Geospatial Location Identification Using LLM and Classifier Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for identifying geospatial locations based on user input, such as large language models (LLMs), face challenges with high costs, accuracy issues, and 'hallucinations' that lead to inaccurate results.
Innovation Solution
The proposed method involves receiving user input, querying a large language model (LLM) with specific parameters to identify potential POI categories, and post-processing the output to extract relevant POI category labels. This method also combines LLMs with other interpretation methods to reduce costs and increase accuracy, and treats 'hallucinations' as features to update POI databases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If large language models (LLMs) are used to interpret user queries and identify geospatial locations, then the ability to handle ambiguous and incomplete user input is improved, but the processing cost and time increase significantly
Solution Approach 1:
The system segments the query processing into multiple stages: first using a lightweight classifier to determine query intent (address search, POI search, or both), then applying LLM only to ambiguous queries that require category interpretation. This segmentation reduces overall LLM usage while maintaining handling capability for complex queries.
Solution Approach 2:
The system applies partial action by using LLM only when necessary (for ambiguous queries) rather than for all queries. The classifier handles straightforward queries efficiently, while LLM provides enhanced interpretation only when needed, avoiding excessive computational resource consumption.
2Measurement precision
If LLMs are used to identify POI categories from user input, then the accuracy of location identification is improved, but hallucinations occur leading to inaccurate results
Solution Approach 1:
The system incorporates feedback mechanisms where the output of the LLM is fed back into the classification system for validation. The classifier verifies whether the LLM-generated POI category makes sense in context, and can request reprocessing or correction when hallucinations are detected, thereby improving reliability while maintaining accuracy.
Solution Approach 2:
The classifier acts as an intermediary between user input and LLM processing. It filters and prepares queries, selecting only those that truly require LLM interpretation, thereby reducing the impact of hallucinations on overall system performance and maintaining higher reliability.
3Measurement precision
If multiple factors are used to rank and disambiguate search results, then the relevance of results to user intent is improved, but the complexity of the processing system increases
Solution Approach 1:
The ranking and disambiguation process is segmented into multiple independent stages: initial filtering by classifier, LLM-based category identification for ambiguous queries, and subsequent ranking using stored POI information. This modular segmentation reduces overall system complexity while maintaining high result relevance through coordinated operation of simpler components.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method of identifying at least one geospatial location based on user input is described. The method comprises: receiving (S100) a string (10) from a user as the user input; using (5200) the string (10) for identifying at least one point-of-interest, POI; and identifying (S300), as the at least one geospatial location, a geospatial location (50) corresponding to the at least one identified POI. Using (S200) the string (10) comprises: querying, based on at least part of the string and a predetermined plurality of POI categories, a large language model, LLM (40), for output relating to identification of at least one POI category label. Furthermore, a method of updating POI information is provided, which may advantageously use selections by a user received when executing the former method.