Long-Text Matching With Semantic Sentence-Bag Aggregation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing text matching technologies face challenges in maintaining accuracy, especially with long texts, leading to reduced matching precision due to redundancy and inefficiency in processing article-length texts.

Innovation Solution

Constructing sentence bags with similar semantics, fusing their semantic vectors based on weight values to generate an aggregation vector, and calculating a matching result using this vector to improve accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If text matching is performed on long texts using existing techniques, then the matching process can handle article-length texts, but the matching accuracy is significantly reduced due to redundancy and inefficiency

Engineering Contradiction:
Improvetext matching accuracyVSAvoidprocessing complexity for long texts
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments long texts into multiple sentence bags, where each sentence bag contains sentences with similar semantics. This segmentation approach breaks down the complex task of matching long texts into smaller, more manageable units, thereby maintaining matching accuracy while reducing processing complexity. The sentence bags are then processed independently and aggregated to produce the final matching result.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If existing text matching techniques are applied to long texts, then the texts can be processed, but redundant information causes reduced matching precision

Engineering Contradiction:
Improvematching precisionVSAvoidredundant information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent extracts and groups sentences with similar semantics into sentence bags, effectively separating meaningful information from redundant information. By clustering semantically similar sentences together and selecting representative sentences for each bag, the method removes redundant information while preserving the essential semantic content needed for accurate text matching.

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If long texts are processed directly without segmentation, then the complete text information is available, but processing efficiency and matching accuracy decrease

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidmatching accuracy
Core Design Contradiction:
ProductivityVSMeasurement precision

Solution Approach 1:

The patent divides long texts into sentence bags containing semantically similar sentences, enabling efficient parallel processing of multiple smaller units while maintaining the semantic relationships within each bag. This segmentation improves processing efficiency by reducing the computational burden on individual processing units while preserving matching accuracy through semantic clustering and aggregation of results.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent merges sentences with similar semantics into sentence bags and then aggregates the results from multiple sentence bags to produce the final matching decision. This merging approach preserves the semantic relationships and contextual information across the original text while enabling efficient processing through structured organization and result aggregation.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4617922A1Text matching method and apparatus, computer-readable storage medium, and terminal
Publication Date: 2025.09.17 UNIDT (SHANGHAI) CO LTD
  • EP4617922A1 patent drawingFigure 1
  • EP4617922A1 patent drawingFigure 2~4
  • EP4617922A1 patent drawing

AI summary

A text matching method and apparatus, a computer-readable storage medium, and a terminal. The method comprises: acquiring texts to be matched, the texts to be matched comprising a first text and a second text; constructing a plurality of bags of sentences, each bag of sentences comprising a plurality of sentences having similar semantics, the plurality of sentences being from the first text and/or the second text, and the semantics represented by different bags of sentences being different or non-similar; according to the weight value of each bag of sentences, fusing semantic vectors of the plurality of bags of sentences so as to obtain an aggregation vector, the weight value of each bag of sentences being used for representing the association degree between said bag of sentences and other bags of sentences; and according to the aggregation vector, calculating a matching result of the first text and the second text. The solution provided by the present application can improve text matching accuracy.