Search

CN-122020248-A - Label recommending method and vehicle

CN122020248ACN 122020248 ACN122020248 ACN 122020248ACN-122020248-A

Abstract

The application relates to the field of data classification, and particularly discloses a tag recommendation method and a vehicle, wherein the method comprises the steps of extracting multidimensional data features of data to be marked, wherein the data features comprise text features, image features and structural data features, carrying out feature fusion on the multidimensional data features to obtain fusion feature vectors, determining recommendation scores of the data to be marked and all tags based on a pre-built tag association model and the fusion feature vectors, and determining target tags corresponding to the data to be marked based on the recommendation scores. The method can comprehensively capture the content and the structural information of the data by extracting various modal characteristics including texts, images and structured data, and can generate unified fusion characteristic vectors by carrying out fusion processing on the multidimensional characteristics, so that the effective integration and complementation of the multisource information are realized. And calculating a recommendation score based on the pre-constructed label association model and the fusion feature vector, so that accurate inference from data feature to label mapping is realized.

Inventors

  • ZHU HE

Assignees

  • 长城汽车股份有限公司

Dates

Publication Date
20260512
Application Date
20260128

Claims (10)

  1. 1. A tag recommendation method, comprising: extracting multidimensional data features of data to be marked, wherein the multidimensional data features comprise text features, image features and structured data features; feature fusion is carried out on the multidimensional data features so as to obtain fusion feature vectors; Determining the recommendation score of the data to be marked and each label based on a pre-constructed label association model and the fusion feature vector; and determining a target label corresponding to the data to be marked based on the recommendation score.
  2. 2. The method according to claim 1, wherein the extracting the multidimensional data feature of the data to be marked comprises: Keyword extraction is carried out on text content based on word frequency statistics and weight of inverse document frequency, so that first text characteristics are obtained; mapping the text content to a continuous vector space to obtain a context semantic representation to obtain a second text feature; Extracting bottom visual characteristics of an image, and obtaining high-level semantic characteristics by pooling and weighting output characteristics of a pre-training model, wherein the bottom visual characteristics comprise at least one of color characteristics, texture characteristics and shape characteristics; encoding the structured fields and extracting statistical features among the fields to obtain structured data features; the field codes comprise at least one of one-hot codes and field name embedded codes, and the statistical features comprise at least one of mean, variance, number of unique values and correlation coefficients among fields.
  3. 3. The method of claim 1, wherein feature fusing the multi-dimensional data features to obtain a fused feature vector comprises: carrying out weighted fusion on the multidimensional data features to obtain weighted feature vectors; and/or splicing the feature vectors of different modes to obtain spliced feature vectors; and/or, calculating interaction feature vectors among the different modality feature vectors; and generating the fusion feature vector based on at least one of the weighted feature vector, the spliced feature vector and the interaction feature vector.
  4. 4. The method of claim 1, wherein prior to determining the recommendation score for the data to be labeled and each label based on the pre-constructed label association model and the fused feature vector, the method further comprises: Based on the history marking record, constructing multi-dimensional interaction data, wherein the multi-dimensional interaction data is used for representing interaction intensity among users, data and labels; Decomposing the multi-dimensional interaction data into a plurality of low-dimensional factor matrices, wherein the low-dimensional factor matrices at least comprise a user factor matrix, a data factor matrix and a label factor matrix; and constructing a tag association model based on the plurality of low-dimensional factor matrices.
  5. 5. The method according to claim 4, wherein determining the recommendation score of the data to be marked and each tag based on the pre-constructed tag association model and the fusion feature vector specifically comprises: constructing a graph structure comprising data nodes, label nodes and user nodes, wherein edges in the graph structure are used for representing interaction relations among the nodes; Based on the label association model, performing representation learning on nodes in the graph structure; The fusion feature vector is used as the representation of the data to be marked in the potential space corresponding to the label association model; And determining the recommendation score of the data to be marked and each label based on the fusion feature vector and the learned graph structure.
  6. 6. The method according to claim 5, wherein determining the recommendation score of the data to be labeled and each label based on the fused feature vector and the learned graph structure specifically includes: inputting the fusion feature vector into a data node representation in the graph structure; Message transmission and node representation updating are carried out through a multi-layer graph convolution network, and updated data node representation and label node representation are obtained; Calculating cosine similarity between the data node representation and each tag node representation; and converting the cosine similarity into a probability score through a preset function, and taking the probability score as the recommendation score.
  7. 7. The method of claim 5, wherein after determining the target tag corresponding to the data to be marked based on the recommendation score, the method further comprises: responding to the received confirmation or correction operation of the user to the target label, taking the user operation as a new historical marking record, and updating the new historical marking record into the multidimensional interactive data; Based on the updated multidimensional interaction data, incremental updating is carried out on the user factor matrix, the data factor matrix and the label factor matrix; and updating the node representation and the parameters of the graph structure based on the updated factor matrix.
  8. 8. The method of claim 5, wherein the determining the target tag corresponding to the data to be marked based on the recommendation score comprises: sorting the labels in descending order according to the recommended score; and selecting a preset positive integer number of labels from the sorting result to be output as the target labels.
  9. 9. The method of claim 5, wherein the determining the target tag corresponding to the data to be marked based on the recommendation score comprises: Prompting a user to input a new label in response to the recommendation scores of all labels being lower than a preset threshold; Adding the new label input by the user to a system label set; and updating the interaction record of the new label and the data to be marked into the multidimensional interaction data as newly-added data.
  10. 10. A vehicle, characterized by comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores instructions executable by the at least one processor to enable the at least one processor to perform the tag recommendation method of any one of claims 1 to 8.

Description

Label recommending method and vehicle Technical Field The application relates to the field of data classification, in particular to a tag recommendation method and a vehicle. Background In the current data management, content retrieval and knowledge organization fields, the tagging of data objects (such as documents, images, structured data sets, etc.) is a widely applied technical means, aiming at classifying, describing and indexing data by semantic keywords, thereby improving discoverability, manageability and subsequent utilization value of the data. Traditional tagging approaches rely primarily on manual operations, often requiring users to manually select and associate one or more tags from a predefined set of tags to target data by their own understanding of the data content. With the rapid growth of data scale and the increasing diversity of data types (such as text, image and table coexistence), the manually-guided mode gradually exposes the problems of low efficiency, difficult guarantee of label quality, insufficient utilization of a label system and the like. Disclosure of Invention In view of the above problems, the present disclosure provides a tag recommendation method and a vehicle for overcoming the above problems or at least partially solving the above problems, where the technical solution is as follows: A tag recommendation method is characterized by comprising the steps of extracting multi-dimensional data features of data to be marked, wherein the multi-dimensional data features comprise text features, image features and structured data features, carrying out feature fusion on the multi-dimensional data features to obtain fusion feature vectors, determining recommendation scores of the data to be marked and each tag based on a pre-built tag association model and the fusion feature vectors, and determining target tags corresponding to the data to be marked based on the recommendation scores. By integrating the modal characteristics in the text, the image and the structured data, the recommendation deviation caused by incomplete single modal information is avoided. By fusing multi-modal features, a more comprehensive data description can be generated. By utilizing the pre-trained label association model, complex real-time calculation is not needed for each recommendation, and the recommendation score can be directly output on the basis of the fusion characteristic fast matching history mode. According to the method, traditional manual feature matching and rule judgment are converted into automatic model reasoning based on unified feature vectors, so that the operation steps and judgment time of manual marking are directly reduced, the marking efficiency is improved, and the consistency of recommendation logics of different data types is ensured. In one example, the method for extracting multidimensional data features of data to be marked specifically comprises the steps of extracting keywords from text content based on word frequency statistics and weights of inverse document frequency to obtain first text features, mapping the text content to a continuous vector space to obtain context semantic representations to obtain second text features, combining the first text features and the second text features to obtain text features, extracting bottom visual features of an image, and obtaining high-level semantic features by pooling and weighting output features of a pre-training model, wherein the bottom visual features comprise at least one of color features, texture features and shape features, combining the bottom visual features with the high-level semantic features to obtain image features, encoding structured fields and extracting statistical features among fields to obtain structured data features, field encoding comprises at least one of single-hot encoding and field name embedding encoding, and the statistical features comprise at least one of mean value, variance, unique value quantity and correlation coefficient among fields. On the text characteristics, the key word characteristics and the semantic characteristics are combined, so that the system can capture the dominant key words in the document and understand the deep semantic of the dominant key words, and the problems of professional terms and synonyms can be solved. For the image, the bottom layer features such as color, texture and the like and the high-level semantic features are extracted at the same time, so that the system can recognize the style of the image and also recognize specific objects in the image. For the structured data, through field coding and statistical feature extraction, tables, attribute lists and the like can be effectively processed, and the correlation among fields can be identified. The refined feature extraction provides high-quality and information-complementary raw materials for subsequent fusion, and directly improves the characterization precision of feature vectors. In one example, the featu