A customer conversation hot word recognition method based on large model technology

Abstract

This invention relates to a method for identifying hot words in customer conversations based on large-scale model technology. The identification method is as follows: model training and construction; users upload batches of historical conversation texts or access conversation streams in real time through the system interface; the system automatically segments the input text into sentences, labels roles, and unifies the encoding format; the trained model processes the preprocessed conversation text; conversation summaries are generated based on hot word graphs; and the results are visualized and output. Compared with existing technologies, this invention achieves accurate identification and classification of the deep meaning of hot words through a multi-dimensional tagging system and semantic understanding; it couples hot word identification and summary generation through multi-task learning and semantic graphs, making the generated summaries both focused and coherent; it supports automated and intelligent analysis of massive conversation data, greatly improving the efficiency and depth of customer sentiment analysis, and providing direct data support for product optimization, service improvement, and risk warning.

Description

Technical Field

[0001] This invention relates to the field of digital data processing technology, and specifically to a method for identifying hot words in customer conversations based on large model technology. Background Technology

[0002] As customer service expands, telephone and online customer service channels generate massive amounts of text and voice conversation data daily. These conversations contain high-value information such as customer feedback, business inquiries, complaints and suggestions, and product defects. Traditional analysis methods mainly rely on keyword statistics and simple sentiment analysis, which are insufficient for deeply understanding the contextual semantics and true intent of the conversations, and have the following limitations:

[0003] (a) The keyword method cannot identify different expressions of the same question and is prone to missing key information.

[0004] (ii) The summary lacks a sense of hierarchy because it cannot distinguish between the facts stated by the client, the demands expressed, and the emotions revealed.

[0005] (iii) For complex and lengthy conversations, it is difficult to extract the core contradictions and development context. The generation of summaries depends on manual labor, which is inefficient and highly subjective.

[0006] Therefore, there is an urgent need for an intelligent method that can deeply understand the content of a conversation, automatically identify the core focus, and generate a structured summary. Summary of the Invention

[0007] To overcome the problems of shallow semantic understanding, inaccurate summary generation, and low efficiency in existing technologies for customer conversation analysis, this invention provides a customer conversation hot word recognition method based on large model technology.

[0008] This method integrates hot word recognition technology with semantic understanding and sequence information, and combines it with a multi-task learning summary generation model to achieve in-depth analysis and intelligent summarization of customer conversations, significantly improving analysis efficiency and insight accuracy.

[0009] The specific technical solution of the present invention is as follows:

[0010] A method for identifying hot words in customer conversations based on large model technology includes the following steps:

[0011] S100: Model training and construction; specifically including the following processes:

[0012] S110: Construct a multi-dimensional conversation analysis tag system; Based on historical conversation data and business knowledge, the system sorts out multi-dimensional analysis dimensions, including core customer needs, root causes of problems, products / services involved, customer emotional polarity, and conversation resolution status, and defines fine-grained tags for each dimension;

[0013] S120: Conversation data preprocessing and augmentation; cleaning, sentence segmentation, and customer and customer service role identification of the original conversation text; data augmentation using methods such as back translation and synonym replacement to improve the model's generalization ability;

[0014] S130: Multi-task joint training and fine-tuning based on a large model; using a pre-trained model based on Qwen2.5-32B as the foundation, a multi-task learning framework is constructed, and the following tasks are trained simultaneously:

[0015] S131: Hot word recognition and classification task; The model identifies key entities and phrases in the conversation, classifies them according to the dimensions defined in S110, and calculates their importance weight in the conversation.

[0016] S132: Summary generation task; train the model in a sequence-to-sequence manner to generate structured summaries containing core issues, causes of problems, processing procedures and results;

[0017] S200: Conversation data input and preprocessing; users upload batch historical conversation texts or access conversation streams in real time through the system interface; the system automatically segments and roles the input text and standardizes the encoding format;

[0018] S300: Multi-level hot word recognition and weight calculation; using the model trained by S100 to process the preprocessed conversation text: specifically including the following processes:

[0019] S310: Basic hot word extraction; preliminary screening based on word frequency, inverse text frequency index, and part of speech;

[0020] S320: Semantic-enhanced hot word recognition; using a finely tuned large model to encode the conversation context, it identifies compound words, phrases and implicit expressions that are important in specific contexts, such as "abnormal electricity bill" which may be associated with "excessive power of electrical appliances" or "electricity bill arrears", and assigns them dimension labels.

[0021] S330: Dynamic weight calculation; comprehensively considers the frequency and position of hot words, as well as the semantic association strength with the customer's core needs and emotional polarity, to calculate the final weight value of each hot word; the position is such as the beginning and end of the conversation, or where customer service repeatedly asks questions;

[0022] S400: Conversation summary generation based on hot word graph; constructs a hot word semantic graph of the conversation by taking the hot words with weights and dimension labels identified in S300; uses this graph as key information to drive the fine-tuned large model to generate structured summaries, ensuring that the summaries cover the key information points represented by the core hot words and logically reproduce the main body of the conversation.

[0023] S500: Results visualization and output.

[0024] Preferably, in step S100, LoRA (Local Efficiency Adjustment) is used to efficiently fine-tune the large model, with the rank lora_rank set to 32, the scaling factor lora_alpha set to 16, and the dropout rate lora_dropout set to 0.05.

[0025] Preferably, in step S132, the training uses a weighted loss function to balance the learning objectives of different tasks; the optimizer is AdamW, and the learning rate uses a warm-up cosine annealing strategy.

[0026] Preferably, in step S300, the dynamic calculation formula for hot word weight is:

[0027] Weight(t)=λ1*TF-IDF(t)+λ2*PositionBias(t)+λ3*SemanticRelevance(t,KeyIntent)

[0028] Where t represents a hot word, λ1, λ2, λ3 are adjustable weight coefficients, TF-IDF(t) is the term frequency-inverse document frequency, PositionBias(t) is the importance bonus based on the position of the hot word in the conversation, and SemanticRelevance(t,KeyIntent) is the semantic relevance score between the hot word and the core conversation intent identified by the model.

[0029] Preferably, in step S500, the system front-end interface provides one or more of the following result display methods:

[0030] The identified hot words and their weights and dimensions are displayed in the form of word clouds or sorted lists;

[0031] Displays automatically generated structured text summaries;

[0032] It provides an interactive link between hot words and corresponding content in the summary. Users can click on hot words to locate the relevant parts in the original conversation and the summary.

[0033] Preferred options also include:

[0034] S600: Abstract quality assessment and model iteration; The system provides an interface for business experts to score or modify the quality of the generated abstracts. This feedback data will be automatically collected and used to continuously and incrementally train the model in S100 to continuously optimize the abstract generation effect.

[0035] Compared with existing technologies, the beneficial effects of this invention are as follows: through a multi-dimensional tagging system and semantic understanding, it achieves accurate identification and classification of the deep meaning of hot words, surpassing simple word frequency statistics; by coupling hot word identification and summary generation through multi-task learning and semantic graphs, the generated summaries are both focused and coherent; and it supports automated and intelligent analysis of massive conversation data, greatly improving the efficiency and depth of customer sentiment analysis, and providing direct data support for product optimization, service improvement, and risk warning. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the process of a customer conversation hot word recognition method based on large model technology according to the present invention. Detailed Implementation

[0037] In this embodiment, after preprocessing by S200, a customer conversation text contains the following fragment:

[0038] Customer: My electricity bill suddenly went up a lot this month, even though I didn't do anything! Customer Service: Have you recently installed any new electrical appliances, such as electric blankets or electric heaters? Customer: I did use an electric heater, but the electricity bill shouldn't be this outrageous! This is definitely overcharging!

[0039] After processing by S300, the model identified high-weight hot words including: abnormal electricity bills (weight 0.95, dimension: core customer demands); high-power appliances (weight 0.88, dimension: root cause of the problem); and complaints (weight 0.90, dimension: customer emotional polarity).

[0040] In S400, the model uses the aforementioned hot words as a guide to generate a structured summary: "[Core Request] Customer complains of an abnormally high electricity bill this month. [Cause of the Problem] Upon investigation, the customer has recently been using an electric heater, which is suspected to be the result of using a high-power appliance, leading to the abnormal electricity bill. [Handling Status] The customer is agitated and requests further investigation. [Resolution Status] Pending."

[0041] Subsequently, the system highlights the aforementioned hot keywords on the S500's front-end interface and associates them with the summary content, while also providing a complete summary report download function. Customer service personnel can quickly grasp the main issues in batch conversations by observing the distribution of hot keywords.

[0042] A method for identifying hot words in customer conversations based on large model technology includes the following steps:

[0043] S100: Model Training and Construction; This step is fundamental to building core analytical capabilities, and its implementation includes the following sub-steps:

[0044] S110: Construct a multi-dimensional conversation analysis tag system;

[0045] Based on the analysis of massive amounts of historical customer service conversation data and combined with the experience of business experts, a complete tagging system has been systematically constructed. This system includes five main dimensions and corresponding subcategories:

[0046] Core customer needs include subcategories such as "fee inquiries", "service processing", "complaints and suggestions", "information inquiries", and "fault reporting".

[0047] Root cause dimensions of the problem include: "use of high-power electrical appliances", "abnormal electricity bills", "customer misunderstandings", "system billing errors", and "service process issues".

[0048] Customer emotional polarity dimension: using a three-level classification standard: negative, neutral, and positive;

[0049] Session resolution status dimensions include "Resolved", "Pending Follow-up", "Appeal Required", and "Unresolved".

[0050] S120: Session Data Preprocessing and Enhancement. This step standardizes the raw session data, specifically including:

[0051] Data cleaning; removing special characters, garbled text, and irrelevant symbols from the conversation, and standardizing numerical expressions such as dates and amounts;

[0052] Sentence segmentation and role recognition; employing a combination of rule-based and model-based methods to accurately segment dialogue turns and identify the speaker roles for customers / customer service representatives;

[0053] Data augmentation: Training samples with the same semantics but different expressions are generated by using "Chinese→English→Chinese" back-translation technology, while the scale of training data is expanded by using techniques such as synonym replacement and random insertion.

[0054] S130: Multi-task joint training and fine-tuning based on a large model. This step employs advanced parameter-efficient fine-tuning techniques, specifically implemented as follows:

[0055] a. Base model selection: Qwen2.5-32B was selected as the base pre-trained model, which performs well in Chinese language understanding and generation tasks.

[0056] b. Low-rank adaptive LoRA configuration: To balance model performance and training efficiency, a LoRA adapter is introduced, with key parameters set as follows: rank lora_rank: 32; scaling factor lora_alpha: 16; dropout rate lora_dropout: 0.05.

[0057] c. Multi-task learning framework:

[0058] Hot word recognition and classification task: A sequence labeling method is used, employing the BIO labeling system. The model predicts the corresponding label and importance weight for each input token. The loss function is designed as follows:

[0059] L_ner = α1 * L_ce + α2* L_mse;

[0060] Where L_ce is the cross-entropy loss for classification, L_mse is the mean squared error loss for weight prediction, and α1 and α2 are the balancing hyperparameters.

[0061] Summary generation task: Using an encoder-decoder architecture, train a model to generate structured summary text. The loss function is the standard cross-entropy loss.

[0062] d. Training parameter settings:

[0063] Batch size: 16;

[0064] Training epochs: 3;

[0065] Optimizer: AdamW, with a learning rate using cosine annealing scheduling and a warm-up strategy;

[0066] Total loss: L_total = μ1 * L_ner + μ2 * L_summary, where μ1 and μ2 are task weight coefficients.

[0067] S200: Session data input and preprocessing;

[0068] Users submit session data to be analyzed via RESTful API or web interface. The system supports both batch text file upload and real-time data stream access. The input data is uniformly encoded into UTF-8, and then the preprocessing pipeline established in S120 is invoked to output a structured session sequence.

[0069] S300: Multi-level hot word identification and weight calculation. This step is the core analysis process and employs a three-level processing flow:

[0070] S310: Basic hot word extraction: Initial keyword extraction is performed based on TF-IDF and TextRank algorithms, and a custom business dictionary is used to ensure accurate identification of professional terms;

[0071] S320: Semantically enhanced hot word recognition: Input the preprocessed conversation text into the model trained by S130, and use its powerful contextual understanding ability to identify concepts and phrases that are important in specific dialogue contexts and assign corresponding dimension labels.

[0072] S330: Dynamic weight calculation: using the comprehensive weight calculation formula:

[0073] Weight(t)=λ1*TF-IDF(t)+λ2*PositionBias(t)+λ3*SemanticRelevance(t,KeyIntent)

[0074] PositionBias(t) assigns different weight coefficients based on the position of the term in the session. For example, important statements at the beginning and end of the session have higher weights. SemanticRelevance(t,KeyIntent) calculates the semantic relevance of the term to the core intent of the entire session.

[0075] S400: Conversation summarization based on hot word graph; it constructs a hot word semantic graph from the weighted and labeled hot word set output by S300, where nodes represent hot words and edges represent co-occurrence relationships or semantic association strength. This graph is used as an attention guidance mechanism to drive the summarization model to focus on key information points and generate a structured summary containing four elements: "core demand - problem cause - solution - resolution status".

[0076] S500: Result visualization and output; the system provides a web-based interactive visualization interface, with main functions including:

[0077] Hot word cloud: Dynamically renders a keyword cloud based on weights, and supports filtering by dimension tags;

[0078] Structured summary display: Clearly presents the four elements of the summary, with key information highlighted;

[0079] Interactive features: Supports linked location of hot words with the original text and abstract, making it convenient for users to trace the source of analysis results.

[0080] S600: Abstract quality assessment and model iteration.

[0081] A closed-loop mechanism for human feedback is established, allowing business experts to rate the system-generated summaries on a quality scale of 1-5 points or make direct modifications via the interface. This feedback data is automatically incorporated into the model iteration dataset, enabling periodic incremental training of the model and continuous optimization of its performance. Simultaneously, the system records various performance metrics, such as hot word recognition accuracy and summary ROUGE score, for monitoring model performance.

Claims

1. A method for identifying hot words in customer conversations based on large model technology, characterized in that, Includes the following steps: S100: Model training and construction; specifically including the following processes: S110: Construct a multi-dimensional conversation analysis tag system; Based on historical conversation data and business knowledge, the system sorts out multi-dimensional analysis dimensions, including core customer needs, root causes of problems, products / services involved, customer emotional polarity, and conversation resolution status, and defines fine-grained tags for each dimension; S120: Conversation data preprocessing and augmentation; cleaning, sentence segmentation, and speaker role identification of the original conversation text; data augmentation using back-translation and synonym replacement methods to improve the model's generalization ability; S130: Multi-task joint training and fine-tuning based on large models; A multi-task learning framework was constructed based on a pre-trained model of Qwen2.5-32B, and the following tasks were trained simultaneously: S131: Hot word recognition and classification task; The model identifies key entities and phrases in the conversation, classifies them according to the dimensions defined in S110, and calculates their importance weight in the conversation. S132: Summary generation task; train the model in a sequence-to-sequence manner to generate structured summaries containing core issues, causes of problems, processing procedures and results; S200: Conversation data input and preprocessing; users upload batch historical conversation texts or access conversation streams in real time through the system interface; the system automatically segments and roles the input text and standardizes the encoding format; S300: Multi-level hot word recognition and weight calculation; using the model trained by S100 to process the preprocessed conversation text: specifically including the following processes: S310: Basic hot word extraction; preliminary screening based on word frequency, inverse text frequency index, and part of speech; S320: Semantically enhanced hot word recognition; utilizes a finely tuned large model to encode the conversation context, identify compound words, phrases and implicit expressions that are important in specific contexts, and assign dimension labels to them; S330: Dynamic weight calculation; comprehensively considers the frequency and position of hot words, as well as the semantic association strength with the core needs and emotional polarity of customers, to calculate the final weight value of each hot word; S400: Conversation summary generation based on hot word graph; construct a hot word semantic graph of the conversation by taking the hot words with weights and dimension labels identified in S300; use this graph as key information to drive the fine-tuned large model to generate structured summaries, ensuring that the summaries cover the key information points represented by the core hot words and logically reproduce the main body of the conversation. S500: Results visualization and output.

2. The customer conversation hot word recognition method based on large model technology according to claim 1, characterized in that, In S100, the parameter efficiency fine-tuning technology LoRA is adopted, with the rank lora_rank set to 32, the scaling factor lora_alpha set to 16, and the dropout rate lora_dropout set to 0.

05.

3. The customer conversation hot word recognition method based on large model technology according to claim 1, characterized in that, In step S132, a weighted loss function is used for training to balance the learning objectives of different tasks; AdamW is selected as the optimizer, and a cosine annealing strategy with warm-up is used for the learning rate.

4. The customer conversation hot word recognition method based on large model technology according to claim 1, characterized in that, In S300, the dynamic calculation formula for hot word weight is as follows: Weight(t)=λ1*TF-IDF(t)+λ2*PositionBias(t)+λ3*SemanticRelevance(t,KeyIntent) Where t represents a hot word, λ1, λ2, λ3 are adjustable weight coefficients, TF-IDF(t) is the term frequency-inverse document frequency, PositionBias(t) is the importance bonus based on the position of the hot word in the conversation, and SemanticRelevance(t,KeyIntent) is the semantic relevance score between the hot word and the core intent of the conversation identified by the model.

5. The customer conversation hot word recognition method based on large model technology according to claim 1, characterized in that, In S500, the system front-end interface provides one or more of the following result display methods: The identified hot words and their weights and dimensions are displayed in the form of word clouds or sorted lists; Displays automatically generated structured text summaries; It provides an interactive link between hot words and corresponding content in the summary. Users can click on hot words to locate the relevant parts in the original conversation and the summary.

6. The customer conversation hot word recognition method based on large model technology according to any one of claims 1-5, characterized in that, Also includes: S600: Abstract quality assessment and model iteration; The system provides an interface for business experts to score or modify the generated summaries. This feedback data will be automatically collected and used to continuously and incrementally train the model in S100, thereby constantly optimizing the summary generation effect.

Images