Ranking Model Training Using Candidate Text Summaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ranking models in search engines, particularly those based on pretrained language models, face efficiency issues when dealing with long candidate texts, leading to high computational resource consumption and low ranking efficiency.
Innovation Solution
The method involves training a ranking model by acquiring sample pairs with label information, determining a target summary for each candidate text, and inputting the search text, title text, and target summary into an initial language model to update its network parameters based on relevance scores.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a pretrained language model-based ranking model is used to rank candidate texts, then ranking accuracy is improved, but computational resource consumption increases and ranking efficiency decreases
Solution Approach 1:
The patent segments the candidate text into a title part and a summary part, where only these segmented portions are fed into the language model for ranking evaluation. This segmentation reduces the input length to the model while preserving the most important information, thereby improving ranking efficiency without significantly compromising accuracy.
Solution Approach 2:
The patent extracts key components (title and summary) from the complete candidate text, discarding the full text content. This extraction approach retains the essential information needed for accurate ranking while eliminating redundant content that would consume computational resources, thus resolving the contradiction between accuracy and efficiency.
2Reliability
If the entire candidate text is input into the language model for training, then model performance is improved, but training time and computational cost increase
Solution Approach 1:
During training, the patent segments candidate texts into title and summary portions, using only these segments as input to the language model. This segmentation reduces the training data processing burden while maintaining the essential information needed for model learning, thereby reducing training time without significantly degrading model performance.
Solution Approach 2:
The patent applies partial action by using only the necessary portions (title and summary) of the candidate text for training, rather than processing the entire text. This partial processing approach is sufficient for achieving good model performance while dramatically reducing computational cost and training time.
Data Source
AI summary
A method of training a ranking model, and an electronic device, which relate to technical fields of natural language processing and intelligent search. The method includes: in training the ranking model, firstly acquiring a plurality of first sample pairs and respective label information; for each first sample pair, inputting a first search text, a first title text of a first candidate text, and a first target summary corresponding to the first candidate text into an initial language model to obtain a second relevance score corresponding to the each first sample pair; then using the first target summary to replace the first candidate text to participate in the training of the ranking model, and updating at least one network parameter of the initial language model according to the label information and the second relevance score corresponding to each first sample pair.