Logistics data visualization interaction method based on large model

Abstract

The invention discloses a logistics data visualization interaction method based on a large model, and belongs to the technical field of data interaction. Comprising a voice input module, a voice recognition module, a semantic understanding and intention recognition module, a secondary intention matching function module, a data query and processing module, a visual display module, a voice feedback module and an interaction and optimization module. Compared with the traditional intention label matching technology, the intention label matching method has the advantages of high accuracy, high robustness, wide application range, low cost and high processing speed in the aspects of learning capability, universality, automation and parallel computing capability.

Description

Technical Field

[0001] The present invention belongs to the technical field of data interaction, and in particular relates to a logistics data visualization interaction method based on a large model. Background Art

[0002] Visual interaction refers to the process of exploring, analyzing and manipulating data by converting data or information into visual forms such as line charts, bar charts, scatter plots, maps, etc.

[0003] Traditional large-model intent labels cannot be accurately mapped to localized business system operations. The main reason is that different localized business systems have unique operating logic and business rules, which often involve complex business processes and specific industry knowledge. In addition, when processing natural language, large models may be affected by language diversity, insufficient context understanding, and data sparsity, which may cause deviations between the intent labels they generate and the actual operational requirements of localized business systems. There may also be mismatches in the interface and data format between large models and localized business systems. These factors work together to make it difficult for large-model intent labels to be directly applied to localized business system operations.

[0004] The traditional method of mapping localized business intent labels to data query statements has obvious drawbacks. Due to the lack of a dedicated evaluation model, these methods often cannot accurately determine whether the generated query statements meet the business needs of a specific scenario when converting intent labels into data query statements, which seriously affects the accuracy and relevance of the query results. Secondly, without the guidance of an evaluation model, these methods often lack pertinence and flexibility when choosing to use semantic analysis, machine learning, or knowledge graphs to construct queries.

[0005] These problems together lead to the shortcomings of traditional methods in mapping localized business intent labels to data query statements, making it difficult to meet the needs of modern business systems. Summary of the invention

[0006] The purpose of the present invention is to provide a logistics data visualization interaction method based on a large model to solve the problems raised in the above background technology.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solutions: a logistics data visualization interaction method based on a large model, comprising a voice input module, a voice recognition module, a semantic understanding and intention recognition module, a secondary intention matching function module, a data query and processing module, a visualization display module, a voice feedback module and an interaction and optimization module.

[0008] Preferably, the user of the voice input module issues voice commands through the device to express the query, analysis or operation requirements for logistics data.

[0009] Preferably, in the speech recognition module, after the system receives the user's speech signal, the speech recognition model is used to convert the speech into text.

[0010] Preferably, the text converted by the speech recognition module in the semantic understanding and intent recognition module is sent to a large model for semantic understanding and intent recognition, and the large model is based on contextual knowledge in the logistics field and pre-trained semantic representation.

[0011] Preferably, the secondary intent matching function module is based on a large model. First, the large model is used to parse the user input to obtain multiple intent labels, and then candidate secondary intents are generated according to the intent labels returned by the large model. The candidate secondary intents are sorted, and finally the secondary intent with the highest score is selected as the final intent label.

[0012] Preferably, the data query and processing module queries the corresponding data from the logistics database according to the user's intention, and performs necessary analysis and processing to generate the information required by the user.

[0013] Preferably, in the visualization module, the system generates corresponding visualization charts and interfaces based on the query results to provide more intuitive and easy-to-understand information.

[0014] Preferably, in the voice feedback module, the system feeds back the visualization results to the user in the form of voice.

[0015] Preferably, in the interaction and optimization module, the user performs further voice interaction based on the system's feedback, and continuously optimizes the model's recognition capability and interaction experience.

[0016] The present invention has the following advantages:

[0017] 1. Support voice commands and natural language interaction. Users can interact with the visualization system in a more natural way, which reduces the usage threshold;

[0018] 2. The intent label secondary matching technology based on the big model, with the advantages of the big model in language knowledge, context information learning, learning ability, versatility, automation and parallel computing ability, shows higher accuracy, stronger robustness, wider application areas, lower cost and faster processing speed compared with the traditional intent label matching technology. This technology can significantly improve the accuracy of intent analysis, and by building a bridge for communication between the big model intent label and the localized business, it can accurately meet the personalized needs of the user's localized business system and achieve more efficient and accurate communication and services;

[0019] 3. The method of dynamically parsing data queries based on adaptive tag intent not only innovatively builds a semantic evaluation model to accurately evaluate the rationality of the mapped query statements, but also cleverly integrates semantic analysis, machine learning, knowledge graphs and other methods to automatically build query statements. This method shows higher flexibility and adaptability, can more accurately understand user intent, and generate more demand-oriented query statements, thereby greatly improving the efficiency and accuracy of data queries. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0021] The specific implementation modes of the present invention are further explained in detail below by describing the embodiments with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0022] like Figure 1 As shown, the present invention provides a logistics data visualization interaction method based on a large model, including a voice input module, a voice recognition module, a semantic understanding and intention recognition module, a secondary intention matching function module, a data query and processing module, a visualization display module, a voice feedback module and an interaction and optimization module.

[0023] In this embodiment, the user of the voice input module issues voice commands through the device to express the query, analysis or operation requirements for logistics data, uses a computer microphone to collect voice signals, and pre-processes the collected voice signals to convert analog voice signals into digital signals. Usually, an appropriate sampling rate and quantization bit number are used to divide the continuous voice signal into short frames for subsequent time-frequency analysis and feature extraction.

[0024] In this embodiment, in the speech recognition module, after the system receives the user's speech signal, it converts the speech into text using the speech recognition model, performs time-frequency analysis techniques such as short-time Fourier transform on the speech signal, converts the speech signal from the time domain to the frequency domain, and extracts the spectral features of the speech. The system loads a pre-trained acoustic model. The acoustic model is usually trained based on a large amount of speech data and corresponding text labels, and is used to map acoustic features to corresponding text units. The system loads a pre-trained language model. The language model is used to describe the probability distribution of the text sequence, taking into account the contextual relationship, grammatical rules and semantic information of the vocabulary. According to the results of the acoustic model matching, the language model is used to search for paths and find the most likely text sequence. This is usually achieved through a dynamic programming algorithm such as the Viterbi algorithm, which comprehensively considers the score of the acoustic model and the score of the language model. After the acoustic model matching and the language model decoding, the system outputs the recognized text result. This text result is a transcription of the user's speech signal, which provides a basis for subsequent natural language understanding and visual interaction.

[0025] In this embodiment, the text obtained by the semantic understanding and intention recognition module after conversion by the speech recognition module is sent to the big model for semantic understanding and intention recognition. The big model is based on the contextual knowledge and pre-trained semantic representation in the logistics field, and is a key link in natural language processing. It involves in-depth analysis of the converted text input to understand the user's true intentions and needs.

[0026] In the logistics visualization interactive system, this step is particularly important because the system needs to accurately understand the user's query, request or instruction in order to provide correct data analysis and visualization display, and pre-process the text obtained from the speech recognition step. This includes removing irrelevant characters and punctuation marks, performing word form restoration (restoring words to their basic form) or part-of-speech tagging (assigning grammatical categories to each word) and other operations. Pre-processing helps improve the accuracy of subsequent semantic understanding.

[0027] The system performs word segmentation on the preprocessed text, that is, dividing the continuous text into separate vocabulary units. At the same time, the named entity recognition technology is used to identify key entities in the text, such as logistics order number, cargo name, location, etc. These entities are crucial to understanding the user's intention.

[0028] The converted semantic representation is fed into a pre-trained large model for inference. Large models are usually built based on deep learning technology and have strong representation learning and generalization capabilities. In the field of logistics, large models are trained using a large amount of logistics data and knowledge to learn how to parse user intent from semantic representations.

[0029] In this embodiment, the secondary intention matching function module is based on a large model. First, the large model is used to parse the user input to obtain multiple intention labels, and then candidate secondary intentions are generated according to the intention labels returned by the large model, and the candidate secondary intentions are sorted. Finally, the secondary intention with the highest score is selected as the final intention label. After the large model infers the semantic representation, one or more intention labels will be output. These labels describe the operation or query type corresponding to the user's query, request or instruction. Traditional natural language parsing intent labels usually use rules or machine learning models to match user input and intent labels. These methods usually rely on manually defined rules or features, so they may not be able to handle complex semantics and grammatical structures well. The large model can learn a large amount of text data and use contextual information to better understand user intent. The large model can capture the subtle differences in language and combine it with contextual information to more accurately parse user intent. Compared with the traditional practice of directly parsing intent labels from user input, secondary intent matching based on the large model has higher accuracy, stronger robustness and a wider range of applications. Specifically, secondary intent matching based on the big model can use the language knowledge and contextual information learned by the big model to improve the accuracy of intent parsing. It can also meet the needs of different business scenarios for the definition and semantic disambiguation of intent labels through secondary intent matching, and map them to specific localized business operations or queries, while improving the security of the business.

[0030] In this embodiment, the data query and processing module queries the corresponding data from the logistics database according to the user's intention, and performs necessary analysis and processing to generate the information required by the user. The system performs corresponding operations or data queries according to the parsed user intention, and feeds back the results to the visualization system. At the same time, the system will continuously adjust and optimize the parameters and rules of the large model according to the user's feedback and behavior to improve the accuracy of semantic understanding and intention recognition. According to the user's intention, the system constructs the corresponding database query statement. These query statements may involve SQL or other database query languages, which are used to extract data from the logistics database. Traditional practices usually rely on fixed rules or templates when mapping intent tags to dynamic SQL queries. These rules or templates are pre-defined, and each intent tag corresponds to one or more fixed SQL query statements. When the system recognizes an intent tag, it will directly use the SQL template corresponding to the tag for query. However, this traditional method has some limitations and lacks flexibility: fixed templates are difficult to cope with complex and changing query requirements, especially in the field of logistics industry where business scenarios are complex and changing. High maintenance cost: As the business develops and changes, it may be necessary to continuously modify and add rules or templates, which increases the difficulty and cost of maintenance. Limited performance optimization: Fixed SQL templates may not fully utilize the database's query optimization technology, resulting in poor query performance. To this end, we have built an adaptive SQL parsing model that measures the semantic similarity between tag intent query statements by evaluating semantic similarity, and then chooses to use semantic analysis, machine learning, and knowledge graphs to parse SQL, improving the flexibility, accuracy, and efficiency of parsing.

[0031] Organize the queried and processed data into a format suitable for visualization. This includes operations such as filtering, sorting, grouping, or transforming data to ensure that the data can be correctly mapped to the visualization elements. Choose the appropriate visualization method based on the data type and display objectives. For time series data, you can use a line chart to show trends; for categorical data, you can use a bar chart or pie chart to compare the proportions of different categories; for geographic spatial data, you can use a map to show the distribution, etc.

[0032] In this embodiment, the system in the visualization display module generates corresponding visualization charts and interfaces based on the query results to provide more intuitive and easy-to-understand information, and designs a visualization interface according to the personalized needs of the user, including selecting appropriate colors, fonts, icons and layouts, etc., to provide intuitive and easy-to-understand information. At the same time, considering the user's interaction habits, ensure that the interface is concise and easy to operate.

[0033] Use visualization tools or libraries to create charts based on data and design requirements. This includes setting axes, adding legends, setting titles and labels, etc., so that users can clearly understand the information conveyed by the chart.

[0034] In this embodiment, in the voice feedback module, the system feeds back the visualization results to the user in the form of voice.

[0035] In this embodiment, the user in the interaction and optimization module performs further voice interaction based on the system's feedback, and continuously optimizes the model's recognition capability and interaction experience.

[0036] Example

[0037] (1) User voice input: Users issue voice commands through the device computer to express their needs for querying, analyzing or operating logistics data.

[0038] (2) The speech signal is converted from the time domain to the frequency domain to extract the spectral features of the speech. The system loads a pre-trained acoustic model. The acoustic model is usually trained based on a large amount of speech data and corresponding text labels to map the acoustic features to the corresponding text units. The system loads a pre-trained language model. The language model is used to describe the probability distribution of text sequences, taking into account the contextual relationship, grammatical rules and semantic information of the vocabulary. According to the results of the acoustic model matching, the language model is used to search for paths to find the most likely text sequence. This is usually achieved through a dynamic programming algorithm (such as the Viterbi algorithm), which comprehensively considers the score of the acoustic model and the score of the language model. After the acoustic model matching and language model decoding, the system outputs the recognized text results.

[0039] (3) Semantic understanding and intent recognition. The converted text is fed into the big model for semantic understanding and intent recognition. The big model uses the contextual knowledge of the logistics field and pre-trained semantic representation. The intent label format returned by the big model interface is usually in JSON format, which includes the following fields:

[0040] Intent: the name of the intent, which indicates the operation that the user wants to complete or the information that the user wants to obtain;

[0041] Confidence: Confidence, which indicates the model's assessment of the accuracy of the prediction of the intent;

[0042] Slots: Slots represent the parameters or information required by the intent;

[0043] In order to better match the intent labels returned by the large model, a secondary intent classifier is matched to connect the intent labels returned by the interface of the large model with the action intents between localized businesses. The intent labels returned by the interface of the large model are universal, while the action intents between localized businesses are specific. Therefore, secondary intent classification is required to map the intent labels returned by the large model to the action intents between localized businesses. To this end, we trained a personalized classifier for intents related to the logistics industry based on the BERT semantic model to improve the intent hit rate of the intent labels returned by the large model to the localized businesses. The algorithms involved are as follows:

[0044] Naive Bayes is used for secondary intent classification. For the semantic label "Query the freight rate trend in the past week", the formula for using Naive Bayes for secondary intent classification is as follows:

[0045] P(y|x)=P(y)*ΠP(x_i|y) / P(x);

[0046] P(y|x): the probability that semantic label x belongs to intent y;

[0047] P(y): prior probability of intention y;

[0048] P(x_i|y): The probability of feature x_i appearing under intent y

[0049] P(x): marginal probability of semantic label x;

[0050] For the semantic tag "Query the freight rate trend in the past week", we can break it down into the following features:

[0051] Keywords: freight rate, trend, one week;

[0052] Parts of speech: noun, verb, adjective;

[0053] Syntactic structure: subject-verb-object structure, verb-object structure;

[0054] Therefore, we can rewrite the above formula as;

[0055] P(y|x)=P(y)*P(freight rate|y)*P(trend|y)*P(week|y)*P(noun|y)*P(verb|y)*P(adjective|y)*P(subject-verb-object structure|y)*P(verb-object structure|y) / P(x);

[0056] Among them, P(x) can be calculated by the following formula;

[0057] P(x)=ΣP(y)*ΠP(x_i|y);

[0058] P(y) can be calculated from the training data;

[0059] P(x_i|y) can be calculated by the following formula, where P(x_i|y) represents the probability of feature x_i appearing under condition y:

[0060] P(x_i|y)=Count(x_i,y) / Count(y);

[0061] Among them, Count(x_i,y) represents the number of times feature x_i appears under intention y, and Count(y) represents the number of times intention y appears;

[0062] The advantages of using Naive Bayes for secondary intent classification are that it is simple to understand and has high computational efficiency. The disadvantages are that the assumption of feature independence is a bit strict and it is a bit overfitting.

[0063] (4) Mapping of data queries. Converting the intent tag matching output by the large model into specific queries is a key step. Currently, there are mainly semantic analysis, machine learning, and knowledge graph methods. For this purpose, we have built an adaptive evaluation model to improve and identify when to use different mapping methods to better meet the needs of business scenarios:

[0064] (a) Semantic similarity can be used to measure the semantic similarity between label intent and SQL statements, so as to choose whether to use semantic analysis or machine learning to parse SQL. The following formula is used to calculate semantic similarity: semantic similarity = (A∩B) / (A∪B), where A represents the semantics of label intent and B represents the semantics of SQL statements;

[0065] (b) Data quality: Data quality can be used to measure data reliability, so as to choose whether to use semantic analysis, machine learning, or knowledge graph to parse SQL. The following formula is used to evaluate data quality:

[0066] Data quality = (accuracy + completeness + consistency + timeliness) / 4;

[0067] (c) Requirements: Requirements can be used to measure the real-time requirements, so as to choose whether to use semantic analysis, machine learning, or knowledge graph to parse SQL. The following formula is used to calculate the real-time requirements:

[0068] Real-time requirement = (real-time + accuracy + consistency) / 3;

[0069] Among them, real-time, accuracy, and consistency are the three dimensions of real-time requirements;

[0070] (d) Cost: You can use the cost to measure the cost of parsing SQL using different methods, so as to choose whether to use semantic analysis, machine learning, or knowledge graph to parse SQL. Use the following formula to calculate the cost:

[0071] Cost = (manpower cost + time cost + resource cost) / 3;

[0072] Among them, labor cost, time cost and resource cost are the three dimensions of cost.

[0073] If the semantic similarity is high, the data quality is high, the real-time requirement is low, and the cost is low, semantic analysis can be used to parse SQL.

[0074] If the semantic similarity is high, the data quality is high, the real-time requirement is high, and the cost is high, machine learning can be used to parse SQL.

[0075] If the semantic similarity is high, the data quality is low, the real-time requirement is low, and the cost is low, the knowledge graph can be used to parse SQL.

[0076] If the semantic similarity is high, the data quality is low, the real-time requirement is high, and the cost is high, machine learning can be used to parse SQL.

[0077] If the semantic similarity is low, the data quality is high, the real-time requirement is low, and the cost is low, semantic analysis or knowledge graph can be used to parse SQL.

[0078] If the semantic similarity is low, the data quality is high, the real-time requirement is high, and the cost is high, machine learning can be used to parse SQL.

[0079] If the semantic similarity is low, the data quality is low, the real-time requirement is low, and the cost is low, the knowledge graph can be used to parse SQL.

[0080] If the semantic similarity is low, the data quality is low, the real-time requirement is high, and the cost is high, machine learning can be used to parse SQL.

[0081] (5) Organize the queried and processed data into a format suitable for visualization. This may include operations such as filtering, sorting, grouping, or transforming data to ensure that the data can be correctly mapped to the visualization elements. According to the type of data and the display goal, choose the appropriate visualization method;

[0082] (6) Design a visual interface based on the user's personalized needs. This includes selecting appropriate colors, fonts, icons, and layouts to provide intuitive and easy-to-understand information. At the same time, consider the user's interaction habits and ensure that the interface is concise and easy to operate;

[0083] (7) Use visualization tools or libraries to create charts based on data and design requirements. This includes setting coordinate axes, adding legends, setting titles and labels, etc., so that users can clearly understand the information conveyed by the chart;

[0084] (8) Add interactive features to charts, such as mouse hover prompts, data filtering, zooming and panning, to provide a richer and deeper interactive experience. These interactive features can help users explore data more deeply and discover patterns and trends hidden in the data;

[0085] (9) Integrate the created visualization chart into the logistics visualization interactive system to ensure seamless connection with other components and functions. Then, conduct system testing to check the accuracy and performance of the visualization display and ensure that users can use the function smoothly.

[0086] The present invention is described above by way of example in conjunction with the accompanying drawings. It is obvious that the specific implementation of the present invention is not limited to the above-mentioned method. As long as various non-substantial improvements are made using the inventive concept and technical solution of the present invention, or the inventive concept and technical solution are directly applied to other occasions without improvement, they are all within the protection scope of the present invention.

Claims

1. A logistics data visualization interactive method based on a large model, characterized by: It includes voice input module, voice recognition module, semantic understanding and intention recognition module, secondary intention matching function module, data query and processing module, visualization display module, voice feedback module and interaction and optimization module.

2. The method for interactive visualization of logistics data based on a large model according to claim 1, characterized in that: The user of the voice input module issues voice commands through the device to express the query, analysis or operation requirements for logistics data.

3. The method for interactive visualization of logistics data based on a large model according to claim 1, characterized in that: In the speech recognition module, after the system receives the user's speech signal, it uses the speech recognition model to convert the speech into text.

4. The method for interactive visualization of logistics data based on a large model according to claim 1, characterized in that: The semantic understanding and intention recognition module, after the text converted by the speech recognition module is sent to the big model for semantic understanding and intention recognition, the big model is based on the contextual knowledge of the logistics field and pre-trained semantic representation.

5. The method for interactive visualization of logistics data based on a large model according to claim 4 is characterized by: The secondary intent matching function module is based on a large model. It first uses the large model to parse the user input to obtain multiple intent labels, then generates candidate secondary intents based on the intent labels returned by the large model, and sorts the candidate secondary intents. Finally, it selects the secondary intent with the highest score as the final intent label.

6. The method for interactive visualization of logistics data based on a large model according to claim 1, characterized in that: The data query and processing module queries the corresponding data from the logistics database according to the user's intention, and performs necessary analysis and processing to generate the information required by the user.

7. The method for interactive visualization of logistics data based on a large model according to claim 1 is characterized by: In the visualization module, the system generates corresponding visualization charts and interfaces based on the query results to provide more intuitive and easy-to-understand information.

8. The method for interactive visualization of logistics data based on a large model according to claim 1 is characterized by: In the voice feedback module, the system feeds back the visualization results to the user in the form of voice.

9. The method for interactive visualization of logistics data based on a large model according to claim 1, characterized in that: In the interaction and optimization module, users conduct further voice interaction based on system feedback, continuously optimizing the model's recognition capability and interactive experience.

Images