CN-122022915-A - Method for generating second-hand car text

CN122022915ACN 122022915 ACN122022915 ACN 122022915ACN-122022915-A

Abstract

The application discloses a second-hand vehicle document generation method, which comprises the steps of obtaining vehicle structural information, characteristic tag information and a vehicle source image set after obtaining a vehicle source identifier, determining key images of description parts such as appearance, interior decoration and power from the image set, constructing part prompt information which comprises task constraint, content constraint and length constraint and is injected with structural information for each description part, calling vision understanding based on the key images and the part prompt information to generate a service to generate a part document and cleaning, fusing the structural information, the characteristic tag information and the part document to generate a summary document, and outputting a structured document object comprising summary fields and part fields. The method has the advantages that the content sources of the text are limited, the constrained generation and the structured output of the parts are realized, the consistency and the controllability of the graphics and texts are improved, the assumption and the quality fluctuation are reduced, and the method is suitable for batch generation.

Inventors

WANG CHEN
Pei Xugeng
LIN QUAN

Assignees

北京好车帅帅科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (9)

1. A method for generating a second hand case, the method comprising: acquiring a vehicle source identifier of a document to be generated; Acquiring structural information and characteristic tag information of a vehicle based on the vehicle source identifier, and acquiring a vehicle source image set corresponding to the vehicle; Determining key images corresponding to different description positions of a vehicle respectively from the vehicle source image set, wherein the different description positions at least comprise an appearance position, an interior position and a power position; Respectively constructing part prompt information aiming at the different description parts, wherein the part prompt information at least comprises task constraint, content constraint and output length constraint, and at least one part of the structural information is injected into the part prompt information; based on the key images and the corresponding position prompt information, calling a visual understanding generating service to respectively generate position documents of the description positions, and cleaning the contents of the position documents; Fusing the structured information, the characteristic tag information and each part document to generate a summary document; And outputting a structured document object, wherein the structured document object at least comprises a summary field and a part field corresponding to the different description parts, the summary field is used for representing the summary document, and the part field is used for representing the corresponding part document.
2. The method of claim 1, wherein the structured information comprises at least one of vehicle model information, registration or placard time, mileage, body color, displacement, or fuel type of a vehicle, and comprises a list of vehicle configuration items or a set of vehicle attribute parameters.
3. The method of claim 1, wherein determining the key image from the set of source images comprises: based on a preset image type identifier or shooting angle identifier, respectively selecting corresponding key images for different description positions in the vehicle source image set; and carrying out parameter adjustment on the image access address of the key image to generate an image address for calling the visual understanding generation service.
4. The method of claim 2, wherein the constructing of the location hint information comprises: Different prompting templates are adopted for different description positions respectively; Setting an output length upper limit threshold in the prompt template; Setting content constraints in the prompt template, wherein the content constraints at least comprise constraints generated based on corresponding key image content and prohibiting the introduction of configuration descriptions not contained in the structural information and the characteristic label information; And setting example guide texts in the prompt templates, and selecting corresponding prompt templates according to the fuel type and/or the power type aiming at the power part.
5. The method of claim 1, wherein invoking the visual understanding generating service to generate the part documents for each of the description parts is concurrent invoking; the content cleansing includes at least removing emoji characters and/or topic label characters.
6. The method of claim 1, wherein invoking the visual understanding generating service comprises a failed retry mechanism comprising performing a retry at a preset number of retries and a preset retry interval when invoking a failure; and under the condition that the retry still fails, switching to calling a text generation service for at least one description part to generate a corresponding part text, and returning empty contents for the rest description parts and recording error information.
7. The method of claim 1, wherein generating the summary document comprises: Constructing at least vehicle registration or branding time and driving mileage, each part document and the characteristic tag information in the structured information into summary input, wherein the characteristic tag information is formatted into a text form of tag titles and tag contents; Building summary prompt information based on the summary input and calling a text generation service to generate the summary document; The summary prompt information comprises constraints of outputting length, expressing kiss, and prohibiting outputting emoticons and/or topic labels; and setting a temperature parameter when the text generation service is called.
8. The method of claim 1, wherein fusing the structured information, the feature tag information, and each of the portion documents comprises a semantic level fusion process; the semantic level fusion process includes keyword extraction of the fused text based on word frequency and inverse document frequency TF-IDF to generate a set of selling point vocabulary entries for generating the summary document, and incorporating the set of selling point vocabulary entries into the summary input.
9. The method according to claim 1, wherein the method further comprises: performing a field integrity check on the structured document object, the field integrity check including at least detecting that any one of the summary field and the location field is empty and marking a generation failure; Generating logs for at least one of structured information acquisition, key image determination, visual understanding generation service call, text generation service call and field integrity check; And storing the position prompt information, the summary prompt information and the corresponding generation result into a database.

Description

Method for generating second-hand car text Technical Field The application relates to the technical field of second-hand car text generation, in particular to a second-hand car text generation method. Background With the development of internet second-hand vehicle transaction platforms, automobile electronic commerce and vehicle source management systems, vehicle documents (such as vehicle outlines, appearance descriptions, interior descriptions, power and engine descriptions, etc.) in vehicle source display pages have become important factors affecting user stay time, consultation conversion rate and transaction efficiency. Particularly, under the business scenes of rapid increase of the number of vehicle sources, high new frequency and diversified display channels, how to quickly and stably produce the second-hand vehicle text which has real content, special expression and differentiated selling points becomes a technical direction of general attention of the industry. In the prior art, the generation of the vehicle text mainly comprises modes of manual editing and writing, automatic filling based on a fixed template, automatic writing based on a general text generation model and the like. The method has the advantages that the method is good in manual writing, the professional experience of editors and the time cost are greatly dependent, the method is difficult to adapt to the large-scale and high-frequency updating requirements of vehicle sources, the template filling mode is usually spliced only according to structural data such as vehicle types, mileage, license time, configuration tables and the like, the mechanical and homogeneous expression forms are obvious, the real information such as appearance flaws, interior trim states and detail textures of the vehicles are difficult to cover, the general text generation model can improve the language fluency, but the reliable understanding and constraint on the contents of real images of the vehicles are often lacking, and the problems such as inconsistent picture and text, speculative configuration or state and the like are easy to occur. In addition, the difference of the image quality, the shooting angle and the information integrity of different vehicle sources is larger, quality fluctuation, insufficient stability and lack of uniform field output and an effective content verification mechanism are easy to occur when the conventional scheme is generated in batches, and the generated result is difficult to be directly used for online display. Therefore, in the automatic generation of a second-hand vehicle source document, the consistency of a vehicle source image and the document content is difficult to ensure, the expression of the real state and the detail selling points of the vehicle is insufficient, and the stability and the controllability of batch generation are poor, so that the problem to be solved is needed. Disclosure of Invention The application provides a method for generating a second-hand vehicle document, which aims to solve the problems that in the automatic generation of the second-hand vehicle source document, the consistency of a vehicle source image and the document content is difficult to ensure, the expression of the real state and the detail selling point of a vehicle is insufficient, and the stability and the controllability of batch generation are poor in the prior art. A second hand case generation method, the method comprising: acquiring a vehicle source identifier of a document to be generated; Acquiring structural information and characteristic tag information of a vehicle based on the vehicle source identifier, and acquiring a vehicle source image set corresponding to the vehicle; Determining key images corresponding to different description positions of a vehicle respectively from the vehicle source image set, wherein the different description positions at least comprise an appearance position, an interior position and a power position; Respectively constructing part prompt information aiming at the different description parts, wherein the part prompt information at least comprises task constraint, content constraint and output length constraint, and at least one part of the structural information is injected into the part prompt information; based on the key images and the corresponding position prompt information, calling a visual understanding generating service to respectively generate position documents of the description positions, and cleaning the contents of the position documents; Fusing the structured information, the characteristic tag information and each part document to generate a summary document; And outputting a structured document object, wherein the structured document object at least comprises a summary field and a part field corresponding to the different description parts, the summary field is used for representing the summary document, and the part field is used for representing the corresponding p