A semantic recognition method and system for intelligent calling

Abstract

The present invention relates to the field of speech recognition technology, and in particular to a semantic recognition method and system for intelligent calls. The present invention first obtains the voice text, response time and answer text of each call statement of the user's call; further, according to the similar features of the voice text in the target user's call and each type of answer text, obtains the matching coefficient between the target user's call and each type of answer text, and determines the answer text category corresponding to the target user's call; further, according to the similar features of the voice text in the adjacent user calls of the same answer text category, the similar features and response time of the call statement in the target user's call and the corresponding answer text, and the similar features of the adjacent call statements, combined with the matching coefficient, accurately evaluates the system's corrected recognition efficiency for the target user's call from multiple aspects, so that the system can timely discover the bottleneck of recognition efficiency, thereby optimizing the system in a targeted manner and improving the intelligence level of the system.

Description

Technical Field

[0001] The present invention relates to the field of speech recognition technology, and in particular to a semantic recognition method and system for intelligent calling. Background Art

[0002] With the advancement of algorithms, semantic recognition technology has evolved from the initial simple command recognition to today's system that can understand natural language. Through semantic recognition technology, the police communication smart call system can quickly convert the user's voice input into text and extract key information from it, such as emergency situation, location, time, etc., so as to speed up the response. Based on the extracted information, the system can automatically classify the call and route it to the most appropriate department or person; at the same time, for common problems, the system can provide self-service options to guide users to solve problems through simple steps, reducing the pressure on manual customer service.

[0003] The rule-based semantic recognition method relies on pre-defined rules and templates to parse the voice or text input by the user. It is difficult to handle the complex, diverse and personalized user voice input in practice, and it is easy to cause inaccurate semantic recognition. At the same time, judging the efficiency of semantic recognition is also an important part of evaluating service quality and optimizing the police smart call system. The existing police smart call system used for policing lacks a feedback link for judging the efficiency of semantic recognition, or relies on simple similarity indicators such as edit distance to evaluate the efficiency of semantic recognition, resulting in inaccurate judgment of recognition efficiency and inaccurate feedback to the system, which makes it impossible to repair or optimize the system in a timely manner. Summary of the invention

[0004] In order to solve the technical problem of inaccurate recognition efficiency evaluation in police communication smart calls, the purpose of the present invention is to provide a semantic recognition method and system for smart calls. The technical solutions adopted are as follows:

[0005] A semantic recognition method for smart calling, the method comprising:

[0006] Acquire the voice data of the user's call and convert it into voice text, and obtain the response time and answer text of each call sentence in the voice data;

[0007] Select any user call as a target user call; obtain a matching coefficient between the target user call and each type of answer text according to similar features between the voice text in the target user call and each type of answer text; determine the answer text category corresponding to the target user call according to all the matching coefficients corresponding to the target user call;

[0008] Based on the similar features of the voice texts in adjacent user calls corresponding to the same answer text category, combined with the matching coefficient, the semantic recognition efficiency of the target user call is obtained; based on the similar features and response time of the voice texts of the call statements of the first preset number of sentences in the target user call and the corresponding answer text, as well as the similar features of the adjacent call statements in the first second preset number of sentences, combined with the semantic recognition efficiency, the corrected recognition efficiency of the target user call is obtained.

[0009] Furthermore, the method for obtaining the matching coefficient includes:

[0010] Select any type of answer text as the target category answer; in the voice text called by the target user, use the voice text of the call sentence corresponding to the answer text of the target category answer as the target category voice text;

[0011] Obtaining an average value of cosine similarities between keyword vectors of all target category speech texts and keyword vectors corresponding to the answer texts as a first matching sub-coefficient between the target user call and the target category answer;

[0012] Obtaining the Jaccard similarity coefficient between the set of keywords of the target category voice text and the set of keywords corresponding to the answer text as a second matching sub-coefficient between the target user call and the target category answer;

[0013] The product of the first matching sub-coefficient and the second matching sub-coefficient is used as the matching coefficient between the target user call and the target category answer.

[0014] Furthermore, the method for determining the answer text category corresponding to the target user call according to all the matching coefficients corresponding to the target user call includes:

[0015] The answer text category with the largest matching coefficient corresponding to the target user call is used as the answer text category corresponding to the target user call.

[0016] Furthermore, the method for obtaining the semantic recognition efficiency includes:

[0017] Taking the historical adjacent user calls of the target user with the same answer text category as the target adjacent user calls;

[0018] Matching the speech text of the call sentence in the target user's call with the speech text of the call sentence in the target adjacent user's call to obtain a matching bigram; arbitrarily combining the keyword vectors of the two speech texts in all the matching bigrams to obtain a bigram vector group, and taking the average value of the cosine similarities of all the bigram vector groups of all the matching bigrams as the first efficiency parameter;

[0019] The reciprocal of the absolute value of the difference between the maximum matching coefficient of the target user call and the maximum matching coefficient of the target adjacent user call and the sum of the preset positive parameter divided by zero is taken as the second efficiency parameter;

[0020] The product of the first efficiency parameter and the second efficiency parameter is normalized and used as the semantic recognition efficiency of the target user call.

[0021] Furthermore, the method for obtaining the modified recognition efficiency includes:

[0022] In the target user call, obtaining the product of the Jaccard similarity coefficient of the keyword of the voice text of the first preset number of call sentences and the keyword corresponding to the answer text as the first correction parameter;

[0023] Acquire the product of the response time of the first preset number of call sentences as the second correction parameter;

[0024] Obtaining a mean value of cosine similarities of sentence vectors of adjacent call sentences in the first second preset number of sentences as a third correction parameter;

[0025] The product of the semantic recognition efficiency called by the target user, the first correction parameter and the third correction parameter is used as the numerator, the second correction parameter is used as the denominator, and the fractional ratio is used as the corrected recognition efficiency called by the target user.

[0026] Furthermore, the second preset number is at least 2.

[0027] Furthermore, the speech data is converted into speech text through a hidden Markov model and a Gaussian mixture model.

[0028] Furthermore, the keywords are extracted by using a TF-IDF algorithm.

[0029] Furthermore, the preset positive division by zero parameter is 1.

[0030] The present invention also proposes a semantic recognition system for smart calls, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any one of the steps of the semantic recognition method for smart calls.

[0031] The present invention has the following beneficial effects:

[0032] The present invention first obtains the voice data of the user's call and converts it into voice text, obtains the response time and answer text of each call sentence in the voice data, and provides a data basis for subsequent analysis; further, by comparing the similar features of the voice text of the target user's call with various answer texts, the matching coefficient of the target user's call and each type of answer text is obtained, and the answer text category corresponding to the target user's call is determined, which can effectively identify the user's specific needs and determine the answer text category corresponding to the target user's call. The system can identify which categories of calls are accurately identified, which is convenient for subsequent targeted optimization and adjustment, and gradually improves the overall recognition efficiency; further, according to the adjacent user calls corresponding to the same answer text category The similar features of the speech text in the target user's call are combined with the matching coefficient to obtain the semantic recognition efficiency of the target user's call. Adjacent calls with higher similarity and higher matching coefficients indicate that the system is accurate in understanding semantics. The system can evaluate the recognition efficiency of the target user's call from multiple angles. Multi-level judgments can effectively avoid the recognition efficiency deviation that may be caused by a single indicator. Further, according to the similar features and response time of the speech text of the call statement of the first preset number of sentences in the target user's call and the corresponding answer text, as well as the similar features of the adjacent call statements in the first second preset number of sentences, combined with the semantic recognition efficiency, the corrected recognition efficiency of the target user's call is obtained from the semantic expression coherence, system response speed and accuracy. The present invention classifies user calls, and uses the similarity, semantic expression coherence, system response speed and accuracy between similar adjacent user calls to accurately evaluate the recognition efficiency of the system from multiple aspects, so that the system can promptly discover the bottleneck of recognition efficiency, which is convenient for targeted optimization of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 A flowchart of a semantic recognition method for smart calling provided by one embodiment of the present invention;

[0035] Figure 2 A flowchart of a method for acquiring semantic recognition efficiency provided by an embodiment of the present invention;

[0036] Figure 3 A flowchart of a method for obtaining a corrected recognition efficiency provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the semantic recognition method and system for intelligent calling proposed by the present invention, its specific implementation method, structure, features and effects, in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0038] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0039] The following is a detailed description of a specific scheme of a semantic recognition method and system for intelligent calling provided by the present invention in conjunction with the accompanying drawings.

[0040] See also Figure 1 , which shows a flowchart of a semantic recognition method for smart calling provided by an embodiment of the present invention, specifically comprising:

[0041] Step S1: Acquire the voice data of the user's call and convert it into voice text, and obtain the response time and answer text of each call sentence in the voice data.

[0042] In one embodiment of the present invention, the user's voice signal is collected by a microphone or other equipment, and the input voice signal is pre-processed by pre-emphasis, framing, windowing, etc. to obtain the voice data of the user's call, so as to eliminate noise and interference and improve the accuracy of voice recognition. At the same time, in order to facilitate natural language processing, the voice data is converted into voice text. Considering that the response time of each call statement directly reflects the speed at which the system responds after semantic recognition, by comparing the answer text with the voice text, the degree of understanding of the user's intention and the accuracy of the feedback of the police communication smart call system can be analyzed, providing a data basis for subsequent analysis.

[0043] In other embodiments of the present invention, recording software may be used to directly record sound, and then remove background noise in the recording to improve voice quality and remove interference for subsequent semantic recognition.

[0044] It should be noted that in one embodiment of the present invention, the voice data is converted into voice text through a hidden Markov model and a Gaussian mixture model. In other embodiments of the present invention, the voice text can also be obtained through a Whisper model and the like. These are all existing technologies and will not be described in detail here.

[0045] It should be noted that in one embodiment of the present invention, the answer text is generated by a large language model based on Prompt, and a natural and fluent voice is generated by the TTS system; the voice data of manual responses to police calls in the past three years is obtained as training data to provide services for the police communication smart call system. In other embodiments of the present invention, deep learning models can also be trained, such as convolutional neural networks, to respond to the call statements of users when they call and generate answer text.

[0046] It should be noted that, in one embodiment of the present invention, there are multiple types of answer texts in the police communication smart call, mainly including confirmation answer texts, which are used to confirm the identity of the police receiver or the caller, police information and other important contents; guidance answer texts, such as in the summary of the alarm process, in order to obtain detailed information, guiding answers may be given. Instruction answer texts, such as issuing instructions, generally contain specific action time and content. Feedback answer texts usually contain the progress and results of the police handling, and timely feedback.

[0047] Step S2: Select any user call as the target user call; obtain the matching coefficient between the target user call and each type of answer text based on the similar features between the voice text in the target user call and each type of answer text; determine the answer text category corresponding to the target user call based on all matching coefficients corresponding to the target user call.

[0048] Since the user purpose, content, method and user habits of each user call are different, the processing efficiency of the Police Communication Intelligent Call System may also fluctuate, so any user call is selected as the target user call and analyzed one by one.

[0049] Taking into account that each type of answer text in the police communication smart call system corresponds to different communication intentions and operational requirements, by comparing the similar features of the voice text of the target user's call with various answer texts, it is possible to effectively identify the user's specific needs and determine the answer text category corresponding to the target user's call. The system can identify which categories of calls are accurately identified, which facilitates subsequent targeted optimization and adjustment, and gradually improves the overall recognition efficiency. Therefore, based on the similar features of the voice text in the target user's call and each type of answer text, the matching coefficient between the target user's call and each type of answer text is obtained; based on all the matching coefficients corresponding to the target user's call, the answer text category corresponding to the target user's call is determined.

[0050] Preferably, in one embodiment of the present invention, any type of answer text is selected as the target category answer, and the voice text of the target user's call and the answer text category are compared one by one; secondly, in the voice text of the target user's call, the voice text of the call sentence corresponding to the answer text of the target category answer is used as the target category voice text, and the call sentence answered by the answer text of the target category answer is determined;

[0051] Considering that the greater the cosine similarity of the keyword vectors between the texts, the more obvious the similarity features between the texts, and the greater the matching degree between the target user call and the target category answer, the average value of the cosine similarity between the keyword vectors of all target category speech texts and the keyword vectors of the corresponding answer texts is obtained as the first matching sub-coefficient between the target user call and the target category answer;

[0052] Considering that the larger the Jaccard similarity coefficient between the keyword sets of the text, the more relevant the voice text is to the answer text, the more obvious the similarity feature is, and the greater the matching degree between the target user call and the target category answer, the Jaccard similarity coefficient between the keyword set of the target category voice text and the keyword set of the corresponding answer text is obtained as the second matching sub-coefficient between the target user call and the target category answer;

[0053] The product of the first matching sub-coefficient and the second matching sub-coefficient is used as the matching coefficient between the target user call and the target category answer.

[0054] As an example, the calculation formula of the matching coefficient includes:

[0055] ;

[0056] in, Indicates the sequence number of the target user's call; The serial number representing the answer of the target category; Indicates The second target user calls and The matching coefficient of the target category answer; Indicates The second target user calls and The number of target category speech texts corresponding to the target category answers; Indicates The second target user calls and The serial number of the target category speech text corresponding to the target category answer; Indicates The second target user calls and In the target category speech text corresponding to the target category answer, the cosine similarity between the keyword vector of the nth target category speech text and the corresponding answer text; Indicates The second target user calls and The average value of the cosine similarity between the keyword vector of the target category speech text corresponding to the target category answer and the keyword vector of the corresponding answer text is also represented by the first The second target user calls and The first matching sub-coefficient corresponding to the target category answer; Indicates The second target user calls and A set of keywords of the target category speech text corresponding to the target category answers; A set of keywords representing the answer text corresponding to the target category speech text; Used to obtain the Jaccard similarity coefficient; Indicates The second target user calls and The second matching sub-coefficient corresponding to the target category answer.

[0057] In the calculation formula of the matching coefficient, the cosine similarity is used to represent the similarity between keyword vectors of different texts. The average value of the cosine similarity between the keyword vectors of all target category speech texts and the keyword vectors of the corresponding answer texts represents the similarity features between the speech text in the target user's call and the answer text of the target category answer. The larger the cosine similarity, the more obvious the similarity features between the texts, the greater the matching degree between the target user's call and the target category answer, and the greater the matching coefficient. The larger the Jaccard similarity coefficient between the keyword sets, the more relevant the speech text and the answer text, the more obvious the similarity features, the greater the matching degree between the target user's call and the target category answer, and the greater the matching coefficient.

[0058] It should be noted that, usually, the reply text generated by the police communication system has a strong correlation with the corresponding voice text, and the cosine similarity of keywords is usually a non-negative value, so the average value of the cosine similarity is directly used as the first matching sub-coefficient; wherein, the method for obtaining the average value of the cosine similarity between the keyword vector of a voice text and the keyword vector of a reply text is: any keyword vector of the voice text is arbitrarily combined with the keyword vectors of all reply texts to form a binary vector group, and the average value of the cosine similarities of all binary vector groups is used as the average value of the cosine similarity between the keyword vector of a voice text and the keyword vector of a reply text; for example, the voice text A of the nth call sentence has keyword vectors a1 and a2, and the reply text B has keyword vectors b1 and b2, then there are binary vector groups a1-b1, a1-b2, a2-b1 and a2-b2, then take the average value of the cosine similarities of the combinations of a1-b1, a1-b2, a2-b1 and a2-b2 as the corresponding ;

[0059] In another embodiment of the present invention, the cosine similarity can be linearly normalized to obtain the first matching sub-coefficient; the method for obtaining the average value of the cosine similarities between the keyword vector of a speech text and the keyword vector of a reply text can also be: taking the maximum value of the cosine similarities between any keyword vector of the speech text and the keyword vectors of all reply texts as the cosine similarity of the corresponding keyword vector of the speech text, and taking the average value of the cosine similarities corresponding to all keyword vectors of the speech text as the average value of the cosine similarities between the keyword vector of a speech text and the keyword vector of a reply text; for example, the speech text A of the nth call sentence has keyword vectors a1 and a2, and the reply text B has keyword vectors b1 and b2, and at the same time, the cosine similarity of the a1-b1 combination is greater than the cosine similarity of the a1-b2 combination, and the cosine similarity of the a2-b2 combination is greater than the cosine similarity of the a2-b1 combination, then take the average value of the cosine similarities of the a1-b1 and a2-b2 combinations as the corresponding .

[0060] It should be noted that, in one embodiment of the present invention, keywords are extracted and word vectors are obtained by the TF-IDF algorithm; in other embodiments of the present invention, implementers may also extract keywords and obtain word vectors by the TextRank algorithm or the Word2Vec algorithm, which and the Jaccard similarity coefficient are technical means well known to those skilled in the art and will not be described in detail here.

[0061] Preferably, in one embodiment of the present invention, considering that the larger the matching coefficient is, the more the target user call belongs to the corresponding answer text category, the answer text category with the largest matching coefficient corresponding to the target user call is used as the answer text category corresponding to the target user call.

[0062] Step S3: Based on the similar features of the voice texts in adjacent user calls corresponding to the same answer text category, combined with the matching coefficient, the semantic recognition efficiency of the target user call is obtained; based on the similar features and response time of the voice texts of the call statements of the first preset number of sentences in the target user call and the corresponding answer text, as well as the similar features of the adjacent call statements in the first second preset number of sentences, combined with the semantic recognition efficiency, the corrected recognition efficiency of the target user call is obtained.

[0063] Taking into account that the semantics between adjacent calls of the same type usually have certain similarities and correlations, the system can reuse the last answer or provide updated information, which can save time for semantic recognition and thus improve recognition efficiency; the matching coefficient reflects the degree of matching between the target user call and the answer text category. Adjacent calls with higher similarity and higher matching coefficients indicate that the system's semantic understanding is accurate. The system can evaluate the recognition efficiency of the target user call from multiple angles. Multi-level judgments can effectively avoid the recognition efficiency deviation that may be caused by a single indicator, thereby supporting the continuous optimization of the algorithm and ensuring that the system can respond and adapt to new recognition needs at any time, thereby improving the overall recognition effect and user satisfaction. Therefore, the semantic recognition efficiency of the target user call is obtained based on the similar features of the voice text in adjacent user calls corresponding to the same answer text category, combined with the matching coefficient.

[0064] Preferably, in one embodiment of the present invention, the method for acquiring semantic recognition efficiency includes:

[0065] See also Figure 2 , which shows a flow chart of a method for acquiring semantic recognition efficiency provided by an embodiment of the present invention, specifically comprising:

[0066] Step S301: historical adjacent user calls of the same answer text category as the target user calls are taken as target adjacent user calls.

[0067] First, obtain the calls of similar adjacent users called by the target user to determine the objects for comparison and analysis.

[0068] Step S302: Match the speech text of the call statement in the target user's call with the speech text of the call statement in the target adjacent user's call to obtain a matching bigram; arbitrarily combine the keyword vectors of the two speech texts in all matching bigrams to obtain a bigram vector group, and use the average cosine similarity of all bigram vector groups of all matching bigrams as the first efficiency parameter.

[0069] Considering that the larger the cosine similarity, the more similar the two corresponding vectors are; the larger the cosine similarity of the binary combination of all keyword vectors between the voice texts, the greater the similarity between the voice texts; further, the larger the cosine similarity between the voice texts of the two call processes, the more similar the two calls are, and the higher the semantic recognition efficiency of the system, so the first efficiency parameter is obtained. The calculation formula of the first efficiency parameter includes:

[0070] ;

[0071] in, Indicates the sequence number of the target user's call; Indicates The first efficiency parameter of the secondary target user call; Indicates The number of call statements called by the secondary target user; Indicates The sequence number of the calling statement called by the secondary target user; Indicates The number of call statements called by the target adjacent user called by the secondary target user; Indicates The sequence number of the call statement of the target adjacent user called by the secondary target user; Indicates The target user calls The voice text of the call sentence is the same as the voice text of the target adjacent user in the call. The average cosine similarity of all binary vector groups corresponding to the speech text of the sentence call sentence.

[0072] In the calculation formula of the first efficiency parameter, the similarity features between keyword vectors are represented by cosine similarity. The first efficiency parameter is obtained by taking the average cosine similarity of all binary vector groups of all matching binary groups in the target user call and the target adjacent user call, which represents the similarity features of the speech texts in adjacent user calls corresponding to the same answer text category, thereby providing a basis for obtaining the semantic recognition efficiency.

[0073] Similarly, in other embodiments of the present invention, the cosine similarity of the binary vector group may be normalized to obtain the first efficiency parameter.

[0074] Step S303: taking the reciprocal of the absolute value of the difference between the maximum matching coefficient of the target user call and the maximum matching coefficient of the target adjacent user call and the sum of the preset positive parameter divided by zero, and using the result as the second efficiency parameter.

[0075] Taking into account that the answer text categories corresponding to the target user call and the target adjacent user call are determined by their respective corresponding maximum matching coefficients, the maximum matching coefficient of the target user call and the maximum matching coefficient of the target adjacent user call are analyzed; considering that the smaller the absolute value of the difference between the data values, the smaller the difference between the data, and the more similar the data are after taking the inverse, while avoiding the denominator being zero, a preset division by zero increase parameter is added. Based on this, the absolute value of the difference between the maximum matching coefficient of the target user call and the maximum matching coefficient of the target adjacent user call is taken the inverse of the sum of the preset division by zero positive parameter and used as the second efficiency parameter.

[0076] As an example, the divide-by-zero positive parameter is preset to 1.

[0077] Step S304: normalize the product of the first efficiency parameter and the second efficiency parameter to obtain the result as the semantic recognition efficiency of the target user call.

[0078] The first efficiency parameter evaluates the similar features between the target user call and the target adjacent user call from the perspective of cosine similarity of the vectors. The larger the first efficiency parameter, the more obvious the similar features. The second efficiency parameter evaluates the similar features between the target user call and the target adjacent user call from the perspective of the matching coefficient. The larger the second efficiency parameter, the more obvious the similar features. At the same time, the more similar the adjacent calls of the same type are, the more the Police Communication Intelligent Call System can reuse the last answer or provide updated information, and the higher the semantic recognition efficiency. Therefore, the product of the first efficiency parameter and the second efficiency parameter is normalized and used as the semantic recognition efficiency of the target user call.

[0079] It should be noted that the normalization can be performed specifically by a linear normalization method.

[0080] Since there will be guidance in police smart calls, for example, after the first sentence is said, there will be a response, and the second sentence is based on the content after the response, until the third sentence, the semantic trend of the voice text of the call statement is similar to the semantic trend of the corresponding answer. Therefore, the accuracy of the semantic recognition of the first few sentences is very important. At the same time, the response time directly reflects the semantic recognition efficiency of the system response, so the similar features and response time of the voice text of the call statement of the first preset number of sentences in the target user's call and the corresponding answer text are combined for correction; it is also considered that the similar features between the voice texts of the first few sentences represent the coherence of the user's semantic expression and reflect the consistency of the content of the conversation, so the similar features of the adjacent call sentences in the first second preset number of sentences are combined with the semantic recognition efficiency to obtain the corrected recognition efficiency of the target user's call.

[0081] Preferably, in one embodiment of the present invention, the method for obtaining the modified recognition efficiency includes:

[0082] See also Figure 3 , which shows a flow chart of a method for obtaining a modified recognition efficiency provided by an embodiment of the present invention, specifically comprising:

[0083] Step S311: in a target user call, the product of the Jaccard similarity coefficients of the keywords of the voice text of the first preset number of call sentences and the keywords of the corresponding answer text is obtained as the first correction parameter.

[0084] Considering that the larger the Jaccard similarity coefficient between the keywords of the voice text of the call statement and the keywords of the corresponding answer text, the higher the similarity between the voice text and the corresponding answer text, and the more accurate the system's response to the call statement, the product of the Jaccard similarity coefficients of the keywords of the voice text of the call statement of the first preset number of sentences and the keywords of the corresponding answer text is obtained as the first correction parameter.

[0085] As an example, the first preset number is 2, and the product of the Jaccard similarity coefficient of the first call sentence and the Jaccard similarity coefficient of the second call sentence is used as the first correction parameter.

[0086] Step S312: Obtain the product of the response times of the first preset number of call sentences as the second correction parameter.

[0087] As an example, the product of the response time of the first call statement and the response time of the second call statement is used as the second correction parameter.

[0088] Step S313: obtaining the mean value of cosine similarities of sentence vectors of adjacent call sentences in the first second preset number of sentences as the third correction parameter.

[0089] Considering that the greater the cosine similarity of the sentence vectors between the call sentences, the more similar the call sentences are, the third correction parameter is obtained based on this.

[0090] As an example, the second preset number is 3, and the sentence vectors of the first three call statements are obtained by the TF-IDF weighted average method, the cosine similarity between the sentence vector of the first call statement and the sentence vector of the second call statement is obtained, and the cosine similarity between the sentence vector of the second call statement and the sentence vector of the third call statement is obtained, and the average of the two cosine similarities is used as the third correction parameter.

[0091] It should be noted that in other embodiments of the present invention, the implementer may also perform linear normalization on the cosine similarity between sentence vectors to obtain a third correction parameter; the implementer may also obtain sentence vectors through the BERT model, which and the TF-IDF weighted average method for obtaining sentence vectors are both existing technologies and will not be described in detail.

[0092] Step S314: taking the product of the semantic recognition efficiency of the target user's call, the first correction parameter and the third correction parameter as the numerator, taking the second correction parameter as the denominator, and taking the fractional ratio as the corrected recognition efficiency of the target user's call.

[0093] As an example, the calculation formula for correcting the recognition efficiency includes:

[0094] ;

[0095] in, Indicates the sequence number of the target user's call; Indicates The semantic recognition efficiency of the secondary target user's call; Indicates Corrected recognition efficiency of secondary target user calls; Indicates The Jaccard similarity coefficient of the first call sentence of the secondary target user; Indicates The Jaccard similarity coefficient of the second call sentence of the secondary target user; Indicates The first correction parameter of the secondary target user call; Indicates The response time of the first call statement of the secondary target user; Indicates The response time of the second call statement of the secondary target user; Indicates A second correction parameter for the secondary target user call; Indicates The cosine similarity between the sentence vector of the first call sentence and the sentence vector of the second call sentence of the secondary target user; Indicates The cosine similarity between the sentence vector of the second call sentence and the sentence vector of the third call sentence for the secondary target user; Indicates The third correction parameter of the secondary target user call.

[0096] In the calculation formula of the corrected recognition efficiency, the larger the first correction parameter is, the more accurate the response of the Police Communication Intelligent Call System to the first two sentences of the user's call is, the higher the efficiency of the system's semantic recognition is, and the higher the corrected recognition efficiency is; the smaller the second correction parameter is, the shorter the response time of the system's response is, the higher the efficiency of the system's semantic recognition is, and the higher the corrected recognition efficiency is; the larger the third correction parameter is, the more fluent and consistent the user's semantic expression is in the conversation, the more conducive it is to the system's accurate semantic recognition, the higher the efficiency of the system's semantic recognition is, and the higher the corrected recognition efficiency is; the greater the semantic recognition efficiency is, from the perspective of the similarity of the system's processing of similar adjacent user needs, the higher the efficiency of the system's semantic recognition is, and the higher the corrected recognition efficiency is.

[0097] It should be noted that it takes time for the system to respond, so the denominator of the calculation formula for correcting the recognition efficiency must not be zero.

[0098] It should be noted that, in the case where the number of call statements in a user call is less than the second preset number, the data sample of the user call is too small, and the user call record is determined to be invalid, and the corrected recognition efficiency is not calculated.

[0099] After obtaining the corrected recognition efficiency corresponding to each user call, the system can promptly discover the bottleneck of recognition efficiency, and can optimize the Police Smart Call System in a targeted manner based on the corresponding answer text category, continuously improve the Police Smart Call System, and improve the efficiency of semantic recognition. Through the Police Smart Call System, the rapid transmission of information and the issuance of accurate instructions can be achieved, ensuring that police issues are resolved in a timely and effective manner.

[0100] As an example, the correction recognition efficiency is linearly normalized and the threshold is set to 0.6. If the normalized correction recognition efficiency of the current user's call is less than 0.6, it is determined that the police communication intelligent call system has defects in the semantic recognition of the current user's call, and the communication record is sent to the manual channel for review, and the current calling user is contacted manually in time to solve the police problem; at the same time, the communication record, the normalized correction recognition efficiency and the answer text category are sent to the system maintenance personnel to optimize the system in a targeted manner. For example, for the scenario of instruction-type answer text, the system can focus on optimizing the recognition speed and accuracy; for the scenario of guidance-type answer text, the system may optimize the semantic understanding ability of multiple rounds of dialogue.

[0101] An embodiment of the present invention also provides a semantic recognition system for smart calls, the system includes a memory, a processor and a computer program, wherein the memory is used to store the corresponding computer program, the processor is used to run the corresponding computer program, and when the computer program runs in the processor, it can implement a semantic recognition method for smart calls described in steps S1-S3.

[0102] In summary, in response to the technical problem of inaccurate recognition efficiency evaluation in smart calls of the police communication system, the present invention proposes a semantic recognition method and system for smart calls. The present invention first obtains the voice text, response time and answer text of each call statement of the user's call; further, based on the similar features between the voice text in the target user's call and each type of answer text, obtains the matching coefficient between the target user's call and each type of answer text, and determines the answer text category corresponding to the target user's call; further, based on the similar features of the voice text in the calls of adjacent users of the same answer text category, the similar features and response time of the call statement in the target user's call and the corresponding answer text, and the similar features of adjacent call statements, combined with the matching coefficient, obtains the corrected recognition efficiency of the target user's call. The present invention accurately evaluates the system's corrected recognition efficiency for the target user's call from multiple aspects, so that the system can promptly discover the bottleneck of recognition efficiency, thereby optimizing the system in a targeted manner and improving the intelligence level of the system.

[0103] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0104] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

Claims

1. A semantic recognition method for smart calling, characterized in that: The method comprises: Acquire the voice data of the user's call and convert it into voice text, and obtain the response time and answer text of each call sentence in the voice data; Select any user call as a target user call; obtain a matching coefficient between the target user call and each type of answer text according to similar features between the voice text in the target user call and each type of answer text; determine the answer text category corresponding to the target user call according to all the matching coefficients corresponding to the target user call; Based on the similar features of the voice texts in adjacent user calls corresponding to the same answer text category, combined with the matching coefficient, the semantic recognition efficiency of the target user call is obtained; based on the similar features and response time of the voice texts of the call statements of the first preset number of sentences in the target user call and the corresponding answer text, as well as the similar features of the adjacent call statements in the first second preset number of sentences, combined with the semantic recognition efficiency, the corrected recognition efficiency of the target user call is obtained.

2. A semantic recognition method for smart calling according to claim 1, characterized in that: The method for obtaining the matching coefficient includes: Select any type of answer text as the target category answer; in the voice text called by the target user, use the voice text of the call sentence corresponding to the answer text of the target category answer as the target category voice text; Obtaining an average value of cosine similarities between keyword vectors of all target category speech texts and keyword vectors corresponding to the answer texts as a first matching sub-coefficient between the target user call and the target category answer; Obtaining the Jaccard similarity coefficient between the set of keywords of the target category voice text and the set of keywords corresponding to the answer text as a second matching sub-coefficient between the target user call and the target category answer; The product of the first matching sub-coefficient and the second matching sub-coefficient is used as the matching coefficient between the target user call and the target category answer.

3. The semantic recognition method for smart calling according to claim 2, characterized in that: The method for determining the answer text category corresponding to the target user call according to all the matching coefficients corresponding to the target user call includes: The answer text category with the largest matching coefficient corresponding to the target user call is used as the answer text category corresponding to the target user call.

4. The semantic recognition method for smart calling according to claim 2, characterized in that: The method for acquiring the semantic recognition efficiency includes: Taking the historical adjacent user calls of the target user with the same answer text category as the target adjacent user calls; Matching the speech text of the call sentence in the target user's call with the speech text of the call sentence in the target adjacent user's call to obtain a matching bigram; arbitrarily combining the keyword vectors of the two speech texts in all the matching bigrams to obtain a bigram vector group, and taking the average value of the cosine similarities of all the bigram vector groups of all the matching bigrams as the first efficiency parameter; The reciprocal of the absolute value of the difference between the maximum matching coefficient of the target user call and the maximum matching coefficient of the target adjacent user call and the sum of the preset positive parameter divided by zero is taken as the second efficiency parameter; The product of the first efficiency parameter and the second efficiency parameter is normalized and used as the semantic recognition efficiency of the target user call.

5. The semantic recognition method for smart calling according to claim 2, characterized in that: The method for obtaining the corrected recognition efficiency includes: In the target user call, obtaining the product of the Jaccard similarity coefficient of the keyword of the voice text of the first preset number of call sentences and the keyword corresponding to the answer text as the first correction parameter; Acquire the product of the response time of the first preset number of call sentences as the second correction parameter; Obtaining a mean value of cosine similarities of sentence vectors of adjacent call sentences in the first second preset number of sentences as a third correction parameter; The product of the semantic recognition efficiency called by the target user, the first correction parameter and the third correction parameter is used as the numerator, the second correction parameter is used as the denominator, and the fractional ratio is used as the corrected recognition efficiency called by the target user.

6. The semantic recognition method for smart calling according to claim 1, characterized in that: The second preset number is at least 2.

7. The semantic recognition method for smart calling according to claim 1, characterized in that: The speech data is converted into speech text through a hidden Markov model and a Gaussian mixture model.

8. The semantic recognition method for smart calling according to claim 2, characterized in that: The keywords are extracted by TF-IDF algorithm.

9. The semantic recognition method for smart calling according to claim 4, characterized in that: The preset zero division positive parameter is 1.

10. A semantic recognition system for smart calling, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of a semantic recognition method for intelligent calling as described in any one of claims 1 to 9 are implemented.

Images