A large-model-driven intelligent voice robot interaction method and system

Abstract

This invention relates to the field of natural language processing technology, specifically to a large-model-driven intelligent voice robot interaction method and system, comprising: acquiring several words from the current word sequence and historical word sequences; acquiring the contextual semantic matching degree of words in the current word sequence based on the distance between words in the current word sequence and historical word sequences and the semantic similarity between words; acquiring the structural fit degree of each word in the current word sequence based on the adjacent words of each word in the current word sequence and historical word sequences; and acquiring the reliability of words in the current word sequence by combining the user's current interaction round with the contextual semantic matching degree of words in the current word sequence, thereby acquiring the user's actual expressive intent. This invention improves the accuracy of recognizing the user's true expressive intent by analyzing the user's current and historical speech signals to acquire the reliability of words in the current speech signal.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a large-model-driven intelligent voice robot interaction method and system. Background Technology

[0002] With the continuous development of artificial intelligence and speech recognition technology, intelligent service systems based on voice interaction are gradually being applied to various scenarios such as intelligent customer service, smart home control, medical consultation, and information inquiry. Users can interact with the system through natural language using voice, and the system can recognize and interpret the user's speech to provide corresponding services or feedback. In existing technologies, speech signals are usually first converted into text information using speech recognition technology, and then semantic analysis of the text content is performed based on natural language processing technology to determine the user's expressed intent. However, in the actual interaction process of intelligent voice robots, users' speech expressions often have strong colloquial characteristics, and due to differences in dialect accents or non-standard Mandarin expressions among users from different regions, speech recognition systems are prone to homonym substitution, semantic deviation, or incomplete expression when converting speech signals into text information, making it difficult for the system to accurately understand the user's true expressed intent. Summary of the Invention

[0003] To address the problem that existing speech recognition technologies cannot accurately identify the user's true intentions in real-world interactive scenarios of intelligent voice robots, the present invention aims to provide a large-model-driven intelligent voice robot interaction method and system. The specific technical solution adopted is as follows: Acquire the user's current and historical voice signals, record the user's current interaction round, and convert the current and historical voice signals into current word sequences and historical word sequences respectively, and obtain several words from the current word sequences and historical word sequences; Based on the distance between words in the current word sequence and words in the historical word sequence, the temporal cumulative weight of words in the current word sequence is obtained. Combined with the semantic similarity between words in the current word sequence and words in the historical word sequence, the contextual semantic matching degree of words in the current word sequence is obtained. Based on the adjacent words of each word in the current word sequence and the adjacent words of each word in the historical word sequence, obtain the structural fit of each word in the current word sequence; based on the current user interaction round, obtain the contextual semantic matching weight and structural fit weight, and combine the contextual semantic matching and structural fit of the words in the current word sequence to obtain the reliability of the words in the current word sequence. Based on the reliability of the words in the current word sequence, the user's actual expressive intent is obtained.

[0004] Preferably, the step of acquiring the user's current and historical voice signals, recording the user's current interaction round, and converting the current and historical voice signals into a current word sequence and a historical word sequence, respectively, and acquiring several words from the current word sequence and the historical word sequence includes: When a user interacts with the intelligent voice robot, the microphone array integrated in the intelligent voice robot collects the user's voice signal and records the current interaction round with the intelligent voice robot, which is denoted as the current interaction round. The voice signal collected at the current moment and the signal collected in the preset historical range are respectively denoted as the current voice signal and the historical voice signal. The speech recognition algorithm converts the current speech signal and the historical speech signal into the current word sequence and the historical word sequence, respectively, and then uses the word segmentation algorithm to obtain a number of words from the current word sequence and the historical word sequence.

[0005] Preferably, obtaining the temporal cumulative weight of words in the current word sequence based on the distance between words in the current word sequence and words in the historical word sequence includes: For any word in the current word sequence, words in the historical word sequence that are the same as the words in the current word sequence are recorded as feature words; based on the temporal distance between the historical speech signal and the current speech signal corresponding to each feature word, the temporal cumulative weight of the words in the current word sequence is obtained; The temporal distance is negatively correlated with the temporal cumulative weight.

[0006] Preferably, obtaining the contextual semantic matching degree of words in the current word sequence includes: Any word in the current word sequence is designated as the target word, and any word in any historical word sequence is designated as the reference word. The word vectors of the target word and the reference word are obtained respectively using the word2vec algorithm. Based on the cosine similarity between the word vectors of the target word and the reference word, and combined with the temporal cumulative weight of the target word, the temporal decay semantic correlation between the target word and the reference word is obtained. The temporal decay semantic correlation between the target word and the control word is positively correlated with the cosine similarity between the word vectors of the target word and the control word; the temporal decay semantic correlation between the target word and the control word is positively correlated with the temporal cumulative weight of the target word. The contextual semantic matching degree of the target word is obtained based on the temporal decay semantic correlation between the target word and all words in all historical word sequences.

[0007] Preferably, obtaining the contextual semantic matching degree of the target word includes: If the temporal decay semantic correlation between the target word and the reference word is less than 0, the temporal decay semantic correlation between the target word and the reference word is set to 0. The mean of the temporal decay semantic correlation between the target word and all words in all historical word sequences is used as the contextual semantic matching degree of the target word.

[0008] Preferably, the step of obtaining the structural fit of each word in the current word sequence based on the adjacent words of each word in the current word sequence and the adjacent words of each word in the historical word sequence includes: For any word in the current word sequence, based on the neighboring words of the word in the current word sequence, the probability statistics of the neighboring words of all words in all historical sequences are performed using the N-gram model to obtain the conditional probability of the word in the current word sequence under the corresponding adjacent context conditions, which is used as the structural fit of the word in the current word sequence.

[0009] Preferably, obtaining the contextual semantic matching weight and structural fit weight based on the user's current interaction round includes: An initial context semantic matching degree weight and amplification coefficient are preset. The context semantic matching degree weight is obtained based on the user's current interaction round, the initial context semantic matching degree weight and amplification coefficient. The contextual semantic matching weights are all positively correlated with the user's current interaction round, the initial contextual semantic matching weight, and the growth coefficient, respectively. The difference between 1 and the contextual semantic matching weight is used as the structural fit weight.

[0010] Preferably, the reliability of obtaining the words in the current word sequence includes: For any word in the current word sequence, the reliability of the word in the current word sequence is obtained by weighting and summing the contextual semantic matching degree and the structural fit degree of the word in the current word sequence through the contextual semantic matching degree weight and the structural fit degree weight.

[0011] Preferably, obtaining the user's actual expressive intent based on the reliability of words in the current word sequence includes: The reliability of each word in the current word sequence is used as the label of each word in the current word sequence to obtain the current word sequence with labels. The current word sequence with labels is converted into natural language prompt word format and input into the large language model. The large language model outputs the intent recognition result of the current word sequence to obtain the user's actual expression intent.

[0012] A large-model-driven intelligent voice robot interaction system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of any one of the large-model-driven intelligent voice robot interaction methods described above.

[0013] This invention has the following beneficial effects: By acquiring the user's current speech signal and historical speech signal, and converting them into current word sequences and historical word sequences respectively, this invention models the user's multi-turn interaction context at the lexical level. Based on this, it obtains the temporal cumulative weight of words by calculating the distance between words in the current word sequence and the historical word sequence, and combines this with the semantic similarity between words to obtain the contextual semantic matching degree of words, thereby quantifying the semantic association between the current expression and the historical context. Furthermore, it obtains the structural fit degree by analyzing the adjacent words of each word in the current word sequence and combining the relationships between adjacent words in the historical word sequence, to represent the continuity and consistency of the user's language expression at the syntactic structure level. Simultaneously, it dynamically obtains the contextual semantic matching degree weight and the structural fit degree weight according to the user's current interaction round, thereby adaptively fusing the semantic and structural information of words to obtain the reliability of each word, thus improving the accuracy of recognizing the user's true expressive intent and enhancing the understanding ability and interactive experience of intelligent voice robots in human-computer interaction. Attached Figure Description

[0014] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating a large-model-driven intelligent voice robot interaction method provided in one embodiment of the present invention; Figure 2 This is a structural block diagram of a large-model-driven intelligent voice robot interaction system according to the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a large-model-driven intelligent voice robot interaction method and system proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] The specific solution of the large-model-driven intelligent voice robot interaction method provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0019] Please see Figure 1 The diagram illustrates a flowchart of a large-model-driven intelligent voice robot interaction method according to an embodiment of the present invention. The method includes: Step S101: Obtain the user's current and historical voice signals, record the user's current interaction round, and convert the current and historical voice signals into current word sequences and historical word sequences respectively, and obtain several words from the current word sequences and historical word sequences.

[0020] It should be noted that this embodiment, as a large-model-driven intelligent voice robot interaction method, specifically identifies the user's current true intention by analyzing the current voice signal and historical voice signals. To fully preserve the user's voice expression information during continuous dialogue, thereby providing the necessary contextual basis for subsequent analysis of the user's intention, this embodiment first acquires the user's current voice signal and historical voice signal, and records the user's current interaction round. Further, the voice signal is converted into a word sequence through voice recognition processing, and the word sequence is segmented to obtain several words. This transforms the original voice signal into computable language units, enabling the system to quantitatively analyze the temporal relationship, semantic association, and structural features between the current and historical sentences at the lexical level. This provides basic data support for subsequent calculation of the temporal cumulative weight, contextual semantic matching degree, and structural adaptability of words, thereby accurately identifying the user's true intention.

[0021] Step S102: Based on the distance between words in the current word sequence and words in the historical word sequence, obtain the temporal cumulative weight of words in the current word sequence, and combine the semantic similarity between words in the current word sequence and words in the historical word sequence to obtain the contextual semantic matching degree of words in the current word sequence.

[0022] It should be noted that in actual interaction with intelligent voice robots, users' speech often has certain colloquial characteristics. Furthermore, due to regional differences in accents or non-standard Mandarin pronunciation, speech recognition systems are prone to issues such as homophone substitution, semantic discrepancies, or incomplete expression when converting speech signals into text. Additionally, in continuous voice interaction scenarios, the user's current statement is often a contextual expression built upon previous statements. If semantic analysis is only performed on the current statement, it's easy to overlook the semantic information already expressed by the user in previous interactions, leading to inaccurate understanding of the user's true intentions. Therefore, this step... By analyzing the temporal distance between words in the current word sequence and those in the historical word sequence, the system obtains the temporal cumulative weight of words in the current word sequence, thereby quantifying the influence of historical semantic information on the understanding of the current sentence. Furthermore, by combining the semantic similarity between words in the current word sequence and those in the historical word sequence, the system obtains the contextual semantic matching degree of words in the current word sequence. This enables the system to analyze the semantic relationship between the current expression and the historical expression at the semantic level, thereby evaluating the semantic rationality of the current words based on the continuity of the historical context. This provides a basis for subsequent calculation of word reliability and identification of the user's true expressive intent.

[0023] Step S103: Based on the adjacent words of each word in the current word sequence and the adjacent words of each word in the historical word sequence, obtain the structural fit of each word in the current word sequence; based on the current user interaction round, obtain the contextual semantic matching weight and structural fit weight, and combine the contextual semantic matching and structural fit of the words in the current word sequence to obtain the reliability of the words in the current word sequence.

[0024] It should be noted that during the interaction with the intelligent voice robot, the speech recognition system, when converting user speech into text words, may encounter semantic deviations due to dialectal accents or colloquial expressions. Furthermore, speech recognition errors may lead to abnormal word order or misidentification of individual words, resulting in the recognized words having inconsistent sentence structures with normal language expression patterns. In other words, relying solely on semantic similarity between words is insufficient to comprehensively determine the appropriateness of the current word in a sentence. Therefore, this step analyzes the adjacent words of each word in the current word sequence and combines this with the adjacent words of each word in the historical word sequence to obtain the structural fit of the current word in the corresponding context structure. This allows for a structural assessment of the current word's position within the sentence from a linguistic structure perspective. The system assesses the reasonableness of sentences. Furthermore, since user interaction rounds differ during continuous voice interaction, the degree of dependence of the current sentence on historical context also varies. Therefore, different weights are assigned to contextual semantic matching and structural fit based on the user's current interaction round, enabling the system to dynamically adjust the influence of semantic and structural information on vocabulary evaluation. Based on this, the reliability of each word is obtained by weighted fusion of contextual semantic matching and structural fit in the current word sequence. This comprehensively reflects the credibility of the current word in terms of both semantic relevance and linguistic structural reasonableness, providing a basis for subsequent identification of the user's true expressive intent based on word reliability.

[0025] Step S104: Based on the reliability of the words in the current word sequence, obtain the user's actual expressive intent.

[0026] It should be noted that after obtaining the reliability of each word in the current word sequence, semantic combination and semantic parsing can be performed based on words with high reliability to determine the user's actual expression intention. In this way, even when the user's voice interaction expression is not standardized, the core semantic information of the user's expression can be accurately extracted, improving the accuracy of the voice robot in recognizing the user's intention, and thus improving the overall voice interaction effect.

[0027] Step S201: Obtain the user's current and historical voice signals, record the user's current interaction round, and convert the current and historical voice signals into current word sequences and historical word sequences respectively, and obtain several words from the current word sequences and historical word sequences.

[0028] Specifically, when a user interacts with the intelligent voice robot, the microphone array integrated in the intelligent voice robot collects the user's voice signal and records the current interaction round between the user and the intelligent voice robot, which is denoted as the current interaction round. The voice signal collected at the current moment and the signal collected in a preset historical range are respectively denoted as the current voice signal and the historical voice signal. The specific value of the preset historical range can be set according to the actual situation. This implementation does not make a hard requirement. In this embodiment, the historical range is described using the 10 interaction rounds closest to the current time as an example. The current speech signal and the historical speech signal are converted into the current word sequence and the historical word sequence respectively by the speech recognition algorithm, and a word segmentation algorithm is used to obtain a number of words in the current word sequence and the historical word sequence. Since the speech recognition algorithm and the word segmentation algorithm are well-known existing technologies, they will not be described in detail in this embodiment.

[0029] It should be noted that in actual voice interaction scenarios, users' expressions often have obvious contextual continuity. Different statements in the same interaction task usually have semantic succession or supplementary relationships. Relying solely on the voice signal at a single moment for semantic parsing can easily overlook the information already expressed by the user in previous interactions, thus affecting the completeness of semantic understanding. Therefore, this embodiment simultaneously acquires the user's current voice signal and historical voice signals within a preset historical range, and records the current interaction rounds between the user and the intelligent voice robot. The voice signals are uniformly converted into a word sequence and the word information is extracted. This allows subsequent processing to perform correlation analysis between the current statement and historical statements at a unified word level, thereby improving the accuracy of the voice robot's understanding of the user's true intentions and enhancing the overall coherence and stability of voice interaction.

[0030] Step S202: Based on the distance between words in the current word sequence and words in the historical word sequence, obtain the temporal cumulative weight of words in the current word sequence, and combine the semantic similarity between words in the current word sequence and words in the historical word sequence to obtain the contextual semantic matching degree of words in the current word sequence.

[0031] Specifically, for any word in the current word sequence, words in the historical word sequence that are the same as the words in the current word sequence are recorded as feature words; based on the temporal distance between the historical speech signal corresponding to each feature word and the current speech signal (the specific method for obtaining the temporal distance between the historical speech signal corresponding to each feature word and the current speech signal is: the distance between the start timestamp of the historical speech signal corresponding to each feature word and the start timestamp of the current speech signal is used as the temporal distance between the historical speech signal corresponding to each feature word and the current speech signal), the temporal cumulative weight of the words in the current word sequence is obtained; The temporal distance is negatively correlated with the temporal cumulative weight.

[0032] As an example, the method for obtaining the temporal cumulative weights of the words in the current word sequence is as follows: For the The first feature word, for the first feature word The temporal distance between the historical speech signal corresponding to each feature word and the current speech signal is linearly normalized (a maximum and minimum value normalization function can be used for normalization, and its maximum and minimum values ​​can be obtained based on historical data or prior experience. Adjusting, calibrating, or optimizing the maximum and minimum values ​​does not constitute a limitation of this invention, and the final normalized value range is [0, 1]). The result of the normalization is denoted as the i-th feature word. Feature distance of each feature word; Based on the feature distance of each feature word, the temporal cumulative weight of the words in the current word sequence is obtained, and the specific calculation formula is as follows: In the formula, This represents the cumulative temporal weight of the words in the current word sequence; Indicates the number of feature words; Indicates the first Feature distance of each feature word; This represents the preset time decay coefficient. The specific value can be set according to the actual situation. This embodiment does not make a hard requirement. In this embodiment, it is used as... Let's take an example to illustrate; This embodiment uses an exponential function with the natural constant as the base. The model is used to represent the inverse proportional relationship and for normalization processing. As input to the model, implementers can set inverse proportional functions and normalization functions according to the actual situation.

[0033] It should be noted that in actual voice interaction, users often repeatedly mention certain keywords around a certain topic in continuous dialogue, such as location, object, or operation. These words may appear multiple times in different rounds of interaction, forming cross-round contextual relationships. Therefore, this embodiment retrieves feature words that are the same as words in the current word sequence from the historical word sequence, and performs time-attenuation statistical processing on these feature words in combination with the temporal distance between the corresponding historical speech signal and the current speech signal. This obtains the temporal cumulative weight of the current word in the historical context. By introducing a time attenuation function, the contribution of historical semantic information that is closer in time to the current word is greater, while the influence of historical semantic information that is farther in time gradually weakens. This preserves effective historical contextual information while suppressing the interference of premature statements on the current semantic judgment, thereby more accurately characterizing the importance of the current word in the entire dialogue context and providing a more stable foundation for the subsequent calculation of contextual semantic matching degree and word reliability.

[0034] Furthermore, any word in the current word sequence is designated as the target word, and any word in any historical word sequence is designated as the reference word. The word vectors of the target word and the reference word are obtained respectively using the word2vec algorithm. Based on the cosine similarity between the word vectors of the target word and the reference word, combined with the temporal cumulative weight of the target word, the temporal decay semantic correlation between the target word and the reference word is obtained. Since the word2vec algorithm is a well-known existing technology, it will not be described in detail in this embodiment. The temporal decay semantic correlation between the target word and the control word is positively correlated with the cosine similarity between the word vectors of the target word and the control word; the temporal decay semantic correlation between the target word and the control word is positively correlated with the temporal cumulative weight of the target word. If the temporal decay semantic correlation between the target word and the reference word is less than 0, the temporal decay semantic correlation between the target word and the reference word is set to 0. The mean of the temporal decay semantic correlation between the target word and all words in all historical word sequences is used as the contextual semantic matching degree of the target word.

[0035] As an example, the specific formula for calculating the contextual semantic matching degree of the target word is as follows: In the formula, Indicates the semantic matching degree of the target word in context; Indicates the number of historical word sequences; Indicates the first The number of words in a historical vocabulary sequence; The word vector representing the target word and the first word represents the word vector of the target word. The first in the historical vocabulary sequence Cosine similarity between word vectors of each word; Indicates the time-series cumulative weight of the target word; This indicates the number of words in all historical word sequences; This represents the function that takes the maximum value. This represents a preset, extremely small positive number. Its specific value can be set according to actual circumstances; this embodiment does not impose strict requirements. In this embodiment, it is used as... The purpose of this example is to avoid the denominator being zero during fraction operations, which could lead to calculation errors. The linear normalization function can be normalized using a maximum and minimum value normalization function. The maximum and minimum values ​​can be obtained based on historical data or prior experience. Adjusting, calibrating, or optimizing the maximum and minimum values ​​does not constitute a limitation of this invention. The final normalized value range is [0, 1].

[0036] It should be noted that in actual voice interaction, users often do not always use the exact same words to express the same meaning. Instead, they may use synonyms, near-synonyms, or semantically related expressions. For example, they may use different expressions to describe the same object or behavior in different rounds of interaction. If the association between the current word and the historical context is established solely through word-for-word matching, it is difficult to effectively identify these semantically equivalent or similar expressions, leading to deviations in the voice robot's understanding of the user's true intent. Therefore, this embodiment uses the word2vec algorithm to map words into word vectors in a vector space, and calculates the cosine similarity between the word vectors of the target word and historical words to quantify the semantic closeness between them. This is combined with the aforementioned acquired target word... The temporal cumulative weighting of the tagging term modulates the semantic similarity by temporal decay, making historical semantic information closer to the current interaction round contribute more to the contextual semantic matching degree of the target word, while the influence of historical semantic information from a more distant time gradually decreases. When the semantic similarity between the target word and the reference word is negative, it indicates that the two have opposite or irrelevant relationships in the semantic space. Therefore, the corresponding temporal decay semantic correlation is truncated to 0 to avoid irrelevant semantics interfering with the calculation of contextual matching degree. By statistically analyzing the temporal decay semantic correlation between the target word and all historical words and taking the average, the contextual semantic matching degree of the current word in the overall historical context is quantified, thereby providing a more stable and context-aware semantic basis for the subsequent evaluation of word reliability.

[0037] Step S203: Based on the adjacent words of each word in the current word sequence and the adjacent words of each word in the historical word sequence, obtain the structural fit of each word in the current word sequence; based on the current user interaction round, obtain the contextual semantic matching weight and structural fit weight, and combine the contextual semantic matching and structural fit of the words in the current word sequence to obtain the reliability of the words in the current word sequence.

[0038] Specifically, for any word in the current word sequence, based on the neighboring words of the word in the current word sequence, the probability statistics of the neighboring words of all words in all historical sequences are performed using the N-gram model to obtain the conditional probability of the word in the current word sequence under the corresponding adjacent context conditions, which is used as the structural fit of the word in the current word sequence. Since the N-gram model is a well-known existing technology, it will not be described in detail in this embodiment.

[0039] It should be noted that in actual voice interaction, even if the speech recognition system can correctly identify individual words, their arrangement in a sentence may still be affected by factors such as spoken expression habits, pauses, word order adjustments, or recognition errors. This can lead to deviations between the structural position of some words in the current word sequence and the normal semantic expression structure. Therefore, this embodiment introduces an N-gram language model to perform probability statistics on the adjacency relationships between words in historical interaction corpora, establish the probability distribution of word occurrence under different contextual structural conditions, and depict the structural rationality of words in specific contexts. If a word in the current word sequence and its adjacent word combination have a high probability of occurrence in historical corpora, then the usage of that word in the current context structure has a high consistency with the common expression structure in historical corpora, and its structural fit is relatively high. This allows for a quantitative evaluation of the rationality of words in the current sentence from the perspective of language structure, thus providing a structural basis for subsequent comprehensive evaluation of word reliability by combining contextual semantic matching.

[0040] Furthermore, an initial context semantic matching degree weight and an amplification coefficient are preset. The specific values ​​of the initial context semantic matching degree weight and the amplification coefficient can be set according to the actual situation. This embodiment does not make a hard requirement. In this embodiment, the initial context semantic matching degree weight is equal to 0.5 and the amplification coefficient is equal to 0.1 as an example. The context semantic matching degree weight is obtained according to the user's current interaction round, the initial context semantic matching degree weight and the amplification coefficient. The contextual semantic matching weights are all positively correlated with the user's current interaction round, the initial contextual semantic matching weight, and the growth coefficient, respectively. The difference between 1 and the contextual semantic matching weight is used as the structural fit weight.

[0041] As an example, the specific formula for calculating the weight of contextual semantic matching is as follows: In the formula, Indicates the weight of semantic matching degree in context; Indicates the current user interaction round; This indicates the weight of the initial semantic matching degree in the context. This indicates the preset amplification factor; This represents an exponential function with the natural constant as its base.

[0042] It should be noted that during voice interaction, the importance of semantic and sentence structure information in understanding user intent is not constant across different interaction stages. In the initial stage of interaction, due to the limited historical context information available to the system, the semantic matching relationship between words in the current sentence has not yet formed a stable context. At this time, over-reliance on semantic association information may lead to semantic judgment bias due to insufficient context samples. Therefore, in the initial stage of interaction, by giving relatively high weight to structural fit, the rationality of language structure can be prioritized for word selection, thereby ensuring the stability of the basic expression structure of the sentence. As the number of user interaction rounds increases, the historical voice and context information accumulated by the system gradually becomes richer, and the semantic association relationship between the current words and historical words gradually stabilizes. At this time, the reference value of contextual semantic information in intent understanding gradually increases, thus enabling a more stable and accurate characterization of the credibility of words in the current sentence at different interaction stages.

[0043] Furthermore, for any word in the current word sequence, the contextual semantic matching degree and structural fit degree of the word in the current word sequence are weighted and summed using the contextual semantic matching degree weight and the structural fit degree weight to obtain the reliability of the word in the current word sequence.

[0044] It should be noted that this embodiment uses a weighted fusion of the contextual semantic matching degree and structural fit degree of each word in the current word sequence to comprehensively evaluate the credibility of word expression from two dimensions: semantic relevance and language structure rationality. This yields the reliability of each word in the current sentence. This corrects for possible homonym misidentification, dialectal differences, or word order deviations that may occur during speech recognition, enabling subsequent intent recognition processes to be analyzed based on more reliable lexical information, thereby improving the accuracy and stability of overall speech interaction understanding.

[0045] Step S204: Based on the reliability of the words in the current word sequence, obtain the user's actual expressive intent.

[0046] Specifically, the reliability of each word in the current word sequence is used as a label for each word in the current word sequence to obtain the current word sequence with labels. The current word sequence with labels is then converted into a natural language prompt (Prompt) format and input into a large language model (LLM). The large language model outputs the intent recognition result of the current word sequence to obtain the user's actual expression intent. Since the conversion of the natural language prompt (Prompt) format and the large language model (LLM) are both well-known existing technologies, they will not be described in detail in this embodiment.

[0047] It should be noted that this embodiment uses the reliability of each word in the current word sequence as a label to attach to the corresponding word. This allows the credibility of different words in the semantic understanding process to be explicitly expressed when inputting into the large language model. As a result, when performing semantic reasoning and intent recognition, the large language model can prioritize the information of words with higher reliability and reduce the interference of words with lower reliability on semantic understanding. In this way, without changing the original structure of the large language model, the word reliability information constructed in the aforementioned steps is used to enhance the model input. This enables the model to more accurately infer the user's true intention even under complex speech expressions, dialect interference, or recognition errors, thereby further improving the accuracy of intent recognition and the stability of interaction in the intelligent voice robot interaction process.

[0048] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0049] Please see Figure 2 The diagram illustrates a structural block diagram of a large-model-driven intelligent voice robot interaction system according to an embodiment of the present invention. The system includes a memory 202, a processor 201, and a computer program 2021 stored in the memory 202 and executable on the processor. When the processor 201 executes the computer program 2021, it implements steps S101 to S104 of the large-model-driven intelligent voice robot interaction method.

[0050] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A large model driven intelligent voice robot interaction method, characterized in that, The method comprises: acquiring current and historical voice signals of a user, recording a current interaction round of the user, and converting the current and historical voice signals into a current vocabulary sequence and a historical vocabulary sequence respectively, and acquiring a plurality of vocabularies in the current and historical vocabulary sequences; obtaining a time sequence cumulative weight of a vocabulary in the current vocabulary sequence according to a distance between vocabularies in the current and historical vocabulary sequences, and obtaining a context semantic matching degree of the vocabulary in the current vocabulary sequence in combination with a semantic similarity between the vocabulary in the current vocabulary sequence and the vocabulary in the historical vocabulary sequence; obtaining a structure adaptation degree of each vocabulary in the current vocabulary sequence in combination with adjacent vocabularies of each vocabulary in the historical vocabulary sequence according to the adjacent vocabularies of each vocabulary in the current vocabulary sequence, and obtaining a reliability of the vocabulary in the current vocabulary sequence in combination with the context semantic matching degree and the structure adaptation degree of the vocabulary in the current vocabulary sequence according to a current interaction round of the user; obtaining an actual expression intention of the user according to the reliability of the vocabulary in the current vocabulary sequence.

2. The large model-driven intelligent voice robot interaction method of claim 1, wherein, The acquisition of the current and historical voice signals of the user, the recording of the current interaction round of the user, and the conversion of the current and historical voice signals into the current vocabulary sequence and the historical vocabulary sequence, and the acquisition of a plurality of vocabularies in the current and historical vocabulary sequences comprise: when the user interacts with the intelligent voice robot, the voice signal of the user is collected through the microphone array integrated in the intelligent voice robot, and the current interaction round of the user with the intelligent voice robot is recorded, which is denoted as the current interaction round of the user; the voice signal collected at the current time and the signal collected in the preset historical range are denoted as the current voice signal and the historical voice signal respectively; the current voice signal and the historical voice signal are converted into the current vocabulary sequence and the historical vocabulary sequence respectively through a voice recognition algorithm, and a plurality of vocabularies in the current and historical vocabulary sequences are acquired by using a word segmentation algorithm.

3. The large model-driven intelligent voice robot interaction method of claim 1, wherein, The acquisition of the time sequence cumulative weight of the vocabulary in the current vocabulary sequence according to the distance between the vocabularies in the current and historical vocabulary sequences comprises: for any vocabulary in the current vocabulary sequence, the vocabulary in the historical vocabulary sequence that is the same as the vocabulary in the current vocabulary sequence is denoted as a feature word; the time sequence cumulative weight of the vocabulary in the current vocabulary sequence is obtained according to the time sequence distance between the historical voice signal corresponding to each feature word and the current voice signal; wherein the time sequence distance and the time sequence cumulative weight are negatively correlated.

4. The large model-driven intelligent voice robot interaction method of claim 1, wherein, The acquisition of the context semantic matching degree of the vocabulary in the current vocabulary sequence comprises: any vocabulary in the current vocabulary sequence is denoted as a target word, and any vocabulary in any historical vocabulary sequence is denoted as a control word; the word vector of the target word and the word vector of the control word are obtained respectively through a word2vec algorithm; the time sequence decay semantic correlation degree between the target word and the control word is obtained according to the cosine similarity between the word vector of the target word and the word vector of the control word in combination with the time sequence cumulative weight of the target word; The temporal decay semantic correlation between the target word and the control word is positively correlated with the cosine similarity between the word vectors of the target word and the control word; the temporal decay semantic correlation between the target word and the control word is positively correlated with the temporal cumulative weight of the target word. The contextual semantic matching degree of the target word is obtained based on the temporal decay semantic correlation between the target word and all words in all historical word sequences.

5. The large model driven intelligent voice robot interaction method of claim 4, wherein, The acquisition of the contextual semantic matching degree of the target word includes: If the temporal decay semantic correlation between the target word and the reference word is less than 0, the temporal decay semantic correlation between the target word and the reference word is set to 0. The mean of the temporal decay semantic correlation between the target word and all words in all historical word sequences is used as the contextual semantic matching degree of the target word.

6. The large model-driven intelligent voice robot interaction method of claim 1, wherein, The step of obtaining the structural fit of each word in the current word sequence based on its neighboring words and in conjunction with the neighboring words in the historical word sequence includes: For any word in the current word sequence, based on the neighboring words of the word in the current word sequence, the probability statistics of the neighboring words of all words in all historical sequences are performed using the N-gram model to obtain the conditional probability of the word in the current word sequence under the corresponding adjacent context conditions, which is used as the structural fit of the word in the current word sequence.

7. The large model-driven intelligent voice robot interaction method of claim 1, wherein, The process of obtaining the contextual semantic matching weight and structural fit weight based on the user's current interaction round includes: An initial context semantic matching degree weight and amplification coefficient are preset. The context semantic matching degree weight is obtained based on the user's current interaction round, the initial context semantic matching degree weight and amplification coefficient. The contextual semantic matching weights are all positively correlated with the user's current interaction round, the initial contextual semantic matching weight, and the growth coefficient, respectively. The difference between 1 and the contextual semantic matching weight is used as the structural fit weight.

8. The large model-driven intelligent voice robot interaction method of claim 1, wherein, The reliability of obtaining words in the current word sequence includes: For any word in the current word sequence, the reliability of the word in the current word sequence is obtained by weighting and summing the contextual semantic matching degree and the structural fit degree of the word in the current word sequence through the contextual semantic matching degree weight and the structural fit degree weight.

9. The large model-driven intelligent voice robot interaction method of claim 1, wherein, The step of obtaining the user's actual expressive intent based on the reliability of words in the current word sequence includes: The reliability of each word in the current word sequence is used as the label of each word in the current word sequence to obtain the current word sequence with labels. The current word sequence with labels is converted into natural language prompt word format and input into the large language model. The large language model outputs the intent recognition result of the current word sequence to obtain the user's actual expression intent.

10. A large model driven intelligent voice robot interaction system, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a large model-driven intelligent voice robot interaction method as described in any one of claims 1-9.

Images