LLM Compound Word Splitting With Prompt Voting for Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Search engines struggle to accurately split compound words in languages like German, leading to incorrect search results due to improper word splitting, which affects vector-based and keyword-based systems, especially in domain-specific contexts.
Innovation Solution
Utilize a large language model (LLM) with varied prompts and a voting technique to generate and select the most appropriate word split for compound words, incorporating domain-specific data to enhance contextual relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional word splitting methods are used, then the system is simple and fast, but the accuracy of compound word splitting is poor leading to incorrect search results
Solution Approach 1:
The patent uses a large language model to generate multiple candidate word splits by copying and adapting from its training data on compound words. The LLM replicates the splitting patterns it learned from corpus data, generating several possible splits that can then be evaluated and selected, thereby achieving high accuracy without requiring complex custom training for each domain
Solution Approach 2:
The patent introduces a voting mechanism as an intermediary between the LLM's generated splits and the final selected split. This voting system acts as a mediator that aggregates results from multiple LLM generations and selects the most appropriate split based on frequency or other criteria, adding a layer of decision-making that improves accuracy while keeping the overall system architecture relatively simple
2Reliability
If domain-specific models are trained to improve word splitting accuracy, then the relevance of search results improves, but the training time and computational resources increase
Solution Approach 1:
The patent employs a universal large language model that can handle multiple domains and languages without requiring separate domain-specific training. The LLM's pre-trained knowledge enables it to generate appropriate word splits across different domains by adapting to context through prompting, eliminating the need for time-consuming domain-specific training while maintaining high search result relevance
Solution Approach 2:
The patent adjusts prompt parameters and temperature settings to control the LLM's output for different domains. By changing prompt formulations and generation parameters rather than retraining the model, the system adapts to domain-specific requirements efficiently, reducing training time while maintaining or improving search relevance through parameter optimization
3Measurement precision
If multiple prompts are used to generate candidate word splits, then the variety and quality of splits improves, but the computational cost and processing time increase
Solution Approach 1:
The patent generates multiple candidate word splits using the LLM with varied prompts and then applies a voting mechanism to select the final split. This partial action approach generates several possibilities rather than exhaustively searching all possible splits, achieving sufficient quality through a manageable number of LLM calls while avoiding the excessive computational cost of complete enumeration
Solution Approach 2:
The voting mechanism automatically selects the best word split based on frequency or predefined criteria without requiring manual intervention or complex evaluation algorithms. The system serves itself by using the aggregated results from multiple LLM generations to make the final decision, reducing the need for additional computational resources for human evaluation or complex post-processing
Data Source
AI summary
The technology relates to determining word splits for compound words using a large language model (LLM). It can be used to enhance search engine performance in languages where words are often combined as compound words, such as German and Dutch. An example method involves prompting the LLM with different prompts to generate multiple candidate word splits for a compound word. A voting technique is applied to select the most appropriate word split. The method may include using different LLM temperatures and compound word-word split pairs from a domain-specific dataset as examples within the prompts. The voting technique may identify the word split that appears most frequently. If no majority, the method selects a candidate word split based on the number of splits, either the highest or lowest, and in some cases, selects a random word split from candidate word splits with the highest or lowest number of splits.