Extraction analysis method and system based on semantic expression understanding

Abstract

The invention belongs to the technical field of semantic recognition, and discloses an extraction analysis method and system based on semantic expression understanding. The method comprises the following steps: performing semantic analysis and feature extraction on an original query statement input by a user in an interactive interface to generate a structured semantic unit; performing interaction scene judgment on the structured semantic unit based on a preset multi-dimensional judgment condition and then outputting a scene identifier; calling a corresponding target intention recognition strategy according to the scene identifier; inputting the structured semantic unit into a domain classification model, and outputting a plurality of candidate intentions and corresponding initial confidence; and determining a target user intention from the candidate intentions based on an interaction clarification result of the user and an intention query rule by using the initial confidence and a preset confidence threshold, and generating a standardized intention recognition result. According to the mode, the user semantics can be deeply understood, the recognition strategy is dynamically adjusted according to the context, and accurate, efficient and sustainable extraction analysis is carried out on the user intention through man-machine cooperation and closed-loop feedback.

Description

Technical Field

[0001] This invention relates to the field of semantic recognition technology, and in particular to an extraction and analysis method and system based on semantic expression understanding. Background Technology

[0002] With the widespread adoption of applications such as intelligent customer service, voice assistants, and intelligent search, enabling machines to accurately understand users' natural language expressions and identify their core intentions has become crucial for improving the human-computer interaction experience. Existing user intent recognition methods mainly suffer from the following limitations: Rule-based methods rely on experts manually writing a large number of pattern matching rules. While this approach achieves high accuracy in small-scale, fixed scenarios, it suffers from poor flexibility, high maintenance costs, and difficulty in adapting to new expressions or business domains.

[0003] Deep learning-based methods utilize classification models (such as CNN, RNN, and BERT) for end-to-end intent classification. While these methods exhibit strong generalization capabilities, their decision-making process is like a "black box," lacking effective interactive clarification mechanisms for predictions with low confidence. Furthermore, these models are typically static, making it difficult to perform real-time self-optimization using online feedback data, resulting in "cold start" and insensitivity to new intents.

[0004] Hybrid approaches, while combining rules and models, are typically loosely structured and lack adaptability to different scenarios. They lack a unified framework to dynamically select the most appropriate analysis strategy based on the context and scenario of the dialogue, and they cannot form an effective closed-loop optimization to improve long-term performance when faced with ambiguous or polysemous queries.

[0005] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0006] The main objective of this invention is to provide an extraction and analysis method and system based on semantic expression understanding, which aims to solve the technical problems of existing intent recognition technologies, such as poor flexibility, lack of interactive clarification capabilities, and inability to continuously self-optimize.

[0007] To achieve the above objectives, the present invention provides an extraction and analysis method based on semantic expression understanding, which includes the following steps: Obtain the original query statement entered by the user in the interactive interface; The original query statement is semantically parsed and its features are extracted to generate a structured semantic unit, which includes a core action, a target entity, and modifying attributes. The structured semantic unit is subjected to interaction scenario judgment based on preset multidimensional judgment conditions, and a scenario identifier is output. The corresponding target intent recognition strategy is called from the intent recognition strategy library according to the scene identifier. The target intent recognition strategy includes a domain classification model, intent probing rules, and result assembly template. The structured semantic unit is input into the domain classification model, and one or more candidate intentions and their corresponding initial confidence scores are output. Based on the initial confidence level and the preset confidence threshold, the intent probing rules are applied to interact with the user for clarification; Based on the interactive clarification results, the target user intent is determined from the candidate intents, and a standardized intent recognition result is generated by assembling a template according to the results.

[0008] In one embodiment, the step of performing semantic parsing and feature extraction on the original query statement to generate structured semantic units includes: The original query statement is segmented, stop words are removed, and part-of-speech tags are applied to generate a standard token sequence; Based on the standard token sequence, the entities corresponding to the original query statement are extracted using a named entity recognition model, and a list of target entities is output. Dependency parsing is performed on the standard token sequence to identify the core verb or verb phrase in the sentence, which is used as the core action; Identify and extract adjectives, adverbs, or prepositional phrases associated with the core action and each entity in the target entity list, wherein the adjectives, adverbs, or prepositional phrases are used as modifying attributes; The core action, the target entity list, and the modifier attributes are associated and encapsulated to obtain a structured semantic unit.

[0009] In one embodiment, the preset multidimensional judgment conditions include a first dimension, a second dimension, a third dimension, a fourth dimension, and a fifth dimension. The step of judging the interaction scenario of the structured semantic unit based on the preset multidimensional judgment conditions and outputting a scenario identifier includes: Based on the first dimension, determine whether the core action in the structured semantic unit belongs to a predefined set of specific business actions; Based on the second dimension, it is determined whether the entities in the target entity list of the structured semantic unit match the core concepts in the preset domain ontology library; Based on the third dimension, determine whether there is a historical intent in the current dialogue session, and analyze the contextual constraint relationship of the historical intent on the current query; The source channel of the original query statement is determined based on the fourth dimension, and the source channel includes voice assistant, online customer service or search box; The fifth dimension is used to determine the current round of the dialogue and whether follow-up questions and clarifications have been triggered in previous rounds. The scene identifier is output based on the combined judgment results from the five dimensions.

[0010] In one embodiment, the step of applying the intent probing rules to the user for clarification based on the initial confidence level and a preset confidence threshold includes: The initial confidence level is compared with a first confidence threshold and a second confidence threshold corresponding to a preset confidence threshold, respectively. The first confidence threshold is the upper limit of the threshold, and the second confidence threshold is the lower limit of the threshold. If a single candidate intent has an initial confidence level higher than the first confidence level threshold, no follow-up questioning will be initiated, and the candidate intent will be directly identified as the intent to be executed. If the initial confidence level of all candidate intentions is lower than the second confidence level threshold, the intention is determined to be unclear, and the general clarification script in the intention follow-up rule set is invoked to request the user to restate the statement. If there are multiple candidate intentions and their initial confidence levels are between the second confidence threshold and the first confidence threshold, then the selective clarification script for this ambiguous situation is invoked to list the possible intentions for the user to choose from. Receive clarification feedback from the user, update the candidate intent and its confidence level, and record the feedback information in the context of the current dialogue session.

[0011] In one embodiment, the method further includes: The identified target user intent is bound to the session ID of the current conversation and stored; Analyze the logical sequence relationship between the target user's intent and previous intents in the same session to construct the intent achievement path; Based on the user's feedback on follow-up questions, the accuracy of each step in the intent recognition process is calculated, and the performance index of the target intent recognition strategy is recorded. The domain classification model in the target intent recognition strategy is incrementally trained and optimized based on the performance index. The optimized domain classification model is updated to the intent recognition strategy library for use in subsequent calls to target intent recognition strategies.

[0012] In one embodiment, incrementally training the domain classification model in the target intent recognition strategy based on the performance index includes: Obtain training samples within an optimization cycle, the training samples including the structured semantic units, the target user intent confirmed by user clarification feedback, and the corresponding session context features; The training samples are subjected to data augmentation and denoising to generate an incremental training dataset; Load the current domain classification model and model parameters from the intent recognition strategy library; The loaded model is incrementally trained using the incremental training dataset. The incrementally trained domain classification model is evaluated using a validation set. Once the evaluation metrics meet the preset improvement criteria, the incrementally trained domain classification model is updated to the intent recognition strategy library.

[0013] Furthermore, to achieve the above objectives, this invention also proposes an extraction and analysis system based on semantic expression understanding. This system is applied to the extraction and analysis method based on semantic expression understanding described above. The system includes: The acquisition module is used to acquire the original query statement entered by the user in the interactive interface; The extraction module is used to perform semantic parsing and feature extraction on the original query statement to generate structured semantic units, which include core actions, target entities and modifying attributes. The judgment module is used to judge the interaction scenario of the structured semantic unit based on preset multi-dimensional judgment conditions and output the scenario identifier; The calling module is used to call the corresponding target intent recognition strategy from the intent recognition strategy library according to the scene identifier. The target intent recognition strategy includes a domain classification model, intent probing rules, and result assembly template. The input module is used to input the structured semantic units into the domain classification model and output one or more candidate intentions and their corresponding initial confidence scores. The interaction module is used to interact with the user and clarify based on the initial confidence level and the preset confidence level threshold, applying the intent probing rules. The output module is used to determine the target user intent from the candidate intents based on the interaction clarification results, and to assemble a template according to the results to generate a standardized intent recognition result.

[0014] In one embodiment, the extraction module is used to perform word segmentation, stop word removal, and part-of-speech tagging on the original query statement to generate a standard token sequence; Based on the standard token sequence, the entities corresponding to the original query statement are extracted using a named entity recognition model, and a list of target entities is output. Dependency parsing is performed on the standard token sequence to identify the core verb or verb phrase in the sentence, which is used as the core action; Identify and extract adjectives, adverbs, or prepositional phrases associated with the core action and each entity in the target entity list, wherein the adjectives, adverbs, or prepositional phrases are used as modifying attributes; The core action, the target entity list, and the modifier attributes are associated and encapsulated to obtain a structured semantic unit.

[0015] Furthermore, to achieve the above objectives, the present invention also proposes an extraction and analysis device based on semantic expression understanding. The extraction and analysis device based on semantic expression understanding includes: a memory, a processor, and an extraction and analysis program based on semantic expression understanding stored in the memory and executable on the processor. The extraction and analysis program based on semantic expression understanding is configured to implement the steps of the extraction and analysis method based on semantic expression understanding as described above.

[0016] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing an extraction and analysis program based on semantic expression understanding, wherein when the extraction and analysis program based on semantic expression understanding is executed by a processor, it implements the steps of the extraction and analysis method based on semantic expression understanding as described above.

[0017] This invention generates structured semantic units by performing semantic parsing and feature extraction on the original query entered by the user in the interactive interface; after judging the interaction scenario of the structured semantic units based on preset multi-dimensional judgment conditions, a scenario identifier is output; the corresponding target intent recognition strategy is invoked according to the scenario identifier; the structured semantic units are input into a domain classification model, which outputs several candidate intents and their corresponding initial confidence scores; based on the initial confidence scores and preset confidence thresholds, intent probing rules are applied, along with the user's interaction clarification results, to determine the target user intent from the candidate intents, and a standardized intent recognition result is generated. This approach enables a deep understanding of user semantics, dynamic adjustment of the recognition strategy based on context, and accurate, efficient, and continuously evolving extraction and analysis of user intent through human-computer collaboration and closed-loop feedback. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the first embodiment of the extraction and analysis method based on semantic expression understanding of the present invention; Figure 2 This is a structural block diagram of the first embodiment of the semantic expression understanding-based extraction and analysis system of the present invention.

[0019] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0020] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0021] This invention provides an extraction and analysis method based on semantic expression understanding, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the extraction and analysis method based on semantic expression understanding according to the present invention.

[0022] In this embodiment, the extraction and analysis method based on semantic expression understanding includes the following steps: Step S10: Obtain the original query statement entered by the user in the interactive interface.

[0023] In this embodiment, the execution entity is a semantic expression understanding-based extraction and analysis device. This semantic expression understanding-based extraction and analysis device has functions such as data processing, data communication, and program execution. The semantic expression understanding-based extraction and analysis device can be a computer terminal device or other network device, or other devices with similar functions. This embodiment does not limit this.

[0024] It's important to note that with the widespread adoption of applications like intelligent customer service, voice assistants, and intelligent search, enabling machines to accurately understand users' natural language expressions and identify their core intentions has become crucial for improving the human-computer interaction experience. Existing user intent recognition methods mainly suffer from the following limitations: Rule-based methods: These rely on experts manually writing a large number of pattern matching rules. While this method achieves high accuracy in small-scale, fixed scenarios, it lacks flexibility, has high maintenance costs, and struggles to adapt to new expressions or business domains. Deep learning-based methods: These utilize classification models (such as CNN, RNN, and BERT) for end-to-end intent classification. Although these methods have strong generalization capabilities, their decision-making process is like a "black box," lacking effective interactive clarification mechanisms for predictions with low confidence. Furthermore, the models are typically static, making it difficult to utilize online feedback data for real-time self-optimization, resulting in "cold start" and insensitivity to new intents. Hybrid methods: While combining rules and models, these methods are usually loosely structured and have poor scenario adaptability. They lack a unified framework to dynamically select the most appropriate analysis strategy based on the context and scenario of the dialogue, and they cannot form an effective closed-loop optimization to improve long-term performance when faced with ambiguous or polysemous queries.

[0025] To address the aforementioned technical issues, this embodiment performs semantic parsing and feature extraction on the original query statement input by the user in the interactive interface to generate structured semantic units. Based on preset multi-dimensional judgment conditions, the structured semantic units are used to determine the interaction scenario and output a scenario identifier. The corresponding target intent recognition strategy is invoked based on the scenario identifier. The structured semantic units are input into a domain classification model, which outputs several candidate intents and their corresponding initial confidence scores. Based on the initial confidence scores and preset confidence thresholds, intent probing rules are applied, along with the user's interaction clarification results, to determine the target user intent from the candidate intents and generate standardized intent recognition results. This approach enables a deep understanding of user semantics, dynamic adjustment of recognition strategies based on context, and accurate, efficient, and continuously evolving extraction and analysis of user intent through human-computer collaboration and closed-loop feedback. Specifically, it can be implemented as follows.

[0026] In this specific implementation, it is necessary to first obtain the original query statement entered by the user on the interactive interface. The original query statement entered by the user can be received through the interactive interface (such as a chat window or voice interface), for example, "Book me a cheap flight to Beijing tomorrow morning".

[0027] Step S20: Perform semantic parsing and feature extraction on the original query statement to generate structured semantic units.

[0028] It should be noted that the structured semantic unit in this embodiment includes core actions, target entities, and modifying attributes.

[0029] In specific implementation, the process of generating structured semantic units involves: segmenting the original query statement, removing stop words, and performing part-of-speech tagging to generate a standard token sequence; based on the standard token sequence, extracting the entities corresponding to the original query statement using a named entity recognition model and outputting a target entity list; performing dependency parsing on the standard token sequence to identify the core verb or verb phrase in the sentence, which serves as the core action; identifying and extracting adjectives, adverbs, or prepositional phrases associated with the core action and each entity in the target entity list, which serve as modifying attributes; and associating and encapsulating the core action, the target entity list, and the modifying attributes to obtain the structured semantic unit.

[0030] It should be noted that, taking the above example, word segmentation and part-of-speech tagging decompose the sentence into a token sequence, such as [help / v, I / r, order / v, one / m, tomorrow / t, morning / t, fly / v, Beijing / ns, of / u, cheap / a, ticket / n]. Named entity recognition: extracts entities such as "tomorrow morning" (time), "Beijing" (location), and "ticket" (item) to form a target entity list. Dependency parsing: identifies the core verb "order" as the core action. Extracting modifier attributes: identifies the adjective "cheap" related to "ticket" as a modifier attribute. The encapsulation process involves associating and encapsulating the core action: order, the target entity list: [time: tomorrow morning, destination: Beijing, item: ticket], and the modifier attribute: [price: cheap] into a structured semantic unit. This unit provides a standardized, semantically rich information carrier for subsequent processing.

[0031] Step S30: Based on preset multi-dimensional judgment conditions, perform interaction scenario judgment on the structured semantic unit and output the scenario identifier.

[0032] In this embodiment, the preset multi-dimensional judgment conditions include a first dimension, a second dimension, a third dimension, a fourth dimension, and a fifth dimension. Different dimensions are judged differently, and the scene identifier is obtained by combining the judgment results of each dimension. The specific process is as follows: Based on the first dimension, it is judged whether the core action in the structured semantic unit belongs to a predefined set of specific business actions; based on the second dimension, it is judged whether the entities in the target entity list in the structured semantic unit match the core concepts in the preset domain ontology library; based on the third dimension, it is judged whether there is a historical intent in the current dialogue session, and the contextual constraint relationship of the historical intent on the current query is analyzed; based on the fourth dimension, it is judged whether the source channel of the original query statement is the source channel, which includes voice assistant, online customer service, or search box; based on the fifth dimension, it is judged whether the current dialogue round and whether follow-up questions and clarifications have been triggered in the previous rounds; and the scene identifier is output by combining the judgment results of the five dimensions.

[0033] It's important to note that the first dimension targets the core action, such as determining if "book" belongs to the "reservation" business action set. The second dimension targets the target entity, such as matching the entities "flight tickets" and "Beijing" to the "travel" domain ontology. The third dimension targets the historical context, such as checking if a prior intent (e.g., "query flights") exists in the current session; if so, the current intent may be a deepening of that prior intent. The fourth dimension targets the source channel, such as determining if the query comes from "online customer service." The fifth dimension targets the dialogue round, such as determining if this is the first inquiry. Based on the above, the system outputs a scenario identifier, such as Travel_Booking_FirstRound.

[0034] Step S40: Call the corresponding target intent recognition strategy from the intent recognition strategy library according to the scene identifier.

[0035] In the specific implementation, after obtaining the scene identifier, the corresponding target intent recognition strategy can be invoked by combining it with the intent recognition strategy library. For example, based on the scene identifier Travel_Booking_FirstRound, the corresponding target intent recognition strategy is invoked from the intent recognition strategy library. This strategy includes: Domain classification model: a fine-grained intent classification model specifically for the "travel booking" scenario; Intent follow-up rules: defining how to follow up when the intent is ambiguous (e.g., when the confidence levels of "book a flight" and "book a hotel" are similar); Result assembly template: defining the fields that the standardized output should include, such as {Intent: "Flight Booking", Slot: {Date: "AA", Destination: "BB", Budget: "CC"}}.

[0036] Step S50: Input the structured semantic unit into the domain classification model and output one or more candidate intentions and their corresponding initial confidence scores.

[0037] In the specific implementation, structured semantic units are input into the domain classification model. The model may output multiple candidate intentions and their initial confidence scores, such as: [flight booking: 0.75, hotel booking: 0.20, transportation query: 0.05].

[0038] Step S60: Based on the initial confidence level and the preset confidence level threshold, apply the intent probing rules to interact with the user for clarification.

[0039] In specific implementation, the interactive clarification process includes comparing the initial confidence level with a first confidence threshold and a second confidence threshold corresponding to a preset confidence threshold, where the first confidence threshold is the upper limit and the second confidence threshold is the lower limit. If the initial confidence level of a single candidate intent is higher than the first confidence threshold, no follow-up questioning is initiated, and the candidate intent is directly identified as the intent to be executed. If the initial confidence level of all candidate intents is lower than the second confidence threshold, the intent is determined to be unclear, and the general clarification script in the intent follow-up questioning rule set is invoked to request the user to restate the intent. If there are multiple candidate intents and their initial confidence levels are between the second confidence threshold and the first confidence threshold, the selective clarification script for this ambiguous situation is invoked to list possible intents for the user to choose from. The user's clarification feedback is received, the candidate intents and their confidence levels are updated, and the feedback information is recorded in the context of the current dialogue session.

[0040] It should be noted that in this embodiment, the thresholds can be set as a first confidence threshold of 0.8 and a second confidence threshold of 0.6. Since the confidence level of 0.75 is between the second confidence threshold and the first confidence threshold, and there are other candidate intentions, the system triggers a selective clarification statement, such as: "Do you want to book a flight or check a hotel?" The system receives the user's feedback "book a flight" and updates the confidence level of the candidate intention (flight booking: 0.95) accordingly, while recording this interaction in the session context.

[0041] Step S70: Determine the target user intent from the candidate intents based on the interaction clarification results, and assemble a template according to the results to generate a standardized intent recognition result.

[0042] In the specific implementation, based on the user's clarification feedback, the system determines the target user's intent as "flight booking". Subsequently, according to the result assembly template, slot information is extracted from the structured semantic units to generate the final result: {intent: "flight booking", confidence level: 0.95, slot: {date: "tomorrow morning", destination: "Beijing", budget: "cheap"}}.

[0043] In one embodiment, closed-loop optimization can also be implemented. Specifically, the identified target user intent is bound to and stored with the session ID of the current dialogue session; the logical sequence relationship between the target user intent and the preceding intents in the same session is analyzed to construct the intent achievement path; based on the user's feedback to follow-up questions, the accuracy of each step in the intent recognition process is calculated, and the performance index of the target intent recognition strategy is recorded; the domain classification model in the target intent recognition strategy is incrementally trained according to the performance index for optimization and adjustment; and the optimized domain classification model is updated to the intent recognition strategy library for subsequent invocation of the target intent recognition strategy.

[0044] The step of incrementally training the domain classification model in the target intent recognition strategy based on the performance index includes: acquiring training samples within an optimization cycle, wherein the training samples include the structured semantic units, the target user intent confirmed by user clarification feedback, and the corresponding session context features; performing data augmentation and denoising processing on the training samples to generate an incremental training dataset; loading the current domain classification model and model parameters from the intent recognition strategy library; incrementally training the loaded model using the incremental training dataset; and evaluating the incrementally trained domain classification model using a validation set, wherein the incrementally trained domain classification model is updated to the intent recognition strategy library after the evaluation index meets the preset improvement criteria.

[0045] It should be noted that intent context association involves, for example, binding the current intent to the session ID and analyzing its sequential relationship in the preceding dialogue. Performance evaluation includes, for example, calculating the accuracy of the recognition (the user ultimately confirmed the intent) and recording the policy performance metrics. Incremental training includes, for example, sample collection: saving the structured semantic units of this interaction and the user's confirmed target user intent as high-quality training samples.

[0046] Data Processing: Data augmentation (such as synonym replacement) and denoising are performed on the samples. Model Training: The current domain classification model is loaded and fine-tuned using a new incremental training dataset to enhance its ability to distinguish between "flight booking" and "hotel booking". Model Evaluation and Deployment: The new model is evaluated using a validation set. If the performance improvement meets the target, it is updated back to the intent recognition strategy library, replacing the old model. Through the above implementation methods, this invention achieves a complete closed loop from semantic understanding to policy execution and self-optimization, significantly improving the accuracy, adaptability, and long-term service capability of intent recognition.

[0047] Furthermore, this embodiment illustrates the above process with a practical application scenario example. For instance, its application in a smart coffee machine voice assistant scenario. User input: The user says to the coffee machine, "A warm latte, not too strong." The semantic parsing process is as follows: Core action: make; Target entity list: [Drink: Latte]; Modifying attributes: [Temperature: Warm; Concentration: Weak]. Scenario judgment process: The core action "make" is judged to belong to the "drink making" scenario, the entity "latte" belongs to the "coffee" domain, the source is "voice assistant," and the scenario identifier Beverage_Making_Voice is output. Strategy invocation: The target intent recognition strategy for drink making is invoked, and its domain classification model can distinguish intents such as "make coffee" and "query status." Classification and confidence are output by the model, for example, [Make coffee: 0.70, Modify order: 0.25]. The initial confidence of 0.70 does not reach the high threshold (0.85). Interaction clarification: The system triggers a follow-up question: "Okay, I'll make you a warm, weak latte. Please say 'yes' to confirm." The user replies: "Yes." The result is as follows: The system generates the final intent: {Intent: "Make coffee", Slot: {Type: "Latte", Temperature: "Warm", Strength: "Weak"}}, and drives the coffee machine to execute the task. The closed-loop optimization is as follows: This successful interaction is recorded as a training sample. The system discovers a strong correlation between modifiers such as "warm" and "weak" and the intent to "make coffee." Through incremental training and model optimization, future recognition of similar expressions will be more accurate and faster (potentially without needing further questioning).

[0048] This embodiment generates structured semantic units by performing semantic parsing and feature extraction on the original query entered by the user in the interactive interface; after judging the interaction scenario based on preset multi-dimensional judgment conditions, a scenario identifier is output; the corresponding target intent recognition strategy is invoked according to the scenario identifier; the structured semantic units are input into a domain classification model, which outputs several candidate intents and their corresponding initial confidence scores; based on the initial confidence scores and preset confidence thresholds, intent probing rules are applied, along with the user's interaction clarification results, to determine the target user intent from the candidate intents, and a standardized intent recognition result is generated. This approach enables a deep understanding of user semantics, dynamic adjustment of the recognition strategy based on context, and accurate, efficient, and continuously evolving extraction and analysis of user intent through human-computer collaboration and closed-loop feedback.

[0049] Furthermore, this embodiment of the invention also proposes a storage medium storing an extraction and analysis program based on semantic expression understanding. When the extraction and analysis program based on semantic expression understanding is executed by a processor, it implements the steps of the extraction and analysis method based on semantic expression understanding as described above.

[0050] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the semantic expression understanding-based extraction and analysis system of the present invention.

[0051] like Figure 2 As shown, the semantic expression understanding-based extraction and analysis system proposed in this embodiment of the invention includes: The acquisition module 10 is used to acquire the original query statement entered by the user in the interactive interface; Extraction module 20 is used to perform semantic parsing and feature extraction on the original query statement to generate structured semantic units, wherein the structured semantic units include core actions, target entities and modifying attributes; The judgment module 30 is used to judge the interaction scenario of the structured semantic unit based on preset multi-dimensional judgment conditions and output the scenario identifier; The calling module 40 is used to call the corresponding target intent recognition strategy from the intent recognition strategy library according to the scene identifier. The target intent recognition strategy includes a domain classification model, intent inquiry rules and result assembly template. Input module 50 is used to input the structured semantic unit into the domain classification model and output one or more candidate intentions and their corresponding initial confidence scores; The interaction module 60 is used to interact with the user and clarify based on the initial confidence level and the preset confidence level threshold, applying the intent probing rules. Output module 70 is used to determine the target user intent from the candidate intents based on the interaction clarification results, and to assemble a template according to the results to generate a standardized intent recognition result.

[0052] This embodiment generates structured semantic units by performing semantic parsing and feature extraction on the original query entered by the user in the interactive interface; after judging the interaction scenario based on preset multi-dimensional judgment conditions, a scenario identifier is output; the corresponding target intent recognition strategy is invoked according to the scenario identifier; the structured semantic units are input into a domain classification model, which outputs several candidate intents and their corresponding initial confidence scores; based on the initial confidence scores and preset confidence thresholds, intent probing rules are applied, along with the user's interaction clarification results, to determine the target user intent from the candidate intents, and a standardized intent recognition result is generated. This approach enables a deep understanding of user semantics, dynamic adjustment of the recognition strategy based on context, and accurate, efficient, and continuously evolving extraction and analysis of user intent through human-computer collaboration and closed-loop feedback.

[0053] In one embodiment, the extraction module is used to perform word segmentation, stop word removal, and part-of-speech tagging on the original query statement to generate a standard token sequence; Based on the standard token sequence, the entities corresponding to the original query statement are extracted using a named entity recognition model, and a list of target entities is output. Dependency parsing is performed on the standard token sequence to identify the core verb or verb phrase in the sentence, which is used as the core action; Identify and extract adjectives, adverbs, or prepositional phrases associated with the core action and each entity in the target entity list, wherein the adjectives, adverbs, or prepositional phrases are used as modifying attributes; The core action, the target entity list, and the modifier attributes are associated and encapsulated to obtain a structured semantic unit.

[0054] This application embodiment also provides an extraction and analysis device based on semantic expression understanding, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other through the communication bus. The memory is used to store the extraction and analysis program based on semantic expression understanding. When the processor executes the program stored in the memory, it implements the above-mentioned extraction and analysis method based on semantic expression understanding.

[0055] The communication bus mentioned in the semantic expression understanding-based extraction and analysis device can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0056] The communication interface is used for communication between the aforementioned semantic expression understanding-based extraction and analysis device and other devices.

[0057] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0058] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0059] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0060] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0061] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0062] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0063] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0064] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0065] In addition, for technical details not described in detail in this embodiment, please refer to the extraction and analysis method based on semantic expression understanding provided in any embodiment of the present invention, which will not be repeated here.

[0066] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0067] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0068] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0069] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

[0070] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above method.

Claims

1. A semantic expression understanding-based extraction and analysis method, characterized in that, The extraction and analysis method based on semantic expression understanding includes: Obtain the original query statement entered by the user in the interactive interface; The original query statement is semantically parsed and its features are extracted to generate a structured semantic unit, which includes a core action, a target entity, and modifying attributes. The structured semantic unit is subjected to interaction scenario judgment based on preset multidimensional judgment conditions, and a scenario identifier is output. The corresponding target intent recognition strategy is called from the intent recognition strategy library according to the scene identifier. The target intent recognition strategy includes a domain classification model, intent probing rules, and result assembly template. The structured semantic unit is input into the domain classification model, and one or more candidate intentions and their corresponding initial confidence scores are output. Based on the initial confidence level and the preset confidence threshold, the intent probing rules are applied to interact with the user for clarification; Based on the interactive clarification results, the target user intent is determined from the candidate intents, and a standardized intent recognition result is generated by assembling a template according to the results.

2. The extraction and analysis method based on semantic expression understanding as described in claim 1, characterized in that, The step of performing semantic parsing and feature extraction on the original query statement to generate structured semantic units includes: The original query statement is segmented, stop words are removed, and part-of-speech tags are applied to generate a standard token sequence; Based on the standard token sequence, the entities corresponding to the original query statement are extracted using a named entity recognition model, and a list of target entities is output. Dependency parsing is performed on the standard token sequence to identify the core verb or verb phrase in the sentence, which is used as the core action; Identify and extract adjectives, adverbs, or prepositional phrases associated with the core action and each entity in the target entity list, wherein the adjectives, adverbs, or prepositional phrases are used as modifying attributes; The core action, the target entity list, and the modifier attributes are associated and encapsulated to obtain a structured semantic unit.

3. The extraction and analysis method based on semantic expression understanding as described in claim 1, characterized in that, The preset multidimensional judgment conditions include a first dimension, a second dimension, a third dimension, a fourth dimension, and a fifth dimension. The interaction scenario judgment of the structured semantic unit based on the preset multidimensional judgment conditions, and the output of a scenario identifier, include: Based on the first dimension, determine whether the core action in the structured semantic unit belongs to a predefined set of specific business actions; Based on the second dimension, it is determined whether the entities in the target entity list of the structured semantic unit match the core concepts in the preset domain ontology library; Based on the third dimension, determine whether there is a historical intent in the current dialogue session, and analyze the contextual constraint relationship of the historical intent on the current query; The source channel of the original query statement is determined based on the fourth dimension, and the source channel includes voice assistant, online customer service or search box; The fifth dimension is used to determine the current round of the dialogue and whether follow-up questions and clarifications have been triggered in previous rounds. The scene identifier is output based on the combined judgment results from the five dimensions.

4. The extraction and analysis method based on semantic expression understanding as described in claim 1, characterized in that, The step of applying the intent probing rules to the user for clarification based on the initial confidence level and a preset confidence threshold includes: The initial confidence level is compared with a first confidence threshold and a second confidence threshold corresponding to a preset confidence threshold, respectively. The first confidence threshold is the upper limit of the threshold, and the second confidence threshold is the lower limit of the threshold. If a single candidate intent has an initial confidence level higher than the first confidence level threshold, no follow-up questioning will be initiated, and the candidate intent will be directly identified as the intent to be executed. If the initial confidence level of all candidate intentions is lower than the second confidence level threshold, the intention is determined to be unclear, and the general clarification script in the intention follow-up rule set is invoked to request the user to restate the statement. If there are multiple candidate intentions and their initial confidence levels are between the second confidence threshold and the first confidence threshold, then the selective clarification script for this ambiguous situation is invoked to list the possible intentions for the user to choose from. Receive clarification feedback from the user, update the candidate intent and its confidence level, and record the feedback information in the context of the current dialogue session.

5. The extraction and analysis method based on semantic expression understanding as described in claim 4, characterized in that, The method further includes: The identified target user intent is bound to the session ID of the current conversation and stored; Analyze the logical sequence relationship between the target user's intent and previous intents in the same session to construct the intent achievement path; Based on the user's feedback on follow-up questions, the accuracy of each step in the intent recognition process is calculated, and the performance index of the target intent recognition strategy is recorded. The domain classification model in the target intent recognition strategy is incrementally trained and optimized based on the performance index. The optimized domain classification model is updated to the intent recognition strategy library for use in subsequent calls to target intent recognition strategies.

6. The extraction and analysis method based on semantic expression understanding as described in claim 5, characterized in that, The incremental training of the domain classification model in the target intent recognition strategy based on the performance index includes: Obtain training samples within an optimization cycle, the training samples including the structured semantic units, the target user intent confirmed by user clarification feedback, and the corresponding session context features; The training samples are subjected to data augmentation and denoising to generate an incremental training dataset; Load the current domain classification model and model parameters from the intent recognition strategy library; The loaded model is incrementally trained using the incremental training dataset. The incrementally trained domain classification model is evaluated using a validation set. Once the evaluation metrics meet the preset improvement criteria, the incrementally trained domain classification model is updated to the intent recognition strategy library.

7. A semantic expression understanding-based extraction and analysis system, characterized in that, The semantic expression understanding-based extraction and analysis system is applied to the semantic expression understanding-based extraction and analysis method as described in any one of claims 1 to 6, wherein the system comprises: The acquisition module is used to acquire the original query statement entered by the user in the interactive interface; The extraction module is used to perform semantic parsing and feature extraction on the original query statement to generate structured semantic units, which include core actions, target entities and modifying attributes. The judgment module is used to judge the interaction scenario of the structured semantic unit based on preset multi-dimensional judgment conditions and output the scenario identifier; The calling module is used to call the corresponding target intent recognition strategy from the intent recognition strategy library according to the scene identifier. The target intent recognition strategy includes a domain classification model, intent probing rules, and result assembly template. The input module is used to input the structured semantic units into the domain classification model and output one or more candidate intentions and their corresponding initial confidence scores. The interaction module is used to interact with the user and clarify based on the initial confidence level and the preset confidence level threshold, applying the intent probing rules. The output module is used to determine the target user intent from the candidate intents based on the interaction clarification results, and to assemble a template according to the results to generate a standardized intent recognition result.

8. The extraction and analysis system based on semantic expression understanding as described in claim 7, characterized in that, The extraction module is used to segment the original query statement, remove stop words and perform part-of-speech tagging, and generate a standard token sequence. Based on the standard token sequence, the entities corresponding to the original query statement are extracted using a named entity recognition model, and a list of target entities is output. Dependency parsing is performed on the standard token sequence to identify the core verb or verb phrase in the sentence, which is used as the core action; Identify and extract adjectives, adverbs, or prepositional phrases associated with the core action and each entity in the target entity list, wherein the adjectives, adverbs, or prepositional phrases are used as modifying attributes; The core action, the target entity list, and the modifier attributes are associated and encapsulated to obtain a structured semantic unit.

9. A semantic expression understanding-based extraction and analysis device, characterized in that, The semantic expression understanding-based extraction and analysis device includes: a memory, a processor, and a semantic expression understanding-based extraction and analysis program stored in the memory and executable on the processor, wherein the semantic expression understanding-based extraction and analysis program is configured to implement the steps of the semantic expression understanding-based extraction and analysis method as described in any one of claims 1 to 6.

10. A storage medium, characterized in that, The storage medium stores an extraction and analysis program based on semantic expression understanding, which, when executed by a processor, implements the steps of the extraction and analysis method based on semantic expression understanding as described in any one of claims 1 to 6.

Images