Intelligent mobile phone model recommendation system for coprocessing large and small models

Abstract

The invention, which relates to the technical field of artificial intelligence, discloses a mobile phone model intelligent recommendation system based on coprocessing of large and small models, comprising an intention classification module for performing intention scene distinguishing and marking intention labels; the entity identification module is used for extracting product brand and model entities from the text content input by the user; the model reasoning module is used for matching standard names in the dictionary and outputting candidate sets of models and brands; according to the model recommendation system, a large and small model cooperation mechanism is introduced, the small model is used for rapid filtering and matching, possible candidate models are rapidly screened out, the large model is responsible for deep semantic understanding of a complex scene, and a complex recommendation scene is supplemented and corrected; therefore, standardized mobile phone models can be more accurately mapped, diversified input expressions are accurately analyzed, the flexibility of the system in coping with different input forms is remarkably improved, and efficient and accurate mobile phone model identification recommendation is realized.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent recommendation system for mobile phone models that uses collaborative processing of large and small models. Background Art

[0002] In the field of intelligent customer service and mobile phone sales, it is crucial to quickly and accurately identify and standardize the mobile phone models mentioned by users. Especially in customer service systems, when users query information about specific mobile phone models, the system needs to be able to accurately understand user needs. However, due to the complexity and update frequency of mobile phone models, relying solely on traditional keyword matching or a single large model for entity extraction often fails to meet actual application needs. As mobile phone brands and models continue to innovate, mobile phone model naming is no longer limited to simple combinations of numbers and letters, but also includes diversified expressions such as brand abbreviations, series abbreviations, special suffixes, etc. At the same time, in order to cope with users' non-standard input, intelligent customer service systems face huge challenges in processing various descriptions, abbreviations and spoken expressions. For example, users may mention mobile phone models with abbreviations, aliases or even vague descriptions, making it difficult for traditional extraction and mapping methods to cover all expressions. In addition, different mobile phone brands have their own naming rules, which further increases the complexity of accurate identification and matching of the system; We found that the prior art has the following defects: 1. Insufficient coverage of keyword matching: Most existing mobile phone model identification methods rely on keyword matching technology, which matches by pre-building a dictionary containing brands and models. However, due to the endless emergence of new products in the mobile phone market, such dictionaries need to be updated frequently, with high maintenance costs, and it is difficult to cover all newly launched models in real time. In addition, keyword matching methods often have problems with recognition errors or omissions when encountering various nicknames and abbreviations of models; 2. The amount of data based on entity extraction is limited: The current entity extraction method shows certain potential in the field of mobile phone model recognition, especially when trained with a corpus containing a certain amount of data. However, due to the wide variety of mobile phone models, a single dataset is difficult to cover all possible expressions, resulting in insufficient model training effect. At the same time, different users may describe the same model in different forms (such as using abbreviations or nicknames), making it difficult for existing methods to cope with various non-standardized input expressions in practical applications; 3. Efficiency issues of pure large models: In the specific task of mobile phone model recognition, directly relying on large language models for entity extraction can improve coverage, but the extraction of candidate models one by one for each user input brings huge computational overhead. Large models consume a lot of computing resources and time when processing long texts or complex sentences, especially in high-concurrency environments, which affects the system's response speed and user experience. This method has a high processing cost and is difficult to meet the needs of real-time response; The above defects indicate the inadequacies of current mobile phone model recognition technology in keyword matching, entity extraction and large-scale model applications. Summary of the invention

[0003] The purpose of the present invention is to provide an intelligent recommendation system for mobile phone models with collaborative processing of large and small models. By introducing a collaborative mechanism of large and small models, efficient and accurate mobile phone model identification and standardized mapping and recommendation can be achieved to solve the problems raised in the above background technology.

[0004] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a mobile phone model intelligent recommendation system for collaborative processing of large and small models, comprising: The intent classification module recognizes the text content input by the user, distinguishes the intent scenarios, and marks the corresponding intent category labels; An entity recognition module extracts product brand and model entities from the text content input by the user, defines labels for the entity content and non-entity content of the brand and model, and reversely optimizes the entity recognition module; The model inference module maintains a dictionary containing brand standard names and alias mappings, and matches the brand and model entities extracted by the above entity recognition module with the standard names in the dictionary according to the intent category labels, and performs vector similarity matching at the same time, and outputs a candidate set of models and brands; The model recommendation module determines the input context based on the text content input by the user, and uses the thinking chain prompt words built according to the input context to dynamically maintain and filter the candidate set generated in the above model reasoning module, and recommend the brand model that best meets the user's needs.

[0005] Preferably, the intent classification module and the entity recognition module both include a data collection module for data augmentation processing, and the specific method is as follows: First, collect text content containing various product brands and models from various platforms to build relevant data sets. The collected samples should cover the real product names of multiple brands and models. Secondly, the collected data is annotated and stored as a structured JSON format file, with each piece of data containing several fields; Finally, the dataset is augmented to enhance the coverage of expressions, including synonym replacement, abbreviation expansion, and brand generalization by replacing brands or models to expand data samples.

[0006] Preferably, each piece of data in the intention classification module includes an intention tag field whose value is "recommended product model" or "non-recommended product model"; Each piece of data in the entity recognition module includes an entity list tag field containing an entity category, an entity value, and a start and end position in the text content.

[0007] Preferably, the intent classification module further includes a classification model training module for performing intent classification training using the collected data, and the specific method is as follows: S101: Feature vector extraction: Use the semantic vector extraction model Bert to vectorize the collected data to ensure the semantic and contextual information of the collected data. After processing, each piece of data will be converted into a vector of fixed length; S102: Input the feature vector to the classification model, and input the extracted feature vector to the DistilBERT model. The DistilBERT model hierarchically processes the input feature vector, gradually extracts more advanced semantic and contextual information, and outputs a fixed-length feature representation that retains important semantic information in the sentence. S103: The feature vector is passed to a fully connected layer. The feature representation is passed to a fully connected layer, which performs a linear transformation on the feature, maps the high-dimensional feature vector to a low-dimensional feature vector for classification, obtains the logits value of each category, and finally outputs the logits value of the category; S104: The logits value is passed to the Softmax layer, and the logits value is converted into probability values ​​of two categories to determine whether the input text belongs to the "recommended product model" or "non-recommended product model" category; S105: Calculate the cross entropy loss, compare the output probability of the softmax layer with the actual label, calculate the cross entropy loss, measure the difference between the model prediction and the true label, and output the loss value of a single sample; S106: Back propagation optimization model, using the calculated loss value, update the model parameters through the back propagation algorithm (such as Adam optimizer), so that the model gradually learns the characteristics of each category, makes the loss function converge, and thus optimizes the model performance. After the training is completed, save the final model parameters, deploy and use, and record them as M 推荐分类 .

[0008] Preferably, the entity recognition module further includes an entity recognition training module for performing entity recognition training using the collected data. The specific steps of extracting brand and model entities from the text content are as follows: S201: Use the Bert model to vectorize the collected data at the word level, convert each sentence into a sequence, and each word corresponds to a feature vector; S202: Input the word vector sequence output by BERT into the Bidirectional Long Short-Term Memory (BiLSTM) layer to capture the context sequence information of the sentence and output the word-level feature representation containing the context information; S203: The feature representation of the above output is passed to the conditional random field (CRF) layer to model the dependency between labels. The CRF layer calculates the scores of all possible label sequences, selects the label sequence with the highest score as the prediction result, and outputs the predicted label of each word. S204: Calculate the loss function, input the predicted label sequence and the real label sequence, use the log-likelihood loss function of the CRF layer, calculate the negative log-likelihood loss, measure the degree of fit of the model to the training data, calculate the difference between the label sequence predicted by the model and the real label sequence, and output the loss value; S205: Use the calculated loss value to update the model parameters through the back propagation algorithm (such as Adam optimizer), adjust the model parameters according to the loss gradient, so that the model gradually learns the characteristics of the entity. After the training is completed, save the final model parameters and deploy them for use, denoted as M 品牌类型实体抽取 .

[0009] Preferably, the entity recognition module defines labels in a BIO annotation manner to annotate each word, specifically: - B-Brand: the first word of the brand entity; - I-Brand: internal word of the brand entity; - B-Model: the beginning word of the model entity; - I-model: internal word of model entity; - O: non-entity word.

[0010] Preferably, the model inference module includes dynamic maintenance and update of brand standard name and alias mapping dictionary, and the specific method includes the following steps: S301: Data source access and regular update, access to data from various platforms, real-time acquisition of corresponding product brands, series and models, including standard names and alias data to enrich the original dictionary, data cleaning and format unification of the acquired data, the cleaned and unified data enters the standard dictionary candidate entry library D 候选 , for further alias extraction and frequency analysis; S302: Frequency-based alias extraction and mapping. After data cleaning, the system performs a 候选 The models in the dictionary are counted for frequency, high-frequency aliases are identified, and the user's non-standardized description is associated with the standard models in the dictionary through a mapping mechanism. This step includes frequency analysis, high-frequency alias mapping, and clustering of unmatched aliases: High-frequency alias frequency analysis: D collected in each update 候选 The model performs frequency statistics and records the number of occurrences of each description. If the number of occurrences of a description in a day exceeds the set frequency threshold, it is a high-frequency alias, recorded as Alia high-freq , take the alias as a candidate entry and enter the next processing flow; High-frequency alias mapping: The system uses a pre-trained language model (such as BERT or Word2Vec) to generate a vector representation VEC for the standard name and alias of each description in the dictionary H , and stored in the vector database for subsequent matching; Alia high-freq Generate a vector, denoted as Vec Alia_h , the high frequency alias vector Vec Alia_h With the dictionary vector Vec H Perform cosine similarity calculation. If the similarity between the high-frequency alias and the existing model in the dictionary is higher than the threshold, it is marked as "matched high-frequency alias"; otherwise, it is marked as "unmatched high-frequency alias"; Clustering of unmatched aliases: For unmatched high-frequency aliases, the system will automatically group them using a clustering algorithm, classify similar descriptions into the same category, and select the description with the highest frequency as the standard description, specifically: The unmatched high-frequency aliases are processed using clustering algorithms such as DBScan, and the descriptions with high similarity are grouped into the same cluster; In each cluster, the description with the highest frequency is selected as the standard description based on the frequency information, and the remaining descriptions are used as aliases of the standard description; Combined with the brand information in the dictionary, brand information is added to the standard description generated by clustering, and the clustering results will be used as new candidate entries; S303: Dictionary update and optimization: the matched high-frequency aliases are directly added to the alias list of the dictionary to expand the coverage of the dictionary. For the candidate entries generated by the unmatched high-frequency aliases, they are added to the candidate queue Q and officially added to the dictionary after manual review and confirmation. At the same time, the system will regularly clean up the descriptions with low frequency of use in the dictionary and move them to the backup word library. Cleaning up low-frequency descriptions: Setting a threshold based on the frequency of use (e.g., fewer than 10 occurrences in a year), and moving low-frequency descriptions from the dictionary to the backup word library; Backup word library call: The cleaned descriptions are still stored in the backup word library for use in special queries to avoid accidental deletion of users' personalized expressions.

[0011] Preferably, the model inference module further includes matching the brand model based on keywords, and the specific matching method steps are: First, keyword matching is performed, and multi-layer regular expression matching is used to quickly screen the brand or model words that appear in the sentence to form a list of candidate brands and models. Regular expressions are constructed based on the brands and models that already exist in the dictionary: Brand expression: construct a brand matching regular expression for the brand names and their aliases in the system; Model expression: construct a model matching regular expression for the standard models and aliases in the system; filter the user input statements using the regular expressions to extract candidate brands and models that meet the conditions; Secondly, vector similarity matching is performed to further expand the candidate set, and the parsing effect is improved through semantic matching. The entity recognition module is used to extract brand and model entities from the user input text content, and the vector V is extracted. input , and then use the vector Vec generated by the standard name-alias dictionary H , for V input With Vec H Calculate the cosine similarity, extract similar models, and finally form a candidate set of product models and brands.

[0012] Preferably, the model recommendation module includes a reasoning analysis module for user needs, and the specific reasoning analysis method steps are: S401: Input data preparation, collecting all historical conversation content of the user in the current session, including previous inquiries, preferences, negative conditions and other information, and obtaining the user's latest input and candidate set; S402: Context judgment: judging the user's input context based on the user's historical conversations and current questions, specifically: initial demand expression, progressive demand expression, and negative condition expression; S403: Select appropriate reasoning process and thinking chain prompt words, select the corresponding reasoning process according to the situation judgment results, and design thinking chain prompt words to guide the model to reason according to the expected logic; S404: Execute large model reasoning, pass the constructed thinking chain prompt words and input data to the model, execute the reasoning process, and obtain the output result of the model; S405: Result analysis and output: Analyze the output of the model, obtain the final recommended model list and recommendation reasons and feed it back to the user.

[0013] Preferably, the construction of the thought chain prompt words includes: Role setting: clarify the role and tasks of the model; Task description: Describe in detail the task objectives that the model needs to complete; Input instructions: clarify the structure and content of input data; Reasoning steps of thought chain: describe step by step the reasoning steps that the model needs to carry out; Output format: Specify the format of the output results to ensure consistency and ease of parsing.

[0014] Preferably, the model recommendation module also includes adjusting and analyzing the candidate set that ultimately forms the product model and brand. When there are too many models in the candidate list, the module will further screen and reduce redundancy, retaining only the core models that best match the user's description; when all relevant models of the user's needs cannot be identified, the module will parse the details of the user's input through the model, supplement the missing candidate models, and ensure that the user's needs are not ignored.

[0015] In summary, the beneficial effects of the present invention are: 1. Collaborative optimization recognition mechanism: A collaborative mechanism of large and small models is introduced. First, the small model is used for rapid filtering and matching to quickly screen out possible candidate models to improve subsequent processing efficiency and reduce computing resource consumption. Subsequently, the large model is responsible for deep semantic understanding of complex scenarios, thereby accurately parsing diverse input expressions, supplementing and correcting complex recommendation scenarios, and more accurately mapping them to standardized mobile phone models. This collaborative optimization mechanism significantly improves the system's flexibility in dealing with different input forms, and achieves efficient and accurate mobile phone model recognition.

[0016] 2. Enhanced contextual understanding: Based on the collaborative framework of large and small models, the present invention can better capture and understand users' various descriptions and contextual information of mobile phone models. This method is not only based on a single vocabulary match, but can also combine the context and context to supplement and correct complex recommendation scenarios through large models, so as to more accurately map to standardized mobile phone models.

[0017] 3. Dynamic update and industry adaptation: Taking into account the update frequency of mobile phone models and the diversity of naming rules of different brands, the present invention introduces a dynamic update mechanism and industry adaptation strategy. This method allows the model to adapt to the launch of new models and update the dictionary in a timely manner to maintain efficient and accurate recognition and matching capabilities, thereby ensuring that the accuracy and practicality of recognition are always maintained in the rapidly changing mobile phone industry environment. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0019] Figure 1 A schematic diagram of the process framework structure of a mobile phone model intelligent recommendation system for collaborative processing of large and small models according to the present invention; Figure 2 A schematic diagram of a classification model training process framework in a mobile phone model intelligent recommendation system for collaborative processing of large and small models according to the present invention; Figure 3 A schematic diagram of a process framework of dictionary maintenance in a mobile phone model intelligent recommendation system for collaborative processing of large and small models according to the present invention; Figure 4 A schematic diagram of a training process framework for word entity recognition in a mobile phone model intelligent recommendation system for collaborative processing of large and small models according to the present invention; Figure 5 A schematic diagram of an example of mobile phone model inquiry in a mobile phone model intelligent recommendation system for collaborative processing of large and small models according to the present invention; Figure 6 This is a service interface display diagram of a mobile phone model intelligent recommendation system for collaborative processing of large and small models according to the present invention. DETAILED DESCRIPTION

[0020] The present invention will now be further described in detail in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. These drawings are all simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic way, and therefore they only show the structures related to the present invention.

[0021] To facilitate understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0022] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0023] Any feature disclosed in this specification (including any additional claims, abstract and drawings), unless otherwise stated, may be replaced by other equivalent or alternative features with similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0024] Combine the following Figure 1-6 The present invention is described in detail. An embodiment of the present invention is: an intelligent recommendation system for mobile phone models with collaborative processing of large and small models, including an intent classification module, an entity recognition module and a model reasoning module as small models of the system, and a model recommendation module as a large model of the system. A collaborative mechanism of large and small models is introduced to quickly filter and match the small model to quickly screen out possible candidate models, so as to improve subsequent processing efficiency and reduce computing resource consumption. Subsequently, the large model is responsible for deep semantic understanding of complex scenes, supplements and corrects complex recommendation scenes, so as to more accurately map them to standardized mobile phone models, thereby accurately parsing diversified input expressions.

[0025] Step 1: Build a user recommendation intent classification module The task of the intent classification module is to accurately identify whether the user input belongs to the "recommended mobile phone model" scenario, thereby filtering out irrelevant sentences and optimizing system resources. The input is the user's text content, and the output is the intent category label of the sentence. This module consists of two key parts: the data collection module and the classification model training module; 1. Data collection module: This module is mainly used to generate and process high-quality training data to ensure the accuracy of the intent classification model. This module specifically includes the following steps: 1.1 Data Collection From the product recommendation data of e-commerce platforms, the real conversation records of intelligent customer service, and the common mobile phone model consultation texts of users, sentences containing recommended mobile phone models and sentences not containing recommended mobile phone models are collected to build a data set. The collected samples should cover real mobile phone names of various brands and models as much as possible, such as "iPhone 14 Pro Max", "Samsung Galaxy S23 Ultra", etc.

[0026] An example of a sentence containing a mobile phone model is as follows: “Please recommend an iPhone with good value for money.” "I want to buy a phone with a good camera. Is there any suitable one? For example, Huawei P60 Pro." After the collected data is manually annotated, it is stored as a structured JSON file. Each piece of data contains two fields: text: The user's input sentence.

[0027] label: intent label, the value is "recommended mobile phone model" or "non-recommended mobile phone model".

[0028] An example of a possible demonstration is as follows: [ { "text": "Please recommend an iPhone 14 Pro Max with good photography effects.", "label": "Recommended mobile phone models" }, { "text": "What will the weather be like tomorrow?", "label": "Not recommended phone model" } ] 1.2 Data Augmentation In order to improve the generalization ability of the model, the data set is augmented to enhance the coverage of multiple expressions. For example: 1. Synonym replacement: Replace "recommendation" with "introduction", "suggestion", etc. to generate multiple expressions.

[0029] 2. Abbreviation expansion: Expand “cost-effectiveness” to “performance-price ratio”, etc., so that the model can better understand sentence changes.

[0030] 3. Brand generalization: Expand the data sample by replacing the brand or model, such as "recommend a replacement for Samsung GalaxyS23 Ultra".

[0031] 2. Classification model training module refer to Figure 2 This module is mainly used to use the training data collected in the previous step to train the relevant intent classification model to identify whether the current buyer's question belongs to the category of "recommended mobile phone models". A specific feasible training step is as follows: 2.1 Feature vector extraction Word vector initialization: Use the semantic vector extraction model Bert to vectorize each sentence to ensure that the semantic and contextual information in the sentence is captured. After processing, each sentence will be converted into a fixed-length vector of 768 dimensions. Bert is a classic pre-trained language representation model in the field of natural language recognition.

[0032] 2.2 Classification Model Training 2.2.1 Input feature vector to classification model Input the feature vector extracted in the previous step into the DistilBERT model. DistilBERT is a lightweight BERT variant suitable for efficient text classification tasks.

[0033] Input: The feature vector (the 768-dimensional vector generated by BERT in step 1) represents the contextual semantic information of each sentence.

[0034] Process: The DistilBERT model processes the input feature vector hierarchically; DistilBERT gradually extracts and processes input information through multiple layers of Transformer encoders. Each layer of Transformer encoder performs complex transformations on the input feature vector, gradually extracting more advanced semantic and contextual information.

[0035] Output: A fixed-length feature representation that retains important semantic information in the sentence.

[0036] 2.2.2 Feature vector is passed to the fully connected layer The feature representation output by DistilBERT is passed to a fully connected layer. This layer performs a linear transformation on the features, mapping the high-dimensional feature vector to a low-dimensional feature vector for classification. This simplifies the feature representation and makes it more suitable for subsequent classification tasks.

[0037] Input: Feature representation output by the DistilBERT model (contextual semantic features of the sentence).

[0038] Process: The fully connected layer performs a linear transformation on the input features to obtain the logits value for each category.

[0039] (The logits value is the raw output that has not been processed by an activation function (such as softmax), and is usually used to calculate loss functions and make probability predictions) Output: The logits value of the category, which is used for subsequent probability calculation.

[0040] 2.2.3 Logits value is passed to the Softmax layer The output logits value of the fully connected layer is passed to the softmax layer, and the logits value is converted into probability values ​​of two categories to determine whether the input text belongs to the "recommended mobile phone model" or "non-recommended mobile phone model" category.

[0041] Input: logits values ​​(raw scores for class predictions) from the fully connected layer.

[0042] Process: The softmax layer converts logits values ​​into category probabilities, calculated as follows: Output: Probability distribution of two categories, used for loss calculation

[0043] Among them, P(y i |x) is the input x belonging to category y i The probability is the logits value of the category, and N is the total number of categories (here 2).

[0044] 2.2.4 Calculating Cross Entropy Loss The output probability of the softmax layer is compared with the actual label, and the cross entropy loss is calculated to measure the difference between the model prediction and the true label. The formula of the cross entropy loss function is as follows:

[0045] Input: Class probability distribution and actual label output by softmax.

[0046] Process: Calculate the cross entropy loss based on the predicted probability and the true label to generate the loss value.

[0047] Output: loss value of a single sample, used for backpropagation.

[0048] 2.2.5 Back-propagation optimization model Using the calculated loss value, the model parameters are updated through a back-propagation algorithm such as the Adam optimizer. The Adam optimizer adjusts the parameters based on the gradient information in each iteration to make the loss function converge, thereby optimizing the model performance.

[0049] Input: loss value for a single example.

[0050] Process: The Adam optimizer adjusts the model parameters according to the loss gradient, so that the model gradually learns the characteristics of each category.

[0051] Output: optimized model parameters.

[0052] After training is completed, save the final model parameters and deploy them for use, recording them as M 推荐分类 .

[0053] It should be noted that Adam (Adaptive Moment Estimation) is a widely used gradient descent optimization algorithm, especially suitable for training deep learning models. It combines the advantages of two gradient descent methods: AdaGrad (which can adapt the learning rate of each parameter and is suitable for handling sparse gradients, such as those common in natural language and image recognition tasks) and RMSProp (which uses the root mean square of the gradient to adjust the learning rate, which helps to handle non-steady-state objective functions).

[0054] Step 2: Build the mobile phone model entity recognition module The task of the mobile phone model entity recognition module is to accurately extract the mobile phone brand and model entities from the text input by the user, so as to facilitate subsequent recommendation and analysis. The input is the user's text content, and the output is the entity information containing the brand and model. This module consists of two key parts: the data collection module and the entity recognition model training module.

[0055] 1. Data Collection Module With similar data collection as in the first step, this module is mainly used to generate and process high-quality training data to ensure the accuracy of the entity recognition model. It includes the following steps: 1.1 Data Collection We collect sentences containing various mobile phone brands and models from product descriptions, user reviews, intelligent customer service conversation records, and mobile phone review websites on e-commerce platforms to build a data set.

[0056] The collected data is manually annotated to mark the brand and model entities in the sentences and stored as a structured JSON format file. Each piece of data contains two fields: text: The user's input sentence.

[0057] entities: A list containing entities, each entity contains the entity category (brand or model), entity value, and start and end position in the sentence.

[0058] One possible example data is as follows: [ { "text": "I want to buy a Huawei P50 Pro with good photography effect.", "entities": [ {"entity": "Brand", "value": "Huawei", "start": 11, "end":13}, {"entity": "Model", "value": "P50 Pro", "start": 13, "end": 20} ] }, ] 1.2 Data Augmentation In order to improve the generalization ability of the model, the data set is augmented to enhance the coverage of multiple expressions. For example: 1.2.1. Synonymous sentence generation: Re-state the sentence using different expressions, such as replacing "I want to buy" with "I want to purchase" or "I need".

[0059] 1.2.2. Synonym replacement: Replace "recommend" with "introduce", "suggest", etc. to generate multiple expressions.

[0060] 1.2.3. Brand and model replacement: Expand the data samples by replacing the brand or model in the sentence. For example, replace "Huawei P50 Pro" with "Xiaomi 11 Ultra".

[0061] 1.2.4. Misspelling and colloquialism processing: Add common misspellings or colloquial expressions, such as "thirteen promax", "Samsung Galaxy".

[0062] 2. Entity Recognition Model Training Module Reference Figure 4 , this module mainly uses the training data collected in the previous step to train an entity recognition model for extracting brand and model entities from sentences. The specific training steps are as follows: 2.1 Feature Vector Extraction Use the pre-trained BERT model to perform word-level vectorization on each sentence to obtain the context semantic representation of each word. After processing, each sentence will be converted into a sequence, and each word corresponds to a 768-dimensional vector.

[0063] 2.2 Entity Recognition Model Training 2.2.1: Input the feature vector into the BiLSTM layer Input the word vector sequence output by BERT into the bidirectional long short-term memory network (BiLSTM) layer to capture the context sequence information of the sentence.

[0064] Input: The word vector sequence generated by BERT.

[0065] Process: BiLSTM processes the word vector sequence through two LSTM layers in the forward and backward directions to obtain the bidirectional context information of each word.

[0066] Output: Word-level feature representation containing context information.

[0067] 2.2.2: Transfer the feature vector to the CRF layer (sequence labeling task) Transfer the feature representation output by BiLSTM to the conditional random field (CRF) layer to model the dependencies between labels, ensuring that the output label sequence is legal and globally optimal. (The CRF layer can be combined with the bidirectional long short-term memory network (BiLSTM) to improve the performance of the model in sequence labeling tasks) Input: The word-level feature representation output by BiLSTM.

[0068] Process: The CRF layer calculates the scores of all possible label sequences and selects the label sequence with the highest score as the prediction result.

[0069] Output: The predicted label of each word, indicating its corresponding entity category.

[0070] 2.2.3: Define labeling scheme The BIO annotation method is used to annotate each word so that the model can learn to recognize the boundaries and categories of entities.

[0071] Tag definition: - B-Brand: The first word of the brand entity.

[0072] - I-Brand: Internal word of the brand entity.

[0073] - B-Model: The beginning word of the model entity.

[0074] - I-Model: Internal word of the model entity.

[0075] - O: non-entity word.

[0076] One possible example is as follows: Sentence: "I want to buy a Huawei P50 Pro that takes good photos." Marking: "OOOOOO B-Brand B-Model I-Model".

[0077] 2.2.4: Calculating the loss function The log-likelihood loss function of the CRF layer is used to calculate the difference between the label sequence predicted by the model and the actual label sequence.

[0078] Input: predicted label sequence and true label sequence.

[0079] Process: Calculate the negative log-likelihood loss to measure how well the model fits the training data.

[0080] Output: loss value, used for backpropagation.

[0081] 2.2.5: Back-propagation optimization model Using the calculated loss value, the model parameters are updated through a back-propagation algorithm (such as the Adam optimizer) so that the model can gradually learn the characteristics of the entity.

[0082] Input: loss value.

[0083] Process: The Adam optimizer adjusts the model parameters according to the loss gradient to optimize the model performance.

[0084] Output: optimized model parameters.

[0085] After training is completed, the final model parameters are saved and deployed for use, denoted as M 品牌类型实体抽取 .

[0086] Step 3: Build a mobile phone model inference module based on keyword matching refer to Figure 3 The goal of the mobile phone model reasoning module based on keyword matching is to map the brand and model entities extracted by the entity recognition module (step 2) to the standard mobile phone brand and model names in the market. By maintaining a dictionary containing standard name and alias mapping, it ensures that various non-standard expressions entered by users can be accurately matched to the correct mobile phone model. This module contains two key modules: the construction module of the standard mobile phone brand and model dictionary and the dictionary matching reasoning module.

[0087] 3.1 Building blocks of the standard mobile phone brand and model dictionary The main goal of this module is to build a mapping dictionary between mobile phone models and brands. Specifically, this dictionary is dedicated to maintaining the mapping relationship between the standard names of mobile phone models in the market and various aliases. Through a regular dynamic update mechanism, the latest model and description mapping is covered to improve the accuracy of the system. A hierarchical tree structure is used to achieve efficient query and ensure fast and accurate retrieval performance, denoted as H 标准名-别名映射 .

[0088] One possible dictionary storage result and demonstration example is as follows: The dictionary storage structure is a hierarchical structure, including: level by level division by brand, series, and model to form a tree structure.

[0089] Brand layer: Contains a list of standard brand names and their aliases.

[0090] Series layer: Under each brand, contains the series name and its aliases of the brand.

[0091] Model layer: Under each series, it contains the specific model name and its alias.

[0092] Hash index: At each level, a hash table (HashMap) is used to store standard name and alias mapping to improve search efficiency.

[0093] iPhone (Brand) ├── Aliases: ["Apple", "Apple", "iPhone"] ├── iPhone 15 (series) │ ├── Aliases: ["15th generation", "Fifteenth generation"] │ ├── iPhone15 (standard name) │ │ ├── Aliases: ["i15", "iPhone15", "15"] Generally speaking, each mobile phone brand will launch new products every year or even every quarter. Therefore, the establishment and maintenance of the dictionary is not a one-time project. It is necessary to establish a dynamic update mechanism to ensure that the system dictionary can reflect the latest mobile phone models and aliases in the market in real time to improve the accuracy and practicality of the system's analysis. A specific implementation step includes: data source access and regular updates, frequency-based alias extraction and mapping, and dictionary updates.

[0094] 3.1.1 Data source access and regular updates 3.1.1.1: Data source access E-commerce platform data: The system is connected to the interfaces of mainstream e-commerce platforms to obtain the latest mobile phone brands, series and models on the shelves in real time, including standard names and multiple aliases.

[0095] A possible demonstration data is as follows, { "Brand": "iPhone", "Series": "iPhone15", "Model": "iPhone15 Pro Max", "Properties": {"Memory": "128GB", "Color": "Black"} } Customer service system conversation records: Get the actual conversation records of users. 推荐分类 It is judged as the "recommend mobile phone model" intention, and after M 品牌类型实体抽取 Entities are extracted, but ultimately no reasonable mapping is found in the existing dictionary. These unrecognized descriptions can enrich the dictionary aliases through subsequent analysis.

[0096] 3.1.1.2: Regular updates Automatically obtain the latest data from the data source regularly to ensure the timeliness and completeness of the dictionary. Set up an automatic pull mechanism to regularly obtain new data from e-commerce platforms and customer service systems.

[0097] E-commerce platform data extraction: Pull the latest mobile phone brand and model information regularly every week to update dictionary content.

[0098] Customer service conversation record analysis: Analyze new user conversation records every week and extract unmatched descriptions.

[0099] 3.1.1.3: Data cleaning and format unification Character removal: remove noise symbols and irrelevant characters (such as emojis, special punctuation) to ensure consistency. For example, "iPhone 15 Pro Max!" is converted to "iPhone15ProMax"; Unify capitalization: Unify brand and model descriptions in uppercase to avoid matching problems caused by inconsistent capitalization.

[0100] One possible format of the converted data is as follows: [ {"Brand": "IPHONE", "Model": "IPHONE15PROMAX"}, {"Brand": "IPHONE", "Model": "IPHONE15PRO"} ] The cleaned data enters the standard dictionary candidate entry library D 候选 , for further alias extraction and frequency analysis. Among them, a possible candidate library storage format is as follows, { "Candidate entry library": [ {"Brand": "IPHONE", "Series": "IPHONE15", "Model": "IPHONE15PROMAX"} } 3.2.2 Frequency-based alias extraction and mapping After data cleaning, the system selects candidate database D 候选 The frequency of the models in the dictionary is counted, and the high-frequency aliases are identified. Then, the user's non-standardized description is associated with the standard models in the dictionary through a mapping mechanism. This step includes frequency analysis, high-frequency alias mapping, and clustering of unmatched aliases.

[0101] 3.2.2.1 High-frequency alias identification For each update, D 候选 The model performs frequency statistics and records the number of occurrences of each description. If the number of occurrences of a description in a day exceeds the set frequency threshold, it is a high-frequency alias, recorded as Alia high-freq , and take the alias as a candidate entry and enter the next processing flow.

[0102] A possible demonstration example is as follows, enter: { "iPhone15PM": 130, "Huawei P50P": 100, "pingguo15ProMax": 10 } Process: Set the threshold to 100, filter by threshold, and select high-frequency aliases.

[0103] Output: [iPhone15PM, Huawei P50P] 3.2.2.2 High-frequency alias mapping a. The system uses a pre-trained language model (such as BERT or Word2Vec) to generate a vector representation for each description (standard name and alias) in the dictionary and stores it in a vector database for subsequent matching.

[0104] b. For high frequency alias Alia high-freq Generate a vector, denoted as Vec Alia_h .

[0105] c. High frequency alias vector Vec Alia_h With the dictionary vector Vec H Perform cosine similarity calculation.

[0106] For example, the high frequency alias "P50P", the vector is Vec Alia_h1 , which can be compared with the standard name "Huawei P50Pro" in the dictionary, vector Vec H2 Calculate the similarity. The similarity calculation method can be expressed as:

[0107] If the similarity between the high-frequency alias and the existing model in the dictionary is higher than the threshold, it is marked as "matched high-frequency alias"; otherwise, it is marked as "unmatched high-frequency alias". For example, if COS(Vec Alia_h1 ,Vec H2 ) is greater than or equal to the threshold, "P50P" is added to the "matched high-frequency alias" queue and the matching information is recorded; if COS(Vec Alia_h1 ,Vec H2 ) is less than the threshold, "P50P" is added to the "unmatched high-frequency alias" list.

[0108] A possible demonstration example is as follows: enter: [P50P, P50Pr, IQO09, IQOONER9, IQOONEO9,13ProMax, iPhone13PM] Process: Extract the vector of the input model, compare it with the dictionary standard name vector, and output the matching result.

[0109] Output: { "Matched high frequency alias": { "P50P": {"Brand":"Huawei", "Standard Name": ""P50Pro"} #Matching information "P50Pr": {"Brand":"Huawei", "Standard Name": ""P50Pro"} #Matching information "Unmatched high frequency alias": [“IQO0NEO9”, “IQOONEO9”, “IQOONER9”, “13ProMax”, “iPhone13PM”] } 3.2.2.3 Clustering of unmatched aliases For unmatched high-frequency aliases, the system will automatically group them using a clustering algorithm, classify similar descriptions into the same category, and select the description with the highest frequency as the standard description. The specific steps are as follows: 1) Unmatched alias clustering: The unmatched high-frequency aliases are processed using clustering algorithms such as DBScan, and descriptions with high similarity are grouped into the same cluster.

[0110] 2) Standard description selection: In each cluster, the description with the highest frequency is selected as the standard description based on the frequency information, and the remaining descriptions are used as aliases of the standard description.

[0111] 3) Brand information supplement: Combined with dictionary H 标准名-别名映射 The brand information in is used to supplement the standard description generated by clustering.

[0112] A possible demonstration example is as follows: enter: [P50P, P50Pr, IQO09, IQOONER9,13ProMax, iPhone13PM] Process: Use clustering algorithms such as DBScan to obtain clustering information; Output: corresponding standard description, brand and alias cluster set: For example: "Standard description": "IQOONEO9", "Brand": "IQOO", "Alias": [IQO0NEO9,IQOONER9]; The clustering results will be used as new candidate entries and enter the next manual review process to ensure accuracy.

[0113] 3.2.3 Dictionary update and optimization Dictionary updating and optimization is based on the clustering results of high-frequency alias matches and unmatched aliases obtained in the previous steps. The dictionary is dynamically expanded and low-frequency descriptions are regularly cleaned up to maintain the efficiency and accuracy of the dictionary.

[0114] Step 1: Dictionary Update Matched high-frequency aliases: Add matched high-frequency aliases directly to the alias list of the dictionary to expand the coverage of the dictionary.

[0115] New candidate entries: For candidate entries generated by unmatched high-frequency aliases, they are added to the candidate queue Q and officially added to the dictionary after manual review and confirmation.

[0116] Step 2: Dictionary optimization and cleaning To ensure the simplicity and efficiency of the dictionary, the system will regularly clean up the less frequently used descriptions in the dictionary and move them to the backup vocabulary.

[0117] Cleaning up low-frequency descriptions: Set a threshold based on the frequency of use (e.g., fewer than 10 occurrences in a year) and move low-frequency descriptions from the dictionary to a backup vocabulary.

[0118] Backup word library call: The cleaned descriptions are still stored in the backup word library for use in special queries to avoid accidental deletion of users' personalized expressions.

[0119] 3.2 Dictionary Matching Reasoning Module The goal of the dictionary matching reasoning module is to receive the mobile phone brand and model entities input by the user, use the system's standard name-alias dictionary, combine keyword matching and vector similarity calculation technology, quickly identify and extract potential mobile phone model standard name candidate sets, and provide accurate candidate information for subsequent processing.

[0120] 3.2.1 Keyword Matching Use multi-layer regular expression matching to quickly filter brand or model words appearing in sentences to form a list of candidate mobile phone models.

[0121] Regular expression construction: Build regular expressions based on brands and models already in the dictionary for quick screening.

[0122] 1) Brand expression: Construct a brand matching regular expression for the brand names and their aliases already in the system.

[0123] 2) Model expression: Construct a regular expression for model matching based on the existing standard models and aliases in the system.

[0124] Keyword matching: The user input sentence is filtered using the above regular expression to extract candidate brands and models that meet the conditions.

[0125] One possible example is as follows: User input: Is there an iPhone 15 series, or a Galaxy S23?

[0126] The system's dictionary vocabulary and constructed regular expressions are as follows: Brand regular expression: brands = r"(iPhone|Samsung)" Model regular expression: models = r"(iPhone15ProMax|GalaxyS23|Neo9)" Matching process: 1) M 推荐分类 Determine the intention to recommend a mobile phone model and proceed to the next step.

[0127] 2) M 品牌类型实体抽取 Extract the entity candidate list: ["iPhne15", "Galaxy S23"] 3) After matching the regular expression, the output is: { "GalaxyS23": {"method": "Keyword matching", "brand": "Samsung"} } Among them, due to the spelling error in the input "iPhne15", it cannot be matched with the standard name "iPhone15" in the dictionary, and only "Galaxy S23" is matched successfully.

[0128] 3.2.2 Vector Similarity Matching To address insufficient matches due to spelling errors or description variations, the system uses a vector similarity matching algorithm to further expand the candidate set and improve parsing results through semantic-level matching.

[0129] a. For the user input question, extract the vector V according to the mobile phone model entity extracted in step 2 input .

[0130] b. The vector Vec generated based on the standard name-alias dictionary generated in step 3 H .

[0131] c. V input and Vec H Calculate cosine similarity and extract similar models.

[0132] One possible example is as follows: User input: Is there an iPhone 15 series, or a Galaxy S23?

[0133] After step 3.2.1, the result is { "GalaxyS23": {"method": "Keyword matching", "brand": "Samsung"}} In the keyword matching, "iPhne15" failed to match the standard model, so the system switched to vector matching: Vector similarity calculation: The system calculates the cosine similarity between "iPhne15" and the standard model "iPhone15" in the dictionary and its related variants (such as "iPhone15Plus", "iPhone15Pro", "iPhone15ProMax").

[0134] Similarity screening: Assuming that the similarity threshold is set to 0.8, the similarity between "iPhne15" and multiple models in the "iPhone15" series is greater than 0.8, so "iPhne15" is mapped to multiple standard models.

[0135] Finally, all the outputs are combined to get { "iPhone15": {"method": "Vector Matching", "brand": "Apple"}, "iPhone15ProMax": {"method": "Vector Match", "brand": "Apple"}, "GalaxyS23": {"method": "Keyword matching", "brand": "Samsung"} } Through the above matching steps, the system forms the final candidate set of mobile phone models and brands based on the user input question. The final output not only contains the standard model candidate set, but also annotates the matching method (keyword matching or vector matching) and brand information, providing accurate candidate information for subsequent processing modules, enhancing the robustness and accuracy of the system. Step 4: Build a mobile phone model recommendation module based on large model analysis The mobile phone model recommendation module based on large model parsing aims to dynamically maintain and filter the candidate mobile phone model entity set through deep semantic understanding of user input, ensuring that the recommendation results are closely aligned with the user's real needs. This module mainly consists of two parts, the model precision parsing module and the user demand reasoning module, which are used for accurate screening of candidate entities and context reasoning respectively.

[0136] 4.1 Model Accurate Analysis Module The model precision parsing module is mainly used to supplement and accurately filter the candidate mobile phone model entity list generated in the previous step (step 3), especially in the following situations: Too many candidate entities: When there are too many models in the candidate list, the module will further filter to reduce redundancy and retain only the core models that best match the user's description.

[0137] Missing candidate entities: When the previous step fails to identify all relevant models of user needs, the module will parse the details of the user input through the large model and supplement the missing candidate models to ensure that user needs are not ignored.

[0138] A specific implementation method is to use the understanding ability of a large language model similar to GPT to cleverly construct a prompt word based on a thought chain for reasoning in the large language model.

[0139] One of the optional large language models is the Qwen-MAX model released by the Alibaba Tongyi Qianwen team.

[0140] One of the feasible prompt word templates based on thought chain is as follows You are an intelligent assistant. Based on the list of potential candidate mobile phone models and the question entered by the user, please extract all the mobile phone models and brands mentioned by the user as completely and accurately as possible.

[0141] # Input description Candidate mobile phone models: User input: # Steps 1. Analyze the brand and model in the user input and identify any ambiguous expressions.

[0142] 2. Filter irrelevant phone models from the candidate list.

[0143] 3. Supplement the list of models that meet the description.

[0144] 4. Output the filtered candidate mobile phone model list.

[0145] # Output format JSON format, the demonstration format is as follows: ```json { "iPhone15": {"method": "Vector Match", "brand": "Apple"} } A specific demonstration example is as follows: Suppose the user enters the question: Is there an iPhone 15 series or Galaxy S23?

[0146] Based on the result of step 3 and the preset prompt word template, the prompt words finally input into the large model are as follows: You are an intelligent assistant. Based on the list of potential candidate mobile phone models and the question entered by the user, please extract all the mobile phone models and brands mentioned by the user as completely and accurately as possible.

[0147] # Input description Candidate mobile phone models: { "iPhone15": {"method": "Vector Matching", "brand": "Apple"}, "iPhone15ProMax": {"method": "Vector Match", "brand": "Apple"}, "GalaxyS23": {"method": "Keyword matching", "brand": "Samsung"} } User input: Is there an iPhone 15 series, or a Galaxy S23? # Steps 1. Analyze the brand and model in the user input and identify any ambiguous expressions.

[0148] 2. Filter irrelevant phone models from the candidate list.

[0149] 3. Supplement the list of models that meet the description.

[0150] 4. Output the filtered candidate mobile phone model list.

[0151] # Output format JSON format, the demonstration format is as follows: ```json { "iPhone15": {"method": "Vector Match", "brand": "Apple"} } 4.2 User Demand Reasoning Module The core of this module is to use a large model to perform in-depth semantic analysis on the user's historical conversations, current questions, and candidate mobile phone model lists, dynamically screen and recommend the mobile phone models that best meet the user's needs. This module pays special attention to the user's contextual expression, progressive needs, limiting conditions, brand preferences, etc., to ensure that the final recommendation results accurately match the user's intentions.

[0152] A specific implementation step is as follows: 4.2.1 Input Data Preparation Context: Collects all the historical conversation content of the user in the current session, including previous inquiries, preferences, negative conditions, etc. Current Query: Get the latest input from the user, which may contain new requirements or conditions; Candidate Models: The list of candidate mobile phone models obtained from the previous step (model precision analysis module).

[0153] 4.2.2 Situational Judgment According to the user's historical conversations and current questions, the user's input context is judged, which is mainly divided into the following categories: Initial demand expression: The user first expresses a demand, and there is no relevant information in the historical conversation; Progressive demand expression: Users gradually refine their demands and add new conditions in multiple rounds of dialogue; Negative condition expression: The user explicitly indicates that certain models or conditions are not required.

[0154] 4.2.3 Choose appropriate reasoning process and thinking chain prompt words According to the results of situational judgment, select the corresponding reasoning process and design professional thinking chain prompts to guide the large model to reason according to the expected logic.

[0155] 4.2.4 Constructing Thinking Chain Prompt Words Design prompts based on the thought chain for the selected scenario to guide the big model to reason step by step and ensure a closed logic loop. The prompts should include: Role setting: clarify the role and tasks of the model; Task description: Describe in detail the task objectives that the model needs to complete; Input instructions: clarify the structure and content of input data; Reasoning steps of thought chain: describe step by step the reasoning steps that the model needs to carry out; Output format: Specify the format of the output results to ensure consistency and ease of parsing.

[0156] 4.2.5 Performing Inference on Large Models The constructed thought chain prompt words and input data are passed to the large model, the reasoning process is executed, and the output result of the model is obtained.

[0157] 4.2.6 Result analysis and output Parse the output of the large model to obtain the final list of recommended mobile phone models and reasons for recommendation, which can be used by subsequent modules or directly fed back to users.

[0158] A specific feasible judgment logic and prompt word template for each scenario is as follows: 1) Initial demand expression Judgment logic: The current user question is judged as the "recommend mobile phone model" intent for the first time in the current user session.

[0159] A possible example dialogue is as follows: {"above": [], "Current sentence": Is there an iPhone 15 series, or a Galaxy S23} The selected prompt word template is as follows: **Role**: You are a professional smartphone recommendation assistant.

[0160] **Task**: Based on the user's current needs, filter and recommend the mobile phone model that best meets the user's needs from the list of candidate mobile phone models.

[0161] **enter**: - **Current Issue**: "{Current Issue}" - **Candidate mobile phone model list**: {Candidate mobile phone model list} **Thinking chain reasoning steps**: 1. **Requirements analysis**: Analyze the current problem and extract the main points of user needs (such as brand, series, model).

[0162] 2. **Candidate screening**: Filter matching models from the candidate phone model list based on the requirements.

[0163] 3. **Generate recommendations**: Sort the screening results and provide a list of recommended models.

[0164] 4. **Generate recommendation reasons**: Based on the screening process, briefly explain the reasons for the recommendation.

[0165] **Output Format**: ```json { "Recommended mobile phone model list": [...], "Recommendation reason": "..." } 2) Progressive demand expression Judgment logic: The current user question is judged to have the intent of "recommending mobile phone models", and there is already a question in the above text that is judged to have the intent of "recommending mobile phone models", but the current user question does not extract a valid entity after the above steps.

[0166] A possible example dialogue is as follows: {"above":["Is there an iPhne15 series, or Galaxy S23"], "Current sentence": "I want pro"} The selected prompt word template is as follows: **Role**: You are a professional smartphone recommendation assistant.

[0167] **Task**: Dynamically adjust the list of candidate mobile phone models based on the user's historical conversations and current questions, and recommend the mobile phone model that best meets the user's needs.

[0168] **enter**: - **Historical Dialogue**: {Historical Dialogue} - **Current Issue**: "{Current Issue}" - **Candidate mobile phone model list**: {Candidate mobile phone model list} **Thinking chain reasoning steps**: 1. **Comprehensive needs analysis**: Combine historical conversations and current issues to summarize all user needs and conditions.

[0169] 2. **Conditional analysis**: Identify the user’s incremental needs, such as new restrictions or preferences.

[0170] 3. **Candidate update**: Based on comprehensive requirements, select eligible models from the candidate phone model list.

[0171] 4. **Generate recommendations**: Sort the screening results and provide a list of recommended models.

[0172] 5. **Generate recommendation reasons**: Based on the screening process, explain the basis for the recommendation.

[0173] **Output Format**: ```json { "Recommended mobile phone model list": [...], "Recommendation reason": "..." } 3) Negative conditional expression Judgment logic: The current user question contains negative keywords, and a list of candidate mobile phone entities already exists.

[0174] A possible example dialogue is as follows: {"above":["Is there an iPhne15 series, or Galaxy S23"], "Current sentence": "I don't want Samsung anymore"} The selected prompt word template is as follows: **Role**: You are a professional smartphone recommendation assistant.

[0175] **Task**: Based on the user's negative needs, eliminate unqualified models from the list of candidate mobile phone models and recommend the mobile phone model that best meets the user's needs.

[0176] **enter**: - **Historical Dialogue**: {Historical Dialogue} - **Current Issue**: "{Current Issue}" - **Candidate mobile phone model list**: {Candidate mobile phone model list} **Thinking chain reasoning steps**: 1. **Negative demand analysis**: Identify models or conditions that users clearly do not need.

[0177] 2. **Candidate elimination**: Remove rejected models from the candidate list.

[0178] 3. **Generate Recommendations**: Provides a list of remaining recommended models.

[0179] 4. **Generate recommendation reasons**: Explain the basis for exclusion and recommendation.

[0180] **Output Format**: ```json { "Recommended mobile phone model list": [...], "Recommendation reason": "..." } Through the above detailed implementation steps and logical processes, we use the thinking chain prompt words to design: based on the thinking chain prompt words, we guide the big model to think step by step, deeply understand user needs, and ensure that the reasoning process is clear and controllable; Contextualized reasoning process: Design corresponding reasoning processes and prompts for different user input scenarios (initial demand, progressive demand, negative conditions, etc.) to ensure comprehensive and accurate processing; Dynamic candidate set maintenance: In multiple rounds of conversations, the candidate mobile phone model list is updated in real time to keep pace with user needs; This enables the 4.2 User Demand Reasoning Module to efficiently and accurately analyze users' real needs and provide professional mobile phone model recommendations.

[0181] The above is only a specific implementation of the invention, but the protection scope of the invention is not limited to it. Any changes or substitutions that are not conceived through creative work should be included in the protection scope of the invention. Therefore, the protection scope of the invention should be based on the protection scope defined in the claims.

Claims

1. An intelligent mobile phone model recommendation system with collaborative processing of large and small models, characterized by: include The intent classification module recognizes the text content input by the user, distinguishes the intent scenarios, and marks the corresponding intent category labels; An entity recognition module extracts product brand and model entities from the text content input by the user, defines labels for the entity content and non-entity content of the brand and model, and reversely optimizes the entity recognition module; The model inference module maintains a dictionary containing brand standard names and alias mappings, and matches the brand and model entities extracted by the above entity recognition module with the standard names in the dictionary according to the intent category labels, and performs vector similarity matching at the same time, and outputs a candidate set of models and brands; The model recommendation module determines the input context based on the text content input by the user, and uses the thinking chain prompt words built according to the input context to dynamically maintain and filter the candidate set generated in the above model reasoning module, and recommend the brand model that best meets the user's needs.

2. According to claim 1, the intelligent mobile phone model recommendation system for large and small model collaborative processing is characterized by: The intent classification module and the entity recognition module both include a data collection module for data augmentation processing, and the specific method is as follows: First, collect text content containing various product brands and models from various platforms to build relevant data sets. The collected samples should cover the real product names of multiple brands and models. Secondly, the collected data is annotated and stored as a structured JSON format file. Each piece of data in the intent classification module includes an intent label field whose value is "recommended product model" or "non-recommended product model", and each piece of data in the entity recognition module includes an entity list label field containing an entity category, an entity value, and the start and end positions in the text content; Finally, the dataset is augmented to enhance the coverage of expressions, including synonym replacement, abbreviation expansion, and brand generalization by replacing brands or models to expand data samples.

3. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 2 is characterized by: The intent classification module also includes a classification model training module that uses the collected data to perform intent classification training. The specific method is as follows: S101: Feature vector extraction: Use the semantic vector extraction model Bert to vectorize the collected data. Each piece of data will be converted into a vector of fixed length. S102: Input the feature vector to the classification model, and input the extracted feature vector to the DistilBERT model. The DistilBERT model hierarchically processes the input feature vector, gradually extracts more advanced semantic and contextual information, and outputs a fixed-length feature representation that retains important semantic information in the sentence. S103: The feature vector is passed to a fully connected layer. The feature representation is passed to a fully connected layer, which performs a linear transformation on the feature, maps the high-dimensional feature vector to a low-dimensional feature vector for classification, obtains the logits value of each category, and finally outputs the logits value of the category; S104: The logits value is passed to the Softmax layer, and the logits value is converted into probability values ​​of two categories to determine whether the input text belongs to the "recommended product model" or "non-recommended product model" category; S105: Calculate the cross entropy loss, compare the output probability of the softmax layer with the actual label, calculate the cross entropy loss, measure the difference between the model prediction and the true label, and output the loss value of a single sample; S106: Back propagation optimization model, using the calculated loss value, update the model parameters through the back propagation algorithm, so that the model gradually learns the characteristics of each category, so that the loss function converges, thereby optimizing the model performance. After the training is completed, save the final model parameters, deploy and use, and record them as M 推荐分类 .

4. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 3 is characterized by: The entity recognition module also includes an entity recognition training module for performing entity recognition training using the collected data. The specific steps of extracting brand and model entities from text content for recognition training are as follows: S201: Use the Bert model to vectorize the collected data at the word level, convert each sentence into a sequence, and each word corresponds to a feature vector; S202: Input the word vector sequence output by Bert into the bidirectional long short-term memory network layer to capture the context sequence information of the sentence and output the word-level feature representation containing the context information; S203: passing the above output feature representation to the conditional random field layer to model the dependency between labels. The conditional random field layer calculates the scores of all possible label sequences, selects the label sequence with the highest score as the prediction result, and outputs the predicted label of each word. S204: Calculate the loss function, input the predicted label sequence and the real label sequence, use the log-likelihood loss function of the conditional random field layer, calculate the negative log-likelihood loss, measure the degree of fit of the model to the training data, calculate the difference between the label sequence predicted by the model and the real label sequence, and output the loss value; S205: Using the calculated loss value, update the model parameters through the back propagation algorithm, adjust the model parameters according to the loss gradient, so that the model gradually learns the characteristics of the entity. After the training is completed, save the final model parameters and deploy them for use, denoted as M 品牌类型实体抽取 .

5. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 1 is characterized by: The model inference module includes dynamic maintenance and update of brand standard name and alias mapping dictionary, and the specific method includes the following steps: S301: Data source access and regular update, access to data from various platforms, real-time acquisition of corresponding product brands, series and models, including standard names and alias data to enrich the original dictionary, data cleaning and format unification of the acquired data, the cleaned and unified data enters the standard dictionary candidate entry library D 候选 , for further alias extraction and frequency analysis; S302: Frequency-based alias extraction and mapping. After data cleaning, the system performs a 候选 The frequency of the models in the dictionary is counted to identify the high-frequency aliases, and the user's non-standard description is associated with the standard models in the dictionary through a mapping mechanism; S303: Dictionary update and optimization. The matched high-frequency aliases are directly added to the alias list of the dictionary to expand the coverage of the dictionary. For the candidate entries generated by the unmatched high-frequency aliases, they are added to the candidate queue and officially added to the dictionary after manual review and confirmation. At the same time, the system will regularly clean up the less frequently used descriptions in the dictionary and move them to the backup vocabulary.

6. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 5, characterized in that: The specific steps of frequency-based alias extraction and mapping are as follows: High-frequency alias frequency analysis: D collected in each update 候选 The model performs frequency statistics and records the number of occurrences of each description. If the number of occurrences of a description in a day exceeds the set frequency threshold, it is a high-frequency alias, recorded as Alia high-freq , take the alias as a candidate entry and enter the next processing flow; High-frequency alias mapping: The system uses a pre-trained language model to generate a vector representation VEC for the standard name and alias of each description in the dictionary H , and stored in the vector database; Alia high-freq Generate a vector, denoted as Vec Alia_h , the high frequency alias vector Vec Alia_h With the dictionary vector Vec H Perform cosine similarity calculation. If the similarity between the high-frequency alias and the existing model in the dictionary is higher than the threshold, it is marked as "matched high-frequency alias"; otherwise, it is marked as "unmatched high-frequency alias"; Clustering of unmatched aliases: For unmatched high-frequency aliases, the system will automatically group them through a clustering algorithm, classify similar descriptions into the same category, and select the description with the highest frequency as the standard description. Specifically: Perform clustering algorithms such as DBScan on unmatched high-frequency aliases, classify descriptions with high similarity into the same cluster, and in each cluster, select the description with the highest frequency as the standard description based on the frequency information, and the remaining descriptions as aliases of the standard description. Combined with the brand information in the dictionary, the brand information is supplemented for the standard description generated by clustering, and the clustering results will be used as new candidate entries.

7. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 6 is characterized by: The model inference module also includes matching the brand model based on keywords. The specific matching method steps are as follows: First, keyword matching is performed, and multi-layer regular expression matching is used to quickly screen the brand or model words that appear in the sentence to form a list of candidate brands and models, and regular expressions are constructed based on the brands and models that already exist in the dictionary; Brand expression: construct a brand matching regular expression for the brand names and their aliases in the system; Model expression: construct a model matching regular expression for the standard models and aliases in the system; filter the user input statements using the regular expressions to extract candidate brands and models that meet the conditions; Secondly, vector similarity matching is performed to further expand the candidate set, and the parsing effect is improved through semantic matching. The entity recognition module is used to extract brand and model entities from the user input text content, and the vector V is extracted. input , and then use the vector Vec generated by the standard name-alias dictionary H , for V input With Vec H Calculate the cosine similarity, extract similar models, and finally form a candidate set of product models and brands.

8. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 1 is characterized by: The model recommendation module includes a reasoning and analysis module for user needs. The specific reasoning and analysis method steps are as follows: S401: Input data preparation, collecting all historical conversation content of the user in the current session, including previous inquiries, preferences, and negation condition information, and obtaining the user's latest input and candidate set; S402: Context judgment: judging the user's input context based on the user's historical conversations and current questions, specifically: initial demand expression, progressive demand expression, and negative condition expression; S403: Select appropriate reasoning process and thinking chain prompt words, select the corresponding reasoning process according to the situation judgment results, and design thinking chain prompt words to guide the model to reason according to the expected logic; S404: Execute the large model reasoning, pass the constructed thinking chain prompt words and input data to the large model, execute the reasoning process, and obtain the output result of the model; S405: Result analysis and output: Analyze the output of the model, obtain the final recommended model list and recommendation reasons and feed it back to the user.

9. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 8, characterized in that: The construction of the thought chain prompt words includes: Role setting: clarify the role and tasks of the model; Task description: Describe in detail the task objectives that the model needs to complete; Input instructions: clarify the structure and content of input data; Reasoning steps of thought chain: describe step by step the reasoning steps that the model needs to carry out; Output format: Specify the format of the output results to ensure consistency and ease of parsing.

10. The mobile phone model intelligent recommendation system for large and small model collaborative processing according to claim 9, characterized in that: The model recommendation module also includes adjusting and analyzing the candidate set that ultimately forms the product model and brand. When there are too many models in the candidate list, the module will further screen and reduce redundancy, retaining only the core models that best match the user's description; when all relevant models that meet the user's needs cannot be identified, the module will use the model to parse the details entered by the user and supplement the missing candidate models.

Images