How can Large Language Models (LLMs) be aligned with human intentions and values? A typical solution is to gather human preference on model outputs and finetune the LLMs accordingly while ensuring that updates do not deviate too far from a reference model. Recent approaches, such as direct preference optimization (DPO), have eliminated the need for unstable and sluggish reinforcement learning optimization by introducing close-formed supervised losses. However, a significant limitation of the current approach is its design for a single reference model only, neglecting to leverage the collective power of numerous pretrained LLMs. To overcome this limitation, we introduce a novel closed-form formulation for direct preference optimization using multiple reference models. The resulting algorithm, Multi-Reference Preference Optimization (MRPO), leverages broader prior knowledge from diverse reference models, substantially enhancing preference learning capabilities compared to the single-reference DPO. Our experiments demonstrate that LLMs finetuned with MRPO generalize better in various preference data, regardless of data scarcity or abundance. Furthermore, MRPO effectively finetunes LLMs to exhibit superior performance in several downstream natural language processing tasks such as HH-RLHF, GSM8K and TruthfulQA.

Data-Free Knowledge Distillation (DFKD) has made significant recent strides by transferring knowledge from a teacher neural network to a student neural network without accessing the original data. Nonetheless, existing approaches encounter a significant challenge when attempting to generate samples from random noise inputs, which inherently lack meaningful information. Consequently, these models struggle to effectively map this noise to the ground-truth sample distribution, resulting in prolonging training times and low-quality outputs. In this paper, we propose a novel Noisy Layer Generation method (NAYER) which re-locates the random source from the input to a noisy layer and utilizes the meaningful constant label-text embedding (LTE) as the input. LTE is generated by using the language model once, and then it is stored in memory for all subsequent training processes. The significance of LTE lies in its ability to contain substantial meaningful inter-class information, enabling the generation of high-quality samples with only a few training steps. Simultaneously, the noisy layer plays a key role in addressing the issue of diversity in sample generation by preventing the model from overemphasizing the constrained label information. By reinitializing the noisy layer in each iteration, we aim to facilitate the generation of diverse samples while still retaining the method's efficiency, thanks to the ease of learning provided by LTE. Experiments carried out on multiple datasets demonstrate that our NAYER not only outperforms the state-of-the-art methods but also achieves speeds 5 to 15 times faster than previous approaches. The code is available at https://github.com/tmtuan1307/nayer.

Aspect-based meeting transcript summarization aims to produce multiple summaries, each focusing on one aspect of content in a meeting transcript. It is challenging as sentences related to different aspects can mingle together, and those relevant to a specific aspect can be scattered throughout the long transcript of a meeting. The traditional summarization methods produce one summary mixing information of all aspects, which cannot deal with the above challenges of aspect-based meeting transcript summarization. In this paper, we propose a two-stage method for aspect-based meeting transcript summarization. To select the input content related to specific aspects, we train a sentence classifier on a dataset constructed from the AMI corpus with pseudo-labeling. Then we merge the sentences selected for a specific aspect as the input for the summarizer to produce the aspect-based summary. Experimental results on the AMI corpus outperform many strong baselines, which verifies the effectiveness of our proposed method.

Digital documents often contain images and scanned text. Parsing such visually-rich documents is a core task for work-flow automation, but it remains challenging since most documents do not encode explicit layout information, e.g., how characters and words are grouped into boxes and ordered into larger semantic entities. Current state-of-the-art layout extraction methods are challenged by such documents as they rely on word sequences to have correct reading order and do not exploit their hierarchical structure. We propose LayerDoc, an approach that uses visual features, textual semantics, and spatial coordinates along with constraint inference to extract the hierarchical layout structure of documents in a bottom-up layer-wise fashion. LayerDoc recursively groups smaller regions into larger semantic elements in 2D to infer complex nested hierarchies. Experiments show that our approach outperforms competitive baselines by 10-15% on three diverse datasets of forms and mobile app screen layouts for the tasks of spatial region classification, higher-order group identification, layout hierarchy extraction, reading order detection, and word grouping.

It is well-known that deep neural networks (DNNs) are susceptible to adversarial attacks, which presents the most severe fragility of the deep learning system. Despite achieving impressive performance, most of the current state-of-the-art classifiers remain highly vulnerable to carefully crafted imperceptible, adversarial perturbations. Recent research attempts to understand neural network attack and defense have become increasingly urgent and important. While rapid progress has been made on this front, there is still an important theoretical gap in achieving guaranteed bounds on attack/defense models, leaving uncertainty in the quality and certified guarantees of these models. To this end, we systematically address this problem in this paper. More specifically, we formulate attack and defense in a generic setting where there exists a family of adversaries (i.e., attackers) for attacking a family of classifiers (i.e., defenders). We develop a novel class of f-divergences suitable for measuring divergence among multiple distributions. This equips us to study the interactions between attackers and defenders in a countervailing game where we formulate a joint risk on attack and defense schemes. This is followed by our key results on guaranteed upper and lower bounds on this risk that can provide a better understanding of the behaviors of those parties from the attack and defense perspectives, thereby having important implications to both attack and defense sides. Finally, benefited from our theory, we propose an empirical approach that bases on a global view to defend against adversarial attacks. The experimental results conducted on benchmark datasets show that the global view for attack/defense if exploited appropriately can help to improve adversarial robustness.

Keyphrase Prediction (KP) is an established NLP task, aiming to yield representative phrases to summarize the main content of a given document. Despite major progress in recent years, existing works on KP have mainly focused on formal texts such as scientific papers or weblogs. The challenges of KP in informal-text domains are not yet fully studied. To this end, this work studies new challenges of KP in transcripts of videos, an understudied domain for KP that involves informal texts and non-cohesive presentation styles. A bottleneck for KP research in this domain involves the lack of high-quality and large-scale annotated data that hinders the development of advanced KP models. To address this issue, we introduce a large-scale manually-annotated KP dataset in the domain of live-stream video transcripts obtained by automatic speech recognition tools. Concretely, transcripts of 500+ hours of videos streamed on the behance.net platform are manually labeled with important keyphrases. Our analysis of the dataset reveals the challenging nature of KP in transcripts. Moreover, for the first time in KP, we demonstrate the idea of improving KP for long documents (i.e., transcripts) by feeding models with paragraph-level keyphrases, i.e., hierarchical extraction. To foster future research, we will publicly release the dataset and code.

Puneet Mathur, Vlad Morariu, Verena Kaynig-Fittkau, Jiuxiang Gu, Franck Dernoncourt, Quan Tran, Ani Nenkova, Dinesh Manocha, Rajiv Jain. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Unsupervised domain adaptation (UDA) aims to transfer knowledge from a model trained on a labeled source domain to an unlabeled target domain. To this end, we propose in this paper a novel cycle class-consistent model based on optimal transport (OT) and knowledge distillation. The model consists of two agents, a teacher and a student cooperatively working in a cycle process under the guidance of the distributional optimal transport and distillation manner. The OT distance is designed to bridge the gap between the distribution of the target data and a distribution over the source class-conditional distributions. The optimal probability matrix then provides pseudo labels to learn a teacher that achieves a good classification performance on the target domain. Knowledge distillation is performed in the next step in which the teacher distills and transfers its knowledge to the student. And finally, the student produces its prediction for the optimal transport step. This process forms a closed cycle in which the teacher and student networks are simultaneously trained to conduct transfer learning from the source to the target domain. Extensive experiments show that our proposed method outperforms existing methods, especially the class-aware and OT-based ones on benchmark datasets including Office-31, Office-Home, and ImageCLEF-DA.

Tuan Lai, Heng Ji, ChengXiang Zhai, Quan Hung Tran. 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.

Jianguo Zhang, Trung Bui, Seunghyun Yoon, Xiang Chen, Zhiwei Liu, Congying Xia, Quan Hung Tran, Walter Chang, Philip Yu. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Multi-source Domain Adaptation (MSDA) is more practical but challenging than the conventional unsupervised domain adaptation due to the involvement of diverse multiple data sources. Two fundamental challenges of MSDA are: (i) how to deal with the diversity in the multiple source domains and (ii) how to cope with the data shift between the target domain and the source domains. In this paper, to address the first challenge, we propose a theoretical-guaranteed approach to combine domain experts locally trained on its own source domain to achieve a combined multi-source teacher that globally predicts well on the mixture of source domains. To address the second challenge, we propose to bridge the gap between the target domain and the mixture of source domains in the latent space via a generator or feature extractor. Together with bridging the gap in the latent space, we train a student to mimic the predictions of the teacher expert on both source and target examples. In addition, our approach is guaranteed with rigorous theory offered insightful justifications of how each component influences the transferring performance. Extensive experiments conducted on three benchmark datasets show that our proposed method achieves state-of-the-art performances to the best of our knowledge.

Using the principle of imitation learning and the theory of optimal transport we propose in this paper a novel model for unsupervised domain adaptation named Teacher Imitation Domain Adaptation with Optimal Transport (TIDOT). Our model includes two cooperative agents: a teacher and a student. The former agent is trained to be an expert on labeled data in the source domain, whilst the latter one aims to work with unlabeled data in the target domain. More specifically, optimal transport is applied to quantify the total of the distance between embedded distributions of the source and target data in the joint space, and the distance between predictive distributions of both agents, thus by minimizing this quantity TIDOT could mitigate not only the data shift but also the label shift. Comprehensive empirical studies show that TIDOT outperforms existing state-of-the-art performance on benchmark datasets.

Tuan Lai, Heng Ji, Trung Bui, Quan Hung Tran, Franck Dernoncourt, Walter Chang. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Multi-source domain adaptation (DA) is more challenging than conventional DA because the knowledge is transferred from several source domains to a target domain. To this end, we propose in this paper a novel model for multi-source DA using the theory of optimal transport and imitation learning. More specifically, our approach consists of two cooperative agents: a teacher classifier and a student classifier. The teacher classifier is a combined expert that leverages knowledge of domain experts that can be theoretically guaranteed to handle perfectly source examples, while the student classifier acting on the target domain tries to imitate the teacher classifier acting on the source domains. Our rigorous theory developed based on optimal transport makes this cross-domain imitation possible and also helps to mitigate not only the data shift but also the label shift, which are inherently thorny issues in DA research. We conduct comprehensive experiments on real-world datasets to demonstrate the merit of our approach and its optimal transport based imitation learning viewpoint. Experimental results show that our proposed method achieves state-of-the-art performance on benchmark datasets for multi-source domain adaptation including Digits-five, Office-Caltech10, and Office-31 to the best of our knowledge.

Keyphrase extraction is the task of extracting a small set of phrases that best describe a document. Most existing benchmark datasets for the task typically have limited numbers of annotated documents, making it challenging to train increasingly complex neural networks. In contrast, digital libraries store millions of scientific articles online, covering a wide range of topics. While a significant portion of these articles contain keyphrases provided by their authors, most other articles lack such kind of annotations. Therefore, to effectively utilize these large amounts of unlabeled articles, we propose a simple and efficient joint learning approach based on the idea of self-distillation. Experimental results show that our approach consistently improves the performance of baseline models for keyphrase extraction. Furthermore, our best models outperform previous methods for the task, achieving new state-of-the-art results on two public benchmarks: Inspec and SemEval-2017.

Interpretability and explainability of deep neural networks are challenging due to their scale, complexity, and the agreeable notions on which the explaining process rests.Previous work, in particular, has focused on representing internal components of neural networks through humanfriendly visuals and concepts.On the other hand, in real life, when making a decision, human tends to rely on similar situations and/or associations in the past.Hence arguably, a promising approach to make the model transparent is to design it in a way such that the model explicitly connects the current sample with the seen ones, and bases its decision on these samples.Grounded on that principle, we propose in this paper an explainable, evidence-based memory network architecture, which learns to summarize the dataset and extract supporting evidences to make its decision.Our model achieves state-of-the-art performance on two popular question answering datasets (i.e.TrecQA and WikiQA).Via further analysis, we show that this model can reliably trace the errors it has made in the validation step to the training instances that might have caused these errors.We believe that this error-tracing capability provides significant benefit in improving dataset quality in many applications.

Acronyms are the short forms of phrases that facilitate conveying lengthy sentences in documents and serve as one of the mainstays of writing.Due to their importance, identifying acronyms and corresponding phrases (i.e., acronym identification (AI)) and finding the correct meaning of each acronym (i.e., acronym disambiguation (AD)) are crucial for text understanding.Despite the recent progress on this task, there are some limitations in the existing datasets which hinder further improvement.More specifically, limited size of manually annotated AI datasets or noises in the automatically created acronym identification datasets obstruct designing advanced highperforming acronym identification models.Moreover, the existing datasets are mostly limited to the medical domain and ignore other domains.In order to address these two limitations, we first create a manually annotated large AI dataset for scientific domain.This dataset contains 17,506 sentences which is substantially larger than previous scientific AI datasets.Next, we prepare an AD dataset for scientific domain with 62,441 samples which is significantly larger than previous scientific AD dataset.Our experiments show that the existing state-of-the-art models fall far behind human-level performance on both datasets proposed by this work.In addition, we propose a new deep learning model which utilizes the syntactical structure of the sentence to expand an ambiguous acronym in a sentence.The proposed model outperforms the state-of-the-art models on the new AD dataset, providing a strong baseline for future research on this dataset 1 .

In a task-oriented dialog system, the goal of dialog state tracking (DST) is to monitor the state of the conversation from the dialog history. Recently, many deep learning based methods have been proposed for the task. Despite their impressive performance, current neural architectures for DST are typically heavily-engineered and conceptually complex, making it difficult to implement, debug, and maintain them in a production setting. In this work, we propose a simple but effective DST model based on BERT. In addition to its simplicity, our approach also has a number of other advantages: (a) the number of parameters does not grow with the ontology size (b) the model can operate in situations where the domain ontology may change dynamically. Experimental results demonstrate that our BERT-based model outperforms previous methods by a large margin, achieving new state-of-the-art results on the standard WoZ 2.0 dataset1. Finally, to make the model small and fast enough for resource-restricted systems, we apply the knowledge distillation method to compress our model. The final compressed model achieves comparable results with the original model while being 8x smaller and 7x faster.

Tuan Lai, Quan Hung Tran, Trung Bui, Daisuke Kihara. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Quan Hung Tran, Tuan Lai, Gholamreza Haffari, Ingrid Zukerman, Trung Bui, Hung Bui. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers). 2018.

We propose a novel hierarchical Recurrent Neural Network (RNN) for learning sequences of Dialogue Acts (DAs).The input in this task is a sequence of utterances (i.e., conversational contributions) comprising a sequence of tokens, and the output is a sequence of DA labels (one label per utterance).Our model leverages the hierarchical nature of dialogue data by using two nested RNNs that capture long-range dependencies at the dialogue level and the utterance level.This model is combined with an attention mechanism that focuses on salient tokens in utterances.Our experimental results show that our model outperforms strong baselines on two popular datasets, Switchboard and MapTask; and our detailed empirical analysis highlights the impact of each aspect of our model.

This paper introduces a novel training/decoding strategy for sequence labeling.Instead of greedily choosing a label at each time step, and using it for the next prediction, we retain the probability distribution over the current label, and pass this distribution to the next prediction.This approach allows us to avoid the effect of label bias and error propagation in sequence learning/decoding.Our experiments on dialogue act classification demonstrate the effectiveness of this approach.Even though our underlying neural network model is relatively simple, it outperforms more complex neural models, achieving state-of-the-art results on the MapTask and Switchboard corpora.

Quan Hung Tran, Ingrid Zukerman, Gholamreza Haffari. Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016.