WO2020077895A1 - 签约意向判断方法、装置、计算机设备和存储介质 - Google Patents

签约意向判断方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2020077895A1
WO2020077895A1 PCT/CN2019/070814 CN2019070814W WO2020077895A1 WO 2020077895 A1 WO2020077895 A1 WO 2020077895A1 CN 2019070814 W CN2019070814 W CN 2019070814W WO 2020077895 A1 WO2020077895 A1 WO 2020077895A1
Authority
WO
WIPO (PCT)
Prior art keywords
scene
speech
text
fraud risk
contract
Prior art date
Application number
PCT/CN2019/070814
Other languages
English (en)
French (fr)
Inventor
臧磊
傅婧
郭鹏程
Original Assignee
深圳壹账通智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳壹账通智能科技有限公司 filed Critical 深圳壹账通智能科技有限公司
Publication of WO2020077895A1 publication Critical patent/WO2020077895A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/40Scenes; Scene-specific elements in video content
    • G06V20/41Higher-level, semantic clustering, classification or understanding of video scenes, e.g. detection, labelling or Markovian modelling of sport events or news items
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/40Spoof detection, e.g. liveness detection
    • GPHYSICS
    • G10MUSICAL INSTRUMENTS; ACOUSTICS
    • G10LSPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00Speech recognition
    • G10L15/26Speech to text systems
    • GPHYSICS
    • G10MUSICAL INSTRUMENTS; ACOUSTICS
    • G10LSPEECH ANALYSIS OR SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING; SPEECH OR AUDIO CODING OR DECODING
    • G10L25/00Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00
    • G10L25/48Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use
    • G10L25/51Speech or voice analysis techniques not restricted to a single one of groups G10L15/00 - G10L21/00 specially adapted for particular use for comparison or discrimination

Definitions

  • This application relates to a method, device, computer equipment, and storage medium for judging the intention of signing a contract.
  • a method, apparatus, computer equipment, and storage medium for judging a contract intention are provided.
  • a method of judging signing intention includes:
  • Extract a speech feature vector from the speech data and input the extracted speech feature vector into a trained fraud risk prediction model to obtain a fraud risk coefficient value;
  • the signing intention judgment result is obtained according to the text matching score, the fraud risk coefficient value, the face recognition score, and the target review rule.
  • a signing intention judgment device includes:
  • Voice data acquisition module for acquiring voice data from the terminal
  • the keyword matching module is used for performing speech recognition on the speech data to obtain corresponding speech text, extracting keywords from the speech text, matching the extracted keywords with preset keywords corresponding to the preset text, and obtaining text Matching score
  • the speech feature vector extraction module is used to extract a speech feature vector from the speech data, and input the extracted speech feature vector into a trained fraud risk prediction model to obtain a fraud risk coefficient value;
  • An image data obtaining module configured to obtain image data from the terminal, perform face recognition on the obtained image data, and obtain a face recognition score
  • a target audit rule acquisition module used to acquire a scene identifier corresponding to the current business scene, and obtain a corresponding target audit rule according to the scene identifier
  • the signing intention judgment module is used to obtain a signing intention judgment result based on the text matching score, the fraud risk coefficient value, the face recognition score, and the target review rule.
  • a computer device includes a memory and one or more processors.
  • Computer-readable instructions are stored in the memory.
  • the steps of the method for judging a contract intention provided in any embodiment of the present application are implemented.
  • One or more non-volatile computer-readable storage media storing computer-readable instructions, which when executed by one or more processors, cause the one or more processors to implement any one of the embodiments of the present application Provide the steps of the signing intention judgment method.
  • FIG. 1 is an application scenario diagram of a method for determining an intention to sign a contract according to one or more embodiments.
  • FIG. 2 is a schematic flowchart of a method for judging a contract intention according to one or more embodiments.
  • FIG. 3 is a schematic flowchart of step S210 according to one or more embodiments.
  • FIG. 4 is a block diagram of a signing intention judgment device according to one or more embodiments.
  • Figure 5 is a block diagram of a computer device in accordance with one or more embodiments.
  • the signing intention judgment method provided by this application can be applied to the application environment shown in FIG. 1.
  • the terminal 102 communicates with the server 104 through the network through the network, and the terminal 102 is used to record voice and video of the client.
  • the server 104 first obtains voice data from the terminal, then performs voice recognition on the voice data to obtain corresponding voice text, extracts keywords from the voice text, and matches the extracted keywords with preset keywords corresponding to presets to obtain a text matching score ; Then extract the voice feature vector from the voice data, and input the extracted voice feature vector into the trained fraud risk prediction model to obtain the fraud risk coefficient value; obtain image data from the terminal, and perform face recognition on the obtained image data Recognize and get face recognition scores; get the scene identification corresponding to the current business scene, and obtain the corresponding target review rules according to the scene identification; get the signing intention judgment result based on the text matching score, fraud risk coefficient value, face recognition score and review rules.
  • the terminal 102 may be, but not limited to, various personal computers, notebook computers, smart phones, tablet computers, and portable wearable devices.
  • the server 104 may be implemented by an independent server or a server cluster composed of multiple servers.
  • a method for determining the intention to sign a contract is provided.
  • the method is applied to the server in FIG. 1 as an example for illustration, and includes the following steps:
  • Step S202 Acquire voice data from the terminal.
  • a predetermined text of a given contracted customer is given, such as an informed consent form
  • the customer reads the text aloud
  • the terminal records the customer's voice and video
  • the server can obtain the recorded voice data from the terminal.
  • the server obtains the voice data corresponding to the entire text uploaded by the terminal after the contracted client reads the entire text; or it may acquire the voice data from the terminal at a certain time interval during the process of reading by the contracted client, for example , You can stipulate that the contracted user will pause after reading a complete sentence.
  • the terminal detects that the blank period of voice exceeds the preset threshold, it will send the currently recorded voice data to the server.
  • the currently recorded voice data refers to from The voice data that was recorded after the voice data was sent last time.
  • Step S204 Perform speech recognition on the speech data to obtain corresponding speech text, extract keywords from the speech text, and match the extracted keywords with the preset keywords corresponding to the preset text to obtain a text matching score.
  • the preset text refers to the given aloud text, which is the original text that needs to be read aloud when the contracted customer performs dual recording of voice and video.
  • keyword extraction needs to be performed in advance, and the extracted keyword is used as the preset keyword.
  • the server obtains the voice data recorded by the terminal, it first preprocesses the voice data, including noise reduction and voice enhancement, and then performs voice recognition on the preprocessed voice data to obtain voice text, and then performs key processing on the obtained voice text.
  • Word extraction to obtain keywords corresponding to the voice text, match these keywords with preset keywords, and calculate the matching degree to obtain a text matching score.
  • any speech recognition method in the prior art may be used, which will not be repeated here.
  • extracting keywords from the speech text includes: segmenting the speech text, calculating the TF-IDF weights for each word obtained from the word segmentation, and then sorting each word in descending order according to the TF-IDF weights, and ranking the top As a keyword.
  • matching the extracted keywords with the preset keywords corresponding to the preset text to obtain a text matching score includes: first calculating the number of words matching the preset keywords among the keywords corresponding to the voice text , The text matching score is obtained according to the ratio of the number of words on the match to the total number of words of the preset keyword.
  • the keywords corresponding to the speech text are A, B, and D.
  • the preset keywords include A, B, C, D, and E.
  • Step S206 Extract a voice feature vector from the voice data, and input the extracted voice feature vector into the trained fraud risk prediction model to obtain a fraud risk coefficient value.
  • the fraud risk prediction model is used to predict the fraud risk coefficient value of the contracted customer based on the voice data.
  • the fraud risk coefficient value is used to characterize the possible fraud risk of the contracted customer. The greater the value of the fraud risk coefficient, the greater the fraud risk of the contracted customer Big.
  • the fraud risk prediction model can be obtained through the training of supervised machine learning models based on historical data. Machine learning models include but are not limited to SVM (Support Vector Machine), logistic regression models, decision trees, etc.
  • the server may use MFCC (Mel-Frequency, Cepstral, Coefficients, Mel Cepstral Coefficients) to extract voice features from the voice data to obtain corresponding feature coefficients, and vectorize the feature coefficients to obtain Corresponding speech feature vector.
  • MFCC Mel-Frequency, Cepstral, Coefficients, Mel Cepstral Coefficients
  • Step S208 Acquire image data from the terminal, perform face recognition on the acquired image data, and obtain a face recognition score.
  • the server may obtain one frame or several consecutive frames of image data from the terminal at preset time intervals, perform face recognition on the obtained image data respectively, obtain multiple face recognition scores, and then perform multiple face recognition on the obtained The score is averaged to obtain the final face recognition score.
  • the server when the server performs face recognition on the acquired image data, it first performs face detection on the picture data to obtain a face image, and then preprocesses the obtained face image.
  • the preprocessing process mainly includes the face Light compensation, grayscale transformation, histogram equalization, normalization, geometric correction, filtering, and sharpening of the image, and then feature extraction of the pre-processed face image, and finally the extracted feature data and the contract
  • the features of the face images pre-saved by the user are compared.
  • the face images pre-saved by the contracted user may be ID card images, or images collected during face recognition by calling the face recognition interface of the Ministry of Public Security.
  • Step S210 Obtain a scene identifier corresponding to the current business scene, and obtain a corresponding target audit rule according to the scene identifier.
  • the business scenario refers to the scenario when signing the contract, including but not limited to the scenarios related to loans, leases, various insurance purchases, and credit card processing.
  • the current business scenario refers to the business scenario corresponding to the current contract of the contracted customer.
  • the scene identifier is used to uniquely identify the business scene.
  • the audit rules refer to the rules used to judge the customer's intention to sign a contract. Different scenarios have different audit rules.
  • the server obtains the scene identifier corresponding to the current business scene, which may be sent by the terminal to the server before the voice and video dual recording is performed, or it may be acquired by the server from the terminal after the dual recording ends.
  • the server may first search the database according to the scene identifier. If the current business scene is a business scene that has already occurred, the corresponding audit rules have been saved in the database, so The corresponding audit rules can be searched directly according to the scene identifier to obtain the target audit rule; in other embodiments, the current business scenario is the first occurrence of the business scenario, and the corresponding contract template and contract elements can be found based on the scene identifier, based on the search The obtained contract elements use the scene classification model corresponding to the contract template to obtain the scene category to which the current business scene belongs, obtain the preset audit rule corresponding to the scene category, and obtain the target audit rule.
  • Step S212 Obtain a signing intention judgment result based on the text matching score, fraud risk coefficient value, face recognition score, and target review rules.
  • the server obtains the text matching score and fraud risk coefficient value from the voice data and image data obtained from the terminal, and combines with the audit rules corresponding to the current business scenario to obtain the signing intention judgment result corresponding to the signing customer.
  • the classification of can be determined in advance, for example, “determine to sign a contract with one's own will”, “determine to sign a contract with a person other than one's will”, “suspect to sign a contract with non-personal will”.
  • the server may first compare the text matching score, fraud risk coefficient value, and face recognition score with the respective thresholds in the target review rule, and determine the text matching score, fraud risk coefficient value, and The signing intention judgment result corresponding to the face recognition score is based on the obtained three signing intention judgment results and the preset rules in the target review rules to obtain the final signing intention judgment result.
  • the text matching score is a1
  • the fraud risk coefficient value is a2
  • the face recognition score is a3.
  • the review rule corresponding to this business scenario is to match the text matching score a1 with the corresponding threshold A1 compares to obtain the first judgment result, compares the fraud risk coefficient a2 with the corresponding threshold to obtain the second judgment result, and compares the face recognition score a3 with the corresponding threshold A3 to obtain the third judgment result, when the three judgment results
  • the final result is determined to be "confirmed that they did not intend to sign a contract”
  • the final result is determined to be "I am willing to sign the contract" in addition to the above two situations, the final result is determined to be "suspected not to sign the contract.”
  • the text matching score, the fraud risk coefficient value, and the face recognition score may be sent to the preset terminal, respectively, according to the signing intention judgment result corresponding to the text matching score returned by the terminal, and the fraud risk coefficient value corresponding to The signing intent judgment result corresponding to the signing intent judgment result and the face recognition score corresponding to the signing intent judgment result and the preset rules in the target review rule get the final signing intent judgment result, for example, when the terminal returns the three corresponding to the fraud risk coefficient value of the signing contract
  • the result of the intent judgment is "I am willing to sign the contract”
  • the final result of the intent to sign the contract is determined to be "I confirm that I am willing to sign the contract.”
  • the final judgment result of the contract signing intention is "determined as non-intentional signing contract”.
  • the final result is "suspected non-self signing contract”.
  • the voice data and the image data are sent to the preset terminal for review.
  • the server obtains the corresponding voice text by acquiring voice data from the terminal, performing voice recognition on the voice data, extracting keywords from the voice text, and performing the preset keywords corresponding to the preset text on the extracted keywords Match, get the text matching score, extract the voice feature vector from the voice data, input the extracted voice feature vector into the trained fraud risk prediction model, get the fraud risk coefficient value, obtain image data from the terminal, and obtain the image data from the terminal.
  • extracting keywords from the speech text includes: segmenting the speech text to obtain a segmentation result; calculating feature weights for each word in the segmentation result, and sorting each word in the segmentation result according to the feature weights; Select keywords for sorting results.
  • the speech text can be divided into complete sentences according to punctuation marks, and then the segmentation of each segmented sentence can be performed.
  • the word segmentation method of string matching can be used to perform segmentation of each segmented sentence, such as positive To the maximum matching method, the word string in a segmented sentence is segmented from left to right; or, the reverse maximum matching method is to segment the string in a segmented sentence from right to left; or, the shortest Path word segmentation, the number of words required to be cut out in a string in a segmented sentence is the least; or, the two-way maximum matching method, which performs word segmentation matching in both forward and reverse directions.
  • Word sense word segmentation is a word segmentation method for machine speech judgment, using syntactic and semantic information to deal with ambiguity to segment words.
  • the server calculates feature weights for each word in the word segmentation result. Specifically, first calculate the word frequency TF of each word, which can be calculated by referring to the following formula:
  • Word Frequency TF The number of times a word appears in the document / total number of words in the document;
  • the server may sort each word in the word segmentation result according to the feature weights, and then select keywords according to the sorting result. For example, each word may be sorted in descending order according to the feature weight, and then a preset number of words that are ranked first are selected as keywords.
  • selecting keywords by calculating feature weights can make the selected keywords more accurate.
  • obtaining the scene identifier corresponding to the current business scene, and obtaining the corresponding audit rule according to the scene identifier includes:
  • Step S302 Search for the corresponding contract template and contract elements according to the scene identifier.
  • a contract template is set in advance.
  • the contract template refers to a template obtained by extracting the fixed format and / or fixed fields of an existing contract
  • the contract elements refer to the variables that make up the contract Field types, for example, in loan-related scenarios, contract elements can include borrowers, ID numbers, addresses, lenders, etc., writing contract elements into a contract template can get a blank electronic contract.
  • the scene identification establishes a mapping relationship with the contract template and the contract element respectively, and the corresponding contract template and contract element can be found according to the scene identification.
  • Step S304 Based on the contract elements, a scene classification model corresponding to the contract template is used to obtain a scene category corresponding to the current business scene.
  • the scene classification model is used to classify various business scenarios to obtain corresponding scene categories.
  • the scene categories are pre-defined according to needs.
  • a type of business scenario with the same audit rules can be divided into a scene category To get multiple scene categories.
  • the training step of the scene classification model includes: acquiring historical contract elements and corresponding scene categories corresponding to each contract template, using historical contract elements as input samples, and using corresponding scene categories as expected output samples for model training To get the scene classification model corresponding to each contract template.
  • supervised machine learning models can be used for training, such as SVM (Support Vector Machine, Support Vector Machine), logistic regression model, decision tree, etc.
  • model training least squares and gradients can be used Algorithms such as descent. It can be understood that in this embodiment, a scene classification model can be trained for each contract template, or a unified scene classification model can be trained for multiple contract models, such as multiple different formats but with the same or similar fixed fields.
  • the contract template trains a unified scene classification model.
  • Step S306 Acquire a preset audit rule corresponding to the scene category, and use the preset audit rule as the target audit rule.
  • a corresponding audit rule is set in advance, and when the server obtains the scenario category corresponding to the current business scenario, the server may use the audit rule corresponding to the scenario category as the target audit rule.
  • the scene classification corresponding to the current business scene is obtained through the scene classification model, and the audit rules corresponding to the scene categories are used as the target audit rules, which can improve the efficiency and accuracy of acquiring the target audit rules.
  • the scene classification model corresponding to the contract template is used based on the contract elements, and before the scene category corresponding to the current business scene is obtained, the method includes: obtaining personal information corresponding to the current contracted user ; Based on the contract elements, use the scene classification model corresponding to the contract template to obtain the scene category corresponding to the current business scene, including: based on personal information, the contract elements use the scene classification model corresponding to the contract template to obtain the scene category corresponding to the current business scene.
  • Personal information of contracted customers including gender, age, occupation, salary, etc.
  • extracting speech feature vectors from speech data includes: extracting speech features from speech data using MFCC ((Mel-scale FrequencyCepstral Coefficients)) to obtain corresponding feature parameters; The parameters are vectorized to obtain the corresponding speech feature vector.
  • MFCC (Mel-scale FrequencyCepstral Coefficients)
  • FFT Fast Fourier Transform
  • the group obtains the Mel spectrum, and performs cepstrum analysis on the Mel spectrum, including taking the logarithm, and implementing the inverse transform through DCT (Discrete, cosine, transform), and taking the 2nd to 13th coefficients after DCT as the characteristic coefficients
  • the feature coefficients are vectorized to obtain speech feature vectors.
  • the speech feature vector is extracted through MFCC, and the obtained speech feature vector can more accurately reflect the features of speech.
  • the generation step of the fraud risk prediction model includes: acquiring a preset amount of historical voice data and corresponding historical fraud risk coefficient values from the voice database; extracting historical voice feature vectors from the historical voice data; converting historical voice features The vector is used as an input sample, and the corresponding historical fraud risk coefficient value is used as the expected output sample for model training to obtain a trained fraud risk prediction model.
  • the historical voice data in the voice database is the voice data whose fraud risk coefficient value has been determined, so it can be used as a training sample for machine learning.
  • the historical voice data can be extracted from these historical voice data Feature vector, the extracted historical speech feature vector is used as the input sample during model training, and the corresponding fraud risk coefficient value is used as the expected output sample for model training.
  • the training process is the process of continuously adjusting the parameters of the model.
  • a stochastic gradient algorithm can be used for model training.
  • the cost function J ( ⁇ ) needs to be minimized.
  • the cost function can be expressed by the following formula:
  • m is the number of sample features in the training set
  • x (i) is the input historical speech feature vector
  • y (i) is the expected fraud risk coefficient value
  • h ⁇ (x (i) ) is the actual output of each training Value of the fraud risk coefficient of, where:
  • ⁇ T x is equal to the sum of the product of historical speech feature vectors and parameters.
  • steps in the flowcharts of FIGS. 2-3 are displayed in order according to the arrows, the steps are not necessarily executed in the order indicated by the arrows. Unless clearly stated in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in FIGS. 2-3 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed and completed at the same time, but may be executed at different times. These sub-steps or stages The execution order of is not necessarily sequential, but may be executed in turn or alternately with at least a part of other steps or sub-steps or stages of other steps.
  • a signing intention judgment device 400 including: a voice data acquisition module 402, a keyword matching module 404, a voice feature vector extraction module 406, an image data acquisition module 408, a target The audit rule acquisition module 410 and the signing intention judgment module 412, in which:
  • the voice data obtaining module 402 is used to obtain voice data from the terminal;
  • the keyword matching module 404 is used for performing speech recognition on the speech data to obtain corresponding speech text, extracting keywords from the speech text, matching the extracted keywords with the preset keywords corresponding to the preset text, and obtaining a text matching score;
  • the speech feature vector extraction module 406 is used to extract speech feature vectors from speech data, input the extracted speech feature vectors into the trained fraud risk prediction model, and obtain the fraud risk coefficient value;
  • the image data obtaining module 408 is used to obtain image data from the terminal, perform face recognition on the obtained image data, and obtain a face recognition score;
  • the target audit rule acquisition module 410 is used to acquire the scene identifier corresponding to the current business scene, and obtain the corresponding target audit rule according to the scene identifier;
  • the signing intention judgment module 412 is used to obtain the signing intention judgment result based on the text matching score, fraud risk coefficient value, face recognition score, and target review rules.
  • the keyword matching module 404 is also used to segment the speech text to obtain a segmentation result; calculate feature weights for each word in the segmentation result, and sort each word in the segmentation result according to the feature weight; according to the sorting Select keywords as a result.
  • the target review rule acquisition module 410 is used to find the corresponding contract template and contract elements according to the scene identifier; based on the contract elements, the scene classification model corresponding to the contract template is used to obtain the scene category corresponding to the current business scenario; Corresponding to the preset audit rules, the preset audit rules are used as the target audit rules.
  • the above device further includes: a personal information acquisition module for acquiring personal information corresponding to the currently contracted user; a target review rule acquisition module 410 is also used for adopting a scene classification model corresponding to the contract template based on personal information and contract elements To get the scene category corresponding to the current business scene.
  • the speech feature vector extraction module 406 is further used to extract speech features from the speech data using Mel cepstrum coefficients to obtain corresponding feature parameters; vectorizing the feature parameters to obtain corresponding speech feature vectors.
  • the above device further includes: a fraud risk prediction model generation module for acquiring a preset amount of historical voice data and corresponding historical fraud risk coefficient values from the voice database; extracting historical voice feature vectors from the historical voice data ; Use historical speech feature vectors as input samples, and use corresponding historical fraud risk coefficient values as expected output samples for model training to obtain a trained fraud risk prediction model.
  • a fraud risk prediction model generation module for acquiring a preset amount of historical voice data and corresponding historical fraud risk coefficient values from the voice database; extracting historical voice feature vectors from the historical voice data ; Use historical speech feature vectors as input samples, and use corresponding historical fraud risk coefficient values as expected output samples for model training to obtain a trained fraud risk prediction model.
  • Each module in the above-mentioned signing intention judgment device may be implemented in whole or in part by software, hardware, or a combination thereof.
  • the above modules may be embedded in the hardware or independent of the processor in the computer device, or may be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • a computer device is provided.
  • the computer device may be a server, and its internal structure may be shown in FIG. 5.
  • the computer device includes a processor, memory, network interface, and database connected by a system bus.
  • the processor of the computer device is used to provide computing and control capabilities.
  • the memory of the computer device includes a non-volatile computer-readable storage medium and internal memory.
  • the non-volatile computer-readable storage medium stores an operating system, computer-readable instructions, and a database.
  • the internal memory provides an environment for the operation of the operating system and computer-readable instructions in the non-volatile storage medium.
  • the database of the computer device is used to store data such as voice data, image data, and audit rules.
  • the network interface of the computer device is used to communicate with external terminals through a network connection. When the computer readable instructions are executed by the processor, a method for judging the signing intention is realized.
  • FIG. 5 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer equipment to which the solution of the present application is applied. Include more or less components than shown in the figure, or combine certain components, or have a different arrangement of components.
  • a computer device includes a memory and one or more processors.
  • the memory stores computer-readable instructions.
  • the one or more processors perform the following steps: obtain voice from the terminal Data; perform speech recognition on the speech data to obtain the corresponding speech text, extract keywords from the speech text, match the extracted keywords with the preset keywords corresponding to the preset text, and obtain a text matching score; extract speech from the speech data Feature vector, input the extracted voice feature vector into the trained fraud risk prediction model to obtain the fraud risk coefficient value; obtain image data from the terminal, perform face recognition on the obtained image data, and obtain a face recognition score; Obtain the scene identification corresponding to the current business scene, and obtain the corresponding target review rules according to the scene identification; obtain the signing intention judgment result based on the text matching score, fraud risk coefficient value, face recognition score, and target review rules.
  • extracting keywords from the speech text includes: segmenting the speech text to obtain a segmentation result; calculating feature weights for each word in the segmentation result, and sorting each word in the segmentation result according to the feature weights; Select keywords for sorting results.
  • obtaining the scene identifier corresponding to the current business scene and obtaining the corresponding audit rules according to the scene identifier include: searching for the corresponding contract template and contract elements according to the scene identifier; using the scene classification model corresponding to the contract template based on the contract elements, Obtain the scene category corresponding to the current business scene; obtain the preset audit rule corresponding to the scene category, and use the preset audit rule as the target audit rule.
  • the scene classification model corresponding to the contract template is used based on the contract elements, and before the scene category corresponding to the current business scene is obtained, the processor also implements the following steps when executing the computer-readable instructions: acquiring personal information corresponding to the current contracted user; Based on the contract elements, the scene classification model corresponding to the contract template is used to obtain the scene category corresponding to the current business scene, including: based on the personal information, the contract element uses the scene classification model corresponding to the contract template to obtain the scene category corresponding to the current business scene.
  • extracting the speech feature vector from the speech data includes: extracting the speech feature from the speech data using Mel cepstrum coefficients to obtain corresponding feature parameters; vectorizing the feature parameters to obtain the corresponding speech feature vector.
  • the processor also implements the following steps when executing computer-readable instructions: acquiring a preset amount of historical voice data and corresponding historical fraud risk coefficient values from the voice database; extracting historical voice feature vectors from the historical voice data; The historical speech feature vector is used as the input sample, and the corresponding historical fraud risk coefficient value is used as the expected output sample for model training to obtain the trained fraud risk prediction model.
  • One or more non-volatile computer-readable storage media storing computer-readable instructions.
  • the one or more processors perform the following steps: obtain from the terminal Voice data; perform voice recognition on the voice data to obtain the corresponding voice text, extract keywords from the voice text, match the extracted keywords with the preset keywords corresponding to the preset text, and obtain a text matching score; extract from the voice data Speech feature vector, input the extracted speech feature vector into the trained fraud risk prediction model to obtain the fraud risk coefficient value; obtain image data from the terminal, and perform face recognition on the obtained image data to obtain a face recognition score Obtain the scene identification corresponding to the current business scene, and obtain the corresponding target review rules according to the scene identification; obtain the signing intention judgment result based on the text matching score, fraud risk coefficient value, face recognition score, and target review rules.
  • extracting keywords from the speech text includes: segmenting the speech text to obtain a segmentation result; calculating feature weights for each word in the segmentation result, and sorting each word in the segmentation result according to the feature weights; Select keywords for sorting results.
  • obtaining the scene identifier corresponding to the current business scene and obtaining the corresponding audit rules according to the scene identifier include: searching for the corresponding contract template and contract elements according to the scene identifier; using the scene classification model corresponding to the contract template based on the contract elements, Obtain the scene category corresponding to the current business scene; obtain the preset audit rule corresponding to the scene category, and use the preset audit rule as the target audit rule.
  • the scene classification model corresponding to the contract template is used based on the contract elements, and before the scene category corresponding to the current business scene is obtained, the computer readable instructions are also executed by the processor to implement the following steps: obtain personal information corresponding to the current contracted user ; Based on the contract elements, use the scene classification model corresponding to the contract template to obtain the scene category corresponding to the current business scene, including: based on personal information, the contract elements use the scene classification model corresponding to the contract template to obtain the scene category corresponding to the current business scene.
  • extracting the speech feature vector from the speech data includes: extracting the speech feature from the speech data using Mel cepstrum coefficients to obtain corresponding feature parameters; vectorizing the feature parameters to obtain the corresponding speech feature vector.
  • the following steps are also implemented: obtaining a preset amount of historical voice data and corresponding historical fraud risk coefficient values from the voice database; and extracting historical voice feature vectors from the historical voice data ; Use historical speech feature vectors as input samples, and use corresponding historical fraud risk coefficient values as expected output samples for model training to obtain a trained fraud risk prediction model.
  • Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM random access memory
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDRSDRAM double data rate SDRAM
  • ESDRAM enhanced SDRAM
  • SLDRAM synchronous chain (Synchlink) DRAM
  • RDRAM direct RAM
  • DRAM direct memory bus dynamic RAM
  • RDRAM memory bus dynamic RAM

Abstract

一种签约意向判断方法,包括:从终端获取语音数据;对所述语音数据进行语音识别得到对应的语音文本,对所述语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;从所述语音数据中提取语音特征向量,将提取到的所述语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;从所述终端获取图像数据,对获取到的所述图像数据进行人脸识别,得到人脸识别分数;获取当前业务场景对应的场景标识,根据所述场景标识获取对应的目标审核规则;及根据所述文本匹配分数、所述欺诈风险系数值、所述人脸识别分数及所述目标审核规则得到签约意向判断结果。

Description

签约意向判断方法、装置、计算机设备和存储介质
相关申请的交叉引用
本申请要求于2018年10月16日提交中国专利局,申请号为2018112027200,申请名称为“签约意向判断方法、装置、计算机设备和存储介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及一种签约意向判断方法、装置、计算机设备和存储介质。
背景技术
现代社会中,需要进行签约的场景非常多,与办理保险业务、借贷业务、租赁业务时,都需要与保险公司、租赁公司等进行签约。
传统技术中,在签约前往往缺乏双录服务。应监管要求,现在大量签约场景需增加视频语音双录功能,然而,发明人意识到,目前的双录材料往往需要后续人工审核校验来判断签约人的签约意向,不仅效率低下而且准确性并不高。
发明内容
根据本申请公开的各种实施例,提供一种签约意向判断方法、装置、计算机设备和存储介质。
一种签约意向判断方法包括:
从终端获取语音数据;
对所述语音数据进行语音识别得到对应的语音文本,对所述语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;
从所述语音数据中提取语音特征向量,将提取到的所述语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;
从所述终端获取图像数据,对获取到的所述图像数据进行人脸识别,得到人脸识别分数;
获取当前业务场景对应的场景标识,根据所述场景标识获取对应的目标审核规则;及
根据所述文本匹配分数、所述欺诈风险系数值、所述人脸识别分数及所述目标审核规则得到签约意向判断结果。
一种签约意向判断装置包括:
语音数据获取模块,用于从终端获取语音数据;
关键词匹配模块,用于对所述语音数据进行语音识别得到对应的语音文本,对所述语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;
语音特征向量提取模块,用于从所述语音数据中提取语音特征向量,将提取到的所述语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;
图像数据获取模块,用于从所述终端获取图像数据,对获取到的所述图像数据进行人脸识别,得到人脸识别分数;
目标审核规则获取模块,用于获取当前业务场景对应的场景标识,根据所述场景标识获取对应的目标审核规则;及
签约意向判断模块,用于根据所述文本匹配分数、所述欺诈风险系数值、所述人脸识别分数及所述目标审核规则得到签约意向判断结果。
一种计算机设备,包括存储器和一个或多个处理器,存储器中存储有计算机可读指令,计算机可读指令被处理器执行时实现本申请任意一个实施例中提供的签约意向判断方法的步骤。
一个或多个存储有计算机可读指令的非易失性计算机可读存储介质,计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器实现本申请任意一个实施例中提供的签约意向判断方法的步骤。
本申请的一个或多个实施例的细节在下面的附图和描述中提出。本申请的其它特征和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更清楚地说明本申请实施例中的技术方案,下面将对实施例中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其它的附图。
图1为根据一个或多个实施例中签约意向判断方法的应用场景图。
图2为根据一个或多个实施例中签约意向判断方法的流程示意图。
图3为根据一个或多个实施例中步骤S210的流程示意图。
图4为根据一个或多个实施例中签约意向判断装置的框图。
图5为根据一个或多个实施例中计算机设备的框图。
具体实施方式
为了使本申请的技术方案及优点更加清楚明白,以下结合附图及实施例,对本申请进行进一步详细说明。应当理解,此处描述的具体实施例仅仅用以解释本申请,并不用于限定本申请。
本申请提供的签约意向判断方法,可以应用于如图1所示的应用环境中。终端102通过网络与服务器104通过网络进行通信,终端102用于对客户进行语音和视频的录制。服务器104首先从终端获取语音数据,然后对语音数据进行语音识别得到对应的语音文本,对语音文本提取关键词,将提取的关键词与预设对应的预设关键词进行匹配,得到文本匹配分数;接着从语音数据中提取语音特征向量,将提取到的语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;从终端获取图像数据,对获取到的图像数据进行人脸识别,得到人脸识别分数;获取当前业务场景对应的场景标识,根据场景标识获取对应的目标审核规则;根据文本匹配分数、欺诈风险系数值、人脸识别分数及审核规则得到签约意向判断结果。
终端102可以但不限于是各种个人计算机、笔记本电脑、智能手机、平板电脑和便携 式可穿戴设备,服务器104可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
在一些实施例中,如图2所示,提供了一种签约意向判断方法,以该方法应用于图1中的服务器为例进行说明,包括以下步骤:
步骤S202,从终端获取语音数据。
本实施例中,给定签约客户一段预设的文本,例如知情同意书等,客户朗读文本,终端对客户进行语音和视频的录制,服务器可以从终端获取录制的语音数据。具体来说,可以是在签约客户朗读完整个文本后,服务器获取终端上传的整个文本对应的语音数据;也可以是在签约客户朗读的过程中,服务器间隔一定的时间从终端获取语音数据,例如,可以约定签约用户每读完一句完整的句子,进行停顿,当终端检测到语音空白时段超过预设阈值时,将当前录制好的语音数据发送至服务器,当前录制好的语音数据指的是从上一次发送完语音数据开始录制的语音数据。
步骤S204,对语音数据进行语音识别得到对应的语音文本,对语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数。
具体地,预设文本指的是给定的朗读文本,是签约客户进行语音视频双录时需要朗读的原始文本。在本实施例中,对于预设文本,需要预先进行关键词提取,提取的关键词作为预设关键词。当服务器获取到终端录制的语音数据时,首先对语音数据进行预处理,包括降噪和语音增强等,然后对预处理后的语音数据进行语音识别得到语音文本,然后将得到的语音文本进行关键词提取,得到该语音文本对应的关键词,将这些关键词与预设关键词进行匹配,并计算匹配度,得到文本匹配分数。语音识别可采用现有技术中任意语音识别方法,本申请在此不再赘述。
在一些实施例中,对语音文本提取关键词,包括:对语音文本进行分词,对分词得到的各个词计算TF-IDF权重,然后根据TF-IDF权重对各个词进行降序排列,将排序靠前的预设数量的词作为关键词。
在一些实施例中,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数,包括:首先计算语音文本对应的关键词中与预设关键词相匹配的词数量,根据匹配上的词数量与预设关键词的总词数的比值得到文本匹配分数。举例来说,语音文本 对应的关键词为A、B、D,预设关键词包括A、B、C、D、E,则匹配词数量为3,文本匹配分数为:3/5x100=60。
步骤S206,从语音数据中提取语音特征向量,将提取到的语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值。
欺诈风险预测模型用于根据语音数据对签约客户的欺诈风险系数值进行预测,欺诈风险系数值用于表征签约客户可能存在的欺诈风险,欺诈风险系数值越大,表示签约客户存在的欺诈风险越大。欺诈风险预测模型可基于历史数据通过有监督的机器学习模型训练得到,机器学习模型包括但不限于SVM(Support Vector Machine,支持向量机)、逻辑回归模型、决策树等等。
在一些实施例中,服务器获取到语音数据后,可以利用MFCC(Mel-Frequency Cepstral Coefficients,梅尔倒谱系数)对语音数据进行语音特征提取,得到对应的特征系数,将特征系数矢量化,得到对应的语音特征向量。
步骤S208,从终端获取图像数据,对获取到的图像数据进行人脸识别,得到人脸识别分数。
具体地,服务器可以按照预设的时间间隔从终端获取一帧或者连续几帧图像数据,对获取到的图像数据分别进行人脸识别,得到多个人脸识别分数,然后对得到的多个人脸识别分数取平均值,得到最终的人脸识别分数。具体来说,服务器对获取到的图像数据进行人脸识别时,首先对图片数据进行人脸检测,得到人脸图像,然后对得到的人脸图像进行预处理,预处理过程主要包括对人脸图像的光线补偿、灰度变换、直方图均衡化、归一化、几何校正、滤波以及锐化等,接着对预处理后的人脸图像进行特征提取,最后将提取到的特征数据与该签约用户预先保存的人脸图像的特征进行比对,签约用户预先保存的人脸图像可以是身份证图像,或者是通过调用公安部人脸识别接口进行人脸识别时采集的图像。
步骤S210,获取当前业务场景对应的场景标识,根据场景标识获取对应的目标审核规则。
具体地,业务场景指的是签约时的场景,包括但不限于是与贷款、租赁、各种保险购买、信用卡办理等相关的场景,当前业务场景指的是签约客户当前签约对应的业务场景, 场景标识用于唯一标识业务场景。审核规则指的是用于判断客户签约意向的规则,不同的场景类别,审核规则不同。服务器获取当前业务场景对应的场景标识,可以是终端在进行语音视频双录之前向服务器发送的,也可以是在双录结束后,服务器主动从终端获取的。
进一步,在一些实施例中,服务器获取到场景标识后,可首先根据场景标识去数据库中进行查找,如果当前业务场景为已经出现过的业务场景,其对应的审核规则已保存在数据库中,因此可以直接根据场景标识查找对应的审核规则得到目标审核规则;在另一些实施例中,当前业务场景为第一次出现的业务场景,可根据场景标识查找到对应的合同模板及合同要素,基于查找到的合同要素采用合同模板对应的场景分类模型得到当前业务场景所属的场景类别,获取该场景类别对应的预设审核规则,得到目标审核规则。
步骤S212,根据文本匹配分数、欺诈风险系数值、人脸识别分数及目标审核规则得到签约意向判断结果。
具体地,服务器通过从终端获取到的语音数据、图像数据分别得到文本匹配分数、欺诈风险系数值后,结合当前业务场景对应的审核规则可以得到签约客户对应的签约意向判断结果,签约意向判断结果的分类可以事先确定,例如可以是,“确定为本人意愿签约”、“确定为非本人意愿签约”、“疑似非本人意愿签约”。
在一些实施例中,服务器可以先将文本匹配分数、欺诈风险系数值、人脸识别分数分别与目标审核规则中各自对应的阈值进行比较,根据比较结果分别确定文本匹配分数、欺诈风险系数值、人脸识别分数对应的签约意向判断结果,根据得到的三个签约意向判断结果及目标审核规则中的预设规则得到最终的签约意向判断结果。
举个例子,在车辆租赁对应的业务场景中,文本匹配分数为a1、欺诈风险系数值为a2、人脸识别分数a3,该业务场景对应的审核规则为,将文本匹配分数a1与对应的阈值A1进行比较得到第一判断结果,将欺诈风险系数a2与对应的阈值进行比较得到第二判断结果,将人脸识别分数a3与对应的阈值A3进行比较得到第三判断结果,当三个判断结果均为“确定为非本人意愿签约”时,判定最终的结果为“确定为非本人意愿签约”,当三个判断结果均为“确定为本人意愿签约”时,判定最终的结果为“确定为本人意愿签约”,除上述两种情形外的情形,判定最终的结果为“疑似非本人意愿签约”。
在另一些实施例中,可以是将文本匹配分数、欺诈风险系数值、人脸识别分数分别 发送至预设的终端,根据终端返回的文本匹配分数对应的签约意向判断结果、欺诈风险系数值对应的签约意向判断结果及人脸识别分数对应的签约意向判断结果及目标审核规则中的预设规则得到最终的签约意向判断结果,例如,当终端返回的三者对应的欺诈风险系数值对应的签约意向判断结果为“本人意愿签约”时,判定最终的签约意向判断结果为“确定为本人意愿签约”,当终端返回的三者对应的欺诈风险系数值对应的签约意向判断结果均为“非本人意愿签约”时,判定最终的签约意向判断结果为“确定为非本人意向签约”,除上述两种情形外的情形,判定最终的结果为“疑似非本人意愿签约”。
在一些实施例中,当最终的结果为“疑似非本人意愿签约”时,将语音数据、图像数据一起发送至预设的终端进行复核。
上述签约意向判断方法中,服务器通过从终端获取语音数据,对语音数据进行语音识别得到对应的语音文本,对语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数,从语音数据中提取语音特征向量,将提取到的语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值,从终端获取图像数据,对获取到的图像数据进行人脸识别,得到人脸识别分数,获取当前业务场景对应的场景标识,根据场景标识获取对应的目标审核规则,根据文本匹配分数、欺诈风险系数值、人脸识别分数及目标审核规则得到签约意向判断结果,通过本申请的方法,可以实现双录视频的自动审核,提高了签约意向的判断效率和准确性。
在一些实施例中,对语音文本提取关键词,包括:对语音文本进行分词,得到分词结果;对分词结果中的各个词计算特征权重,根据特征权重对分词结果中的各个词进行排序;根据排序结果选取关键词。
具体地,可首先根据标点符号将语音文本分成一条条完整的语句,再对各个切分的语句进行分词处理,如可利用字符串匹配的分词方法对各个切分的语句进行分词处理,如正向最大匹配法,把一个切分的语句中的字符串从左至右来分词;或者,反向最大匹配法,把一个切分的语句中的字符串从右至左来分词;或者,最短路径分词法,一个切分的语句中的字符串里面要求切出的词数是最少的;或者,双向最大匹配法,正反向同时进行分词匹配。还可利用词义分词法对各个切分的语句进行分词处理,词义分词法是一种机器语音判断的分词方法,利用句法信息和语义信息来处理歧义现象来分词。
进一步,服务器对分词结果中的各个词计算特征权重。具体地,首先计算各个词的词频TF,可参考如下公式进行计算:
词频TF=某个词在文档中出现的次数/文档的总词数;
然后,计算各个词的逆文档词频IDF,可参考如下公式进行计算:
逆文档词频IDF=错误!未找到引用源。);
最后,计算词频TF与逆文档词频IDF的乘积得到特征权重。
进一步,计算得到特征权重后,服务器根据特征权重可以对分词结果中各个词进行排序,然后根据排序结果选取关键词。例如可以根据特征权重对各个词进行降序排列,然后选取排序靠前的预设数量的词作为关键词。
上述实施例中,通过计算特征权重选取关键词,可以使得选择的关键词更加准确。
在一些实施例中,如图3所示,获取当前业务场景对应的场景标识,根据场景标识获取对应的审核规则,包括:
步骤S302,根据场景标识查找对应的合同模板及合同要素。
具体地,客户在进行签约之前需要先确定待签约的电子合同。在本实施例中,为提高电子合同的生成效率,预先设置了合同模板,合同模板指的是抽取已有合同的固定格式和/或固定字段得到的模板,合同要素指的是组成合同的变量字段类型,例如,在与贷款有关的场景中,合同要素可以包括借款人、身份证号、住址、贷款人等等,将合同要素写入合同模板中可以得到空白的电子合同。本实施例中,场景标识分别与合同模板、合同要素建立了映射关系,根据场景标识可以查找到对应的合同模板及合同要素。
步骤S304,基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。
具体地,场景分类模型用于对各个业务场景进行分类,得到对应的场景类别,场景类别根据需要进行事先划定,本实施例中,可以将审核规则相同的一类业务场景划分为一个场景类别,得到多个场景类别。
在一些实施例中,场景分类模型的训练步骤包括:获取各个合同模板对应的历史合同要素及对应的场景类别,将历史合同要素作为输入样本,将对应的场景类别作为期望的输出样本进行模型训练,得到各个合同模板对应的场景分类模型。在进行模型训练时,可 采用有监督的机器学习模型进行训练,例如SVM(Support Vector Machine,支持向量机)、逻辑回归模型、决策树等,在进行模型训练时,可采用最小二乘法和梯度下降等算法。可以理解,本实施例中,可以分别针对每一个合同模板训练一个场景分类模型,也可以对多个合同模型训练一个统一的场景分类模型,如将格式不同、但固定字段相同或相近的多个合同模板训练一个统一的场景分类模型。
步骤S306,获取场景类别对应的预设审核规则,将预设审核规则作为目标审核规则。
具体地,对于每一个场景类别,事先设定了对应的审核规则,服务器在获取到当前业务场景对应的场景类别时,可以将该场景类别对应的审核规则作为目标审核规则。
上述实施例中,通过场景分类模型来获取当前业务场景对应的场景类别,将场景类别对应的审核规则作为目标审核规则,可以提高目标审核规则获取的效率和准确率。
在一些实施例中,为进一步确保获取的审核规则的准确性,基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别之前,方法包括:获取当前签约用户对应的个人信息;基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别,包括:基于个人信息、合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。签约客户的个人信息,包括性别、年龄、职业、薪资等。
在一些实施例中,从语音数据中提取语音特征向量,包括:利用MFCC((Mel-scale Frequency Cepstral Coefficients,梅尔倒谱系数)对语音数据进行语音特征提取,得到对应的特征参数;将特征参数矢量化,得到对应的语音特征向量。
具体地,首先对先对语音进行预加重、分帧和加窗,对每一个短时分析窗,通过FFT(Fast Fourier Transform,快速傅利叶转换)得到对应的频谱,将得到的频谱通过Mel滤波器组得到Mel频谱,在Mel频谱上面进行倒谱分析,包括取对数,通过DCT(Discrete cosine transform,离散余弦变换)来实现逆变换,取DCT后的第2个到第13个系数作为特征系数,进一步,将特征系数矢量化,得到语音特征向量。
上述实施例中,通过MFCC提取语音特征向量,得到的语音特征向量能够更加准确的反映语音的特征。
在一些实施例中,欺诈风险预测模型的生成步骤包括:从语音数据库中获取预设数 量的历史语音数据及对应的历史欺诈风险系数值;对历史语音数据提取历史语音特征向量;将历史语音特征向量作为输入样本,将对应的历史欺诈风险系数值作为期望的输出样本进行模型训练,得到已训练的欺诈风险预测模型。
具体地,语音数据库中的历史语音数据为欺诈风险系数值已经确定的语音数据,因此可作为机器学习的训练样本,服务器语音数据库中选取到历史语音数据后,可对这些历史语音数据提取历史语音特征向量,将提取的历史语音特征向量作为模型训练时的输入样本,将其对应的欺诈风险系数值作为期望的输出样本进行模型训练,训练的过程就是对模型的参数进行不断调整的过程。在进行模型训练时,可以采用有监督的模型训练方式,例如逻辑回归模型、SVM(Support Vector Machine,支持向量机)以及贝叶斯模型等。
在一些实施例中,以SVM为例,可以采用随机梯度算法进行模型训练,在梯度下降过程中需要使得代价函数J(θ)最小,在一些实施例中,代价函数可以采用以下公式表示:
Figure PCTCN2019070814-appb-000001
m表示训练集中样本特征的个数,x (i)为输入的历史语音特征向量,y (i)表示期望输出的欺诈风险系数值,h θ(x (i))表示每次训练的实际输出的欺诈风险系数值,其中:
Figure PCTCN2019070814-appb-000002
错误!未找到引用源。,即θ Tx等于历史语音特征向量与参数的乘积的和。
应该理解的是,虽然图2-3的流程图中的各个步骤按照箭头的指示依次显示,但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明,这些步骤的执行并没有严格的顺序限制,这些步骤可以以其它的顺序执行。而且,图2-3中的至少一部分步骤可以包括多个子步骤或者多个阶段,这些子步骤或者阶段并不必然是在同一时刻执行完成,而是可以在不同的时刻执行,这些子步骤或者阶段的执行顺序也不必然是依次进行,而是可以与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
在一些实施例中,如图4所示,提供了一种签约意向判断装置400,包括:语音数据获取模块402、关键词匹配模块404、语音特征向量提取模块406、图像数据获取模块408、目标审核规则获取模块410和签约意向判断模块412,其中:
语音数据获取模块402用于从终端获取语音数据;
关键词匹配模块404用于对语音数据进行语音识别得到对应的语音文本,对语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;
语音特征向量提取模块406用于从语音数据中提取语音特征向量,将提取到的语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;
图像数据获取模块408用于从终端获取图像数据,对获取到的图像数据进行人脸识别,得到人脸识别分数;
目标审核规则获取模块410用于获取当前业务场景对应的场景标识,根据场景标识获取对应的目标审核规则;
签约意向判断模块412用于根据文本匹配分数、欺诈风险系数值、人脸识别分数及目标审核规则得到签约意向判断结果。
在一些实施例中,关键词匹配模块404还用于对语音文本进行分词,得到分词结果;对分词结果中的各个词计算特征权重,根据特征权重对分词结果中的各个词进行排序;根据排序结果选取关键词。
在一些实施例中,目标审核规则获取模块410用于根据场景标识查找对应的合同模板及合同要素;基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别;获取场景类别对应的预设审核规则,将预设审核规则作为目标审核规则。
在一些实施例中,上述装置还包括:个人信息获取模块,用于获取当前签约用户对应的个人信息;目标审核规则获取模块410还用于基于个人信息、合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。
在一些实施例中,语音特征向量提取模块406还用于利用梅尔倒谱系数对语音数据进行语音特征提取,得到对应的特征参数;将特征参数矢量化,得到对应的语音特征向量。
在一些实施例中,上述装置还包括:欺诈风险预测模型生成模块,用于从语音数据库 中获取预设数量的历史语音数据及对应的历史欺诈风险系数值;对历史语音数据提取历史语音特征向量;将历史语音特征向量作为输入样本,将对应的历史欺诈风险系数值作为期望的输出样本进行模型训练,得到已训练的欺诈风险预测模型。
关于签约意向判断装置的具体限定可以参见上文中对于签约意向判断方法的限定,在此不再赘述。上述签约意向判断装置中的各个模块可全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中,也可以以软件形式存储于计算机设备中的存储器中,以便于处理器调用执行以上各个模块对应的操作。
在一些实施例中,提供了一种计算机设备,该计算机设备可以是服务器,其内部结构图可以如图5所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。该计算机设备的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性计算机可读存储介质、内存储器。该非易失性计算机可读存储介质存储有操作系统、计算机可读指令和数据库。该内存储器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储语音数据、图像数据、审核规则等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令被处理器执行时以实现一种签约意向判断方法。
本领域技术人员可以理解,图5中示出的结构,仅仅是与本申请方案相关的部分结构的框图,并不构成对本申请方案所应用于其上的计算机设备的限定,具体的计算机设备可以包括比图中所示更多或更少的部件,或者组合某些部件,或者具有不同的部件布置。
一种计算机设备,包括存储器和一个或多个处理器,存储器中储存有计算机可读指令,计算机可读指令被处理器执行时,使得一个或多个处理器执行以下步骤::从终端获取语音数据;对语音数据进行语音识别得到对应的语音文本,对语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;从语音数据中提取语音特征向量,将提取到的语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;从终端获取图像数据,对获取到的图像数据进行人脸识别,得到人脸识别分 数;获取当前业务场景对应的场景标识,根据场景标识获取对应的目标审核规则;根据文本匹配分数、欺诈风险系数值、人脸识别分数及目标审核规则得到签约意向判断结果。
在一些实施例中,对语音文本提取关键词,包括:对语音文本进行分词,得到分词结果;对分词结果中的各个词计算特征权重,根据特征权重对分词结果中的各个词进行排序;根据排序结果选取关键词。
在一些实施例中,获取当前业务场景对应的场景标识,根据场景标识获取对应的审核规则,包括:根据场景标识查找对应的合同模板及合同要素;基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别;获取场景类别对应的预设审核规则,将预设审核规则作为目标审核规则。
在一些实施例中,基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别之前,处理器执行计算机可读指令时还实现以下步骤:获取当前签约用户对应的个人信息;基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别,包括:基于个人信息、合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。
在一些实施例中,从语音数据中提取语音特征向量,包括:利用梅尔倒谱系数对语音数据进行语音特征提取,得到对应的特征参数;将特征参数矢量化,得到对应的语音特征向量。
在一些实施例中,处理器执行计算机可读指令时还实现以下步骤:从语音数据库中获取预设数量的历史语音数据及对应的历史欺诈风险系数值;对历史语音数据提取历史语音特征向量;将历史语音特征向量作为输入样本,将对应的历史欺诈风险系数值作为期望的输出样本进行模型训练,得到已训练的欺诈风险预测模型。
一个或多个存储有计算机可读指令的非易失性计算机可读存储介质,计算机可读指令被一个或多个处理器执行时,使得一个或多个处理器执行以下步骤::从终端获取语音数据;对语音数据进行语音识别得到对应的语音文本,对语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;从语音数据中提取语音特征向量,将提取到的语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险 系数值;从终端获取图像数据,对获取到的图像数据进行人脸识别,得到人脸识别分数;获取当前业务场景对应的场景标识,根据场景标识获取对应的目标审核规则;根据文本匹配分数、欺诈风险系数值、人脸识别分数及目标审核规则得到签约意向判断结果。
在一些实施例中,对语音文本提取关键词,包括:对语音文本进行分词,得到分词结果;对分词结果中的各个词计算特征权重,根据特征权重对分词结果中的各个词进行排序;根据排序结果选取关键词。
在一些实施例中,获取当前业务场景对应的场景标识,根据场景标识获取对应的审核规则,包括:根据场景标识查找对应的合同模板及合同要素;基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别;获取场景类别对应的预设审核规则,将预设审核规则作为目标审核规则。
在一些实施例中,基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别之前,计算机可读指令被处理器执行时还实现以下步骤:获取当前签约用户对应的个人信息;基于合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别,包括:基于个人信息、合同要素采用合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。
在一些实施例中,从语音数据中提取语音特征向量,包括:利用梅尔倒谱系数对语音数据进行语音特征提取,得到对应的特征参数;将特征参数矢量化,得到对应的语音特征向量。
在一些实施例中,计算机可读指令被处理器执行时还实现以下步骤:从语音数据库中获取预设数量的历史语音数据及对应的历史欺诈风险系数值;对历史语音数据提取历史语音特征向量;将历史语音特征向量作为输入样本,将对应的历史欺诈风险系数值作为期望的输出样本进行模型训练,得到已训练的欺诈风险预测模型。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,是可以通过计算机可读指令来指令相关的硬件来完成,所述的计算机可读指令可存储于一非易失性计算机可读取存储介质中,该计算机可读指令在执行时,可包括如上述各方法的实施例的流程。本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用,均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、 可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限,RAM以多种形式可得,诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上实施例的各技术特征可以进行任意的组合,为使描述简洁,未对上述实施例中的各个技术特征所有可能的组合都进行描述,然而,只要这些技术特征的组合不存在矛盾,都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式,其描述较为具体和详细,但并不能因此而理解为对发明专利范围的限制。应当指出的是,对于本领域的普通技术人员来说,在不脱离本申请构思的前提下,还可以做出若干变形和改进,这些都属于本申请的保护范围。因此,本申请专利的保护范围应以所附权利要求为准。

Claims (20)

  1. 一种签约意向判断方法,包括:
    从终端获取语音数据;
    对所述语音数据进行语音识别得到对应的语音文本,对所述语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;
    从所述语音数据中提取语音特征向量,将提取到的所述语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;
    从所述终端获取图像数据,对获取到的所述图像数据进行人脸识别,得到人脸识别分数;
    获取当前业务场景对应的场景标识,根据所述场景标识获取对应的目标审核规则;及
    根据所述文本匹配分数、所述欺诈风险系数值、所述人脸识别分数及所述目标审核规则得到签约意向判断结果。
  2. 根据权利要求1所述的方法,其特征在于,所述对所述语音文本提取关键词,包括:
    对所述语音文本进行分词,得到分词结果;
    对所述分词结果中的各个词计算特征权重,根据所述特征权重对所述分词结果中的各个词进行排序;及
    根据排序结果选取关键词。
  3. 根据权利要求1所述的方法,其特征在于,所述获取当前业务场景对应的场景标识,根据所述场景标识获取对应的审核规则,包括:
    根据所述场景标识查找对应的合同模板及合同要素;
    基于所述合同要素采用所述合同模板对应的场景分类模型,得到所述当前业务场景对应的场景类别;及
    获取所述场景类别对应的预设审核规则,将所述预设审核规则作为目标审核规则。
  4. 根据权利要求3所述的方法,其特征在于,在所述基于所述合同要素采用所述合同模板对应的场景分类模型,得到当前业务场景对应的场景类别之前,所述方法包括:
    获取当前签约用户对应的个人信息;及
    所述基于所述合同要素采用所述合同模板对应的场景分类模型,得到当前业务场景对应的场景类别,包括:
    基于所述个人信息、所述合同要素采用所述合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。
  5. 根据权利要求1所述的方法,其特征在于,所述从所述语音数据中提取语音特征向量,包括:
    利用梅尔倒谱系数对所述语音数据进行语音特征提取,得到对应的特征参数;及
    将所述特征参数矢量化,得到对应的语音特征向量。
  6. 根据权利要求1至5任意一项所述的方法,其特征在于,所述欺诈风险预测模型的生成步骤包括:
    从语音数据库中获取预设数量的历史语音数据及对应的历史欺诈风险系数值;
    对所述历史语音数据提取历史语音特征向量;及
    将所述历史语音特征向量作为输入样本,将对应的所述历史欺诈风险系数值作为期望的输出样本进行模型训练,得到已训练的欺诈风险预测模型。
  7. 一种签约意向判断装置,包括:
    语音数据获取模块,用于从终端获取语音数据;
    关键词匹配模块,用于对所述语音数据进行语音识别得到对应的语音文本,对所述语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;
    语音特征向量提取模块,用于从所述语音数据中提取语音特征向量,将提取到的所述语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;
    图像数据获取模块,用于从所述终端获取图像数据,对获取到的所述图像数据进行人脸识别,得到人脸识别分数;
    目标审核规则获取模块,用于获取当前业务场景对应的场景标识,根据所述场景标识获取对应的目标审核规则;及
    签约意向判断模块,用于根据所述文本匹配分数、所述欺诈风险系数值、所述人脸识 别分数及所述目标审核规则得到签约意向判断结果。
  8. 根据权利要求7所述的装置,其特征在于,目标审核规则获取模块用于根据所述场景标识查找对应的合同模板及合同要素;基于所述合同要素采用所述合同模板对应的场景分类模型,得到所述当前业务场景对应的场景类别;获取所述场景类别对应的预设审核规则,将所述预设审核规则作为目标审核规则。
  9. 一种计算机设备,包括存储器及一个或多个处理器,所述存储器中储存有计算机可读指令,所述计算机可读指令被所述一个或多个处理器执行时,使得所述一个或多个处理器执行以下步骤:
    从终端获取语音数据;
    对所述语音数据进行语音识别得到对应的语音文本,对所述语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;
    从所述语音数据中提取语音特征向量,将提取到的所述语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;
    从所述终端获取图像数据,对获取到的所述图像数据进行人脸识别,得到人脸识别分数;
    获取当前业务场景对应的场景标识,根据所述场景标识获取对应的目标审核规则;及
    根据所述文本匹配分数、所述欺诈风险系数值、所述人脸识别分数及所述目标审核规则得到签约意向判断结果。
  10. 根据权利要求9所述的计算机设备,其特征在于,所述处理器执行所述计算机可读指令时还执行以下步骤:
    对所述语音文本进行分词,得到分词结果;
    对所述分词结果中的各个词计算特征权重,根据所述特征权重对所述分词结果中的各个词进行排序;及
    根据排序结果选取关键词。
  11. 根据权利要求9所述的计算机设备,其特征在于,所述处理器执行所述计算机可读指令时还执行以下步骤:
    根据所述场景标识查找对应的合同模板及合同要素;
    基于所述合同要素采用所述合同模板对应的场景分类模型,得到所述当前业务场景对应的场景类别;及
    获取所述场景类别对应的预设审核规则,将所述预设审核规则作为目标审核规则。
  12. 根据权利要求11所述的计算机设备,其特征在于,所述处理器执行所述计算机可读指令时还执行以下步骤:
    获取当前签约用户对应的个人信息;及
    基于所述个人信息、所述合同要素采用所述合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。
  13. 根据权利要求9所述的计算机设备,其特征在于,所述处理器执行所述计算机可读指令时还执行以下步骤:
    利用梅尔倒谱系数对所述语音数据进行语音特征提取,得到对应的特征参数;及
    将所述特征参数矢量化,得到对应的语音特征向量。
  14. 根据权利要求9至13任意一项所述的计算机设备,其特征在于,所述处理器执行所述计算机可读指令时还执行以下步骤:
    从语音数据库中获取预设数量的历史语音数据及对应的历史欺诈风险系数值;
    对所述历史语音数据提取历史语音特征向量;及
    将所述历史语音特征向量作为输入样本,将对应的所述历史欺诈风险系数值作为期望的输出样本进行模型训练,得到已训练的欺诈风险预测模型。
  15. 一个或多个存储有计算机可读指令的非易失性计算机可读存储介质,所述计算机可读指令被一个或多个处理器执行时,使得所述一个或多个处理器执行以下步骤:
    从终端获取语音数据;
    对所述语音数据进行语音识别得到对应的语音文本,对所述语音文本提取关键词,将提取的关键词与预设文本对应的预设关键词进行匹配,得到文本匹配分数;
    从所述语音数据中提取语音特征向量,将提取到的所述语音特征向量输入到已训练的欺诈风险预测模型中,得到欺诈风险系数值;
    从所述终端获取图像数据,对获取到的所述图像数据进行人脸识别,得到人脸识别分 数;
    获取当前业务场景对应的场景标识,根据所述场景标识获取对应的目标审核规则;及
    根据所述文本匹配分数、所述欺诈风险系数值、所述人脸识别分数及所述目标审核规则得到签约意向判断结果。
  16. 根据权利要求15所述的存储介质,其特征在于,所述计算机可读指令被所述处理器执行时还执行以下步骤:
    对所述语音文本进行分词,得到分词结果;
    对所述分词结果中的各个词计算特征权重,根据所述特征权重对所述分词结果中的各个词进行排序;及
    根据排序结果选取关键词。
  17. 根据权利要求15所述的存储介质,其特征在于,所述计算机可读指令被所述处理器执行时还执行以下步骤:
    根据所述场景标识查找对应的合同模板及合同要素;
    基于所述合同要素采用所述合同模板对应的场景分类模型,得到所述当前业务场景对应的场景类别;及
    获取所述场景类别对应的预设审核规则,将所述预设审核规则作为目标审核规则。
  18. 根据权利要求17所述的存储介质,其特征在于,所述计算机可读指令被所述处理器执行时还执行以下步骤:
    获取当前签约用户对应的个人信息;及
    基于所述个人信息、所述合同要素采用所述合同模板对应的场景分类模型,得到当前业务场景对应的场景类别。
  19. 根据权利要求15所述的存储介质,其特征在于,所述计算机可读指令被所述处理器执行时还执行以下步骤:
    利用梅尔倒谱系数对所述语音数据进行语音特征提取,得到对应的特征参数;及
    将所述特征参数矢量化,得到对应的语音特征向量。
  20. 根据权利要求15至19任意一项所述的存储介质,其特征在于,所述计算机可读指令被所述处理器执行时还执行以下步骤:
    从语音数据库中获取预设数量的历史语音数据及对应的历史欺诈风险系数值;
    对所述历史语音数据提取历史语音特征向量;及
    将所述历史语音特征向量作为输入样本,将对应的所述历史欺诈风险系数值作为期望的输出样本进行模型训练,得到已训练的欺诈风险预测模型。
PCT/CN2019/070814 2018-10-16 2019-01-08 签约意向判断方法、装置、计算机设备和存储介质 WO2020077895A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201811202720.0 2018-10-16
CN201811202720.0A CN109543516A (zh) 2018-10-16 2018-10-16 签约意向判断方法、装置、计算机设备和存储介质

Publications (1)

Publication Number Publication Date
WO2020077895A1 true WO2020077895A1 (zh) 2020-04-23

Family

ID=65843802

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/070814 WO2020077895A1 (zh) 2018-10-16 2019-01-08 签约意向判断方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN109543516A (zh)
WO (1) WO2020077895A1 (zh)

Cited By (29)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111581344A (zh) * 2020-04-26 2020-08-25 腾讯科技(深圳)有限公司 一种接口信息审核方法、装置、计算机设备及存储介质
CN111753516A (zh) * 2020-06-29 2020-10-09 平安国际智慧城市科技股份有限公司 文本查重处理方法、装置、计算机设备及计算机存储介质
CN111984779A (zh) * 2020-09-10 2020-11-24 支付宝(杭州)信息技术有限公司 一种对话文本分析方法、装置、设备和可读介质
CN112015869A (zh) * 2020-09-14 2020-12-01 支付宝(杭州)信息技术有限公司 一种待发布文本的风险检测方法、装置及设备
CN112015747A (zh) * 2020-09-14 2020-12-01 支付宝(杭州)信息技术有限公司 数据上传方法及装置
CN112037796A (zh) * 2020-09-08 2020-12-04 平安科技(深圳)有限公司 一种数据处理方法、装置、设备及介质
CN112100379A (zh) * 2020-09-15 2020-12-18 北京百度网讯科技有限公司 用于处理数据的方法、装置、设备以及存储介质
CN112100357A (zh) * 2020-09-24 2020-12-18 腾讯科技(深圳)有限公司 引导语的生成方法及装置、电子设备、计算机存储介质
CN112215087A (zh) * 2020-09-21 2021-01-12 浙江数链科技有限公司 图片审核方法、装置、计算机设备和存储介质
CN112257425A (zh) * 2020-09-29 2021-01-22 国网天津市电力公司 一种基于数据分级模型的电力数据分析方法及系统
CN112307235A (zh) * 2020-05-09 2021-02-02 支付宝(杭州)信息技术有限公司 前端页面元素的命名方法、装置及电子设备
CN112416476A (zh) * 2020-11-25 2021-02-26 武汉联影医疗科技有限公司 工作流执行方法、装置、计算机设备和存储介质
CN112492343A (zh) * 2020-12-16 2021-03-12 浙江大华技术股份有限公司 一种视频直播监控方法及相关装置
CN112507936A (zh) * 2020-12-16 2021-03-16 平安银行股份有限公司 图像信息审核方法、装置、电子设备及可读存储介质
CN112508717A (zh) * 2020-12-01 2021-03-16 中国人寿保险股份有限公司 一种影像信息的审核方法、装置、电子设备及存储介质
CN112685594A (zh) * 2020-12-24 2021-04-20 中国人民解放军战略支援部队信息工程大学 基于注意力的弱监督语音检索方法及系统
CN112861662A (zh) * 2021-01-22 2021-05-28 平安科技(深圳)有限公司 基于人脸和交互文本的目标对象行为预测方法及相关设备
CN113094520A (zh) * 2021-03-29 2021-07-09 四川新网银行股份有限公司 一种检查电子合同的方法
CN113362072A (zh) * 2021-06-30 2021-09-07 平安普惠企业管理有限公司 风控数据处理方法、装置、电子设备及存储介质
CN113723759A (zh) * 2021-07-30 2021-11-30 北京淇瑀信息科技有限公司 基于设备意向度和设备风险度为设备提供互联网服务的方法及装置
CN113793597A (zh) * 2021-09-15 2021-12-14 云知声智能科技股份有限公司 一种语音识别方法、装置、电子设备和存储介质
CN113806505A (zh) * 2021-09-09 2021-12-17 科大讯飞股份有限公司 要素比对方法、装置、电子设备和存储介质
CN114140140A (zh) * 2020-09-03 2022-03-04 中国移动通信集团浙江有限公司 一种场景筛选方法、装置及设备
CN114282882A (zh) * 2021-11-15 2022-04-05 深圳市东信时代信息技术有限公司 素材工单审核方法、装置、设备及存储介质
CN114640606A (zh) * 2020-12-01 2022-06-17 中移物联网有限公司 一种物联网卡终端的异常处理方法及控制器
CN115526605A (zh) * 2022-10-21 2022-12-27 金恒智控管理咨询集团股份有限公司 基于企业内部控制管理的审批方法及系统
CN116405300A (zh) * 2023-04-18 2023-07-07 无锡锡商银行股份有限公司 一种基于场景的在线协议签署安全分析系统及方法
CN117313723A (zh) * 2023-11-28 2023-12-29 广州云趣信息科技有限公司 一种基于大数据的语义分析方法、系统和存储介质
CN117392694A (zh) * 2023-12-07 2024-01-12 支付宝(杭州)信息技术有限公司 数据处理方法、装置及设备

Families Citing this family (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110147981A (zh) * 2019-04-12 2019-08-20 深圳壹账通智能科技有限公司 基于文本分析的合同风险审核方法、装置及终端设备
CN110134830A (zh) * 2019-04-15 2019-08-16 深圳壹账通智能科技有限公司 视频信息数据处理方法、装置、计算机设备和存储介质
CN112069806B (zh) * 2019-05-21 2024-04-05 杭州海康威视数字技术股份有限公司 简历筛选方法、装置、电子设备及存储介质
CN110427453B (zh) * 2019-05-31 2024-03-19 平安科技(深圳)有限公司 数据的相似度计算方法、装置、计算机设备及存储介质
CN110287318B (zh) * 2019-06-06 2021-09-17 秒针信息技术有限公司 业务操作的检测方法及装置、存储介质、电子装置
CN110442712B (zh) * 2019-07-05 2023-08-22 创新先进技术有限公司 风险的确定方法、装置、服务器和文本审理系统
CN110443187B (zh) * 2019-07-31 2022-08-26 上海秒针网络科技有限公司 特征信息的记录方法及装置
CN110689225A (zh) * 2019-08-26 2020-01-14 深圳壹账通智能科技有限公司 基于外呼的企业金融风险画像创建方法及相关设备
CN110503555A (zh) * 2019-08-28 2019-11-26 中国工商银行股份有限公司 交易数据的复核方法、装置和服务器
CN110717650A (zh) * 2019-09-06 2020-01-21 平安医疗健康管理股份有限公司 单据数据处理方法、装置、计算机设备和存储介质
CN110751553A (zh) * 2019-10-24 2020-02-04 深圳前海微众银行股份有限公司 潜在风险对象的识别方法、装置、终端设备及存储介质
CN111124925B (zh) * 2019-12-25 2024-04-05 斑马网络技术有限公司 基于大数据的场景提取方法、装置、设备和存储介质
CN111598041B (zh) * 2020-05-25 2023-05-02 青岛联合创智科技有限公司 一种用于物品查找的图像生成文本方法
CN112258317B (zh) * 2020-10-30 2022-11-11 深圳壹账通智能科技有限公司 基于人工智能的线上信贷方法、装置、计算机设备及介质
CN112163585B (zh) * 2020-11-10 2023-11-10 上海七猫文化传媒有限公司 文本的审核方法、装置、计算机设备及存储介质
CN112632989B (zh) * 2020-12-29 2023-11-03 中国农业银行股份有限公司 一种合同文本中风险信息的提示方法、装置及设备
CN112818150B (zh) * 2021-01-22 2024-05-07 天翼视联科技有限公司 一种图片内容审核方法、装置、设备和介质
CN112802460B (zh) * 2021-04-14 2021-10-19 中国科学院国家空间科学中心 一种基于语音处理的空间环境预报系统
CN113449506A (zh) * 2021-06-29 2021-09-28 未鲲(上海)科技服务有限公司 一种数据检测方法、装置、设备及可读存储介质
CN113658585B (zh) * 2021-08-13 2024-04-09 北京百度网讯科技有限公司 语音交互模型的训练方法、语音交互方法及装置
CN114095282B (zh) * 2022-01-21 2022-04-15 杭银消费金融股份有限公司 一种基于短文本特征提取的风控处理方法及设备
CN114898271A (zh) * 2022-05-26 2022-08-12 中国平安人寿保险股份有限公司 视频内容监控方法、装置、设备及介质
CN114926831A (zh) * 2022-05-31 2022-08-19 平安普惠企业管理有限公司 基于文本识别方法、装置、电子设备及可读存储介质
CN117094816B (zh) * 2023-10-18 2024-02-06 天津市职业大学 一种基于大数据的企业财务评估方法及系统

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101697514A (zh) * 2009-10-22 2010-04-21 中兴通讯股份有限公司 一种身份验证的方法及系统
US20170034226A1 (en) * 2015-07-29 2017-02-02 International Business Machines Corporation Group chat with dynamic background images and content from social media
CN108205526A (zh) * 2016-12-20 2018-06-26 百度在线网络技术(北京)有限公司 一种确定综合语义信息的方法与装置

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2003263559A (ja) * 2002-03-11 2003-09-19 Hiromasa Asakawa プリペイドカードを用いた保険契約処理システム
US20100306249A1 (en) * 2009-05-27 2010-12-02 James Hill Social network systems and methods
CN103279868B (zh) * 2013-05-22 2016-08-17 兰亭集势有限公司 一种自动识别欺诈订单的方法和装置
CN106506454B (zh) * 2016-10-10 2019-11-12 江苏通付盾科技有限公司 欺诈业务识别方法及装置
WO2018104834A1 (en) * 2016-12-07 2018-06-14 Yogesh Chunilal Rathod Real-time, ephemeral, single mode, group & auto taking visual media, stories, auto status, following feed types, mass actions, suggested activities, ar media & platform
CN107395488A (zh) * 2017-06-08 2017-11-24 深圳市金立通信设备有限公司 一种识别风险账户的方法及终端
CN107680602A (zh) * 2017-08-24 2018-02-09 平安科技(深圳)有限公司 语音欺诈识别方法、装置、终端设备及存储介质
CN107919137A (zh) * 2017-10-25 2018-04-17 平安普惠企业管理有限公司 远程审批方法、装置、设备及可读存储介质
CN107885870B (zh) * 2017-11-24 2019-04-16 北京神州泰岳软件股份有限公司 一种业务文档公式提取方法及装置
CN108520398B (zh) * 2018-03-28 2021-10-01 施益斌 建筑设计招投标签约系统
CN108399510A (zh) * 2018-04-16 2018-08-14 法信公证云(厦门)科技有限公司 一种合同风险管控方法和设备

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101697514A (zh) * 2009-10-22 2010-04-21 中兴通讯股份有限公司 一种身份验证的方法及系统
US20170034226A1 (en) * 2015-07-29 2017-02-02 International Business Machines Corporation Group chat with dynamic background images and content from social media
CN108205526A (zh) * 2016-12-20 2018-06-26 百度在线网络技术(北京)有限公司 一种确定综合语义信息的方法与装置

Cited By (46)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111581344A (zh) * 2020-04-26 2020-08-25 腾讯科技(深圳)有限公司 一种接口信息审核方法、装置、计算机设备及存储介质
CN112307235A (zh) * 2020-05-09 2021-02-02 支付宝(杭州)信息技术有限公司 前端页面元素的命名方法、装置及电子设备
CN112307235B (zh) * 2020-05-09 2024-02-20 支付宝(杭州)信息技术有限公司 前端页面元素的命名方法、装置及电子设备
CN111753516A (zh) * 2020-06-29 2020-10-09 平安国际智慧城市科技股份有限公司 文本查重处理方法、装置、计算机设备及计算机存储介质
CN111753516B (zh) * 2020-06-29 2024-04-16 平安国际智慧城市科技股份有限公司 文本查重处理方法、装置、计算机设备及计算机存储介质
CN114140140B (zh) * 2020-09-03 2023-03-21 中国移动通信集团浙江有限公司 一种场景筛选方法、装置及设备
CN114140140A (zh) * 2020-09-03 2022-03-04 中国移动通信集团浙江有限公司 一种场景筛选方法、装置及设备
CN112037796A (zh) * 2020-09-08 2020-12-04 平安科技(深圳)有限公司 一种数据处理方法、装置、设备及介质
CN111984779A (zh) * 2020-09-10 2020-11-24 支付宝(杭州)信息技术有限公司 一种对话文本分析方法、装置、设备和可读介质
CN112015747B (zh) * 2020-09-14 2024-04-16 支付宝(杭州)信息技术有限公司 数据上传方法及装置
CN112015747A (zh) * 2020-09-14 2020-12-01 支付宝(杭州)信息技术有限公司 数据上传方法及装置
CN112015869A (zh) * 2020-09-14 2020-12-01 支付宝(杭州)信息技术有限公司 一种待发布文本的风险检测方法、装置及设备
CN112100379A (zh) * 2020-09-15 2020-12-18 北京百度网讯科技有限公司 用于处理数据的方法、装置、设备以及存储介质
CN112100379B (zh) * 2020-09-15 2023-07-28 北京百度网讯科技有限公司 用于处理数据的方法、装置、设备以及存储介质
CN112215087A (zh) * 2020-09-21 2021-01-12 浙江数链科技有限公司 图片审核方法、装置、计算机设备和存储介质
CN112100357A (zh) * 2020-09-24 2020-12-18 腾讯科技(深圳)有限公司 引导语的生成方法及装置、电子设备、计算机存储介质
CN112257425A (zh) * 2020-09-29 2021-01-22 国网天津市电力公司 一种基于数据分级模型的电力数据分析方法及系统
CN112416476A (zh) * 2020-11-25 2021-02-26 武汉联影医疗科技有限公司 工作流执行方法、装置、计算机设备和存储介质
CN112416476B (zh) * 2020-11-25 2023-03-24 武汉联影医疗科技有限公司 工作流执行方法、装置、计算机设备和存储介质
CN112508717A (zh) * 2020-12-01 2021-03-16 中国人寿保险股份有限公司 一种影像信息的审核方法、装置、电子设备及存储介质
CN114640606A (zh) * 2020-12-01 2022-06-17 中移物联网有限公司 一种物联网卡终端的异常处理方法及控制器
CN112507936B (zh) * 2020-12-16 2024-04-23 平安银行股份有限公司 图像信息审核方法、装置、电子设备及可读存储介质
CN112492343A (zh) * 2020-12-16 2021-03-12 浙江大华技术股份有限公司 一种视频直播监控方法及相关装置
CN112507936A (zh) * 2020-12-16 2021-03-16 平安银行股份有限公司 图像信息审核方法、装置、电子设备及可读存储介质
CN112492343B (zh) * 2020-12-16 2023-11-10 浙江大华技术股份有限公司 一种视频直播监控方法及相关装置
CN112685594B (zh) * 2020-12-24 2022-10-04 中国人民解放军战略支援部队信息工程大学 基于注意力的弱监督语音检索方法及系统
CN112685594A (zh) * 2020-12-24 2021-04-20 中国人民解放军战略支援部队信息工程大学 基于注意力的弱监督语音检索方法及系统
CN112861662A (zh) * 2021-01-22 2021-05-28 平安科技(深圳)有限公司 基于人脸和交互文本的目标对象行为预测方法及相关设备
CN112861662B (zh) * 2021-01-22 2023-09-01 平安科技(深圳)有限公司 基于人脸和交互文本的目标对象行为预测方法及相关设备
CN113094520A (zh) * 2021-03-29 2021-07-09 四川新网银行股份有限公司 一种检查电子合同的方法
CN113362072B (zh) * 2021-06-30 2023-09-08 成都一蟹科技有限公司 风控数据处理方法、装置、电子设备及存储介质
CN113362072A (zh) * 2021-06-30 2021-09-07 平安普惠企业管理有限公司 风控数据处理方法、装置、电子设备及存储介质
CN113723759A (zh) * 2021-07-30 2021-11-30 北京淇瑀信息科技有限公司 基于设备意向度和设备风险度为设备提供互联网服务的方法及装置
CN113806505B (zh) * 2021-09-09 2024-04-16 科大讯飞股份有限公司 要素比对方法、装置、电子设备和存储介质
CN113806505A (zh) * 2021-09-09 2021-12-17 科大讯飞股份有限公司 要素比对方法、装置、电子设备和存储介质
CN113793597A (zh) * 2021-09-15 2021-12-14 云知声智能科技股份有限公司 一种语音识别方法、装置、电子设备和存储介质
CN114282882A (zh) * 2021-11-15 2022-04-05 深圳市东信时代信息技术有限公司 素材工单审核方法、装置、设备及存储介质
CN114282882B (zh) * 2021-11-15 2023-04-14 深圳市东信时代信息技术有限公司 素材工单审核方法、装置、设备及存储介质
CN115526605B (zh) * 2022-10-21 2024-03-08 金恒智控管理咨询集团股份有限公司 基于企业内部控制管理的审批方法及系统
CN115526605A (zh) * 2022-10-21 2022-12-27 金恒智控管理咨询集团股份有限公司 基于企业内部控制管理的审批方法及系统
CN116405300B (zh) * 2023-04-18 2024-01-23 无锡锡商银行股份有限公司 一种基于场景的在线协议签署安全分析系统及方法
CN116405300A (zh) * 2023-04-18 2023-07-07 无锡锡商银行股份有限公司 一种基于场景的在线协议签署安全分析系统及方法
CN117313723A (zh) * 2023-11-28 2023-12-29 广州云趣信息科技有限公司 一种基于大数据的语义分析方法、系统和存储介质
CN117313723B (zh) * 2023-11-28 2024-02-20 广州云趣信息科技有限公司 一种基于大数据的语义分析方法、系统和存储介质
CN117392694A (zh) * 2023-12-07 2024-01-12 支付宝(杭州)信息技术有限公司 数据处理方法、装置及设备
CN117392694B (zh) * 2023-12-07 2024-04-19 支付宝(杭州)信息技术有限公司 数据处理方法、装置及设备

Also Published As

Publication number Publication date
CN109543516A (zh) 2019-03-29

Similar Documents

Publication Publication Date Title
WO2020077895A1 (zh) 签约意向判断方法、装置、计算机设备和存储介质
CN109165840B (zh) 风险预测处理方法、装置、计算机设备和介质
CN110021439B (zh) 基于机器学习的医疗数据分类方法、装置和计算机设备
CN109272396B (zh) 客户风险预警方法、装置、计算机设备和介质
WO2021169111A1 (zh) 简历筛选方法、装置、计算机设备和存储介质
CN109871446B (zh) 意图识别中的拒识方法、电子装置及存储介质
CN108536800B (zh) 文本分类方法、系统、计算机设备和存储介质
WO2020077896A1 (zh) 提问数据生成方法、装置、计算机设备和存储介质
CN109829629B (zh) 风险分析报告的生成方法、装置、计算机设备和存储介质
WO2021114840A1 (zh) 基于语义分析的评分方法、装置、终端设备及存储介质
WO2020147395A1 (zh) 基于情感的文本分类处理方法、装置和计算机设备
CN109063921B (zh) 客户风险预警的优化处理方法、装置、计算机设备和介质
WO2020114100A1 (zh) 一种信息处理方法、装置和计算机存储介质
WO2021204017A1 (zh) 文本意图识别方法、装置以及相关设备
CN110909784B (zh) 一种图像识别模型的训练方法、装置及电子设备
WO2021164205A1 (zh) 基于身份识别的数据审核方法、装置和计算机设备
WO2019041528A1 (zh) 新闻情感方向判断方法、电子设备及计算机可读存储介质
CN110377618B (zh) 裁决结果分析方法、装置、计算机设备和存储介质
CN114493902A (zh) 多模态信息异常监控方法、装置、计算机设备及存储介质
CN110362592B (zh) 裁决指引信息推送方法、装置、计算机设备和存储介质
CN107533672A (zh) 模式识别装置、模式识别方法以及程序
CN112288279A (zh) 基于自然语言处理和线性回归的业务风险评估方法和装置
CN116955559A (zh) 问答匹配方法、装置、电子设备及存储介质
CN114283429A (zh) 素材工单数据处理方法、装置、设备及存储介质
CN115358817A (zh) 基于社交数据的智能产品推荐方法、装置、设备及介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19873974

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 05/08/2021)

122 Ep: pct application non-entry in european phase

Ref document number: 19873974

Country of ref document: EP

Kind code of ref document: A1