Unified construction method for intention recognition system and semantic vector recall system

Abstract

The invention discloses a unified construction method for an intention recognition system and a semantic vector recall system, relates to the field of computer processing, and aims at uniformly constructing the intention recognition and semantic vector recall system by aligning a user vector, a material implicit vector and an intention implicit vector in a natural semantic space. Besides, the material implicit vector is projected to the user vector in a vector projection mode, and the similarity between the material implicit vector and the user vector is calculated, so that long semantic information of visual content and short semantic information of user search terms are subjected to consistent alignment. Therefore, a plurality of functional modules on a whole search link are replaced, the research and development and operation cost is greatly reduced, and a search system is obviously improved in the aspects of user intention estimation and integrity and accuracy of preference material recall.

Description

Technical Field

[0001] The present invention relates to the field of computer processing, and in particular to a unified construction method of an intention recognition system and a semantic vector recall system. Background Art

[0002] In current search systems, most personalized search vector recall and intent recognition systems are independently built based on large language models or multimodal large models, using large-scale data and complex algorithms to understand and predict users' search intentions and preferences. First, user vectors are generated based on the user's historical behavior (such as search, click, conversion, etc.) and real-time interaction (such as input search terms, filter conditions, etc.) on the search system; then, two systems are independently built for the two application scenarios of intent recognition and vector recall, namely the intent recognition system and the semantic vector recall system.

[0003] First, the intent recognition system searches for the predefined user intent types and ranges in the scenario and builds an intent library. Then, based on the user vector, it predicts the probability values ​​in the intent library and obtains the user search intent through topk or probability thresholds. The semantic vector recall system is based on the user vector and the material latent vector for the materials to be distributed by the search system, and performs similarity matching around relevance and conversion goals. In the above scheme, the independent construction, development, and operation of the two systems are not only costly and time-consuming, but also affect the search system's estimation of user intent and the completeness and accuracy of recalling preferred materials. Summary of the invention

[0004] The purpose of the present invention is to provide a unified construction method of an intention recognition system and a semantic vector recall system, which aims to overcome the problems existing in the prior art.

[0005] To achieve the purpose, the present invention provides the following technical solutions: A unified construction method for an intent recognition system and a semantic vector recall system, generating a first text vector through a text component of a large picture-text multimodal model for interactive text information input by a user into a search system; generating a first visual vector through a visual component of a large picture-text multimodal model for visual marketing content clicked or / and downloaded by a user; generating a user vector representing the user's search intent from the first text vector and the first visual vector by using an attention mechanism technology around the real-time interaction between the user and the search system; and further comprising the following steps: Generate latent intent vectors for the text-based intent labeling system through the text component of the large multimodal model. For the text information of the visual marketing content, a second text vector is generated through the text component of the graphic-text multimodal model; For the visual images of visual marketing content, a second visual vector is generated through the visual components of the graphic and text multimodal model; By using a learnable alpha parameter, the second text vector and the second visual vector are fused as a material latent vector; The user vector, intent latent vector and material latent vector are aligned in the natural semantic space to uniformly construct the intent recognition system and the semantic vector recall system.

[0006] Furthermore, the user vector, the intention latent vector and the material latent vector are aligned in the natural semantic space. Specifically, the user vector is used as the alignment target, and a two-layer MLP is used to bridge the intention latent vector and the material latent vector to the user vector space.

[0007] Furthermore, before aligning the user vector, the intention latent vector and the material latent vector, the following steps are also included: projecting the material latent vector to the user vector by means of vector projection, and calculating the similarity between the material latent vector and the user vector.

[0008] Furthermore, the data generated by users in the entire search system are collected, including the interactive text information and the visual marketing content, to provide a data basis for the multimodal large model of pictures and texts.

[0009] Furthermore, the text information of the visual marketing content includes categories, tags and pictures.

[0010] These data come from the data generated in the entire link of user search, including search information flow, screening information flow and related recommendation information flow, which can reflect user behaviors such as search, browsing, click and download.

[0011] Furthermore, for intent labels that are not within the scope of the intent label system, inference prediction is performed through zero-shot learning.

[0012] Compared with the prior art, the present invention has the following beneficial effects: First, the present invention aligns the user vector, material latent vector and intent latent vector in the natural semantic space, that is, multi-modal alignment of user behavior data (ID mode), intent recognition (natural language mode) and visual marketing content (image mode), thereby uniformly constructing an intent recognition and semantic vector recall system. It can be seen that compared with the traditional solution, the present invention replaces multiple functional modules in the entire search link, greatly reducing R&D and operating costs; and has a significant improvement in the completeness and accuracy of user intent estimation and preferred material recall.

[0013] Secondly, in the present invention, the material latent vector is first projected to the user vector by means of vector projection, and the similarity between the material latent vector and the user vector is calculated, so that the long semantic information of the visual content is aligned with the short semantic information of the user's search term, thereby further improving the system's completeness and accuracy in estimating user intent and recalling preferred materials.

[0014] Third, in addition to this, since the user vector, material latent vector and intent latent vector are aligned in the same natural language space, the unified construction of the intent recognition system and the semantic vector recall system can perform inference and prediction in a zero-shot manner for intentions that are not within the predefined intent type and scope, further expanding the application scope of intent recognition and further reducing R&D costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0016] The specific implementation of the present invention is described below with reference to the accompanying drawings. In order to fully understand the present invention, many details are described below, but for those skilled in the art, the present invention can be implemented without these details.

[0017] like Figure 1 As shown in the figure, a unified construction method of an intent recognition system and a semantic vector recall system includes the following aspects: (1) Data collection Collect data generated by user search, browsing, clicking and downloading behaviors in the entire search system. These data come from search information flow, screening information flow and related recommendation information flow, etc., providing a data foundation for the multimodal large model of pictures and texts.

[0018] This data includes, but is not limited to, interactive text information that users enter into the search system, and visual marketing content that users click on and / or download.

[0019] Furthermore, the data of visual marketing content is divided into text information and visual images. The text information of visual marketing content includes but is not limited to titles, tags, categories, etc.

[0020] (2) Data processing and vector generation (2.1) For interactive text information input by a user into the search system, a first text vector is generated through the text component of the graphic-text multimodal model; for visual marketing content clicked or / and downloaded by a user, a first visual vector is generated through the visual component of the graphic-text multimodal model; (2.2) Generate a user vector representing the user's search intent from the first text vector and the first visual vector based on the real-time interaction between the user and the search system through the attention mechanism technology; the attention mechanism technology includes but is not limited to Target Attention, Zero-Attention and other technologies (2.3) Generate latent intent vectors for the text-based intent labeling system through the text component of the large multimodal model. (2.4) For the text information of the visual marketing content, a second text vector is generated through the text component of the image-text multimodal model; for the visual image of the visual marketing content, a second visual vector is generated through the visual component of the image-text multimodal model; (2.5) fusing the second text vector and the second visual vector as a material latent vector through a learnable alpha parameter; (3) Multimodal alignment The user vector of step (2.2), the intention latent vector of step (2.3), and the material latent vector of step (2.5) are aligned in the natural semantic space, so as to uniformly construct the intention recognition system and the semantic vector recall system.

[0021] In a specific embodiment, the following method is specifically used to align the user vector, the intention latent vector and the material latent vector: taking the user vector as the alignment target, a two-layer MLP is used to bridge the intention latent vector and the material latent vector to the user vector space.

[0022] In the step (3), before aligning the user vector, the intention latent vector and the material latent vector, the consistency alignment of the long and short semantic information is also included, specifically: projecting the material latent vector to the user vector by means of vector projection, and calculating the similarity between the material latent vector and the user vector.

[0023] To further illustrate the consistent alignment of long and short semantic information: user search terms are usually short, and visual marketing content will have different degrees of semantic information depending on the richness of its visual elements.

[0024] For example, the user search term is "blue K12 poster"; The semantic information of visual content A is “blue, K12 education, course promotion, poster”; The semantic information of visual content B is “blue poster”.

[0025] The material latent vector of visual content A contains all the requirements of the user's search term, but exceeds it. The material latent vector of visual content B contains part of the user's search term requirements. From the perspective of vector similarity calculation (such as vector cosine and Euclidean distance), the similarity score between visual content B and the user's search term is higher.

[0026] To solve this problem, the present invention projects the material latent vectors of visual content A and visual content B to the user vector by means of vector projection, and calculates their similarity with the user vector, so as to achieve the purpose of selecting visual content B. That is, both visual contents A and B are subjected to vector similarity calculation on the modulus scale of the user vector.

[0027] In addition, since the user vector, material latent vector and intent latent vector are aligned in the same natural language space, the uniformly constructed graph recognition system and semantic vector recall system can perform inference and prediction through zero-shot learning for intent labels that are not within the scope of the intent label system, and identify user intent.

[0028] It is only a specific implementation method of the present invention, but the design concept of the present invention is not limited to this. Any non-substantial changes to the present invention using this concept shall be deemed as an infringement of the protection scope of the present invention.

Claims

1. A unified construction method of an intention recognition system and a semantic vector recall system, generating a first text vector through a text component of a large multimodal model of images and texts for interactive text information input by a user into a search system; For visual marketing content clicked or / and downloaded by users, a first visual vector is generated through the visual component of the text-image multimodal model; a user vector representing the user's search intention is generated from the first text vector and the first visual vector through the attention mechanism technology around the real-time interaction between the user and the search system, which is characterized by: The following steps are also included: Generate latent intent vectors for the text-based intent labeling system through the text component of the large multimodal model. For the text information of the visual marketing content, a second text vector is generated through the text component of the large picture-text multimodal model; for the visual image of the visual marketing content, a second visual vector is generated through the visual component of the large picture-text multimodal model; through the learnable alpha parameter, the second text vector and the second visual vector are fused as the material latent vector; The user vector, intent latent vector and material latent vector are aligned in the natural semantic space to uniformly construct the intent recognition system and the semantic vector recall system.

2. The unified construction method of the intention recognition system and the semantic vector recall system according to claim 1, characterized in that: The user vector, the intention latent vector and the material latent vector are aligned in the natural semantic space. Specifically, the user vector is used as the alignment target, and the intention latent vector and the material latent vector are bridged to the user vector space through a two-layer MLP.

3. The unified construction method of the intention recognition system and the semantic vector recall system according to claim 1, characterized in that: Before aligning the user vector, the intention latent vector and the material latent vector, the following steps are also included: projecting the material latent vector to the user vector by means of vector projection, and calculating the similarity between the material latent vector and the user vector.

4. The unified construction method of the intention recognition system and the semantic vector recall system according to claim 1, characterized in that: Collect the data generated by users in the entire link of the search system, including the interactive text information and the visual marketing content, to provide a data foundation for the multimodal large model of pictures and texts.

5. The unified construction method of the intention recognition system and the semantic vector recall system according to claim 4 is characterized by: The text information of the visual marketing content includes categories, tags and pictures.

6. According to the unified construction method of the intent recognition system and the semantic vector recall system described in claim 1, the intent labels that are not within the scope of the intent label system are inferred and predicted through zero-shot learning.

Images