Papers by Chengyu Dong
Revisiting and Advancing Chinese Natural Language Understanding with Accelerated Heterogeneous Knowledge Pre-training (2022.emnlp-industry)
Copied to clipboard
Taolin Zhang, Junwei Dong, Jianing Wang, Chengyu Wang, Ang Wang, Yinghui Liu, Jun Huang, Yong Li, Xiaofeng He
| Challenge: | Existing knowledge-enhanced pre-trained language models (KEPLMs) can capture internal knowledge, but can't understand external background knowledge. |
| Approach: | They propose to use Chinese knowledge-enhanced pre-trained language models to improve context-aware representations via learning from structured relations in knowledge bases. |
| Outcome: | Experiments show that Chinese knowledge-enhanced pre-trained language models outperform strong baselines over various benchmark NLP tasks and in different model sizes. |
Debiasing Made State-of-the-art: Revisiting the Simple Seed-based Weak Supervision for Text Classification (2023.emnlp-main)
Copied to clipboard
| Challenge: | Recent advances in weakly supervised text classification focus on designing sophisticated methods to turn high-level human heuristics into quality pseudo-labels. |
| Approach: | They propose to use a seed matching-based method to generate quality pseudo-labels by deleting the seed words present in the matched input text. |
| Outcome: | The proposed method can be improved significantly by deleting the seed words in the matched input text with a high deletion ratio. |
Text Grafting: Near-Distribution Weak Supervision for Minority Classes in Text Classification (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent work generates pseudo labels by mining texts similar to the class names from the raw corpus, but there is a high risk that LLMs cannot generate in-distribution data, leading to ungeneralizable classifiers. |
| Approach: | They propose to use LLMs to generate pseudo labels by mining masked templates from corpus . they then use state-of-the-art LLM to synthesize near-distribution texts falling into minority classes . |
| Outcome: | The proposed framework improves on the previous methods for extremely weak-supervised text classification. |
Can Textual Unlearning Solve Cross-Modality Safety Alignment? (2024.findings-emnlp)
Copied to clipboard
Trishna Chakraborty, Erfan Shayegani, Zikui Cai, Nael Abu-Ghazaleh, M. Salman Asif, Yue Dong, Amit Roy-Chowdhury, Chengyu Song
| Challenge: | integrating new modalities into large language models creates new attack surface . existing safety training techniques like SFT and RLHF are not feasible in multi-modal settings . |
| Approach: | They explore whether unlearning in the textual domain can be effective for cross-modality safety alignment. |
| Outcome: | The proposed approach reduces the Attack Success Rate (ASR) to less than 8% and preserves the utility. |
BFClass: A Backdoor-free Text Classification Framework (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Various trigger design strategies have been explored to attack text classifiers, however, defending such attacks remains an open problem. |
| Approach: | They propose a backdoor-free training framework that poisons a subset of training data by injecting trigger patterns and setting their labels as the target labels. |
| Outcome: | The proposed framework can detect all the triggers, remove 95% of poisoned training samples with very limited false alarms, and achieve almost the same performance as the models trained on benign training data. |
HEAL: An Empirical Study on Hallucinations in Embodied Agents Driven by Large Language Models (2025.findings-emnlp)
Copied to clipboard
Trishna Chakraborty, Udita Ghosh, Xiaopan Zhang, Fahim Faisal Niloy, Yue Dong, Jiachen Li, Amit Roy-Chowdhury, Chengyu Song
| Challenge: | Large language models are increasingly being adopted as the cognitive core of embodied agents. |
| Approach: | They propose a systematic study of hallucinations in large language models . they aim to understand to what extent hallucinos occur, what types trigger them . |
| Outcome: | The proposed model can induce hallucinations up to 40 higher than base prompts . the model fails to resolve scene-task inconsistencies, the study finds . |
Evaluating the Smooth Control of Attribute Intensity in Text Generation with LLMs (2024.findings-acl)
Copied to clipboard
| Challenge: | Controllable text generation is increasingly tailored to individual preferences. |
| Approach: | They propose to evaluate the attribute intensity of text generated by large language models on five different attributes for error, variation of the generated sentence's intensities and relevance to the generation questions. |
| Outcome: | The proposed methods are based on Elo rating system and GPT4 and are able to be trained without training. |
SELFOOD: Self-Supervised Out-Of-Distribution Detection via Learning to Rank (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for supervised OOD detection require expensive manual annotation of in-distribution and OOD samples. |
| Approach: | They propose a self-supervised OOD detection method that requires only in-distribution samples as supervision. |
| Outcome: | Experiments with multiple classifiers on coarse- and fine-grained datasets show the proposed method performs well in both coarse-and fine-grid settings. |
“Average” Approximates “First Principal Component”? An Empirical Analysis on Representations from Neural Language Models (2021.emnlp-main)
Copied to clipboard
| Challenge: | Contextualized representations have been used in various NLP tasks, but their nature remains a mystery. |
| Approach: | They propose to use a property to estimate the power of contextualized representations . they show that the average representation shares almost the same direction as the first principal component . |
| Outcome: | The proposed representations share the same direction as the first principal component . the results suggest that the property is intrinsic to the distribution of representations . |
LOPS: Learning Order Inspired Pseudo-Label Selection for Weakly Supervised Text Classification (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Weakly-supervised text classification methods are noisy due to their heuristic nature . selection of correct pseudo-labels has a huge potential for performance boost . |
| Approach: | They propose a pseudo-label selection method that takes learning order into account . they propose to select samples that are learnt earlier based on their pseudo-labels . |
| Outcome: | The proposed method is ineffective and unstable due to erroneous predictions from poorly calibrated models. |