Automated optimization framework for accurate extraction of tabular data from diverse PDF documents using large language models
The disclosure relates to a self-correcting artificial intelligence (AI) framework that receives PDF documents for extracting and validating tabular data. The framework includes a data extractor module that reconstructs tabular data into a structured dataframe using a data extraction algorithm. A comparator module validates the extracted data by querying the dataframe and generating a reconstruction loss calculation. An error correction module iteratively adjusts the data extractor module's parameters to minimize reconstruction loss. An adaptive learner module continuously adapts the framework to new data, monitoring a dual-model architecture comprising a primary large language model (LLM) module and a secondary model. The primary LLM module generates ground truth and synthetic data to train the secondary model, which determines the confidence level of the extracted data. The framework outputs final extracted data after validation and error correction, ensuring accurate representation of the original tabular content in the PDF documents.
1 . A computer-implemented method comprising:
receiving one or more PDF documents by a self-correcting artificial intelligence (AI) framework for extracting and validating tabular data from the PDF documents;
extracting and reconstructing the tabular data in a format as extracted data from the PDF documents into a structured dataframe using a data extractor module;
validating the extracted data by querying the structured dataframe using a comparator module, wherein the comparator module compares the extracted data against predefined criteria and generates a reconstruction loss calculation;
adjusting parameters of the data extractor module based on the reconstruction loss calculation using an error correction module, wherein the error correction module iteratively minimizes the reconstruction loss;
continuously adapting the framework to new data using an adaptive learner module, wherein the adaptive learner module monitors the performance of a dual-model architecture and triggers retraining when a high reconstruction loss is detected, and wherein the dual-model architecture comprises a primary large language model (LLM) module and a secondary model;
generating ground truth data and synthetic data using the primary LLM module and a synthetic data generator module, wherein the ground truth data and synthetic data are used to train the secondary model;
determining a confidence level of the extracted data using the secondary model; and
outputting final extracted data of the data extraction module after validation and error correction processes, ensuring the final extracted data accurately represents an original format of the tabular content of the PDF documents.
2 . The method of claim 1 , wherein the data extractor module employs a parameter-based data extraction algorithm to read bounding box information from the PDF documents.
3 . The method of claim 2 , wherein the data extractor module processes both searchable and scanned PDF documents, wherein the data extractor module processes the scanned PDF documents using OCR technology to extract the tabular content of the PDF documents.
4 . The method of claim 1 , wherein the data extractor module employs a column clustering algorithm to ensure proper column alignment in an EXCEL file.
5 . The method of claim 4 , wherein the column clustering algorithm calculates a content density of each column to ensure proper alignment in the EXCEL file.
6 . The method of claim 1 , wherein the comparator module includes a reconstruction loss calculation to measure accuracy of the extracted data.
7 . The method of claim 6 , wherein the reconstruction loss calculation is based on a loss function that compares ground truth answers and the confidence level of the extracted data;
and wherein the error correction module constructs a sample space from the extracted data to identify the extracted data that minimizes the loss function.
8 . The method of claim 7 , wherein the synthetic data generator module creates additional data fields from the same PDF and/or other past PDFs to mimic a correct data distribution and generate incorrect information for training the secondary model.
9 . The method of claim 7 , wherein the secondary model evaluates question, value, and context variables to provide a probabilistic confidence outcome representing a confidence in a correctness of the ground truth answers.
10 . The method of claim 1 , wherein the primary LLM module generates both correct (positive) and incorrect (negative) question-answer pairs from the tabular data in the PDF document to simulate hallucinatory behavior to mitigate hallucinatory responses.
11 . The method of claim 1 , wherein the error correction module operates in a gradient descent manner to iteratively minimize the reconstruction loss.
12 . A system comprising:
a self-correcting artificial intelligence (AI) framework configured to receive one or more PDF documents for extracting and validating tabular data from the PDF documents;
a data extractor module configured to extract and reconstruct the tabular data into a structured dataframe using a parameter-based data extraction algorithm;
a comparator module configured to validate the extracted data by querying the structured dataframe and comparing the extracted data against predefined criteria, generating a reconstruction loss calculation;
an error correction module configured to adjust parameters of the data extractor module based on the reconstruction loss calculation, iteratively minimizing the reconstruction loss;
an adaptive learner module configured to continuously adapt the apparatus to new data by monitoring the performance of a dual-model architecture and triggering retraining when a high reconstruction loss is detected, wherein the dual-model architecture comprises a primary large language model (LLM) module and a secondary model;
wherein the primary LLM module is configured to generate ground truth data and use a synthetic data generator module to generate synthetic data, wherein the ground truth data and the synthetic data are used to train the secondary model;
wherein the secondary model is configured to determine a confidence level of the extracted data;
wherein the data extractor module employs a column clustering algorithm to ensure proper column alignment in the extracted data; and
wherein the framework outputs final extracted data after validation and error correction processes, ensuring the final extracted data accurately represents an original format of the tabular content of the PDF documents.
13 . The system of claim 12 , wherein the data extractor module employs a parameter-based data extraction algorithm to read bounding box information from the PDF documents.
14 . The system of claim 13 , wherein the data extractor module processes both searchable and scanned PDF documents, wherein the data extractor module processes the scanned PDF documents using OCR technology to extract the tabular content of the PDF documents.
15 . The system of claim 12 , wherein the data extractor module employs a column clustering algorithm to ensure proper column alignment in an EXCEL file; and
wherein the column clustering algorithm calculates a content density of each column to ensure proper alignment in the EXCEL file.
16 . The system of claim 12 , wherein the comparator module includes a reconstruction loss calculation to measure accuracy of the extracted data.
17 . The system of claim 16 , wherein the reconstruction loss calculation is based on a loss function that compares ground truth answers and the confidence level of the extracted data; and
wherein the error correction module constructs a sample space from the extracted data to identify the extracted data that minimizes the loss function.
18 . The system of claim 17 , wherein the secondary model evaluates question, value, and context variables to provide a probabilistic confidence outcome representing a confidence in a correctness of the ground truth answers.
19 . The system of claim 12 , wherein the primary LLM module generates both correct (positive) and incorrect (negative) question-answer pairs from the tabular data in the PDF document to simulate hallucinatory behavior to mitigate hallucinatory responses.
20 . A non-transitory computer-readable storage medium storing instructions that when executed by a computer cause the computer to perform a method comprising:
receiving one or more PDF documents by a self-correcting artificial intelligence (AI) framework for extracting and validating tabular data from the PDF documents;
extracting and reconstructing the tabular data in a format as extracted data from the PDF documents into a structured dataframe using a data extractor module;
validating the extracted data by querying the structured dataframe using a comparator module, wherein the comparator module compares the extracted data against predefined criteria and generates a reconstruction loss calculation;
adjusting parameters of the data extractor module based on the reconstruction loss calculation using an error correction module, wherein the error correction module iteratively minimizes the reconstruction loss;
continuously adapting the framework to new data using an adaptive learner module, wherein the adaptive learner module monitors the performance of a dual-model architecture and triggers retraining when a high reconstruction loss is detected, and
wherein the dual-model architecture comprises a primary large language model (LLM) module and a secondary model;
generating ground truth data and synthetic data using the primary LLM module and a synthetic data generator module, wherein the ground truth data and synthetic data are used to train the secondary model;
determining a confidence level of the extracted data using the secondary model; and
outputting final extracted data of the data extraction module after validation and error correction processes, ensuring the final extracted data accurately represents an original format of the tabular content of the PDF documents.