Automobile product matching system and method based on large language model

Abstract

The invention relates to the technical field of semantic understanding, in particular to an automobile product matching system and method based on a large language model. The method comprises the following steps: receiving selling point information of a target vehicle type input by a vehicle enterprise, and constructing target selling point features according to semantics of the selling point information; obtaining a plurality of recommendation interest circle layers output by the large language model based on target selling point features; according to matching records of a plurality of historical interest circle layers and selling point features, obtaining distances of different interest circle layers; constructing a circle layer association graph of each interest circle layer according to the distance; mapping the plurality of recommended interest circle layers into a circle layer association graph, and selecting extended interest circle layers within a set range around the plurality of recommended interest circle layers; and pushing to a user terminal of the interest circle layer. According to the invention, the appropriate crowd can be matched for the automobile product and accurately pushed to the terminal of the crowd, so that the system operation efficiency is improved, and the effect of accurate recommendation is achieved while network resources are saved.

Description

Technical Field

[0001] The present invention relates to the technical field of semantic understanding, and in particular to an automobile product matching system and method based on a large language model. Background Art

[0002] With the development of the Internet and big data technology, car companies will choose to release new models through the Internet. In the existing technology, car companies will publish the description, selling points and other data and pictures of new models on public websites. Users can view the information of new models by logging into the website.

[0003] In the prior art, users obtain vehicle model information by actively viewing the information, which lacks a personalized recommendation function. This solution requires that vehicle model information be published indiscriminately to each user's terminal, resulting in a waste of network resources and failing to achieve a good promotion effect.

[0004] In view of this, the present invention is proposed. Summary of the invention

[0005] The purpose of the present invention is to provide an automobile product matching system and method based on a large language model, which can match automobile products to suitable people and accurately push them to the terminals of the people, thereby improving the system operation efficiency, saving network resources and achieving the effect of accurate recommendation.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for matching automobile products based on a large language model, comprising: The large language model is preliminarily trained using a first training sample, wherein the first training sample includes a selling point information sample and an interest circle information label to which the user belongs; the first training sample is obtained by matching the selling point information sample with the interest circle information label according to the semantic similarity; Build a dynamic profile of each interest circle in multiple interest circles; Receive selling point information of a target vehicle model input by a car manufacturer, and construct target selling point features according to the semantics of the selling point information; According to the target selling point feature, a second training sample is selected from the data set, including a selling point feature sample and a dynamic portrait label; the data set is constructed according to the user's purchase behavior of multiple car models; Inputting the target selling point feature and the second training sample into a large language model, retraining the large language model that has been preliminarily trained using the second training sample, and obtaining a plurality of recommended interest circles output by the large language model based on the target selling point feature; According to the matching records of multiple interest circles and selling point features in history, the distances between different interest circles are obtained; and according to the distances, a circle association graph of each interest circle is constructed; Mapping the multiple recommended interest circles into a circle association graph, and selecting an extended interest circle within a set range around the multiple recommended interest circles; The target vehicle model information is pushed to user terminals of multiple recommended interest circles and extended interest circles through the network.

[0007] In a second aspect, the present invention provides an automobile product matching system based on a large language model, comprising: A preprocessing module is used to perform preliminary training on the large language model using a first training sample, wherein the first training sample includes a selling point information sample and an interest circle information label to which the user belongs; the first training sample is obtained by matching the selling point information sample with the interest circle information label according to the semantic similarity; and a dynamic portrait of each interest circle in a plurality of interest circles is constructed; An input module is used to receive the selling point information of the target vehicle model input by the car manufacturer, and construct the target selling point features according to the semantics of the selling point information; A recommendation module is used to filter out a second training sample from a data set according to the target selling point feature, including a selling point feature sample and a dynamic portrait label; the data set is constructed according to the user's purchasing behavior for multiple car models; the target selling point feature and the second training sample are input into a large language model, the large language model that has been preliminarily trained is retrained using the second training sample, and multiple recommended interest circles output by the large language model are obtained based on the target selling point feature; the distances between different interest circles are obtained according to the historical matching records of multiple interest circles and selling point features; a circle association graph of each interest circle is constructed according to the distance; the multiple recommended interest circles are mapped into the circle association graph, and an extended interest circle within a set range around the multiple recommended interest circles is selected; The output module is used to push the target vehicle model information to user terminals of multiple recommended interest circles and extended interest circles through the network.

[0008] Compared with the prior art, the present invention has the following beneficial effects: In terms of the overall idea, the present invention inputs the target selling point features, uses the matching ability of the large language model to obtain the recommended interest circles of the first stage, and then obtains the extended interest circles of the second stage based on the circle association graph, so that a more comprehensive and accurate interest circle that matches the target selling point features is obtained through two-stage processing, thereby achieving the purpose of accurate recommendation and saving network resources. The present invention adopts a two-stage training method of preliminary training and retraining for the large language model. The first training sample is constructed by language similarity, so that the large language model has the ability to semantically associate from selling points to interest circles. The second training sample is constructed by the dynamic portrait of the selling point features and the interest circles, so that the large language model can learn the matching relationship between the selling point and the interest circle from a deep level, and when selecting the second training sample, the large language model can match the recommended interest circle that is closer to the target selling point features through the comparison training method of positive and negative samples, thereby improving the accuracy of the recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0010] Figure 1 is a flow chart of an automobile product matching method based on a large language model provided by an embodiment of the present invention; Figure 2 It is a structural diagram of an automobile product matching system based on a large language model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The following is a description of exemplary embodiments of the present invention in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and conciseness, the description of well-known functions and structures is omitted in the following description.

[0012] Example 1 The method provided in this embodiment is applicable to the case where automobile product information is pushed to a matching user terminal based on the selling point information (or promotional information) of the vehicle model. This method is executed by an automobile product matching system based on a large language model, see Figure 1 , the method comprises the following steps.

[0013] S110: Perform preliminary training on the large language model using the first training sample.

[0014] This embodiment does not limit the type of large language model, which can be an LLM model. The first training sample includes a selling point information sample and an interest circle information label to which the user belongs. The selling point information is text and data related to the selling point of the vehicle model. For example, a 360-degree panoramic image and a driving range of 800km are used as samples input to the large language model, which are called selling point information samples. The interest circle information to which the user belongs includes description information of the interest circle. For example, the description information of the self-driving circle includes: a group that is keen on self-driving travel and likes to explore natural scenery and foreign culture. The members of this circle not only enjoy driving itself, but also love to plan travel routes and pursue independent travel experiences. The interest circle information is used as a label for training the large language model, which is called an interest circle information label. In the actual training process, different numbers are used to represent different interest circle information labels. The label has a corresponding relationship with the sample, that is, a certain selling point information will attract users in certain interest circles. The corresponding relationship of the first training sample is obtained by matching the semantic similarity between the selling point information sample and the interest circle information label. For example, the selling point information is suitable for self-driving and has off-road capability, which has a high semantic similarity with the descriptive information of the self-driving circle "enthusiastic about self-driving travel", and thus constitutes the first training sample.

[0015] The selling point information samples are embedded to obtain feature vectors, which are input into the large language model. The parameters of the large language model are adjusted to make the output of the large language model close to the interest circle information label, thus obtaining the large language model after preliminary training.

[0016] S120. Construct a dynamic portrait of each interest circle in the multiple interest circles.

[0017] The dynamic portrait D of each interest circle represents the characteristics of the interest circle, and as the focus or interest points of users in the interest circle change, the dynamic portrait of the interest circle changes dynamically over time.

[0018] S130: receiving selling point information of a target vehicle model input by a car manufacturer, and constructing target selling point features according to the semantics of the selling point information.

[0019] The car company inputs the selling point information of a target car model through the terminal, such as large space and energy saving. This embodiment performs semantic extraction and analysis on the selling point information to form a feature vector, which is called the target selling point feature.

[0020] S140. Filter out a second training sample from the data set according to the target selling point features and the dynamic portrait.

[0021] The large language model after initial training only has the function of matching interest circles according to descriptive selling point information. The second training sample (including selling point feature samples and dynamic portrait labels) is used to retrain the large language model after initial training. By semantically extracting and parsing the selling point information, the selling point features can better reflect the characteristics of the selling points; by constructing dynamic portraits based on descriptive interest circle information, the characteristics of the interest circle can also be better reflected. The selling point feature samples are input into the large language model, and the parameters of the large language model are continuously adjusted so that the output of the large language model approaches the dynamic portrait labels. The dynamic portrait labels have a one-to-one correspondence with the interest circles. Therefore, the interest circles corresponding to the target selling point features can be obtained using the trained large language model.

[0022] The following describes the process of obtaining the second training sample: First, a data set is constructed based on the user's purchase behavior for multiple car models. For example, if the user who purchased the M car model belongs to the self-driving circle and the animation circle, then the selling point feature samples of the M car model and the dynamic portrait labels of the self-driving circle and the animation circle constitute the second training sample. Since the data set stores the selling point features of multiple car models and the features of multiple dynamic portraits, and the target selling point feature is the feature of a certain car model, in order to enable the large language model to have a more accurate matching ability for the car model, it is necessary to screen out the second training sample (including positive samples and negative samples) with a certain semantic similarity to the target selling point feature from the data set. Specifically, the semantic similarity between the target selling point feature and each selling point feature in the database is calculated, and the selling point feature samples and the corresponding dynamic portrait labels with a semantic similarity greater than the high threshold are used as positive samples, and the selling point feature samples and the corresponding dynamic portrait labels with a semantic similarity less than the low threshold are used as negative samples. Among them, the high threshold is 90%, the low threshold is 10%, and it is assumed that the target selling point feature is represented by T4. BERT is used to semantically represent T4 to obtain the feature vector of the selling point. , see formula (1), the selling point features in the dataset are also processed by BERT into feature vectors . Using cosine similarity ( ) Calculate the similarity between T4 and the selling point feature vector in the data set, see formula (2). The selling point features and corresponding dynamic portrait labels with a similarity greater than 90% are taken as positive samples, and the selling point features and corresponding dynamic portrait labels with a similarity less than 10% are taken as negative samples.

[0023] ; (1) ; (2) S150, inputting the target selling point feature and the second training sample into the large language model, retraining the large language model that has been initially trained using the second training sample, and obtaining multiple recommended interest circles output by the large language model based on the target selling point feature.

[0024] Specifically, fill the target selling point features into the question part of the prompt template, fill the second training sample into the example part of the prompt template, and fill the recommended interest circle into the answer part of the prompt template. For example, the prompt template includes: the selling points of the car models that users are interested in can be summarized as {T4} after analysis. Please perform correlation analysis based on the dynamic portraits {D} of different interest circles to predict which interest circles are the audience of this selling point. Here are some examples: {positive sample, negative sample}. Among them, {} is the part that needs to be filled with content.

[0025] The filled prompt template is input into the large language model, and it is pre-agreed that the example in the prompt template is taken out as the second training sample. The language model is trained again using the second training sample. After the training is completed, the target selling point feature T4 is extracted from the prompt template as the input of the large language model. The large language model obtains multiple dynamic portraits based on the target selling point features, and then obtains the interest circles corresponding to the dynamic portraits, which are called recommended interest circles. Exemplarily, the large language model is called to obtain TOP-K recommended interest circles, where K can be 3.

[0026] S160. Obtain distances between different interest circles based on historical matching records of multiple interest circles and selling point features; and construct a circle association graph for each interest circle based on the distances.

[0027] This step is mainly to construct a circle association graph based on the distances between different interest circles, so as to expand the recommended interest circles later. Since the data set stores the selling point features of multiple car models and the features of multiple dynamic portraits, this embodiment obtains the historical matching records of multiple interest circles and selling point features from the aforementioned data set. First, for the aforementioned matching records, the selling point features with similar semantics are merged. It can be understood that when the selling point features are merged, the corresponding interest circles will also be merged. Count the number of times different interest circles are jointly recommended (that is, the number of times they appear in the recommended circle of the same selling point feature), as well as the similarity of the static portraits S of different interest circles; the distances of different interest circles are obtained based on the number and similarity. Use formula (3) and formula (4) to calculate the distances of different interest circles. .

[0028] ; (3) ; (4) in, and These are two static portraits of two interest circles. It is the result of subtracting the similarity of two interest circles from 1. It is an interest circle The number of co-occurrences of is a fixed hyperparameter.

[0029] Finally, we get the circle association diagram ,in Nodes representing interest circles, Indicates the distance between interest circles.

[0030] S170: Map multiple recommended interest circles to a circle association graph, and select an extended interest circle within a set range around the multiple recommended interest circles.

[0031] Determine the nodes consistent with each recommended interest circle in the circle association graph; take any node as the center, and traverse the nodes within the set range around it as an extended interest circle. Assuming that the self-driving circle is a recommended interest circle, draw a circle with the self-driving circle node as the center and the set range (for example, 5) in the circle association graph, and take the nodes within the circle that are directly / indirectly connected to the self-driving circle node as the extended interest circle.

[0032] S180: Push the target vehicle model information to user terminals of multiple recommended interest circles and extended interest circles through the network.

[0033] The target vehicle model information includes the selling points, vehicle configuration, pictures, etc. The relationship between users and interest circles can be pre-built, for example, the social groups on the self-driving website are used as users of the self-driving circle.

[0034] Optionally, the number of interest circles to be promoted is preset, for example, 3. If the sum of the number of recommended interest circles and extended interest circles is less than or equal to 3, the target vehicle model information is directly pushed to the user terminals of multiple recommended interest circles and extended interest circles. If the sum of the number of recommended interest circles and extended interest circles is greater than 3, it is necessary to obtain 3 interest circles through a weighted sampling algorithm, for example, weighted sampling of extended interest circles and recommended interest circles is performed based on the historical push records to the user terminal.

[0035] The following formula is used to sample the circles: ; (5) ; (6) in, is the sampling weight of the i-th interest circle, represents the i-th interest circle, is a fixed hyperparameter, It is to the interest circle The number of times the vehicle model information is pushed to the user terminal. The more times the information is pushed, the lower the sampling weight, which is conducive to improving the diversification of the push interest circles. represents the weighted sampling algorithm, Represents the sampling weight of each interest circle, Represents the set of recommended interest circles and extended interest circles. Indicates the interest circles returned by the sampling.

[0036] In terms of the overall idea, the present invention inputs the target selling point features, uses the matching ability of the large language model to obtain the recommended interest circles of the first stage, and then obtains the extended interest circles of the second stage based on the circle association graph, so that a more comprehensive and accurate interest circle that matches the target selling point features is obtained through two-stage processing, thereby achieving the purpose of accurate recommendation and saving network resources. The present invention adopts a two-stage training method of preliminary training and retraining for the large language model. The first training sample is constructed by language similarity, so that the large language model has the ability to semantically associate from selling points to interest circles. The second training sample is constructed by the dynamic portrait of the selling point features and the interest circles, so that the large language model can learn the matching relationship between the selling point and the interest circle from a deep level, and when selecting the second training sample, the large language model can match the recommended interest circle that is closer to the target selling point features through the comparison training method of positive and negative samples, thereby improving the accuracy of the recommendation.

[0037] Since the selling points of car models (car brand marketing) are strongly correlated with some circles, such as the self-driving circle, cycling circle, and modification circle, it is difficult to avoid the problem of homogeneous recommendation results by directly performing prompt engineering on the large language model. That is, given different descriptions of car model selling points, the model will recommend similar circles and lacks the ability to discover the correlation between different circles. To this end, a two-stage interest circle recommendation method is proposed. Based on the large language model prompt engineering, the correlation analysis of different circles is performed for the large amount of recommendation data generated in actual production scenarios to improve the diversity and comprehensiveness of the model recommendation results.

[0038] Example 2 Based on the above embodiment, this embodiment optimizes the process of obtaining dynamic portraits. Constructing a dynamic portrait of each interest circle in multiple interest circles includes the following three steps: Step 1: Obtain user information I of multiple users, and identify the interest circles to which the users belong and static portraits S of the interest circles based on the user information I. The user information includes age, region, interests, consumption, and catalysts (active social media) and other information.

[0039] The portrait modeling model is, for example, the LLM model. Use user information to build a prompt template , input the prompt template into the portrait modeling model (such as the LLM model) to obtain the static portrait output by the portrait modeling model. The prompt template is, for example: "The user information is {I}. Please conduct portrait modeling for this circle, analyze user behavior from the perspective of sociology, and infer the possible needs of this circle group for car model characteristics based on their behavior. Here are some examples: {}" The variables in {} can be automatically replaced. Examples are as follows: Self-driving circle refers to a group of people who are keen on self-driving travel and like to explore natural scenery and foreign cultures. Members of this circle not only enjoy driving itself, but also love to plan travel routes and pursue independent travel experiences. The static portrait of the self-driving circle includes: the self-driving circle is mainly composed of young and middle-aged people and family groups, with a balanced gender distribution, and usually has a certain economic strength and free time. Members of the self-driving circle prefer free and flexible travel methods and enjoy the exploration process on the road, especially through diverse travel locations such as natural scenery and cultural relics. This group likes to share experiences such as self-driving routes, travel scenery and travel equipment on social platforms. The consumption focus of users in the self-driving circle is on high-quality, reliable and comfortable travel equipment, such as camping equipment, car refrigerators, portable chargers, etc. The vehicle model demand of the self-driving circle is also part of the static portrait. The self-driving circle group attaches great importance to the high passability, comfort and long-distance driving performance of the vehicle, and especially prefers SUVs or off-road vehicles that are suitable for various road conditions. Large space and flexible storage design are important needs to meet the needs of equipment storage during self-driving travel. This group focuses on the reliability, safety and fuel economy of the vehicle, and prefers safety features such as body stability control system, panoramic image, automatic driving assistance, etc. In addition, members of the self-driving circle also have high requirements for configurations such as in-vehicle navigation system, long-distance driving assistance, and seat comfort to improve the comfort and convenience of long-distance journeys.

[0040] Step 2: Obtain hot topic data in interest circles.

[0041] Specifically, relevant data is crawled from social media platforms (data is crawled according to catalysts), and natural language processing technology is used for data cleaning, text tokenization, and word frequency statistics to identify the hot topic data M of each interest circle. Among them, data cleaning is to preprocess the original data and remove irrelevant information (such as punctuation marks, emojis, extra spaces, advertising information, etc.). This method uses regular expressions (Regex) to remove URLs, HTML tags, special characters, and filter out stop words (such as "de", "shi", etc.) to reduce interference from irrelevant vocabulary. Text tokenization is to divide continuous text into several words or phrases. This method uses the Jieba tokenization tool, and its core principle is: based on the DAG (directed acyclic graph) search of the prefix dictionary, and combined with the dynamic programming algorithm to find the maximum probability path, so as to achieve the optimal tokenization result. Word frequency statistics is to count the tokenized text to identify the words with higher frequencies. This method is implemented using the TF-IDF (term frequency-inverse document frequency) algorithm.

[0042] Step 3: Use the static portrait S of the interest circle and the hot topic data M to construct a prompt template, and input the prompt template into the portrait modeling model to obtain the dynamic portrait output by the portrait modeling model. The dynamic portrait includes the vehicle model requirements of the user.

[0043] Construct a prompt template It is required that the portrait modeling model can be based on the static portrait and the hot topic data M to analyze the interest hotspots and vehicle model requirements of different circles, so as to perform periodic dynamic updates on the static portrait based on the social media hot data.

[0044] The core idea of constructing the prompt template is as follows: "The user information is { }, and in the past period of time, the hot topics that the users in this circle have paid attention to on social media are { }. Please perform portrait modeling on this circle, analyze the user behavior from the perspective of sociology, etc., summarize its interest hotspots, and infer the possible requirements of this circle group for the characteristics of vehicle models according to their behavior."

[0045] According to the actual input data and , format the prompt template , and call the portrait modeling model to perform portrait modeling to obtain the dynamic portrait of the circle .

[0046] This embodiment has the following technical effects: There is a gap between the existing portrait modeling methods and actual production scenarios. The features obtained by modeling are often too general and lack relevance to specific application requirements. For example, the age distribution, regional distribution and other information obtained by modeling analysis are different from the interest distribution related to the selling points of vehicle models required in actual application scenarios. To address this problem, this embodiment proposes a circle portrait modeling method for vehicle model demand analysis to ensure high relevance and high availability between the modeled features and actual production requirements.

[0047] As time goes by, the focus and interests of the circle may shift. In order to better adapt to the dynamic changes in circle characteristics, it is necessary to periodically model the circle dynamically. To address this problem, this embodiment proposes a method for periodically and dynamically updating static circle portraits based on social media hotspot data, thereby effectively capturing the interest changes of the current circle.

[0048] Example 3 Based on the above embodiment, this embodiment optimizes the process of constructing the target selling point feature, and constructs the target selling point feature according to the semantics of the selling point information, including the following operations: Step 1: Perform syntax parsing and semantic analysis on the selling point information T1 to obtain the selling point-description tuple.

[0049] Selling point-description tuples T2 are extracted from the selling point information T1, and semantically similar tuples are merged.

[0050] Step 2: Use the automobile knowledge base to detect and explain the bigrams, forming selling point-description-explanation triplets.

[0051] Based on the existing list of automotive proper nouns, the RAG (retrieval-augmented generation) architecture is used to explain the proper nouns in the automotive field. If the proper noun exceeds the scope of local knowledge, the search tool is called to obtain the corresponding noun explanation from the Internet, and finally the automotive knowledge base is constructed. The proper noun detection is performed on the selling point-description tuple. If the tuple involves a proper noun in the automotive field, the noun explanation is performed through the automotive knowledge base to obtain the selling point-description-explanation triple T3.

[0052] Step 3: Conduct correlation analysis between the triples and automobile brand marketing to form a selling point-description-explanation-correlation quadruple, and retain the quadruple with correlation greater than the set threshold as the target selling point feature.

[0053] A correlation analysis is performed on the selling point-description-explanation triple and automobile brand marketing, that is, whether the triple belongs to the category of automobile brand marketing is determined, and a correlation score is obtained. For example, the probability that the triple belongs to automobile brand marketing is determined by a classification model, and the probability is used as the correlation score. The classification model is pre-trained using multiple triple samples belonging to automobile brand marketing. Triples with correlations less than or equal to a set threshold (for example, 70%) are eliminated, and triplets with correlations greater than the set threshold are retained, and specific correlation scores are concatenated at the end to form a selling point-description-explanation-correlation quadruple T4.

[0054] This embodiment has the following technical effects: Since the selling point information input by users is uncertain, in order to eliminate the interference of problems such as irrelevant to automobile brand marketing, repeated emphasis of a single selling point, and grammatical redundancy in the selling point information, a text description normalization method for vehicle model selling points is proposed to parse and normalize the selling point information input by users, thereby reducing the performance loss caused by the uncertainty of the selling point information input by users.

[0055] Example 3 This embodiment provides an automobile product matching system based on a large language model. Figure 2 ,include: A preprocessing module is used to perform preliminary training on the large language model using a first training sample, wherein the first training sample includes a selling point information sample and an interest circle information label to which the user belongs; the first training sample is obtained by matching the selling point information sample with the interest circle information label according to the semantic similarity; and a dynamic portrait of each interest circle in a plurality of interest circles is constructed; An input module is used to receive the selling point information of the target vehicle model input by the car manufacturer, and construct the target selling point features according to the semantics of the selling point information; A recommendation module is used to filter out a second training sample from a data set according to the target selling point feature, including a selling point feature sample and a dynamic portrait label; the data set is constructed according to the user's purchasing behavior for multiple car models; the target selling point feature and the second training sample are input into a large language model, the large language model that has been preliminarily trained is retrained using the second training sample, and multiple recommended interest circles output by the large language model are obtained based on the target selling point feature; the distances between different interest circles are obtained according to the historical matching records of multiple interest circles and selling point features; a circle association graph of each interest circle is constructed according to the distance; the multiple recommended interest circles are mapped into the circle association graph, and an extended interest circle within a set range around the multiple recommended interest circles is selected; The output module is used to push the target vehicle model information to user terminals of multiple recommended interest circles and extended interest circles through the network.

[0056] The system provided in this embodiment can execute the methods provided in embodiments 1 to 3 and have corresponding technical effects.

[0057] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution disclosed in the present invention can be achieved, and this document does not limit this.

[0058] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for matching automobile products based on a large language model, characterized in that: include: The large language model is preliminarily trained using a first training sample, where the first training sample includes a selling point information sample and an interest circle information label to which the user belongs; The first training sample is obtained by matching the selling point information sample with the semantic similarity of the interest circle information label; Build a dynamic profile of each interest circle in multiple interest circles; Receive selling point information of a target vehicle model input by a car manufacturer, and construct target selling point features according to the semantics of the selling point information; According to the target selling point feature, a second training sample is selected from the data set, including a selling point feature sample and a dynamic portrait label; the data set is constructed according to the user's purchase behavior of multiple car models; Inputting the target selling point feature and the second training sample into a large language model, retraining the large language model that has been preliminarily trained using the second training sample, and obtaining a plurality of recommended interest circles output by the large language model based on the target selling point feature; According to the matching records of multiple interest circles and selling point features in history, the distances between different interest circles are obtained; and according to the distances, a circle association graph of each interest circle is constructed; Mapping the multiple recommended interest circles into a circle association graph, and selecting an extended interest circle within a set range around the multiple recommended interest circles; The target vehicle model information is pushed to user terminals of multiple recommended interest circles and extended interest circles through the network.

2. The automobile product matching method based on a large language model according to claim 1, characterized in that: Build dynamic portraits of each interest circle in multiple interest circles, including: Obtain user information of multiple users, identify interest circles to which the users belong based on the user information, and static portraits of the interest circles; Obtain hot topic data in interest circles; According to the static portrait of the interest circle and the hot topic data, a dynamic portrait of the interest circle is obtained.

3. The automobile product matching method based on a large language model according to claim 2 is characterized in that: Identify the interest circle to which the user belongs according to the user information, and the static portrait of the interest circle includes: Using the user information to construct a prompt template, inputting the prompt template into a portrait modeling model, and obtaining a static portrait output by the portrait modeling model; According to the static portrait of the interest circle and the hot topic data, a dynamic portrait of the interest circle is obtained, including: The static portrait of the interest circle and the hot topic data are used to construct a prompt template, and the prompt template is input into the portrait modeling model to obtain a dynamic portrait output by the portrait modeling model.

4. The automobile product matching method based on a large language model according to claim 1, characterized in that: Constructing target selling point features according to the semantics of the selling point information, including: Performing syntax parsing and semantic analysis on the selling point information to obtain a selling point-description tuple; Using the automobile knowledge base to perform proper noun detection and interpretation on the bigrams to form selling point-description-explanation triplets; A correlation analysis is performed on the triples and automobile brand marketing to form a selling point-description-explanation-correlation quadruple, and the quadruple with a correlation greater than a set threshold is retained as the target selling point feature.

5. The automobile product matching method based on a large language model according to claim 1, characterized in that: According to the target selling point features, a second training sample is selected from the data set, including: Calculate the semantic similarity between the target selling point features and the selling point features in the database, take the selling point feature samples with semantic similarity greater than the high threshold and the corresponding dynamic portrait labels as positive samples, and take the selling point feature samples with semantic similarity less than the low threshold and the corresponding dynamic portrait labels as negative samples.

6. The automobile product matching method based on a large language model according to claim 1, characterized in that: The target selling point feature and the second training sample are input into the large language model, the large language model that has been preliminarily trained is retrained using the second training sample, and multiple recommended interest circles output by the large language model are obtained based on the target selling point feature, including: Fill the target selling point feature into the question part of the prompt template, fill the second training sample into the example part of the prompt template, and fill the recommended interest circle into the answer part of the prompt template; The filled prompt template is input into the large language model so that the large language model is retrained with the second training sample, and a plurality of recommended interest circles output by the large language model are obtained based on the target selling point features.

7. The automobile product matching method based on a large language model according to claim 1, characterized in that: Based on the historical matching records of multiple interest circles and selling point features, the distances between different interest circles are obtained, including: Merge semantically similar selling point features; Count the number of times different interest circles recommend something together, and the similarity of static portraits of different interest circles; The distances between different interest circles are obtained according to the number and similarity.

8. The automobile product matching method based on a large language model according to claim 7 is characterized in that: Mapping the multiple recommended interest circles to a circle association graph, and selecting an extended interest circle within a set range around the multiple recommended interest circles, including: Determine the nodes consistent with each recommended interest circle in the circle association graph; Taking any node as the center, traverse the nodes within the surrounding set range as the expanded interest circle.

9. The automobile product matching method based on a large language model according to claim 8, characterized in that: After traversing the nodes within the surrounding set range as the extended interest circle, it also includes: According to the historical push records to the user terminal, weighted sampling is performed on the extended interest circle and the recommended interest circle.

10. An automobile product matching system based on a large language model, characterized in that: include: A preprocessing module, configured to perform preliminary training on the large language model using a first training sample, wherein the first training sample includes a selling point information sample and an interest circle information label to which the user belongs; The first training sample is obtained by matching the selling point information sample with the semantic similarity of the interest circle information label; constructing a dynamic portrait of each interest circle in the multiple interest circles; An input module is used to receive the selling point information of the target vehicle model input by the car manufacturer, and construct the target selling point features according to the semantics of the selling point information; A recommendation module is used to filter out a second training sample from a data set according to the target selling point feature, including a selling point feature sample and a dynamic portrait label; the data set is constructed according to the user's purchasing behavior for multiple car models; the target selling point feature and the second training sample are input into a large language model, the large language model that has been preliminarily trained is retrained using the second training sample, and multiple recommended interest circles output by the large language model are obtained based on the target selling point feature; the distances between different interest circles are obtained according to the historical matching records of multiple interest circles and selling point features; a circle association graph of each interest circle is constructed according to the distance; the multiple recommended interest circles are mapped into the circle association graph, and an extended interest circle within a set range around the multiple recommended interest circles is selected; The output module is used to push the target vehicle model information to user terminals of multiple recommended interest circles and extended interest circles through the network.

Images