IP Library › Granted Patent US 12,488,179
Granted Patent B2
US 12,488,179 · App. 18/197,120 · Granted Dec 2, 2025

Quality controlled paraphrase generation

Inventors: Elron Bandel (Jerusalem, IL); Liat Ein-Dor (Tel-Aviv, IL); Ranit Aharonov (Ramat Hasharon, IL); Michal Shmueli-Scheuer (Tel-Aviv, IL); Ilya Shnayderman (Jerusalem, IL)
Assignee: International Business Machines Corporation
G06F40/166G06F40/284G06F40/30
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 12,488,179
App. No.
18/197,120
Granted
Dec 2, 2025
Kind
B2
Abstract

A computer-implemented method including: receiving, as input, a dataset comprising training pairs (s, t), wherein each training pair comprises (i) a source sentence s and (ii) a target paraphrase t of the source sentences; at a training stage, training a machine learning model on the dataset, to obtain a trained quality-controlled paraphrase generator model, wherein during the training stage, each of the training pairs is associated with a predicted control vector representing a predicted paraphrase quality of the source sentence in the training pair; and at an inference stage, inferencing the trained quality-controlled paraphrase generator model on an input sentence, wherein the input sentence is associated with an input quality control vector, to obtain an output paraphrase of the input sentence which conforms to the quality control vector.

Claims (41)

1 . A method comprising:

receiving, as input, a dataset comprising a plurality of training pairs, wherein each of the plurality of training pairs includes a source sentence and a target paraphrase of the source sentence;

training, at a training stage, a machine learning model on the dataset to generate a trained quality-controlled paraphrase generator model, wherein during the training stage each of the plurality of training pairs are associated with a predicted control vector representing a predicted paraphrase quality of the source sentence; and

inferencing, at an inference stage, the trained quality-controlled paraphrase generator model on an input sentence, wherein the input sentence includes an input quality control vector for controlling an output paraphrase, wherein the input quality control vector is a three-dimensional vector including three quality dimensions, a desired semantic similarity, a desired syntactic distance, and a desired lexical distance, and wherein requested quality values of the three quality dimensions are set by a user.

2 . The method of claim 1 , further comprising:

generating the input quality control vector, wherein the input quality control vector is a sum of an expected quality values and an offset vector, wherein the vector is a differential between the requested quality values set by the user and the expected quality values.

3 . The method of claim 2 , wherein the machine learning model includes a quality predictor configured to predict the expected quality values of paraphrases of a sentence.

4 . The method of claim 3 , wherein the quality predictor is a regressor configured to predict the expected quality values of the input sentence.

5 . The method of claim 2 , wherein the offset vector represents a sentence-independent level of difficulty generating paraphrases for the input sentence.

6 . The method of claim 1 , further comprising:

adjusting the requested quality values set by the user based on a predicted paraphrase quality associated with the input sentence.

7 . The method of claim 6 , wherein the adjustment assumes that a quality distribution of all paraphrases of the input sentence are normally distributed around a sentence-dependent mean and a variance is independent, and wherein, given the input sentence, a difficulty to generate a paraphrase of the requested quality values is dominated by the quality distribution rather than the requested quality values.

8 . A computer system comprising:

a processor set;

one or more computer-readable storage media; and

program instructions stored on the one or more computer-readable storage media to cause the processor set to perform operations comprising:

receiving, as input, a dataset comprising a plurality of training pairs, wherein each of the plurality of training pairs includes a source sentence and a target paraphrase of the source sentence;

training, at a training stage, a machine learning model on the dataset to generate a trained quality-controlled paraphrase generator model, wherein during the training stage each of the plurality of training pairs are associated with a predicted control vector representing a predicted paraphrase quality of the source sentence; and

inferencing, at an inference stage, the trained quality-controlled paraphrase generator model on an input sentence, wherein the input sentence includes an input quality control vector for controlling an output paraphrase, wherein the input quality control vector is a three-dimensional vector including three quality dimensions, a desired semantic similarity, a desired syntactic distance, and a desired lexical distance, and wherein requested quality values of the three quality dimensions are set by a user.

9 . The computer system of claim 8 , wherein the operations further comprise:

generating the input quality control vector, wherein the input quality control vector is a sum of an expected quality values and an offset vector, wherein the vector is a differential between the requested quality values set by the user and the expected quality values.

10 . The computer system of claim 9 , wherein the machine learning model includes a quality predictor configured to predict the expected quality values of paraphrases of a sentence.

11 . The computer system of claim 10 , wherein the quality predictor is a regressor configured to predict the expected quality values of the input sentence.

12 . The computer system of claim 9 , wherein the offset vector represents a sentence-independent level of difficulty generating paraphrases for the input sentence.

13 . The computer system of claim 8 , wherein the operations further comprise:

adjusting the requested quality values set by the user based on a predicted paraphrase quality associated with the input sentence.

14 . The computer system of claim 13 , wherein the adjustment assumes that a quality distribution of all paraphrases of the input sentence are normally distributed around a sentence-dependent mean and a variance is independent, and wherein, given the input sentence, a difficulty to generate a paraphrase of the requested quality values is dominated by the quality distribution rather than the requested quality values.

15 . A computer program product comprising:

one or more computer-readable storage media; and

program instructions stored on the one or more computer-readable storage media to perform operations comprising:

receiving, as input, a dataset comprising a plurality of training pairs, wherein each of the plurality of training pairs includes a source sentence and a target paraphrase of the source sentence;

training, at a training stage, a machine learning model on the dataset to generate a trained quality-controlled paraphrase generator model, wherein during the training stage each of the plurality of training pairs are associated with a predicted control vector representing a predicted paraphrase quality of the source sentence; and

inferencing, at an inference stage, the trained quality-controlled paraphrase generator model on an input sentence, wherein the input sentence includes an input quality control vector for controlling an output paraphrase, wherein the input quality control vector is a three-dimensional vector including three quality dimensions, a desired semantic similarity, a desired syntactic distance, and a desired lexical distance, and wherein requested quality values of the three quality dimensions are set by a user.

16 . The computer program product of claim 15 , wherein the operations further comprise:

generating the input quality control vector, wherein the input quality control vector is a sum of an expected quality values and an offset vector, wherein the vector is a differential between the requested quality values set by the user and the expected quality values.

17 . The computer program product of claim 16 , wherein the machine learning model includes a quality predictor configured to predict the expected quality values of paraphrases of a sentence.

18 . The computer program product of claim 17 , wherein the quality predictor is a regressor configured to predict the expected quality values of the input sentence.

19 . The computer program product of claim 16 , wherein the offset vector represents a sentence-independent level of difficulty generating paraphrases for the input sentence.

20 . The computer program product of claim 15 , wherein the operations further comprise:

adjusting the requested quality values set by the user based on a predicted paraphrase quality associated with the input sentence.

21 . The computer program product of claim 20 , wherein the adjustment assumes that a quality distribution of all paraphrases of the input sentence are normally distributed around a sentence-dependent mean and a variance is independent, and wherein, given the input sentence, a difficulty to generate a paraphrase of the requested quality values is dominated by the quality distribution rather than the requested quality values.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded May 15, 2023
From: BANDEL, ELRON; EIN-DOR, LIAT; AHARONOV, RANIT; SHMUELI-SCHEUER, MICHAL; SHNAYDERMAN, ILYA
To: INTERNATIONAL BUSINESS MACHINES CORPORATION
Reel/Frame 063636/0889 →
Continuity (1)
Related Publication 20240386188A1 · Nov 21, 2024
References Cited (31)
US 11042712B2 · Prakash · 2021 [cited by applicant]
US 11328017B2 · Gupta · 2022 [cited by applicant]
US 20040088156A1 · Kanagasabai · 2004 [cited by examiner]
US 20180061408A1 · Andreas · 2018 [cited by applicant]
US 20180329882A1 · Bennett · 2018 [cited by examiner]
US 20190042663A1 · Gupta · 2019 [cited by examiner]
US 20200250377A1 · Jiang · 2020 [cited by applicant]
US 20220067285A1 · Sellam · 2022 [cited by examiner]
A. Chitra and A. Rajkumar, “Evolutionary approach for building efficient Paraphrase Recognizers,” 2011 World Congress on Information and Communication Technologies, Mumbai, India, 2011, pp. 661-666. (Year: 2011). [cited by examiner]
X. Zhang, W. Rong, J. Liu, C. Tian and Z. Xiong, “Convolution neural network based syntactic and semantic aware paraphrase identification,” 2017 International Joint Conference on Neural Networks (IJCNN), Anchorage, AK, … [cited by examiner]
Ankush Gupta et al, “A Deep Generative Framework for Paraphrase Generation”; Online at: https://arxiv.org/pdf/1709.05074.pdf, Sep. 15, 2017. [cited by applicant]
Elron Bandel et al, “Quality Controlled Paraphrase Generation”; Online at: https://arxiv.org/abs/2203.10940, Apr. 1, 2022. [cited by applicant]
Mingda Chen et al, “Controllable Paraphrase Generation with a Syntactic Exemplar”; In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 5972-5984, Jul. 2019. [cited by applicant]
Sonal Garg et al, “Unsupervised Contextual Paraphrase Generation using Lexical Control and Reinforcement Learning”; Online at: https://arxiv.org/abs/2103.12777, Mar. 23, 2021. [cited by applicant]
Tanya Goyal et al, “Neural Syntactic Preordering for Controlled Paraphrase Generation”; In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 238-252, Jul. 5-10, 2020. [cited by applicant]
Bao et al., “Generating Sentences from Disentangled Syntactic and Semantic Spaces”, arXiv:1907.05789v1 [cs.CL], Jul. 6, 2019, 11 pages. [cited by applicant]
Callison-Burch et al., “ParaMetric: An Automatic Evaluation Metric for Paraphrasing”, Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), Aug. 2008, pp. 97-104. [cited by applicant]
Clark et al., “Electra: Pre-training Text Encoders as Discriminators Rather than Generators”, arXiv:2003.10555v1 [cs.CL], Mar. 23, 2020, 18 pages. [cited by applicant]
Hosking et al., “Factorising Meaning and Form for Intent-Preserving Paraphrasing”, arXiv:2105.15053v1 [cs.CL], May 31, 2021, 14 pages. [cited by applicant]
Iyyer et al., “Adversarial Example Generation with Syntactically Controlled Paraphrase Networks”, arXiv:1804.06059v1 [cs.CL], Apr. 17, 2018, 11 pages. [cited by applicant]
Li et al., “Transformers to Learn Hierarchical Contexts in Multiparty Dialogue for Span-based Question Answering”, Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Jul. 2020, pp. … [cited by applicant]
Liu et al., “Unsupervised Paraphrasing by Simulated Annealing”, roceedings of the 58th Annual Meeting of the Association for Computational Linguistics, Jul. 5-10, 2020, pp. 302-312. [cited by applicant]
Nicula et al., “Automated Paraphrase Quality Assessment Using Language Models and Transfer Learning”, Computers, Dec. 6, 2021, 16 pages. [cited by applicant]
Niu et al., “Unsupervised Paraphrasing with Pretrained Language Models”, arXiv:2010.12885v2 [cs.CL], Sep. 10, 2021, 15 pages. [cited by applicant]
Park et al., “Paraphrase Diversification Using Counterfactual Debiasing”, AAAI Technical Track: Natural Language Processing, Jul. 17, 2019, pp. 6883-6891. [cited by applicant]
Qian et al., “Exploring Diverse Expressions for Paraphrase Generation”, Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Langua… [cited by applicant]
Raffel et al., “Exploring the limits of transfer learning with a unified text-to-text transformer”, arXiv:1910.10683v4 [cs.LG], Sep. 19, 2023, 67 pages. [cited by applicant]
Siddique et al., “Unsupervised Paraphrasing via Deep Reinforcement Learning”, arXiv:2007.02244 [cs.CL], Jul. 5, 2020, 10 pages. [cited by applicant]
Thompson et al., “Paraphrase Generation as Zero-Shot Multilingual Translation: Disentangling Semantic Similarity from Lexical and Syntactic Diversity”, arXiv:2008.04935v2 [cs.CL], Oct. 28, 2020, 10 pages. [cited by applicant]
Vijayakumar et al., “Diverse Beam Search for Improved Description of Complex Scenes”, Thirty-Second AAAI Conference on Artificial Intelligence, Apr. 27, 2018, pp. 7371-7379. [cited by applicant]
Zeng et al., “User-Oriented Paraphrase Generation With Keywords Controlled Network”, IEEE Access, Jun. 14, 2019, pp. 80542-80551. [cited by applicant]