Intelligent return visit method and device for fraud-related number verification

Abstract

The invention provides an intelligent return visit method and device for fraud-related number verification, and the method comprises the steps: in a user return visit process, under the condition that an intention recognition result of intention recognition based on user reply information does not correspond to a verbal skill branch included in a main verbal skill process tree; and obtaining a reply result based on the natural language generation model and replying the reply result to the return visit user. According to the embodiment of the invention, the knowledge fusion-based natural language generation model oriented to fraud-related number return visit verification is provided, the pre-training large model is used as the ability of training and prediction, and the friendliness of a return visit process is improved, so that the response rate is improved, the complaint rate is reduced, and the possibility of wrong fraud-related judgment response is reduced; on the basis of an intention recognition model oriented to fraud-related number return visit verification, the dialogue is pulled back to the main verbal skill process tree in time, and the probability of effective reply is improved.

Description

Technical Field

[0001] The present invention relates to the field of information security technology, and in particular to an intelligent return visit method and device for verifying fraudulent numbers. Background Art

[0002] In order to prevent numbers released through various channels from being directly or indirectly used by fraudsters and to protect the property safety of the public, the anti-fraud business departments of current operators mainly focus on data dimensions such as call behavior, basic personal information, SMS behavior, number card location, and terminal information. Based on expert rules and superimposed data models, they achieve quasi-real-time model matching and fraudulent number analysis, thereby controlling and shutting down risky numbers. However, fraudulent numbers identified under this model often lack accuracy and are prone to being blocked by mistake, increasing the complaint rate and re-opening rate. Summary of the invention

[0003] In view of the problems existing in the prior art, the present invention provides an intelligent return visit method and device for verifying fraudulent numbers.

[0004] The present invention provides an intelligent return visit method for verifying fraudulent numbers, the method comprising: The user response information of the returning user obtained during the return visit is input into the intent recognition model to recognize the intention of the returning user, and based on the intention recognition result and the pre-established main speech process tree, the next round of speech response is entered; In the case where the intention recognition result does not correspond to the speech branch included in the main speech process tree, the conversation history information and related knowledge information are input into the natural language generation model, and the reply result output by the natural language generation model is replied to the returning user; Based on the intention recognition model, binary classification emotion recognition is performed to obtain emotion recognition results, and based on the obtained emotion recognition results, it is determined to enter the next round of speech response or end the return visit process.

[0005] Optionally, the natural language generation model is obtained by training based on dialogue history information samples and relevant knowledge information samples.

[0006] Optionally, the relevant knowledge information is information selected from a knowledge base through a knowledge selection model based on the conversation history information; The relevant knowledge information samples are information samples selected from a knowledge base through a knowledge selection model based on the dialogue history information samples.

[0007] Optionally, the method further comprises: In the main speech process tree, a dynamic text-to-speech component is used to synthesize the dynamic information of the return visit in real time and reply to the return visit user.

[0008] Optionally, the step of constructing the knowledge base includes: Acquire knowledge base data from at least one source, wherein the knowledge base data is a structured triple set, including a head entity node, an entity relationship node, and a tail entity node; wherein the head entity node represents a specific knowledge entity, the tail entity node represents an entity attribute feature or another entity, and the entity relationship node represents a connection between the head entity node and the tail entity node; Decomposing the knowledge base data, performing word segmentation processing to obtain corresponding head vectors, relationship vectors, and tail vectors, and obtaining key / value representations corresponding to the knowledge base data based on the head vectors, relationship vectors, and tail vectors; Based on the key / value representation corresponding to the knowledge base data, the weight information from different sources and the conversation history information, a unified embedding corresponding to each knowledge base data in the knowledge base is obtained.

[0009] Optionally, obtaining the key / value representation corresponding to the knowledge base data based on the head vector, the relationship vector and the tail vector includes: Calculate the average vector of the head vector, relationship vector and tail vector respectively, and get [ , , 】; Will , ] Perform vector averaging again to get the key representation Key m ,Will , ] Perform vector averaging again to get the value representation .

[0010] Optionally, the method further comprises: The knowledge base is screened based on a word frequency-inverse document frequency value.

[0011] Optionally, the screening of the knowledge base based on the word frequency-inverse document frequency value specifically includes: Obtaining word frequency-inverse file frequency values ​​of the first word set corresponding to the conversation history information, and sorting them; A second word set is selected based on the ranking result, and a filtered knowledge base is obtained based on the vector similarity between the second word set and the knowledge base data.

[0012] Optionally, the relevant knowledge information is selected from a knowledge base through a knowledge selection model based on the conversation history information, specifically including: The conversation history information and the user reply information are respectively concatenated with the vector of each piece of knowledge in the knowledge base to obtain a historical probability distribution and a posterior reply probability distribution, which represent the probability of selecting each piece of knowledge; Based on the cross entropy loss function, the historical probability distribution is made to continuously narrow the gap with the posterior response probability distribution to determine the target knowledge vector.

[0013] Optionally, the historical probability distribution is obtained based on a dialog history embedding corresponding to the dialog history information, and the dialog history embedding is obtained based on a word vector sequence corresponding to the dialog history information and an attention factor corresponding to each word vector.

[0014] The present invention also provides an electronic device, including 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, it implements any of the above-mentioned intelligent return visit methods for verifying fraudulent numbers.

[0015] The invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, it implements any of the above-mentioned intelligent follow-up methods for verifying fraudulent numbers.

[0016] The present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements any of the above-mentioned intelligent return visit methods for verifying fraudulent numbers.

[0017] The present invention provides an intelligent return visit method and device for fraudulent number verification. The method includes the following steps: during the user return visit, when the intention recognition result based on the user's reply information does not correspond to the speech branch included in the main speech process tree, a reply result is obtained based on a natural language generation model and replied to the return visit user. The embodiment of the present invention proposes a knowledge fusion-based NLG model for return visit verification of fraudulent numbers, which uses a pre-trained large model as a training and prediction capability, increases the friendliness of the return visit process, thereby improving the response rate, reducing the complaint rate, and reducing the possibility of erroneous fraudulent judgment responses; based on the intention recognition model for return visit verification of fraudulent numbers, the conversation is pulled back to the main speech process tree in a timely manner, thereby increasing the probability of an effective reply. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0019] Figure 1 A schematic diagram of the architecture of an intelligent return visit system for fraudulent number verification provided by an embodiment of the present invention.

[0020] Figure 2 A flow chart of an intelligent return visit method for fraudulent number verification provided by an embodiment of the present invention.

[0021] Figure 3 A schematic diagram of the main speech flow tree provided for an embodiment of the present invention.

[0022] Figure 4 A schematic diagram of an intent recognition model provided in an embodiment of the present invention.

[0023] Figure 5 This is a diagram of the overall architecture of the knowledge selection model provided in an embodiment of the present invention.

[0024] Figure 6 A schematic diagram of constructing an NLG model based on knowledge fusion provided in an embodiment of the present invention.

[0025] Figure 7 A schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] In order to solve the problems of the judgment method based on expert rules and data models, the existing technology has introduced a method of intelligent return visit to the caller of the fraudulent number. Through the guidance of robots, potential victims are asked to confirm whether they have been exposed to such fraudulent information recently and whether it comes from the fraudulent number. The current intelligent return visit methods in this field mainly include the following: (I) Solution 1: Establish a speech process tree, broadcast the pre-recorded recordings of each node, and the user feedbacks the intention information by pressing keys, and finally obtains the judgment result at the leaf node.

[0028] (II) Solution 2: Establish a conversation flow tree, introduce dynamic information such as call time, fraud number, fraud type, etc. into the conversation through dynamic text to speech (TTS) and report it in full. At the same time, obtain user intent through regular expression matching after speech transcription and natural language understanding (NLU) technology to finally get the judgment result.

[0029] (III) Solution 3: Establish an intelligent agent based on the big model, set the background and role for the intelligent agent, that is, act as an agent for the anti-fraud verification and return visit task; adopt a few-shot prompt learning or thinking chain mode of a small sample, so that the big model can understand the tasks it should undertake while learning to understand the user's intentions and guide the user to answer the corresponding questions, step by step approaching the final judgment result. At the same time, each time the return visit task is started, dynamic information such as the call time, fraud number, and fraud type are introduced in the prompt.

[0030] (1) Solution 1 does not introduce dynamic information such as the call time, fraud number, and fraud type used to associate specific follow-up tasks, resulting in inaccurate follow-up results.

[0031] (2) In solutions 1 and 2, once the user's answer exceeds the preset nodes of the speech process tree, they can only enter the default speech branch or use the backup knowledge base to respond. The follow-up process is not user-friendly and can easily make users feel disgusted, thereby reducing the response rate, increasing the complaint rate, and increasing the possibility of incorrect responses.

[0032] (3) Solution 3: Using a large model agent will increase the friendliness of multi-round conversations. However, due to the divergent thinking characteristics and hallucination defects of the large model, as well as the uncertainty of the user's answer, if there is no correlation with the samples in the few-shot, it is likely to gradually move away from the goal of the revisit task, resulting in no effective judgment results.

[0033] (4) Solution 3 relies on the generation of general responses obtained by a large model through learning from a very large corpus. It does not incorporate key knowledge about anti-fraud business, company information, authorization information, the source of this information, and other anti-fraud verification vertical scenarios. The answer of "security" from the large model will prevent users from receiving a professional and reasonable response, thereby giving rise to user complaints and increasing the error rate of judgment.

[0034] In view of the problems arising from the above-mentioned schemes, the embodiments of the present invention provide an intelligent return visit method for fraud-related number verification. The speech process tree is integrated with the active generation of the model, and the natural language generation (NLG) is dynamically introduced to carry out the dialogue when the speech branch is exceeded; the dynamic parameters of the speech process tree and the complete dialogue of NLG are generated respectively by dynamic TTS; the intention / emotion recognition model based on Albert ALBert is built to identify the user's response and the mapping of the corresponding intention branch; the professional knowledge base is introduced in the active generation part of the model, and a knowledge selection model based on the pre-trained model is established for the model to select accurate knowledge information in combination with contextual information; on this basis, a knowledge fusion NLG large model based on the pre-trained large model is established, so that the large model can be used to widely understand the user's intention and give a general answer. At the same time, it can give a more accurate reply when the user involves professional questions related to anti-fraud verification; finally, the emotion classification is used to transfer the conversation to the main task process itself, so that the embodiment of the present invention can solve the defects and problems of the above-mentioned existing schemes. The use of the embodiment of the present invention in the intelligent anti-fraud system can significantly improve the intelligence level and performance indicators of the system.

[0035] Figure 1 A schematic diagram of the intelligent return visit system architecture for fraudulent number verification provided by an embodiment of the present invention, such as Figure 1 As shown, the embodiment of the present invention is mainly explained from several parts, such as the construction of the main speech process tree, the construction of the fraud intention / emotion recognition model based on Albert, the import of embedding and coarse screening of the industry knowledge base, the construction of the knowledge selection model, and the construction of the NLG model based on knowledge fusion.

[0036] Figure 2 A flow chart of an intelligent return visit method for fraudulent number verification provided by an embodiment of the present invention, such as Figure 2 As shown, the method includes: Step 200: input the user response information of the return user obtained during the return visit into the intention recognition model to perform intention recognition of the return user, and enter the next round of speech response based on the intention recognition result and the pre-established main speech process tree; First, establish the main conversation process tree. Based on the pre-established main conversation process tree, start the return visit process. After receiving the user's reply information, use the intent recognition model to perform multi-classification intent recognition judgment. After obtaining the intent recognition result, follow the settings of the main conversation process tree and divert to the corresponding intent branch to enter the next round of conversation response.

[0037] Step 201: When the intention recognition result does not correspond to the speech branch included in the main speech process tree, the conversation history information and related knowledge information are input into the natural language generation (NLG) model, and the reply result output by the NLG model is replied to the returning user; If an intention other than the main speech process tree occurs, the embodiment of the present invention uses the conversation history information and related knowledge information as input to the NLG model, obtains the reply result output by the NLG model, and replies to the returning user.

[0038] Step 202: Perform binary classification emotion recognition based on the intention recognition model to obtain emotion recognition results, and determine whether to enter the next round of speech response or end the return visit process based on the acquired emotion recognition results.

[0039] After the intention recognition model performs binary classification emotion recognition on the returning user to obtain the emotion recognition result, it determines to continue the return visit process if the emotion recognition result is positive, or determines to postpone the return visit process if the emotion recognition result is negative.

[0040] The intelligent return visit method provided by the embodiment of the present invention increases the friendliness of the return visit process, thereby improving the response rate, reducing the complaint rate, and reducing the possibility of erroneous fraud judgment responses; based on the intention recognition model for return visit verification of fraudulent numbers, the conversation is pulled back to the main speech process tree in a timely manner, thereby increasing the probability of an effective reply. In the process of return visits, if there is an intention outside the main speech process tree, it is necessary to call the knowledge selection model to select knowledge vectors that are strongly related to the current context and context, and splice them with historical conversations to generate replies as a pre-trained NLG model. Specifically, the conversation history information and related knowledge information are input into the NLG model, and the reply result output by the NLG model is replied to the return visit user, and the emotion recognition result obtained by the binary emotion recognition based on the intention recognition model, according to the user's emotion classification, enter the original node to repeat the speech, or end the conversation, and finally obtain the judgment result of whether it is fraudulent at the leaf node of the main speech process tree.

[0041] Furthermore, the natural language generation model can be obtained by iterative training based on the dialogue history information samples and the relevant knowledge information samples in advance. The relevant knowledge information samples can be information samples selected from the knowledge base through the knowledge selection model based on the dialogue history information samples. The relevant knowledge information described in step 201 is information selected from the knowledge base through the knowledge selection model based on the dialogue history information.

[0042] In the intelligent follow-up method for fraudulent number verification provided by the embodiment of the present invention, when the intention recognition result based on the user's reply information does not correspond to the speech branch included in the main speech process tree, the reply result is obtained based on the natural language generation model and replied to the follow-up user. The embodiment of the present invention proposes a knowledge-fusion-based NLG model for follow-up verification of fraudulent numbers, which uses a pre-trained large model as a training and prediction capability, increases the friendliness of the follow-up process, thereby improving the response rate, reducing the complaint rate, and reducing the possibility of erroneous fraud judgment responses; based on the intention recognition model for follow-up verification of fraudulent numbers, the conversation is pulled back to the main speech process tree in a timely manner, which increases the probability of an effective reply.

[0043] In the method embodiment of the present invention, an NLG model for follow-up verification of fraudulent numbers is proposed. Combined with the main speech process tree, when the intention recognition model determines that it exceeds the established intention branch, that is, when the intention recognition result does not correspond to the speech branch included in the main speech process tree, an NLG model based on knowledge fusion is used to generate dialogue, which solves the problems of the unfriendly follow-up process in the prior art, which easily makes users feel disgusted, thereby reducing the response rate, increasing the complaint rate, and increasing the possibility of incorrect responses.

[0044] In the method embodiment of the present invention, an intent recognition model for follow-up verification of fraudulent numbers is proposed. After exceeding the intent branch and the large model gives an answer, the user's emotional judgment is performed, and the conversation is promptly pulled back to the main speech process tree, avoiding the natural divergent thinking characteristics and illusion defects of the large model in the prior art, gradually moving away from the goal of the follow-up task, and resulting in the problem of not being able to obtain effective judgment results.

[0045] In the above method embodiment, the method further includes: in the main speech process tree, applying a dynamic text-to-speech TTS component to synthesize the dynamic information of the return visit in real time, and replying to the return visit user. Figure 3 A schematic diagram of the main speech process tree provided by an embodiment of the present invention, such as Figure 3 As shown: Among them, the "R" node represents the robot broadcast node, and "[]" represents the use of real-time TTS to synthesize dynamic parameters.

[0046] After user feedback and automatic speech recognition (ASR) speech transcription, the intent recognition model performs multi-classification intent recognition and binary classification emotion recognition of the user; For each intention recognition point in the process tree, there can be a related robot broadcast node to connect; for the intention nodes outside the process tree, the NLG model is used to superimpose the knowledge base to generate replies, and combined with the previous emotion recognition, the next node is selected. If it is a negative emotion, it directly enters the end node. If it is a positive emotion, it enters the current layer of robot to repeat the broadcast and intention recognition.

[0047] The embodiment of the present invention proposes a main speech flow tree for follow-up verification of fraudulent numbers, and integrates a dynamic TTS component to synthesize dynamic information such as the call time of the caller and the called party, the fraud number, and the fraud type in real time, thereby solving the problem of inaccurate follow-up results in the prior art and improving the accuracy of the follow-up verification.

[0048] Figure 4 Schematic diagram of the intent recognition model provided by the embodiment of the present invention. Build an intent recognition model based on Albert and perform fine-tuning on the two tasks of sentiment classification and intent recognition, such as Figure 4 shown.

[0049] (1) Input the ASR-transcribed and pre-processed text into AlBert to obtain the embedding of the word segmentation.

[0050] (2) Use the fully connected layer to reduce the dimension and map the hidden layer vector dimension to the output layer dimension.

[0051] (3) Fine-tuning Task 1: For the first position of the vector, use the Sigmoid function activation to convert the probability to the interval [0, 1] for sentiment classification; (4) Fine-tuning Task 2: For the remaining positions, use the Softmax function to activate and distribute the probabilities of all classifications in the interval [0, 1], and make the sum of the probabilities 1 for intent recognition.

[0052] Loss1 and Loss2 correspond to the loss functions of fine-tuning tasks 1 and 2 respectively, and both are fitted using the cross entropy loss function, as shown below.

[0053]

[0054] The loss functions of the two fine-tuning tasks are weighted and combined to obtain the loss function of this model. Among them, w is a hyperparameter, and 0.65 is taken as the optimal value after experiments.

[0055]

[0056] In the above method embodiment, the knowledge base is constructed based on the following steps: Acquire knowledge base data from at least one source, wherein the knowledge base data is a structured triple set, including a head entity node, an entity relationship node, and a tail entity node; wherein the head entity node represents a specific knowledge entity, the tail entity node represents an entity attribute feature or another entity, and the entity relationship node represents a connection between the head entity node and the tail entity node; Decomposing the knowledge base data, performing word segmentation processing to obtain corresponding head vectors, relationship vectors, and tail vectors, and obtaining key / value representations corresponding to the knowledge base data based on the head vectors, relationship vectors, and tail vectors; Based on the key / value representation corresponding to the knowledge base data, the weight information from different sources and the conversation history information, a unified embedding corresponding to each knowledge base data in the knowledge base is obtained.

[0057] The step of obtaining the key / value representation corresponding to the knowledge base data based on the head vector, the relationship vector and the tail vector includes: Calculate the average vector of the head vector, relationship vector and tail vector respectively, and get [ , , 】; Will , ] Perform vector averaging again to get the key representation Key m ,Will , ] Perform vector averaging again to get the value representation .

[0058] Specifically, first, a unified import of multi-structured knowledge base is performed. The knowledge base Z introduced into the NLG model in the embodiment of the present invention is a structured triple set, which is composed of a head entity node (Head Entity), an entity relationship node (Relation), and a tail entity node (Tail Entity). The head entity represents a specific knowledge entity, the tail entity node represents an entity attribute feature, or another entity, and the entity relationship node represents the connection between the head entity and the tail entity, such as [H "Mobile Company", R "Has the responsibility of supervision, analysis, identification, and disposal", T "Fraudulent number"]; [H "Cancel air ticket", R "Air ticket related services will not be notified and processed through the mobile number, but will use the official number segment, and there will be no abnormal behavior such as adding friends", T "Fraudulent behavior"].

[0059] The embodiment of the present invention collects knowledge base data from at least one source to maximize the addition of vertical knowledge in the anti-fraud industry, operator industry, return visit criteria, and cross-fields to the large model, as shown in the following table.

[0060]

[0061] Then, the overall knowledge base is embedded based on the key-value expression fusion weights.

[0062] In order to better express triple knowledge information from different sources, different contents, and different purposes in the same semantic space, fully capture the deep representation of the knowledge base, and highlight the industry knowledge categories that you want the big model to focus on while combining historical semantics, the embodiment of the present invention uses a key-value representation-based model to embed the knowledge base as a whole, and combines weight information from different sources and conversation history data. The steps are as follows.

[0063] (1) Decompose the mth triplet of knowledge [H, R, T], perform word segmentation and embedding on it respectively, and use Word2Vec as the word embedding algorithm. For each piece of knowledge, we will get , , ; (2) Yes , , Find the average vector respectively, and get [ , , 】.

[0064]

[0065] (3) , ] Perform vector averaging again to obtain .Will , ] Perform vector averaging again to obtain . Then each piece of knowledge can be used [ express.

[0066] (4) Obtain fusion head vector, relationship vector, and conversation history , weight of each industry The coding index .

[0067]

[0068] (5) Get the final knowledge base unified embedding .

[0069]

[0070] The method embodiment further includes a step of screening the knowledge base based on a term frequency-inverse document frequency value TF-IDF, including: Obtaining TF-IDF values ​​of the first word set corresponding to the conversation history information and sorting them; A second word set is selected based on the ranking result, and a filtered knowledge base is obtained based on the vector similarity between the second word set and the knowledge base data.

[0071] Specifically, in the scenario of anti-fraud return visit verification, the knowledge base is too large, and the computational overhead and time complexity required for subsequent deep learning are unbearable. In addition, the introduction of too much irrelevant peripheral noise data will affect the learning quality of the neural network, and thus affect the performance indicators of the final model. Therefore, the embodiment of the present invention introduces a rough screening of the knowledge base based on TF-IDF, and the process is as follows: (1) Based on conversation history Perform word segmentation, remove punctuation and stop words, and obtain the word set V.

[0072] (2) Calculate the TF-IDF value for each word segment and word set V and sort them.

[0073] Select word set O based on the sorting results, and select knowledge base subset based on vector similarity ,in , the unified embedding representation is calculated as .

[0074] In the above method embodiment, the relevant knowledge information is selected from the knowledge base through a knowledge selection model based on the conversation history information, specifically including: The conversation history information and the user reply information are concatenated with the vector of each piece of knowledge in the knowledge base to obtain the historical probability distribution Z a and the posterior probability distribution Z b , indicating that each piece of knowledge Z is selected i probability; Based on the cross entropy loss function, the historical probability distribution Z a Keep shrinking and the posterior probability distribution Z b The gap between the historical probability distribution Z and the historical probability distribution Z is obtained, and the most appropriate knowledge vector is selected as the target knowledge vector. a is the conversation history embedding H corresponding to the conversation history information e Obtained, the H e is based on the word vector sequence corresponding to the conversation history information and the attention factor corresponding to each word vector Obtained.

[0075] The embodiment of the present invention proposes a knowledge selection model for follow-up verification of fraudulent numbers, which fully integrates the expression level of the current context in the entire conversation history, and gives an attention factor to the conversation history sequence, so as to accurately select appropriate knowledge base entries, so that the large model can generate "guided replies" that are constantly close to the answer to the final question, thereby solving the problem that the original knowledge base will have a large amount of irrelevant noise data, affecting the learning and prediction quality of the generative neural network, and then affecting the performance indicators of the final model, as well as the accuracy of fraudulent number verification and increasing the user complaint rate. It improves the accuracy of the reply and the accuracy of fraudulent number verification, and reduces the user complaint rate.

[0076] Specifically, the construction of the knowledge selection model is introduced below. Figure 5 The overall architecture diagram of the knowledge selection model provided by the embodiment of the present invention is as follows: Figure 5 As shown in Figure 1, it is mainly composed of two parts: dialogue history and human response embedding and knowledge fusion selection. The dialogue history and human response vectors are concatenated with the knowledge base vector to obtain the historical probability distribution Z a and the posterior probability distribution Z b , indicating that each piece of knowledge Z is selected i Based on the cross entropy loss function, let the historical probability distribution Z a Keep shrinking and posterior recovery probability distribution Z b , so as to obtain the most likely knowledge probability distribution and select the most appropriate knowledge vector as the target knowledge vector.

[0077] (a) Dialogue history and human response embedding.

[0078] In the anti-fraud follow-up scenario, users need to get closer to the answer to the final question as quickly as possible and give accurate "guiding replies" instead of giving "divergent replies" with higher security characteristics in order to satisfy users. Therefore, it is necessary to highlight the importance of the current context and fully integrate the recent reply vector into the encoding of the overall semantic representation of the conversation history, so as to obtain accurate historical conversation embeddings that are biased towards the current session, provide accurate replies to user feedback, and provide guiding robot answers. The specific process is as follows: (1) Use represents the conversation history sequence, The final conversation vector is obtained by removing stop words, preprocessing, and word segmentation from the conversation history to obtain a word vector sequence. .

[0079] (2) The Transformer-based pre-trained language model ERINE is selected as the embedding model. The core idea of ​​the ERINE model is to integrate structured knowledge (such as knowledge graphs) with text information to enhance the model's ability to understand language. Input ERINE_Base to get the encoded output CLS vector , representing the current context representation.

[0080]

[0081] (3) Order is the Query vector in the Attention mechanism, For the Key vector in the Attention mechanism, solve each word vector for the current context The attention factor is denoted as , represents the influence of the i-th word segment on the current session.

[0082]

[0083] (4) Then Perform weighted summation to obtain the final dialogue history encoding .

[0084]

[0085] In addition, the user's response sequence needs to be independently encoded for subsequent supervised learning.

[0086] (1) Use Represents the user reply sequence, removes stop words, performs preprocessing, and performs word segmentation to obtain a word vector sequence .

[0087] (2) Sequence word vectors Directly input the ERINE_Base model to extract and embed the vector representation to obtain the reply vector .

[0088]

[0089] 2. Knowledge fusion selection

[0090] In the scenario of follow-up verification of fraudulent numbers, the essence of knowledge selection is to capture the semantic information of the current context, match the knowledge graph that is strongly related to the current context, and find key knowledge items that can enable the large model to reduce hallucinations, accurately respond, and quickly guide the answers to questions, so as to integrate them into the input of the subsequent dialogue generation model. The specific process is as follows.

[0091] (1) Embedding the conversation history And the knowledge embedding after coarse screening After concatenation, the fully connected layer is used to map the neurons in the output dimension, where the output dimension is the knowledge item k after coarse screening. Softmax activation is used to obtain ,in Represents the importance of the k-th piece of knowledge, that is, the probability of being selected. Represents the probability distribution of selection based on historical dialogue knowledge. , are the parameters that need to be learned iteratively.

[0092] softmax(Linear(concat( )))

[0093] (2) Similarly, embedding the user reply sequence Perform the same operation and obtain . represents the probability distribution of knowledge selection based on the user's posterior response sequence.

[0094] softmax(Linear(concat( )))

[0095] (3) Based on the posterior response information As a basis for supervision, Continuously reduce the probability distribution gap between it and the knowledge, so that the selection of knowledge is more effective and accurate, and the cross entropy loss function is obtained as follows.

[0096]

[0097] Figure 6 Schematic diagram of NLG model construction based on knowledge fusion provided by an embodiment of the present invention, and NLG model architecture for return visit verification of fraudulent numbers Figure 6 As shown, the embodiment of the present invention adopts the GLM-4 large language model based on GPT as the pre-training model base, splices the dialogue history sequence and the sequence after knowledge selection and inputs them into the training, and masks the dialogue vector after the current prediction position during the fine-tuning process, and also uses the cross entropy loss function as the optimization basis. It represents the probability of selecting each word in the vocabulary when generating the i-th word, where y is the correct generated vector.

[0098]

[0099] The embodiment of the present invention proposes a classification import, encoding, rough screening and knowledge selection model for the industry knowledge base for the return visit verification of fraudulent numbers as the input item of the generative large model, so as to obtain accurate professional replies, solving the problem that the answer of the large model "security" in the third solution does not allow users to get a professional and reasonable response, thereby giving rise to user complaints and increasing the error rate of judgment. The importance of sources from different fields of the knowledge base and the semantic information of the dialogue context are fully considered, and the chaotic knowledge base is uniformly imported, represented and purified, solving the problem that the knowledge base is too large, the computational overhead and time complexity required in subsequent deep learning cannot be borne, and the knowledge base has many sources, resulting in uneven organization, data quality and importance, reducing the user complaint rate and the error rate of fraud judgment in the return visit.

[0100] Figure 7 A schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention, such as Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the intelligent return visit method for fraudulent number verification provided by the above embodiments.

[0101] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on this understanding, the technical solution of the present invention can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0102] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the intelligent follow-up method for fraudulent number verification provided in the above-mentioned embodiments.

[0103] On the other hand, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, and the computer program is used by a processor to execute the intelligent return visit method for fraudulent number verification provided in the above-mentioned embodiments.

[0104] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0105] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent return visit method for fraudulent number verification, characterized in that: The method comprises: The user response information of the returning user obtained during the return visit is input into the intent recognition model to recognize the intention of the returning user, and based on the intention recognition result and the pre-established main speech process tree, the next round of speech response is entered; In the case where the intention recognition result does not correspond to the speech branch included in the main speech process tree, the conversation history information and related knowledge information are input into the natural language generation model, and the reply result output by the natural language generation model is replied to the returning user; Based on the intention recognition model, binary classification emotion recognition is performed to obtain emotion recognition results, and based on the obtained emotion recognition results, it is determined to enter the next round of speech response or end the return visit process.

2. The intelligent return visit method for fraudulent number verification according to claim 1 is characterized in that: The natural language generation model is obtained by training based on dialogue history information samples and relevant knowledge information samples.

3. The intelligent return visit method for fraudulent number verification according to claim 2 is characterized in that: The relevant knowledge information is information selected from a knowledge base through a knowledge selection model based on the conversation history information; The relevant knowledge information samples are information samples selected from a knowledge base through a knowledge selection model based on the dialogue history information samples.

4. The intelligent return visit method for fraudulent number verification according to claim 1 is characterized in that: The method further comprises: In the main speech process tree, a dynamic text-to-speech component is used to synthesize the dynamic information of the return visit in real time and reply to the return visit user.

5. The intelligent return visit method for fraudulent number verification according to claim 3 is characterized in that: The steps of constructing the knowledge base include: Acquire knowledge base data from at least one source, wherein the knowledge base data is a structured triple set, including a head entity node, an entity relationship node, and a tail entity node; wherein the head entity node represents a specific knowledge entity, the tail entity node represents an entity attribute feature or another entity, and the entity relationship node represents a connection between the head entity node and the tail entity node; Decomposing the knowledge base data, performing word segmentation processing to obtain corresponding head vectors, relationship vectors, and tail vectors, and obtaining key / value representations corresponding to the knowledge base data based on the head vectors, relationship vectors, and tail vectors; Based on the key / value representation corresponding to the knowledge base data, the weight information from different sources and the conversation history information, a unified embedding corresponding to each knowledge base data in the knowledge base is obtained.

6. The intelligent return visit method for fraudulent number verification according to claim 5 is characterized in that: The obtaining the key / value representation corresponding to the knowledge base data based on the head vector, the relationship vector and the tail vector includes: Calculate the average vector of the head vector, relationship vector and tail vector respectively, and get [ , , 】; Will , ] Perform vector averaging again to get the key representation Key m ,Will , ] Perform vector averaging again to get the value representation .

7. The intelligent return visit method for fraudulent number verification according to claim 5 or 6, characterized in that: The method further comprises: The knowledge base is screened based on a word frequency-inverse document frequency value.

8. The intelligent return visit method for fraudulent number verification according to claim 7 is characterized in that: The screening of the knowledge base based on the word frequency-inverse document frequency value specifically includes: Obtaining word frequency-inverse file frequency values ​​of the first word set corresponding to the conversation history information, and sorting them; A second word set is selected based on the ranking result, and a filtered knowledge base is obtained based on the vector similarity between the second word set and the knowledge base data.

9. The intelligent return visit method for fraudulent number verification according to claim 1 is characterized in that: The relevant knowledge information is selected from the knowledge base through a knowledge selection model based on the conversation history information, specifically including: The conversation history information and the user reply information are respectively concatenated with the vector of each piece of knowledge in the knowledge base to obtain a historical probability distribution and a posterior reply probability distribution, which represent the probability of selecting each piece of knowledge; Based on the cross entropy loss function, the historical probability distribution is made to continuously narrow the gap with the posterior response probability distribution to determine the target knowledge vector.

10. The intelligent return visit method for fraudulent number verification according to claim 9 is characterized in that: The historical probability distribution is obtained based on the dialogue history embedding corresponding to the dialogue history information, and the dialogue history embedding is obtained based on the word vector sequence corresponding to the dialogue history information and the attention factor corresponding to each word vector.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the intelligent return visit method for verifying fraudulent numbers as described in any one of claims 1 to 10 is implemented.

12. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the intelligent return visit method for verifying fraudulent numbers as described in any one of claims 1 to 10 is implemented.

13. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the intelligent return visit method for verifying fraudulent numbers as described in any one of claims 1 to 10 is implemented.

Images