A method and system for quantitatively measuring cross-scene speech dialogue behavior rules

Abstract

This invention discloses a quantitative measurement method and system for cross-scenario voice dialogue behavior patterns, relating to the field of language processing technology. The method includes: acquiring the voice dialogue stream in the target scenario, parsing turn information and intent tags, constructing an intent sequence and recording response delays; calculating information entropy based on the probability of occurrence of different intents and outputting an uncertainty coefficient; obtaining the intent transition confidence sequence at each adjacent position through an intent transition confidence model; using the intent transition confidence sequence to perform weighted correction on the initial measurement to obtain a temporal dependency correction measurement; determining a stiffness measurement based on the frequency distribution of directed intent transition pairs and performing a secondary correction to obtain a topological stiffness correction measurement; and fusing the temporal dependency correction measurement and the topological stiffness correction measurement to output a comprehensive measurement index. This invention solves the technical problem in existing technologies where reliance on manually preset templates leads to high misjudgment rates and severely distorted measurement results.

Description

Technical Field

[0001] This invention relates to the field of language processing technology, specifically to a quantitative measurement method and system for cross-scenario voice dialogue behavior patterns. Background Technology

[0002] In different scenarios of voice interaction, dialogue behavior often exhibits specific structured patterns, which reflect the certainty and predictability of intent flow during the interaction process. Therefore, it is necessary to quantify the regularity of dialogue behavior to facilitate the design optimization of dialogue systems, abnormal interaction detection, and user experience evaluation.

[0003] Traditional methods primarily rely on manually preset dialogue state nodes, keyword frequency statistics, and fixed path matching counts. They roughly measure the structural convergence of behavior by calculating the overlap rate or edit distance between the actual interaction trajectory and the preset template. However, due to the rigidity and lack of adaptability of the preset template, it is prone to misjudging the actual underlying structural evolution as matching deviation when facing open scenarios or natural interaction changes. This leads to severely distorted measurement results and fails to objectively reflect the dynamic nature of dialogue behavior.

[0004] Therefore, there is a need for a quantitative measurement method and system that can quantify the regularity intensity of speech dialogue behavior and adaptively capture the dynamic evolution characteristics of the underlying dialogue structure. Summary of the Invention

[0005] This invention provides a quantitative measurement method and system for cross-scenario voice dialogue behavior patterns, which solves the technical problems of high misjudgment rate and serious distortion of measurement results caused by the reliance on manually preset templates in the prior art.

[0006] In a first aspect, the present invention provides a method for quantitatively measuring the patterns of cross-scenario voice dialogue behavior, the method comprising:

[0007] Acquire the voice dialogue stream in the target scenario, parse the round information and the intent tag corresponding to each round in the voice dialogue stream, construct the intent sequence with round as the time unit, and record the response delay between adjacent rounds;

[0008] The information entropy is calculated based on the probability of occurrence of different intentions in the intention sequence. The information entropy is used as a quantification value of the uncertainty of dialogue intention invocation, and the uncertainty coefficient is output.

[0009] The intent sequence is input into the intent transition confidence model, and the intent transition confidence sequence at each adjacent position is output. The intent transition confidence model is constructed based on the frequency distribution of intent transitions in historical dialogue data across scenarios.

[0010] The difference between the uncertainty coefficient and the conditional entropy of adjacent intentions is used as the initial measure of temporal dependence, and the initial measure of temporal dependence is weighted and corrected using the intention transfer confidence sequence to obtain the corrected measure of temporal dependence.

[0011] Based on the frequency distribution of directed intention transition pairs in the intention sequence, the stiffness metric of the state transition skeleton is determined, and the stiffness metric is further corrected using the distribution characteristics of the response delay to obtain the topological stiffness correction metric.

[0012] By integrating the temporal dependency correction metric and the topological stiffness correction metric, a comprehensive metric index is output to quantify the intensity of the regularity of voice dialogue behavior in the target scenario.

[0013] Secondly, the present invention provides a quantitative measurement system for cross-scenario voice dialogue behavior patterns, the system comprising:

[0014] An intent sequence module is constructed to acquire the voice dialogue stream in the target scenario, parse the round information and the intent tag corresponding to each round in the voice dialogue stream, construct an intent sequence with round as the time unit, and record the response delay between adjacent rounds.

[0015] The uncertainty coefficient acquisition module is used to calculate the information entropy based on the probability of occurrence of different intentions in the intention sequence, use the information entropy as a quantification value of the uncertainty of the dialogue intention call, and output the uncertainty coefficient.

[0016] The intent transfer confidence model construction module is used to input the intent sequence into the intent transfer confidence model and output the intent transfer confidence sequence at each adjacent position, wherein the intent transfer confidence model is constructed based on the intent transfer frequency distribution in cross-scene historical dialogue data;

[0017] The weighted correction module is used to take the difference between the uncertainty coefficient and the conditional entropy of adjacent intentions as the initial measure of temporal dependence, and to use the intention transfer confidence sequence to perform weighted correction on the initial measure of temporal dependence to obtain the temporal dependence correction measure.

[0018] The secondary correction module is used to determine the stiffness metric of the state transition skeleton based on the frequency distribution of directed intention transition pairs in the intention sequence, and to perform secondary correction on the stiffness metric using the distribution characteristics of the response delay to obtain the topological stiffness correction metric.

[0019] The comprehensive metric index fusion module is used to fuse the temporal dependency correction metric and the topological stiffness correction metric to output a comprehensive metric index, thereby quantifying the intensity of the regularity of voice dialogue behavior in the target scenario.

[0020] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0021] This invention provides a quantitative measurement method and system for cross-scenario voice dialogue behavior patterns. First, it transforms the original unstructured voice dialogue stream into a structured temporal symbol sequence and synchronous temporal features, reducing subsequent computational complexity and obtaining intent recognition results. Second, it characterizes the dispersion or concentration of various intent calls by the dialogue system from a macro-statistical perspective of intent distribution, avoiding the loss of global distribution information due to relying solely on local transition patterns. Third, it constructs an intent transition confidence model to obtain intent transition confidence, overcoming the inability to distinguish the reliability differences of different transition positions. Furthermore, it uses the uncertainty coefficient and adjacent intent conditional entropy to obtain an initial measure of temporal dependency, and uses the arithmetic mean of intent transition confidence as a correction weight for weighted correction, suppressing the interference of abnormal transitions on temporal dependency evaluation. Finally, it quantifies the structural rigidity of the dialogue state transition skeleton, obtains the response delay variation coefficient, and performs a secondary correction on the initial stiffness measure to obtain a topological stiffness correction measure. Ultimately, by integrating the temporal constraint strength of intent flow and the rigidity dimension of state transition skeleton structure, a comprehensive metric index will be generated to quantify the regularity of voice dialogue behavior and improve the interpretability and comparability of the metric results. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a method for quantifying and measuring cross-scenario voice dialogue behavior patterns provided by the present invention.

[0023] Figure 2 This is a logical diagram of a method for quantitatively measuring cross-scenario voice dialogue behavior patterns provided by the present invention.

[0024] Figure 3 This is a schematic diagram of the structure of a quantitative measurement system for cross-scenario voice dialogue behavior patterns provided by the present invention.

[0025] In the attached diagram, the components represented by each number are as follows:

[0026] The module includes: Module 11 for constructing the intent sequence; Module 12 for obtaining the uncertainty coefficient; Module 13 for constructing the intent transfer confidence model; Module 14 for weighted correction; Module 15 for secondary correction; and Module 16 for fusion of comprehensive measurement index. Detailed Implementation

[0027] This invention provides a quantitative measurement method and system for cross-scenario voice dialogue behavior patterns, which solves the technical problems of high misjudgment rate and serious distortion of measurement results caused by the reliance on manually preset templates in the prior art.

[0028] The present invention will now be described in detail with reference to the accompanying drawings.

[0029] Example 1, as Figure 1 , Figure 2 As shown, this invention provides a method for quantitatively measuring the behavioral patterns of cross-scenario voice dialogue, the method comprising:

[0030] S100: Obtain the voice dialogue stream in the target scenario, parse the round information and the intent label corresponding to each round in the voice dialogue stream, construct the intent sequence with round as the time unit, and record the response delay between adjacent rounds;

[0031] In this embodiment of the invention, the target scenario refers to the specific application environment in which the dialogue behavior patterns are to be quantified, including but not limited to intelligent customer service dialogue scenarios, in-vehicle voice interaction scenarios, medical consultation dialogue scenarios, or online education and tutoring dialogue scenarios; the voice dialogue stream is a continuous voice signal data stream generated by one or more parties participating in the interaction in the target scenario.

[0032] Specifically, voice acquisition devices deployed in the target scenario are used to capture real-time voice dialogue streams. Voice activity detection is performed on the acquired voice dialogue streams to distinguish between voice and non-voice periods, and speaker separation is also performed. The speech segments of different participants are differentiated to obtain voice activity segments and their corresponding time boundaries. To prevent the same speaker from being segmented into multiple segments due to short pauses, adjacent voice activity segments belonging to the same speaker and with an interval length below the fusion threshold are merged into independent rounds to prevent erroneous segmentation. After fusion processing, each independent round is assigned a round number, start timestamp, and end timestamp. Automatic speech recognition technology is used to convert the independent rounds into text sequences, and semantic matching is performed between the text sequences and intent labels. After completing intent label recognition for all rounds, the intent labels are arranged sequentially according to the round number from smallest to largest, constructing an intent sequence with rounds as the time unit. Simultaneously, for any two adjacent independent rounds, the start timestamp of the later round is subtracted from the end timestamp of the earlier round to obtain the response delay value between the adjacent rounds, which is stored sequentially as a response delay sequence.

[0033] Step S100 in the method of this embodiment of the invention includes:

[0034] Real-time capture of voice dialogue streams using voice acquisition devices deployed within the target scenario;

[0035] The speech dialogue stream is subjected to speech activity detection and speaker separation to obtain multiple speech activity segments arranged in time sequence and the time boundary corresponding to each speech activity segment;

[0036] Adjacent speech activity segments belonging to the same speaker and with an interval duration less than the fusion threshold are merged into independent rounds, and each independent round is assigned a round number, a start timestamp, and an end timestamp.

[0037] For the speech signal in each independent round, a text sequence is obtained through speech recognition. The text sequence is then semantically matched with the scene intent tag set to output the intent tag corresponding to the independent round.

[0038] Arrange the intent tags corresponding to each independent round in ascending order of round number to construct an intent sequence with round as the time unit;

[0039] For any two adjacent independent rounds, subtract the end timestamp of the previous round from the start timestamp of the later round to obtain the response delay between adjacent rounds, and store the response delays between all adjacent rounds in sequence according to the round number as a response delay sequence.

[0040] In this embodiment of the invention, firstly, a voice acquisition device deployed within the target scene—that is, a device in the target scene used to pick up dialogue voice signals, including but not limited to microphone arrays, built-in microphones of smart terminals, voice recorders, or IoT devices with voice acquisition capabilities—captures the voice dialogue stream in real time. The voice acquisition device uses a preset sampling rate, such as 16kHz, to capture ambient sounds within the scene in real time and converts the analog sound wave signals into a digital pulse-code modulated audio stream.

[0041] Secondly, speech activity detection and speaker separation are performed on the speech dialogue stream. Specifically, a speech activity detection algorithm analyzes the audio stream frame by frame, calculating features such as short-time energy, zero-crossing rate, or speech probability based on deep learning, thereby distinguishing between speech frames and non-speech frames and merging consecutive speech frames into multiple speech activity segments. Speaker separation technology, in multi-person dialogue audio, classifies the speakers according to their identities based on differences in acoustic features, separating speech segments belonging to different speakers, and sorting them chronologically to obtain multiple speech activity segments arranged in chronological order and the corresponding time boundaries of each speech activity segment. Speech activity detection is used to identify time periods containing human voice signals in a continuous audio stream and exclude non-speech segments such as silence and background noise.

[0042] Next, all speech segments belonging to the same speaker are traversed in chronological order. The interval between two adjacent segments is calculated, i.e., the difference between the end time of the previous segment and the start time of the next segment. If this interval is less than a preset fusion threshold, the two segments are merged, considered as part of the same round, and the end timestamp of that round is updated. If the interval is greater than or equal to the fusion threshold, the previous segment is considered an independent round. After all merging operations are completed, for each final independent round, an incremental round number is assigned according to its chronological order in the dialogue, and the start and end timestamps of the round as a whole are finally determined. The fusion threshold is a preset time length value, such as 1 second or 1.5 seconds, used to determine whether two adjacent speech segments belonging to the same speaker should be considered as consecutive speeches in the same round. It can adaptively determine the value based on statistical analysis of historical dialogue data. It extracts the interval duration samples of adjacent speech segments from the historical dialogue data of the target scene and fits a bimodal distribution. The intersection of the probability density curves of the Gaussian components of the corresponding round pause and round switching is used as the initial value. The search is performed with this candidate value as the center, and the value that makes the round division most reasonable is selected as the fusion threshold.

[0043] Furthermore, the speech recognition engine is invoked for the speech signal within each independent round to convert the speech into a text sequence. The recognized text sequence is then normalized to obtain normalized text. This normalized text is then semantically matched with each intent tag in the scene intent tag set, and the intent tag corresponding to that round is determined based on the degree of matching.

[0044] Furthermore, the round number and intent label assigned to each round are obtained. The intent labels corresponding to each independent round are arranged sequentially according to the round number in ascending order, constructing an intent sequence with round as the time unit.

[0045] Finally, for any two adjacent independent rounds, the start timestamp of the later round is subtracted from the end timestamp of the earlier round. The difference is the response delay between the adjacent rounds. All response delays between adjacent rounds are then stored sequentially by round number as a response delay sequence. The length of the response delay sequence is one less than the length of the intention sequence, and it can be used for topology stiffness correction metrics.

[0046] In the method of this embodiment of the invention, for the speech signal in each independent round, a text sequence is obtained through speech recognition, the text sequence is semantically matched with a scene intent tag set, and the intent tag corresponding to the independent round is output, including:

[0047] The text sequence obtained from the speech recognition is subjected to text normalization processing to obtain normalized text, and the normalized text is segmented into words, stop words are removed, and a content word sequence is obtained.

[0048] For each intent tag in the scene intent tag set, obtain a set of synonym expressions associated with the intent tag. The scene intent tag set is constructed by deduplicating and summarizing all the labeled intent tags in the historical dialogue data and retaining the intent tags that appear more than a frequency threshold in the current target scene.

[0049] The step of determining the frequency threshold includes:

[0050] The frequency of each candidate intent tag in the historical dialogue data is counted, and the mean and standard deviation of the frequency of the tag occurrences are extracted.

[0051] The frequency threshold is the sum of the frequency mean and the frequency standard deviation.

[0052] Calculate the word hit rate between the content word sequence and the synonym expression word set, as well as the longest consecutive hit length of the hit words in the content word sequence. Then, sum the normalized values ​​of the word hit rate and the longest consecutive hit length in a weighted manner to obtain the semantic matching base score.

[0053] The probability of occurrence of the intent tag is statistically analyzed from the historical dialogue data, and combined with the semantic matching base score to obtain the prior corrected matching score;

[0054] Compare the prior corrected matching scores of all intent tags in the scene intent tag set, and select the intent tag corresponding to the largest prior corrected matching score as the intent tag output for the independent round.

[0055] In this embodiment of the invention, firstly, the text sequence obtained from speech recognition within the current independent round is subjected to text normalization processing to obtain normalized text. Then, the normalized text is segmented into words, dividing the continuous character sequence into independent lexical units and removing stop words to obtain a content word sequence. Text normalization processing refers to the process of standardizing the original text output by speech recognition, including but not limited to: converting full-width characters to half-width characters, standardizing the expression format of dates and numbers, correcting common recognition errors, removing meaningless filler words, and standardizing punctuation marks. Stop words refer to common words that frequently appear in different texts but contribute very little to semantic intent judgment.

[0056] Secondly, for each intent tag in the scene intent tag set, a set of synonyms associated with the intent tag is obtained. This set of synonyms refers to a collection of different words or phrases that can express the same intent, covering the diverse wording a user might use to express the same intent. These can be obtained by mining high-frequency words co-occurring with the target intent tag from historical dialogue data, or by using human expert annotation. The scene intent tag set is constructed by deduplicating and summarizing all annotated intent tags in historical dialogue data, retaining intent tags that appear more than a frequency threshold in the current target scene.

[0057] The steps for determining the frequency threshold include:

[0058] The frequency of each candidate intent tag in historical dialogue data was counted, and then the mean and standard deviation of the tag frequency were extracted.

[0059] The frequency threshold is calculated by summing the frequency mean and the frequency standard deviation. Candidate intent tags that appear more frequently than this threshold are retained, forming a set of scene intent tags for the current target scene. That is, frequency threshold = frequency mean + frequency standard deviation.

[0060] Next, the lexical hit rate between the content word sequence and the synonym expression set, as well as the longest consecutive hit length of the hit words in the content word sequence, are calculated. The normalized values ​​of the lexical hit rate and the longest consecutive hit length are then weighted and summed to obtain the semantic matching base score. The lexical hit rate refers to the number of words in the content word sequence that hit the synonym expression set of the current intent tag. Dividing the number of hit words by the total number of words in the content word sequence reflects the density of the match between the content word sequence and the target intent tag at the lexical level. The longest consecutive hit length refers to the length of the longest consecutive subsequence that hits the synonym expression set in the content word sequence.

[0061] Specifically, since the longest consecutive hit length is a count value, its range depends on the length of the content word sequence. Therefore, it needs to be normalized, for example, by dividing the longest consecutive hit length by the total number of words in the content word sequence, so that it is on the same order of magnitude as the word hit rate. The word hit rate and the normalized longest consecutive hit length are then weighted and summed to obtain the semantic matching base score. The weight coefficients of the weighted summation can be preset and adjusted according to the actual scenario. For example, the word hit rate can be assigned a weight of 0.6, and the longest consecutive hit length can be assigned a weight of 0.4, to balance the contribution of the overall word matching density and the intensity of consecutive expressions to the matching result.

[0062] Furthermore, in dialogue scenarios, the frequency of occurrence of different intent tags often varies significantly. Ignoring the prior distribution of the intent itself may lead to low-frequency intents consistently being at a disadvantage in matching, even if their textual expression is very clear, they are easily misjudged as high-frequency intents. Therefore, the occurrence probability of intent tags is statistically analyzed from historical dialogue data: occurrence probability = number of occurrences of intent tag / total number of occurrences of all intent tags. This is then combined with the semantic matching base score to obtain the prior corrected matching score. Wherein, the prior corrected matching score = semantic matching base score × occurrence probability.

[0063] Finally, the prior corrected matching scores of all intent tags in the scene intent tag set are compared, and the intent tag corresponding to the largest prior corrected matching score is selected as the intent tag output for the independent round.

[0064] Specifically, for each independent round of content word sequence, a priori corrected matching score is calculated between it and each intent tag in the scene intent tag set, resulting in multiple matching scores equal to the number of intent tags. The intent tag with the highest score is selected as the final intent recognition result for that round.

[0065] In this embodiment of the invention, continuous audio streams are accurately segmented through speech activity detection, speaker separation, and fusion threshold merging, avoiding over-segmentation due to short pauses or under-segmentation due to improper silence threshold settings. Intent tags for each round are obtained through semantic matching, transforming high-dimensional speech signals into low-dimensional discrete symbol sequences, reducing subsequent computational complexity. Response delay provides a data foundation for the regularity evaluation of fused interaction rhythm information, ensuring the integrity and temporal accuracy of the link information. Subsequently, the final intent recognition result is obtained through text normalization, frequency thresholding, synonym set matching, longest consecutive hit analysis, and prior probability correction, making the final intent recognition result more adaptable to open scenarios.

[0066] S200: Calculate the information entropy based on the probability of occurrence of different intentions in the intention sequence, use the information entropy as a quantification value of the uncertainty of dialogue intention invocation, and output the uncertainty coefficient;

[0067] In this embodiment of the invention, the occurrence probability is the proportion of the frequency of a specific intent label appearing in the intent sequence to the total length of the intent sequence; information entropy is a fundamental quantity in information theory used to measure the uncertainty of random variables.

[0068] Specifically, the intent sequence is traversed, and the frequency of intent tags appearing in the intent sequence within the scene intent tag set is counted to determine the total length of the intent sequence. The total number of intents in the scene intent tag set and their corresponding frequency thresholds are then obtained. To improve the smoothness of the probability estimation and avoid infinite values ​​in the information entropy calculation due to zero-frequency issues, additive smoothing techniques are used to correct the occurrence probabilities of intent tags, and the fused occurrence probability of each intent tag is calculated. After obtaining the fused occurrence probabilities of all intent tags, the information entropy is calculated according to the definition of information entropy, and this calculation result is output as the uncertainty coefficient.

[0069] Step S200 in the method of this embodiment of the invention includes:

[0070] Traverse the intent sequence, count the frequency of each intent tag in the scene intent tag set in the intent sequence, and determine the total length of the intent sequence;

[0071] Obtain the total number of intents in the scene intent tag set, and the frequency threshold corresponding to the scene intent tag set;

[0072] For each intent tag, the probability of fusion occurrence is calculated based on the additive smoothing relationship between the occurrence frequency, the total length of the intent sequence, the total number of intents, and the frequency threshold.

[0073] The information entropy is calculated based on the fusion occurrence probability of all intent labels, and the calculation result is output as the uncertainty coefficient.

[0074] In this embodiment of the invention, firstly, the intent sequence is traversed, the frequency of each intent tag in the scene intent tag set in the intent sequence is counted, and the total length of the intent sequence is determined.

[0075] Specifically, a counting vector of the same size as the scene intent tag set is initialized, with all counts initially set to zero. Each intent tag in the intent sequence is traversed sequentially, and for each intent tag found, the corresponding position in the counting vector is incremented by 1. After traversal, each element in the counting vector corresponds to the frequency of an intent tag appearing in the current dialogue. Simultaneously, the total length of the intent sequence is recorded, i.e., the total number of rounds in the dialogue. For example, suppose the intent sequence of a dialogue is [Query menu, Query menu, Place order, Place order, Place order, Confirm address, End dialogue], with a total length of 7. The scene intent tag set contains 6 intent tags: [Greeting, Query menu, Place order, Confirm address, End dialogue, Cancel order]. After traversal, the frequency of each intent tag is calculated.

[0076] Secondly, obtain the total number of intents in the scene intent tag set, and the frequency threshold corresponding to the scene intent tag set. The total number of intents is the number of intent tags contained in the scene intent tag set.

[0077] Next, for each intent tag, the probability of fusion occurrence is calculated based on the additive smoothing relationship between occurrence frequency, total length of intent sequence, total number of intents, and frequency threshold.

[0078] In actual intent sequences, due to the limited number of dialogue turns, some intent tags may never appear in the current dialogue. Setting the occurrence probability directly to zero would distort the estimation of information about intents that did not appear in the information entropy calculation. Therefore, additive smoothing is used to correct the occurrence probability estimation. For each intent tag, the fused occurrence probability is calculated using the sum of the occurrence frequency and a frequency threshold as the numerator, and the sum of the total length and the total number of intents multiplied by the frequency threshold as the denominator.

[0079] Finally, the information entropy is calculated based on the fused occurrence probabilities of all intent tags, and the result is output as the uncertainty coefficient. A larger information entropy value indicates a more uniform distribution of the occurrence probabilities of each intent tag in the intent sequence, and a higher uncertainty in intent invocation; a smaller information entropy value indicates a more concentrated intent distribution, and a lower uncertainty.

[0080] Specifically, information entropy Where p is the probability of the fused occurrence of the i-th intent tag, the logarithm is usually calculated using the natural logarithm e or base 2, and the corresponding entropy unit is knight. Substituting the calculated fused occurrence probabilities of all intent tags into this formula, taking the logarithm of each probability value, multiplying it by the probability itself, and then summing all the products and taking the opposite, yields the information entropy value of the dialogue intent sequence, i.e., the uncertainty coefficient.

[0081] In the method of this embodiment of the invention, the training step of the intent transfer confidence model includes:

[0082] Extract multiple intent sequences corresponding to dialogues from historical dialogue data across different scenarios, traverse adjacent intent transition pairs in all intent sequences, and count the original transition frequency of each intent transition pair.

[0083] The training sample set is formed by using the transfer pairs whose original transfer frequency is higher than the frequency threshold as positive samples and the transfer pairs whose original transfer frequency is zero as negative samples.

[0084] Construct an intent transition confidence model based on a sequence-to-sequence confidence estimation architecture, wherein the architecture includes a sequence encoder for encoding an intent sequence into a temporal hidden state sequence, and a confidence decoder for decoding the temporal hidden state sequence into transition confidence at each adjacent position.

[0085] Using the intent sequence fragments in the training sample set as input and the positive and negative sample labels at corresponding adjacent positions as supervision signals, the intent transfer confidence model is trained end-to-end in a supervised manner using a binary classification cross-entropy loss function.

[0086] When the classification accuracy of the intent transfer confidence model on the independent validation set for positive and negative samples exceeds the accuracy threshold, training is stopped, and the pre-training of the intent transfer confidence model is completed.

[0087] When the intent transfer confidence model receives any intent sequence, it sequentially inputs the temporal hidden state combination corresponding to each adjacent intent transfer pair in the intent sequence into the confidence decoder, and arranges the intent transfer confidence estimates output by the confidence decoder in the order of adjacent positions to form an intent transfer confidence sequence.

[0088] In this embodiment of the invention, firstly, multiple intent sequences corresponding to dialogues are extracted from cross-scene historical dialogue data. Adjacent intent transition pairs in all intent sequences are traversed. For each intent sequence, two adjacent intent tags constitute an ordered transition pair. For example, the sequence [intent A, intent B, intent C] contains two transition pairs: (intent A → intent B) and (intent B → intent C). By traversing all adjacent positions in the intent sequences of all historical dialogues and accumulating the occurrence count of each directed transition pair, the original transition frequency of each intent transition pair can be obtained. Here, an intent transition pair refers to an ordered binary tuple composed of intent tags corresponding to any two adjacent rounds in the intent sequence, including the preceding intent and the succeeding intent.

[0089] Secondly, a training sample set is formed by using transition pairs with an original transition frequency higher than a frequency threshold as positive samples and transition pairs with an original transition frequency of zero as negative samples. The input for each sample is a corresponding intent sequence fragment, and the label indicates whether the transition pair is a sample. Positive samples represent intent transition patterns that have occurred in real dialogues and have statistical significance; negative samples represent intent transition pairs that have never been observed in historical data, which may be illogical or violate dialogue norms. By constructing positive and negative sample sets, the intent transition confidence model can distinguish between reasonable and unreasonable intent transitions.

[0090] Next, an intent transition confidence model based on a sequence-to-sequence confidence estimation architecture is constructed. This architecture includes a sequence encoder for encoding the intent sequence into a temporal hidden state sequence, and a confidence decoder for decoding the temporal hidden state sequence into transition confidence at each adjacent position. The sequence encoder can employ any of the following: a recurrent neural network, a long short-term memory network, a gated recurrent unit, or a Transformer encoder. It receives the intent sequence and outputs the hidden state vector corresponding to each time step. The confidence decoder receives the hidden state vector corresponding to each time step output by the encoder and, for each adjacent position, combines the hidden states corresponding to the preceding and following positions. A fully connected layer maps the combination of hidden state vectors from adjacent time steps to the interval between 0 and 1, outputting the intent transition confidence estimate for the corresponding adjacent position. The intent transition confidence value ranges from [0,1]. A higher value indicates that the transition occurs more frequently and is more typical in historical data; a lower value indicates that the transition is rarer or more anomalous.

[0091] For example, suppose a Transformer encoder is used as the sequence encoder, which includes a multi-layer multi-head self-attention mechanism and a feedforward network, enabling effective modeling of long-distance dependencies. For the input intent sequence [greeting, querying menu, placing an order, confirming address, ending conversation], the encoder maps it to 5 hidden state vectors [h1, h2, h3, h4, h5]. For the first adjacent position, the confidence decoder concatenates h1 and h2 and inputs them into a fully connected layer, outputting the confidence estimate of that transition; for the second adjacent position, h2 and h3 are concatenated and input into the fully connected layer, and so on, so that the confidence estimate at each position utilizes the contextual information of the entire sequence, improving the accuracy of the estimate.

[0092] Furthermore, using intent sequence fragments from the training sample set as input and positive and negative sample labels at corresponding adjacent positions as supervision signals, the intent transfer confidence model is trained end-to-end in a supervised manner using a binary cross-entropy loss function.

[0093] Specifically, during training, the training sample set inputs its intent sequence fragments into the intent transfer confidence model, which calculates the transfer confidence estimate for each adjacent position. For positions labeled as positive samples, the desired confidence value output by the intent transfer confidence model is as close to 1 as possible; for positions labeled as negative samples, the desired confidence value is as close to 0 as possible. A binary classification cross-entropy loss function is used to measure the difference between the predicted confidence and the true label. Where y is the true label, 1 represents a positive sample, 0 represents a negative sample, and T is the confidence estimate of the model output. The total loss is obtained by summing or averaging the loss values ​​over all training samples and all labeled locations. Then, the backpropagation algorithm is used to calculate the gradient of the loss with respect to each trainable parameter in the model, and the parameters are updated using an optimizer such as Adam. The training process continues for multiple iterations until the model's loss value on the training set tends to stabilize.

[0094] Furthermore, when the classification accuracy of the intent transfer confidence model on the independent validation set for positive and negative samples exceeds the accuracy threshold, training is stopped, and the pre-training of the intent transfer confidence model is completed.

[0095] Specifically, to prevent overfitting and ensure the model has good generalization ability, a portion of historical data not used in training should be reserved as an independent validation set during training. After each training epoch, samples from the validation set are input into the model, and the model's classification accuracy on the validation set is calculated. When the validation accuracy reaches a preset accuracy threshold, such as 95%, it indicates that the model has sufficiently strong discriminative ability, and training can be stopped.

[0096] Once the model has completed pre-training, it can be used in the actual inference stage. When the intent transfer confidence model receives any intent sequence, the sequence encoder encodes the temporal hidden state corresponding to each adjacent intent transfer pair in the intent sequence into a temporal hidden state sequence, which is then input into the confidence decoder. The confidence decoder calculates and outputs the transfer confidence estimate for each adjacent position, and arranges the intent transfer confidence estimates in the order of adjacent positions to form the intent transfer confidence sequence.

[0097] In this embodiment of the invention, the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set are obtained; for each intent tag, the probability of occurrence is calculated based on the additive smoothing relationship between the frequency of occurrence, the total length of the intent sequence, the total number of intents, and the frequency threshold; and the additive smoothing is fused with the historical frequency threshold to calculate the uncertainty coefficient, which reflects the actual intent call distribution of the current dialogue, and the probability of occurrence is corrected to avoid estimation bias caused by zero-frequency intents in small sample dialogues.

[0098] S300: Input the intent sequence into the intent transition confidence model and output the intent transition confidence sequence at each adjacent position, wherein the intent transition confidence model is constructed based on the intent transition frequency distribution in cross-scene historical dialogue data;

[0099] In this embodiment of the invention, the intent sequence is used as input data and fed into a pre-trained intent transition confidence model. The intent transition confidence model establishes a mapping relationship from the context of the intent sequence to the transition confidence of each adjacent position by learning the co-occurrence frequency of a large number of intent transition pairs in historical dialogue data. When the model receives the current intent sequence, it sequentially traverses each adjacent position in the sequence, extracts the preceding and succeeding intents corresponding to that position, and combines this with the contextual intent information surrounding that position. Using the trained encoder-decoder architecture, through the encoding-decoding calculation process within the intent transition confidence model, it outputs a corresponding transition confidence estimate for each adjacent position. Arranging the confidence estimates output at all adjacent positions in the order of their adjacent positions constitutes the intent transition confidence sequence. Each confidence value in this sequence reflects the typicality or credibility of the corresponding local transition in historical experience.

[0100] In this embodiment of the invention, the intent transfer confidence model constructed based on historical frequency distribution obtains the intent transfer confidence at each adjacent transfer position using a sequence-level encoding method, providing a data basis for subsequent correction of temporal dependencies.

[0101] S400: The difference between the uncertainty coefficient and the conditional entropy of adjacent intentions is used as the initial measure of temporal dependence, and the initial measure of temporal dependence is weighted and corrected using the intention transfer confidence sequence to obtain the temporal dependence correction measure.

[0102] In this embodiment of the invention, the adjacent intention conditional entropy refers to the information entropy corresponding to the conditional probability distribution of the subsequent intention under the condition of knowing the preceding intention; the initial measure of temporal dependency refers to the measure constructed based on the difference between the uncertainty coefficient and the adjacent intention conditional entropy.

[0103] Specifically, based on the frequency statistics of each intent transition pair in historical dialogue data, the conditional probability distribution of each subsequent intent under the condition of the preceding intent is determined, and the conditional entropy of adjacent intents is calculated based on the conditional probability distribution. Subsequently, the uncertainty coefficient is subtracted from the conditional entropy of the adjacent intent as the initial measure of temporal dependency. All confidence values ​​in the intent transition confidence sequence are extracted, and the arithmetic mean of the confidence values ​​is calculated as the correction weight. Finally, the initial measure of temporal dependency is multiplied by the correction weight to obtain the corrected measure of temporal dependency.

[0104] Step S400 in the method of this embodiment of the invention includes:

[0105] Based on the frequency statistics of each intent transition pair in the historical dialogue data, the conditional probability distribution of the subsequent intent under each preceding intent is determined, and the conditional entropy of adjacent intents is calculated using the conditional probability distribution.

[0106] Based on the uncertainty coefficient and the adjacent intention conditional entropy, determine the initial measure of temporal dependency;

[0107] Extract the confidence value at each adjacent position in the intent transfer confidence sequence, and calculate the arithmetic mean of all confidence values ​​as the correction weight;

[0108] The initial time-series dependency metric is multiplied by the corrected weight, and the product is output as the corrected time-series dependency metric.

[0109] In this embodiment of the invention, firstly, all intent sequences in the historical dialogue data are traversed, and based on the frequency statistics of each intent transition pair in the historical dialogue data, the conditional probability distribution of subsequent intents under each preceding intent is determined, and the conditional entropy of adjacent intents is calculated.

[0110] Secondly, based on the uncertainty coefficient and the conditional entropy of adjacent intentions, an initial measure of temporal dependency is determined: Initial measure of temporal dependency = Uncertainty coefficient - Conditional entropy of adjacent intentions. This initial measure reflects the difference between the marginal uncertainty of the entire intention sequence and the conditional uncertainty under the known previous intention. A larger difference indicates a stronger temporal dependency in the current dialogue's intention sequence, meaning a stronger regularity in dialogue behavior; a smaller difference indicates a lack of stable transition patterns between intentions, and a tendency towards random or chaotic behavior.

[0111] Next, the confidence value at each adjacent position in the intent transition confidence sequence is extracted, and the arithmetic mean of all confidence values ​​is calculated as the adjustment weight. The average confidence reflects the typicality of the transitions in the current dialogue at the overall level. A high average confidence indicates that the transition patterns in the current dialogue mostly conform to typical patterns in historical experience, and the estimation of its temporal dependencies is relatively reliable. A low average confidence indicates that there are many atypical transitions in the current dialogue, which may be caused by scene changes, individual differences, or annotation noise, and in this case, downward adjustment is necessary.

[0112] Finally, the initial time dependency metric is multiplied by the adjusted weights, and the product is output as the adjusted time dependency metric. Wherein, the adjusted time dependency metric = initial time dependency metric × adjusted weights.

[0113] If the correction weight is high, the initial metric is largely preserved; if the correction weight is low, the initial metric is significantly reduced, indicating that the transition pattern in the dialogue deviates from the typical pattern, and some signals in the initial metric may originate from noise, and should not be assigned too high a confidence level.

[0114] In the method of this embodiment of the invention, based on the frequency statistics of each intent transition pair in the historical dialogue data, the conditional probability distribution of the subsequent intent under each preceding intent condition is determined, and the conditional entropy of adjacent intents is calculated using the conditional probability distribution, including:

[0115] Traverse all intent sequences in the historical dialogue data, count the frequency of occurrence of each subsequent intent under each preceding intent condition, and construct a condition frequency matrix with all intents as rows and all intents as columns.

[0116] Obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set;

[0117] Additive smoothing is performed on the occurrence frequency of each condition in the conditional frequency matrix. The smoothed conditional frequency is divided by the total smoothed frequency of the row to obtain the conditional probability value of each subsequent intention under each preceding intention, thus forming a conditional probability matrix.

[0118] For each row in the conditional probability matrix, multiply all conditional probability values ​​in the row by the logarithm of the corresponding conditional probability value, sum all the products and take the opposite number to obtain the single-intent conditional entropy of the corresponding preceding intent;

[0119] The total frequency of each preceding intent is counted from the historical dialogue data, and the total frequency is divided by the sum of the total frequencies of all preceding intents in the historical dialogue data to obtain the marginal probability of each preceding intent.

[0120] The marginal probabilities of all preceding intentions are summed with the corresponding single intention conditional entropy, and the summation result is output as the adjacent intention conditional entropy.

[0121] In this embodiment of the invention, during actual dialogue, the transfer between intents forms a directed graph structure, where nodes are intent labels and directed edges represent the transfer relationship from a preceding intent to a subsequent intent. To quantify the bias of dialogue transfers, it is necessary to count the frequency of occurrence of all adjacent intent transfer pairs in the current intent sequence.

[0122] First, iterate through all intent sequences in the historical dialogue data, count the frequency of occurrence of each subsequent intent under each preceding intent condition, and construct a condition frequency matrix with all intents as rows and all intents as columns.

[0123] Specifically, it iterates through all intent sequences in the historical dialogue data, using all intent tags in the scene intent tag set as row and column indices, and initializes a zero-based square matrix with the number of rows and columns equal to the total number of intents as the initial form of the conditional frequency matrix; it then scans each pair of adjacent intent transitions in each intent sequence, using the preceding intent as the row coordinate and the subsequent intent as the column coordinate, and increments the count value at the corresponding position in the conditional frequency matrix by one unit; after completing the traversal of all historical intent sequences, the element value in the i-th row and j-th column of the conditional frequency matrix represents the frequency of occurrence of the condition when the i-th intent tag is the preceding intent and the j-th intent tag is the subsequent intent in all historical dialogue data.

[0124] Secondly, obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set.

[0125] Secondly, the size of the scene intent tag set is N. Additive smoothing is performed on the frequency of occurrence of each condition in the conditional frequency matrix. Specifically, for each element C in the conditional frequency matrix C... ij Its smoothed frequency is C' ij =C ij +(1 / R), where the smoothing parameter is 1 / R and R is the frequency threshold. Divide the smoothed conditional frequency by the total smoothed frequency of the row to obtain the conditional probability value of each subsequent intention under each preceding intention. Specifically, for each row i, calculate the total smoothed frequency of that row = S + [(1 / R) × N], where S is the sum of the original frequencies of that row. Finally, divide the smoothed frequency by the total row frequency to obtain the conditional probability p. ij =[C ij +(1 / R)] / [S+(1 / R)×N]. Performing this calculation on all rows of the matrix yields an N×N conditional probability matrix, where the sum of the probabilities in each row equals 1.

[0126] Furthermore, for each row of the conditional probability matrix, multiply all conditional probability values ​​in the row by the logarithm of the corresponding conditional probability value, sum all products, and take the negative of the sum to obtain the single-intent conditional entropy for the corresponding preceding intent. Single-intent conditional entropy p ij Let N be the conditional probability and N be the size of the scene intent tag set. The logarithm is usually calculated using the natural logarithm e or base 2, and the natural logarithm can be used. A larger entropy value indicates a more dispersed distribution of subsequent intents under a given preceding intent, with higher uncertainty, meaning a weaker predictive ability of the preceding intent for the following intent; a smaller entropy value indicates a more concentrated distribution of subsequent intents, with lower uncertainty, meaning a stronger predictive ability of the preceding intent for the following intent.

[0127] Furthermore, the total frequency of each preceding intent is calculated from the historical dialogue data. This total frequency is then divided by the sum of the total frequencies of all preceding intents in the historical dialogue data to obtain the marginal probability of each preceding intent. Marginal probability = total frequency / sum of the total frequencies of all preceding intents, reflecting the natural frequency of that preceding intent in the entire dialogue system.

[0128] Finally, the products of the marginal probabilities of all preceding intentions and the corresponding single-intention conditional entropies are summed, and the sum is output as the conditional entropy of adjacent intentions. Adjacent Intention Conditional Entropy , where H c For the single-intention conditional entropy, P e denoted as marginal probability, and N as the size of the scene intent label set.

[0129] In this embodiment of the invention, the difference between the uncertainty coefficient and the conditional entropy of adjacent intentions is used as the initial measure of temporal dependency. The arithmetic mean of the confidence of all intentions in the intentions transfer confidence sequence is used as the correction weight to perform a weighted correction on the initial measure of temporal dependency. This effectively suppresses the interference of low-confidence abnormal transfers on the temporal dependency assessment and improves the robustness of the measurement results.

[0130] S500: Based on the frequency distribution of directed intention transition pairs in the intention sequence, determine the stiffness metric of the state transition skeleton, and use the distribution characteristics of the response delay to perform a secondary correction on the stiffness metric to obtain the topological stiffness correction metric.

[0131] In this embodiment of the invention, a directed intent transition pair refers to an irreversible transition pair consisting of two adjacent rounds in an intent sequence; a state transition skeleton is a topological graph consisting of all intent tags in the scene intent tag set as nodes and the directed transition relationships between intents as edges; stiffness metric is an indicator that quantifies the structural stiffness of the state transition skeleton; and the distribution characteristics of response delay are the statistical characteristics presented by the response delay sequence.

[0132] Specifically, the process iterates through all adjacent intent transition pairs in the intent sequence, counts the frequency of each directed intent transition pair, and constructs a directed transition frequency matrix with all intent tags in the scene intent tag set as row and column indices. The total number of intents in the scene intent tag set and their corresponding frequency thresholds are obtained, and additive smoothing is applied to each occurrence frequency in the frequency matrix. After smoothing, the transition probability from each preceding intent to each subsequent intent is obtained, forming a directed transition probability matrix. The sum of squares of all transition probability values ​​in this directed transition probability matrix is ​​calculated, and combined with the total number of intents, the average transition concentration is determined. All response delay values ​​in the response delay sequence are extracted. Finally, a topological stiffness correction metric is determined based on the initial stiffness metric and the delay variation coefficient.

[0133] Step S500 in the method of this embodiment of the invention includes:

[0134] Traverse all adjacent intent transition pairs in the intent sequence, count the frequency of occurrence of each directed intent transition pair, and construct a directed transition frequency matrix with all intent tags in the scene intent tag set as row and column indices;

[0135] Obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set, and perform additive smoothing on each occurrence frequency in the directed transition frequency matrix;

[0136] Divide the smoothed frequency by the total smoothed frequency of the row to obtain the transition probability from each preceding intention to the subsequent intention, forming a directed transition probability matrix.

[0137] Calculate the sum of squares of all transition probability values ​​in the directed transition probability matrix, and combine it with the total number of intentions to determine the average transition concentration, which serves as the initial stiffness measure of the state transition skeleton.

[0138] Extract all response delay values ​​from the response delay sequence, calculate the mean response delay and the standard deviation of the response delay, and use the ratio of the standard deviation of the response delay to the mean response delay as the delay variation coefficient;

[0139] Based on the initial stiffness metric and the delay variation coefficient, a topological stiffness correction metric is determined, wherein the topological stiffness correction metric is positively correlated with the initial stiffness metric and negatively correlated with the delay variation coefficient.

[0140] In this embodiment of the invention, firstly, all adjacent intention transition pairs in the intention sequence are traversed, the frequency of occurrence of each directed intention transition pair is counted, and a directed transition frequency matrix is ​​constructed with all intention tags in the scene intention tag set as row and column indices.

[0141] Secondly, iterate through all adjacent positions in the current intent sequence, obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set, and perform additive smoothing on each occurrence frequency in the directed transition frequency matrix.

[0142] Specifically, suppose the scene intent tag set contains M intent tags, arranged in a fixed order, such as alphabetical or business logic order. Initialize an M×M all-zero matrix, where the row index represents the preceding intent and the column index represents the succeeding intent. Traverse all adjacent positions in the current intent sequence to form directed transition pairs, and increment the corresponding row and column element values ​​in the matrix by 1. After traversing the directed transition frequency matrix, each element represents the number of times it appears in the current intent sequence.

[0143] Next, the smoothing parameter is taken as the reciprocal of the frequency threshold, adding the smoothing parameter to the original frequency. The smoothed frequency is divided by the total smoothed frequency of the row to obtain the transition probability from each preceding intention to the subsequent intention: Transition probability = Smoothed frequency / Total smoothed frequency of the row. Performing the same calculation on all rows yields the directed transition probability matrix, where the sum of the probabilities in each row equals 1.

[0144] Furthermore, the sum of squares of all transition probability values ​​in the directed transition probability matrix is ​​calculated, and the sum of squares is divided by the total number of intentions M to obtain the average transition concentration, which serves as the initial stiffness measure of the state transition skeleton. When the transition probabilities are completely concentrated on the same subsequent intention, the average transition concentration is 1; when the transition probabilities are evenly distributed across all M×M transition pairs, the average transition concentration is 1 / M. The range of the average transition concentration is [1 / M, 1], with a larger value indicating a more concentrated distribution and stronger stiffness. The closer the initial stiffness measure is to 1, the more rigid the intention transition skeleton of the entire dialogue system is, and the stronger the behavioral regularity; the closer the initial stiffness measure is to 1 / M, the looser the transition skeleton is, and the more random the behavior.

[0145] Furthermore, all response delay values ​​in the response delay sequence are extracted, and the mean and standard deviation of the response delay are calculated. The standard deviation is the square root of the sum of squared deviations of each response delay value from its arithmetic mean. The ratio of the standard deviation to the mean of the response delay is used as the coefficient of variation. A smaller coefficient of variation indicates a more stable response delay and a more regular time rhythm; a larger coefficient of variation indicates more drastic fluctuations in the response delay and a more chaotic time rhythm.

[0146] Finally, based on the initial stiffness metric and the lag coefficient of variation, the topology stiffness correction metric is determined. Topology stiffness correction metric = Initial stiffness metric / (1 + Lag coefficient of variation). The topology stiffness correction metric is positively correlated with the initial stiffness metric and negatively correlated with the lag coefficient of variation; that is, an increase in the lag coefficient of variation leads to a decrease in the topology stiffness correction metric, and a decrease in the lag coefficient of variation leads to an increase in the topology stiffness correction metric.

[0147] In this embodiment of the invention, the structural rigidity of the dialogue state transition skeleton is quantified from the concentration of the transition probability matrix based on the frequency distribution of directed intent transition pairs in the intent sequence; the response delay variation coefficient is obtained by using the distribution map features of the response delay, and the initial stiffness measure is corrected a second time to obtain the topological stiffness correction measure, thus integrating information from the semantic and temporal levels.

[0148] S600: Integrate the temporal dependency correction metric and the topological stiffness correction metric to output a comprehensive metric index, thereby quantifying the intensity of the regularity of voice dialogue behavior in the target scenario.

[0149] In this embodiment of the invention, the comprehensive metric index is the final output quantified value, and its magnitude characterizes the strength of the behavioral regularity of the current voice dialogue stream in the target scenario.

[0150] Specifically, the temporal dependency correction metric and the topological stiffness correction metric are obtained, multiplied and fused together. After the fusion calculation is completed, the comprehensive metric index is used as the output of the quantitative result of the regularity intensity of voice dialogue behavior in the final target scenario.

[0151] In the method of this embodiment of the invention, after fusing the time-dependency correction metric and the topological stiffness correction metric to output a comprehensive metric index, the method further includes:

[0152] The comprehensive metric indexes output by multiple dialogue streams at different time periods under the same target scenario are obtained and a comprehensive metric index set is formed.

[0153] Calculate the arithmetic mean of all comprehensive measurement indices in the comprehensive measurement index set to obtain the comprehensive measurement benchmark value;

[0154] Calculate the difference between each comprehensive metric index in the comprehensive metric index set and the comprehensive metric benchmark value, sum the absolute values ​​of each difference and divide by the total number of comprehensive metric indices in the comprehensive metric index set to obtain the comprehensive metric dispersion.

[0155] The overall metric dispersion is output as a cross-dialogue consistency metric for the regularity of dialogue behavior in the target scenario.

[0156] In this embodiment of the invention, firstly, the comprehensive metric indexes output by multiple dialogue streams corresponding to multiple different time periods within the same target scenario are obtained to quantify the behavioral regularity of the dialogue. A comprehensive metric index set is constructed by collecting the comprehensive metric indices corresponding to dialogue streams generated at multiple different time periods. The collection strategy can be flexibly set according to actual needs. For example, several dialogue streams generated at fixed time periods each day for seven consecutive days within the same scenario can be collected; alternatively, dialogue streams from different dates within the same scenario can be collected. After processing, the comprehensive metric index set is obtained by arranging them in chronological order.

[0157] Secondly, the arithmetic mean of all comprehensive metric indices in the comprehensive metric index set is calculated to obtain the comprehensive metric benchmark value, which is used to reflect the average intensity level of the regularity of dialogue behavior in the target scenario.

[0158] Next, the mean absolute deviation is used as the metric for overall dispersion. The difference between each overall metric index in the overall metric index set and the overall metric benchmark value is calculated. Then, the absolute values ​​of each difference are summed and divided by the total number of overall metric indices in the overall metric index set to obtain the overall metric dispersion. The smaller the overall metric dispersion, the more consistent the regularity among the dialogue flows; the larger the overall metric dispersion, the greater the difference in regularity among the dialogue flows.

[0159] Finally, the overall metric dispersion is output as a cross-dialogue consistency metric for the regularity of dialogue behavior in the target scenario. The smaller the overall metric dispersion, the closer the regularity assessment results of each dialogue stream in different time periods in the target scenario are to the consistency, and the dialogue behavior pattern in the target scenario has high repeatability and predictability in different time windows; the larger the overall metric dispersion, the more significant the fluctuation in the regularity assessment results of each dialogue stream in different time periods in the target scenario, and the greater the difference in the dialogue behavior pattern in the target scenario in different time windows.

[0160] In this embodiment of the invention, the dimension reflecting the temporal constraint strength of intent flow and the dimension reflecting the rigidity of the state transition skeleton structure are fused to generate a comprehensive metric index, thereby completing the multi-perspective and multi-level quantification of the regularity of voice dialogue behavior in the target scenario and improving the interpretability and comparability of the metric results.

[0161] Through the above specific implementation methods, the embodiments of the present invention achieve the following technical effects:

[0162] In this embodiment of the invention, firstly, continuous audio streams are accurately segmented through speech activity detection, speaker separation, and fusion threshold merging, avoiding over-segmentation due to short pauses or under-segmentation due to improper silence threshold settings. Then, intent tags for each round are obtained through semantic matching, transforming high-dimensional speech signals into low-dimensional discrete symbol sequences, reducing subsequent computational complexity. Response delay provides a data foundation for evaluating the regularity of fused interaction rhythm information, ensuring the integrity of link information and temporal accuracy. Subsequently, the final intent recognition result is obtained through text normalization, frequency thresholding, synonym set matching, longest consecutive hit analysis, and prior probability correction, making the final intent recognition result more adaptable to open scenarios.

[0163] Secondly, obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set; for each intent tag, calculate the fused occurrence probability based on the additive smoothing relationship between the occurrence frequency, the total length of the intent sequence, the total number of intents, and the frequency threshold; and fuse the additive smoothing with the historical frequency threshold to calculate the uncertainty coefficient, which reflects the actual intent call distribution of the current dialogue, and corrects the occurrence probability to avoid estimation bias caused by zero-frequency intents in small sample dialogues.

[0164] Furthermore, the intent transfer confidence model, constructed based on historical frequency distribution, obtains the intent transfer confidence at each adjacent transfer position using sequence-level encoding, providing a data foundation for subsequent correction of temporal dependencies.

[0165] Furthermore, the difference between the uncertainty coefficient and the conditional entropy of adjacent intentions is used as the initial measure of temporal dependency. The arithmetic mean of the confidence of all intention transfers in the intention transfer confidence sequence is used as the correction weight to perform a weighted correction on the initial measure of temporal dependency. This effectively suppresses the interference of low-confidence abnormal transfers on the assessment of temporal dependency and improves the robustness of the measurement results.

[0166] Furthermore, based on the frequency distribution of directed intent transition pairs in the intent sequence, the structural rigidity of the dialogue state transition skeleton is quantified from the concentration of the transition probability matrix; the response delay variation coefficient is obtained by using the distribution map features of the response delay, and the initial stiffness measure is corrected twice to obtain the topological stiffness correction measure, thus integrating information from the semantic and temporal levels.

[0167] Finally, the dimension reflecting the strength of the temporal constraints of intent flow and the dimension reflecting the rigidity of the state transition skeleton structure are merged to generate a comprehensive metric index. This completes the multi-perspective and multi-level quantification of the regularity of voice dialogue behavior in the target scenario, improving the interpretability and comparability of the metric results.

[0168] Example 2, as Figure 3 As shown, based on the same inventive concept as the method for quantifying cross-scenario voice dialogue behavior patterns provided in Embodiment 1, this embodiment of the invention also provides a system for quantifying cross-scenario voice dialogue behavior patterns, the system comprising:

[0169] The intent sequence construction module 11 is used to acquire the voice dialogue stream in the target scene, parse the round information and the intent tag corresponding to each round in the voice dialogue stream, construct the intent sequence with round as the time unit, and record the response delay between adjacent rounds.

[0170] The uncertainty coefficient acquisition module 12 is used to calculate the information entropy based on the probability of occurrence of different intentions in the intention sequence, use the information entropy as the quantification value of the uncertainty of the dialogue intention call, and output the uncertainty coefficient.

[0171] The intent transfer confidence model construction module 13 is used to input the intent sequence into the intent transfer confidence model and output the intent transfer confidence sequence at each adjacent position, wherein the intent transfer confidence model is constructed based on the intent transfer frequency distribution in cross-scene historical dialogue data;

[0172] The weighted correction module 14 is used to use the difference between the uncertainty coefficient and the adjacent intention conditional entropy as the initial measure of temporal dependency, and to use the intention transfer confidence sequence to weight and correct the initial measure of temporal dependency to obtain the temporal dependency correction measure.

[0173] The secondary correction module 15 is used to determine the stiffness metric of the state transition skeleton based on the frequency distribution of the directed intention transition pairs in the intention sequence, and to perform secondary correction on the stiffness metric using the distribution characteristics of the response delay to obtain the topological stiffness correction metric.

[0174] The comprehensive metric index fusion module 16 is used to fuse the temporal dependency correction metric and the topological stiffness correction metric, and output a comprehensive metric index to quantify the intensity of the regularity of voice dialogue behavior in the target scenario.

[0175] In one embodiment, an intent sequence module 11 is constructed for:

[0176] Real-time capture of voice dialogue streams using voice acquisition devices deployed within the target scenario;

[0177] The speech dialogue stream is subjected to speech activity detection and speaker separation to obtain multiple speech activity segments arranged in time sequence and the time boundary corresponding to each speech activity segment;

[0178] Adjacent speech activity segments belonging to the same speaker and with an interval duration less than the fusion threshold are merged into independent rounds, and each independent round is assigned a round number, a start timestamp, and an end timestamp.

[0179] For the speech signal in each independent round, a text sequence is obtained through speech recognition. The text sequence is then semantically matched with the scene intent tag set to output the intent tag corresponding to the independent round.

[0180] Arrange the intent tags corresponding to each independent round in ascending order of round number to construct an intent sequence with round as the time unit;

[0181] For any two adjacent independent rounds, subtract the end timestamp of the previous round from the start timestamp of the later round to obtain the response delay between adjacent rounds, and store the response delays between all adjacent rounds in sequence according to the round number as a response delay sequence.

[0182] Specifically, for the speech signal within each independent round, a text sequence is obtained through speech recognition. This text sequence is then semantically matched with a set of scene intent tags to output the intent tag corresponding to that independent round, including:

[0183] The text sequence obtained from the speech recognition is subjected to text normalization processing to obtain normalized text, and the normalized text is segmented into words, stop words are removed, and a content word sequence is obtained.

[0184] For each intent tag in the scene intent tag set, obtain a set of synonym expressions associated with the intent tag. The scene intent tag set is constructed by deduplicating and summarizing all the labeled intent tags in the historical dialogue data and retaining the intent tags that appear more than a frequency threshold in the current target scene.

[0185] The step of determining the frequency threshold includes:

[0186] The frequency of each candidate intent tag in the historical dialogue data is counted, and the mean and standard deviation of the frequency of the tag occurrences are extracted.

[0187] The frequency threshold is the sum of the frequency mean and the frequency standard deviation.

[0188] Calculate the word hit rate between the content word sequence and the synonym expression word set, as well as the longest consecutive hit length of the hit words in the content word sequence. Then, sum the normalized values ​​of the word hit rate and the longest consecutive hit length in a weighted manner to obtain the semantic matching base score.

[0189] The probability of occurrence of the intent tag is statistically analyzed from the historical dialogue data, and combined with the semantic matching base score to obtain the prior corrected matching score;

[0190] Compare the prior corrected matching scores of all intent tags in the scene intent tag set, and select the intent tag corresponding to the largest prior corrected matching score as the intent tag output for the independent round.

[0191] In one embodiment, the uncertainty coefficient acquisition module 12 is used for:

[0192] Traverse the intent sequence, count the frequency of each intent tag in the scene intent tag set in the intent sequence, and determine the total length of the intent sequence;

[0193] Obtain the total number of intents in the scene intent tag set, and the frequency threshold corresponding to the scene intent tag set;

[0194] For each intent tag, the probability of fusion occurrence is calculated based on the additive smoothing relationship between the occurrence frequency, the total length of the intent sequence, the total number of intents, and the frequency threshold.

[0195] The information entropy is calculated based on the fusion occurrence probability of all intent labels, and the calculation result is output as the uncertainty coefficient.

[0196] The training steps of the intent transfer confidence model include:

[0197] Extract multiple intent sequences corresponding to dialogues from historical dialogue data across different scenarios, traverse adjacent intent transition pairs in all intent sequences, and count the original transition frequency of each intent transition pair.

[0198] The training sample set is formed by using the transfer pairs whose original transfer frequency is higher than the frequency threshold as positive samples and the transfer pairs whose original transfer frequency is zero as negative samples.

[0199] Construct an intent transition confidence model based on a sequence-to-sequence confidence estimation architecture, wherein the architecture includes a sequence encoder for encoding an intent sequence into a temporal hidden state sequence, and a confidence decoder for decoding the temporal hidden state sequence into transition confidence at each adjacent position.

[0200] Using the intent sequence fragments in the training sample set as input and the positive and negative sample labels at corresponding adjacent positions as supervision signals, the intent transfer confidence model is trained end-to-end in a supervised manner using a binary classification cross-entropy loss function.

[0201] When the classification accuracy of the intent transfer confidence model on the independent validation set for positive and negative samples exceeds the accuracy threshold, training is stopped, and the pre-training of the intent transfer confidence model is completed.

[0202] When the intent transfer confidence model receives any intent sequence, it sequentially inputs the temporal hidden state combination corresponding to each adjacent intent transfer pair in the intent sequence into the confidence decoder, and arranges the intent transfer confidence estimates output by the confidence decoder in the order of adjacent positions to form an intent transfer confidence sequence.

[0203] In one embodiment, the weighted correction module 14 is used for:

[0204] Based on the frequency statistics of each intent transition pair in the historical dialogue data, the conditional probability distribution of the subsequent intent under each preceding intent is determined, and the conditional entropy of adjacent intents is calculated using the conditional probability distribution.

[0205] Based on the uncertainty coefficient and the adjacent intention conditional entropy, determine the initial measure of temporal dependency;

[0206] Extract the confidence value at each adjacent position in the intent transfer confidence sequence, and calculate the arithmetic mean of all confidence values ​​as the correction weight;

[0207] The initial time-series dependency metric is multiplied by the corrected weight, and the product is output as the corrected time-series dependency metric.

[0208] Specifically, based on the frequency statistics of each intent transition pair in the historical dialogue data, the conditional probability distribution of subsequent intents under each preceding intent is determined, and the conditional entropy of adjacent intents is calculated using the conditional probability distribution, including:

[0209] Traverse all intent sequences in the historical dialogue data, count the frequency of occurrence of each subsequent intent under each preceding intent condition, and construct a condition frequency matrix with all intents as rows and all intents as columns.

[0210] Obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set;

[0211] Additive smoothing is performed on the occurrence frequency of each condition in the conditional frequency matrix. The smoothed conditional frequency is divided by the total smoothed frequency of the row to obtain the conditional probability value of each subsequent intention under each preceding intention, thus forming a conditional probability matrix.

[0212] For each row in the conditional probability matrix, multiply all conditional probability values ​​in the row by the logarithm of the corresponding conditional probability value, sum all the products and take the opposite number to obtain the single-intent conditional entropy of the corresponding preceding intent;

[0213] The total frequency of each preceding intent is counted from the historical dialogue data, and the total frequency is divided by the sum of the total frequencies of all preceding intents in the historical dialogue data to obtain the marginal probability of each preceding intent.

[0214] The marginal probabilities of all preceding intentions are summed with the corresponding single intention conditional entropy, and the summation result is output as the adjacent intention conditional entropy.

[0215] In one embodiment, the secondary correction module 15 is used for:

[0216] Traverse all adjacent intent transition pairs in the intent sequence, count the frequency of occurrence of each directed intent transition pair, and construct a directed transition frequency matrix with all intent tags in the scene intent tag set as row and column indices;

[0217] Obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set, and perform additive smoothing on each occurrence frequency in the directed transition frequency matrix;

[0218] Divide the smoothed frequency by the total smoothed frequency of the row to obtain the transition probability from each preceding intention to the subsequent intention, forming a directed transition probability matrix.

[0219] Calculate the sum of squares of all transition probability values ​​in the directed transition probability matrix, and combine it with the total number of intentions to determine the average transition concentration, which serves as the initial stiffness measure of the state transition skeleton.

[0220] Extract all response delay values ​​from the response delay sequence, calculate the mean response delay and the standard deviation of the response delay, and use the ratio of the standard deviation of the response delay to the mean response delay as the delay variation coefficient;

[0221] Based on the initial stiffness metric and the delay variation coefficient, a topological stiffness correction metric is determined, wherein the topological stiffness correction metric is positively correlated with the initial stiffness metric and negatively correlated with the delay variation coefficient.

[0222] In one embodiment, the comprehensive metric index fusion module 16, after fusing the time-series dependency correction metric and the topological stiffness correction metric and outputting the comprehensive metric index, further includes:

[0223] The comprehensive metric indexes output by multiple dialogue streams at different time periods under the same target scenario are obtained and a comprehensive metric index set is formed.

[0224] Calculate the arithmetic mean of all comprehensive measurement indices in the comprehensive measurement index set to obtain the comprehensive measurement benchmark value;

[0225] Calculate the difference between each comprehensive metric index in the comprehensive metric index set and the comprehensive metric benchmark value, sum the absolute values ​​of each difference and divide by the total number of comprehensive metric indices in the comprehensive metric index set to obtain the comprehensive metric dispersion.

[0226] The overall metric dispersion is output as a cross-dialogue consistency metric for the regularity of dialogue behavior in the target scenario.

[0227] Compared to existing technologies, this approach first reduces subsequent computational complexity by transforming the original unstructured speech dialogue stream into a structured sequence of temporal symbols and synchronous temporal features, thus obtaining intent recognition results. Secondly, it characterizes the dispersion or concentration of various intent calls within the dialogue system from a macro-statistical perspective of intent distribution, avoiding the loss of global distribution information due to reliance solely on local transition patterns. Thirdly, it constructs an intent transition confidence model to obtain intent transition confidence, overcoming the inability to distinguish reliability differences between different transition positions. Furthermore, it utilizes the uncertainty coefficient and adjacent intent conditional entropy to obtain an initial measure of temporal dependency, and uses the arithmetic mean of intent transition confidence as a correction weight for weighted correction, suppressing the interference of abnormal transitions on temporal dependency evaluation. Further, it quantifies the structural rigidity of the dialogue state transition skeleton, obtains the response delay variation coefficient, and performs a secondary correction on the initial stiffness measure to obtain a topological stiffness correction measure. Finally, it integrates the temporal constraint strength of intent flow and the structural rigidity dimension of the state transition skeleton to generate a comprehensive metric index, quantifying the regularity strength of speech dialogue behavior and improving the interpretability and comparability of the measurement results.

Claims

1. A quantitative measurement method for cross-scenario voice dialogue behavior patterns, characterized in that, The method includes: Acquire the voice dialogue stream in the target scenario, parse the round information and the intent tag corresponding to each round in the voice dialogue stream, construct the intent sequence with round as the time unit, and record the response delay between adjacent rounds; The information entropy is calculated based on the probability of occurrence of different intentions in the intention sequence. The information entropy is used as a quantification value of the uncertainty of dialogue intention invocation, and the uncertainty coefficient is output. The intent sequence is input into the intent transition confidence model, and the intent transition confidence sequence at each adjacent position is output. The intent transition confidence model is constructed based on the frequency distribution of intent transitions in historical dialogue data across scenarios. The difference between the uncertainty coefficient and the conditional entropy of adjacent intentions is used as the initial measure of temporal dependence, and the initial measure of temporal dependence is weighted and corrected using the intention transfer confidence sequence to obtain the corrected measure of temporal dependence. Based on the frequency distribution of directed intention transition pairs in the intention sequence, the stiffness metric of the state transition skeleton is determined, and the stiffness metric is further corrected using the distribution characteristics of the response delay to obtain the topological stiffness correction metric. By integrating the temporal dependency correction metric and the topological stiffness correction metric, a comprehensive metric index is output to quantify the intensity of the regularity of voice dialogue behavior in the target scenario.

2. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 1, characterized in that, Acquire the voice dialogue stream in the target scenario, parse the round information and the intent label corresponding to each round in the voice dialogue stream, construct the intent sequence with rounds as the time unit, and record the response delay between adjacent rounds, including: Real-time capture of voice dialogue streams using voice acquisition devices deployed within the target scenario; The speech dialogue stream is subjected to speech activity detection and speaker separation to obtain multiple speech activity segments arranged in time sequence and the time boundary corresponding to each speech activity segment; Adjacent speech activity segments belonging to the same speaker and with an interval duration less than the fusion threshold are merged into independent rounds, and each independent round is assigned a round number, a start timestamp, and an end timestamp. For the speech signal in each independent round, a text sequence is obtained through speech recognition. The text sequence is then semantically matched with the scene intent tag set to output the intent tag corresponding to the independent round. Arrange the intent tags corresponding to each independent round in ascending order of round number to construct an intent sequence with round as the time unit; For any two adjacent independent rounds, subtract the end timestamp of the previous round from the start timestamp of the later round to obtain the response delay between adjacent rounds, and store the response delays between all adjacent rounds in sequence according to the round number as a response delay sequence.

3. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 2, characterized in that, For the speech signal within each independent round, a text sequence is obtained through speech recognition. The text sequence is then semantically matched with a set of scene intent tags to output the intent tag corresponding to that independent round, including: The text sequence obtained from the speech recognition is subjected to text normalization processing to obtain normalized text, and the normalized text is segmented into words, stop words are removed, and a content word sequence is obtained. For each intent tag in the scene intent tag set, obtain a set of synonym expressions associated with the intent tag. The scene intent tag set is constructed by deduplicating and summarizing all the labeled intent tags in the historical dialogue data and retaining the intent tags that appear more than a frequency threshold in the current target scene. The step of determining the frequency threshold includes: The frequency of each candidate intent tag in the historical dialogue data is counted, and the mean and standard deviation of the frequency of the tag occurrences are extracted. The frequency threshold is the sum of the frequency mean and the frequency standard deviation. Calculate the word hit rate between the content word sequence and the synonym expression word set, as well as the longest consecutive hit length of the hit words in the content word sequence. Then, sum the normalized values ​​of the word hit rate and the longest consecutive hit length in a weighted manner to obtain the semantic matching base score. The probability of occurrence of the intent tag is statistically analyzed from the historical dialogue data, and combined with the semantic matching base score to obtain the prior corrected matching score; Compare the prior corrected matching scores of all intent tags in the scene intent tag set, and select the intent tag corresponding to the largest prior corrected matching score as the intent tag output for the independent round.

4. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 1, characterized in that, Information entropy is calculated based on the probability of occurrence of different intentions in the intent sequence. This information entropy is then used as a quantification of the uncertainty of the dialogue intent invocation, and an uncertainty coefficient is output, including: Traverse the intent sequence, count the frequency of each intent tag in the scene intent tag set in the intent sequence, and determine the total length of the intent sequence; Obtain the total number of intents in the scene intent tag set, and the frequency threshold corresponding to the scene intent tag set; For each intent tag, the probability of fusion occurrence is calculated based on the additive smoothing relationship between the occurrence frequency, the total length of the intent sequence, the total number of intents, and the frequency threshold. The information entropy is calculated based on the fusion occurrence probability of all intent labels, and the calculation result is output as the uncertainty coefficient.

5. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 1, characterized in that, The training steps for the intent transfer confidence model include: Extract multiple intent sequences corresponding to dialogues from historical dialogue data across different scenarios, traverse adjacent intent transition pairs in all intent sequences, and count the original transition frequency of each intent transition pair. The training sample set is formed by using the transfer pairs whose original transfer frequency is higher than the frequency threshold as positive samples and the transfer pairs whose original transfer frequency is zero as negative samples. Construct an intent transition confidence model based on a sequence-to-sequence confidence estimation architecture, wherein the architecture includes a sequence encoder for encoding an intent sequence into a temporal hidden state sequence, and a confidence decoder for decoding the temporal hidden state sequence into transition confidence at each adjacent position. Using the intent sequence fragments in the training sample set as input and the positive and negative sample labels at corresponding adjacent positions as supervision signals, the intent transfer confidence model is trained end-to-end in a supervised manner using a binary classification cross-entropy loss function. When the classification accuracy of the intent transfer confidence model on the independent validation set for positive and negative samples exceeds the accuracy threshold, training is stopped, and the pre-training of the intent transfer confidence model is completed. When the intent transfer confidence model receives any intent sequence, it sequentially inputs the temporal hidden state combination corresponding to each adjacent intent transfer pair in the intent sequence into the confidence decoder, and arranges the intent transfer confidence estimates output by the confidence decoder in the order of adjacent positions to form an intent transfer confidence sequence.

6. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 1, characterized in that, The difference between the uncertainty coefficient and the conditional entropy of adjacent intentions is used as the initial measure of temporal dependency. The initial measure of temporal dependency is then weighted and corrected using the intention transition confidence sequence to obtain a corrected measure of temporal dependency, including: Based on the frequency statistics of each intent transition pair in the historical dialogue data, the conditional probability distribution of the subsequent intent under each preceding intent is determined, and the conditional entropy of adjacent intents is calculated using the conditional probability distribution. Based on the uncertainty coefficient and the adjacent intention conditional entropy, determine the initial measure of temporal dependency; Extract the confidence value at each adjacent position in the intent transfer confidence sequence, and calculate the arithmetic mean of all confidence values ​​as the correction weight; The initial time-series dependency metric is multiplied by the corrected weight, and the product is output as the corrected time-series dependency metric.

7. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 6, characterized in that, Based on the frequency statistics of each intent transition pair in the historical dialogue data, the conditional probability distribution of subsequent intents under each preceding intent is determined, and the conditional entropy of adjacent intents is calculated using the conditional probability distribution, including: Traverse all intent sequences in the historical dialogue data, count the frequency of occurrence of each subsequent intent under each preceding intent condition, and construct a condition frequency matrix with all intents as rows and all intents as columns. Obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set; Additive smoothing is performed on the occurrence frequency of each condition in the conditional frequency matrix. The smoothed conditional frequency is divided by the total smoothed frequency of the row to obtain the conditional probability value of each subsequent intention under each preceding intention, thus forming a conditional probability matrix. For each row in the conditional probability matrix, multiply all conditional probability values ​​in the row by the logarithm of the corresponding conditional probability value, sum all the products and take the opposite number to obtain the single-intent conditional entropy of the corresponding preceding intent; The total frequency of each preceding intent is counted from the historical dialogue data, and the total frequency is divided by the sum of the total frequencies of all preceding intents in the historical dialogue data to obtain the marginal probability of each preceding intent. The marginal probabilities of all preceding intentions are summed with the corresponding single intention conditional entropy, and the summation result is output as the adjacent intention conditional entropy.

8. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 1, characterized in that, Based on the frequency distribution of directed intention transition pairs in the intention sequence, a stiffness metric for the state transition skeleton is determined. The stiffness metric is then corrected a second time using the distribution characteristics of the response delay to obtain a topological stiffness correction metric, including: Traverse all adjacent intent transition pairs in the intent sequence, count the frequency of occurrence of each directed intent transition pair, and construct a directed transition frequency matrix with all intent tags in the scene intent tag set as row and column indices; Obtain the total number of intents in the scene intent tag set and the frequency threshold corresponding to the scene intent tag set, and perform additive smoothing on each occurrence frequency in the directed transition frequency matrix; Divide the smoothed frequency by the total smoothed frequency of the row to obtain the transition probability from each preceding intention to the subsequent intention, forming a directed transition probability matrix. Calculate the sum of squares of all transition probability values ​​in the directed transition probability matrix, and combine it with the total number of intentions to determine the average transition concentration, which serves as the initial stiffness measure of the state transition skeleton. Extract all response delay values ​​from the response delay sequence, calculate the mean response delay and the standard deviation of the response delay, and use the ratio of the standard deviation of the response delay to the mean response delay as the delay variation coefficient; Based on the initial stiffness metric and the delay variation coefficient, a topological stiffness correction metric is determined, wherein the topological stiffness correction metric is positively correlated with the initial stiffness metric and negatively correlated with the delay variation coefficient.

9. The method for quantitatively measuring cross-scenario voice dialogue behavior patterns according to claim 1, characterized in that, After integrating the temporal dependency correction metric and the topological stiffness correction metric to output a comprehensive metric index, the method further includes: The comprehensive metric indexes output by multiple dialogue streams at different time periods under the same target scenario are obtained and a comprehensive metric index set is formed. Calculate the arithmetic mean of all comprehensive measurement indices in the comprehensive measurement index set to obtain the comprehensive measurement benchmark value; Calculate the difference between each comprehensive metric index in the comprehensive metric index set and the comprehensive metric benchmark value, sum the absolute values ​​of each difference and divide by the total number of comprehensive metric indices in the comprehensive metric index set to obtain the comprehensive metric dispersion. The overall metric dispersion is output as a cross-dialogue consistency metric for the regularity of dialogue behavior in the target scenario.

10. A quantitative measurement system for cross-scenario voice dialogue behavior patterns, characterized in that, A system for implementing a quantitative measurement method for cross-scenario voice dialogue behavior patterns as described in any one of claims 1-9, the system comprising: An intent sequence module is constructed to acquire the voice dialogue stream in the target scenario, parse the round information and the intent tag corresponding to each round in the voice dialogue stream, construct an intent sequence with round as the time unit, and record the response delay between adjacent rounds. The uncertainty coefficient acquisition module is used to calculate the information entropy based on the probability of occurrence of different intentions in the intention sequence, use the information entropy as a quantification value of the uncertainty of the dialogue intention call, and output the uncertainty coefficient. The intent transfer confidence model construction module is used to input the intent sequence into the intent transfer confidence model and output the intent transfer confidence sequence at each adjacent position, wherein the intent transfer confidence model is constructed based on the intent transfer frequency distribution in cross-scene historical dialogue data; The weighted correction module is used to take the difference between the uncertainty coefficient and the conditional entropy of adjacent intentions as the initial measure of temporal dependence, and to use the intention transfer confidence sequence to perform weighted correction on the initial measure of temporal dependence to obtain the temporal dependence correction measure. The secondary correction module is used to determine the stiffness metric of the state transition skeleton based on the frequency distribution of directed intention transition pairs in the intention sequence, and to perform secondary correction on the stiffness metric using the distribution characteristics of the response delay to obtain the topological stiffness correction metric. The comprehensive metric index fusion module is used to fuse the temporal dependency correction metric and the topological stiffness correction metric to output a comprehensive metric index, thereby quantifying the intensity of the regularity of voice dialogue behavior in the target scenario.

Images