Dynamic intent chain modeling and causal cleaning method and system in multi-round dialogue scenario
By preprocessing multi-turn dialogue data and using one-hot encoding for intent recognition, combined with Bayesian network causal modeling, the dynamic correlation problem of intent recognition in multi-turn dialogue scenarios is solved, improving the rationality of responses and user satisfaction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHENGDU YUNDING INTELLIGENT CONTROL TECH CO LTD
- Filing Date
- 2025-07-31
- Publication Date
- 2026-07-21
Smart Images
Figure CN121146044B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing technology, specifically to a method and system for dynamic intent chain modeling and causal cleansing in multi-turn dialogue scenarios. Background Technology
[0002] Natural Language Processing (NLP) is a core branch of artificial intelligence, aiming to enable computers to understand, generate, and interact with human language. With breakthroughs in deep learning technology, NLP has made significant progress in tasks such as machine translation, sentiment analysis, question answering systems, and dialogue generation. Especially in multi-turn dialogue scenarios, accurately identifying user intent, understanding contextual semantics, and generating coherent responses has become crucial for improving the human-computer interaction experience.
[0003] However, traditional intent recognition methods are usually based on text features from a single round of input. Intent recognition relies on static models, lacks dynamic correlation capabilities, and is limited by the lack of causal reasoning mechanisms, resulting in low response rationality.
[0004] To address the aforementioned issues, it is necessary to propose a method and system for dynamic intent chain modeling and causal cleansing in multi-turn dialogue scenarios. Summary of the Invention
[0005] The purpose of this invention is to solve the problems existing in the background art, and to propose a method and system for dynamic intent chain modeling and causal cleansing in multi-turn dialogue scenarios.
[0006] The objective of this invention can be achieved through the following technical solutions: A method and system for dynamic intent chain modeling and causal cleansing in multi-turn dialogue scenarios, including a method and system for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios.
[0007] In a first aspect, the present invention provides a method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios, including the following steps: Step 1: Multi-turn dialogue data collection and preprocessing; User voice input data is collected through a speech recognition device and input interface. A mapping from voice input data to text data is established using the CTTC loss function, converting the voice input data into text data. The converted text data undergoes preliminary preprocessing, including noise removal, annotation, word segmentation, and speech signal conversion. Semantic recognition is then performed on the preliminary preprocessing results, assigning a one-hot code representing the meaning of each word or phrase.
[0008] The original speech signal x(t) contained in the speech input data is collected, and speech recognition is performed by mapping from speech input data to text data, which is then converted into text data S=ASR(x(t)).
[0009] As a preferred embodiment of the present invention, the text data is preprocessed by using regular expressions and preset duplicate character removal rules to remove irrelevant characters and duplicate symbols from the text data.
[0010] As a preferred embodiment of the present invention, the pre-processed text data is segmented into several sub-words or phrases, and semantic recognition is performed on all sub-words or phrases based on natural language processing. Based on the semantic recognition results, a one-hot code representing the meaning of each sub-word or phrase is assigned.
[0011] Step 2: Intent Recognition and Intent Sequence Modeling; Semantic recognition is performed using one-hot encoding to identify user intents within the dialogue and obtain semantic information. A multi-turn dialogue intent sequence model is then constructed through intent sequence modeling to concatenate the semantic information. Based on this intent sequence model, a fast mapping from voice input data to user intents is established, bypassing text data conversion.
[0012] The user's input intent is classified using a pre-trained classification model to obtain the intent label corresponding to each one-hot encoding. Each intent label corresponds to a pre-defined convolutional neural network. Each one-hot encoding is then input into its corresponding convolutional neural network according to its intent label for semantic understanding, yielding semantic information.
[0013] By concatenating semantic information using an intent sequence model based on an LSTM temporal neural network, multiple intents in the dialogue are linked together in chronological order to form a dynamic intent sequence. Each intent sequence not only represents the current user's need but is also related to the user's previous intent.
[0014] Based on the intent sequence model, a fast mapping is established to directly obtain the user's intent from the speech input data, skipping the text data conversion. The specific process is as follows: Acquire speech input data and its corresponding user intent, and extract audio features from the speech input data, including Mel-frequency cepstral coefficients, pitch, volume, frequency, and power spectral density. Establish a direct mapping from the audio features of the speech input data to the intent sequence in the intent sequence model. Whenever new speech input data is recognized, calculate its similarity to the audio features of historical speech input data. Obtain the matching results of each historical speech input data and intent sequence, and generate the matching probability of matching each intent sequence based on the similarity of the audio features of the new speech input data and historical speech input data. Extract the intent sequence with the highest matching probability with the new speech input data. If the matching probability is greater than a preset threshold, the match is considered successful, and the new speech input data is directly matched with the intent sequence with the highest matching probability.
[0015] Step 3: Causal relationship modeling and intention reasoning; Based on the user's intent sequence, a causal relationship graph is constructed, and the user's next intent is inferred and calculated to determine the next interactive operation to be performed.
[0016] Based on the collected intent sequences, causal modeling is performed using Bayesian networks to deduce the causal relationships between the various intents.
[0017] According to the causal inference formula of Bayesian networks:
[0018] Perform causal modeling and calculate the impact of the previous intent. The probability of inferring the next intention X. , among them For the parent node variable, among which Let be the conditional probability of each intention i, and n be the number of intentions contained in the intention sequence.
[0019] Step 4: Signal matching and operation execution; Based on the causal reasoning results, corresponding execution signals are matched, and the information contained in the execution signals drives specific interactive operations of the interactive device, including system responses, user prompts, and interactive data output. After each operation, the dialogue state is further adjusted based on user feedback, and a user satisfaction score is obtained. The satisfaction score is then used to optimize the intent reasoning results.
[0020] Obtain the causal inference results and map them into a set of candidate signals. These signals contain the operations that the system needs to perform and the information that the system needs to return to the user.
[0021] The rules engine filters candidate signals, prioritizes them, and resolves conflicts. When multiple intents correspond to multiple candidate signals, the signals are prioritized based on priority, context, and the user's current needs. Signals with a priority greater than a preset threshold are executed first, while other candidate signals are delayed or canceled.
[0022] After each operation, feedback data is collected, including: the number of times the user clicked the "Satisfied" button through explicit feedback, the time interval between the user's next voice input data generation and the frequency of repeating the previous voice input data through implicit feedback, and a satisfaction score is obtained by weighted summation of the feedback data. The priority ranking rules of the rule engine are modified based on the satisfaction score. If the satisfaction score is lower than a preset threshold, the priority of candidate signals with a current priority greater than the preset threshold is reduced.
[0023] Secondly, the present invention provides a dynamic intent sequence modeling and causal cleansing system for multi-turn dialogue scenarios, including a request data acquisition module, a preprocessing module, an intent sequence modeling module, an intent reasoning module, and an execution and feedback optimization module.
[0024] The request data acquisition module collects user voice input data through a speech recognition device and input interface, and establishes a mapping from voice input data to text data using the CTTC loss function, converting the voice input data into text data. The converted text data undergoes preliminary preprocessing, including noise removal, annotation, text segmentation, and speech signal conversion. Semantic recognition is then performed on the preliminary preprocessing results, assigning a one-hot code representing the meaning of each word or phrase.
[0025] The original speech signal x(t) contained in the speech input data is collected, and speech recognition is performed through the mapping from speech input data to text data, which is then converted into text data S=ASR(x(t))).
[0026] The preprocessing module performs preliminary preprocessing on the text data, removing irrelevant characters and repeated symbols from the text data using regular expressions and preset duplicate character removal rules.
[0027] As a preferred embodiment of the present invention, the pre-processed text data is segmented into several sub-words or phrases, and semantic recognition is performed on all sub-words or phrases based on natural language processing. Based on the semantic recognition results, a one-hot code representing the meaning of each sub-word or phrase is assigned.
[0028] The intent sequence modeling module performs semantic recognition based on one-hot encoding, identifying user intents within the dialogue and obtaining semantic information. It then constructs a multi-turn dialogue intent sequence model through intent sequence modeling, concatenating the semantic information. Based on the intent sequence model, it establishes a fast mapping from voice input data, bypassing text data conversion, to directly obtain user intents.
[0029] The user's input intent is classified using a pre-trained classification model to obtain the intent label corresponding to each one-hot encoding. Each intent label corresponds to a pre-defined convolutional neural network. Each one-hot encoding is then input into its corresponding convolutional neural network according to its intent label for semantic understanding, yielding semantic information.
[0030] By concatenating semantic information using an intent sequence model based on an LSTM temporal neural network, multiple intents in the dialogue are linked together in chronological order to form a dynamic intent sequence. Each intent sequence not only represents the current user's need but is also related to the user's previous intent.
[0031] Based on the intent sequence model, a fast mapping is established to directly obtain the user's intent from the speech input data, skipping the text data conversion. The specific process is as follows: Acquire speech input data and its corresponding user intent, and extract audio features from the speech input data, including Mel-frequency cepstral coefficients, pitch, volume, frequency, and power spectral density. Establish a direct mapping from the audio features of the speech input data to the intent sequence in the intent sequence model. Whenever new speech input data is recognized, calculate its similarity to the audio features of historical speech input data. Obtain the matching results of each historical speech input data and intent sequence, and generate the matching probability of matching each intent sequence based on the similarity of the audio features of the new speech input data and historical speech input data. Extract the intent sequence with the highest matching probability with the new speech input data. If the matching probability is greater than a preset threshold, the match is considered successful, and the new speech input data is directly matched with the intent sequence with the highest matching probability.
[0032] The intent reasoning module constructs a causal relationship graph based on the user's intent sequence and performs inference calculations for the user's next intent to determine the next interactive operation to be performed.
[0033] Based on the collected intent sequences, causal modeling is performed using Bayesian networks to deduce the causal relationships between the various intents.
[0034] According to the causal inference formula of Bayesian networks:
[0035] Perform causal modeling and calculate the impact of the previous intent. The probability of inferring the next intention X. , among them For the parent node variable, among which Let be the conditional probability of each intention i, and n be the number of intentions contained in the intention sequence.
[0036] The execution and feedback optimization module matches corresponding execution signals based on the results of causal reasoning, and drives specific interactive operations of the interactive device based on the information contained in the execution signals, including system responses, user prompts, and interactive data output. After each execution operation, the dialogue state is further adjusted based on user feedback, and a user satisfaction score is obtained. The satisfaction score is then used to optimize the intent reasoning results.
[0037] Obtain the causal inference results and map them into a set of candidate signals. These signals contain the operations that the system needs to perform and the information that the system needs to return to the user.
[0038] The rules engine filters candidate signals, prioritizes them, and resolves conflicts. When multiple intents correspond to multiple candidate signals, the signals are prioritized based on priority, context, and the user's current needs. Signals with a priority greater than a preset threshold are executed first, while other candidate signals are delayed or canceled.
[0039] After each operation, feedback data is collected, including: the number of times the user clicked the "Satisfied" button through explicit feedback, the time interval between the user's next voice input data generation and the frequency of repeating the previous voice input data through implicit feedback, and a satisfaction score is obtained by weighted summation of the feedback data. The priority ranking rules of the rule engine are modified based on the satisfaction score. If the satisfaction score is lower than a preset threshold, the priority of candidate signals with a current priority greater than the preset threshold is reduced.
[0040] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention collects and preprocesses multi-turn dialogue data, assigns one-hot codes to each sub-word or phrase, and then performs intent recognition and sequence modeling based on these codes, which can accurately identify user intent; at the same time, it establishes a fast mapping from voice input data to directly obtain user intent by skipping text data conversion, which improves intent recognition efficiency. 2. This invention constructs a causal relationship graph based on the user intent sequence, uses Bayesian networks for causal modeling, derives the causal relationship between each intent, calculates the probability of inferring the next intent from the previous intent, and thus accurately infers the user's next intent and the interaction operation to be performed. 3. This invention matches and executes signal-driven interactive operations based on causal reasoning results, collects user feedback data to calculate satisfaction scores, and corrects the priority ranking rules of the rule engine based on the scores, continuously optimizing interactive operations and improving user satisfaction and experience. Attached Figure Description
[0041] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation
[0042] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] Please see Figure 1 As shown, the method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios includes the following steps: Step 1: Multi-turn dialogue data collection and preprocessing; User voice input data is collected through a speech recognition device and input interface. A mapping from voice input data to text data is established using the CTTC loss function, converting the voice input data into text data. The converted text data undergoes preliminary preprocessing, including noise removal, annotation, word segmentation, and speech signal conversion. Semantic recognition is then performed on the preliminary preprocessing results, assigning a one-hot code representing the meaning of each word or phrase.
[0044] The original speech signal x(t) contained in the speech input data is collected, and speech recognition is performed by mapping from speech input data to text data, which is then converted into text data S=ASR(x(t)).
[0045] Furthermore, the text data undergoes preliminary preprocessing, using regular expressions and preset duplicate character removal rules to remove irrelevant characters and repeated symbols from the text data.
[0046] Furthermore, the pre-processed text data is segmented into several sub-words or phrases, and semantic recognition is performed on all sub-words or phrases based on natural language processing. Based on the semantic recognition results, a one-hot code representing the meaning of each sub-word or phrase is assigned.
[0047] Step 2: Intent Recognition and Intent Sequence Modeling; Semantic recognition is performed using one-hot encoding to identify user intents within the dialogue and obtain semantic information. A multi-turn dialogue intent sequence model is then constructed through intent sequence modeling to concatenate the semantic information. Based on this intent sequence model, a fast mapping from voice input data to user intents is established, bypassing text data conversion.
[0048] The user's input intent is classified using a pre-trained classification model to obtain the intent label corresponding to each one-hot encoding. Each intent label corresponds to a pre-defined convolutional neural network. Each one-hot encoding is then input into its corresponding convolutional neural network according to its intent label for semantic understanding, yielding semantic information.
[0049] By concatenating semantic information using an intent sequence model based on an LSTM temporal neural network, multiple intents in the dialogue are linked together in chronological order to form a dynamic intent sequence. Each intent sequence not only represents the current user's need but is also related to the user's previous intent.
[0050] Based on the intent sequence model, a fast mapping is established to directly obtain the user's intent from the speech input data, skipping the text data conversion. The specific process is as follows: Acquire speech input data and its corresponding user intent, and extract audio features from the speech input data, including Mel-frequency cepstral coefficients, pitch, volume, frequency, and power spectral density. Establish a direct mapping from the audio features of the speech input data to the intent sequence in the intent sequence model. Whenever new speech input data is recognized, calculate its similarity to the audio features of historical speech input data. Obtain the matching results of each historical speech input data and intent sequence, and generate the matching probability of matching each intent sequence based on the similarity of the audio features of the new speech input data and historical speech input data. Extract the intent sequence with the highest matching probability with the new speech input data. If the matching probability is greater than a preset threshold, the match is considered successful, and the new speech input data is directly matched with the intent sequence with the highest matching probability.
[0051] An intent sequence is specifically a sequence of actions performed by a user. For example: close an entertainment streaming app → check today's weather → check tomorrow morning's weather → set an alarm → change the alarm time.
[0052] Step 3: Causal relationship modeling and intention reasoning; Based on the user's intent sequence, a causal relationship graph is constructed, and the user's next intent is inferred and calculated to determine the next interactive operation to be performed.
[0053] Based on the collected intent sequences, causal modeling is performed using Bayesian networks to deduce the causal relationships between the various intents.
[0054] For example, an intention might be generated by a previous action or piece of information, and the system needs to deduce this causal relationship in order to execute the next action. Specifically, if a user expresses the intention to turn on GPS location, the next action might be to open a navigation app or check the weather.
[0055] According to the causal inference formula of Bayesian networks:
[0056] Perform causal modeling and calculate the impact of the previous intent. The probability of inferring the next intention X. , among them For the parent node variable, among which Let be the conditional probability of each intention i, and n be the number of intentions contained in the intention sequence.
[0057] Step 4: Signal matching and operation execution; Based on the causal reasoning results, corresponding execution signals are matched, and the information contained in the execution signals drives specific interactive operations of the interactive device, including system responses, user prompts, and interactive data output. After each operation, the dialogue state is further adjusted based on user feedback, and a user satisfaction score is obtained. The satisfaction score is then used to optimize the intent reasoning results.
[0058] Obtain the causal inference results and map them into a set of candidate signals. These signals contain the operations that the system needs to perform and the information that the system needs to return to the user.
[0059] The rules engine filters candidate signals, prioritizes them, and resolves conflicts. When multiple intents correspond to multiple candidate signals, the signals are prioritized based on priority, context, and the user's current needs. Signals with a priority greater than a preset threshold are executed first, while other candidate signals are delayed or canceled.
[0060] After each operation, feedback data is collected, including: the number of times the user clicked the "Satisfied" button through explicit feedback, the time interval between the user's next voice input data generation and the frequency of repeating the previous voice input data through implicit feedback, and a satisfaction score is obtained by weighted summation of the feedback data. The priority ranking rules of the rule engine are modified based on the satisfaction score. If the satisfaction score is lower than a preset threshold, the priority of candidate signals with a current priority greater than the preset threshold is reduced.
[0061] Please see Figure 2As shown, the dynamic intent sequence modeling and causal cleansing system for multi-turn dialogue scenarios includes a request data acquisition module, a preprocessing module, an intent sequence modeling module, an intent reasoning module, and an execution and feedback optimization module.
[0062] The request data acquisition module collects user voice input data through a speech recognition device and input interface, and establishes a mapping from voice input data to text data using the CTTC loss function, converting the voice input data into text data. The converted text data undergoes preliminary preprocessing, including noise removal, annotation, text segmentation, and speech signal conversion. Semantic recognition is then performed on the preliminary preprocessing results, assigning a one-hot code representing the meaning of each word or phrase.
[0063] The original speech signal x(t) contained in the speech input data is collected, and speech recognition is performed through the mapping from speech input data to text data, which is then converted into text data S=ASR(x(t))).
[0064] The preprocessing module performs preliminary preprocessing on the text data, removing irrelevant characters and repeated symbols from the text data using regular expressions and preset duplicate character removal rules.
[0065] Furthermore, the pre-processed text data is segmented into several sub-words or phrases, and semantic recognition is performed on all sub-words or phrases based on natural language processing. Based on the semantic recognition results, a one-hot code representing the meaning of each sub-word or phrase is assigned.
[0066] The intent sequence modeling module performs semantic recognition based on one-hot encoding, identifying user intents within the dialogue and obtaining semantic information. It then constructs a multi-turn dialogue intent sequence model through intent sequence modeling, concatenating the semantic information. Based on the intent sequence model, it establishes a fast mapping from voice input data, bypassing text data conversion, to directly obtain user intents.
[0067] The user's input intent is classified using a pre-trained classification model to obtain the intent label corresponding to each one-hot encoding. Each intent label corresponds to a pre-defined convolutional neural network. Each one-hot encoding is then input into its corresponding convolutional neural network according to its intent label for semantic understanding, yielding semantic information.
[0068] By concatenating semantic information using an intent sequence model based on an LSTM temporal neural network, multiple intents in the dialogue are linked together in chronological order to form a dynamic intent sequence. Each intent sequence not only represents the current user's need but is also related to the user's previous intent.
[0069] Based on the intent sequence model, a fast mapping is established to directly obtain the user's intent from the speech input data, skipping the text data conversion. The specific process is as follows: Acquire speech input data and its corresponding user intent, and extract audio features from the speech input data, including Mel-frequency cepstral coefficients, pitch, volume, frequency, and power spectral density. Establish a direct mapping from the audio features of the speech input data to the intent sequence in the intent sequence model. Whenever new speech input data is recognized, calculate its similarity to the audio features of historical speech input data. Obtain the matching results of each historical speech input data and intent sequence, and generate the matching probability of matching each intent sequence based on the similarity of the audio features of the new speech input data and historical speech input data. Extract the intent sequence with the highest matching probability with the new speech input data. If the matching probability is greater than a preset threshold, the match is considered successful, and the new speech input data is directly matched with the intent sequence with the highest matching probability.
[0070] The intent reasoning module constructs a causal relationship graph based on the user's intent sequence and performs inference calculations for the user's next intent to determine the next interactive operation to be performed.
[0071] Based on the collected intent sequences, causal modeling is performed using Bayesian networks to deduce the causal relationships between the various intents.
[0072] According to the causal inference formula of Bayesian networks:
[0073] Perform causal modeling and calculate the impact of the previous intent. The probability of inferring the next intention X. , among them For the parent node variable, among which Let be the conditional probability of each intention i, and n be the number of intentions contained in the intention sequence.
[0074] The execution and feedback optimization module matches corresponding execution signals based on the results of causal reasoning, and drives specific interactive operations of the interactive device based on the information contained in the execution signals, including system responses, user prompts, and interactive data output. After each execution operation, the dialogue state is further adjusted based on user feedback, and a user satisfaction score is obtained. The satisfaction score is then used to optimize the intent reasoning results.
[0075] Obtain the causal inference results and map them into a set of candidate signals. These signals contain the operations that the system needs to perform and the information that the system needs to return to the user.
[0076] The rules engine filters candidate signals, prioritizes them, and resolves conflicts. When multiple intents correspond to multiple candidate signals, the signals are prioritized based on priority, context, and the user's current needs. Signals with a priority greater than a preset threshold are executed first, while other candidate signals are delayed or canceled.
[0077] After each operation, feedback data is collected, including: the number of times the user clicked the "Satisfied" button through explicit feedback, the time interval between the user's next voice input data generation and the frequency of repeating the previous voice input data through implicit feedback, and a satisfaction score is obtained by weighted summation of the feedback data. The priority ranking rules of the rule engine are modified based on the satisfaction score. If the satisfaction score is lower than a preset threshold, the priority of candidate signals with a current priority greater than the preset threshold is reduced.
[0078] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.
[0079] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims means any combination and all possible combinations of one or more of the associated listed items, and includes such combinations; The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios, characterized in that, Includes the following steps: Step 1: Multi-turn dialogue data collection and preprocessing; The system collects user voice input data through a voice recognition device and input interface, and establishes a mapping from voice input data to text data using the CTTC loss function, converting the voice input data into text data. The converted text data undergoes preliminary preprocessing, including noise removal, annotation, text segmentation, and voice signal conversion. The system then performs semantic recognition on the preliminary preprocessing results of the text data, assigning a one-hot code representing the meaning of each subword or phrase. Step 2: Intent Recognition and Intent Sequence Modeling; Semantic recognition is performed based on one-hot encoding to identify user intents contained in the dialogue and obtain semantic information; a multi-turn dialogue intent sequence model is constructed through intent sequence modeling to concatenate the semantic information; and a fast mapping is established based on the intent sequence model to directly obtain user intents from voice input data without text data conversion. The specific process for identifying user intent contained in a dialogue is as follows: The intent of user input is classified by a pre-trained classification model to obtain the intent label corresponding to each one-hot code; each intent label corresponds to a pre-set convolutional neural network. Each one-hot encoding is input into the corresponding convolutional neural network according to its own intent label, and semantic understanding is performed to obtain semantic information. Semantic information is concatenated by an intent sequence model based on LSTM temporal neural network, linking multiple intents in the dialogue in chronological order to form a dynamic intent sequence; each intent sequence not only represents the current user's need, but is also related to the user's previous intent. Step 3: Causal relationship modeling and intention reasoning; Based on the user intent sequence, a causal relationship graph is constructed, and the user's next intent is inferred and calculated to determine the next interactive operation to be performed; based on the collected intent sequence, causal modeling is performed through Bayesian network to deduce the causal relationship between each intent. Step 4: Signal matching and operation execution; Based on the results of causal reasoning, corresponding execution signals are matched, and the specific interactive operations of the interactive device are driven by the information contained in the execution signals, including system response, user prompts, and interactive data output. After each operation, the dialogue state is further adjusted based on user feedback, and user satisfaction scores are obtained. The intention reasoning results are then optimized based on the satisfaction scores.
2. The method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios according to claim 1, characterized in that, The specific process of converting voice input data into text data is as follows: The original speech signal x(t) contained in the speech input data is collected, and speech recognition is performed through the mapping from speech input data to text data, which is then converted into text data S=ASR(x(t)). The text data undergoes preliminary preprocessing, using regular expressions and preset duplicate character removal rules to remove irrelevant characters and repeated symbols from the text data; The text data that has undergone preliminary preprocessing is segmented into words, and the text S is segmented into several sub-words or phrases. Semantic recognition is performed on all sub-words or phrases based on natural language processing. Based on the semantic recognition results, a one-hot code representing the meaning of each sub-word or phrase is assigned.
3. The method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios according to claim 1, characterized in that, The specific process of establishing a speech input data skipping text data conversion based on the intent sequence model is as follows: Acquire speech input data and its corresponding user intent; extract audio features from the speech input data, including Mel-frequency cepstral coefficients, pitch, volume, frequency, and power spectral density; establish a direct mapping from the audio features of the speech input data to the intent sequence in the intent sequence model; whenever new speech input data is identified, calculate its similarity to the audio features of historical speech input data; obtain the matching results of each historical speech input data and the intent sequence, and generate the matching probability of matching each intent sequence based on the similarity of the audio features of the new speech input data and the historical speech input data; Extract the intent sequence with the highest matching probability to the new voice input data. If the matching probability is greater than a preset threshold, the match is considered successful, and the intent sequence with the highest matching probability to the new voice input data is directly matched.
4. The method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios according to claim 1, characterized in that, The specific process of causal modeling using Bayesian networks is as follows: According to the causal inference formula of Bayesian networks: Calculation through the previous intention The probability of inferring the next intention X. , among them For the parent node variable, among which Let be the conditional probability of each intention i, and n be the number of intentions contained in the intention sequence.
5. The method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios according to claim 1, characterized in that, The specific process of matching the corresponding execution signal based on the result of causal reasoning is as follows: Obtain the causal inference results and map them into a set of candidate signals. These signals contain the operations that the system needs to perform and the information that the system needs to return to the user. The rule engine filters candidate signals, prioritizes and resolves conflicts among them. When multiple intents correspond to multiple candidate signals, the candidate signals are sorted according to priority, context and the user's current needs. Signals with a priority greater than a preset threshold will be executed first, while other candidate signals will be delayed or canceled.
6. The method for dynamic intent sequence modeling and causal cleansing in multi-turn dialogue scenarios according to claim 1, characterized in that, The specific process of optimizing the intention reasoning results through satisfaction rating is as follows: After each operation, feedback data is collected, including: the number of times the user clicks the "Satisfied" button through explicit feedback, the time interval between the user's next voice input data generation and the frequency of repeating the previous voice input data through implicit feedback, and a satisfaction score is obtained by weighted summation of the feedback data; the priority ranking rules of the rule engine are modified based on the satisfaction score; if the satisfaction score is lower than a preset threshold, the priority of candidate signals with a current priority greater than the preset threshold is reduced.
7. Dynamic intent sequence modeling and causal cleansing system for multi-turn dialogue scenarios; It includes a request data acquisition module, a preprocessing module, an intent sequence modeling module, an intent reasoning module, and an execution and feedback optimization module, characterized in that... ； The request data acquisition module collects the user's voice input data through the speech recognition device and input interface, and establishes a mapping from voice input data to text data through the CTTC loss function, converting the voice input data into text data; performs preliminary preprocessing on the converted text data, including noise removal, annotation, text segmentation and speech signal conversion; performs semantic recognition on the preliminary preprocessing results of the text data, assigning a one-hot code representing the meaning of each subword or phrase; The preprocessing module performs preliminary preprocessing on the text data, removing irrelevant characters and repeated symbols from the text data using regular expressions and preset duplicate character removal rules; The intent sequence modeling module performs semantic recognition based on one-hot encoding, identifies user intents contained in the dialogue, and obtains semantic information; it also constructs an intent sequence model for multi-turn dialogues through intent sequence modeling, and concatenates the semantic information; based on the intent sequence model, it establishes a fast mapping from voice input data to directly obtain user intents by skipping text data conversion; The specific process for identifying user intent contained in a dialogue is as follows: The intent of user input is classified by a pre-trained classification model to obtain the intent label corresponding to each one-hot code; each intent label corresponds to a pre-set convolutional neural network. Each one-hot encoding is input into the corresponding convolutional neural network according to its own intent label, and semantic understanding is performed to obtain semantic information. Semantic information is concatenated by an intent sequence model based on LSTM temporal neural network, linking multiple intents in the dialogue in chronological order to form a dynamic intent sequence; each intent sequence not only represents the current user's need, but is also related to the user's previous intent. The intent reasoning module constructs a causal relationship graph based on the user's intent sequence, and performs the next intent reasoning calculation to determine the next interactive operation to be performed. The execution and feedback optimization module matches the corresponding execution signals based on the results of causal reasoning, and drives the specific interactive operations of the interactive device based on the information contained in the execution signals, including system response, user prompts, and interactive data output. After each execution operation, the module further adjusts the dialogue state based on user feedback, obtains the user's satisfaction score, and optimizes the intent reasoning results based on the satisfaction score.