IP Library › Granted Patent US 12,682,171
Granted Patent B2
US 12,682,171 · App. 18/479,051 · Granted Jul 14, 2026

Context disambiguation using deep neural networks

Inventors: Rodrigo Reis Alves (Campinas, BR); Angelo Moore (Dunboyne, IE); Valdir Salustino Guimaraes (Americana, BR); Daniela Arrigoni (Seregno, IT); Vasanthi M. Gopal (Plainsboro, NJ)
Assignee: International Business Machines Corporation
G06F40/30G06F40/186
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 12,682,171
App. No.
18/479,051
Filed
Sep 30, 2023
Granted
Jul 14, 2026
Kind
B2
Art Unit
2655
USPC
704/9
Abstract

A computer-implemented process for updating an electronic document includes the following operations. Using a preprocessor, preprocessing is performed on the electronic document to generate a computer data structure. The computer data structure is evaluated using a word sense disambiguation (WSD) engine and a deep neural network to determine a context of a sentence within the electronic document. Based upon the context, a determination is made that a word within the sentence is a word of interest. The sentence is rewritten using a mitigation engine and a large language model to generate a revised sentence that does not include the word of interest. A determination is made that the revised sentence does not include any other word of interest; and the electronic document is updated to include the revised sentence.

Claims (70)

1 . A computer-implemented method for updating an electronic document employing a preprocessor, a word sense disambiguation (WSD) engine, and a mitigation engine, comprising:

performing preprocessing on the electronic document, using the preprocessor, to generate a computer data structure;

identifying, using a knowledge base that stores a mapping between one or more words of interest and one or more contexts, that a sentence within the electronic document includes a potential word of interest by identifying at least one context of the one or more contexts in which the potential word of interest is a word of interest;

determining, by evaluating the computer data structure using the WSD engine and a deep neural network, a particular context of the potential word of interest using the knowledge base;

determining, based upon the particular context, that the potential word of interest is the word of interest;

rewriting the sentence, using the mitigation engine and a large language model, to generate a revised sentence that does not include the word of interest;

determining, using a similarity engine to compare a contextual meaning associated with the sentence and the revised sentence, that the revised sentence is sufficiently similar to the sentence;

determining that the revised sentence does not include any other word of interest; and

updating the electronic document to include the revised sentence responsive to determining that the revised sentence is sufficiently similar to the sentence and responsive to determining that the revised sentence does not include the any other word of interest.

2 . The method of claim 1 , wherein

the deep neural network includes an Adaptive Skip-gram (AdaGram) model.

3 . The method of claim 1 , wherein

a graphical user interface is presented to a user, and

the graphical user interface is configured to prompt the user to generate the revised sentence.

4 . The method of claim 1 , wherein

the revised sentence is generated using a modification template.

5 . The method of claim 4 , wherein

the modification template is automatically selected by the mitigation engine.

6 . The method of claim 4 , wherein

a graphical user interface is presented to a user, and

the graphical user interface is configured to provide word level explainability regarding the sentence.

7 . The method of claim 4 , wherein

a second modification template is used to generate the revised sentence after an initial revised sentence was determined to meet a similarity evaluation.

8 . The method of claim 4 , wherein

a second modification template is used to generate the revised sentence after an initial revised sentence was determined to contain other words of interest.

9 . A computer hardware system for updating an electronic document, comprising:

a hardware processor including a preprocessor, a word sense disambiguation (WSD) engine, and a mitigation engine configured to perform the following executable operations:

performing preprocessing on the electronic document, using the preprocessor, to generate a computer data structure;

identifying, using a knowledge base that stores a mapping between one or more words of interest and one or more contexts, that a sentence within the electronic document includes a potential word of interest by identifying at least one context of the one or more contexts in which the potential word of interest is a word of interest;

determining, by evaluating the computer data structure using the WSD engine and a deep neural network, a particular context of the potential word of interest using the knowledge base;

determining, based upon the particular context, that the potential word of interest is the word of interest;

rewriting the sentence, using the mitigation engine and a large language model, to generate a revised sentence that does not include the word of interest;

determining, using a similarity engine to compare a contextual meaning associated with the sentence and the revised sentence, that the revised sentence is sufficiently similar to the sentence;

determining that the revised sentence does not include any other word of interest; and

updating the electronic document to include the revised sentence responsive to determining that the revised sentence is sufficiently similar to the sentence and responsive to determining that the revised sentence does not include the any other word of interest.

10 . The system of claim 9 , wherein

the deep neural network includes an Adaptive Skip-gram (AdaGram) model.

11 . The system of claim 9 , wherein

a graphical user interface is presented to a user, and

the graphical user interface is configured to prompt the user to generate the revised sentence.

12 . The system of claim 9 , wherein

the revised sentence is generated using a modification template.

13 . The system of claim 12 , wherein

the modification template is automatically selected by the mitigation engine.

14 . The system of claim 12 , wherein

a graphical user interface is presented to a user, and

the graphical user interface is configured to provide word level explainability regarding the sentence.

15 . The system of claim 12 , wherein

a second modification template is used to generate the revised sentence after an initial revised sentence was determined to meet a similarity evaluation.

16 . The system of claim 12 , wherein

a second modification template is used to generate the revised sentence after an initial revised sentence was determined to contain other words of interest.

17 . A computer program product, comprising:

a computer readable storage medium having stored therein program code for updating an electronic document,

the program code, which when executed by a computer hardware system including a preprocessor, a word sense disambiguation (WSD) engine, and a mitigation engine, cause the computer hardware system to perform:

performing preprocessing on the electronic document, using the preprocessor, to generate a computer data structure;

identifying, using a knowledge base that stores a mapping between one or more words of interest and one or more contexts, that a sentence within the electronic document includes a potential word of interest by identifying at least one context of the one or more contexts in which the potential word of interest is a word of interest;

determining, by evaluating the computer data structure using the WSD engine and a deep neural network, a particular context of the potential word of interest using the knowledge base;

determining, based upon the particular context, that the potential word of interest is the word of interest;

rewriting the sentence, using the mitigation engine and a large language model, to generate a revised sentence that does not include the word of interest;

determining, using a similarity engine to compare a contextual meaning associated with the sentence and the revised sentence, that the revised sentence is sufficiently similar to the sentence;

determining that the revised sentence does not include any other word of interest; and

updating the electronic document to include the revised sentence responsive to determining that the revised sentence is sufficiently similar to the sentence and responsive to determining that the revised sentence does not include the any other word of interest.

18 . The computer program product of claim 17 , wherein

the deep neural network includes an Adaptive Skip-gram (AdaGram) model.

19 . The computer program product of claim 17 , wherein

the revised sentence is generated using a modification template.

20 . The computer program product of claim 17 , wherein

a second modification template is used to generate the revised sentence after an initial revised sentence was determined to:

meet a similarity evaluation or

contain other words of interest.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Sep 30, 2023
From: REIS ALVES, RODRIGO; MOORE, ANGELO; SALUSTINO GUIMARAES, VALDIR; ARRIGONI, DANIELA; GOPAL, VASANTHI M.
To: INTERNATIONAL BUSINESS MACHINES CORPORATION
Reel/Frame 065083/0176 →
Continuity (1)
Related Publication 20250111160A1 · Apr 3, 2025
References Cited (42)
US 10594757B1 · Shevchenko · 2020 [cited by examiner]
US 10635750B1 · Epstein et al. · 2020 [cited by applicant]
US 10861439B2 · Doyle et al. · 2020 [cited by applicant]
US 11610061B2 · Eisenschlos · 2023 [cited by examiner]
US 12271697B2 · van Dam · 2025 [cited by examiner]
US 20050049852A1 · Chao · 2005 [cited by applicant]
US 20140358519A1 · Mirkin · 2014 [cited by examiner]
US 20170286376A1 · Mugan · 2017 [cited by examiner]
US 20190303435A1 · Herr · 2019 [cited by examiner]
US 20210004432A1 · Li et al. · 2021 [cited by applicant]
US 20210073224A1 · Zhao · 2021 [cited by examiner]
US 20210149996A1 · Bellegarda · 2021 [cited by applicant]
US 20210294974A1 · Gao et al. · 2021 [cited by applicant]
US 20220058339A1 · Archuleta · 2022 [cited by examiner]
US 20220100962A1 · Akhalwaya et al. · 2022 [cited by applicant]
US 20220121879A1 · Goyal · 2022 [cited by examiner]
US 20220215047A1 · Banipal et al. · 2022 [cited by applicant]
US 20220405482A1 · Tam · 2022 [cited by applicant]
US 20230137209A1 · Nangi · 2023 [cited by examiner]
US 20230153546A1 · Peleg · 2023 [cited by examiner]
US 20240303247A1 · Sokolov · 2024 [cited by examiner]
US 20250053725A1 · Zhou · 2025 [cited by examiner]
WO 2023059561A1 · 2023 [cited by applicant]
WO 2025068001A1 · 2025 [cited by applicant]
Inglesias et al. “A Toxic Style Transfer Method Based on the Delete-Retrieve-Generate Framework Exploiting Toxic Semantic Similarity”. Appl. Sci. 2023, 13(15), 8590, Published Jul. 26, 2023 (Year: 2023). [cited by examiner]
Lopukhina et al. Regular Polysemy: from sense vectors to sense patterns. Proceedings of the Workshop on Cognitive Aspects of the Lexicon, pp. 19-23, Osaka, Japan, Dec. 11-17, 2016 (Year: 2016). [cited by examiner]
Logacheva et al. “ParaDetox: Detoxification with Parallel Data”. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics vol. 1: Long Papers, pp. 6804-6818 May 22-27, 2022 (Year: 2022). [cited by examiner]
Anand et al. “Context Aware Query Rewriting for Text Rankers using LLM”. arXiv:2308.16753v1 [cs.IR] Aug. 31, 2023 (Year: 2023). [cited by examiner]
Bott et al. “Can Spanish Be Simpler? LexSIS: Lexical Simplification for Spanish” Proceedings of COLING 2012: Technical Papers, pp. 357-374 (Year: 2012). [cited by examiner]
Feng et al. “Sentence Simplification via Large Language Models”. arXiv:2302.11957v1 [cs.CL] Feb. 23, 2023 (Year: 2023). [cited by examiner]
Lu et al. “Facilitating Fine-grained Detection of Chinese Toxic Language: Hierarchical Taxonomy, Resources, and Benchmarks”. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics vol. 1… [cited by examiner]
Francesca et al. “Beyond word embeddings: A survey”, Information Fusion, Jan. 1, 2023, pp. 418-436, vol. 89, Issue C. [cited by applicant]
Gupta et al. “Improving Document Classification with Multi-Sense Embeddings ”, arXiv:1911.07918 [cs.CL], Nov. 18, 2019, 8 pages. [cited by applicant]
International Searching Authority, “Notification of Transmittal of the International Search Report and the Written Opinion of the International Searching Authority, or Declaration,” Patent Cooperation Treaty, Dec. 13, 2… [cited by applicant]
Ayetiran, E.F. et al., “An optimized Lesk-based algorithm for word sense disambiguation,” Open Computer Science, Aug. 1, 2016, vol. 8, No. 1, pp. 165-172. [cited by applicant]
Huang, L. et al., “GlossBERT: BERT for word sense disambiguation with gloss knowledge,” arXiv preprint, arXiv:1908.07245, Aug. 20, 2019, 6 pg. [cited by applicant]
Blevins, T. et al., “Moving down the long tail of word sense disambiguation with gloss-informed biencoders,” arXiv preprint, arXiv:2005.02590, May 6, 2020, 12 pg. [cited by applicant]
Scarlini, B. et al., “With More Contexts Comes Better Performance: Contextualized Sense Embeddings for All-Round Word Sense Disambiguation,” In Proceedings of the 2020 Conference on Empirical Methods in Natural Language… [cited by applicant]
Bartunov, S. et al., “Breaking sticks and ambiguities with adaptive skip-gram,” In Artificial Intelligence and Statistics, May 2, 2016, pp. 130-138, PMLR. [cited by applicant]
Ustalov, D. et al., “An unsupervised word sense disambiguation system for under-resourced languages,” arXiv preprint, arXiv:1804.10686, Apr. 27, 2018, 5 pg. [cited by applicant]
Mell, P. et al., The NIST Definition of Cloud Computing, National Institute of Standards and Technology, U.S. Dept. of Commerce, Special Publication 800-145, Sep. 2011, 7 pg. [cited by applicant]
Bevilacqua, M. et al., “Breaking through the 80% glass ceiling: Raising the state of the art in word sense disambiguation by incorporating knowledge graph information,” In Proceedings of the 58th Annual Meeting of the A… [cited by applicant]