to obtain the topic of the questionnaire answer text, and the text is segmented to generate the topic classification code for each short sentence;

[0009] Sentiment polarity analysis: Use dictionary-based and rule-based methods to perform sentiment polarity analysis to obtain the sentiment tendency of the text of open-ended questions;

[0010] Text intent recognition: predict text intent through the extracted topic words, sentiment polarity, and answer consistency results;

[0011] Generate an analysis report: Encode and merge the identified intent labels with the subject words, sentiment polarity, and answer consistency to generate a model table of "topic classification + sentiment polarity + answer consistency + intent category" to quantify the results of the questionnaire answer text.

[0012] Furthermore, the subject word extraction comprises the following steps:

[0013] Text vectorization: Use the bag-of-words model or TF-IDF to convert the preprocessed text data into numerical form;

[0014] Construct LDA model: Construct LDA model using preprocessed and vectorized data: Determine the number of topics in the model by using the perplexity in the model evaluation metric; α and β are set, where α Controls the sparsity of topic distribution in documents, and β controls the sparsity of word distribution in topics;

[0015] Model training: Use the document-word matrix to train the LDA model, and estimate the document-topic distribution and topic-word distribution through an iterative process;

[0016] Topic extraction: After model training is completed, the model is used to extract keywords for each topic and the topic distribution of each document: each topic consists of several keywords, and each document is represented as a topic distribution;

[0017] Topic classification coding: Quantify the topics determined by the LDA model, select the number of digital coding bits, and encode the topics.

[0018] Furthermore, the calculation formula of the perplexity is:

[0019] ;

[0020] In the formula, D For document sets; It is a document d The probability of all words appearing in ; N d is the number of words in the dth document.

[0021] Furthermore, the model training comprises the following steps:

[0022] Build a document-word matrix: Count each word in the document to generate a vocabulary, and then build a matrix. The rows of the matrix represent documents, the columns represent words, and the values ​​in the matrix represent the number of times the word appears in the corresponding document.

[0023] Training the LDA model: Use the LDA implementation tool and the LDA model in the Gensim library to train the document-word matrix;

[0024] The training process involves selecting model parameters, including the number of topics and the number of iterations; the LDA model generates word distributions for each topic and topic distributions for each document by analyzing the document-word matrix.

[0025] Furthermore, the sentiment polarity analysis includes the following steps:

[0026] Weight allocation of sentiment words: sentiment words in the sentiment dictionary are divided into positive sentiment word list and negative sentiment word list, and the preprocessed text is matched with the positive and negative sentiment word lists in the dictionary one by one. If the match is successful, the word is assigned a corresponding weight;

[0027] The weight of positive sentiment is set to 1, the weight of negative sentiment is set to -1, and the sentiment polarity value is obtained by stacking the number of sentiment words in the text;

[0028] Weight allocation of negative words: The weight of negative words is set to -1. When a negative word is matched, the sentiment polarity value is multiplied by -1. For multiple negative words, the sentiment polarity value is multiplied in sequence.

[0029] Weight distribution of degree adverbs: Degree adverbs are divided into four levels according to their degree, and the weights are set to 2.0, 1.5, 0.8, and 0.5 from strong to weak, representing the strength of emotional polarity;

[0030] Position weight distribution of degree adverbs and negative words: When both degree adverbs and negative words appear in a short sentence, weight distribution is performed based on the position relationship between the degree adverbs and negative words;

[0031] Calculate the weighted value of sentiment polarity: After determining the weights of sentiment words, negative words, and degree adverbs in the answer text, perform a weighted sentiment polarity operation on each topic in the answer text to obtain the sentiment polarity value of each topic in each text, and then make a judgment on the sentiment tendency:

[0032] The answer text S is divided into n short sentences, and the sentiment polarity value of each short sentence is calculated based on the topic of the short sentence. When sentiment words, degree adverbs and negative words appear in the short sentence:

[0033] ;

[0034] Where W is the sentiment polarity value, E is the weight of sentiment words, N is the weight of negative words, D is the weight of degree adverbs, and a is the position weight parameter;

[0035] When the sentiment polarity value is greater than 0, it indicates positive sentiment; when the sentiment polarity value is equal to 0, it indicates neutral sentiment; when the sentiment polarity value is less than 0, it indicates negative sentiment;

[0036] Emotional polarity coding: Emotional polarity is divided into five categories according to the emotional polarity value. The coding range is set to 1-5, which is represented as follows: 5 represents relatively strong positive emotions, 4 represents normal positive emotions, 3 represents neutral emotions, 2 represents negative emotions, and 1 represents strong negative emotions.

[0037] Furthermore, the weight distribution according to the positional relationship between the degree adverb and the negative word includes:

[0038] When the situation of "negative word + degree adverb + sentiment word" appears, the negative sentiment tendency will be weakened, and the position weight parameter is set to 0.5;

[0039] When the situation of "degree adverb + negative word + sentiment word" appears, the negative sentiment tendency will be strengthened, and the position weight parameter is set to 2.

[0040] Furthermore, the text intent recognition comprises the following steps:

[0041] Answer consistency calculation: measure the knowledge association between the topics of semi-open questions and open questions, and calculate the cosine similarity;

[0042] Constructing a data set: Setting a selection threshold for answer consistency, filtering text data that meets the selection threshold based on the similarity value of answer consistency, merging the three features of topic words, sentiment polarity, and answer consistency corresponding to the filtered text into a unified feature vector, and setting the intent label accordingly to construct an intent recognition data set;

[0043] Establish a machine learning model: divide the data set into a training set and a test set, establish a TextCNN model, concatenate the sentiment polarity and answer consistency vectors with the topic vector through the pooling layer of CNN, and input them into the fully connected layer for Softmax classification; combine the topic words and intent labels into the information pattern required by the model, and train the model;

[0044] Model application: Use cross-validation model performance and adjust model parameters to optimize prediction results. Use the trained model to predict the intent of new samples, input the feature vector of the new sample, and output the predicted intent category.

[0045] Furthermore, the cosine similarity is calculated using the following formula:

[0046] ;

[0047] In the formula, Indicates the subject and The cosine similarity of The word vector representing the topic of the answer text of the semi-open question, The word vector representing the topic composition of the answer text of the open-ended question; and Represents various features of two word vectors.

[0048] Furthermore, the generating of the analysis report also includes marking according to a preset answer consistency selection threshold, performing prompt processing in the analysis report, and mining a single answer result.

[0049] Furthermore, after the data collection step, data preprocessing is also included: the questionnaire answer text is segmented and stop words are removed; and according to the purpose of the analysis, symbols or adverbs and function words with emotional color are removed and retained.

[0050] The beneficial effects of the present invention are:

[0051] The present invention first converts the questions in the questionnaire survey into text descriptions in combination with the answer results, and also incorporates the answer results of semi-open questions into the text analysis, so as to mine the potential correlation of the answers of the respondents, and obtain the divergent thinking of the respondents in answering the open questions based on the semi-open question structure, so as to facilitate the mining of targeted answer suggestions; then the text data is sequentially subjected to topic extraction, sentiment polarity analysis and answer consistency calculation, and a data set is created for model training, which expands the text information features and improves the accuracy of the intention prediction results; finally, quantitative statistics are performed based on the generated "topic classification + sentiment polarity + answer consistency + intention category" model, which improves the statistical efficiency and accuracy of the questionnaire survey results, and at the same time, the questionnaire is improved based on the results of answer consistency, providing guidance for the subsequent questionnaire design. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to facilitate understanding by those skilled in the art, the present invention is further described below with reference to the accompanying drawings.

[0053] Figure 1 A schematic diagram of the steps of an AI-based questionnaire survey result analysis method provided by an embodiment of the present invention;

[0054] Figure 2 A schematic diagram of the steps of extracting subject words provided by an embodiment of the present invention;

[0055] Figure 3 A schematic diagram of the steps of sentiment polarity analysis provided by an embodiment of the present invention;

[0056] Figure 4 A schematic diagram of the steps of text intent recognition provided by an embodiment of the present invention. Detailed implementation manners

[0057] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following will, in conjunction with the accompanying drawings and preferred embodiments, elaborate in detail on the specific implementation manners, structures, features and their effects of the present invention as follows.

[0058] This embodiment provides an AI-based method for analyzing questionnaire survey results. As Figure 1 shown, it includes the following steps:

[0059] S1. Collect data: Combine the filling results in the semi-open-ended questions into the questions to generate text, and construct a questionnaire response text by combining the generated text with the response text of the open-ended questions;

[0060] It can be understood that semi-open-ended questions refer to question types mainly based on options, which give structural answers within a preset range through single-choice or multiple-choice to guide the respondents; open-ended questions are not restricted by preset options or limitations, and respondents can freely express their opinions, ideas or experiences. By combining semi-open-ended questions with the response results and converting them into text, it is convenient to match the responses of semi-open-ended questions and open-ended questions later and evaluate the validity of the questionnaire response text.

[0061] S2. Data preprocessing: Perform preprocessing tasks such as word segmentation and stop word removal on the questionnaire response text. In this embodiment, jieba is used for word segmentation of the text, and according to the purpose and requirements of the analysis, symbols, adverbs and function words with enhanced or weakened emotional colors are removed or retained.

[0062] It can be understood that in this embodiment, Chinese word segmentation is adopted, which is the process of splitting a Chinese character sequence into a sequence of words. Through word segmentation technology, it is beneficial to quickly extract key information and provides great convenience for mining text information. In the process of removing stop words, stop words are generally prepositions, conjunctions and pronouns that appear in the document, such as "de", "ne", "yushi", "zhilei", etc., and some symbols, such as "@", "√", "#", etc. all belong to the category of stop words. By removing stop words, the input noise and redundant information are reduced, and the semantic quality and the performance of the task are improved.

[0063] S3. Topic word extraction: Use a topic model for the questionnaire response text to obtain the text topic, and generate a topic classification code for each short sentence after splitting the text. As Figure 2 shown, it includes the following steps:

[0064] S31. Text vectorization: Convert the preprocessed text data into numerical form. The commonly used method is based on the Bag of Words model or TF-IDF (Term Frequency-Inverse Document Frequency). The Bag of Words model counts the frequency of each word in the document. TF-IDF considers word frequency and inverse document frequency to measure the importance of words.

[0065] S32. Build LDA model: Build LDA model using preprocessed and vectorized data:

[0066] By adopting the perplexity in the model evaluation index ( perplexity ) determines the number of topics in the model, calculated as follows:

[0067] ;

[0068] In the formula, D For document sets; It is a document d The probability of all words appearing in ; Nd is the number of words in the dth document.

[0069] It should be noted that the lower the perplexity, the stronger the model's ability to explain the data. However, since the perplexity usually shows a downward trend as the number of topics increases, it is necessary to determine the number of topics based on the inflection point of the perplexity curve.

[0070] For hyperparameters α and β are set, where α controls the sparsity of topic distribution in documents, and β controls the sparsity of word distribution in topics.

[0071] It should be noted that when α When the value is small, the topic distribution of documents will be more sparse, that is, each document is mainly concentrated on a few topics; α When the value is large, the topic distribution of the document will be more uniform, that is, each document contains more different topics. When the β value is small, the word distribution of each topic will be sparser, that is, each topic is mainly concentrated on a few words; when the β value is large, the word distribution of each topic will be more uniform, that is, each topic contains more different words.

[0072] S33, model training: Use the document-word matrix to train the LDA model, and estimate the document-topic distribution and topic-word distribution through an iterative process, including the following steps:

[0073] S331. Construct a document-word matrix: count each word in the document to generate a vocabulary, and then construct a matrix. The rows of the matrix represent documents, the columns represent words, and the values ​​in the matrix represent the number of times the word appears in the corresponding document.

[0074] S332, training LDA model: select LDA implementation tools, such as Gensim or Scikit-learn, etc. In this embodiment, the Gensim library is selected, and the LDA model in the Gensim library is used to train the document-word matrix;

[0075] The training process involves selecting model parameters, including the number of topics, number of iterations, etc. The LDA model generates word distributions for each topic and topic distributions for each document by analyzing the document-word matrix.

[0076] S34, topic extraction: After the model training is completed, the model is used to extract the keywords of each topic and the topic distribution of each document:

[0077] Topic keyword extraction: Each topic consists of several keywords, which are determined according to their weight in the topic.

[0078] Document topic distribution: Each document can be represented as a topic distribution, that is, the weight of each topic in the document.

[0079] The LDA (Latent Dirichlet Allocation) model is an unsupervised machine learning text mining method proposed by Blei et al. based on the PLSI model. It consists of three layers of Bayesian probability structures: document layer, topic layer, and word layer. It extracts a multinomial distribution of implicit topics from the Dirichlet prior distribution of each document, repeatedly samples the generated topics, maps the document to the topic vector space, and then generates a "document-term" for a document.

[0080] It is understandable that by using the topic model to find out the implicit themes between texts and words, it is convenient to conduct a deeper analysis of the answers to open-ended questions, and at the same time provide a basis for judgment for the subsequent consistency analysis of the text answers to semi-open-ended and open-ended questions.

[0081] S35, topic classification coding: Quantify the topics determined by the LDA model, select the number of digital coding bits, and encode the topics.

[0082] S4. Sentiment polarity analysis: Use dictionary-based and rule-based methods to perform sentiment polarity analysis to obtain the sentiment tendency of the open-ended question answer text, such as Figure 3 As shown, the following steps are included:

[0083] S41, weight allocation of sentiment words: the sentiment words in the sentiment dictionary are divided into positive sentiment word list and negative sentiment word list, and the pre-processed text is matched with the positive and negative sentiment word lists in the dictionary one by one, and the corresponding weight is assigned to the word if the match is successful;

[0084] The weight of positive sentiment is set to 1, and the weight of negative sentiment is set to -1. The sentiment polarity value is obtained by sequentially adding up the number of sentiment words in the text.

[0085] S42. Weight allocation of negative words: The weight of negative words is set to -1. When a negative word is matched, the sentiment polarity value is multiplied by -1. For multiple negations, the values ​​are multiplied sequentially.

[0086] It is understandable that a negative word is a word that expresses a negative meaning. Its appearance can reverse the sentiment polarity of a sentence and make its meaning completely opposite to the original. A positive sentiment word plus a negative word will turn the sentiment polarity value into a negative value. Similarly, a negative sentiment word plus a negative word will turn its sentiment polarity value into a positive value.

[0087] S43. Weight distribution of degree adverbs: Degree adverbs are divided into four levels according to their strength, and the weights are set to 2.0, 1.5, 0.8, and 0.5 from strong to weak, representing the strength of emotional polarity.

[0088] Adverbs of degree are a type of adverb that can indicate the strength of a sentiment. The addition of adverbs of degree can greatly expand the amount of information in the original text.

[0089] S44. Position weight distribution of degree adverbs and negative words: When both degree adverbs and negative words appear in a text sentence, weight distribution is performed according to the position relationship between the degree adverbs and negative words:

[0090] When the situation of "negative word + degree adverb + sentiment word" appears, the negative sentiment tendency will be weakened, and the position weight parameter is set to 0.5;

[0091] When the situation of "degree adverb + negative word + sentiment word" appears, the negative sentiment tendency will be strengthened, and the position weight parameter is set to 2.

[0092] It is understandable that when adverbs of degree and negative words appear in a sentence at the same time, their different positions will also affect the emotional tendency. Adverbs of degree before negative words will strengthen the original negative emotional tendency, while adverbs of degree after negative words will weaken the original negative emotional tendency.

[0093] S45, calculating the weighted value of sentiment polarity: after determining the weights of sentiment words, negative words, and degree adverbs in the answer text, perform a weighted sentiment polarity operation on each topic in the answer text to obtain the sentiment polarity value of each topic in each text, and then make a judgment on the sentiment tendency:

[0094] The answer text S is divided into n short sentences, and the sentiment polarity value of each short sentence is calculated based on the topic of the short sentence. When sentiment words, degree adverbs and negative words appear in the short sentence:

[0095] ;

[0096] Where W is the sentiment polarity value, E is the weight of sentiment words, N is the weight of negative words, D is the weight of degree adverbs, and a is the position weight parameter.

[0097] It should be noted that when the sentence does not contain topic words and sentiment words, the sentiment polarity value is set to 0, indicating neutral sentiment. According to the calculation results of the sentiment polarity value, there are three cases:

[0098] When the sentiment polarity value is greater than 0, it indicates positive sentiment; when the sentiment polarity value is equal to 0, it indicates neutral sentiment; when the sentiment polarity value is less than 0, it indicates negative sentiment.

[0099] It should be noted that the sentiment dictionary used in this embodiment is a combination of the HowNet dictionary and other sentiment dictionaries, including a positive sentiment word list, a negative sentiment word list, a negation word list and a degree adverb list.

[0100] S46, emotion polarity coding: emotion polarity is divided into five categories according to the emotion polarity value, and the coding range is set to 1-5, which is represented as follows: 5 represents relatively strong positive emotion, 4 represents normal positive emotion, 3 represents neutral emotion, 2 represents negative emotion, and 1 represents relatively strong negative emotion.

[0101] It can be understood that when the sentiment polarity value is less than or equal to -2.5, the sentiment number is 1; when the sentiment polarity value is greater than -2.5 and less than 0, the sentiment number is 2; when the sentiment polarity value is equal to 0, the sentiment number is 3; when the sentiment polarity value is greater than 0 and less than 2.5, the sentiment number is 4; when the sentiment polarity value is greater than or equal to 2.5, the sentiment number is 5.

[0102] S5. Text intent recognition: Text intent prediction is performed based on the extracted topic words, sentiment polarity, and answer consistency results, such as Figure 4 As shown, the following steps are included:

[0103] S51. Answer consistency calculation: measure the knowledge association between the topics of semi-open questions and open questions, and calculate the cosine similarity:

[0104] ;

[0105] In the formula, Indicates the subject and The cosine similarity of The word vector representing the topic of the answer text of the semi-open question, The word vector representing the topic composition of the answer text of the open-ended question; and Represents various features of two word vectors.

[0106] It can be understood that the similarity value is between 0 and 1, where 0 means that the two topics are completely dissimilar, that is, there is no topic association relationship; when the value approaches 1, it means that the two topics are highly similar in semantics and have a strong knowledge association.

[0107] S52. Construct a data set: set a selection threshold for answer consistency, filter text data that meets the selection threshold based on the similarity value of answer consistency, merge the three features of topic words, sentiment polarity, and answer consistency corresponding to the filtered text into a unified feature vector, set the intent label accordingly, and construct an intent recognition data set.

[0108] Among them, the setting of intent labels is based on the classification topic of the text, which contains multiple keywords. The intent category of the text is judged by multiple keywords, and the intent category is defined based on the questionnaire requirements. For example, for a product satisfaction survey, the intent categories that can be defined include "product satisfaction", "product dissatisfaction", "suggested improvement", "complaint", etc.

[0109] S53. Establish a machine learning model: divide the data set into a training set and a test set, establish a TextCNN model, concatenate the sentiment polarity and answer consistency vectors with the topic vector through the pooling layer of CNN, and input them into the fully connected layer for Softmax classification.

[0110] Combine the keywords and intent labels into the information pattern required by the model and train the model.

[0111] TextCNN (Text Convolutional Neural Network) is a deep learning model for text classification and text feature extraction. Compared with the traditional CNN model network, the overall structure of TextCNN has become simpler. TextCNN contains embedding layer, convolution layer, pooling layer and fully connected layer. The model converts text into word vectors through the embedding layer, then extracts local features of different sizes through the convolution layer, and then performs feature dimensionality reduction through the pooling layer, and finally performs classification or other subsequent tasks through the fully connected layer. The training process of the entire model usually uses the back propagation algorithm for parameter optimization.

[0112] Embedding layer: used to map discrete words in the text to continuous word vector representations. The role of this layer is to convert text data into a vector form that can be understood and processed by the machine. A common method is to use a pre-trained word vector model to map each word to its corresponding word vector.

[0113] Convolutional layer: The convolutional layer is one of the core components of TextCNN. It uses convolution kernels of different sizes to perform sliding window convolution operations on the input word vector sequence. Each convolution kernel extracts local features within the sliding window range. Through multiple convolution kernels, local feature information of different lengths can be captured.

[0114] Pooling layer: The pooling layer is used to reduce the dimension of the features extracted by the convolution layer and extract the most important features. Common pooling operations are maximum pooling or average pooling, which reduces the dimension of the feature vector by selecting the maximum value or calculating the average value on the features output by each convolution kernel.

[0115] Fully connected layer: The fully connected layer receives the feature vector output by the pooling layer as input and maps it to the final output category through a series of fully connected operations. This layer is usually composed of multiple fully connected neurons, each neuron is connected to all neurons in the previous layer, and the activation function of the last layer uses the Softmax function for multi-classification output.

[0116] In this embodiment, the sentiment polarity and answer consistency information are integrated into the intent classification model, enriching the input features of the fully connected layer, highlighting the influence of emotional features and answer status, optimizing the embedding layer word vector of TextCNN and expanding the text information features, and improving the accuracy of the intent prediction results. The answer consistency is calculated to detect the abnormal and contradictory answers of the respondents, thereby eliminating abnormal data and integrating the correlation between the two parts of the answers into the intent prediction model, enhancing the reliability of the data set, and utilizing the potential connections between the questions in the questionnaire to mine personalized suggestions.

[0117] S54. Model application: Use the cross-validation model performance and adjust the model parameters to optimize the prediction effect; use the trained model to predict the intent of new samples, input the feature vector of the new sample, and output the predicted intent category.

[0118] S6. Generate an analysis report: Encode and merge the identified intent labels with the subject words, sentiment polarity, and answer consistency to generate a model table of "topic classification + sentiment polarity + answer consistency + intent category" to quantify the results of the questionnaire answer text.

[0119] During the model training process, data with low text topic similarity are eliminated based on answer consistency. Therefore, the intent recognition results are not accurate when predicting data with low text topic similarity. However, this part of the data is not invalid data. Instead, it is targeted suggestions and services provided by respondents beyond the structured model. Therefore, this part of the data is marked by a preset selection threshold and prompted in the analysis report to dig out more effective survey results.

[0120] The present invention first converts the questions in the questionnaire survey into text descriptions in combination with the answer results, and also incorporates the answer results of semi-open questions into the text analysis, so as to mine the potential correlation of the answers of the respondents, and obtain the divergent thinking of the respondents in answering the open questions based on the semi-open question structure, so as to facilitate the mining of targeted answer suggestions; then the text data is sequentially subjected to topic extraction, sentiment polarity analysis and answer consistency calculation, and a data set is created for model training, which expands the text information features and improves the accuracy of the intention prediction results; finally, quantitative statistics are performed based on the generated "topic classification + sentiment polarity + answer consistency + intention category" model, which improves the statistical efficiency and accuracy of the questionnaire survey results, and at the same time, the questionnaire is improved based on the results of answer consistency, providing guidance for the subsequent questionnaire design.

[0121] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technical personnel in this field can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any brief modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A questionnaire survey result analysis method based on AI, characterized by: The following steps are involved: Collect data: merge the results of the semi-open questions into the questions to generate text, and construct the questionnaire answer text by combining the generated text with the answer text of the open questions; Extraction of keywords: The topic model is used to obtain the topic of the questionnaire answer text, and the text is segmented to generate the topic classification code for each short sentence; Sentiment polarity analysis: Use dictionary-based and rule-based methods to perform sentiment polarity analysis to obtain the sentiment tendency of the text of open-ended questions; Text intent recognition: predict text intent through the extracted topic words, sentiment polarity, and answer consistency results; Generate analysis report: Encode and merge the identified intent labels with subject words, sentiment polarity, and answer consistency to generate a model table of "topic classification + sentiment polarity + answer consistency + intent category" to quantify and count the results of the questionnaire answer text; The subject word extraction comprises the following steps: Text vectorization: Use the bag-of-words model or TF-IDF to convert the preprocessed text data into numerical form; Construct LDA model: Construct LDA model using preprocessed and vectorized data: Determine the number of topics in the model by using the perplexity in the model evaluation metric; α and β are set, where α Controls the sparsity of topic distribution in documents, and β controls the sparsity of word distribution in topics; Model training: Use the document-word matrix to train the LDA model, and estimate the document-topic distribution and topic-word distribution through an iterative process; Topic extraction: After model training is completed, the model is used to extract keywords for each topic and the topic distribution of each document: each topic consists of several keywords, and each document is represented as a topic distribution; Topic classification coding: Quantify the topics determined by the LDA model, select the number of digital coding digits, and encode the topics; The sentiment polarity analysis comprises the following steps: Weight allocation of sentiment words: sentiment words in the sentiment dictionary are divided into positive sentiment word list and negative sentiment word list, and the preprocessed text is matched with the positive and negative sentiment word lists in the dictionary one by one. If the match is successful, the word is assigned a corresponding weight; The weight of positive sentiment is set to 1, the weight of negative sentiment is set to -1, and the sentiment polarity value is obtained by stacking the number of sentiment words in the text; Weight allocation of negative words: The weight of negative words is set to -1. When a negative word is matched, the sentiment polarity value is multiplied by -1. For multiple negative words, the sentiment polarity value is multiplied in sequence. Weight distribution of degree adverbs: Degree adverbs are divided into four levels according to their degree, and the weights are set to 2.0, 1.5, 0.8, and 0.5 from strong to weak, representing the strength of emotional polarity; Position weight distribution of degree adverbs and negative words: When both degree adverbs and negative words appear in a short sentence, weight distribution is performed based on the position relationship between the degree adverbs and negative words; Calculate the weighted value of sentiment polarity: After determining the weights of sentiment words, negative words, and degree adverbs in the answer text, perform a weighted sentiment polarity operation on each topic in the answer text to obtain the sentiment polarity value of each topic in each text, and then make a judgment on the sentiment tendency: The answer text S is divided into n short sentences, and the sentiment polarity value of each short sentence is calculated based on the topic of the short sentence. When sentiment words, degree adverbs and negative words appear in the short sentence: ； Where W is the sentiment polarity value, E is the weight of sentiment words, N is the weight of negative words, D is the weight of degree adverbs, and a is the position weight parameter; When the sentiment polarity value is greater than 0, it indicates positive sentiment; when the sentiment polarity value is equal to 0, it indicates neutral sentiment; when the sentiment polarity value is less than 0, it indicates negative sentiment; Emotion polarity coding: Emotion polarity is divided into five categories according to the emotion polarity value. The coding range is set to 1-5, which is represented as follows: 5 represents relatively strong positive emotion, 4 represents normal positive emotion, 3 represents neutral emotion, 2 represents negative emotion, and 1 represents strong negative emotion; The text intent recognition comprises the following steps: Answer consistency calculation: measure the knowledge association between the topics of semi-open questions and open questions, and calculate the cosine similarity; Constructing a data set: Setting a selection threshold for answer consistency, filtering text data that meets the selection threshold based on the similarity value of answer consistency, merging the three features of topic words, sentiment polarity, and answer consistency corresponding to the filtered text into a unified feature vector, and setting the intent label accordingly to construct an intent recognition data set; Establish a machine learning model: divide the data set into a training set and a test set, establish a TextCNN model, concatenate the sentiment polarity and answer consistency vectors with the topic vector through the pooling layer of CNN, and input them into the fully connected layer for Softmax classification; combine the topic words and intent labels into the information pattern required by the model, and train the model; Model application: Use cross-validation model performance and adjust model parameters to optimize prediction results; use the trained model to predict the intent of new samples, input the feature vector of the new sample, and output the predicted intent category; Generating the analysis report also includes marking based on a preset answer consistency threshold, performing prompt processing in the analysis report, and mining a single answer result; The calculation formula of the perplexity is: ； In the formula, D For document sets; It is a document d The probability of all words appearing in ; N d is the number of words in the dth document; The formula for calculating cosine similarity is as follows: ； In the formula, Indicates the subject and The cosine similarity of The word vector representing the topic of the answer text of the semi-open question, The word vector representing the topic composition of the answer text of the open-ended question; and Represents various features of two word vectors.

2. The AI-based questionnaire survey result analysis method according to claim 1, characterized in that: The model training comprises the following steps: Build a document-word matrix: Count each word in the document to generate a vocabulary, and then build a matrix. The rows of the matrix represent documents, the columns represent words, and the values ​​in the matrix represent the number of times the word appears in the corresponding document. Training the LDA model: Use the LDA implementation tool and the LDA model in the Gensim library to train the document-word matrix; The training process involves selecting model parameters, including the number of topics and the number of iterations; the LDA model generates word distributions for each topic and topic distributions for each document by analyzing the document-word matrix.

3. The method for analyzing questionnaire results based on AI according to claim 1, characterized in that: The weight distribution according to the positional relationship between the degree adverb and the negative word includes: When "negative word + degree adverb + sentiment word" appears, the negative sentiment tendency will be weakened, and the position weight parameter is set to 0.5; When the situation of "degree adverb + negative word + sentiment word" appears, the negative sentiment tendency will be strengthened, and the position weight parameter is set to 2.

4. The method for analyzing questionnaire results based on AI according to claim 1, characterized in that: After the data collection step, data preprocessing is also included: segmenting the questionnaire answer text and removing stop words; And according to the purpose of the analysis, symbols, adverbs and function words with emotional connotations are removed or retained.

Images