Recommendation Embedding Model Using Semantic Text Alignment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional recommendation models struggle to model semantic information and utilize external semantic knowledge, leading to poor recommendation precision due to reliance solely on collaborative signals.

Innovation Solution

A data processing method that combines collaborative information with semantic information by using an embedding network and natural language processing model to generate and update feature representations, incorporating semantic features and improving the recommendation model's predictive capabilities.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a conventional recommendation model uses only collaborative information from the dataset, then the model architecture remains simple and online inference latency is short, but the recommendation precision is poor because semantic information cannot be modeled

Engineering Contradiction:
Improverecommendation precisionVSAvoidmodel complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent merges collaborative filtering information with semantic information from language models into a unified recommendation model. The embedding network integrates both types of information through shared layers and joint training, allowing the model to leverage external semantic knowledge while maintaining collaborative filtering capabilities, thereby improving recommendation precision without excessive complexity increase

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The embedding network is designed with multi-functionality to handle both collaborative filtering tasks and semantic processing tasks. The same network structure processes both types of information, making the model more versatile and reducing overall system complexity compared to using separate models for each information type

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Measurement precision

If a language model is used to process semantic information, then external semantic knowledge can be obtained and recommendation precision improves, but online inference latency increases

Engineering Contradiction:
Improverecommendation precisionVSAvoidonline inference latency
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary action by pre-training the embedding network with semantic information using offline datasets before deployment. During online inference, the pre-trained model directly processes new inputs without requiring real-time semantic processing, thereby reducing online latency while maintaining improved recommendation precision from the semantic training

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The model processing is segmented into offline training phase and online inference phase. Semantic information processing is performed during offline training to update the embedding network, while online inference uses the trained model for fast prediction, separating the computationally intensive semantic processing from real-time requirements

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4700608A1Data processing method and related apparatus
Publication Date: 2026.02.25 HUAWEI TECH CO LTD
  • EP4700608A1 patent drawingFigure 1
  • EP4700608A1 patent drawingFigure 2
  • EP4700608A1 patent drawingFigure 3

AI summary

A data processing method is provided, which may be applied to the field of artificial intelligence. The method includes: obtaining first data, where the first data includes attribute information of a user or attribute information of an item in a plurality of dimensions; generating second data based on the first data, where the second data is a natural language text used to describe the attribute information; obtaining a first feature representation based on the first data through an embedding network, where the first feature representation includes a plurality of embedding vectors, and each embedding vector corresponds to one dimension; obtaining a second feature representation based on the second data through a natural language processing model, where the second feature representation includes a semantic feature of the second data; and updating the embedding network based on a difference between the second feature representation and the first feature representation, to obtain an updated embedding network. In this application, collaborative information and semantic information are combined to perform modeling, so that knowledge of a language model can be fully used, thereby improving prediction effect.