Emotion recognition method for large model in combination with geographic information and internet surfing behaviors

Abstract

The invention discloses an emotion recognition method for a large model in combination with geographic information and internet surfing behaviors, and relates to the technical field of emotion recognition, and the method comprises the steps: S1, data collection and integration; s2, preprocessing the data; s3, feature extraction; s4, training and applying a large model; and S5, evaluating and optimizing a result. According to the method, the limitation of a single data source is overcome, and the emotional state of the student is comprehensively reflected by integrating multi-source data; a large model can be used for understanding context information, emotion can be more accurately interpreted, and the influence of geographical, social and cultural backgrounds on emotion expression and perception is fully considered; student emotion changes can be tracked through a continuous learning and feedback mechanism, the model prediction ability is improved, and the method adapts to emotion dynamic changes.

Description

Technical Field

[0001] The present invention relates to the technical field of emotion recognition, and specifically, to an emotion recognition method combining a large model with geographical information and Internet behavior. Background Art

[0002] The management of college students by colleges and universities covers various aspects such as academic performance, psychology, code of conduct, and career development guidance. It is necessary to pay attention to the unique needs of students based on their individual emotional differences and help students grow through communication and guidance. Students will face various emotional problems during their growth, such as loneliness, anxiety, friendship problems, love problems, etc. For this reason, schools have carried out work such as course selection guidance, performance evaluation, psychological counseling, formulation of rules and regulations, and supervision of club activities.

[0003] At present, the emotional changes of students can be manifested through staying at specific locations and Internet behavior. Combining geographical and Internet behavior to identify emotional needs is of certain significance. However, affected by factors such as geographical environment, social environment, cultural background, social media use, and online activity forms, the data volume is huge, and it is difficult to quickly judge the emotional changes of students.

[0004] There are certain deficiencies in existing emotion recognition methods. For example, when using data deep mining and big data technology to depict the behavior trajectories of students, there are defects in emotion recognition. The emotional state is difficult to capture through behavior data, and its multidimensionality and rapid variability make it impossible for a single trajectory dimension to comprehensively reflect, and traditional big data analysis is difficult to process qualitative data (such as emotions, perceptions, etc.).

[0005] Based on this, a method for emotion recognition combining a large model with geographical information and Internet behavior is now provided, which can eliminate the drawbacks of existing methods. Summary of the Invention

[0006] The purpose of the present invention is to provide an emotion recognition method combining a large model with geographical information and Internet behavior to solve the problems of the shortcomings of modern methods in the background art.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] An emotion recognition method combining a large model with geographical information and Internet behavior, comprising the following steps:

[0009] Step S1: Data collection and integration: Collect the geographical information and Internet behavior data of users. The geographical information includes, but is not limited to, the longitude and latitude where the user is located, and geographical location tags;

[0010] The Internet behavior data includes the websites visited by the user, the stay time on each website, the content of the browsed pages, the interaction behavior on social platforms, and the usage of application programs; Integrate the collected geographical information and Internet behavior data to form a multi-source data set;

[0011] Step S2: Data preprocessing: Clean the integrated multi-source data set to remove noise data, duplicate data, and error data; normalize the data to make different types of data comparable; perform natural language processing operations such as word segmentation and part-of-speech tagging on the online behavior data of the text type.

[0012] Step S3: Feature extraction: Extract features related to sentiment recognition from the preprocessed data. For geographical information, extract features of the geographical location, such as the number of people and the type of venue; for online behavior data, extract features such as the user's browsing preferences and social activity, and combine the extracted features into a feature vector.

[0013] Step S4: Large model training and application: Use the training data set with sentiment labels to train the large model. The large model is a pre-trained model with strong semantic understanding and feature learning capabilities, and a model based on the Transformer architecture can be used; input the feature vector obtained through feature extraction into the trained large model, and through the calculation and analysis of the large model, output the sentiment recognition result of the user. The sentiment recognition result includes but is not limited to sentiment categories such as positive, negative, and neutral.

[0014] Step S5: Result evaluation and optimization: Use evaluation metrics to evaluate the sentiment recognition result; adjust and optimize the parameters of the large model according to the evaluation result, or improve the feature extraction method.

[0015] Based on the above technical solutions, the present invention also provides the following alternative technical solutions:

[0016] In an alternative solution: In the step S1, the collection methods of geographical information include obtaining through the GPS positioning system of the user's mobile device, obtaining through Wi-Fi positioning technology, and obtaining from a third-party geographical information service provider; the collection methods of online behavior data include collecting through browser plugins and the buried point technology of mobile applications.

[0017] In an alternative solution: In the step S2, when performing natural language processing operations on text type data, operations such as removing stop words and performing stemming are also included.

[0018] In an alternative solution: In the step S3, on the basis of extracting features related to sentiment recognition, time features are combined, such as the geographical information and online behavior data of the user at different time periods.

[0019] In an alternative solution: in step S4, when training the large model, a transfer learning method is adopted, and the knowledge learned by the pre-trained model on a large-scale corpus is used to fine-tune the model in combination with a small amount of training data with sentiment labels.

[0020] In an alternative solution: in step S5, when training the large model, manually annotated samples are introduced to verify and calibrate the sentiment recognition results.

[0021] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0022] 1. By overcoming the limitations of a single data source, the present invention comprehensively reflects the emotional state of students by integrating multi-source data.

[0023] 2. The present invention uses a large model to understand context information, interprets emotions more accurately, and fully considers the influence of geographical, social, and cultural backgrounds on emotional expression and perception.

[0024] 3. Through continuous learning and feedback mechanisms, the present invention can track the emotional changes of students, improve the model's prediction ability, and adapt to dynamic emotional changes.

[0025] 4. The present invention can achieve personalized emotion recognition, meet the individual differences of students, and assist schools in precise management and guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the project algorithm flowchart of the present invention.

[0027] Figure 2 It is the schematic diagram of the student emotion recognition system of the present invention.

[0028] Figure 3 It is the schematic diagram of the learning and feedback mechanism of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0029] In order to make the objectives, technical solutions and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0030] In one embodiment, as Figures 1 - 3 shown, an emotion recognition method combining a large model with geographical information and Internet behavior includes the following steps:

[0031] Step S1: Data collection and integration: Collect the geographical information and Internet behavior data of users. The geographical information includes, but is not limited to, the longitude and latitude where the user is located and the geographical location label;

[0032] Online behavior data includes the websites visited by users, the time spent on each website, the content of the pages viewed, the interactive behavior on social platforms, and the use of applications; the collected geographic information and online behavior data are integrated to form a multi-source data set;

[0033] Step S2: Data preprocessing: Clean the integrated multi-source data set to remove noise data, duplicate data, and erroneous data; normalize the data to make different types of data comparable; perform natural language processing operations such as word segmentation and part-of-speech tagging on text-based online behavior data;

[0034] Step S3: Feature extraction: Extract features related to emotion recognition from the preprocessed data. For geographic information, extract features of geographic location, such as flow of people, type of venue, etc.; for online behavior data, extract features such as user browsing preferences and social activity, and combine the extracted features into feature vectors.

[0035] Step S4: training and application of the big model: using a training data set with emotion labels to train the big model, the big model is a pre-trained model with strong semantic understanding and feature learning capabilities, and a model based on the Transformer architecture can be used; the feature vector obtained by feature extraction is input into the trained big model, and the emotion recognition result of the user is output through calculation and analysis of the big model, and the emotion recognition result includes but is not limited to positive, negative, neutral and other emotion categories;

[0036] Step S5: Result evaluation and optimization: Use evaluation indicators to evaluate the emotion recognition results; adjust and optimize the parameters of the large model or improve the feature extraction method according to the evaluation results to improve the accuracy of emotion recognition.

[0037] In one embodiment, in step S1, the geographic information is collected by the GPS positioning system of the user's mobile device, by Wi-Fi positioning technology, or from a third-party geographic information service provider; the Internet behavior data is collected by browser plug-ins and mobile application tracking technology.

[0038] In one embodiment, in step S2, when performing natural language processing on text data, it also includes removing stop words, performing stem extraction, etc. Reduce the interference of irrelevant information on emotion recognition.

[0039] In one embodiment, in step S3, the time features, such as the user's geographic information and online behavior data in different time periods, are combined on the basis of extracting the features related to emotion recognition to more comprehensively reflect the user's emotional state.

[0040] In one embodiment, in step S4, when training the large model, a transfer learning method is adopted. The knowledge learned by the pre-trained model on a large-scale corpus is used to fine-tune the model in combination with a small amount of training data with sentiment labels, improving the training efficiency and model performance.

[0041] In one embodiment, in step S5, when training the large model, manually annotated samples are introduced to verify and calibrate the sentiment recognition results, further improving the reliability of the recognition results.

[0042] The above embodiment discloses a sentiment recognition method combining a large model with geographical information and Internet access behavior. Its specific working principle and process are as follows:

[0043] I. Multidimensional data integration embodiment:

[0044] Suppose we want to analyze the emotional state of students in a certain university on campus. Obtain their location information in places such as classrooms, cafeterias, and libraries through the GPS positioning of students' mobile phones, and use Wi-Fi positioning to assist in precise positioning. Obtain the websites visited by students at these locations through network monitoring, such as learning websites, social platforms, etc., and record the stay time and click-through rate. On social platforms, collect students' comments, likes, and sharing behaviors, and count the frequencies of using various applications such as learning, entertainment, and social networking. Organize and store these data according to the specified data fields (user ID, session ID, etc.) to form a multidimensional data set.

[0045] II. Data context understanding embodiment:

[0046] (I). Build an AI large model framework based on Xinference and Langchain, and locally deploy the Alibaba Tongyi Qianwen large model (qwen-chat 7B) and vector model (bge-base-zh) to construct an AI-based knowledge base system:

[0047] The detailed steps are as follows:

[0048] 1. Deploy the Qwen-Chat 7B model

[0049] Obtain the model

[0050] You need to obtain the qwen-chat 7B model weights from Alibaba Cloud or other official channels. Ensure that the model is used in accordance with the relevant license terms.

[0051] Model loading

[0052] Use Xinference to load the model through the following code:

[0053] python

[0054] from xinference import Model # Load the Qwen - Chat model model = Model("qwen - chat - 7B")

[0055] 2. Deploy the BGE - base - zh vector model

[0056] Download the vector model

[0057] Make sure to download the bge - base - zh model and place it in an accessible path.

[0058] Load the vector model

[0059] Load the vector model using the same method:

[0060] python

[0061] vector_model = Model("bge - base - zh")

[0062] 3. Build a knowledge base system

[0063] Data preparation

[0064] Collect and organize the data required for the knowledge base, such as documents, Q&A pairs, articles, etc., to ensure the validity and integrity of the data. Pre - process the data and convert it into a suitable format (such as JSON, CSV).

[0065] Vectorization

[0066] Use the BGE model to convert text data into vectors for subsequent retrieval.

[0067] python

[0068] def vectorize_documents(documents):

[0069] vectors = []

[0070] for doc in documents:

[0071] vector = vector_model.encode(doc) # Vectorize each document vectors.append(vector)

[0072] return vectors

[0073] documents = ["Document Content 1", "Document Content 2", "Document Content 3"] document_vectors = vectorize_documents(documents)

[0074] Build a question-answering system

[0075] Accept user input

[0076] Use Langchain to build an interactive question-answering system that accepts user questions.

[0077] Vectorize the question

[0078] Vectorize the question entered by the user for comparison with the documents in the knowledge base

[0079] Python

[0080] user_query = "User's question"

[0081] query_vector = vector_model.encode(user_query)

[0082] Similarity search

[0083] Use similarity calculation (such as cosine similarity) to find the content in the knowledge base similar to the user's question.

[0084] python

[0085] def find_similar_documents(query_vector, document_vectors):

[0086] similarities = []

[0087] for vector in document_vectors:

[0088] similarity = compute_similarity(query_vector, vector) # Calculate similarity similarities.append(similarity)

[0089] return similarities

[0090] similar_docs = find_similar_documents(query_vector, document_vectors)

[0091] Return result

[0092] Return the most relevant documents and answers based on similarity;

[0093] (2) Use Qwen-Agent intelligent agent technology to achieve precise recognition of students' emotions through the prompt+tools method and complete API calls. The prompt designs effective user input prompts to ensure that users can clearly express their needs. For example:

[0094] "Please output the emotional changes that students will show when they regularly go shopping at a certain mall after school and simultaneously access a certain online APP", and then use Tools to combine relevant tools for data processing and analysis. For example, use tools to obtain students' online behaviors and geographically process the spatial locations where the behaviors occur, extract spatial coordinate data, and simultaneously associate cases of students' emotional changes when shopping in this mall.

[0095] 1. Design effective prompts

[0096] Multi-round dialogue: Based on the user's initial input, conduct multi-round dialogues to gradually guide the user to provide more details. For example: Initial question: The user hopes to understand certain data. Follow-up question: Can ask about the type of data (such as location, sentiment analysis, statistical information, etc.).

[0097] Selection and integration of tools

[0098] According to the user's final appeal, select appropriate tools for data processing or task execution.

[0099] These tools include:

[0100] Data analysis tools: Through neural network statistical analysis, data visualization, etc. for cases of students' emotional changes and students' online behaviors at specific geographical locations.

[0101] API calls: For example, access Baidu Map's geographic location service, weather data, social media information, etc. Natural language processing tools: Parse the emotions, intentions, or keywords in the text.

[0102] Implementation process The following are the steps to implement a specific example. Assume that the user wants to analyze emotional information based on data of certain locations:

[0103] Receive user input User input: "I want to know the emotional changes of students who use a certain online APP during class in a certain classroom recently."

[0104] Identify requirements and design dialogues

[0105] The program can parse this command and confirm:

[0106] Location: A certain classroom

[0107] Data type: Internet usage behavior of students at a specific location. Objective: Analysis of emotional changes

[0108] Select data for background analysis and obtain recent student emotional change data. Call relevant tool example code:

[0109] python

[0110] def fetch_user_reviews(location):

[0111] Assume calling an API to get user reviews. response = call_external_api(location). return response

[0112] def analyze_sentiment(reviews):

[0113] Use a sentiment analysis toolkit, such as TextBlob or other NLP techniques. sentiments = [analyze(review) for review in reviews]. return sentiments

[0114] User input

[0115] user_input = "I want to know the sentiment of user reviews in Beijing recently."

[0116] location = "Beijing"

[0117] Get user reviews

[0118] reviews = fetch_user_reviews(location)

[0119] Analyze sentiment

[0120] sentiment_results = analyze_sentiment(reviews)

[0121] # Output the result. print(f"Sentiment analysis result of users in {location}: {sentiment_results}")

[0122] Return the result

[0123] Finally, return the processed data to the user in a user-friendly form, for example:

[0124] 70% of the evaluation sentiment of students' online behavior of accessing a certain APP at a specific location is positive, 25% is neutral, and 5% is negative. "This method not only improves the interaction experience but also can flexibly allocate tools and resources according to different needs, thus providing better services.

[0125] III. Examples of establishing a continuous learning and feedback mechanism:

[0126] Taking the sentiment classification task as an example, replace the original fully connected layer of the pre-trained neural network model with a classification layer suitable for sentiment classification. Use the professional corpus dataset of students' geographical locations, online behaviors, and mood changes collected on campus to continue training the model. Set a relatively small learning rate, such as 0.001, and the batch size is 32. During the training process, monitor the loss function and accuracy. If the loss value drops slowly for multiple consecutive rounds or the accuracy no longer improves, adjust the learning rate or other hyperparameters to avoid overfitting and continuously optimize the model.

[0127] IV. Examples of personalized analysis and output:

[0128] Collect long-term online behavior and emotion-related data of each student to construct a personal historical dataset. The large model analyzes the similarity between a student and other students based on features such as behavior patterns and interest preferences in the student's historical data. When it is necessary to identify the current emotional state of a certain student, the model combines the student's historical data and the emotional change rules of similar students to output personalized emotional recognition results, such as "This student may currently be in an anxious state about future career planning due to frequently accessing job hunting websites and staying for a long time."

[0129] As described above, the above are only specific embodiments of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for emotion recognition by combining a large model with geographical information and Internet behavior, characterized in that, It includes the following steps: Step S1: Data collection and integration: Collect the geographical information and Internet usage behavior data of users. The geographical information includes, but is not limited to, the longitude and latitude where the user is located and the geographical location tags; The Internet usage behavior data includes the websites visited by the user, the stay time on each website, the content of the browsed pages, the interaction behaviors on social platforms, and the usage of application programs; Integrate the collected geographical information and Internet usage behavior data to form a multi-source data set; Step S2: Data preprocessing: Clean the integrated multi-source data set to remove the noise data, duplicate data, and error data in it; Perform normalization processing on the data to make different types of data comparable; Perform natural language processing operations such as word segmentation and part-of-speech tagging on the text-type Internet usage behavior data; Step S3: Feature extraction: Extract the features related to sentiment recognition from the preprocessed data. For geographical information, extract the features of the geographical location, such as the number of people flow and the type of venue, etc.; For Internet usage behavior data, extract the browsing preferences and social activity levels of users, etc. Combine the extracted features into a feature vector; Step S4: Large model training and application: Use the training data set with sentiment labels to train the large model. The large model is a pre-trained model with strong semantic understanding and feature learning capabilities, and a model based on the Transformer architecture can be used; Input the feature vector obtained through feature extraction into the trained large model, and through the calculation and analysis of the large model, output the sentiment recognition result of the user. The sentiment recognition result includes, but is not limited to, sentiment categories such as positive, negative, and neutral; Step S5: Result evaluation and optimization: Use evaluation metrics to evaluate the sentiment recognition result; Adjust and optimize the parameters of the large model according to the evaluation result, or improve the feature extraction method.

2. The emotional recognition method combining a large model, geographical information, and Internet access behavior according to claim 1, wherein In the said step S1, the collection methods of geographical information include obtaining through the GPS positioning system of the user's mobile device, obtaining through Wi-Fi positioning technology, and obtaining from a third-party geographical information service provider; The collection methods of Internet usage behavior data include collecting through browser plugins and the buried point technology of mobile application programs.

3. The emotional recognition method combining a large model, geographical information, and Internet access behavior according to claim 1, wherein In the said step S2, when performing natural language processing operations on text-type data, it also includes operations such as removing stop words and performing stemming.

4. The emotional recognition method combining a large model with geographical information and Internet behavior according to claim 1, characterized in that In the said step S3, on the basis of extracting the features related to sentiment recognition, combine time features, such as the geographical information and Internet usage behavior data of users in different time periods.

5. A method for emotion recognition by combining a large model with geographical information and Internet behavior according to claim 1, characterized in that, In the said step S4, when training the large model, adopt the method of transfer learning, and use the knowledge learned by the pre-trained model on a large-scale corpus to fine-tune the model in combination with a small amount of training data with sentiment labels.

6. The emotional recognition method combining a large model, geographical information and Internet access behavior according to claim 1, characterized in that, In the said step S5, introduce manually labeled samples when training the large model to verify and calibrate the sentiment recognition result.

Images