Thai Le
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Thai Le has a total of 107 co-authors. Here are the top co-authors by number of shared publications:
- Dongwon Lee
The Pennsylvania State University, University Park, PA, USA
4 shared publications - Jooyoung Lee
Penn State University, USA
2 shared publications - Adaku Uchendu
Massachusetts Institute of Technology Lincoln Laboratory, Lexington, MA, USA
1 shared publication - Aghiles Salah
Huawei Ireland Research Center, Dublin, Ireland
1 shared publication - Amit Sheth
University of South Carolina
1 shared publication - Arvinder Kang
University of Mississippi, Oxford, Mississippi, USA
1 shared publication
Latest Publications
- 2 July 2024
- Published by Springer Nature
- p. 468-481
The Unexpected Effects of Google Smart Compose on Open-Ended Writing Tasks
- 2 July 2024
- Published by Springer Nature
- p. 468-481
The Unexpected Effects of Google Smart Compose on Open-Ended Writing Tasks
- Open Access
- 18 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 278-283
Toshakhana: A Multidimensional Panjabi Corpus in Gurmukhi Script
Panjabi (also referred to as Punjabi) is a name given to a collection of tonal languages originating in the Punjab area of South Asia. It is the ninth most spoken language in the world - roughly 1.9% of the world population. Panjabi is written in two scripts - Gurmukhi and Shahmukhi. Yet it can be considered a "low resource language" due to lack of basic building blocks of Natural Language Processing (NLP) research. Toshakhana is our attempt to build the first Panjabi corpus in Gurmukhi script with temporal component.
- 18 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 278-283
Open AccessToshakhana: A Multidimensional Panjabi Corpus in Gurmukhi Script
Panjabi (also referred to as Punjabi) is a name given to a collection of tonal languages originating in the Punjab area of South Asia. It is the ninth most spoken language in the world - roughly 1.9% of the world population. Panjabi is written in two scripts - Gurmukhi and Shahmukhi. Yet it can be considered a "low resource language" due to lack of basic building blocks of Natural Language Processing (NLP) research. Toshakhana is our attempt to build the first Panjabi corpus in Gurmukhi script with temporal component.
- Open Access
- 18 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 259-265
Benchmarking Machine Learning Techniques for Bankruptcy Prediction under Benign and Adversarial Behaviors
This research uses machine learning methods to perform company bankruptcy prediction. 10K Financial reports are audited by public auditing firms and submitted by companies to the Security Exchange Commission (SEC). 10K Financial reports are the appropriate and reasonable financial statements including financial variables (numbers) and texts to evaluate a company's financial status yearly. However, people who don't have a business background would find it difficult to read and understand the 10K Financial reports. Therefore, it is useful to have machine learning methods to analyze the financial status or predict bankruptcy on 10K Financial reports datasets. This research uses different time frames 10K Financial reports datasets and compares several machine learning models to classify and predict bankruptcy companies. Based on different dataset's experimental results, this research concluded that the Random Forest model or XGBoost model generates a better performance. The outcomes of comparing adversarial model performance showed that MLP (Multi-layer Perceptron) + FGSM (Fast Gradient Sign Method) generate better performance, dropping the prediction performance in F1 score from 81.63% to 79.01%.
- 18 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 259-265
Open AccessBenchmarking Machine Learning Techniques for Bankruptcy Prediction under Benign and Adversarial Behaviors
This research uses machine learning methods to perform company bankruptcy prediction. 10K Financial reports are audited by public auditing firms and submitted by companies to the Security Exchange Commission (SEC). 10K Financial reports are the appropriate and reasonable financial statements including financial variables (numbers) and texts to evaluate a company's financial status yearly. However, people who don't have a business background would find it difficult to read and understand the 10K Financial reports. Therefore, it is useful to have machine learning methods to analyze the financial status or predict bankruptcy on 10K Financial reports datasets. This research uses different time frames 10K Financial reports datasets and compares several machine learning models to classify and predict bankruptcy companies. Based on different dataset's experimental results, this research concluded that the Random Forest model or XGBoost model generates a better performance. The outcomes of comparing adversarial model performance showed that MLP (Multi-layer Perceptron) + FGSM (Fast Gradient Sign Method) generate better performance, dropping the prediction performance in F1 score from 81.63% to 79.01%.
- Open Access
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 648-658
Tutorials at The Web Conference 2023
This paper summarizes the content of the 28 tutorials that have been given at The Web Conference 2023.
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 648-658
Open AccessTutorials at The Web Conference 2023
This paper summarizes the content of the 28 tutorials that have been given at The Web Conference 2023.
- Open Access
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 3637-3647
Do Language Models Plagiarize?
Past literature has illustrated that language models (LMs) often memorize parts of training instances and reproduce them in natural language generation (NLG) processes. However, it is unclear to what extent LMs “reuse” a training corpus. For instance, models can generate paraphrased sentences that are contextually similar to training samples. In this work, therefore, we study three types of plagiarism (i.e., verbatim, paraphrase, and idea) among GPT-2 generated texts, in comparison to its training data, and further analyze the plagiarism patterns of fine-tuned LMs with domain-specific corpora which are extensively used in practice. Our results suggest that (1) three types of plagiarism widely exist in LMs beyond memorization, (2) both size and decoding methods of LMs are strongly associated with the degrees of plagiarism they exhibit, and (3) fine-tuned LMs’ plagiarism patterns vary based on their corpus similarity and homogeneity. Given that a majority of LMs’ training data is scraped from the Web without informing content owners, their reiteration of words, phrases, and even core ideas from training sets into generated texts has ethical implications. Their patterns are likely to exacerbate as both the size of LMs and their training data increase, raising concerns about indiscriminately pursuing larger models with larger training corpora. Plagiarized content can also contain individuals’ personal and sensitive information. These findings overall cast doubt on the practicality of current LMs in mission-critical writing tasks and urge more discussions around the observed phenomena. Data and source code are available at https://github.com/Brit7777/LM-plagiarism.
- 30 April 2023
- Published by Association for Computing Machinery (ACM)
- p. 3637-3647
Open AccessDo Language Models Plagiarize?
Past literature has illustrated that language models (LMs) often memorize parts of training instances and reproduce them in natural language generation (NLG) processes. However, it is unclear to what extent LMs “reuse” a training corpus. For instance, models can generate paraphrased sentences that are contextually similar to training samples. In this work, therefore, we study three types of plagiarism (i.e., verbatim, paraphrase, and idea) among GPT-2 generated texts, in comparison to its training data, and further analyze the plagiarism patterns of fine-tuned LMs with domain-specific corpora which are extensively used in practice. Our results suggest that (1) three types of plagiarism widely exist in LMs beyond memorization, (2) both size and decoding methods of LMs are strongly associated with the degrees of plagiarism they exhibit, and (3) fine-tuned LMs’ plagiarism patterns vary based on their corpus similarity and homogeneity. Given that a majority of LMs’ training data is scraped from the Web without informing content owners, their reiteration of words, phrases, and even core ideas from training sets into generated texts has ethical implications. Their patterns are likely to exacerbate as both the size of LMs and their training data increase, raising concerns about indiscriminately pursuing larger models with larger training corpora. Plagiarized content can also contain individuals’ personal and sensitive information. These findings overall cast doubt on the practicality of current LMs in mission-critical writing tasks and urge more discussions around the observed phenomena. Data and source code are available at https://github.com/Brit7777/LM-plagiarism.
- Open Access
Perturbations in the Wild: Leveraging Human-Written Text Perturbations for Realistic Adversarial Attack and Defense
We proposes a novel algorithm, ANTHRO, that inductively extracts over 600K human-written text perturbations in the wild and leverages them for realistic adversarial attack. Unlike existing character-based attacks which often deductively hypothesize a set of manipulation strategies, our work is grounded on actual observations from real-world texts. We find that adversarial texts generated by ANTHRO achieve the best trade-off between (1) attack success rate, (2) semantic preservation of the original text, and (3) stealthiness--i.e. indistinguishable from human writings hence harder to be flagged as suspicious. Specifically, our attacks accomplished around 83% and 91% attack success rates on BERT and RoBERTa, respectively. Moreover, it outperformed the TextBugger baseline with an increase of 50% and 40% in terms of semantic preservation and stealthiness when evaluated by both layperson and professional human workers. ANTHRO can further enhance a BERT classifier's performance in understanding different variations of human-written toxic texts via adversarial training when compared to the Perspective API.
Open AccessPerturbations in the Wild: Leveraging Human-Written Text Perturbations for Realistic Adversarial Attack and Defense
We proposes a novel algorithm, ANTHRO, that inductively extracts over 600K human-written text perturbations in the wild and leverages them for realistic adversarial attack. Unlike existing character-based attacks which often deductively hypothesize a set of manipulation strategies, our work is grounded on actual observations from real-world texts. We find that adversarial texts generated by ANTHRO achieve the best trade-off between (1) attack success rate, (2) semantic preservation of the original text, and (3) stealthiness--i.e. indistinguishable from human writings hence harder to be flagged as suspicious. Specifically, our attacks accomplished around 83% and 91% attack success rates on BERT and RoBERTa, respectively. Moreover, it outperformed the TextBugger baseline with an increase of 50% and 40% in terms of semantic preservation and stealthiness when evaluated by both layperson and professional human workers. ANTHRO can further enhance a BERT classifier's performance in understanding different variations of human-written toxic texts via adversarial training when compared to the Perspective API.
- Open Access
- 23 June 2021
- Published by Mathematical Sciences Publishers in Moscow Journal of Combinatorics and Number Theory
- Vol. 10 (2) , 149-165
Essential components in vector spaces over finite fields
- 23 June 2021
- Published by Mathematical Sciences Publishers in Moscow Journal of Combinatorics and Number Theory
- Vol. 10 (2) , 149-165
Open AccessEssential components in vector spaces over finite fields
- Open Access
- 1 January 2021
- Published by Association for Computational Linguistics (ACL)
- p. 3831-3844
A Sweet Rabbit Hole by DARCY: Using Honeypots to Detect Universal Trigger’s Adversarial Attacks
Thai Le, Noseong Park, Dongwon Lee. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.
- 1 January 2021
- Published by Association for Computational Linguistics (ACL)
- p. 3831-3844
Open AccessA Sweet Rabbit Hole by DARCY: Using Honeypots to Detect Universal Trigger’s Adversarial Attacks
Thai Le, Noseong Park, Dongwon Lee. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.