Core Product Term Identification Using Siamese Network Image Similarity

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In e-commerce, identifying the core product term in a commodity title is challenging due to unclear semantics and stacked product terms, with existing methods relying solely on textual information, which limits their ability to generalize and effectively parse syntactic information, especially with non-standard writing forms.

Innovation Solution

A method and system that incorporate image information by using a Siamese network to determine the similarity between a commodity's display image and candidate product terms, overcoming the limitations of textual-based approaches by leveraging image features and clustering algorithms to identify the core product term.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If vocabulary-based rule method is used, then processing speed is fast, but generalization ability is poor and cannot handle new samples

Engineering Contradiction:

Improveprocessing speedVSAvoidgeneralization ability

Core Design Contradiction:

ProductivityVSAdaptability or versatility

Solution Approach 1:

The patent replaces the mechanical vocabulary-matching system with a deep learning neural network system. The neural network automatically learns semantic representations from data, substituting the rigid rule-based mechanical matching with an adaptive intelligent system that generalizes to new products and writing styles.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Solution Approach 2:

The patent transforms the identification approach from discrete vocabulary matching to continuous semantic space representation. By encoding products and titles into vector representations, the system can measure semantic similarity through parameter comparison rather than exact string matching, enabling generalization to new samples.

Inventive Principle:

Principle #35Parameter changes

2Adaptability or versatility

If machine learning method is used, then generalization ability improves, but effectiveness in identifying core product term decreases due to severe product term stacking

Engineering Contradiction:

Improvegeneralization abilityVSAvoidcore product term identification accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent segments the title into multiple candidate product terms and generates separate embeddings for each. This segmentation allows the system to evaluate each term's contribution independently, then aggregate their representations to identify the core product term despite stacking and unclear semantics.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces product embeddings as an intermediary representation between the raw title text and the core product term identification. These embeddings capture semantic meaning and serve as a bridge, allowing the system to resolve ambiguities caused by product term stacking through semantic similarity comparison.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Device complexity

If only textual information is used, then system complexity is low, but identification accuracy decreases due to non-standard writing forms

Engineering Contradiction:

Improvesystem complexityVSAvoididentification accuracy

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

The patent adds the embedding space dimension to the traditional text processing approach. By mapping textual information into a multi-dimensional semantic space, the system can capture nuances and relationships that are not apparent in raw text, improving accuracy for non-standard writing forms while maintaining reasonable system complexity through efficient neural network architectures.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Data Source

PatentUS11741094B2Method and system for identifying core product terms

Publication Date: 2023.08.29 BEIJING JINGDONG SHANGKE INFORMATION TECH CO LTD

AI summary

The invention, which discloses a method and a system for identifying a core product term, relates to the technical field of computers. A specific implementation mode of the method comprises: acquiring a display image of a commodity, determining a plurality of candidate product terms included in the title of the commodity, and determining a plurality of product image sets, wherein the respective ones of the plurality of product image sets are in one-to-one correspondence with the respective ones of the plurality of candidate product terms; for each of the plurality of product image sets, determining, according to the display image and respective images in the product image set, a similarity between the commodity in the display image and the candidate product term corresponding to the product image set to thereby obtain a plurality of similarities; and determining the candidate product term corresponding to a similarity greater than a predetermined threshold among the plurality of similarities as the core product term. The implementation mode incorporates image information into identification of the core product term so that the core product term in the title of the commodity is determined more intuitively and accurately.