Natural language intention recognition system based on rule and large model dynamic collaboration

Abstract

The invention discloses a natural language intention recognition system based on rule and large model dynamic collaboration, and relates to the technical field of artificial intelligence, and the method comprises the following steps: an input preprocessing layer is used for converting a user dialogue voice stream into text information; the complexity judgment module is used for judging whether the text information is complex or not, and the matching module is used for performing rule matching on high-priority rule base rules under a multi-level rule tree architecture; the rule engine module is used for performing intention recognition on the text information based on the rule matching method selected by the matching module under the condition that the text information is not complex so as to obtain an intention recognition result, and the large model processing module is used for performing intention recognition on the text information under the condition that the text information is complex so as to obtain an intention recognition result; and the output layer is used for outputting an intention recognition result. The intention recognition method is beneficial for solving the problems of relatively low intention recognition accuracy, relatively slow response speed and relatively poor processing capability for services in a specific field in the prior art.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a natural language intention recognition system based on dynamic collaboration of rules and large models. Background Art

[0002] Driven by the wave of digitalization, intelligent call systems have become a key hub for communication and interaction between enterprises and customers in fields such as finance, telecommunications, and e-commerce. Intent recognition technology is the core support of this system, and its performance directly determines communication efficiency and service quality.

[0003] Leveraging their powerful natural language understanding capabilities, large models, trained on massive amounts of data, build complex semantic understanding models, significantly improving the accuracy and generalization of intent recognition. However, this technology also has numerous limitations. For one thing, large models require massive computational effort, a complex reasoning process, and a response time generally exceeding 500ms. In intelligent outbound call scenarios with extremely high real-time requirements, this delay prevents the system from responding promptly, significantly impacting outbound call efficiency and customer experience. Furthermore, large models lack effective control over business processes and are unable to enforce constraints on the conversation flow. Conversations often deviate from the preset path. Furthermore, when identifying domain-specific terms, large models are prone to misunderstandings due to issues with the training data or model structure, leading to intent recognition errors.

[0004] Therefore, relying solely on large model technology is difficult to meet the stringent requirements of intelligent call systems for intent recognition accuracy, efficiency, and real-time performance. There is an urgent need for a more advanced, efficient, and accurate intent recognition technology to improve the performance and service quality of intelligent call systems. Summary of the Invention

[0005] In view of this, the present invention proposes a natural language intent recognition system based on dynamic collaboration of rules and large models, which can achieve accurate and efficient recognition of natural language intent.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A natural language intent recognition system based on dynamic collaboration between rules and large models includes an input preprocessing layer, a dynamic routing decision layer, an intent recognition layer, and an output layer, wherein:

[0008] The input preprocessing layer is used to convert the user conversation voice stream into text information;

[0009] The dynamic routing decision layer includes a complexity judgment module and a matching module, wherein the complexity judgment module is used to judge whether the text information is complex, and the matching module is used to match the rules of the high-priority rule base under a multi-level rule tree architecture. The rule matching methods include keyword matching, regular matching and semantic template matching;

[0010] The intent recognition layer includes a rule engine module and a large model processing module, wherein the rule engine module is used to perform intent recognition on the text information based on the rule matching method selected by the matching module when the text information is not complex, so as to obtain an intent recognition result; the large model processing module is used to perform intent recognition on the text information when the text information is complex, so as to obtain an intent recognition result;

[0011] The output layer is used to output the intent recognition result.

[0012] On the basis of the above technical solution, the present invention can also be improved as follows:

[0013] Optionally, the input preprocessing layer is further used to:

[0014] Convert user conversation voice stream into text information;

[0015] Performing cleaning and normalization processing on the text information;

[0016] Record historical intent and conversation content based on context to improve the consistency of multi-round conversations.

[0017] Optionally, the input preprocessing layer is also used to analyze the text information, obtain text length information, grammatical structure information and semantic understanding information of the text information, and determine whether the text information is complex based on the text length information, grammatical structure information and semantic understanding information.

[0018] Optionally, the dynamic routing decision layer is further configured to determine whether the text information has a reversal if the text information is not complex;

[0019] The rule engine module is further configured to perform intent recognition on the text information based on the rule matching method selected by the matching module when the inversion does not exist;

[0020] The large model processing module is further configured to perform intent recognition on the text information based on the large model when there is a reversal.

[0021] Optionally, the dynamic routing decision layer is further used to calculate the intent recognition delay;

[0022] The intent recognition layer is also used to use a degradation strategy to perform intent recognition when the intent recognition delay is greater than a preset threshold.

[0023] Optionally, the matching module is further configured to calculate a basic weight of keyword matching by formula (1) when keyword matching is used to perform rule matching on a high-priority rule base rule in a multi-level rule tree architecture;

[0024] Keyword matching basic weight = business key coefficient × TF-IDF normalized value + user behavior correction formula (1);

[0025] In the formula, TF-IDF is term frequency-inverse document frequency.

[0026] Optionally, the matching module is further configured to calculate the weight of the regular rule by formula (2) when performing rule matching on the high-priority rule base rule in a multi-level rule tree architecture using regular matching;

[0027] The weight of the regular rule = basic weight × complexity coefficient - fuzzy penalty formula (2);

[0028] Where, the basic weight is the weight score divided according to the importance of the rule. Regular rules include key rules, business rules and auxiliary rules.

[0029] The complexity of calculating the number of character sets is calculated using formula (3);

[0030] C char =0.2×N char Formula (3);

[0031] Where C char is the complexity of the number of character sets, N char is the number of character sets;

[0032] The complexity of the number of capture groups is calculated using formula (4);

[0033] C capture =0.3×N capture Formula (4);

[0034] Where C capture is the complexity value of the number of capture groups, N copture is the number of capturing groups;

[0035] The complexity of calculating the number of fixed-length digital blocks is calculated using formula (5);

[0036] C digit =0.1×N digit Formula (5);

[0037] Where C digitis the complexity of the number of fixed-length digital blocks, N digit is the number of fixed-length digital blocks;

[0038] The total complexity is calculated by formula (6);

[0039] Complexity total =C char +C capture +C digit Formula (6);

[0040] The complexity coefficient is calculated by formula (7);

[0041] FinalCoeff=min(1.0+0.1×Complexity total ,1.5) Formula (7);

[0042] Where FinalCoeff is the complexity coefficient, Complexity total is the total complexity;

[0043] The fuzzy penalty is calculated by formula (8);

[0044]

[0045] Where Penalty is the fuzzy penalty, α is the maximum penalty coefficient, N matched is the expected number of capture groups, N expected The actual number of effective capturing groups.

[0046] Optionally, the matching module is further configured to, when semantic template matching is used to perform rule matching on high-priority rule base rules in a multi-level rule tree architecture:

[0047] Determine candidate templates based on accurate text information;

[0048] Extracting the slots defined by each candidate template from the text information;

[0049] Verify whether the extracted slot meets the preset rules. If yes, enter the context verification phase. If not, discard the current candidate template.

[0050] Check whether the conversation context meets the template requirements. If so, generate a structured intent recognition result. If not, mark the current result as pending and enter the follow-up question processing process.

[0051] The confidence of the structured intent recognition result is calculated by formula (9);

[0052] Confidence=SlotFillingRate×0.6+ContextMatch×0.4 Formula (9);

[0053] Where Confidence is the confidence, SlotFillingRate is the filling completeness of the required slots in the template, and ContextMatch is the context matching degree;

[0054] When the confidence level is greater than a specified threshold, the structured intent recognition result is used as the result of the rule matching.

[0055] An electronic device comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein the steps of the system are implemented when the processor executes the computer program.

[0056] A non-transitory computer-readable storage medium stores a computer program, which implements the steps of the system when executed by a processor.

[0057] The present invention has the following advantages:

[0058] The natural language intent recognition system in this invention, based on the dynamic collaboration of rules and a large model, uses a complexity judgment module in the dynamic routing decision layer to select appropriate processing methods for text messages of varying complexity. For simple text, the rule engine module uses a rule-matching method to quickly and efficiently identify intent. For complex text, the large model processing module leverages the model's powerful semantic understanding and processing capabilities to accurately identify intent, thus balancing processing efficiency and accuracy.

[0059] The natural language intent recognition system based on the dynamic collaboration of rules and large models in the present invention adopts a variety of rule matching methods, including keyword matching, regular matching and semantic template matching, and is combined with a large model processing module. A variety of rule matching methods can analyze and match text more comprehensively at the rule level, while the large model can handle more complex semantics and contextual information, making up for the shortcomings of simple rule matching in complex situations. The combination of the two helps to improve the accuracy of intent recognition. It can process simple natural language texts to meet the common daily intent recognition needs, and can also deal with complex, ambiguous or ambiguous text information. It is suitable for a variety of natural language processing application scenarios, such as intelligent customer service, voice assistants, etc., which improves the versatility and practicality of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] For purposes of illustration and not limitation, the present invention will now be described with reference to embodiments thereof and the accompanying drawings, in which:

[0061] Figure 1 A schematic diagram of the main components of a natural language intent recognition system based on dynamic collaboration between rules and a large model in an embodiment of the present invention;

[0062] Figure 2 Schematic diagram of parameter settings for keyword matching in an embodiment of the present invention;

[0063] Figure 3 Schematic diagram of parameter settings for regular expression matching in an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0065] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work should fall within the scope of protection of the present invention.

[0066] It should be noted that the terms "first," "second," and the like in the description of the present invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate for the embodiments of the present invention described herein. In addition, the terms "including," "having," and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatuses.

[0067] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features thereof can be combined with each other. The embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0068] Figure 1 This is a schematic diagram of the main components of the natural language intention recognition system based on dynamic collaboration between rules and large models in an embodiment of the present invention. Figure 2 As shown, the intent recognition model 1 provided by the embodiment of the present invention includes an input preprocessing layer 10 , a dynamic routing decision layer 20 , an intent recognition layer 30 and an output layer 40 .

[0069] The input pre-processing layer 10 is used to convert the user conversation voice stream into text information;

[0070] The input pre-processing layer 10 is further used for:

[0071] Convert user conversation voice stream into text information;

[0072] Performing cleaning and normalization processing on the text information;

[0073] Record historical intent and conversation content based on context to improve the consistency of multi-round conversations.

[0074] The dynamic routing decision layer 20 includes a complexity judgment module and a matching module, wherein the complexity judgment module is used to judge whether the text information is complex, and the matching module is used to match the high-priority rule base rules under a multi-level rule tree architecture, and the rule matching methods include keyword matching, regular matching and semantic template matching;

[0075] The input preprocessing layer 10 is further used to analyze the text information, obtain text length information, grammatical structure information and semantic understanding information of the text information, and determine whether the text information is complex based on the text length information, grammatical structure information and semantic understanding information.

[0076] Set a threshold for text length, such as 50 words. When the user enters more than 50 words, the text is considered complex; when the number of words is less than 50, it is considered simple. For example, if a user enters "Please help me find information on flights from Beijing to Shanghai tomorrow with an economy class price under 1,000 yuan and a departure time in the morning," and the number of words exceeds 50, it is considered complex text. However, if the user only enters "Check the weather in Beijing," the number of words is far less than 50, and it is classified as simple text.

[0077] The complexity of a text can be assessed by analyzing the grammatical structure of the sentence and identifying the number of clauses and modifiers it contains. If a sentence contains multiple clauses, long modifying phrases, or complex parallel structures, then the text is considered complex. For example, "I need to book a hotel in the city center, with a subway station and shopping mall nearby, and with 24-hour hot water and free breakfast, priced between 300 and 500 yuan." The sentence uses multiple modifying elements to describe the hotel, and the sentence structure is complex, so it can be judged as a complex text. In contrast, simple subject-verb-object structure sentences such as "book a hotel" have few modifying elements and are simple texts.

[0078] The difficulty of semantic understanding of the text is evaluated with the help of keywords, domain knowledge, and semantic models. If the text involves knowledge from multiple fields, or uses a large number of professional terms, metaphors, puns, etc., which increases the difficulty of semantic understanding, it will be judged as complex text. For example, when a user asks "Please explain the application of Bell's inequality in the field of information encryption based on the theory of quantum entanglement", the text contains professional terms such as "quantum entanglement" and "Bell's inequality", which are difficult to understand semantically and should be judged as complex text; while easy-to-understand texts such as "Recommend a comedy movie to me" have clear semantics and low difficulty in understanding, and can be classified as simple text.

[0079] The dynamic routing decision layer 20 is further configured to determine whether the text information has a reversal if the text information is not complex;

[0080] The rule engine module is further configured to perform intent recognition on the text information based on the rule matching method selected by the matching module when the inversion does not exist;

[0081] The large model processing module is further configured to perform intent recognition on the text information based on the large model when there is a reversal.

[0082] The dynamic routing decision layer 20 is also used to calculate the intention recognition delay;

[0083] The intention recognition layer 30 is also used to use a degradation strategy to perform intention recognition when the intention recognition delay is greater than a preset threshold.

[0084] The matching module is also used to match the high priority rule base rules in a multi-level rule tree architecture using keyword matching.

[0085] Keyword matching uses AC automaton (multi-mode string matching algorithm) to detect keywords. Keywords are divided into core words and extended words, and the two have different weights.

[0086] Core words are key terms that can directly determine intent classification. Generally, the probability of such words appearing in the corresponding intent is extremely high, and the probability of the word appearing in other intents is very low.

[0087] Extensions are words that help enhance intent confidence. They are typically verbs or quantifiers that frequently appear with the core word. Synonyms are typically grouped together within extensions for management.

[0088] Calculate the basic weight of keyword matching through formula (1);

[0089] Keyword matching basic weight = business key coefficient × TF-IDF normalized value + user behavior correction formula (1);

[0090] In the formula, TF-IDF is term frequency-inverse document frequency.

[0091] In addition, the weight value can be adjusted dynamically during operation:

[0092] When both high success rate and high frequency of use are met, the word is considered a high-quality keyword, and its weight is gradually increased according to the set weight growth factor.

[0093] When the success rate is low or the usage frequency is low, the weight is reduced according to the set weight attenuation factor.

[0094] According to this rule, the weight of the phrase can be defined.

[0095] like Figure 2 As shown, three parameters for specific calculations are displayed. Each parameter includes the calculation method and example value. The following is a detailed explanation:

[0096] Business Critical Factors:

[0097] Calculation Method: Experts perform the analysis and determine the key words. Words are divided into core words and extended words. Core words are assigned a coefficient of 1.0, and extended words are assigned a coefficient of 0.6. This method relies on professional judgment to measure the business-level criticality of a word.

[0098] Example values: The core word coefficient is 1.0 and the expansion word coefficient is 0.6, which clearly shows the coefficient values corresponding to different types of vocabulary.

[0099] TF-IDF normalized value:

[0100] Calculation method: A normalized approach is used. First, the TF-IDF (Term Frequency - Inverse Document Frequency) value of the word is obtained. The minimum value in the entire vocabulary is then subtracted. The resulting difference is then divided by the difference between the maximum and minimum values in the entire vocabulary. This method maps the word's TF-IDF value to a standard range, facilitating comparison and analysis across different vocabularies.

[0101] Example value: The value range is 0.3-0.9, indicating that after the above calculation, the result of this parameter is generally within this range, reflecting the characteristics of the standardized numerical range.

[0102] User behavior correction:

[0103] Calculation method: Based on statistics of user triggering behavior. Take the logarithm of the number of successful user triggers and multiply by 0.1. This calculation method allows for adjustments to relevant parameters based on the actual frequency of user operations, reflecting the impact of user behavior on the results.

[0104] Example value: Taking 10 successful triggers as an example, according to the calculation method log(10)×0.1, the result is +0.2, which intuitively demonstrates the specific calculation process and result presentation format.

[0105] The matching module is also used for matching high-priority rule base rules in a multi-level rule tree architecture using regular matching:

[0106] Regular expression matching: uses a pre-compiled regular expression library to match structured data, and assumes the responsibility of structured pattern recognition in the rule engine. Its main application scenarios include: fixed-format data extraction, context-sensitive recognition, and multi-condition joint judgment.

[0107] The weight of the regularization rule is calculated by formula (2);

[0108] The weight of the regular rule = basic weight × complexity coefficient - fuzzy penalty formula (2);

[0109] Where, the basic weight is the weight score divided according to the importance of the rule. Regular rules include key rules, business rules and auxiliary rules.

[0110] The complexity of calculating the number of character sets is calculated using formula (3);

[0111] C char =0.2×N char Formula (3);

[0112] Where C char is the complexity of the number of character sets, N char is the number of character sets;

[0113] The complexity of the number of capture groups is calculated using formula (4);

[0114] C capture =0.3×N capture Formula (4);

[0115] Where C capture is the complexity value of the number of capture groups, N capture is the number of capturing groups;

[0116] The complexity of calculating the number of fixed-length digital blocks is calculated using formula (5);

[0117] C digit =0.1×N digit Formula (5);

[0118] Where C digit is the complexity of the number of fixed-length digital blocks, N digit is the number of fixed-length digital blocks;

[0119] The total complexity is calculated by formula (6);

[0120] Complexity total =C char +C capture +C digit Formula (6);

[0121] The complexity coefficient is calculated by formula (7);

[0122] FinalCoeff=min(1.0+0.1×Complexity total ,1.5) Formula (7);

[0123] Where FinalCoeff is the complexity coefficient, Complexity total is the total complexity;

[0124] The fuzzy penalty is calculated by formula (8);

[0125]

[0126] Where Penalty is the fuzzy penalty, α is the maximum penalty coefficient, N matched is the expected number of capture groups, N expected The actual number of effective capturing groups.

[0127] like Figure 3 As shown, three parameters related to data capture are shown, as follows:

[0128] α: Value range: A value between 0 and 1, which means that this parameter is a proportional coefficient used to fine-tune the intensity of the effect it represents.

[0129] The example value of 0.3 is a typical value for this parameter in real-world applications. Note: This parameter is primarily used to control the severity of the penalty. In data capture scenarios, when certain expected capture groups are missing, α determines the significance of the negative impact of this omission. The closer the α value is to 1, the more pronounced the negative effect of missing capture groups; the closer the value is to 0, the less pronounced the negative effect. For example, in text information extraction, if α is set too high, the failure to successfully extract key information groups will have a significant adverse impact on the overall results.

[0130] Nexpected: Value range: integer greater than or equal to 1. The minimum value of this parameter is limited to 1, indicating that at least one capture group is required.

[0131] Example value: 3. For example, when extracting dates, if a regular expression is used to capture the year, month, and day information in a date, the value 3 represents the three groups that must be captured. This value is determined based on the specific regular expression design requirements. Note: Nexpected refers to the number of groups that must be captured, as determined by the regular expression design. When processing text data, set this value based on specific requirements and regular expression rules to clearly define the number of key information groups to be extracted.

[0132] Nmatched: The value range is an integer between 0 and Nexpected. Its upper limit is determined by Nexpected, and the lower limit is 0, indicating that no valid group may be successfully captured.

[0133] The example value is 2. This means that the number of groups successfully captured with non-empty content is 2, indicating that two of the expected capture groups met the valid capture criteria. Note: Nmatched represents the number of groups successfully captured with non-empty content. When using regular expressions for data capture, empty strings or invalid matches are filtered out. Nmatched reflects the final number of valid capture results and can be used to evaluate the actual effectiveness of the capture operation.

[0134] The matching module is further used for matching high-priority rule base rules in a multi-level rule tree architecture using semantic template matching:

[0135] Semantic templates are a method of describing user intent through structured patterns, including the following structures:

[0136] Intent definition: Clarify the business actions corresponding to the template;

[0137] Slot declaration: defines the key parameters and their constraints that need to be extracted from the text;

[0138] Sentence template: defines the text expression mode of the intent and supports structured placeholders;

[0139] Context dependencies: declare the prerequisites required to execute the intention;

[0140] Confidence configuration: Set the basic credibility benchmark value for intent recognition;

[0141] The semantic template matching parsing process is as follows:

[0142] Input text: The process starts with receiving text content entered by the user, and the system obtains the original text information to be processed.

[0143] Template screening: Based on a predefined semantic template library, the system uses mechanisms such as keyword indexing and slot type matching to quickly select a set of candidate templates that may be relevant to the input text. This stage only performs preliminary matching and does not involve detailed parsing.

[0144] Candidate template check: If at least one candidate template is found, the process enters the slot filling phase. If no matching template is found, the process terminates and returns an empty result.

[0145] Slot filling: For each candidate template, use regular expressions or deep learning models to extract the template-defined parameters (i.e., "slots") from the text. For example, extract the order ID and address fields from "Modify the address of order 20230815."

[0146] Slot verification: Verify whether the extracted slot meets the preset rules:

[0147] Data type checking (such as date format, number range);

[0148] Business logic verification (such as whether the order exists, whether the city is supported);

[0149] Completeness verification of required fields;

[0150] Verification result determination: If the verification passes, the context verification phase is entered. If the verification fails, the current template candidate is discarded. If all candidate templates fail, an empty result is returned.

[0151] Context validation: Checks whether the conversation context meets the template requirements:

[0152] Pre-order intent matching (e.g., you need to query the order before modifying it)

[0153] Required parameters have been obtained (such as user authentication information)

[0154] Business process status compatibility (e.g. payment process has not timed out)

[0155] Contextual condition determination: If the conditions are met, a structured intent recognition result is generated. If the conditions are not met, the current result is marked as pending and enters the follow-up processing flow.

[0156] Structured result generation: Output intent and slot information according to the defined JSON format

[0157] The confidence of the structured intent recognition result is calculated by formula (9);

[0158] Confidence=SlotFillingRate×0.6+ContextMatch×0.4 Formula (9);

[0159] Where Confidence is the confidence, SlotFillingRate is the filling completeness of the required slots in the template, and ContextMatch is the context matching degree;

[0160]

[0161] When the confidence level is greater than a specified threshold, the structured intent recognition result is adopted.

[0162] To be supplemented: When context verification fails, the missing context elements are recorded and the follow-up mechanism is triggered.

[0163] The intent recognition layer includes a rule engine module and a large model processing module, wherein the rule engine module is used to perform intent recognition on the text information based on the rule matching method selected by the matching module when the text information is not complex, so as to obtain an intent recognition result; the large model processing module is used to perform intent recognition on the text information when the text information is complex, so as to obtain an intent recognition result;

[0164] The output layer 40 is used to output the intention recognition result.

[0165] Figure 4 A schematic diagram of the physical structure of an electronic device provided in an embodiment of the present invention, such as Figure 4 As shown, the electronic device 50 includes: a processor 501 (processor), a memory 502 (memory) and a bus 503;

[0166] The processor 501 and the memory 502 communicate with each other via the bus 503.

[0167] The processor 501 is used to call the program instructions in the memory 502 to execute the methods provided by the above-mentioned method embodiments, so as to execute the methods provided by the implementation methods of the present invention.

[0168] This embodiment provides a non-transitory computer-readable storage medium, which stores computer instructions. The computer instructions enable a computer to execute the method provided by the embodiment of the present invention.

[0169] Those skilled in the art will understand that all or part of the steps of implementing the above-mentioned method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above-mentioned method embodiment; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk, etc. Various storage media that can store program codes.

[0170] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A natural language intention recognition system based on dynamic collaboration between rules and large models, characterized by: It includes input preprocessing layer, dynamic routing decision layer, intent recognition layer and output layer, where: The input preprocessing layer is used to convert the user conversation voice stream into text information; The dynamic routing decision layer includes a complexity judgment module and a matching module, wherein the complexity judgment module is used to judge whether the text information is complex, and the matching module is used to match the rules of the high-priority rule base under a multi-level rule tree architecture. The rule matching methods include keyword matching, regular matching and semantic template matching; The intent recognition layer includes a rule engine module and a large model processing module, wherein the rule engine module is used to perform intent recognition on the text information based on the rule matching method selected by the matching module when the text information is not complex, so as to obtain an intent recognition result; the large model processing module is used to perform intent recognition on the text information when the text information is complex, so as to obtain an intent recognition result; The output layer is used to output the intent recognition result.

2. The system according to claim 1, wherein: The input preprocessing layer is also used to: Convert user conversation voice stream into text information; Performing cleaning and normalization processing on the text information; Record historical intent and conversation content based on context to improve the consistency of multi-round conversations.

3. The system according to claim 1, wherein: The input preprocessing layer is also used to analyze the text information, obtain text length information, grammatical structure information and semantic understanding information of the text information, and determine whether the text information is complex based on the text length information, grammatical structure information and semantic understanding information.

4. The system according to claim 1, wherein: The dynamic routing decision layer is further configured to determine whether the text information has a reversal if the text information is not complex; The rule engine module is further configured to perform intent recognition on the text information based on the rule matching method selected by the matching module when the inversion does not exist; The large model processing module is further configured to perform intent recognition on the text information based on the large model when there is a reversal.

5. The system according to claim 1, wherein: The dynamic routing decision layer is also used to calculate the intention recognition delay; The intent recognition layer is also used to use a degradation strategy to perform intent recognition when the intent recognition delay is greater than a preset threshold.

6. The system according to claim 1, wherein: The matching module is also used to match the high priority rule base rules in a multi-level rule tree architecture using keyword matching. Calculate the basic weight of keyword matching through formula (1); Keyword matching basic weight = business key coefficient × TF-IDF normalized value + user behavior correction formula (1); In the formula, TF-IDF is term frequency-inverse document frequency.

7. The system according to claim 1, wherein: The matching module is also used for matching high-priority rule base rules in a multi-level rule tree architecture using regular matching: The weight of the regularization rule is calculated by formula (2); The weight of the regular rule = basic weight × complexity coefficient - fuzzy penalty formula (2); Where, the basic weight is the weight score divided according to the importance of the rule. Regular rules include key rules, business rules and auxiliary rules. The complexity of calculating the number of character sets is calculated using formula (3); C char =0.2×N char Formula (3); Where C char is the complexity of the number of character sets, N char is the number of character sets; The complexity of the number of capture groups is calculated using formula (4); C apture =0.3×N capture Formula (4); Where C capture is the complexity value of the number of capture groups, N capture is the number of capturing groups; The complexity of calculating the number of fixed-length digital blocks is calculated using formula (5); C digit =0.1×N digit Formula (5); Where C digit is the complexity of the number of fixed-length digital blocks, N digit is the number of fixed-length digital blocks; The total complexity is calculated by formula (6); Complexity total =C char +C capture +C digit Formula (6); The complexity coefficient is calculated by formula (7); FinalCoeff=min(1.0+0.1×Compleaxity total ,1.5) Formula (7); Where FinalCoeff is the complexity coefficient, Complexity total is the total complexity; The fuzzy penalty is calculated by formula (8); Where Penalty is the fuzzy penalty, α is the maximum penalty coefficient, N matched is the expected number of capture groups, N expected The actual number of effective capturing groups.

8. The system according to claim 1, wherein: The matching module is further used for matching high-priority rule base rules in a multi-level rule tree architecture using semantic template matching: Determine candidate templates based on accurate text information; Extracting the slots defined by each candidate template from the text information; Verify whether the extracted slot meets the preset rules. If yes, enter the context verification phase. If not, discard the current candidate template. Check whether the conversation context meets the template requirements. If so, generate a structured intent recognition result. If not, mark the current result as pending and enter the follow-up question processing process. The confidence of the structured intent recognition result is calculated by formula (9); Confidence=SlotFillingRate×0.6+ContextMatch×0.4 Formula (9); Where Confidence is the confidence, SlotFillingRate is the filling completeness of the required slots in the template, and ContextMatch is the context matching degree; When the confidence level is greater than a specified threshold, the structured intent recognition result is used as the result of the rule matching.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: The computer program comprises the system of any one of claims 1 to 8.

10. A non-transitory computer readable medium having a computer program stored thereon, characterized in that: The computer program comprises the system of any one of claims 1 to 8.

Images