US 2026/0300353 A1 · LARGE LANGUAGE MODEL-ASSISTED ENTITY NAME RESOLUTION

Abstract

Systems and methods for resolving entity names combine hybrid search techniques with large language models to improve accuracy in matching business entities. The system includes an index builder that generates and stores both sparse and dense vector embeddings representing entity information. A query component performs hybrid searches using both keyword-based and semantic methods to retrieve entity candidates from the stored embeddings. When high-confidence matches are not found, a virtual entity generator leverages large language models to generate additional entity candidates based on the input data. The system employs a re-ranker to order retrieved candidates based on their relevance to the input query. This approach enables accurate entity resolution even with noisy or incomplete input data by combining traditional search methods with advanced language model capabilities.

A method for entity resolution with adaptive confidence optimization, the method being performed by one or more processors coupled to a vector store storing sparse vector embeddings and dense vector embeddings representing entity records, comprising:

receiving input data comprising noisy entity information;
generating, from the input data, a query sparse vector embedding and a query dense vector embedding;
performing an initial hybrid search of the vector store by:
comparing the query sparse vector embedding with stored sparse vector embeddings using a keyword-based search method;
and
comparing the query dense vector embedding with stored dense vector embeddings using a semantic search method,
to retrieve entity candidates and corresponding confidence scores;
determining a confidence threshold by:
performing a systematic evaluation that tests threshold values across a predetermined range of threshold values to establish baseline performance;
measuring retrieval performance metrics at each tested threshold value to assess effectiveness;
and
validating threshold performance by processing real-world transaction data and analyzing matching performance;
determining that the confidence scores of the retrieved entity candidates fail to meet the confidence threshold;
responsive to determining that the confidence scores of the retrieved entity candidates fail to meet the confidence threshold:
generating virtual entities using a large language model that normalizes the input data by:
reintroducing spaces in entity names;
correcting abbreviations using learned abbreviation patterns;
identifying and expanding single-letter initials;
analyzing truncated names through contextual understanding;
and
restoring proper spacing and capitalization based on learned formatting rules;
generating one or more additional query embeddings from the generated virtual entities;
performing a second hybrid search of the vector store using the one or more additional query embeddings generated from the virtual entities to retrieve additional entity candidates;
forming a combined candidate set comprising the entity candidates retrieved by the initial hybrid search and the additional entity candidates retrieved by the second hybrid search;
re-ranking the combined candidate set using a composite scoring mechanism that assigns configurable weights to multiple ranking criteria including:
string similarity metrics calculated between query field values and candidate entity name field values;
spatial attribute values derived from location data fields when available;
quantitative attribute values extracted from entity database records;
time-based metadata attribute values indicating data recency;
hierarchical relational data representing associations between entity records;
and
categorical classification attribute matching between entity field values;
cross-referencing top-ranked candidates from the re-ranked combined candidate set against third-party commercial entity information databases having structured data fields including an official entity name and at least one of an address, phone number, website, or business identifier;
and
outputting verified entity information with associated confidence scores.

32 more elements

Large language model-assisted entity name resolution

Life of this patent

Large language model-assisted entity name resolution

Life of this patent

Abstract

Claims20