Method of digital document review using factor graph document databases
A method of digital document review using factor graph document databases comprising an extract method and a summarise method, wherein the extract method uses a Large Language Model (LLM) to decompose into variables and elements a document or series of documents and to identify new terms semantically associated with the description of the interest of the desired audience for the document review process, and representing the variables and elements in a factor graph document database, and wherein the summarise method includes indexing the factor graph document database based on terms of interest to the audience of the review, and inferring which of the new variables and elements are relevant to the audience, and reconstructing a human interpretable document out of the variables and elements inferred in the factor graph document database.
1 . A method of digital document review using a factor graph document database according to a summary of needs provided by a user, comprising,
receiving user input data, wherein the input data includes a document written in natural language that is to be reviewed, and further includes a summary of an audience's needs with respect to the document to be reviewed,
decomposing the text of the document to be reviewed according to the summary of the audience's needs into variables and elements for use in a factor graph document database, using a method of comparing data strings,
identifying new variables and elements relevant to the audience's needs that are semantically associated with the variables and elements decomposed from the text of the document to be reviewed, using a method of comparing data strings,
representing the relationship between all the variables, elements and factors in a factor graph document database, using a computer program,
inferring among the new variables those which are most probably related to the variables and elements representing the text of the document to be reviewed,
reconstructing a human interpretable document containing the variable, elements and factors according to the summary of needs and the most probable new variables, using a method of comparing data strings.
2 . A method of claim 1 , wherein the factor graph document database is made of one or more destination nodes representing variables over one or more elements, and of one or more source nodes representing variables over one or more elements, and of one or more edges capable of performing mathematical or logical operations over the elements of the variables represented by the destination nodes and the source nodes.
3 . A method of claim 1 , wherein the method of comparing data strings uses a Large Language Model and a series of engineered prompts.
4 . A method of claim 2 , wherein the inference over variables and elements of the factor graph document database is a message passing scheme using a sum-product algorithm for belief propagation that iterates over the factor graph document database, wherein the messages are propagated between adjacent source and destination nodes, and wherein each iteration updates the value of the variables and elements represented by the nodes until convergence is achieved.