Semantic matching and retrieval method and apparatus and non-transitory computer-readable medium

Patent metadata

Field	Value
Publication number	US-12197479-B2
Application number	US-202318390496-A
Country	US
Kind code	B2
Filing date	Dec 20, 2023
Priority date	Dec 30, 2022
Publication date	Jan 14, 2025
Grant date	Jan 14, 2025

Abstract

Disclosed are a semantic matching and retrieval method and apparatus. The semantic matching and retrieval method includes steps of obtaining both the vector representation of a query text and the vector representation of a document text; obtaining the final vector representation of the query text; obtaining the final vector representation of the document text; calculating, based on the final vector representation of the query text and the final vector representation of the document text, the similarity score between the query text and the document text; and selecting, based on the similarity scores between the query text and a plurality of document texts, a document text matching the query text from the plurality of document texts.

First claim

What is claimed is:

1. A semantic matching and retrieval method comprising:
obtaining both a vector representation of a query text and a vector representation of a document text, wherein, the vector representation of the query text contains a vector representation of each character or word in the query text, and the vector representation of the document text contains a vector representation of each character or word in the document text;
extracting, based on the vector representation of each character or word in the query text, an original feature and a self-attention feature of the query text, extracting, based on the vector representation of each character or word in the query text as well as the vector representation of each character or word in the document text, an interactive attention feature of the query text, and fusing the original feature, the self-attention feature, and the interactive attention feature of the query text to obtain a final vector representation of the query text;
extracting, based on the vector representation of each character or word in the document text, an original feature and a self-attention feature of the document text, extracting, based on the vector representation of each character or word in the document text as well as the vector representation of each character or word in the query text, an interactive attention feature of the document text, and fusing the original feature, the self-attention feature, and the interactive attention feature of the document text to obtain a final vector representation of the document text;
calculating, based on the final vector representation of the query text as well as the final vector representation of the document text, a similarity score between the query text and the document text; and
selecting, based on the similarity scores between the query text and a plurality of document texts, a document text matching the query text from the plurality of document texts.

2. The semantic matching and retrieval method according to claim 1 , wherein,
the extraction of the original feature of the query text includes inputting the vector representation of each character or word in the query text into a first deep learning network to acquire the original feature of the query text output by the first deep learning network, and
the extraction of the original feature of the document text includes inputting the vector presentation of each character or word in the document text into a second deep learning network to acquire the original feature of the document text output by the second deep learning network.

3. The semantic matching and retrieval method according to claim 1 , wherein,
the extraction of the self-attention feature of the query text includes
calculating a self-attention weight of each character or word in the query text with respect to a target character or word in the query text;
performing, based on the self-attention weight of each character or word in the query text with respect to the target character or word in the query text, weighted summation on the vector representation of each character or word in the query text to acquire a self-attention feature of the target character or word in the query text; and
stitching the self-attention feature of each character or word in the query text to acquire the self-attention feature of the query text, and

the extraction of the self-attention feature of the document text includes
calculating a self-attention weight of each character or word in the document text with respect to a target character or word in the document text;
performing, based on the self-attention weight of each character or word in the document text with respect to the target character or word in the document text, weighted summation on the vector representation of each character or word in the document text to acquire a self-attention feature of the target character or word in the document text; and
stitching the self-attention feature of each character or word in the document text to acquire the self-attention feature of the document text.

4. The semantic matching and retrieval method according to claim 1 , wherein,
the extraction of the interactive attention feature of the query text includes
calculating an interactive attention weight of each character or word in the document text with respect to a target character or word in the query text;
performing, based on the interactive attention weight of each character or word in the document text with respect to the target character or word in the query text, weighted summation on the vector representation of each character or word in the document text to acquire an interactive attention feature of the target character or word in the query text; and
stitching the interactive attention feature of each character or word in the query text to acquire the interactive attention feature of the query text, and

the extraction of the interactive attention feature of the document text includes
calculating an interactive attention weight of each character or word in the query text with respect to a target character or word in the document text;
performing, based on the interactive attention weight of each character or word in the query text with respect to the target character or word in the document text, weighted summation on the vector representation of each character or word in the query text to acquire an interactive attention feature of the target character or word in the document text; and
stitching the interactive attention feature of each character or word in the document text to acquire the interactive attention feature of the document text.

5. The semantic matching and retrieval method according to claim 1 , wherein,
the fusion of the original feature, the self-attention feature, and the interactive attention feature of the query text includes performing feature addition, feature stitching, or weighted summation on the original feature, the self-attention feature, and the interactive attention feature of the query text to acquire the final vector representation of the query text, and
the fusion of the original feature, the self-attention feature, and the interactive attention feature of the document text includes performing feature addition, feature stitching, or weighted summation on the original feature, the self-attention feature, and the interactive attention feature of the document text to acquire the final vector representation of the document text.

6. The semantic matching and retrieval method according to claim 1 , wherein,
the calculation of the similarity score between the query text and the document text based on the final vector representation of the query text as well as the final vector representation of the document text includes calculating a cosine distance or a Manhattan distance between the final vector representation of the query text and the final vector representation of the document text to serve as the similarity score between the query text and the document text.

7. A semantic matching and retrieval apparatus comprising:
a first vector obtainment part configured to obtain a vector representation of a query text, wherein, the vector representation of the query text contains a vector representation of each character or word in the query text;
a second vector obtainment part configured to obtain a vector representation of a document text, wherein, the vector representation of the document text contains a vector representation of each character or word in the document text;
a first feature extraction part configured to extract, based on the vector representation of each character or word in the query text, an original feature and a self-attention feature of the query text, extract, based on the vect

Assignees

• Cheng Rui
• Dong Bin
• Jiang Shanshan
• Luo Lu
• Ding Lei
• Ricoh Co Ltd

Inventors

• Cheng Rui
• Dong Bin
• Jiang Shanshan
• Luo Lu
• Ding Lei

Classifications

• G06N3/045 Primary

  Combinations of networks · CPC title

• G06F16/3347 Primary

  using vector based model · CPC title

• Y02D10/00

  Energy efficient computing, e.g. low power processors, power management or thermal management · CPC title

• G06F40/194

  Calculation of difference between files · CPC title

• G06F40/30

  Semantic analysis · CPC title