Natural language framework for contextual entity identification
A system for mitigating language-model hallucinations during entity identification can include a server that stores multiple entity records in a database and, upon receiving from a client a request containing selection criteria, retrieves a first record for the candidate entity. The server forms a first natural-language prompt containing questions that embody the criteria plus the first record data, submits the prompt to a language model, and obtains a first answer. If the first answer alone fails to satisfy the criteria, the server retrieves a second record for the same entity, constructs a second prompt repeating the questions and appending the second record, and obtains a second answer from the model. The server then evaluates the aggregate of the first and second answers; when the combined information fulfills the selection criteria, the server issues an alert identifying the entity.
1 . A system for reducing language model hallucinations for entity identification, the system comprising:
a server comprising one or more processors, configured by instructions stored in memory to:
store a plurality of records in a database, each record containing a set of data for an entity;
receive, from a client device, a request comprising a set of criteria to identify whether the entity satisfies the set of criteria;
generate a first natural language query comprising one or more questions determined to resolve the set of criteria and a first record retrieved from the database using an identifier of the entity;
execute a language model using the first natural language query to generate at least a first answer to the one or more questions;
responsive to determining the first answer does not satisfy the set of criteria, generate a second natural language query comprising the one or more questions and a second record retrieved from the database using the identifier of the entity;
execute the language model using the second natural language query to generate at least a second answer to the one or more questions;
determine at least one of the one or more questions can be resolved with structured data;
responsive to the determination, map the at least one question to a programmatic query configured to query one or more of the plurality of records;
execute the programmatic query to retrieve a third answer from the one or more records;
determine a combination of the first answer, the second answer, and the third answer satisfy the set of criteria; and
generate an alert identifying the entity responsive to the determining the combination of at least the first answer, the second answer, and the third answer satisfy the set of criteria.
2 . A system for reducing language model hallucinations for entity identification, the system comprising:
a server comprising one or more processors, configured by instructions stored in memory to:
store a plurality of records in a database, each record containing a set of data for an entity;
receive, from a client device, a request comprising a set of criteria to identify whether the entity satisfies the set of criteria;
generate one or more questions determined to resolve the set of criteria;
determine at least one of the one or more questions can be resolved with structured data;
responsive to the determination, map the at least one question to a programmatic query configured to query one or more of the plurality of records;
execute the programmatic query to retrieve a first answer from the one or more records;
generate, in response to determining the first answer does not satisfy the set of criteria, a first natural language query comprising the one or more questions determined to resolve the set of criteria and a first record retrieved from the database using an identifier of the entity;
execute a language model using the first natural language query to generate at least a second answer to the one or more questions;
responsive to determining the second answer does not satisfy the set of criteria, generate a second natural language query comprising the one or more questions and a second record retrieved from the database using the identifier of the entity;
execute the language model using the second natural language query to generate at least a third answer to the one or more questions;
determine a combination of at least the second answer and the third answer satisfy the set of criteria; and
generate an alert identifying the entity responsive to the determining the combination of at least the second answer and the third answer satisfy the set of criteria.
3 . The system of claim 1 , wherein the language model is a decoder-only model, wherein the one or more processors are configured to:
execute an encoder-decoder model using the first answer and the second answer to generate a normalized first answer and second answer, and
wherein the one or more processors are configured to determine the combination of the first answer and the second answer satisfy the set of criteria by determining a combination of the normalized first answer and second answer satisfy the set of criteria.
4 . The system of claim 1 , wherein the one or more processors are configured to:
generate, using a second language model, a set of code using the first answer and/or the second answer; and
identify, using the set of code, a first record of the plurality of records based on which the first answer and/or the second answer were generated.
5 . The system of claim 4 , wherein the one or more processors are configured to:
output an identification of the first record to a user interface of the client device.
6 . The system of claim 1 , wherein the one or more processors are configured to execute the language model using the first natural language query to generate at least the first answer to the one or more questions by:
generating, using the language model, an intermediate representation of the first answer generated based on the first record and a confidence score for the intermediate representation of the first answer; and
providing the intermediate representation of the first answer and the confidence score to the language model.
7 . The system of claim 1 , wherein the one or more processors are configured to:
store the first answer in a cache; and
retrieve the first answer from the cache responsive to generating the second answer.
8 . A method, comprising:
storing, by one or more processors, a plurality of records in a database, each record containing a set of data for an entity;
receiving, by the one or more processors, from a client device, a request comprising a set of criteria to identify whether the entity satisfies the set of criteria;
generating, by the one or more processors, a first natural language query comprising one or more questions determined to resolve the set of criteria and a first record retrieved from the database using an identifier of the entity;
executing, by the one or more processors, a language model using the first natural language query to generate at least a first answer to the one or more questions:
responsive to determining the first answer does not satisfy the set of criteria, generating, by the one or more processors, a second natural language query comprising the one or more questions and a second record retrieved from the database using the identifier of the entity;
executing, by the one or more processors, the language model using the second natural language query to generate at least a second answer to the one or more questions;
determining, by the one or more processors, at least one of the one or more questions can be resolved with structured data;
responsive to the determination, mapping, by the one or more processors, the at least one question to a programmatic query configured to query one or more of the plurality of records;
executing, by the one or more processors, the programmatic query to retrieve a third answer from the one or more records;
determining, by the one or more processors, a combination of the first answer, the second answer, and the third answer satisfy the set of criteria; and
generating, by the one or more processors, an alert identifying the entity responsive to the determining that the combination of at least the first answer and the second answer satisfy the set of criteria.
9 . The method of claim 8 , comprising generating, by the one or more processors, the first natural language query comprising the one or more questions and the first record in response to determining the third answer does not satisfy the set of criteria.
10 . The method of claim 8 , wherein the language model is a decoder-only model, the method comprising:
executing, by the one or more processors, an encoder-decoder model using the first answer and the second answer to generate a normalized first answer and second answer,
wherein determining the combination of the first answer and the second answer satisfy the set of criteria comprises determining, by the one or more processors, a combination of the normalized first answer and second answer satisfy the set of criteria.
11 . The method of claim 8 , comprising:
generating, by the one or more processors, using a second language model, a set of code using the first answer and/or the second answer; and
identifying, by the one or more processors using the set of code, a first record of the plurality of records based on which the first answer and/or the second answer were generated.
12 . The method of claim 11 , comprising:
outputting, by the one or more processors, an identification of the first record to a user interface of the client device.
13 . One or more non-transitory machine-readable media storing program instructions that, when executed by one or more processors, causes the one or more processors to perform operations comprising:
storing a plurality of records in a database, each record containing a set of data for an entity;
receiving, from a client device, a request comprising a set of criteria to identify whether the entity satisfies the set of criteria;
generating a first natural language query comprising one or more questions determined to resolve the set of criteria and a first record retrieved from the database using an identifier of the entity;
executing a language model using the first natural language query to generate at least a first answer to the one or more questions:
responsive to determining the first answer does not satisfy the set of criteria, generating a second natural language query comprising the one or more questions and a second record retrieved from the database using the identifier of the entity;
executing the language model using the second natural language query to generate at least a second answer to the one or more questions;
determining at least one of the one or more questions can be resolved with structured data;
responsive to the determination, mapping the at least one question to a programmatic query configured to query one or more of the plurality of records; and
executing the programmatic query to retrieve a third answer from the one or more records;
determining a combination of at least the first answer, the second answer, and the third answer satisfy the set of criteria; and
generating an alert identifying the entity responsive to the determining that the combination of at least the first answer and the second answer satisfy the set of criteria.
14 . The one or more non-transitory machine-readable media of claim 13 , the operations comprising generating the first natural language query comprising the one or more questions and the first record in response to determining the third answer does not satisfy the set of criteria.