Methods and systems for artificial intelligence (AI)-assisted document annotation and training of machine learning-based models for document data extraction are described. The methods and systems described herein take advantage of a continuous machine learning approach to create document processing pipelines that provide accurate and efficient data extraction from documents that include structured text, semi-structured text, unstructured text, or any combination thereof.

Disclosed is a text classification method and a text classification device. The text classification method includes: receiving text data (S1), the text data comprising one or more text semantic units; replacing the text semantic unit with a corresponding text keyword (S2), based on a correspondence between text semantic elements and text keywords; extracting, with a semantic model, a semantic feature of the text keyword (S3); and classifying, with a classification model, the text keyword at least based on the semantic feature, as a classification result of the text data (S4).

An example method of extracting structured information from document images comprises: receiving a document image; detecting a tabular structure within the document image; identifying a plurality of rows of the tabular structure, wherein each row of the plurality of rows comprises one or more lines; for each row of the plurality of rows, identifying a set of field types of one or more fields comprised by each line of the one or more lines comprised by the respective row; detecting, in each line of the one or more lines, a set of fields corresponding to a respective set of field types; and extracting information from the set of fields.

In screening of documents based on a similarity between keywords, it is difficult to exclude documents having similar background but different conclusions. In a document review assistance method executed by a computer system, a storage unit stores data on a plurality of documents, and the document review assistance method includes: a step of creating, by a control unit, a sentence vector based on a sentence included in the plurality of documents; a step of classifying, by the control unit, the plurality of documents into a plurality of clusters based on the created sentence vector; a step of specifying, by the control unit, a subgraph on a network of a word in a first document set included in at least one of the clusters; and a step of controlling, by the control unit, the creation of the sentence vector based on the specified subgraph.

A method and system for training a neural network are provided. The method includes receiving an input image, selecting at least one data augmentation method from a pool of data augmentation methods, generating an augmented image by applying the selected at least one data augmentation method to the input image, and generating a mixed image from the input image and the augmented image.

A system and method for machine learning classification of user sentiment is disclosed. The method includes storing including a plurality of category information. The plurality of category information includes a set of domain-specific category information. The method further includes extracting a plurality of aspects from textual data. The method further includes generating a sentiment by a machine learning model. The method further includes receiving the plurality of aspects and the set of domain-specific category information. The method further includes generating a sentiment based on the plurality of aspects and the set of domain-specific category information

Methods, apparatus, systems and articles of manufacture are disclosed for text extraction from a receipt image. An example non-transitory computer readable medium comprises instructions that, when executed, cause a machine to at least improve region of interest detection efficiency by converting pixels of an input receipt image from a first format to a second format, generate a binary representation of the input receipt image based on the converted pixels, the binary representation of the input receipt image corresponding to saturation values for respective ones of the converted pixels, calculate mirror data from the binary representation of the input receipt image, and cluster the binary representation of the input receipt image to identify a first set of candidate regions of interest, the candidate regions of interest characterized by portions of the binary representation of the input receipt image having saturation values that satisfy a threshold value.

A system and method for automating and improving data extraction from a variety of document types, including both unstructured, structured, and nested content, is disclosed. The system and method incorporate an intelligent machine learning model that is designed to intelligently identify chunks of text, map the fields in the document, and extract multi-record values. The system is designed to operate with little to no human intervention, while offering significant gains in accuracy, data visualization, and efficiency. The architecture applies customized techniques including density-based adaptive text clustering, tabular data extraction based on hierarchical intelligent keyword searches, and natural language processing-based field value selection.

Systems and methods for performing OCR of a series of images depicting text symbols. An example method comprises: receiving a current image of a series of images of an original document, wherein the current image at least partially overlaps with a previous image of the series of images; performing optical symbol recognition (OCR) of the current image to produce an OCR text and a corresponding text layout; identifying, using the OCR text and the corresponding text layout, a plurality of textual artifacts in each of the current image and the previous image, wherein each textual artifact is represented by a sequence of symbols that has a frequency of occurrence within the OCR text falling below a threshold frequency; identifying, in each of the current image and the previous image, a corresponding plurality of base points, wherein each base point is associated with at least one textural artifact of the plurality of textual artifacts; identifying, using coordinates of matching base points in the current image and the previous image, parameters of a coordinate transformation converting coordinates of the previous image into coordinates of the current image; associating, using the coordinate transformation, at least part of the OCR text with a cluster of a plurality of clusters of symbol sequences, wherein the OCR text is produced by processing the current image and wherein the symbol sequences are produced by processing one or more previously received images of the series of images; identifying, for each cluster, a median string representing the cluster of symbol sequences; and producing, using the median string, a resulting OCR text representing at least a portion of the original document.

The present disclosure relates to a system and method to extract information from unstructured image documents. The extraction technique is content-driven and not dependent on the layout of a particular image document type. The disclosed method breaks down an image document into smaller images using the text cluster detection algorithm. The smaller images are converted into text samples using optical character recognition (OCR). Each of the text samples is fed to a trained machine learning model. The model classifies each text sample into one of a plurality of pre-determined field types. The desired value extraction problem may be converted into a question-answering problem using a pre-trained model. A fixed question is formed on the basis of the classified field type. The output of the question-answering model may be passed through a rule-based post-processing step to obtain the final answer.