Large model cue word optimization method and system for city block emotion memory

Abstract

The invention relates to a large model cue word optimization method and system for city block emotion memory, and the method comprises the following steps: obtaining an emotion memory text related to a city updating scene, and carrying out the preprocessing of the text; cue words are designed, and space types and time features are limited; the cue word is finely adjusted, related parameters of emotion polarity and intensity are added in the cue word, a context and a cultural background are introduced, and a timeline or plot development is added, so that the cue word achieves a balance between the complexity and the conciseness of emotion expression; and combining a psychological theory framework to construct a sentiment classification system adapted to a city update scene, extracting sentiment words in the text, and checking and correcting. Compared with the prior art, the method has the advantages that the sentiment analysis precision is improved, and structured sentiment memory data output is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of large models, and in particular to a large model prompt word optimization method and system for urban block emotional memory. Background Art

[0002] Continuing the emotional memory of the city is the front-end work and guiding basis for urban block renewal. Urban renewal not only includes the transformation of buildings and infrastructure, but also involves the continuation of intangible cultural heritage such as emotional memory and social value of the block. As the carrier of urban culture, the block not only witnesses the historical development, but also is an important source of emotional belonging for residents. At present, sentiment analysis technology has been widely used in various fields, especially in social media analysis and public opinion monitoring. However, there is still a lack of research and relevant practical applications in the field of urban renewal.

[0003] For example, patent application CN 118277569 A discloses a large model training method and a text emotion entity recognition method, which includes: inputting a sample text sequence and an emotion label into an initial large model for recognition to obtain a first prediction sequence; inputting a sample text sequence and an entity label representing the category of an entity object in the sample text sequence into the initial large model for recognition to obtain a second prediction sequence; using the sample text sequence, the first prediction sequence, and the second prediction sequence as training samples, and using the emotion label, the entity label, and the joint label as training labels, inputting the initial large model for recognition to obtain multiple prediction sequences, and continuing to iteratively train the initial large model according to the target loss value to obtain a target large model. This method enhances the model's ability to mine the correlation between emotion recognition and entity recognition, and the joint training of the two tasks increases recognition efficiency and accuracy, but there are still deficiencies in multi-level emotion expression processing and structured output. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a large-model prompt word optimization method and system for urban block emotional memory, thereby improving the accuracy of large-model emotional analysis and realizing structured emotional memory data output.

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

[0006] A large model prompt word optimization method for urban block emotional memory includes the following steps:

[0007] Acquire emotional memory texts related to urban renewal scenarios, and pre-process the texts;

[0008] Design prompt words to define the spatial type and time characteristics;

[0009] The prompt words are fine-tuned, relevant parameters of emotional polarity and intensity are added to the prompt words, context and cultural background are introduced, and a timeline or plot development is added, so that the prompt words achieve a balance between the complexity and simplicity of emotional expression.

[0010] Furthermore, the pre-processing steps are as follows:

[0011] Selecting texts that clearly express emotions and are related to urban renewal, architecture and public space from the texts through keyword retrieval, semantic matching and topic clustering algorithms;

[0012] Classify texts that explicitly express emotions and are related to urban renewal, architecture, and public space according to their length and emotional richness;

[0013] Perform text cleaning on the classified text to remove redundant information irrelevant to sentiment analysis and retain the paragraphs directly related to sentiment extraction;

[0014] Unified text format.

[0015] Furthermore, the redundant information includes a variety of background descriptions and non-emotional statements.

[0016] Furthermore, the space type includes space function type and space location characteristics. The space function type includes multiple types of residential living space, production and consumption space, and ecological leisure space. The space location characteristics include multiple types of location characteristics, urban positioning, and historical and cultural background.

[0017] Furthermore, a sentiment classification system adapted to urban renewal scenarios is constructed in combination with a psychological theory framework, and sentiment words in the text are extracted for verification and correction.

[0018] Furthermore, when designing the prompt words, the text is segmented, each paragraph maintains the same attention depth, focuses on place nouns related to sentiment analysis, and introduces an attention verification mechanism.

[0019] Furthermore, the prompt words are used to clarify the contents contained in the table header and the corresponding field scanning rules, and the output of the large model is structured.

[0020] Furthermore, the table header includes a plurality of block names, emotional keywords, emotional intensity, specific memory events and related years.

[0021] Furthermore, the number of entries in the table is dynamically adjusted according to the length of the text.

[0022] According to another aspect of the present invention, a large model prompt word optimization system for urban block emotional memory is provided, comprising:

[0023] A text preprocessing module, used to obtain emotional memory text related to the urban renewal scene and preprocess the text;

[0024] A prompt word design module is used to design prompt words and define spatial types and time characteristics;

[0025] The prompt word optimization module fine-tunes the prompt words, adds relevant parameters of emotional polarity and intensity to the prompt words, introduces context and cultural background, adds a timeline or plot development, finds a balance between the complexity and simplicity of emotional expression of the prompt words, and captures subtle emotional changes.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] 1. This invention significantly improves the task understanding quality of the large model when processing the urban block emotional memory recognition task by optimizing the prompt words. Combining background input adjustment and emotional deep mining strategy, it ensures that the large model maintains high accuracy in the multi-dimensional emotion recognition process and performs more delicate analysis of complex emotional associations.

[0028] 2. The present invention introduces an attention verification mechanism to ensure that the large model can cover all key information when analyzing the text, and confirm whether it pays attention to the emotions and key locations in each paragraph of the text. If there are omissions, the analysis strategy can be adjusted in time to ensure the consistency of the large model output.

[0029] 3. The present invention expands the prompt word optimization strategy from the result output end, guiding the large model to present the sentiment analysis results in a structured manner, such as outputting sentiment classification, sentiment intensity and context matching information in the form of tables. This strategy improves the clarity and operability of the result display, and provides more powerful support for subsequent data analysis and updating of the sentiment library.

[0030] 4. Aiming at the defects of different large models in processing emotional tasks, the present invention proposes a targeted prompt word optimization method, which effectively reduces the common errors of large models in emotion classification, emotion intensity quantification and other links, improves the task adaptability and generalization ability of the model, thereby ensuring stability and consistency in different scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A method flow chart of a large model prompt word optimization method for urban block emotional memory proposed by the present invention;

[0032] Figure 2 This is a structural schematic diagram of a large-model prompt word optimization system for urban block emotional memory proposed by the present invention. DETAILED DESCRIPTION

[0033] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0034] Example 1

[0035] This embodiment provides a large model prompt word optimization method for urban block emotional memory, such as Figure 1 As shown, the following steps are included:

[0036] S1. Obtain emotional memory text related to urban renewal scenarios and preprocess the text.

[0037] This embodiment selects text data closely related to the emotions of urban blocks from newspapers, magazines, books, interview recordings and other materials to ensure data diversity. Through algorithms such as keyword retrieval, semantic matching and topic clustering, texts that clearly express emotions and are related to urban renewal, architecture, public space, etc. are selected.

[0038] The texts are classified according to their length and emotional enrichment. Emotionally rich and long texts are prioritized and segmented using semantic clustering technology to facilitate the structure of content output. Emotionally concentrated but short texts are classified through the emotional tag system to further explore deep emotional features. Long texts with thin emotions are streamlined through automatic summarization technology to retain key emotional expressions. Short texts with less emotions are used as auxiliary data and included in the research system through labeling and classification. The text is then cleaned to remove redundant information that is not related to sentiment analysis, such as background descriptions, non-emotional statements, etc., while retaining paragraphs directly related to sentiment extraction. The automated text cleaning module removes invalid data according to predefined criteria and filters noisy texts through the sentiment density analysis algorithm. These texts will be stored for reference in subsequent analysis.

[0039] In addition, text cleaning also needs to pay attention to the processing of punctuation, formatting symbols, abbreviations and special symbols to ensure that key emotional information is not lost. Finally, the text format is unified to provide a consistent basis for contextual analysis.

[0040] S2. Design prompt words to limit spatial type and time characteristics.

[0041] In the urban renewal scenario, prompt words need to clarify the spatial type and time characteristics. The spatial type includes spatial function type and spatial location characteristics. Spatial function types include residential living space, production and consumption space, ecological leisure space, etc. Each spatial function type corresponds to a different type of emotion. For example, the emotions in the production and consumption space are more related to commercial interests and consumer experience; while the residential space is closely connected with emotions such as family and community identity.

[0042] Spatial location features include location characteristics, urban positioning, historical and cultural background, etc. For example, the emotions of historical districts are often focused on residents' nostalgia and collective memory, while emerging commercial districts are more involved in development, change, etc. Through the limitation of such spatial features, the prompt words can better provide the model with an accurate background for sentiment classification.

[0043] Temporal characteristics include the time point or time stage when emotions appear.

[0044] S3. Fine-tune the prompt words, add relevant parameters of emotional polarity and intensity to the prompt words, introduce context and cultural background, add timeline or plot development, so that the prompt words achieve a balance between the complexity and simplicity of emotional expression.

[0045] First, the input text is segmented, and longer text is divided into multiple small paragraphs with clear themes. Each paragraph should revolve around a central theme and be able to be summarized in a sentence or title. During the analysis process, the large model is required to maintain the same attention depth for each paragraph. Each paragraph should be paid equal attention during analysis to ensure that emotional information is not missed.

[0046] Examples of prompt words for this section are as follows:

[0047] “Please conduct a sentiment analysis task based on the following urban renewal project context: You are a [identity / perspective, such as an urban planner], and you are studying the emotional characteristics of residents in an urban renewal project of the [residential living space / industrial consumption space / ecological leisure space] type. The project is located in [specific urban spatial location], which has a rich historical and cultural background, including [historical sites, important landmarks, unique community culture, etc.]. Residents in the area have deep emotional connections to these historical neighborhoods, especially through the collective memory and experience of social change generated by [time period or event, such as the early 20th century, celebrations].”

[0048] The design of prompt words needs to guide the model to maintain continuous attention to the text to ensure the consistency of sentiment analysis. Prompt words should remind the model to evenly devote analysis attention regardless of the length of the text, and avoid focusing only on the beginning or end of the text. At the same time, due to the targeted nature of the task, it is necessary to focus on location-related nouns and the context of the nouns. Introduce an attention verification mechanism to ensure that the model covers all key information when analyzing the text. Prompt words should also guide the model to recognize the complexity of emotional context, especially the emotional expression conveyed through rhetoric.

[0049] Examples of prompt words for this section are as follows:

[0050] (1) “Please analyze the following text paragraph by paragraph, making sure that the emotional information in each paragraph is given equal attention.”

[0051] (2) “Pay attention to [place] nouns, extract them, and analyze their role in the emotional memory of the city.”

[0052] (3) “Please analyze the emotional expression in the following text, paying special attention to the underlying emotional meaning conveyed through rhetorical devices such as simile, symbol, and metaphor.”

[0053] The design of prompt words should guide the model to recognize the complexity of emotional context, especially the emotional expression conveyed through rhetoric. The prompt words should clearly require the model to handle complex contextual expressions to ensure the comprehensiveness of sentiment analysis.

[0054] When designing prompt words, instruct the model to prioritize metaphors and symbols, and conduct in-depth analysis of their roles in emotional contexts. By fine-tuning the prompt words, guide the model to identify and parse the hidden levels of emotional expression in the text, especially the emotions conveyed through metaphors, symbols, and metaphors. Prompt words need to allow the model to understand the meaning of rhetoric in different contexts, avoiding the model from only analyzing literal emotions. Enhance the model's recognition of figurative and metaphorical language, prompting the model to not only analyze keywords, but also consider the emotional intentions behind the rhetoric.

[0055] Add relevant parameters of sentiment polarity and intensity in the prompt word design, prompting the model to analyze the depth and multidimensional characteristics of sentiment from the semantic intensity and intonation changes of the text. This strategy guides the model to analyze the type and intensity of sentiment in the text, and guides the model to quantify the polarity (positive, negative) and intensity (high, medium, low) of sentiment expression into evaluable parameters through the optimization of prompt words. In the prompt words, it is emphasized that the model not only needs to identify the type of sentiment, but also quantify its intensity and polarity to ensure the accuracy of sentiment classification.

[0056] According to the specific cultural background, the model is prompted to pay attention to the impact of historical culture, social values ​​and other factors on emotional expression, and conduct in-depth analysis in combination with the specific time background. This optimization strategy can help the model more accurately analyze the emotions of people and land in the process of urban renewal, especially those emotional expressions closely related to cultural heritage, such as nostalgia or collective memory, ensuring that the emotional analysis has higher cultural sensitivity and contextual depth.

[0057] Adding a timeline or plot development hint to the prompt words guides the model to analyze the fluidity and complexity of emotions and capture how emotions change as events evolve. Through this optimization strategy, the model can provide a more dynamic and detailed analysis of emotional changes, especially in the process of urban renewal, where the transformation of residents' emotions often involves complex temporal and spatial factors.

[0058] Different large models have different advantages and limitations in sentiment analysis and generation capabilities. In view of these differences, a series of prompt word optimization strategies are proposed to improve the generalization of the results.

[0059] By testing four representative pre-trained language models (GPT-4o, KIMI, Microsoft COPILOT, GLM-4), it is found that they show different characteristics in the sentiment extraction task:

[0060] GPT-4o: It performs well in the refinement of sentiment classification, especially in the multi-level analysis of complex sentiment. However, its overly complex classification may lead to blurred boundaries between sentiment categories, increasing the difficulty of use and understanding.

[0061] KIMI: Its advantage lies in the simplicity and intuitiveness of sentiment classification, but it loses its sophistication when dealing with complex emotions and cannot fully capture delicate emotional expressions, which limits its application in deep sentiment analysis.

[0062] Microsoft COPILOT: It performs well in the intuitiveness of sentiment classification and is suitable for fast sentiment analysis scenarios. However, due to the lack of deeper sentiment classification capabilities, it is insufficient in capturing complex emotions.

[0063] GLM-4: It has powerful sentiment level analysis capabilities and can accurately capture the changes in complex emotions. It is suitable for highly complex sentiment analysis tasks, but overly detailed classification will increase application costs and users may get lost in the emotion categories.

[0064] Based on the characteristics and defects of the above models, a prompt word optimization strategy is formulated to balance the complexity and simplicity of emotions, so as to improve the generalization and practicality of large models in sentiment analysis.

[0065] The design of prompt words needs to strike a balance between the complexity and simplicity of emotional expression. It should be able to handle delicate emotional distinctions without making the classification too complicated. In the task of sentiment extraction in the context of urban renewal, the design should be able to capture subtle emotional changes while ensuring that the model outputs clear and understandable prompt words.

[0066] In order to strike a balance between the complexity and simplicity of the extracted sentiment words, regulations are made from two dimensions: sentiment type and sentiment word length:

[0067] (1) The same sentiment word extracted contains no more than two sentiment types;

[0068] (2) The length of each sentiment word shall not exceed 5 characters.

[0069] Examples of prompt words for this section are as follows:

[0070] “In the sentiment extraction task in the context of urban renewal, clear sentiment classifications such as “place attachment”, “childhood memory”, “cultural nostalgia”, “cultural admiration”, “happy experience”, “community expectation”, etc. are used to ensure that sentiment words are not overly complicated.”

[0071] In another preferred embodiment, it also includes:

[0072] S4. Combine the theoretical framework of psychology to build a sentiment classification system that adapts to urban renewal scenarios, extract sentiment words from the text, and conduct verification and correction.

[0073] This embodiment refers to the Plutchik emotion roulette theory to preliminarily construct a sentiment classification system that adapts to the urban block renewal scenario. According to the characteristics of the urban block renewal scenario, the sentiment in the input text is matched and checked with the Plutchik emotion roulette theory classification standards. And the sentiment is scored according to the intensity of the sentiment. The purpose of this step is to standardize, unify, and subdivide the sentiment types in the characteristics of the urban block renewal scenario. In this process, the basic dimensions of emotion need to be refined into multi-level sentiment levels to help the language model accurately understand the different forms of emotion.

[0074] Examples of prompt words for this section are as follows:

[0075] (1) “According to Plutchik’s emotion wheel theory, emotions can be divided into eight core emotions and their secondary emotions. We hope that the model can understand the hierarchical relationship between these emotions and conduct analysis based on these theoretical frameworks in the subsequent steps.”

[0076] (2) “Based on Ekman’s sentiment framework, extract sentences expressing sentiment in the text and match them with the corresponding basic sentiment categories. Each sentiment category can correspond to a word in Plutchik’s sentiment classification dictionary. Sentiments are divided into major sentiment categories and minor sentiment categories. Each major sentiment category contains several minor sentiment categories to ensure the accuracy and reliability of the classification results. Finally, output an emotional vocabulary that reflects the subtle changes in these emotions.”

[0077] (3) “Ensure the consistency of classification according to the criteria for defining sentiment. Classify the intensity of sentiment in the identified text into three categories: high, medium, and low.”

[0078] S5. Input the preprocessed text, prompt words and sentiment words into the large model to obtain structured sentiment memory recognition results.

[0079] The emotional memory recognition results are presented in a structured table. The prompt words should not only clearly explain the organizational logic of the table, but also explain in detail the contents contained in the table header and the corresponding field scanning rules. The details are as follows:

[0080] Each table header should clearly display the name of the block, emotional keywords, emotional intensity, specific memory events, and related years. The prompt words should guide the model to organize each piece of data according to this logic to ensure that the table content is structured and easy to search. Each table entry should contain 4-6 fields of information to avoid excessive complexity. At the same time, the number of output table entries for each paragraph of text should be dynamically adjusted according to the length of the text.

[0081] The following is an example of header design:

[0082] Block name: This refers to the specific city block name, such as "Sichuan North Road" and "Duolun Road".

[0083] Emotional categories: This refers to the main emotions that residents have towards the neighborhood, such as "nostalgic", "happy" and "sad".

[0084] Emotional subcategories: These refer to the residents’ main emotions towards the neighborhood. Each major emotion category contains several emotional subcategories, such as “place attachment,” “childhood memories,” and “cultural nostalgia.”

[0085] Sentiment Intensity: This is a numerical value that indicates the intensity of a sentiment, which can be classified as high, medium, or low. The higher the sentiment score, the stronger the sentiment.

[0086] Specific memory events: These are specific memory events of the neighborhood that residents have. For example, "Duolun Road" reminds people of the memory of playing here when they were young; "Sichuan North Road" reminds people of the youth time spent with friends.

[0087] Relevant years: This refers to the time period in which these memories occurred. For example, memories about "Sichuan North Road" occurred from the 1920s to the present; memories about "Duolun Road" occurred from the 1990s to the present.

[0088] Field scanning rules are used to ensure the accuracy and consistency of the emotional memory data of urban blocks. The methods include data acquisition, preprocessing, storage and analysis. The integrity and standardization of the data are ensured by verifying the type, length, uniqueness and value range restrictions of each field. The key principles and methods of field scanning are as follows:

[0089] Block Name: Should be a string, no longer than 50 characters, must be unique and non-empty.

[0090] Emotion category: should be a string, such as "nostalgic", "happy", and not empty.

[0091] Emotion subcategory: should be a string, such as "place attachment", "childhood memory", "cultural nostalgia", and should not be empty.

[0092] Sentiment intensity: should be "high, medium, low", and not empty.

[0093] Specific memory event: should be a string with a length of no more than 200 characters and not empty.

[0094] Relevant Years: should be a valid year or year range (such as "1980-1990"), within a reasonable range (such as 1990 to present), and not empty.

[0095] Examples of prompt words for this section are as follows:

[0096] “Please extract the sentiment features of the following text and organize them into a standardized table format. The table should contain fields such as [neighborhood name, major sentiment category, minor sentiment category, sentiment intensity, specific memory event, and related year], and each field should be matched according to the sentiment features in the text. Make sure that each piece of data in the output table contains complete sentiment analysis results, and scan sentiment keywords, sentiment categories, and their context in each paragraph of text.”

[0097] The results of structured emotional memory recognition are shown in Table 1.

[0098] Table 1. Emotional memory table of urban blocks

[0099]

[0100]

[0101]

[0102]

[0103] Example 2

[0104] This embodiment provides a large model prompt word optimization system for urban block emotional memory, the structural diagram is as follows Figure 2 As shown, including:

[0105] A text preprocessing module, used to obtain emotional memory text related to the urban renewal scene and preprocess the text;

[0106] A prompt word design module is used to design prompt words and define spatial types and time characteristics;

[0107] The prompt word optimization module fine-tunes the prompt words, adds relevant parameters of emotional polarity and intensity to the prompt words, introduces context and cultural background, adds a timeline or plot development, finds a balance between the complexity and simplicity of emotional expression of the prompt words, and captures subtle emotional changes.

[0108] The rest is the same as in Example 1.

[0109] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A large model prompt word optimization method for urban block emotional memory, characterized in that: The following steps are involved: Acquire emotional memory texts related to urban renewal scenarios, and pre-process the texts; Design prompt words to define the spatial type and time characteristics; The prompt words are fine-tuned, relevant parameters of emotional polarity and intensity are added to the prompt words, context and cultural background are introduced, and a timeline or plot development is added, so that the prompt words achieve a balance between the complexity and simplicity of emotional expression.

2. The large model prompt word optimization method for urban block emotional memory according to claim 1 is characterized in that: The steps of the pre-processing are as follows: Selecting texts that clearly express emotions and are related to urban renewal, architecture and public space from the texts through keyword retrieval, semantic matching and topic clustering algorithms; Classify texts that explicitly express emotions and are related to urban renewal, architecture, and public space according to their length and emotional richness; Perform text cleaning on the classified text to remove redundant information irrelevant to sentiment analysis and retain the paragraphs directly related to sentiment extraction; Unified text format.

3. The large model prompt word optimization method for urban block emotional memory according to claim 2 is characterized in that: The redundant information includes a variety of background descriptions and non-emotional statements.

4. The large model prompt word optimization method for urban block emotional memory according to claim 1 is characterized in that: The space types include space function types and space location characteristics. The space function types include multiple types of residential living space, production and consumption space, and ecological leisure space. The space location characteristics include multiple types of location characteristics, urban positioning, and historical and cultural backgrounds.

5. The large model prompt word optimization method for urban block emotional memory according to claim 1 is characterized in that: Combined with the theoretical framework of psychology, a sentiment classification system adapted to urban renewal scenarios is constructed, and sentiment words in the text are extracted for verification and correction.

6. The large model prompt word optimization method for urban block emotional memory according to claim 1 is characterized in that: When designing the prompt words, the text is segmented, each paragraph maintains the same attention depth, focuses on place nouns related to sentiment analysis, and introduces an attention verification mechanism.

7. The large model prompt word optimization method for urban block emotional memory according to claim 1 is characterized in that: The prompt words are used to clarify the contents contained in the table header and the corresponding field scanning rules, and the output of the large model is structured.

8. The large model prompt word optimization method for urban block emotional memory according to claim 7 is characterized in that: The header includes a variety of block names, emotional keywords, emotional intensity, specific memory events and related years.

9. The large model prompt word optimization method for urban block emotional memory according to claim 7 is characterized in that: The number of entries in the table is dynamically adjusted according to the length of the text.

10. A large model prompt word optimization system for urban block emotional memory, characterized in that: include: A text preprocessing module, used to obtain emotional memory text related to the urban renewal scene and preprocess the text; A prompt word design module is used to design prompt words and define spatial types and time characteristics; The prompt word optimization module fine-tunes the prompt words, adds relevant parameters of emotional polarity and intensity to the prompt words, introduces context and cultural background, adds a timeline or plot development, finds a balance between the complexity and simplicity of emotional expression of the prompt words, and captures subtle emotional changes.

Images