Systems and methods for detection and correction of OCR text
OCR-text correction system and method embodiments are described. The OCR-text correction embodiments comprise or cooperate with a transformer-based sequence-to-sequence language model. The model is pretrained to denoise corrupted text and is fine-tuned using OCR-correction-specific examples. Text obtained at least in part through OCR is applied to the fine-tuned pretrained transformer model to detect at least one error in a subset of the text. Responsive to detecting the at least one error, the fine-tuned pretrained transformer model outputs an updated subset of the text to correct the at least one error.
1 . A computer-implemented method, comprising:
receiving a document that includes text obtained at least in part through OCR;
applying an adjusted bidirectional-and-auto-regressive-transformers (BART) model to the text to detect at least one error in a subset of the text, the adjusted BART model having been adjusted from a BART model pretrained to perform a non-optical character recognition (non-OCR) task using a first training dataset comprising corrupted text data and the adjusted BART model further being adjusted from the BART model to perform an OCR task using a second training dataset comprising OCR samples; and
generating, by applying the adjusted BART model to the text of the document, an updated subset of the text correcting the at least one error in the subset of the text.
2 . The computer-implemented method of claim 1 , wherein the BART model comprises:
a bidirectional encoder configured to receive the text; and
an autoregressive decoder configured to detect the at least one error in the text and correct the at least one error in the text by predicting original text.
3 . The computer-implemented method of claim 1 , wherein the first training dataset includes one or more of the following: token masking, token deletion, sentence permutation, document rotation, or text infilling.
4 . The computer-implemented method of any of claim 1 , wherein the second training dataset includes monograph and periodical example sentences.
5 . The computer-implemented method of claim 1 , wherein the adjusted BART model is configured to perform detection and correction of the at least one error in a single step.
6 . The computer-implemented method of claim 1 , wherein the adjusted BART model is configured to correct the at least one error in the text without being trained on alignment of characters between inputs to an encoder and outputs to a decoder.
7 . The computer-implemented method of claim 1 , wherein the first training dataset comprises fewer than 1,000 documents.
8 . The computer-implemented method of claim 1 , wherein the adjusted BART model has been fine-tuned from a BART-base checkpoint comprising weights arrived at based on pretraining.
9 . The computer-implemented method of claim 1 , wherein the at least one error includes an undersegmentation error caused by incorrectly combining a plurality of words into a single word by OCR or an oversegmentation error caused by incorrectly segmenting a single word into two separate words by OCR.
10 . The computer-implemented method of claim 1 , wherein the adjusted BART model has been fine-tuned utilizing a hugging face package.
11 . The computer-implemented method of claim 1 , wherein the at least one error includes a missing character error caused by incorrectly omitting a character by OCR or a misrecognized character error caused by incorrectly recognizing a character by OCR.
12 . The computer-implemented method of claim 1 , wherein the at least one error includes a hallucination error caused by incorrectly inserting a non-existing character by OCR.
13 . A computer system for detecting and/or correcting text, comprising:
a processor; and
memory in communication with the processor, the memory configured to store instructions that, when executed by the processor, cause the processor to:
access an adjusted bidirectional-and-auto-regressive-transformers (BART) model, the adjusted BART model having been adjusted from a BART model pretrained to perform a non-optical character recognition (non-OCR) task using a first training dataset comprising corrupted text data and the adjusted BART model further being adjusted from the BART model to perform an OCR task using a second training dataset comprising OCR samples;
provide text obtained at least in part through optical character recognition (OCR);
apply the text to the adjusted BART model to detect at least one error in a subset of the text; and
generate an updated subset of the text by the adjusted BART model correcting the at least one error in the subset of the text.
14 . The computer system of claim 13 , wherein the BART model comprises:
a bidirectional encoder configured to receive the text; and
an autoregressive decoder configured to detect the at least one error in the text and correct the at least one error in the text by predicting original text.
15 . The computer system of claim 13 wherein the first training dataset includes one or more of the following: token masking, token deletion, sentence permutation, document rotation, or text infilling.
16 . The computer system of claim 13 , wherein the second training dataset includes monograph and periodical example sentences.
17 . The computer system of claim 13 , wherein the adjusted BART model is configured to perform detection and correction of the at least one error in a single step.
18 . The computer system of claim 13 , wherein the adjusted BART model is configured to correct the at least one error in the text without being trained on alignment of characters between inputs to an encoder and outputs to a decoder.
19 . The computer system of claim 13 , wherein the first training dataset comprises fewer than 1,000 documents.
20 . A non-transitory computer readable storage medium configured to store code comprising instructions, wherein the instructions, when executed by a processor, cause the processor to:
access an adjusted bidirectional-and-auto-regressive-transformers (BART) model, the adjusted BART model having been adjusted from a BART model pretrained to perform a non-optical character recognition (non-OCR) task using a first training dataset comprising corrupted text data and the adjusted BART model further being adjusted from the BART model to perform an OCR task using a second training dataset comprising OCR samples;
provide text obtained at least in part through optical character recognition (OCR);
apply the text to the adjusted BART model to detect at least one error in a subset of the text; and
generate an updated subset of the text by the adjusted BART model correcting the at least one error in the subset of the text.