Multi-industry research text data analysis model and method based on big data resources
Through the multi-level classification method of big data resources and ReBERTa model, the time-consuming and labor-intensive problem of artificial label formulation in research text data analysis is solved, and efficient and accurate label generation and analysis are achieved.
Patent Information
- Application Number
- CN202510531554.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-25
- Publication Date
- 2025-08-08
AI Technical Summary
The existing research text data analysis methods require artificial label development, which consumes a lot of work and has accuracy problems. Traditional word cloud diagrams and text classification methods cannot accurately understand the label proportion.
Using a multi-level classification method based on big data resources, combined with a standard encode model fine-tuned by ReBERTa, through data cleaning, clustering and similarity algorithms, matching labels are trained step by step and output to adapt to different fields and tasks.
It improves the efficiency and accuracy of research text data analysis, reduces the workload of artificial processing, and achieves fast and accurate label generation.
Smart Images
Figure CN120448542A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of survey data processing methods, and in particular to a multi-industry survey text data analysis model and method based on big data resources. Background Art
[0002] Survey text data consists of responses to open-ended questions, such as "What are your suggestions / opinions?", collected during market research. Survey text data is crucial for obtaining crucial information during the research process, and analyzing it is crucial for obtaining this information.
[0003] Traditional methods for analyzing survey text data typically employ two approaches. The first involves word cloud analysis, which reveals the most common keywords in respondents' answers. These keywords represent the core meaning of their responses. For example, the majority of respondents cited car purchase purposes as commuting, road trips, shopping, and child care (displayed in larger font). However, while word cloud analysis can provide a general understanding of the core meaning of respondents' answers, it cannot reveal the proportions, such as the percentage of respondents who chose "commuting." Therefore, a second approach, text classification, emerged. Text classification involves pre-defined categorization labels, such as commuting, shopping, and child care. Respondents' responses are then categorized into these labels. Finally, the proportion of these labels is calculated. For example, if there are 1,000 respondents and 2,000 labels are counted, a single respondent's answers may have multiple labels, with commuting accounting for 500 of them. Therefore, the proportion of commuting is 500 / 2,000 = 25%.
[0004] Although existing text classification can determine the proportion of each label, it requires labels to be formulated in advance. Due to the limitations of researchers, the formulated labels may be inaccurate or missing. In addition, the workload of classifying each piece of data will be enormous. Summary of the Invention
[0005] One of the purposes of the present invention is to provide a multi-industry survey text data analysis method based on big data resources to solve the problem of huge workload in survey text data analysis.
[0006] The multi-industry research text data analysis method based on big data resources in this solution includes:
[0007] S1: Clean the acquired open-ended question data, classify them into multiple levels according to the research industry, research direction, and research question, and conduct multi-level step-by-step training based on the preset model;
[0008] S2: Use clustering methods to perform unsupervised classification on all answers to any type of open questions, select data with high evaluation indicators, establish cluster labels for each open question, judge and name the cluster labels, and establish classification labels for each open question;
[0009] S3 extracts features based on a standard encoder model fine-tuned with ReBERTa and constructs multi-task classification tasks within the training tasks, adapting to various research fields and directions.
[0010] S4: Output the matching label information.
[0011] The beneficial effects of this program are:
[0012] The answer information of open questions is cleaned, and the answer to each open question is classified, and classification labels are established. Finally, feature extraction is performed to adapt to the classification of different fields and directions, and the label information of the answer information of the open questions is output. This can improve the speed and efficiency of processing the answer information of open questions and reduce the workload caused by manual processing.
[0013] Furthermore, in S1, the data cleaning step is:
[0014] S1.1, process historical case data and open-ended question data to obtain normal data, abnormal emotional data, and random answer data;
[0015] S1.2: Construct a classification algorithm, a clustering algorithm, and a similarity algorithm, and use these three algorithms to label the normal data, the emotionally abnormal data, and the random answer data in S1.1 respectively;
[0016] S1.3, the label with the most votes for each open question obtained by the three algorithms in S1.2 is taken as the final label.
[0017] The beneficial effects are: the data is cleaned using different algorithms, and the label with the most votes is used as the final label, avoiding errors in a certain algorithm for a certain open-ended question, and filtering out content that is randomly answered due to users' negative emotions towards the open-ended questions, thereby reducing subsequent analysis errors caused by abnormal answers to the open-ended questions themselves.
[0018] Furthermore, in S1.2, the similarity algorithm is:
[0019] The encoding model based on ReBERT fine-tuning is used to extract the data features of the case data, and the cosine similarity algorithm is used to calculate the similarity between the data features of the data and the data features of the case data. The vector obtained by embedding it into the vector converted by the encoding model based on ReBERT fine-tuning is stored in the Es database.
[0020] The beneficial effect is that the calculation of the similarity algorithm can accurately judge the answer data of the open questions for accurate classification.
[0021] Furthermore, in S1.3, when three algorithms obtain three different labels for the same open question, the result of the clustering algorithm shall prevail.
[0022] The beneficial effect is that for different labels obtained by the three algorithms, the accuracy of the labeling results can be improved by taking the results of the clustering algorithm as the standard.
[0023] Furthermore, in S1, the classification levels of the open questions include level one, level two, level three, ..., level N, and the level-by-level combined training is level one, level two, level one, level two, level three, ..., and so on.
[0024] The beneficial effect is that the classification level of open questions can cover multiple levels of problems, making the classification more accurate.
[0025] Furthermore, in S3, different prompt words are used for different fields and different tasks, and the prompt words of different fields and different tasks are combined to perform classification processing on multiple tasks simultaneously.
[0026] The beneficial effect is that by setting prompt words to simultaneously carry out classification processing of different fields and different tasks, and combining different fields and different tasks, it is possible to quickly process the huge amount of big data and improve processing speed and efficiency.
[0027] Furthermore, in S3, the emotional color of the answers to the open questions is judged, and the answers to the open questions are divided into three categories: positive, negative and neutral.
[0028] The beneficial effect is that the emotional color classification of the answers to the open questions can accurately judge the user's answers to the open questions, so as to facilitate the subsequent accurate analysis of the answers to the open questions.
[0029] Furthermore, in S4, the similarity between the output label and the given label is calculated using similarity, and the label with the highest similarity is used as the output result label of the data.
[0030] The beneficial effect is that: for the problem that the output result after label classification does not meet the predetermined requirements, similarity calculation is performed, and the one with the highest similarity is used as the output result, thereby improving the probability that the output matches the predetermined requirements.
[0031] A second purpose of the present invention is to provide a multi-industry survey text data analysis model based on big data resources to reduce the workload of survey text data analysis.
[0032] The multi-industry survey text data analysis model based on big data resources is obtained according to the above analysis method. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1 This is a flowchart of an embodiment of a multi-industry survey text data analysis method based on big data resources;
[0034] Figure 2 This is the classification label corresponding to the venue category in the embodiment of the multi-industry survey text data analysis method based on big data resources. DETAILED DESCRIPTION
[0035] The following is further explained in detail through specific implementation methods.
[0036] Example 1
[0037] Multi-industry research text data analysis methods based on big data resources, such as Figure 1 Shown, including:
[0038] S1. Data cleaning is performed on the obtained open-ended question data. The open-ended question data is the answers to the open-ended question. For example, the answer to the open-ended question "What are your suggestions / opinions on the use of vehicle A?" is the open-ended question data. The data cleaning steps are as follows:
[0039] S1.1, process the historical case data and open-ended question data to obtain normal data, abnormal emotional data, and random answer data. The historical case data (i.e., original data) comes from the open-ended question data of the survey text data.
[0040] S1.2, construct a classification algorithm, a clustering algorithm and a similarity algorithm, and use the three algorithms to add labels to the normal data, abnormal emotional data and random answer data in S1.1 respectively. The classification algorithm uses a deep learning algorithm to construct a classification algorithm based on the above three types of label data. After each piece of data passes through the classification algorithm, it will directly obtain a label data, such as a label of none or random answer. The clustering algorithm is: KMeans algorithm using cosine distance as the distance metric, and uses the above label data to construct a 3-category clustering algorithm. After each piece of data passes through the clustering model, it will directly obtain a label data.
[0041] The similarity algorithm is:
[0042] A coding model based on ReBERT fine-tuning is used to extract the data features of the case data. The cosine similarity algorithm is used to calculate the similarity between the data features of the data and the data features of the case data. The cosine similarity algorithm calculates the similarity between the label and the text. Clustering algorithm: kmeans algorithm with cosine distance is used. The data is processed in batches. Pre-encoding: encoding is performed when the text enters the database. During calculation, vectors can be directly extracted from the es database.
[0043] S1.3: The label with the most votes for each open question obtained by the three algorithms in S1.2 is used as the final label. When the three algorithms obtain three different labels for the same open question, the result of the clustering algorithm shall prevail.
[0044] The open questions are classified into multiple levels according to the research industry, research direction and research questions, and multi-level step-by-step combined training is performed based on the preset model. The classification levels of the open questions include level one, level two, level three, ..., level N, and the step-by-step combined training is level one, level two, level one, level two, level three, ..., and so on. The multi-level classification in this embodiment is three-level, and the multi-level classification is classified according to industry, research direction, and specific questions, such as research direction (level one), specific industry (level two), and specific questions (referring to open questions, level three). Research direction refers to the type of research, such as satisfaction, user evaluation, and preliminary research; specific industry refers to which industry the research belongs to, such as public transportation, rail, and hospital; specific questions refer to the specific content of the inquiry, such as opinions, suggestions, and evaluation feelings.
[0045] For each industry, perform the following sub-steps:
[0046] Sub-step 1: Divide the survey text data into different industries, such as automobiles and public transportation. Because the data was collected by different industries, this part of the work has been completed in advance.
[0047] Sub-step 2: Use the encoding model based on ReBERT fine-tuning, the kmeans clustering algorithm, and cosine similarity to calculate the similarity of the text. Select an exception text that meets the preset conditions as the similar text of the text, and use this similar text as the positive sample. The preset condition is that the similarity > 95% is the highest similarity. The specific operation includes the following steps:
[0048] A. First, convert the text into vectors based on the ReBERT fine-tuned encoding model, for example, into a dense vector group such as (0.11234, 1.21245233, ..., 0.2124552);
[0049] B. The cluster centers of the first iteration of the original kmeans clustering algorithm are randomly generated, which has great randomness and consumes a lot of computing resources. In addition, the class labels need to be re-determined after clustering. Therefore, the following method is used to perform cluster analysis according to the required scenario:
[0050] B.1 Use the kmeans clustering algorithm with cosine distance to perform preliminary clustering on the preset proportion of text in sub-step 1 and obtain the center points of each cluster. This step solves the problem of random generation of cluster centers and the need to re-determine the class labels after clustering.
[0051] B.2. Iteratively update the cluster centers using a preset percentage of the remaining text from sub-step 1. Calculate the distance from each cluster center to the remaining text from sub-step 1, assign it to the point closest to it, and then update the cluster centers. Repeat this process multiple times, updating each cluster center, to obtain the optimal cluster center, which is 70% of the text from sub-step 1.
[0052] The preset ratio is 30%. By setting the preset ratio, a good balance can be achieved between reducing computing pressure and machine resources and improving accuracy.
[0053] C. In each cluster, select the text with the highest similarity of >95% as the positive sample, and the one with the lowest similarity as the negative sample;
[0054] D. Update negative samples: Randomly extract a text from other clusters to calculate the similarity, and compare it with the similarity of the negative sample, and select the one with the lowest similarity as the negative sample.
[0055] All the obtained texts are processed according to sub-step 2 to obtain positive samples and negative samples.
[0056] S2, use clustering method to perform unsupervised classification on all answers of any type of open questions, select some data with high evaluation index, and use Kmeans clustering, a commonly used unsupervised method, to select some data with high evaluation index at the same time, to reduce the difference in data quality caused by different researchers' qualities, and establish cluster labels for each open question, judge and name the cluster labels, and establish classification labels for each open question, such as Figure 2 The classification labels for the different "venue categories" are shown.
[0057] After completing the classification of the classification labels, the clustered labels are judged and named by obtaining the classification information of the confirmed classification to establish the classification labels, thereby reducing the number of standard classification labels for a specific open question. Subsequent projects that belong to the same category, such as "Satisfaction - Hospital - Environment", can all use the classification labels of the open question.
[0058] S3, based on the standard encoder model fine-tuned by ReBERTa, extracts features, fine-tunes in advance on data related to the research industry, and constructs multi-task classification tasks in the training tasks to adapt to different research fields and directions. It simultaneously classifies and processes multiple tasks for different fields, different tasks, and combines prompt words of different fields and different tasks. When classifying different fields, a classification model is used to achieve the classification of different tasks. The specific approach is: classify the description sentences of the field and assign different classification tasks according to the classification results. Because the changes in each description sentence are not large, this classification model can achieve an accuracy of more than 98%; such as realizing the medical field: "This is a medical field"; realizing the bus field: "This is a bus field"; when classifying different tasks, such as realizing general classification tasks: "classification tasks"; realizing sentiment classification tasks: "sentiment classification tasks"; when combining different fields with different tasks, such as realizing general medical classification tasks: "This is a classification task related to the medical field"; realizing medical sentiment classification tasks: "This is a sentiment classification task related to the medical field", a special character " is used between task descriptions, texts, and labels.<separator\> " to separate; when implementing common classification tasks in the medical field, examples of labeled data and new data are as follows:
[0059] Original data: "This is a classification task in the medical field <separator>Text: xxxxxxxx <separator>Label: xx; This is a classification task related to the medical field <separator>Text: xxxxxxxx <separator>Label: xx; This is a classification task related to the medical field <separator>Text: xxxxxxxx <separator>Label: xx".
[0060] New data: "This is a classification task in the medical field <separator>Text: xxxxx <separator>Text: xxxxx;".
[0061] The emotional color of the answers to the open-ended questions is judged and divided into three categories: positive, negative and neutral. The sentiment classification task is also included in the above classification model.
[0062] Model training, based on the ReBERTa model, using the data from step 1 for fine-tuning includes the following:
[0063] First, given a main sentence a, a positive sentence p, and a negative sentence n, the network is adjusted by triple loss so that the distance between the main sentence a and the positive sentence p is smaller than the distance between the main sentence a and the negative sentence n. Mathematically, the loss function is represented by minimizing the loss function as follows:
[0064] Loss=max(||s a -s p ||-||s a -s n ||+ε,0);
[0065] s a ,s p ,s n The sentence embeddings representing the main sentence a, positive sentence p, and negative sentence n are vectors converted from the encoding model fine-tuned based on ReBERT, with distance represented by ||·|| and edge parameter ε representing s p With s a The distance is at least greater than s n Close to ε, we get the fine-tuned ReBERTa model.
[0066] Use this fine-tuned ReBERTa model as the similarity extraction model.
[0067] For data storage, we use the open-source distributed database es (Elasticsearch) to store text and the aforementioned ReBERTa model to extract vectors. The es database allows for simultaneous storage and querying of text and vectors without sacrificing performance, achieving improved performance and data management efficiency. S4 outputs the matched tag information and uses similarity calculations to compare the output tag with the given tag. The tag with the highest similarity is used as the output tag for the data.
[0068] Compared with the existing technology, this embodiment cleans the answer information of open questions, and during the cleaning process, it uses a variety of different algorithms to clean, and takes the label with the most votes as the final label, avoiding the error of a certain algorithm in a certain open question, and filtering out the content obtained by random answers due to the user's negative emotions towards the open questions, reducing the subsequent analysis errors caused by abnormal answers to the open questions themselves, and classifying the answers to each open question, establishing classification labels, and finally performing feature extraction to adapt to classifications in different fields and directions, and outputting label information of the answer information of the open questions, which can improve the speed and efficiency of processing the answer information of the open questions and reduce the workload caused by manual processing.
[0069] Example 2
[0070] A multi-industry survey text data analysis model based on big data resources is based on the analysis method of embodiment one, and an analysis model is obtained according to the analysis method of embodiment one.
[0071] The above is only an embodiment of the present invention, and the common knowledge such as the specific structure and characteristics of the scheme is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several variations and improvements can be made, which should also be regarded as the scope of protection of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.< / separator> < / separator> < / separator> < / separator> < / separator> < / separator> < / separator> < / separator>
Claims
1. A multi-industry research text data analysis method based on big data resources, characterized by: include: S1: Clean the acquired open-ended question data, classify them into multiple levels according to the research industry, research direction, and research question, and conduct multi-level step-by-step training based on the preset model; S2: Use clustering methods to perform unsupervised classification on all answers to any type of open questions, select data with high evaluation indicators, establish cluster labels for each open question, judge and name the cluster labels, and establish classification labels for each open question; S3 extracts features based on a standard encoder model fine-tuned with ReBERTa, and constructs multi-task classification tasks within the training tasks to adapt to various research fields and directions; S4: Output the matching label information.
2. The multi-industry research text data analysis method based on big data resources according to claim 1 is characterized by: In S1, the data cleaning steps are: S1.1, process historical case data and open-ended question data to obtain normal data, abnormal emotional data, and random answer data; S1.2: Construct a classification algorithm, a clustering algorithm, and a similarity algorithm, and use these three algorithms to label the normal data, the emotionally abnormal data, and the random answer data in S1.1 respectively; S1.3, the label with the most votes for each open question obtained by the three algorithms in S1.2 is taken as the final label.
3. The multi-industry research text data analysis method based on big data resources according to claim 2 is characterized by: In S1.2, the similarity algorithm is: The encoding model based on ReBERT fine-tuning is used to extract the data features of the case data. The cosine similarity algorithm is used to calculate the similarity between the data features of the data and the data features of the case data. The data is then embedded in the vector converted by the encoding model based on ReBERT fine-tuning and stored in the ES database.
4. The multi-industry research text data analysis method based on big data resources according to claim 2 is characterized by: In S1.3, when three algorithms obtain three different labels for the same open question, the result of the clustering algorithm shall prevail.
5. The multi-industry research text data analysis method based on big data resources according to claim 1 is characterized by: In S1, the classification levels of the open questions include level one, level two, level three, ..., level N, and the level-by-level combined training is level one, level two, level one, level two, level three, ..., and so on.
6. The multi-industry research text data analysis method based on big data resources according to claim 1 is characterized by: In S3, different prompt words or description sentences are used for different fields and different tasks, and the prompt words or description sentences of different fields and different tasks are combined to perform classification processing on multiple tasks simultaneously.
7. The multi-industry research text data analysis method based on big data resources according to claim 1 is characterized by: In S3, the emotional color of the answers to the open questions is judged, and the answers to the open questions are divided into three categories: positive, negative and neutral.
8. The multi-industry research text data analysis method based on big data resources according to claim 1 is characterized by: In S4, the similarity between the output label and the given label is calculated using the similarity, and the label with the highest similarity is used as the output result label of the data.
9. A multi-industry research text data analysis model based on big data resources, characterized by: Obtained according to the analytical method according to any one of claims 1 to 8.