Complaint telephone intelligent analysis method and device based on NLP and ASR

Abstract

The invention provides a complaint telephone intelligent analysis method and device based on NLP and ASR, and relates to the field of artificial intelligence. The method comprises the following steps: acquiring insurance complaint telephone data, and respectively inputting the insurance complaint telephone data into an acoustic model and a semantic model; outputting pronunciation features related to insurance terms based on the acoustic model, and outputting semantic general shooting features related to the insurance field based on the semantic model; performing cross-modal fusion processing on the pronunciation features and the semantic general shooting features, and outputting double-path fusion features; and inputting the two-way fusion feature into a multi-modal complaint classification model, and outputting a complaint analysis result corresponding to the insurance complaint telephone data, the complaint analysis result comprising a complaint type, a complaint amount and a complaint trend. The problems that the real intention of the complaint customer cannot be accurately captured and the complaint classification accuracy is easy to occur due to the fact that the complaint phone in the insurance field is analyzed only by depending on a single-path voice recognition model are solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method and apparatus for intelligent analysis of complaint calls based on NLP and ASR. Background Technology

[0002] Today, in the insurance industry's large-scale customer service system, the number of complaint calls continues to grow, and emotional expressions, complex business terminology, and multi-round conversations have become common phenomena, making the automated identification and classification of complaint content an important requirement for industry operations.

[0003] Existing technologies often analyze insurance complaint calls using speech recognition models that rely solely on a single path. For example, some text classification methods use only the transcribed text output by an acoustic model. These methods input the recognized call text into a traditional shallow semantic classifier, using statistical features or fixed template rules to determine the complaint type, thus outputting the complaint type based on the structured text. However, because these methods ignore crucial acoustic elements in speech expression, such as tone variations, emotional fluctuations, and pronunciation differences, they fail to accurately capture the true intent of complaining customers when faced with complex speech behaviors frequently encountered in insurance customer service scenarios, such as irony, emotional agitation, and homophonic misidentification. This often results in low complaint classification accuracy.

[0004] Therefore, there is an urgent need for a method and device for intelligent analysis of complaint calls based on NLP and ASR. Summary of the Invention

[0005] This application provides a method and apparatus for intelligent analysis of complaint calls based on NLP and ASR, which solves the problem that relying solely on a single-path speech recognition model to analyze complaint calls in the insurance field cannot accurately capture the true intent of complaining customers and is prone to low accuracy in complaint classification.

[0006] The first aspect of this application provides an intelligent analysis method for insurance complaint calls based on NLP and ASR. The method includes: acquiring insurance complaint call data and inputting the insurance complaint call data into an acoustic model and a semantic model respectively; outputting pronunciation features related to insurance terminology based on the acoustic model and outputting semantic generalization features related to the insurance field based on the semantic model; performing cross-modal fusion processing on the pronunciation features and semantic generalization features and outputting dual-path fusion features; inputting the dual-path fusion features into a multimodal complaint classification model and outputting complaint analysis results corresponding to the insurance complaint call data, the complaint analysis results including complaint type, complaint volume, and complaint trend.

[0007] Optionally, before acquiring insurance complaint call data and inputting it into the acoustic and semantic models respectively, the acoustic and semantic models are trained by: constructing an insurance-domain speech-text pairing training set using insurance business speech resource data; the insurance business speech resource data includes customer service recording data, announcer recording data, and synthesized speech data; the insurance-domain speech-text pairing training set includes target insurance terms and target insurance customer service dialogue scenarios; and training the acoustic and semantic models based on the insurance-domain speech-text pairing training set.

[0008] Optionally, the acoustic model is trained based on a speech-text pairing training set in the insurance field. Specifically, this includes: adjusting the parameters of the acoustic model based on the acoustic performance features in the speech-text pairing training set in the insurance field; the acoustic performance features include the pronunciation features and tone features of insurance terms; and jointly optimizing the parameter-adjusted acoustic model based on the speech-text correspondence in the speech-text pairing training set in the insurance field.

[0009] Optionally, before performing cross-modal fusion processing on pronunciation features and semantic overarching features and outputting dual-path fused features, the method further includes: training an insurance domain language model based on insurance domain text resources; and introducing a language score from the insurance domain language model during the generation of pronunciation features to enhance the pronunciation features in the business context.

[0010] Optionally, cross-modal fusion processing is performed on the pronunciation features and semantic overarching features, and dual-path fusion features are output. Specifically, this includes: using an attention mechanism to assign weights to the pronunciation features and semantic overarching features according to a preset fusion strategy; performing cross-modal fusion processing on the pronunciation features and semantic overarching features according to the assigned weights, and outputting dual-path fusion features.

[0011] Optionally, the dual-path fusion features are input into the multimodal complaint classification model, and the complaint analysis results corresponding to the insurance complaint call data are output. Specifically, this includes: based on the dual-path fusion features, the text sentiment score and voice sentiment score corresponding to the insurance complaint call data are output through the multimodal complaint classification model; the complaint risk value is calculated based on the text sentiment score, voice sentiment score, and the interaction term between the text sentiment score and voice sentiment score; and the complaint analysis results are output through the complaint risk value.

[0012] Optionally, after inputting the dual-path fusion features into the multimodal complaint classification model and outputting the complaint analysis results corresponding to the insurance complaint telephone data, the method further includes: determining the risk range corresponding to the complaint risk value, and determining the corresponding complaint handling strategy based on the risk range. Specifically, if the risk range corresponding to the complaint risk value is determined to be at the first risk threshold, the corresponding complaint handling strategy is determined to be the first strategy, which is to execute a low-intervention level complaint handling process; if the risk range corresponding to the complaint risk value is determined to be at the second risk threshold, the corresponding complaint handling strategy is determined to be the second strategy, which is to execute an expedited level complaint handling process; if the risk range corresponding to the complaint risk value is determined to be at the third risk threshold, the corresponding complaint handling strategy is determined to be the third strategy, which is to execute a priority escalation complaint handling process; the first risk threshold, the second risk threshold, and the third risk threshold are multiple non-overlapping intervals divided according to the order of risk level.

[0013] A second aspect of this application provides an intelligent analysis device for complaint calls based on NLP and ASR. The device includes an acquisition module and a processing module, wherein... The acquisition module is used to acquire insurance complaint call data and input the insurance complaint call data into the acoustic model and semantic model respectively; based on the acoustic model, it outputs pronunciation features related to insurance terminology, and based on the semantic model, it outputs semantic generalization features related to the insurance field.

[0014] The processing module is used to perform cross-modal fusion processing on pronunciation features and semantic comprehensive features, and output dual-path fusion features. The dual-path fusion features are input into the multimodal complaint classification model, and the complaint analysis results corresponding to the insurance complaint telephone data are output. The complaint analysis results include complaint type, complaint volume and complaint trend.

[0015] A third aspect of this application provides an electronic device including a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described above.

[0016] A fourth aspect of this application provides a computer-readable storage medium storing a computer program, which is executed by a processor using the method described in any of the foregoing descriptions.

[0017] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages: 1. Acquire insurance complaint call data and input it into the acoustic model and semantic model respectively; output pronunciation features related to insurance terminology based on the acoustic model, and output semantic generalization features related to the insurance field based on the semantic model; perform cross-modal fusion processing on the pronunciation features and semantic generalization features, and output dual-path fusion features; input the dual-path fusion features into the multimodal complaint classification model, and output the complaint type, complaint volume, and complaint trend corresponding to the insurance complaint call data. This allows for high-precision and structured automatic parsing of complaint content based on dual information sources of voice and text without relying on manual analysis of each call. The system can instantly identify the customer's business intent, emotional state, and potential risk changes during real-time calls, and further realize full-link monitoring and trend analysis of the complaint situation, significantly improving the efficiency and accuracy of complaint handling.

[0018] 2. The acoustic model is parameter-adjusted based on the acoustic performance features in the speech-text pairing training set in the insurance field. The acoustic performance features include the pronunciation features and tone features of insurance terms. The acoustic model after parameter adjustment is jointly optimized based on the speech-text correspondence in the speech-text pairing training set in the insurance field. This enables the acoustic model to accurately learn key acoustic patterns such as the pronunciation of terms, tone changes and emotional fluctuations in insurance complaint scenarios, and to form a more stable pronunciation-to-semantics mapping under the speech-text correspondence constraint.

[0019] 3. Based on the dual-path fusion feature, the system outputs text sentiment scores and voice sentiment scores corresponding to insurance complaint call data through a multimodal complaint classification model; it calculates complaint risk values ​​based on text sentiment scores, voice sentiment scores, and the interaction items between text sentiment scores and voice sentiment scores; and it outputs complaint analysis results through complaint risk values. This enables the system to comprehensively examine changes in customer speech content and tone in complaint scenarios with complex emotional expressions, interwoven multi-turn dialogues, and implied business intentions. It achieves a refined quantitative judgment on the intensity of customer emotions, potential risk trends, and the possibility of complaint escalation. This allows the system to not only identify the category of the complaint but also to detect the trend of deteriorating customer emotions in advance and trigger corresponding intervention strategies. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating an intelligent analysis method for complaint calls based on NLP and ASR provided in an embodiment of this application. Figure 2 This is a schematic diagram of a module of an intelligent complaint call analysis device based on NLP and ASR provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0021] Explanation of reference numerals in the attached drawings: 21. Acquisition module; 22. Processing module; 301. Processor; 302. Communication bus; 303. User interface; 304. Network interface; 305. Memory. Detailed Implementation

[0022] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0023] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0024] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0025] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0026] Please refer to Figure 1 The diagram illustrates a flowchart of an intelligent analysis method for complaint calls based on NLP and ASR provided in this application embodiment. The flowchart mainly includes the following steps: S101 to S104.

[0027] Step S101: Obtain insurance complaint call data and input the insurance complaint call data into the acoustic model and semantic model respectively.

[0028] Specifically, in actual insurance business operations, after a customer dials the customer service hotline, the system receives the voice stream in real time throughout the entire conversation between the customer and the agent. The voice recording is transmitted to the backend processing node with a millisecond delay via a front-end recording gateway, enabling the system to receive continuous complaint voice data while the communication is still ongoing. Upon receiving this real-time voice data, the system can synchronously input it into the acoustic and semantic models without manual intervention. The acoustic model processes the customer's pronunciation patterns, tone fluctuations, and speech rhythm during the call, while the semantic model processes the customer's insurance-related statements, word order logic, and complaint expression methods.

[0029] Based on the above scenario, this application achieves fully automated transcription and analysis through high-precision speech recognition (ASR) and natural language processing (NLP), improving processing efficiency by more than 50%. First, insurance complaint call data is acquired. In this process, the system first acquires the insurance complaint call data through the speech processing module, and immediately inputs it into the acoustic model and semantic model respectively when the voice stream arrives, thereby establishing the dual-path feature foundation required for subsequent multimodal analysis. The speech processing module is deployed on the background processing node at the call access side. After receiving continuous voice data, it processes the voice according to a preset audio acquisition specification, including a sampling rate of 16kHz, a bit depth of 16bit, and PCM or WAV audio encoding. After audio acquisition, the speech processing module performs preprocessing procedures on the voice data, including spectral subtraction noise reduction (e.g., window length 25ms, step size 10ms), energy-based speech activity detection, and speaker separation based on the GMM-UBM model, to ensure that the voice data input to the acoustic and semantic models has stable signal quality and effective speech segments.

[0030] In one possible implementation, step S101 further includes: constructing an insurance-domain voice-text pairing training set using insurance business voice resource data; the insurance business voice resource data includes customer service recording data, announcer recording data, and synthesized voice data; the insurance-domain voice-text pairing training set includes target insurance terms and target insurance customer service dialogue scenarios; and training an acoustic model and a semantic model based on the insurance-domain voice-text pairing training set.

[0031] Specifically, the voice resource data for insurance business comes from customer call recordings in real insurance business. These recordings are anonymized to retain voice expression characteristics without including user privacy. Additionally, voice recordings of insurance business scenarios by professional broadcasters and synthesized voice generated from text scripts are used as supplementary data to ensure diversity in the voice sources of the voice-text pairing training set in the insurance field. When constructing the voice-text pairing training set, voice-text segments containing target insurance terms such as lump-sum payment, cash value, underwriting, and claims are selected, along with target insurance customer service dialogues such as sales communication, claims consultation, and customer service processing. The scenario-related speech and text content enables the training set to include common sentence structures, terminology combinations, and customer expression patterns in insurance business. When training the acoustic model, it allows the acoustic model to learn the pronunciation patterns, tone variations, and acoustic expression rules of insurance terms based on the training set. When training the semantic model, it allows the semantic model to learn the semantic structure, word order relationships, and semantic logic of insurance business based on the training set. After training, the acoustic and semantic models can more accurately process the insurance complaint telephone data obtained in step S101, thereby providing a feature foundation consistent with the insurance business scenario for subsequent multimodal analysis.

[0032] Step S102: Output pronunciation features related to insurance terminology based on the acoustic model, and output semantic overarching features related to the insurance field based on the semantic model.

[0033] Specifically, after receiving preprocessed insurance complaint call data, the acoustic model generates pronunciation features that represent the customer's pronunciation, tone changes, and acoustic performance of the target insurance terms during the complaint process, based on the pronunciation patterns of insurance terms learned from the insurance domain speech-text pairing training set. The semantic model performs semantic understanding processing on the text content related to the insurance complaint scenario based on the learning of the semantic structure and word order logic related to insurance business, and generates semantic overarching features to represent the semantic relationships, business intentions, and contextual logic in the insurance domain.

[0034] In one possible implementation, step S102 further includes: constructing an insurance-domain voice-text pairing training set using insurance business voice resource data; the insurance business voice resource data includes customer service recording data, announcer recording data, and synthesized voice data; the insurance-domain voice-text pairing training set includes target insurance terms and target insurance customer service dialogue scenarios; and training an acoustic model and a semantic model based on the insurance-domain voice-text pairing training set.

[0035] Specifically, the datasets used to build the model primarily come from relational databases and graph databases. The relational database uses MySQL 8.0 as its core engine, leveraging its high-concurrency transaction processing capabilities and index optimization mechanisms to ensure real-time writing and stable retrieval performance even under high-frequency data entry scenarios for large-scale insurance complaint records. The graph database uses Neo4j 4.3 to construct a knowledge graph for the insurance domain, storing information such as insurance terms, business process nodes, rule relationships, and complaint paths in a graph structure. This allows the semantic model to enhance context based on graph database query results when processing word order logic and business meaning in customer statements, thereby improving the accuracy and consistency of semantic overarching features in subsequent cross-modal fusion. To clearly distinguish the acoustic and semantic models in terms of learning objectives and feature domains, this implementation adopts differentiated data mapping methods and optimization mechanisms for the two types of models. To train the acoustic model, continuous speech windows containing pronunciation segments of target insurance terms are selected from customer service recording data and constructed as an acoustic input matrix. This matrix represents the dynamic spectral changes of speech over time; the acoustic model is optimized based on the following feature energy alignment algorithm:

[0036] in, For the acoustic model Acoustic parameters under the next iteration For the acoustic model Acoustic parameters under the next iteration This is the set of trainable parameters for the acoustic model, such as convolution kernel coefficients to characterize speech spectrum patterns, temporal mapping weights to describe changes in pronunciation timing, and feature transformation matrices to model acoustic energy distribution. This set of parameters is continuously adjusted during training as the acoustic features of customer service recording data are learned, enabling the acoustic model to capture the pronunciation structure, intonation changes, and prosodic patterns of the target insurance terms, thereby forming the pronunciation features required for subsequent cross-modal fusion. For acoustic learning rate, The pronunciation spectrum of the terms actually appearing in customer service recordings. This is the energy distribution sequence calculated by the acoustic model on the input speech matrix. For the target text tags of the corresponding terms, a text-side energy template is generated after dictionary mapping. This represents the L2 distance. This optimization method enables the acoustic model to learn the correspondence between pronunciation energy patterns and term labels, thereby recognizing the pronunciation patterns of insurance terms in complaint scenarios.

[0037] To train the semantic model, semantic input sequences were extracted from the text corresponding to customer service recordings, the text recorded by announcers, and the synthesized speech scripts. And map it to a semantic dependency graph structure The semantic model is based on the following semantics Figure 1 Consistency algorithm for parameter optimization:

[0038] in, For the semantic model Semantic parameters in the next iteration For the semantic model Semantic parameters in the next iteration This is a set of trainable parameters for the semantic model, such as vector weights for representing semantic dependencies, attention weight matrices for representing contextual associations, and mapping matrices for constructing semantic overarching structures. This set of parameters is continuously iteratively optimized to approximate the real semantic structure in insurance complaint scenarios, enabling the semantic model to extract complaint intent, business logic relationships, and word order patterns at the text level, thereby generating semantic overarching features for cross-modal fusion. For semantic learning rate, This is a semantic input sequence containing target insurance terms and their business logic relationships within the complaint text. This is a semantic dependency graph built based on insurance customer service dialogue scenarios. These are semantic tag sequences used to describe the intent relationships, syntactic structure, and word order logic within the context of a complaint. The semantic distribution generated by the semantic model based on the semantic dependency graph. The target semantic distribution generated for the tags. The KL divergence is the distance between the two distributions. This optimization method enables the semantic model to learn the correspondence between semantic structure and complaint intent. Specifically, the pronunciation features that can be extracted in this embodiment include, but are not limited to, energy distribution sequences characterizing the pronunciation of insurance terminology, spectral energy gradient features reflecting changes in the intensity of a customer's tone in the complaint context, and temporal resonance features describing speech rhythm and prosodic patterns. The semantic overarching features that can be extracted in this embodiment include, but are not limited to, semantic dependency distributions summarizing the logical relationships in insurance business, semantic hierarchical features unifying the representation of syntactic structure and intent links, and context overarching vectors abstracting the semantic backbone of customer requests.

[0039] In one possible implementation, step S102 further includes: adjusting the parameters of the acoustic model based on the acoustic performance features in the insurance domain speech-text pairing training set; the acoustic performance features include the pronunciation features and tone features of insurance terms; and jointly optimizing the parameter-adjusted acoustic model based on the speech-text correspondence in the insurance domain speech-text pairing training set.

[0040] Specifically, in this embodiment, when training the acoustic model, the actual pronunciation segments containing the target insurance terms are first selected from customer service recording data. These segments reflect the pronunciation morphology, tone changes, and speech energy distribution of customers when using insurance terms in real complaint calls. After selecting the above pronunciation segments, the continuous speech segments are constructed into an acoustic input matrix to represent the spectral change pattern of the speech segment in the time dimension, so that the acoustic model can learn the acoustic performance features related to the pronunciation of insurance terms from the dynamic structure of the spectrum.

[0041] During the training of the acoustic model, the parameter set of the acoustic model is used to characterize the local changes and overall structure of the speech spectrum. For example, it includes convolution kernel coefficients to describe information in different frequency bands, time mapping weights to represent the time-dimensional variation of speech pronunciation, and feature transformation matrices to construct the relationship of speech energy distribution. This parameter set is continuously updated during training iterations, and the model minimizes the difference between the energy distribution of the acoustic input matrix and the energy template mapped from the target text label by calculating the L2 distance between them during parameter adjustment. Through this optimization method based on energy distribution differences, the acoustic model can gradually learn the pronunciation and tone characteristics of insurance terms in actual complaint scenarios.

[0042] After initial parameter adjustments, the acoustic model is further jointly optimized based on the one-to-one correspondence between speech and text. In this stage, the model not only learns the acoustic energy changes of speech segments, but also learns the position, semantic background and terminology of the acoustic segment in the text sequence, so as to establish a closer mapping relationship between acoustic features and text labels.

[0043] Based on the above explanation, after training, the acoustic model can not only recognize the accurate pronunciation of insurance terms, but also distinguish the acoustic performance of different tone changes in the context of complaints, thus providing a more business-consistent pronunciation feature foundation for subsequent cross-modal fusion.

[0044] In one possible implementation, step S102 further includes: training an insurance domain language model based on insurance domain text resources; and introducing a language score from the insurance domain language model during the generation of pronunciation features to enhance the pronunciation features in the business context.

[0045] Specifically, this embodiment first constructs an acoustic input matrix using target insurance terms and their actual pronunciation segments from a speech-text pairing training set in the insurance field. Based on this matrix, the trainable parameters of the acoustic model are adjusted so that it can capture the pronunciation change patterns and tone variations of insurance terms in real complaint scenarios during the learning process. After completing the initial parameter adjustment based on acoustic performance features, the acoustic model is then jointly optimized using the one-to-one correspondence between speech and text in the training set. This allows the acoustic model to focus on acoustic structure while strengthening the alignment between term pronunciation and text labels, thereby improving the accuracy of insurance term recognition.

[0046] Based on this embodiment, to further enhance the robustness of the model in insurance terminology recognition, an insurance-domain language model is introduced to correct and enhance the output of the acoustic model. During the recognition process, the pronunciation features generated by the acoustic model are mapped to candidate text sequences, and the contextual probability distribution is provided by the insurance-domain language model. By combining the acoustic model score and the language model score in a shallow fusion manner, the accuracy of insurance terminology recognition can be significantly improved, enabling the model to automatically correct common erroneous expressions in business scenarios. The score calculation for shallow fusion can be expressed as:

[0047] in, The final score after fusion is used to select the optimal text sequence from multiple candidate texts; The score is calculated by the acoustic model based on the input speech segment, which reflects the degree of pronunciation matching between the speech and the candidate word sequence; Context scores generated for language models in the insurance field are used to reflect the rationality of candidate word sequences in the context of insurance business. The fusion weight parameter, whose value is between 0 and 1, is used to control the proportion of influence of the acoustic side and the language side on the final recognition result.

[0048] In the joint optimization phase, the parameters of the acoustic model are further enhanced through training, enabling it to learn the synergistic relationship between acoustic and linguistic features under the drive of a shallow fusion mechanism. The optimization expression used in this phase is:

[0049] in, For the acoustic model in the first The parameter set in the next iteration includes convolution kernel coefficients for capturing the speech spectral structure, time mapping weights for modeling speech temporal variations, and feature transformation matrices for describing energy distribution. The acoustic learning rate is used to control the step size for updating model parameters; The voice matrix extracted from customer service recording data contains pronunciation segments of target insurance terms; This is the energy distribution sequence generated by the acoustic model from the input speech matrix; To generate energy templates for the text labels of target terms in the training set, a dictionary mapping is used. The energy template corresponding to the target text; The L2 distance is used to measure the difference between the acoustic output and the target template. The language model loss is used to measure the reasonableness of candidate text sequences in the insurance semantic space; These are the weighting parameters used to balance the acoustic loss and the language model loss.

[0050] Through the improved algorithm described above, this embodiment can capture the pronunciation and tone features of insurance terms at the acoustic layer, and supplement the prior knowledge of business context by using a language model in the insurance field at the semantic layer. This significantly reduces homophonic misidentification, pronunciation weakening misidentification, and semantic deviation misidentification during the recognition stage, resulting in a significant improvement in the final output text sequence in terms of professional terminology recognition rate, semantic coherence, and business consistency. It also provides high-quality basic features for subsequent complaint intent analysis, emotion recognition, and risk assessment.

[0051] Step S103: Perform cross-modal fusion processing on pronunciation features and semantic overarching features, and output dual-path fused features.

[0052] Specifically, when both the pronunciation features generated by the acoustic model and the semantic features generated by the semantic model are available, the two types of features are fed into the cross-modal fusion module. The pronunciation-side information and semantic-side information are dynamically weighted using a preset attention weighting method, enabling the system to automatically adjust the feature contribution based on the expressive intensity, tone changes, and semantic logic related to insurance business in the complaint speech. After feature weighting, cross-modal mapping is performed on the two types of features, so that the pronunciation pattern of the acoustic layer and the business logic of the semantic layer are jointly expressed in the same feature space, thereby generating dual-path fusion features that can simultaneously reflect the speech performance and the semantic relationship of the text, providing a comprehensive and robust feature foundation for subsequent complaint emotion recognition and risk assessment.

[0053] In one possible implementation, step S103 further includes: using an attention mechanism to assign weights to pronunciation features and semantic overarching features according to a preset fusion strategy; performing cross-modal fusion processing on the pronunciation features and semantic overarching features according to the assigned weights, and outputting dual-path fusion features.

[0054] Specifically, this embodiment, based on the conventional attention mechanism and combined with the particularity of the insurance complaint telephone scenario, introduces cross-modal sensitive factors such as tone intensity, pronunciation stability, semantic conflict degree and terminology confidence. This makes the attention weight not only affected by the pronunciation features and semantic overarching features themselves, but also by the complaint emotion, terminology context consistency and cross-modal conflict degree, thereby constructing a dynamic fusion model that can adapt to real insurance complaint dialogues.

[0055] This embodiment first calculates the credibility evaluation of pronunciation feature vectors and semantic overarching feature vectors in an insurance complaint scenario, and then introduces a cross-modal consistency factor. This enables the system to perceive whether the spoken expression matches the semantics of the text. When a customer exhibits situations such as "neutral text but angry tone" or "strong text but flat tone," this application obtains the cross-modal conflict value through the following methods:

[0056] in, For pronunciation feature vectors; To use semantic mapping functions Mapping semantically encompassing features to a vector in the acoustic domain; This represents the cross-modal deviation between acoustics and semantics; This refers to the speech rhythm and prosody offset, used to measure the fluctuations in customer emotions and the intensity of speech expression. The confidence difference of insurance terms involved in the semantic overarching features on the speech side is used to determine whether the customer has incomplete pronunciation of terms or unclear stress. This is the adjustment factor for cross-modal bias, used to balance the contributions of acoustic bias, prosodic bias, and terminological bias to the degree of conflict. A larger value indicates a greater inconsistency between acoustics and semantics. In this case, the acoustic weight should be increased to avoid the language model becoming overly dominant.

[0057] To achieve dynamic fusion, this application constructs an adaptive attention mechanism, ensuring that the pronunciation feature weights and semantic overarching feature weights depend not only on their internal representations but also on the degree of cross-modal conflict, tone intensity, and term semantic density. The pronunciation-side attention weights are calculated as follows:

[0058] The semantic attention weights are calculated as follows:

[0059] in, This is the attention mapping matrix for pronunciation features, used to project pronunciation features into the attention space; This is the attention mapping matrix for semantically governed features, used to project semantic features into the attention space; It is a multimodal modulator used to control the effects of cross-modal conflict, tone intensity, and semantic density on attention allocation; This is a tone intensity index calculated based on pronunciation features, used to represent the intensity of a customer's emotional changes in a complaint scenario; The terminological density index, calculated based on semantic overarching features, is used to evaluate whether the current text is in a strong semantic region, such as containing multiple consecutive insurance keywords. Through this structure, the system can identify various real-world complaint scenarios, such as "strong speech but weak semantics," "strong semantics but weak speech," and "both strong speech and semantics," and make differentiated fusion decisions.

[0060] After obtaining the dynamic weights, this embodiment generates the final fused features:

[0061] in, This dual-path fusion feature, generated by comprehensively considering the emotional expression, semantic strength of terms, and cross-modal consistency of the current complaint call, enables the system to more completely describe the implied and unspoken meanings in customer complaints, providing high-precision input for subsequent complaint type determination and risk value calculation. This mechanism overcomes the shortcomings of existing algorithms in correctly handling irony, homophonic discrepancies in terminology, and semantic ambiguity, giving the model a dynamic understanding capability for real insurance complaints.

[0062] Step S104: Input the dual-path fusion features into the multimodal complaint classification model and output the complaint analysis results corresponding to the insurance complaint telephone data.

[0063] Specifically, the dual-path fusion features are used as input to the multimodal complaint classification model. Based on the pronunciation features and semantic overarching features contained in the dual-path fusion features, the multimodal complaint classification model comprehensively judges the current insurance complaint telephone data and outputs complaint analysis results that characterize the classification of complaint content, the statistics of the number of complaints over time, and the trend of complaint changes over time. This provides a structured output for the insurance institution's complaint monitoring and operational decision-making. The complaint analysis results include complaint type, complaint volume, and complaint trend.

[0064] The complaint types in the complaint analysis results are used to characterize the current insurance complaint calls in terms of business scope. For example, complaints can be automatically categorized into preset categories such as policy maintenance, claims, sales misrepresentation, and service attitude, or further subdivided into subcategories such as "policy cancellation disputes," "claims progress disputes," and "misunderstandings of policy terms." This allows operations personnel to directly understand which type of business issue each complaint call mainly focuses on. The complaint type output is usually presented in the form of category probability distribution and final decision category to support subsequent business statistics and strategy adjustments.

[0065] The complaint volume in the complaint analysis results is used to statistically analyze the number of complaints within a given time window. For example, it can be used to count the number of calls corresponding to various complaint types by day, week, or month, forming a multi-dimensional statistical indicator of complaint volume. The complaint volume not only includes the total number of complaints, but can also be combined with the complaint type to count the frequency of occurrence of various complaints in different time periods. This helps insurance institutions identify whether there is a concentrated outbreak of complaints for a certain product or a certain process within a specific time period, which can support resource allocation and emergency response.

[0066] The complaint trend analysis results are used to characterize the dynamic trend of complaint volume over time. For example, by smoothing, fitting, and analyzing the rate of change of complaint volume over historical time series, it can be determined whether the total number of complaints and various types of complaints will show an upward, stable, or downward trend in the future. The complaint trend can be represented by trend curves, slopes of change, or inflection points, helping insurance institutions to anticipate potential deterioration in service quality or product issues and whether proactive risk intervention or process optimization is necessary, thereby achieving proactive management of complaint risk. In one possible implementation, step S104 further includes: based on the dual-path fusion features, outputting the text sentiment score and voice sentiment score corresponding to the insurance complaint telephone data through a multimodal complaint classification model; calculating the complaint risk value based on the text sentiment score, voice sentiment score, and the interaction item between the text sentiment score and voice sentiment score; and outputting the complaint analysis result through the complaint risk value.

[0067] Specifically, this embodiment first constructs cross-modal emotion latent variables based on text emotion scores and speech emotion scores, and maps the two types of emotion information to a unified emotion space in the following way:

[0068] in, This represents a cross-modal emotion latent variable, used to characterize the combined emotional state of text and speech in insurance complaint scenarios; The sentiment score for the text; The emotional score for the voice; This represents the extended emotion vector formed by combining text emotion, voice emotion, and their product. This is the latent variable mapping matrix, used to project the emotion extension vector onto the latent emotion space; This is a bias term.

[0069] Building upon this, this embodiment further calculates the degree of inconsistency between speech and semantics, and constructs a semantic conflict penalty term by using the cross-modal difference between speech features and semantic overarching features:

[0070] in, It indicates the degree of semantic conflict and is used to characterize the deviation between the emotional tone of speech and the semantics of text. These are pronunciation characteristics; It is a semantic overarching feature; A mapping function that maps semantic features to acoustic space; It represents the absolute value norm, used to measure whether there is inconsistency between text and speech, such as "speech neutral but strong tone".

[0071] To enable risk assessment to identify highly sensitive terms in the insurance field, this embodiment introduces a term sensitivity adjustment parameter, allowing the system to generate a stronger risk response when identifying highly sensitive terms such as "policy surrender" and "claims dispute."

[0072] in, Characterizing terminology sensitivity; Embedded vectors for insurance-sensitive terms; It features dual-path fusion; This is the activation function used to keep the term sensitivity output stable.

[0073] To capture the risk trend of emotional changes over time during complaint conversations, this embodiment constructs an emotional temporal change quantity to characterize the dynamic process of "emotional escalation or sudden intensity" during a customer's call:

[0074] in, This represents the temporal variation of emotions. This represents the change in text sentiment. This refers to the amount of change in vocal emotion. and This is a temporal attention parameter used to measure the contribution of different sources of emotional change to overall risk change.

[0075] The final complaint risk value is calculated based on the aforementioned cross-modal emotional latent variables, semantic conflict degree, terminology sensitivity, and temporal change in emotion:

[0076] in, This indicates the complaint risk value, used to reflect the emotional complexity and business semantic risks during the call; Risk weights are latent variables used to control for the impact of overall sentiment on risk. This is a semantic conflict risk weight, used to amplify the risk increase caused by verbal inconsistency; This is a terminology sensitivity weight, used to amplify the risk impact of key insurance terms; Weights for temporal changes are used to capture risk increases resulting from changes in emotional trends. This is a risk bias term.

[0077] In one possible implementation, step S104 further includes: determining the risk range corresponding to the complaint risk value, and determining the corresponding complaint handling strategy based on the risk range, specifically including: if it is determined that the risk range corresponding to the complaint risk value is at a first risk threshold, then the corresponding complaint handling strategy is determined to be a first strategy, which is to execute a low-intervention level complaint handling process; if it is determined that the risk range corresponding to the complaint risk value is at a second risk threshold, then the corresponding complaint handling strategy is determined to be a second strategy, which is to execute an expedited level complaint handling process; if it is determined that the risk range corresponding to the complaint risk value is at a third risk threshold, then the corresponding complaint handling strategy is determined to be a third strategy, which is to execute a priority escalation complaint handling process; the first risk threshold, the second risk threshold, and the third risk threshold are multiple non-overlapping intervals divided in order of risk level.

[0078] Specifically, in this embodiment, after calculating the complaint risk value, the system automatically determines the risk range into which the complaint risk value falls based on the relationship between the complaint risk value and preset first, second, and third risk thresholds, and triggers a complaint handling strategy matching the risk level accordingly. When the complaint risk value is within the risk range corresponding to the first risk threshold, the system classifies the current complaint as a low-risk, routine complaint and automatically determines the corresponding complaint handling strategy as the first strategy. The first strategy involves executing a low-intervention complaint handling process, such as assigning the complaint to a regular agent queue and registering, explaining, and providing feedback according to the standard acceptance order. The system mainly focuses on recording the complaint content and forming basic analytical data, without consuming additional high-level resources or introducing special intervention measures, thus ensuring overall service efficiency while controlling processing costs.

[0079] When the complaint risk value falls within the risk range corresponding to the second risk threshold, the system classifies the current complaint as a sensitive complaint with a medium risk level and automatically determines the corresponding complaint handling strategy as the second strategy. The second strategy is to execute an expedited complaint handling process. For example, in the work order generation and circulation process, a shorter response time and processing time limit are assigned to the complaint, and the complaint is prioritized to a complaint specialist or team with higher business experience. At the same time, a processing progress monitoring mark is added to the system so that if the complaint is not resolved on time or the customer's mood continues to deteriorate, the relevant handling personnel can be quickly reminded to make a follow-up visit, thereby intervening before the complaint evolves into a high-risk event.

[0080] When the complaint risk value falls within the risk range corresponding to the third risk threshold, the system classifies the current complaint as a high-risk, critical complaint and automatically determines the corresponding complaint handling strategy as the third strategy. The third strategy is to implement a priority escalation complaint handling process, such as automatically placing the complaint in the highest priority channel, pushing the complaint information to authorized managers or special handling teams, and triggering a multi-level linkage mechanism in the system, so that subsequent verification, review, remedial measures, and customer reassurance work can be initiated in a shorter time. At the same time, the system can internally mark the complaint as a potential major complaint or a potential source of public opinion risk, so as to track it closely in subsequent statistical analysis and trend judgment, thereby building a risk-driven closed-loop complaint management system at the business operation level.

[0081] The first, second, and third risk thresholds are divided into multiple non-overlapping intervals according to the degree of risk from low to high, so that any complaint risk value corresponds to only one clear risk value range, avoiding situations where strategies overlap or handling disagreements occur.

[0082] Please refer to Figure 2 This illustration shows a schematic diagram of a complaint call intelligent analysis device based on NLP and ASR provided in an embodiment of this application. The device includes an acquisition module 21 and a processing module 22, wherein... The acquisition module 21 is used to acquire insurance complaint telephone data and input the insurance complaint telephone data into the acoustic model and the semantic model respectively; based on the acoustic model, it outputs the pronunciation features related to insurance terminology, and based on the semantic model, it outputs the semantic general features related to the insurance field.

[0083] Processing module 22 is used to perform cross-modal fusion processing on pronunciation features and semantic comprehensive features, and output dual-path fusion features; input the dual-path fusion features into the multimodal complaint classification model, and output the complaint analysis results corresponding to the insurance complaint telephone data. The complaint analysis results include complaint type, complaint volume and complaint trend.

[0084] In one possible implementation, the acquisition module 21 is used to train the acoustic model and semantic model before acquiring insurance complaint call data and inputting the insurance complaint call data into the acoustic model and semantic model respectively: constructing an insurance domain voice-text pairing training set through insurance business voice resource data; the insurance business voice resource data includes customer service recording data, announcer recording data, and synthesized voice data; the insurance domain voice-text pairing training set includes target insurance terms and target insurance customer service dialogue scenarios; and training the acoustic model and semantic model based on the insurance domain voice-text pairing training set.

[0085] In one possible implementation, the acquisition module 21 is used to train an acoustic model based on a speech-text pairing training set in the insurance field, specifically including: adjusting the parameters of the acoustic model based on the acoustic performance features in the speech-text pairing training set in the insurance field; the acoustic performance features include pronunciation features and tone features of insurance terms; and jointly optimizing the parameter-adjusted acoustic model based on the speech-text correspondence in the speech-text pairing training set in the insurance field.

[0086] In one possible implementation, before performing cross-modal fusion processing on pronunciation features and semantic overarching features and outputting dual-path fused features, the processing module 22 further includes: training an insurance domain language model based on insurance domain text resources; and introducing a language score from the insurance domain language model during the generation of pronunciation features to enhance the pronunciation features in the business context.

[0087] In one possible implementation, the processing module 22 is used to perform cross-modal fusion processing on pronunciation features and semantic overarching features, and output dual-path fusion features. Specifically, it includes: using an attention mechanism to assign weights to pronunciation features and semantic overarching features according to a preset fusion strategy; performing cross-modal fusion processing on pronunciation features and semantic overarching features according to the assigned weights, and outputting dual-path fusion features.

[0088] In one possible implementation, the processing module 22 is used to input the dual-path fusion features into the multimodal complaint classification model and output the complaint analysis results corresponding to the insurance complaint call data. Specifically, it includes: based on the dual-path fusion features, outputting the text sentiment score and voice sentiment score corresponding to the insurance complaint call data through the multimodal complaint classification model; calculating the complaint risk value based on the text sentiment score, voice sentiment score, and the interaction term between the text sentiment score and voice sentiment score; and outputting the complaint analysis results through the complaint risk value.

[0089] In one possible implementation, after the processing module 22 inputs the dual-path fusion features into the multimodal complaint classification model and outputs the complaint analysis results corresponding to the insurance complaint telephone data, the method further includes: determining the risk range corresponding to the complaint risk value, and determining the corresponding complaint handling strategy based on the risk range. Specifically, if it is determined that the risk range corresponding to the complaint risk value is at a first risk threshold, then the corresponding complaint handling strategy is determined to be the first strategy, which is to execute a low-intervention level complaint handling process; if it is determined that the risk range corresponding to the complaint risk value is at a second risk threshold, then the corresponding complaint handling strategy is determined to be the second strategy, which is to execute an expedited level complaint handling process; if it is determined that the risk range corresponding to the complaint risk value is at a third risk threshold, then the corresponding complaint handling strategy is determined to be the third strategy, which is to execute a priority escalation complaint handling process; the first risk threshold, the second risk threshold, and the third risk threshold are multiple non-overlapping intervals divided in order of risk level.

[0090] It should be noted that the above embodiments of the apparatus are only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided above belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0091] This application also provides an electronic device. (See reference...) Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: at least one processor 301, at least one communication bus 302, a user interface 303, at least one network interface 304, and a memory 305.

[0092] The communication bus 302 is used to enable communication between these components.

[0093] The user interface 303 may include a display screen and a camera. Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0094] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0095] The processor 301 may include one or more processing cores. The processor 301 connects to various parts of the server using various interfaces and lines, and performs various server functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in memory 305, and by calling data stored in memory 305. Optionally, the processor 301 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 301 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content required for display; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 301 and may be implemented as a separate chip.

[0096] The memory 305 may include random access memory (RAM) or read-only memory. Optionally, the memory 305 may include a non-transitory computer-readable storage medium. The memory 305 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 305 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 305 may also be at least one storage device located remotely from the aforementioned processor 301. (Refer to...) Figure 3 The memory 305, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a complaint call intelligent analysis application based on NLP and ASR.

[0097] exist Figure 3In the illustrated electronic device, the user interface 303 is primarily used to provide an input interface for the user and acquire user input data; while the processor 301 can be used to call the NLP and ASR-based intelligent complaint call analysis application stored in the memory 305. When executed by one or more processors 301, the electronic device performs one or more of the methods described in the above embodiments. It should be noted that, for the foregoing method embodiments, for the sake of simplicity, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0098] This application also provides a computer-readable storage medium storing instructions. When executed by one or more processors, these instructions cause an electronic device to perform one or more of the methods described in the above embodiments.

[0099] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0100] In the various embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between apparatuses or units may be electrical or other forms.

[0101] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0102] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0103] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, portable hard drives, magnetic disks, or optical disks.

[0104] The above description is merely an exemplary embodiment disclosed in this application and should not be construed as limiting the scope of this application. Any equivalent changes and modifications made in accordance with the teachings of this application shall still fall within the scope of this application.

[0105] This application is intended to cover any variations, uses, or adaptations disclosed herein that follow the general principles disclosed herein and include common knowledge or customary technical means in the art that are not described in this application.

Claims

1. A method for intelligent analysis of complaint calls based on NLP and ASR, characterized in that, The method includes: Acquire insurance complaint call data, and input the insurance complaint call data into the acoustic model and semantic model respectively; Based on the acoustic model, output pronunciation features related to insurance terminology; and based on the semantic model, output semantic overarching features related to the insurance field. The pronunciation features and the semantic overarching features are subjected to cross-modal fusion processing, and dual-path fused features are output. The dual-path fusion features are input into the multimodal complaint classification model, and the complaint analysis results corresponding to the insurance complaint telephone data are output. The complaint analysis results include complaint type, complaint volume, and complaint trend.

2. The method according to claim 1, characterized in that, Before acquiring insurance complaint call data and inputting it into the acoustic and semantic models respectively, the acoustic and semantic models are trained: A voice-text pairing training set for the insurance field is constructed using voice resource data from insurance business. The voice resource data includes customer service recording data, announcer recording data, and synthesized voice data. The voice-text pairing training set for the insurance field includes target insurance terminology and target insurance customer service dialogue scenarios. The acoustic model and the semantic model are trained based on the speech-text pairing training set in the insurance field.

3. The method according to claim 1, characterized in that, The training of the acoustic model based on the speech-text pairing training set in the insurance field specifically includes: The acoustic model is parameter-adjusted based on the acoustic performance features in the speech-text pairing training set in the insurance field; the acoustic performance features include pronunciation features of insurance terminology and tone features; The acoustic model with adjusted parameters is jointly optimized based on the speech-text correspondence in the speech-text pairing training set in the insurance field.

4. The method according to claim 1, characterized in that, Before performing cross-modal fusion processing on the pronunciation features and the semantic overarching features, and outputting the dual-path fused features, the method further includes: Training an insurance-related language model based on insurance-related text resources; In the process of generating the pronunciation features, a language score from the insurance domain language model is introduced to enhance the pronunciation features in the business context.

5. The method according to claim 1, characterized in that, The process of performing cross-modal fusion processing on the pronunciation features and the semantic overarching features, and outputting dual-path fused features, specifically includes: An attention mechanism is used to assign weights to the pronunciation features and the semantic overarching features according to a preset fusion strategy; The pronunciation features and semantic overarching features are fused across modally according to the assigned weights, and the dual-path fused features are output.

6. The method according to claim 1, characterized in that, The dual-path fusion features are input into a multimodal complaint classification model, and the complaint analysis results corresponding to the insurance complaint call data are output, specifically including: Based on the dual-path fusion features, the text sentiment score and voice sentiment score corresponding to the insurance complaint telephone data are output through the multimodal complaint classification model; The complaint risk value is calculated based on the text sentiment score, the voice sentiment score, and the interaction between the text sentiment score and the voice sentiment score; The complaint analysis results are output based on the complaint risk value.

7. The method according to claim 6, characterized in that, After inputting the dual-path fusion features into the multimodal complaint classification model and outputting the complaint analysis results corresponding to the insurance complaint telephone data, the method further includes: Determine the risk range corresponding to the complaint risk value, and determine the corresponding complaint handling strategy based on the risk range, specifically including: If it is determined that the risk value range corresponding to the complaint risk value is within the first risk threshold, then the corresponding complaint handling strategy is determined to be the first strategy, which is to execute a low-intervention level complaint handling process. If it is determined that the risk value range corresponding to the complaint risk value is within the second risk threshold, then the corresponding complaint handling strategy is determined to be the second strategy, which is to execute the expedited complaint handling process. If it is determined that the risk value range corresponding to the complaint risk value is within the third risk threshold, then the corresponding complaint handling strategy is determined to be the third strategy, which is to execute a complaint handling process that prioritizes escalation; the first risk threshold, the second risk threshold, and the third risk threshold are multiple non-overlapping intervals divided in order of risk level.

8. A complaint call intelligent analysis device based on NLP and ASR, characterized in that, The device includes an acquisition module and a processing module, wherein, The acquisition module is used to acquire insurance complaint telephone data and input the insurance complaint telephone data into an acoustic model and a semantic model respectively; based on the acoustic model, it outputs pronunciation features related to insurance terminology, and based on the semantic model, it outputs semantic overarching features related to the insurance field. The processing module is used to perform cross-modal fusion processing on the pronunciation features and the semantic overarching features, and output dual-path fusion features; input the dual-path fusion features into a multimodal complaint classification model, and output the complaint analysis results corresponding to the insurance complaint telephone data, the complaint analysis results including complaint type, complaint volume and complaint trend.

9. An electronic device, characterized in that, The device includes a processor, a communication bus, a user interface, a network interface, and a memory. The memory is used to store instructions. The user interface and the network interface are used to communicate with other devices. The processor is used to execute the instructions stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1 to 7.

Images