A telephone semantic analysis method based on speech recognition

Abstract

The present invention provides a telephone semantic analysis method based on speech recognition, comprising: obtaining call content and transcribing the call content into text data; extracting keywords from the text data and constructing a keyword list vector according to the extracted keywords; comparing the keyword list vector with each intention feature model respectively to obtain a suspicion score between the keyword list vector and each intention feature model respectively, and determining the intention represented by the call content according to the maximum value of multiple suspicion scores; wherein different intention feature models represent different intentions respectively; the method provided by the present invention converts the call speech into text content through continuous speech transcription technology, and then combines the intention understanding technology to allow the system to analyze the intention of the call content, thereby improving the accuracy of recognizing the intention of the call content.

Description

Technical Field

[0001] The present invention relates to the technical field of speech recognition and semantic understanding, and in particular to a telephone semantic analysis method, device and electronic equipment based on speech recognition. Background Art

[0002] With the development of communication technology, more and more criminals have used mobile and fixed-line communication technologies to conduct illegal propaganda and fraud, causing economic losses to many telephone users and disrupting normal social order.

[0003] Telecom users hope to have a safe calling environment to reduce and eliminate the occurrence of fraudulent calls. National public security and other government departments and telecom operators also hope to combat fraudulent calls at the source through technical means and provide the public with a safe and reliable calling environment.

[0004] Due to the limitation of technology development, the analysis and judgment of fraudulent calls in the early telephone fraud prevention system mainly relied on manual intervention to listen to the recordings for judgment, or automatic judgment by extracting simple keywords through voice recognition content. In the actual system operation, due to the excessive amount of recordings far exceeding the manual processing capacity, a large number of gray lists could not be judged in time, affecting the system's interception function. Extracting simple keywords for automatic judgment resulted in a low accuracy rate in the analysis of fraudulent numbers, which led to a large number of subsequent complaints about wrongly blocked numbers. Summary of the invention

[0005] In view of this, a solution is provided that can overcome at least one of the above-mentioned defects of the prior art.

[0006] On the one hand, the present invention provides a telephone semantic analysis method based on speech recognition, including: obtaining call content and transcribing the call content into text data; extracting keywords from the text data and constructing a keyword list vector based on the extracted keywords; comparing the keyword list vector with each intention feature model respectively to obtain the suspicion score between the keyword list vector and each intention feature model respectively, and determining the intent represented by the call content based on the maximum value of multiple suspicion scores, wherein different intention feature models respectively represent different intentions.

[0007] In one embodiment, before transcribing the call content into text data, the method further includes: performing background Gaussian noise removal processing on the call content.

[0008] In one embodiment, the keyword extraction module converts the text data into a keyword list vector including: segmenting the text data to obtain multiple text segments; filtering the text segments to obtain a keyword node set; obtaining node weights based on the keyword node set and a preset edge set; sorting the node weights in reverse order, and taking at least one of the node weights at the top of the sort as a candidate keyword vector; marking the portion of the text data that is identical to the candidate keyword, and when multiple candidate keywords are adjacent in the text data, merging the adjacent candidate keywords as a multi-word keyword vector; and merging multiple multi-word keyword vectors to obtain the keyword list vector.

[0009] In one embodiment, filtering the text segment includes: performing word segmentation processing on the text segment to obtain word segmentation data; marking the part of speech of the word segmentation data; removing stop words in the word segmentation data and the word segmentation data without preset part of speech marking in the word segmentation data to obtain a keyword node set; wherein the preset part of speech is a preset keyword part of speech.

[0010] In one embodiment, based on the co-occurrence relationship, each keyword node in the keyword node set is processed according to the keyword node set and a preset edge set, a keyword node is connected to another keyword node to form an edge, and based on the number of edges connected to the keyword and the preset threshold, the corresponding keyword nodes whose number of edges does not reach the preset threshold are removed.

[0011] In one embodiment, the remaining keyword nodes are iteratively calculated to obtain the node weights, including:

[0012]

[0013] Among them, V i 、V j Any two nodes in the keyword node set, WS(V i ) is V i The node weight, d is the damping coefficient; w ji 、w jk Respectively represent V i 、V j The weight coefficient, In(V i ) points to V i The collection of out(V j ) points to V j A collection of .

[0014] In one embodiment, the intent understanding module compares the keyword list vector with each of the intent feature models to obtain a suspicion score corresponding to the keyword list vector and each of the intent feature models, including: respectively calculating the relative entropy of the keyword vector and each of the intent feature models, wherein the minimum value among multiple relative entropies corresponds to the maximum value of the suspicion score.

[0015] In one embodiment, respectively calculating the relative entropy between the keyword vector and each of the intention feature models specifically includes:

[0016]

[0017] Among them, P(x) represents the probability distribution of a keyword in the keyword list vector, Q(x) represents the probability distribution of the keyword in the intention feature model, x represents the preset keyword character set in the intention feature model, and KL(P||Q) represents the relative entropy corresponding to P(x) and Q(x).

[0018] Another aspect of the present invention provides a telephone semantic analysis system based on speech recognition, including: a speech recognition module, used to obtain the content of a call, transcribe the call content into text data, and send it to a keyword extraction module; a keyword extraction module, used to convert the text data into a keyword list vector, and send it to an intention understanding module; an intention understanding module, whose feature library includes multiple intention feature models representing different intentions, used to compare the keyword list vector with each intention feature model to obtain the suspicion score corresponding to the keyword list vector and each intention feature model, and determine the intention represented by the call content according to the maximum value of the multiple suspicion scores.

[0019] Another aspect of the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned telephone semantic analysis method based on speech recognition when executing the computer program.

[0020] Compared with the prior art, the present invention has the following technical effects: by extracting semantic attribute features, from the perspective of content cognition, based on speech recognition and semantic understanding technology, a multi-dimensional cognitive model is constructed to analyze and detect fraudulent numbers, thereby maximizing the accuracy of fraudulent call identification and reducing the misjudgment rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The following drawings are only used to illustrate and explain the present invention, and do not limit the scope of the present invention:

[0022] Figure 1A schematic diagram showing a method flow of an embodiment of the present invention; and

[0023] Figure 2 A schematic diagram of a method flow chart of an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that in this specification, the expressions of first, second, etc. are only used to distinguish one feature from another feature, and do not represent any limitation on the feature. Therefore, without departing from the teaching of this application, the first buffer discussed below may also be referred to as the second buffer.

[0026] In the drawings, the thickness, size and shape of objects have been slightly exaggerated for ease of explanation. The drawings are only examples and are not drawn strictly to scale.

[0027] It should also be understood that the terms "comprises", "including", "having", "includes" and / or "comprising", when used in this specification, indicate the presence of the stated features, wholes, steps, operations, elements and / or parts, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, parts and / or combinations thereof. In addition, when expressions such as "at least one of..." appear after a list of listed features, they modify the entire listed features rather than modifying the individual elements in the list. In addition, when describing embodiments of the present application, "may" is used to mean "one or more embodiments of the present application". And, the term "exemplary" is intended to refer to an example or illustration.

[0028] As used herein, the terms "substantially," "approximately," and similar terms are used as terms of approximation, not degree, and are intended to account for the inherent variations in measurements or calculations that would be recognized by those of ordinary skill in the art.

[0029] Unless otherwise defined, all terms (including technical terms and scientific terms) used in this article have the same meaning as commonly understood by ordinary technicians in the field to which this application belongs. It should also be understood that terms (such as terms defined in commonly used dictionaries) should be interpreted as having the same meaning as their meaning in the context of the relevant technology, and will not be interpreted in an idealized or overly formal sense unless explicitly defined in this article.

[0030] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0031] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention is shown; a telephone semantic analysis method based on speech recognition provided by the present invention comprises the following steps:

[0032] S1, obtaining the call content and transcribing the call content into text data;

[0033] S2, extract keywords from text data and construct a keyword list vector based on the extracted keywords;

[0034] S3, compare the keyword list vector with each intent feature model respectively to obtain the suspicion score between the keyword list vector and each intent feature model respectively, and determine the intent represented by the call content according to the maximum value of multiple suspicion scores, wherein different intent feature models represent different intents respectively.

[0035] The method provided by the present invention will be described step by step below.

[0036] S1, obtaining the call content and transcribing the call content into text data.

[0037] In some embodiments, S1 can specifically decode the call content by first serially inputting the call content into the acoustic model, and then inputting the output features of the acoustic model into the language model, calculating the posterior probability of the candidate words using the Bayesian formula, calculating the credibility score, performing sentence segmentation processing, and transcribing the call content into text data for subsequent text processing.

[0038] In some implementations, obtaining the call content and transcribing the call content into text data may specifically adopt one or more of a random model method, an artificial neural network method, or probabilistic grammar analysis.

[0039] Specifically, after sampling the call content to obtain waveform data, it must first be input into the feature extraction module to extract appropriate acoustic feature parameters for subsequent acoustic model use. Good acoustic features should consider the following three factors. First, it should have relatively good distinguishing characteristics so that different modeling units of the acoustic model can be conveniently and accurately modeled. Second, feature extraction can also be considered as a compression encoding process for voice information, which requires eliminating the factors of the channel and the speaker and retaining information related to the content, while also using the lowest possible parameter dimension without losing too much useful information, so as to facilitate efficient and accurate model training. Third, robustness needs to be considered, that is, the ability to resist interference from environmental noise.

[0040] In some implementations, a Hidden Markov Model (HMM) may be used as an acoustic model because HMM has many excellent properties. The state transition model of the HMM model is well suited to the short-term stationary characteristics of human speech, and can conveniently perform statistical modeling on continuously generated observations (conversation content).

[0041] Take Chinese as an example: Chinese is classified into four types according to the pronunciation characteristics of phonemes: consonants, vowels, diphthongs, and complex nasal codas. According to the syllable structure, it is classified into initials and finals, and initials or finals are composed of phonemes. Sometimes, finals with tones are called tone finals. A syllable is composed of a single tone final or an initial consonant and a tone final. A syllable in Chinese is the sound of a Chinese character, that is, a syllable word. Syllable words are composed of words, and finally sentences are composed of words. There are 22 Chinese initial consonants, including zero initial consonants, and 38 finals. According to the classification of phonemes, there are 22 Chinese consonants, 13 monophthongs, 13 diphthongs, and 16 complex nasal codas. At present, the commonly used acoustic model primitives are initials, finals, syllables or words, and different primitives are selected according to different implementation purposes. Chinese has a total of 412 syllables including modal particles, including light-tone characters, and a total of 1,282 syllables. Therefore, when isolated words in a small vocabulary are recognized, words are often selected as primitives. Syllables or finals are often used for modeling in large vocabulary recognition. In continuous speech recognition, finals are often used for modeling due to the influence of co-articulation.

[0042] The most commonly used statistical-based acoustic model is the HMM model. The relevant theories of the HMM model include model structure selection, model initialization, model parameter re-estimation, and corresponding recognition algorithms.

[0043] In the implementation mode provided by the present invention, before transcribing the call content into text data, it is also necessary to perform speech denoising on the call content, specifically, to remove the background Gaussian noise in the call content.

[0044] In some implementations, the speech denoising method may be: a method based on the TF (Time-Frequency) domain, or a method based on the time domain.

[0045] S2, extracting keywords from the text data, and constructing a keyword list vector based on the extracted keywords.

[0046] In the implementation mode provided by the present invention, this step S2 specifically includes: segmenting the text data to obtain multiple text segments; filtering each text segment to obtain a keyword node set composed of multiple keywords, wherein each keyword serves as a keyword node, the set of all keyword nodes is a keyword node set, and all text segments correspond to a keyword node set; obtaining node weights according to the keyword node set and a preset edge set; sorting the node weights in reverse order, and forming a candidate keyword vector (the value of n can be determined according to actual conditions) with at least one of the top node weights, that is, the n node weights ranked in the top n places; marking the part of the text data that is the same as the candidate keyword, and when multiple candidate keywords are adjacent in the text data, merging the adjacent candidate keywords as a multi-word keyword vector; merging multiple multi-word keyword vectors to obtain the keyword list vector.

[0047] In some embodiments, filtering a text segment includes: performing word segmentation processing on the text segment to obtain word segmentation data; marking the part of speech of the word segmentation data; removing stop words in the word segmentation data and the word segmentation data that does not have the preset part of speech marking in the word segmentation data to obtain a keyword node set; wherein the preset part of speech is a preset keyword part of speech.

[0048] In some implementations, the word segmentation process may adopt an algorithm: a word segmentation method based on string matching, a word segmentation method based on understanding, or a word segmentation method based on statistics. Depending on whether it is combined with the part-of-speech tagging process, it can be divided into a simple word segmentation method and an integrated method combining word segmentation and tagging.

[0049] In some implementations, word segmentation processing can be implemented using tools such as Jieba word segmentation as word segmentation.

[0050] In some implementations, the marked part of speech may be one part of speech, or multiple parts of speech may be marked according to actual needs.

[0051] In some embodiments, obtaining the node weight according to the keyword node set and the preset edge set includes: based on the co-occurrence relationship, processing each keyword node in the keyword node set according to the keyword node set and the preset edge set, connecting a keyword node with another keyword node to form an edge, and based on the preset threshold and the number of edges connected to the keyword, removing the corresponding keyword nodes whose number of edges does not reach the preset threshold.

[0052] Specifically, the edge set includes a set of multiple text areas, and each text area is a text area in the text segment after the above filtering. A keyword is selected as the starting point of the text area, and then a keyword after the keyword is selected as the end point of the text area. The current text segment formed is used as a text area in the edge set, and the number of times the start point-end point appear is recorded. The edge set is a set of the above multiple text areas and the corresponding start point-end point appearance times. The specific selection of the number of text areas, the starting point and the end point can be set according to actual needs; in some embodiments, more than two keyword nodes can also constitute the text area.

[0053] Furthermore, according to the number of edges connected to each keyword, based on a preset threshold, corresponding keyword nodes whose number of edges does not reach the threshold are removed. The preset threshold can be represented by K.

[0054] In some implementations, the remaining keyword nodes are iteratively calculated to obtain the node weights as follows:

[0055]

[0056] Among them, V i 、V j Any two nodes in the keyword node set, WS(V i ) is V i The node weight, d is the damping coefficient; w ji 、w jk Respectively represent V i 、V j The preset weight coefficient, In(V i ) points to V i The collection of out(V j ) points to V j A collection of .

[0057] In some embodiments, the value of d may be 0.85.

[0058] In some implementations, a multi-word keyword may also be composed of multiple non-adjacent candidate keywords. When the number of edges formed by multiple candidate keywords reaches K, the multiple candidate keywords may also constitute a multi-word keyword.

[0059] S3, comparing the keyword list vector with each intent feature model respectively to obtain the suspicion scores between the keyword list vector and each intent feature model respectively, and determining the intent represented by the call content according to the maximum value of multiple suspicion scores, wherein different intent feature models represent different intents respectively.

[0060] refer to Figure 2For example, the intent feature model may include: fake e-commerce customer service model, fake online loan model, fake investment and financial management model, and fake order model, and the intentions represented are: fake e-commerce customer service, fake online loan, fake investment and financial management, and fake order; the feature library of each model is configured in the device's processor in advance. Furthermore, the intent feature model can be configured as a feature model that represents any other intention according to actual needs, and this article does not limit this.

[0061] In some embodiments, the intent understanding module compares the keyword list vector with each of the intent feature models to obtain a suspicion score corresponding to the keyword list vector and each of the intent feature models, including: respectively calculating the relative entropy of the keyword vector and each of the intent feature models, wherein the minimum value among multiple relative entropies corresponds to the maximum value of the suspicion score.

[0062] The similarity calculation uses the relative entropy algorithm to measure the distance between the keyword list vector and the model feature library. When the two are the same, their relative entropy is zero. When the difference between the two increases, their relative entropy will also increase. Therefore, relative entropy can be used to compare the similarity of texts. First, count the frequency of the words, and then calculate the relative entropy.

[0063] Assume P(x) and Q(x) are two probability distributions on the random variable x. In the case of discrete and continuous random variables, the relative entropy KL(P||Q) of P(x) and Q(x) is defined as:

[0064]

[0065] Furthermore, relative entropy is used to measure the number of extra bits required on average to encode samples from P using a Q-based code. P represents the true distribution of the data, and Q represents the theoretical distribution, model distribution, or approximate distribution of P. Given a probability distribution of a character set, design a code that requires the least number of bits on average to represent a string of characters. Assuming that this character set is X, and for x∈X, its probability of occurrence is P(x), then the average number of bits required for its optimal code is equal to the entropy H(x) of this character set:

[0066]

[0067] On the same character set, suppose there is another probability distribution Q(x). If the optimal encoding of the probability distribution P(x) is used (that is, the encoding length of character x is equal to ), to encode characters that conform to the distribution P(x), then these characters will use more bits than the ideal case. Relative entropy is used to measure the average number of extra bits used per character in this case, so it can be used to measure the distance between two distributions.

[0068] In the embodiments provided by the present invention, the random variable X does not involve a continuous random variable, but those skilled in the art can understand that, in some embodiments, X may be a continuous random variable.

[0069] In the implementation manner provided by the present invention, respectively calculating the relative entropy of the keyword vector and each of the intention feature models specifically includes:

[0070]

[0071] Among them, P(x) represents the probability distribution of a keyword vector x in the keyword list vector, Q(x) represents the probability distribution of keyword x in the intention feature model, X represents the keyword character set preset in the intention feature model, x is a keyword vector x in the keyword list vector X, and KL(P||Q) represents the relative entropy corresponding to P(x) and Q(x).

[0072] Another aspect of the present invention further provides a speech recognition module for acquiring call content, transcribing the call content into text data, and sending it to a keyword extraction module; a keyword extraction module for converting the text data into a keyword list vector, and sending it to an intent understanding module; an intent understanding module, whose feature library includes multiple intent feature models representing different intents, for comparing the keyword list vector with each intent feature model to obtain a suspicion score corresponding to the keyword list vector and each intent feature model, and determining the intent represented by the call content based on the maximum value of the multiple suspicion scores. Its design is based on the same idea as the method provided above, and will not be elaborated in this article.

[0073] Another aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the telephone semantic analysis method based on speech recognition as described above are implemented. The design of the electronic device is based on the same idea as the method, and will not be elaborated in this article.

[0074] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A telephone semantic analysis method based on speech recognition, characterized in that: include: Acquire call content, and transcribe the call content into text data; Extracting keywords from the text data, and constructing a keyword list vector according to the extracted keywords; Comparing the keyword list vector with each intention feature model respectively to obtain a suspicion score between the keyword list vector and each intention feature model respectively, and determining the intention represented by the call content according to the maximum value of multiple suspicion scores, wherein different intention feature models represent different intentions respectively; The step of constructing a keyword list vector according to the extracted keywords includes: Segmenting the text data to obtain multiple text segments; Filtering the text segment to obtain a keyword node set; Obtaining node weights according to the keyword node set and the preset edge set; sorting the node weights in reverse order, and taking at least one node weight that is at the front of the order as a candidate keyword vector; Marking the same part of the text data as the candidate keyword, and when a plurality of the candidate keywords are adjacent in the text data, merging the adjacent candidate keywords as a multi-word keyword vector; Merge a plurality of the multi-word keyword vectors to obtain the keyword list vector; Wherein, obtaining the node weight according to the keyword node set and the preset edge set includes: Based on the co-occurrence relationship, each keyword node in the keyword node set is processed according to the keyword node set and a preset edge set, a keyword node is connected with another keyword node to form an edge, and according to the number of edges connected to the keyword and based on a preset threshold, the corresponding keyword nodes whose number of edges does not reach the preset threshold are removed; According to the preset damping coefficient, the remaining keyword nodes are iteratively calculated to obtain the node weight of each remaining keyword node.

2. The method according to claim 1, characterized in that Before transcribing the call content into text data, the method further includes: Background Gaussian noise removal is performed on the call content.

3. The method according to claim 1, characterized in that Filtering the text segment includes: Perform word segmentation on the text segment to obtain word segmentation data; Marking the part of speech of the word segmentation data; Stop words in the word segmentation data and word segmentation data without preset part-of-speech tags in the word segmentation data are removed to obtain a keyword node set; wherein the preset part-of-speech is a preset keyword part-of-speech.

4. The method according to claim 1, characterized in that The remaining keyword nodes are iteratively calculated to obtain the node weights including: in, , Any two nodes in the keyword node set, for The node weights, d is the damping coefficient; , Respectively , The weight coefficient of To point to A collection of To point to of gather.

5. The method according to claim 1, characterized in that Comparing the keyword list vector with each of the intention feature models to obtain a suspicion score corresponding to the keyword list vector and each of the intention feature models includes: The relative entropy between the keyword vector and each intention feature model is calculated respectively, wherein the minimum value among multiple relative entropies corresponds to the maximum value of the suspicion score.

6. The method according to claim 5, characterized in that: The respectively calculating the relative entropy between the keyword vector and each of the intention feature models specifically includes: Among them, P(x) represents the probability distribution of a keyword in the keyword list vector, Q(x) represents the probability distribution of the keyword in the intention feature model, x represents the keyword character set preset in the intention feature model, and KL(P||Q) represents the relative entropy corresponding to P(x) and Q(x).

7. A telephone semantic analysis system based on speech recognition, characterized in that: include: A speech recognition module, used to obtain the content of the call, transcribe the call content into text data, and send it to the keyword extraction module; A keyword extraction module, used to convert the text data into a keyword list vector and send it to the intention understanding module; An intention understanding module, whose feature library includes multiple intention feature models representing different intentions, is used to compare the keyword list vector with each intention feature model to obtain a suspicion score corresponding to the keyword list vector and each intention feature model, and determine the intention represented by the call content according to the maximum value of the multiple suspicion scores; Wherein, the keyword extraction module is specifically used to divide the text data into multiple text segments when converting the text data into a keyword list vector; filter the text segments to obtain a keyword node set; obtain node weights according to the keyword node set and a preset edge set; sort the node weights in reverse order, and use at least one node weight that is ranked first as a candidate keyword vector; mark the part of the text data that is the same as the candidate keyword, and when multiple candidate keywords are adjacent in the text data, merge the adjacent candidate keywords as a multi-word keyword vector; merge multiple multi-word keyword vectors to obtain the keyword list vector; wherein, according to the keyword node set and the preset edge set, obtaining the node weight includes: based on the co-occurrence relationship, processing each keyword node in the keyword node set according to the keyword node set and the preset edge set, connecting a keyword node with another keyword node to form an edge, and removing the corresponding keyword nodes whose number of edges does not reach the preset threshold based on the number of edges connected to the keyword and the preset threshold; According to the preset damping coefficient, the remaining keyword nodes are iteratively calculated to obtain the node weight of each remaining keyword node.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the telephone semantic analysis method based on speech recognition as described in any one of claims 1 to 6 are implemented.

Images