# US 12,001,951 B2 · Automated contextual processing of unstructured data

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,001,951 B2
- Document ID: US12001951B2
- Status: Granted
- Kind: B2
- Application number: 17/210,153
- Filed: 2021-03-23
- Published: 2022-09-29
- Granted: 2024-06-04
- Earliest claimed: 2021-03-23
- Assignees: Accenture Global Solutions Limited
- Applicants: Accenture Global Solutions Limited
- Inventors: Kavita V V Ganeshan; Swati Tata; Soujanya Soni; Madhur Bhasini Chaini; Anjani Kumari; Omar Razi; Thyagarajan Delli; Ullas Balan Nambiar; Guanglei Xiong; Sivasubramanian Arumugam Jalajam; Srinivasan Krishnan Rajagopalan; Venkatesan Kamalakannan; Harbhajan Singh
- Web page: https://patentgrep.com/patent/US12001951B2

## CPC

- G06N 3/08: Learning methods
- G06F 16/353: into predefined classes
- G06F 18/22: Matching criteria, e.g. proximity measures
- G06F 40/216: using statistical methods
- G06F 40/232: Orthographic correction, e.g. spell checking or vowelisation
- G06F 40/247: Thesauruses; Synonyms
- G06F 40/274: Converting codes to words; Guess-ahead of partial word inputs
- G06F 40/284: Lexical analysis, e.g. tokenisation or collocates
- G06F 40/289: Phrasal analysis, e.g. finite state techniques or chunking
- G06F 40/30: Semantic analysis
- G06N 3/04: Architecture, e.g. interconnection topology
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/0455: Auto-encoder networks; Encoder-decoder networks
- G06N 3/09: Supervised learning
- G06V 10/774: Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
- G06V 10/82: using neural networks
- G06V 30/18057: {Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]}
- G06V 30/262: using context analysis, e.g. lexical, syntactic or semantic context
- G06V 30/40: Document-oriented image-based pattern recognition
- G06V 30/416: Extracting the logical structure, e.g. chapters, sections or page numbers; Identifying elements of the document, e.g. authors
- G06V 30/418: Document matching, e.g. of document images
- G06N 3/047: Probabilistic or stochastic networks
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N 3/096: Transfer learning
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

A system for providing automated and domain specific contextual processing for context based verification may classify a plurality of extracted parameters from a set of digitized training document to assign a document similarity score with respect to a set of reference documents. The system may automatically detect a domain for the set of digitized training documents based on the document similarity score. The system may load a domain based neural model for the detected domain to generate a plurality of pre-defined contextual parameters. The system may receive a set of input documents and perform a contextual processing of the received set of documents based on the pre-defined contextual parameters to obtain an output in form of a plurality of filtered snippets, each bearing a corresponding rank. The context based verification may be performed based on the plurality of filtered snippets and the corresponding rank.

## Claims

### Claim 1

A system for automated contextual processing, the system comprising:
a processor;
a data trainer coupled to the processor, the data trainer to:
classify, using a classification model, a plurality of extracted parameters from a set of digitized training documents, wherein the classification is performed to assign a document similarity score with respect to a set of reference documents corresponding to a plurality of domains;
detect automatically, a domain for the set of digitized training documents based on the document similarity score; and
load a domain based neural model for the detected domain to generate a plurality of pre-defined contextual parameters specific to the detected domain, the plurality of pre-defined contextual parameters being obtained by extraction of multiple queries from the set of digitized training documents and subsequent processing of the extracted queries;
a contextual processing engine of the processor, the engine to:
receive a set of input documents obtained by digitization of a non-digital documents;
perform, through an AI model, a contextual processing of the received set of input documents based on the pre-defined contextual parameters to obtain an output in form of a plurality of filtered snippets each bearing a corresponding rank, the contextual processing comprising context building, context search and context-based ranking of one or more snippets extracted from the input documents; and
wherein a context-based verification of the unstructured data is performed based on the plurality of filtered snippets and the corresponding rank; and
a hybrid ensemble coupled to the processor, wherein the hybrid ensemble is configured to:
receive, from the contextual processing engine, a first data comprising the plurality of filtered snippets;
receive, from a learning engine coupled to the processor, a second data comprising an updated plurality of filtered snippets;
classify, using one or more models, the first data and the second data in a pre-defined format;
assign, using the one or more models, a pre-defined weight to each of the classified first data and the classified second data;
determine, using the one or more models, a similarity score based on the assigned weights; and
update, using the one or more models, the rank of each snippet in the plurality of filtered snippets to assign an updated rank.

### Claim 2 (depends on 1)

The system as claimed in claim 1, wherein the system comprises the learning engine of the processor to generate the updated plurality of filtered snippets by:
updating, using a neural network based model, the plurality of filtered snippets by performing a context-based detection and context-based update, based on a feedback from a user interface, the feedback indicating preference of at least one of the plurality of filtered snippets, and wherein the neural network based model is an encoder decoder based long short term memory (LSTM) model.

### Claim 3 (depends on 1)

The system as claimed in claim 1, wherein the classification comprises a few-shot text classification, wherein the classification model comprises at least one of an attention-based induction network model, and a Siamese based network model, and wherein the document similarity score is generated by a natural language processing tool.

### Claim 4 (depends on 1)

The system as claimed in claim 1, wherein the document similarity score is generated by comparing the extracted plurality of parameters with corresponding parameters in the set of reference documents.

### Claim 5 (depends on 1)

The system as claimed in claim 1, wherein the plurality of pre-defined contextual parameters are generated by the domain based neural model in combination with a first predefined vocabulary specific to the detected domain, wherein the processing of the extracted queries comprises enrichment of the extracted queries by at least one of synonym enrichment, domain based filtering, and query cleaning.

### Claim 6 (depends on 5)

The system as claimed in claim 5, wherein the synonym enrichment comprises generating one or more synonyms for the extracted queries, wherein the domain-based filtering comprises filtering the extracted queries based on relevance with respect to the detected domain, and the query cleaning comprises obtaining the extracted queries in a predefined format.

### Claim 7 (depends on 1)

The system as claimed in claim 1, wherein the plurality of filtered snippets are generated by the AI model in combination with a second predefined vocabulary specific to the detected domain.

### Claim 8 (depends on 1)

The system as claimed in claim 1, wherein the context building in the contextual processing is performed based on at least one of spelling error detection, abbreviation expansion, alias detection, context window detection, critical word detection, key entity detection, enrichment of fieldname with synonyms, fuzzy querying and phrase search within permissible gap limits to obtain context based results.

### Claim 9 (depends on 8)

The system as claimed in claim 8, wherein the critical word detection is to detect relevance of a word in presence of one or more synonyms.

### Claim 10 (depends on 8)

The system as claimed in claim 8, wherein based on the combination of the plurality of pre-defined contextual parameters from the data trainer and the context-based results obtained from the context building, the context-based search is performed by an elastic search so as to obtain the plurality of filtered snippets.

### Claim 11 (depends on 1)

The system as claimed in claim 1, wherein the context-based ranking includes allocating a score for each snippet and assigning the corresponding rank to the plurality of filtered snippets based on the allocated score using a ranking model.

### Claim 12 (depends on 1)

The system as claimed in claim 1, wherein the system comprises an automated tag generator coupled to the processor, the tag generator to:
generate an annotation using the plurality of filtered snippets obtained from contextual processing, wherein the annotation corresponds to one or more contextual coordinates corresponding to an image document obtained after scanning of the set of input documents,
wherein the annotation is fed to the data trainer for automated named entity recognition (NER).

### Claim 13 (depends on 1)

The system as claimed in claim 1, wherein the plurality of pre-defined contextual parameters and plurality of filtered snippets comprise at least one of entity name, entity identification data, and serial number of service provided by the entity, date of service initiation, service duration, address of entity and operational information.

### Claim 14

A method for automated contextual processing, the method comprising:
classifying, by a processor, using a classification model, a plurality of extracted parameters from a set of digitized training documents, wherein the classification is performed to assign a document similarity score with respect to a set of reference documents corresponding to a plurality of domains;
detecting automatically, by the processor, a domain for the set of digitized training documents based on the document similarity score;
loading, by the processor, a domain based neural model for the detected domain to generate a plurality of pre-defined contextual parameters specific to the detected domain, the plurality of pre-defined contextual parameters being obtained by extraction of multiple queries from the set of digitized training documents and subsequent processing of the extracted queries;
receiving, by the processor, a set of input documents obtained by digitization of a non-digital documents;
performing, by the processor, using an AI model, a contextual processing of the received set of input documents based on the pre-defined contextual parameters to obtain an output in form of a plurality of filtered snippets, each bearing a corresponding rank, the contextual processing comprising context building, context search and context-based ranking of one or more snippets extracted from the input documents;
wherein a context-based verification of the unstructured data is performed based on the plurality of filtered snippets and the corresponding rank;
receiving, by the processor, from a contextual processing engine, a first data comprising the plurality of filtered snippets;
receiving, by the processor, from a learning engine coupled to the processor, a second data comprising an updated plurality of filtered snippets;
classifying, by the processor, using one or more models, the first data and the second data in a pre-defined format;
assigning, by the processor, using the one or more models, a pre-defined weight to each of the classified first data and the classified second data;
determining, by the processor, using the one or more models, a similarity score for the plurality of filtered snippets based on the assigned weights; and
updating, by the processor, using the one or more models, the rank of each snippet in the plurality of filtered snippets to assign an updated rank.

### Claim 15 (depends on 14)

The method as claimed in claim 14, wherein the updated plurality of filtered snippets are generated by:
updating, by the processor, using a neural network based model, the plurality of filtered snippets by performing a context-based detection and context-based update based on a feedback from the user interface, the feedback indicating preference of at least one of the plurality of filtered snippets, and wherein the neural network based model is an encoder decoder based long short term memory (LSTM) model.

### Claim 16 (depends on 14)

The method as claimed in claim 14 comprising:
generating automatically, by the processor, an annotation using the plurality of filtered snippets obtained from contextual processing, wherein the annotation corresponds to one or more contextual coordinates corresponding to an image document obtained after scanning of the set of input documents,
wherein the annotation is fed to the data trainer for enabling automated named entity recognition (NER).

### Claim 17

A non-transitory computer readable medium comprising machine executable instructions that are executable by a processor to:
classify a plurality of extracted parameters from a set of digitized training documents, wherein the classification is performed so as to assign a document similarity score with respect to a set of reference documents corresponding to a plurality of domains;
detect automatically, a domain for the set of digitized training documents based on the document similarity score;
load a domain based neural model for the detected domain to generate a plurality of pre-defined contextual parameters specific to the detected domain, the plurality of pre-defined contextual parameters being obtained by extraction of multiple queries from the set of digitized training documents and subsequent processing of the extracted queries;
receive a set of input documents that are obtained by digitization of a non-digital input documents through optical character recognition (OCR); and
perform a contextual processing of the received set of input documents based on the pre-defined contextual parameters to obtain an output in form of a plurality of filtered snippets each bearing a corresponding rank, the contextual processing comprising context building, context search and context-based ranking of one or more snippets extracted from the input documents;
wherein a context-based verification of the unstructured data is performed based on the plurality of filtered snippets and the corresponding rank;
receive a first data comprising the plurality of filtered snippets;
receive a second data comprising an updated plurality of filtered snippets;
classify using one or more models, the first data and the second data in a pre-defined format;
assign using the one or more models, a pre-defined weight to each of the classified first data and the classified second data;
determine using the one or more models, a similarity score for the plurality of filtered snippets based on the assigned weights; and
update using the one or more models, the rank of each snippet in the plurality of filtered snippets to assign an updated rank.

### Claim 18 (depends on 17)

The non-transitory computer readable medium as claimed in claim 17,
wherein to generate the updated plurality of filtered snippets, the machine executable instructions are executable by a processor to:
update the plurality of filtered snippets by performing a context-based detection and context-based update, based on a feedback from a user interface, the feedback indicating preference of at least one of the plurality of filtered snippets, and wherein the neural network-based model is an encoder decoder based long short-term memory (LSTM) model.