US 12,608,410 B1Granted

Apparatus for preparing responses

AssigneeS2 Strategy, LLC

InventorsAlex Firestone, Brendan Corr, Brian Stone, Brian McLaughlin

Life of this patent

Estimating the term…

Loading patent…

Abstract

A device and method to improve the automation for answering questions in multiple computer forms is described herein. For example, an electronic questionnaire form seeking information from an institutional asset manager in a questionnaire format may require the asset manager to answer the same question for multiple potential clients. The tool described herein stores previous questions and answers, and provides the answer to a question if a near-exact question has previously been asked. If a similar question has been previously asked the tool identifies the closest answers, and then uses a large language model to combine the previous answers into a new answer. The tool may parse a questionnaire into questions and then use the cosine similarity score to determine the closeness of the question to previously answered questions.

Claims20

Dependent claims

An apparatus comprising:

a processor;
a memory communicatively connected to the processor, the memory including a question and answer database, a large language model, a table of stems, a document, a best answer list, and a set of non-transitory machine readable instructions, where the question and answer database includes a question text, an answer text, and a question stem table, the question stem table including a database vector;
where the processor is configured to execute the set of non-transitory machine readable instructions to:
receive the document;
loop through each question in the document;
parse the question into words;
lemmatize each word into the table of stems;
convert the table of stems into a question vector;
loop through each entry in the question and answer database;
calculate a cosine similarity score between the question vector and the database vector for the entry;
if the cosine similarity score is greater than an autocomplete threshold, then save the cosine similarity score and a pointer to the entry in the best answer list;
if the cosine similarity score is greater than a similarity threshold, then save the cosine similarity score and the pointer to the entry in the best answer list;
and
continue the loop through the question and answer database;
if the best answer list is empty, prompt a user for an answer;
if a largest cosine similarity score in the best answer list is less than the autocomplete threshold, prompt the user to select one or more answer texts in the best answer list, and combine the selected answer texts with the question through the large language model to generate the answer;
store the question, the table of stems, the question vector, and the answer in a new entry in the question and answer database;
store the answer in the document;
and
continue the loop through each question in the document.

17 more elements

A method comprising:

receiving a document into a computer memory, the computer memory including a question and answer database, a large language model, a table of stems, the document, a best answer list, and a set of non-transitory machine readable instructions, where the question and answer database includes a question text, an answer text, and a question stem table, the question stem table including a database vector;
looping through each question in the document;
parsing the question into words;
lemmatizing each word into the table of stems;
converting the table of stems into a question vector;
looping through each entry in the question and answer database;
calculating a cosine similarity score between the question vector and the database vector for the entry;
if the cosine similarity score is greater than an autocomplete threshold, then saving the cosine similarity score and a pointer to the entry in the best answer list;
if the cosine similarity score is greater than a similarity threshold, then saving the cosine similarity score and the pointer to the entry in the best answer list;
and
continuing the looping through the question and answer database;
if the best answer list is empty, prompting a user for an answer;
if a largest cosine similarity score in the best answer list is less than the autocomplete threshold, prompting the user to select one or more answer texts in the best answer list, and combining the selected answer texts with the question through the large language model to generate the answer;
storing the question, the table of stems, the question vector, and the answer in a new entry in the question and answer database;
storing the answer in the document;
and
continuing the looping through each question in the document.

14 more elements

Dependencies are inferred from the claim text.