Papers

My doctoral research focuses on understanding semantic knowledge in neural network models trained solely to predict natural language (referred to as language models, or LMs), by drawing on insights from the study of concepts and categories grounded in cognitive science. I propose a framework inspired by 'inductive reasoning,' a phenomenon that sheds light on how humans utilize background knowledge to make inductive leaps and generalize from new pieces of information about concepts and their properties. Drawing from experiments that study inductive reasoning, I propose to analyze semantic inductive generalization in LMs using phenomena observed in human-induction literature, investigate inductive behavior on tasks such as implicit reasoning and emergent feature recognition, and analyze and relate induction dynamics to the learned conceptual representation space.

poster session 3 @ red 5, poster session 1 @ red 5, poster session 3, poster session 1

Goal-oriented dialog policy learning algorithms aim to learn a dialog policy for selecting language actions based on the current dialog state. Deep reinforcement learning methods have been used for dialog policy learning. This work is motivated by the observation that, although dialog is a domain with rich contextual knowledge, reinforcement learning methods are ill-equipped to incorporate such knowledge into the dialog policy learning process. In this paper, we develop a deep reinforcement learning framework for goal-oriented dialog policy learning that learns user preferences from user goal data, while leveraging commonsense knowledge from people. The developed framework has been evaluated using a realistic dialog simulation platform. Compared with baselines from the literature and the ablations of our approach, we see significant improvements in learning efficiency and the quality of the computed action policies.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

In dialogue systems, utterances with similar semantics may have distinctive emotions under different contexts. Therefore, modeling long-range contextual emotional relationships with speaker dependency plays a crucial part in dialogue emotion recognition. Meanwhile, distinguishing the different emotion categories is non-trivial since they usually have semantically similar sentiments. To this end, we adopt supervised contrastive learning to make different emotions mutually exclusive to identify similar emotions better. Meanwhile, we utilize an auxiliary response generation task to enhance the model's ability of handling context information, thereby forcing the model to recognize emotions with similar semantics in diverse contexts. To achieve these objectives, we use the pre-trained encoder-decoder model BART as our backbone model since it is very suitable for both understanding and generation tasks. The experiments on four datasets demonstrate that our proposed model obtains significantly more favorable results than the state-of-the-art model in dialogue emotion recognition. The ablation study further demonstrates the effectiveness of supervised contrastive loss and generative loss.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Current state-of-the-art cross-lingual summarization models employ multi-task learning paradigm, which works on a shared vocabulary module and relies on the self-attention mechanism to attend among tokens in two languages. However, correlation learned by self-attention is often loose and implicit, inefficient in capturing crucial cross-lingual representations between languages. The matter worsens when performing on languages with separate morphological or structural features, making the cross-lingual alignment more challenging, resulting in the performance drop. To overcome this problem, we propose a novel Knowledge-Distillation-based framework for Cross-Lingual Summarization, seeking to explicitly construct cross-lingual correlation by distilling the knowledge of the monolingual summarization teacher into the cross-lingual summarization student. Since the representations of the teacher and the student lie on two different vector spaces, we further propose a Knowledge Distillation loss using Sinkhorn Divergence, an Optimal-Transport distance, to estimate the discrepancy between those teacher and student representations. Due to the intuitively geometric nature of Sinkhorn Divergence, the student model can productively learn to align its produced cross-lingual hidden states with monolingual hidden states, hence leading to a strong correlation between distant languages. Experiments on cross-lingual summarization datasets in pairs of distant languages demonstrate that our method outperforms state-of-the-art models under both high and low-resourced settings.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Current researches on spoken language understanding (SLU) heavily are limited to a simple setting: the plain text-based SLU that takes the user utterance as input and generates its corresponding semantic frames (e.g., intent and slots). Unfortunately, such a simple setting may fail to work in complex real-world scenarios when an utterance is semantically ambiguous, which cannot be achieved by the text-based SLU models. In this paper, we first introduce a new and important task, Profile-based Spoken Language Understanding (ProSLU), which requires the model that not only relies on the plain text but also the supporting profile information to predict the correct intents and slots. To this end, we further introduce a large-scale human-annotated Chinese dataset with over 5K utterances and their corresponding supporting profile information (Knowledge Graph (KG), User Profile (UP), Context Awareness (CA)). In addition, we evaluate several state-of-the-art baseline models and explore a multi-level knowledge adapter to effectively incorporate profile information. Experimental results reveal that all existing text-based SLU models fail to work when the utterances are semantically ambiguous and our proposed framework can effectively fuse the supporting information for sentence-level intent detection and token-level slot filling. Finally, we summarize key challenges and provide new points for future directions, which hopes to facilitate the research.

poster session 3 @ red 5, poster session 8 @ red 5, oral session 8 @ red 5, poster session 3, poster session 8, oral session 8

Capturing visual similarity among images is the core of many computer vision and pattern recognition tasks. This problem can be formulated in such a paradigm called metric learning. Most research in the area has been mainly focusing on improving the loss functions and similarity measures. However, due to the ignoring of geometric structure, existing methods often lead to sub-optimal results. Thus, several recent research methods took advantage of Wasserstein distance between batches of samples to characterize the spacial geometry. Although these approaches can achieve enhanced performance, the aggregation over batches definitely hinders Wasserstein distance's superior measure capability and leads to high computational complexity. To address this limitation, we propose a novel Deep Wasserstein Metric Learning framework, which employs Wasserstein distance to precisely capture the relationship among various images under ranking-based loss functions such as contrastive loss and triplet loss. Our method directly computes the distance between images, considering the geometry at a finer granularity than batch level. Furthermore, we introduce a new efficient algorithm using Sinkhorn approximation and Wasserstein measure coreset. The experimental results demonstrate the improvements of our framework over various baselines in different applications and benchmark datasets.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

The goal of building intelligent dialogue systems has largely been separately pursued under two paradigms: task-oriented dialogue (TOD) systems, which perform task-specific functions, and open-domain dialogue (ODD) systems, which focus on non-goal-oriented chitchat. The two dialogue modes can potentially be intertwined together seamlessly in the same conversation, as easily done by a friendly human assistant. Such ability is desirable in conversational agents, as the integration makes them more accessible and useful. Our paper addresses this problem of fusing TODs and ODDs in multi-turn dialogues. Based on the popular TOD dataset MultiWOZ, we build a new dataset FusedChat, by rewriting the existing TOD turns and adding new ODD turns. This procedure constructs conversation sessions containing exchanges from both dialogue modes. It features inter-mode contextual dependency, i.e., the dialogue turns from the two modes depend on each other. Rich dependency patterns such as co-reference and ellipsis are included. The new dataset, with 60k new human-written ODD turns and 5k re-written TOD turns, offers a benchmark to test a dialogue model's ability to perform inter-mode conversations. This is a more challenging task since the model has to determine the appropriate dialogue mode and generate the response based on the inter-mode context. However, such models would better mimic human-level conversation capabilities. We evaluate two baseline models on this task, including the classification-based two-stage models and the two-in-one fused models. We publicly release FusedChat (https://github.com/tomyoung903/FusedChat) and the baselines to propel future work on inter-mode dialogue systems.

poster session 3 @ red 5, poster session 8 @ red 5, oral session 8 @ red 5, poster session 3, poster session 8, oral session 8

A key trait of daily conversations between individuals is the ability to express empathy towards others, and exploring ways to implement empathy is a crucial step towards human-like dialogue systems. Previous approaches on this topic mainly focus on detecting and utilizing the user's emotion for generating empathetic responses. However, since empathy includes both aspects of affection and cognition, we argue that in addition to identifying the user's emotion, cognitive understanding of the user's situation should also be considered. To this end, we propose a novel approach for empathetic response generation, which leverages commonsense to draw more information about the user's situation and uses this additional information to further enhance the empathy expression in generated responses. We evaluate our approach on EmpatheticDialogues, which is a widely-used benchmark dataset for empathetic response generation. Empirical results demonstrate that our approach outperforms the baseline models in both automatic and human evaluations and can generate more informative and empathetic responses.

poster session 3 @ red 5, poster session 8 @ red 5, oral session 3 @ red 5, poster session 3, poster session 8, oral session 3

Intent identification is the key technology in dialogue system. However, not all online queries are clear or complete. To identify users' intents from those fuzzy queries accurately, this paper proposes a multi-factor classification framework on the query level. Experimental results on our online serving system JIMI demonstrate the effectiveness of our proposed framework.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

A conversation is a sequence of utterances, where each utterance plays a specific discourse role while expressing a particular emotion. This paper proposes a novel method to exploit latent discourse role information of an utterance to determine the emotion it conveys in a conversation. Specifically, we use a variant of the Variational-Autoencoder (VAE) to model the context-aware latent discourse roles of each utterance in an unsupervised way. The latent discourse role representation further equips the utterance representation with a salient clue for more accurate emotion recognition. Our experiments show that our proposed method beats the best-reported performances on three public Emotion Recognition in Conversation datasets. This proves that the discourse role information of an utterance plays an important role in the emotion recognition task, which no previous work has studied.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Pre-trained models have proved to be powerful in enhancing task-oriented dialog systems. However, current pre-training methods mainly focus on enhancing dialog understanding and generation tasks while neglecting the exploitation of dialog policy. In this paper, we propose GALAXY, a novel pre-trained dialog model that explicitly learns dialog policy from limited labeled dialogs and large-scale unlabeled dialog corpora via semi-supervised learning. Specifically, we introduce a dialog act prediction task for policy optimization during pre-training and employ a consistency regularization term to refine the learned representation with the help of unlabeled dialogs. We also implement a gating mechanism to weigh suitable unlabeled dialog samples. Empirical results show that GALAXY substantially improves the performance of task-oriented dialog systems, and achieves new state-of-the-art results on benchmark datasets: In-Car, MultiWOZ2.0 and MultiWOZ2.1, improving their end-to-end combined scores by 2.5, 5.3 and 5.5 points, respectively. We also show that GALAXY has a stronger few-shot ability than existing models under various low-resource settings. For reproducibility, we release the code and data at https://github.com/siat-nlp/GALAXY.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Humans usually have conversations by making use of prior knowledge about a topic and background information of the people whom they are talking to. However, existing conversational agents and datasets do not consider such comprehensive information, and thus they have a limitation in generating the utterances where the knowledge and persona are fused properly. To address this issue, we introduce a call For Customized conversation (FoCus) dataset where the customized answers are built with the user's persona and Wikipedia knowledge. To evaluate the abilities to make informative and customized utterances of pre-trained language models, we utilize BART and GPT-2 as well as transformer-based models. We assess their generation abilities with automatic scores and conduct human evaluations for qualitative results. We examine whether the model reflects adequate persona and knowledge with our proposed two sub-tasks, persona grounding (PG) and knowledge grounding (KG). Moreover, we show that the utterances of our data are constructed with the proper knowledge and persona through grounding quality assessment.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Generating absent keyphrases, which do not appear in the input document, is challenging in the keyphrase prediction task. Most previous works treat the problem as an autoregressive sequence-to-sequence generation task, which demonstrates promising results for generating grammatically correct and fluent absent keyphrases. However, such an end-to-end process with a complete data-driven manner is unconstrained, which is prone to generate keyphrases inconsistent with the input document. In addition, the existing autoregressive decoding method makes the generation of keyphrases must be done from left to right, leading to slow speed during inference. In this paper, we propose a constrained absent keyphrase generation method in a prompt-based learning fashion. Specifically, the prompt will be created firstly based on the keywords, which are defined as the overlapping words between absent keyphrase and document. Then, a mask-predict decoder is used to complete the absent keyphrase on the constraint of prompt. Experiments on keyphrase generation benchmarks have demonstrated the effectiveness of our approach. In addition, we evaluate the performance of constrained absent keyphrases generation from an information retrieval perspective. The result shows that our approach can generate more consistent keyphrases, which can improve document retrieval performance. What’s more, with a non-autoregressive decoding manner, our model can speed up the absent keyphrase generation by 8.67× compared with the autoregressive method.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

This paper introduces a simple yet effective data-centric approach for the task of improving persona-conditioned dialogue agents. Prior model-centric approaches unquestioningly depend on the raw crowdsourced benchmark datasets such as Persona-Chat. In contrast, we aim to fix annotation artifacts in benchmarking, which is orthogonally applicable to any dialogue model. Specifically, we augment relevant personas to improve dialogue dataset/agent, by leveraging the primal-dual structure of the two tasks, predicting dialogue responses and personas based on each other. Experiments on Persona-Chat show that our approach outperforms pre-trained LMs by an 11.7 point gain in terms of accuracy.

poster session 3 @ red 5, poster session 8 @ red 5, oral session 8 @ red 5, poster session 3, poster session 8, oral session 8

Key information extraction (KIE) from document images requires understanding the contextual and spatial semantics of texts in two-dimensional (2D) space. Many recent studies try to solve the task by developing pre-trained language models focusing on combining visual features from document images with texts and their layout. On the other hand, this paper tackles the problem by going back to the basic: effective combination of text and layout. Specifically, we propose a pre-trained language model, named BROS (BERT Relying On Spatiality), that encodes relative positions of texts in 2D space and learns from unlabeled documents with area-masking strategy. With this optimized training scheme for understanding texts in 2D space, BROS shows comparable or better performance compared to previous methods on four KIE benchmarks (FUNSD, SROIE*, CORD, and SciTSR) without relying on visual features. This paper also reveals two real-world challenges in KIE tasks-(1) minimizing the error from incorrect text ordering and (2) efficient learning from fewer downstream examples-and demonstrates the superiority of BROS over previous methods. Code will be available at https://github.com/clovaai.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Transformer-based pre-trained language models like BERT and its variants have recently achieved promising performance in various natural language processing (NLP) tasks. However, the conventional paradigm constructs the backbone by purely stacking the manually designed global self-attention layers, introducing inductive bias and thus leads to sub-optimal. In this work, we make the first attempt to automatically discover novel pre-trained language model (PLM) backbone on a flexible search space containing the most fundamental operations from scratch. Specifically, we propose a well-designed search space which (i) contains primitive math operations in the intra-layer level to explore novel attention structures, and (ii) leverages convolution blocks to be the supplementary for attentions in the inter-layer level to better learn local dependency. To enhance the efficiency for finding promising architectures, we propose an Operation-Priority Neural Architecture Search (OP-NAS) algorithm, which optimizes both the search algorithm and evaluation of candidate models. Specifically, we propose Operation-Priority (OP) evolution strategy to facilitate model search via balancing exploration and exploitation. Furthermore, we design a Bi-branch Weight-Sharing (BIWS) training strategy for fast model evaluation. Extensive experiments show that the searched architecture (named AutoBERT-Zero) significantly outperforms BERT and its variants of different model capacities in various downstream tasks, proving the architecture's transfer and scaling abilities. Remarkably, AutoBERT-Zero-base outperforms RoBERTa-base (using much more data) and BERT-large (with much larger model size) by 2.4 and 1.4 higher score on GLUE test set.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Contextual word representations generated by language models learn spurious associations present in the training corpora. Adversaries can exploit these associations to reverse-engineer the private attributes of entities mentioned in the training corpora. These findings have led to efforts towards minimizing the privacy risks of language models. However, existing approaches lack interpretability, compromise on data utility and fail to provide privacy guarantees. Thus, the goal of my doctoral research is to develop interpretable approaches towards privacy preservation of text representations that maximize data utility retention and guarantee privacy. To this end, I aim to study and develop methods to incorporate steganographic modifications within the vector geometry to obfuscate underlying spurious associations and retain the distributional semantic properties learnt during training.

poster session 3 @ red 5, poster session 1 @ red 5, poster session 3, poster session 1

The task of summarization often requires a non-trivial understanding of the given text at the semantic level. In this work, we essentially incorporate the constituent structure into the single document summarization via the Graph Neural Networks to learn the semantic meaning of tokens. More specifically, we propose a novel hierarchical heterogeneous graph attention network over constituency-based parse trees for syntax-aware summarization. This approach reflects psychological findings that humans will pinpoint specific selection patterns to construct summaries hierarchically. Extensive experiments demonstrate that our model is effective for both the abstractive and extractive summarization tasks on six benchmark datasets from various domains. Moreover, further performance improvement can be obtained by virtue of state-of-the-art pre-trained models. Lastly, we test our model for the code summarization task for border impact.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Identifying breakdowns in ongoing dialogues helps to improve communication effectiveness. Most prior work on this topic relies on human annotated data and data augmentation to learn a classification model. While quality labeled dialogue data requires human annotation and is usually expensive to obtain, unlabeled data is easier to collect from various sources. In this paper, we propose a novel semi-supervised teacher-student learning framework to tackle this task. We introduce two teachers which are trained on labeled data and perturbed labeled data respectively. We leverage unlabeled data to improve classification in student training where we employ two teachers to refine the labeling of unlabeled data through teacher-student learning in a bootstrapping manner. Through our proposed training approach, the student can achieve improvements over single-teacher performance. Experimental results on the Dialogue Breakdown Detection Challenge dataset DBDC5 and Learning to Identify Follow-Up Questions dataset LIF show that our approach outperforms all previous published approaches as well as other supervised and semi-supervised baseline methods.

poster session 3 @ red 5, poster session 8 @ red 5, oral session 3 @ red 5, poster session 3, poster session 8, oral session 3

Recent studies have shown that strong Natural Language Understanding (NLU) models are prone to relying on annotation biases of the datasets as a shortcut, which goes against the underlying mechanisms of the task of interest. To reduce such biases, several recent works introduce debiasing methods to regularize the training process of targeted NLU models. In this paper, we provide a new perspective with causal inference to find out the bias. On one hand, we show that there is an unobserved confounder for the natural language utterances and their respective classes, leading to spurious correlations from training data. To remove such confounder, the backdoor adjustment with causal intervention is utilized to find the true causal effect, which makes the training process fundamentally different from the traditional likelihood estimation. On the other hand, in inference process, we formulate the bias as the direct causal effect and remove it by pursuing the indirect causal effect with counterfactual reasoning. We conduct experiments on large-scale natural language inference and fact verification benchmarks, evaluating on bias sensitive datasets that are specifically designed to assess the robustness of models against known biases in the training data. Experimental results show that our proposed debiasing framework outperforms previous state-of-the-art debiasing methods while maintaining the original in-distribution performance.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Recent advances in large-scale pre-training such as GPT-3 allow seemingly high quality text to be generated from a given prompt. However, such generation systems often suffer from problems of hallucinated facts, and are not inherently designed to incorporate useful external information. Grounded generation models appear to offer remedies, but their training typically relies on rarely-available parallel data where corresponding information-relevant documents are provided for context. We propose a framework that alleviates this data constraint by jointly training a grounded generator and document retriever on the language model signal. The model learns to reward retrieval of the documents with the highest utility in generation, and attentively combines them using a Mixture-of-Experts (MoE) ensemble to generate follow-on text. We demonstrate that both generator and retriever can take advantage of this joint training and work synergistically to produce more informative and relevant text in both prose and dialogue generation.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Human conversations are guided by short-term and long-term goals. We study how to plan short-term goal sequences as coherently as humans do and naturally direct them to an assigned long-term goal in open-domain conversations. Goal sequences are a series of knowledge graph (KG) entity-relation connections generated by KG walkers that traverse through the KG. The existing recurrent and graph attention based KG walkers either insufficiently utilize the conversation states or lack global guidance. In our work, a hierarchical model learns goal planning in a hierarchical learning framework. We present HiTKG, a hierarchical transformer-based graph walker that leverages multiscale inputs to make precise and flexible predictions on KG paths. Furthermore, we propose a two-hierarchy learning framework that employs two stages to learn both turn-level (short-term) and global-level (long-term) conversation goals. Specifically, at the first stage, HiTKG is trained in a supervised fashion to learn how to plan turn-level goal sequences; at the second stage, HiTKG tries to naturally approach the assigned global goal via reinforcement learning. In addition, we propose MetaPath as the backbone method for KG path representation to exploit the entity and relation information concurrently. We further propose Multi-source Decoding Inputs and Output-level Length Head to improve the decoding controllability. Our experiments show that HiTKG achieves a significant improvement in the performance of turn-level goal learning compared with state-of-the-art baselines. Additionally, both automatic and human evaluation prove the effectiveness of the two-hierarchy learning framework for both short-term and long-term goal planning.

poster session 3 @ red 5, poster session 8 @ red 5, oral session 3 @ red 5, poster session 3, poster session 8, oral session 3

Procedural text understanding requires machines to reason about entity states within the dynamical narratives. Current procedural text understanding approaches are commonly entity-wise, which separately track each entity and independently predict different states of each entity. Such an entity-wise paradigm does not consider the interaction between entities and their states. In this paper, we propose a new scene-wise paradigm for procedural text understanding, which jointly tracks states of all entities in a scene-by-scene manner. Based on this paradigm, we propose Scene Graph Reasoner (SGR), which introduces a series of dynamically evolving scene graphs to jointly formulate the evolution of entities, states and their associations throughout the narrative. In this way, the deep interactions between all entities and states can be jointly captured and simultaneously derived from scene graphs. Experiments show that SGR not only achieves the new state-of-the-art performance but also significantly accelerates the speed of reasoning.

poster session 3 @ red 5, poster session 8 @ red 5, poster session 3, poster session 8

Named Entity Recognition models perform well on benchmark datasets but fail to generalize well even in the same domain. The goal of my th esis is to quantify the degree of in-domain generalization in NER, probe models for entity name vs. context learning and finally improve their robustness, focusing on the recognition of ethnically diverse entities and new entities over time when the models are deployed.

poster session 3 @ red 5, poster session 1 @ red 5, poster session 3, poster session 1