COLING 2022 Papers

We present a new multimodal dataset called Visual Recipe Flow, which enables us to learn a cooking action result for each object in a recipe text. The dataset consists of object state changes and the workflow of the recipe text. The state change is represented as an image pair, while the workflow is represented as a recipe flow graph. We developed a web interface to reduce human annotation costs. The dataset allows us to try various applications, including multimodal information retrieval.

Deploying large-scale pre-trained models in the prompt-tuning paradigm has demonstrated promising performance in few-shot learning. Particularly, vision-language pre-training models (VL-PTMs) have been intensively explored in various few-shot downstream tasks. However, most existing works only apply VL-PTMs to visual tasks like image classification, with few attempts being made on language tasks like text classification. In few-shot text classification, a feasible paradigm for deploying VL-PTMs is to align the input samples and their category names via the text encoders. However, it leads to the waste of visual information learned by the image encoders of VL-PTMs. To overcome this drawback, we propose a novel method named Visual Prompt Tuning (VPT). To our best knowledge, this method is the first attempt to deploy VL-PTM in few-shot text classification task. The main idea is to generate the image embeddings w.r.t. category names as visual prompt and then add them to the aligning process. Extensive experiments show that our VPT can achieve significant improvements under both zero-shot and few-shot settings. Importantly, our VPT even outperforms the most recent prompt-tuning methods on five public text classification datasets.

Zero-shot paraphrase generation has drawn much attention as the large-scale high-quality paraphrase corpus is limited. Back-translation, also known as the pivot-based method, is typical to this end. Several works leverage different information as ”pivot” such as language, semantic representation and so on. In this paper, we explore using visual information such as image as the ”pivot” of back-translation. Different with the pipeline back-translation method, we propose visual information guided zero-shot paraphrase generation (ViPG) based only on paired image-caption data. It jointly trains an image captioning model and a paraphrasing model and leverage the image captioning model to guide the training of the paraphrasing model. Both automatic evaluation and human evaluation show our model can generate paraphrase with good relevancy, fluency and diversity, and image is a promising kind of pivot for zero-shot paraphrase generation.

Brain encoding aims at reconstructing fMRI brain activity given a stimulus. There exists a plethora of neural encoding models which study brain encoding for single mode stimuli: visual (pretrained CNNs) or text (pretrained language models). Few recent papers have also obtained separate visual and text representation models and performed late-fusion using simple heuristics. However, previous work has failed to explore the co-attentive multi-modal modeling for visual and text reasoning. In this paper, we systematically explore the efficacy of image and multi-modal Transformers for brain encoding. Extensive experiments on two popular datasets, BOLD5000 and Pereira, provide the following insights. (1) We find that VisualBERT, a multi-modal Transformer, significantly outperforms previously proposed single-mode CNNs, image Transformers as well as other previously proposed multi-modal models, thereby establishing new state-of-the-art. (2) The regions such as LPTG, LMTG, LIFG, and STS which have dual functionalities for language and vision, have higher correlation with multi-modal models which reinforces the fact that these models are good at mimicing the human brain behavior. (3) The supremacy of visio-linguistic models raises the question of whether the responses elicited in the visual regions are affected implicitly by linguistic processing even when passively viewing images. Future fMRI tasks can verify this computational insight in an appropriate experimental setting. We make our code publicly available.

Domain-specific documents cover terminologies and specialized knowledge. This has been the main challenge of domain-specific document retrieval systems. Previous approaches propose domain-adaptation and transfer learning methods to alleviate this problem. However, these approaches still follow the same document representation method in previous approaches; a document is embedded into a single vector. In this study, we propose VKGDR. VKGDR represents a given corpus into a graph of entities and their relations (known as a virtual knowledge graph) and computes the relevance between queries and documents based on the graph representation. We conduct three experiments 1) domain-specific document retrieval, 2) comparison of our virtual knowledge graph construction method with previous approaches, and 3) ablation study on each component of our virtual knowledge graph. From the results, we see that unsupervised VKGDR outperforms baselines in a zero-shot setting and even outperforms fully-supervised bi-encoder. We also verify that our virtual knowledge graph construction method results in better retrieval performance than previous approaches.

Existing works on dialogue summarization often follow the common practice in document summarization and view the dialogue, which comprises utterances of different speakers, as a single utterance stream ordered by time. However, this single-stream approach without specific attention to the speaker-centered points has limitations in fully understanding the dialogue. To better capture the dialogue information, we propose a 2D view of dialogue based on a time-speaker perspective, where the time and speaker streams of dialogue can be obtained as strengthened input. Based on this 2D view, we present an effective two-stream model called ATM to combine the two streams. Extensive experiments on various summarization datasets demonstrate that ATM significantly surpasses other models regarding diverse metrics and beats the state-of-the-art models on the QMSum dataset in ROUGE scores. Besides, ATM achieves great improvements in summary faithfulness and human evaluation. Moreover, results on machine reading comprehension datasets show the generalization ability of the proposed methods and shed light on other dialogue-based tasks. Our code will be publicly available online.

Over a decade, the research field of computational linguistics has witnessed the growth of corpora and models for natural language inference (NLI) for rich-resource languages such as English and Chinese. A large-scale and high-quality corpus is necessary for studies on NLI for Vietnamese, which can be considered a low-resource language. In this paper, we introduce ViNLI (Vietnamese Natural Language Inference), an open-domain and high-quality corpus for evaluating Vietnamese NLI models, which is created and evaluated with a strict process of quality control. ViNLI comprises over 30,000 human-annotated premise-hypothesis sentence pairs extracted from more than 800 online news articles on 13 distinct topics. In this paper, we introduce the guidelines for corpus creation which take the specific characteristics of the Vietnamese language in expressing entailment and contradiction into account. To evaluate the challenging level of our corpus, we conduct experiments with state-of-the-art deep neural networks and pre-trained models on our dataset. The best system performance is still far from human performance (a 14.20% gap in accuracy). The ViNLI corpus is a challenging corpus to accelerate progress in Vietnamese computational linguistics. Our corpus is available publicly for research purposes.

Large pretrained language models offer powerful generation capabilities, but cannot be reliably controlled at a sub-sentential level. We propose to make such fine-grained control possible in pretrained LMs by generating text directly from a semantic representation, Abstract Meaning Representation (AMR), which is augmented at the node level with syntactic control tags. We experiment with English-language generation of three modes of syntax relevant to the framing of a sentence - verb voice, verb tense, and realization of human entities - and demonstrate that they can be reliably controlled, even in settings that diverge drastically from the training distribution. These syntactic aspects contribute to how information is framed in text, something that is important for applications such as summarization which aim to highlight salient information.

Personalized response selection systems are generally grounded on persona. However, a correlation exists between persona and empathy, which these systems do not explore well. Also, when a contradictory or off-topic response is selected, faithfulness to the conversation context plunges. This paper attempts to address these issues by proposing a suite of fusion strategies that capture the interaction between persona, emotion, and entailment information of the utterances. Ablation studies on the Persona-Chat dataset show that incorporating emotion and entailment improves the accuracy of response selection. We combine our fusion strategies and concept-flow encoding to train a BERT-based model which outperforms the previous methods by margins larger than 2.3% on original personas and 1.9% on revised personas in terms of hits@1 (top-1 accuracy), achieving a new state-of-the-art performance on the Persona-Chat dataset

Measuring Sentence Textual Similarity (STS) is a classic task that can be applied to many downstream NLP applications such as text generation and retrieval. In this paper, we focus on unsupervised STS that works on various domains but only requires minimal data and computational resources. Theoretically, we propose a light-weighted Expectation-Correction (EC) formulation for STS computation. EC formulation unifies unsupervised STS approaches including the cosine similarity of Additively Composed (AC) sentence embeddings, Optimal Transport (OT), and Tree Kernels (TK). Moreover, we propose the Recursive Optimal Transport Similarity (ROTS) algorithm to capture the compositional phrase semantics by composing multiple recursive EC formulations. ROTS finishes in linear time and is faster than its predecessors. ROTS is empirically more effective and scalable than previous approaches. Extensive experiments on 29 STS tasks under various settings show the clear advantage of ROTS over existing approaches. Detailed ablation studies prove the effectiveness of our approaches.

Unsupervised question answering is an attractive task due to its independence on labeled data. Previous works usually make use of heuristic rules as well as pre-trained models to construct data and train QA models. However, most of these works regard named entity (NE) as the only answer type, which ignores the high diversity of answers in the real world. To tackle this problem, we propose a novel unsupervised method by diversifying answers, named DiverseQA. Specifically, the proposed method is composed of three modules: data construction, data augmentation and denoising filter. Firstly, the data construction module extends the extracted named entity into a longer sentence constituent as the new answer span to construct a QA dataset with diverse answers. Secondly, the data augmentation module adopts an answer-type dependent data augmentation process via adversarial training in the embedding level. Thirdly, the denoising filter module is designed to alleviate the noise in the constructed data. Extensive experiments show that the proposed method outperforms previous unsupervised models on five benchmark datasets, including SQuADv1.1, NewsQA, TriviaQA, BioASQ, and DuoRC. Besides, the proposed method shows strong performance in the few-shot learning setting.

Naturalness and expressiveness are crucial for audiobook speech synthesis, but now are limited by the averaged global-scale speaking style representation. In this paper, we propose an unsupervised multi-scale context-sensitive text-to-speech model for audiobooks. A multi-scale hierarchical context encoder is specially designed to predict both global-scale context style embedding and local-scale context style embedding from a wider context of input text in a hierarchical manner. Likewise, a multi-scale reference encoder is introduced to extract reference style embeddings at both global and local scales from the reference speech, which is used to guide the prediction of speaking styles. On top of these, a bi-reference attention mechanism is used to align both local-scale reference style embedding sequence and local-scale context style embedding sequence with corresponding phoneme embedding sequence. Both objective and subjective experiment results on a real-world multi-speaker Mandarin novel audio dataset demonstrate the excellent performance of our proposed method over all baselines in terms of naturalness and expressiveness of the synthesized speech.

We propose a new unsupervised method for lexical substitution using pre-trained language models. Compared to previous approaches that use the generative capability of language models to predict substitutes, our method retrieves substitutes based on the similarity of contextualised and decontextualised word embeddings, i.e. the average contextual representation of a word in multiple contexts. We conduct experiments in English and Italian, and show that our method substantially outperforms strong baselines and establishes a new state-of-the-art without any explicit supervision or fine-tuning. We further show that our method performs particularly well at predicting low-frequency substitutes, and also generates a diverse list of substitute candidates, reducing morphophonetic or morphosyntactic biases induced by article-noun agreement.

A shift in data distribution can have a significant impact on performance of a text classification model. Recent methods addressing unsupervised domain adaptation for textual tasks typically extracted domain-invariant representations through balancing between multiple objectives to align feature spaces between source and target domains. While effective, these methods induce various new domain-sensitive hyperparameters, thus are impractical as large-scale language models are drastically growing bigger to achieve optimal performance. To this end, we propose to leverage meta-learning framework to train a neural network-based self-paced learning procedure in an end-to-end manner. Our method, called Meta Self-Paced Domain Adaption (MSP-DA), follows a novel but intuitive domain-shift variation of cluster assumption to derive the meta train-test dataset split based on the self-pacing difficulties of source domain’s examples. As a result, MSP-DA effectively leverages self-training and self-tuning domain-specific hyperparameters simultaneously throughout the learning process. Extensive experiments demonstrate our framework substantially improves performance on target domains, surpassing state-of-the-art approaches. Detailed analyses validate our method and provide insight into how each domain affects the learned hyperparameters.

Recent approaches to Aspect-based Sentiment Analysis (ABSA) take a co-extraction approach to this span-level classification task, performing the subtasks of aspect term extraction (ATE) and aspect sentiment classification (ASC) simultaneously. In this work, we build on recent progress in applying pre-training to this co-extraction task with the introduction of an adaptation of Unsupervised Data Augmentation in semi-supervised learning. As originally implemented, UDA cannot accommodate span-level classification since it relies on advanced data augmentation techniques, such as back-translation, that alter the sequence lengths of the original data and cause index mismatches. We introduce an adaptation of UDA using Masked Language Model (MLM) unmasking that accommodates this index-match constraint and test the approach on standard ABSA benchmark datasets. We show that simple augmentations applied to modest-sized datasets along with consistency training lead to competitive performance with the current ABSA state-of-the-art in the restaurant and laptop domains using only 75% of the training data.

We leverage cross-language data expansion and retraining to enhance neural Event Detection (abbr., ED) on English ACE corpus. Machine translation is utilized for expanding English training set of ED from that of Chinese. However, experimental results illustrate that such strategy actually results in performance degradation. The survey of translations suggests that the mistakenly-aligned triggers in the expanded data negatively influences the retraining process. We refer this phenomenon to “trigger falsification”. To overcome the issue, we apply heuristic rules for regulating the expanded data, fixing the distracting samples that contain the falsified triggers. The supplementary experiments show that the rule-based regulation is beneficial, yielding the improvement of about 1.6% F1-score for ED. We additionally prove that, instead of transfer learning from the translated ED data, the straight data combination by random pouring surprisingly performs better.

Generative question answering (QA) models generate answers to questions either solely based on the parameters of the model (the closed-book setting) or additionally retrieving relevant evidence (the open-book setting). Generative QA models can answer some relatively complex questions, but the mechanism through which they do so is still poorly understood. We perform several studies aimed at better understanding the multi-hop reasoning capabilities of generative QA models. First, we decompose multi-hop questions into multiple corresponding single-hop questions, and find marked inconsistency in QA models’ answers on these pairs of ostensibly identical question chains. Second, we find that models lack zero-shot multi-hop reasoning ability: when trained only on single-hop questions, models generalize poorly to multi-hop questions. Finally, we demonstrate that it is possible to improve models’ zero-shot multi-hop reasoning capacity through two methods that approximate real multi-hop natural language (NL) questions by training on either concatenation of single-hop questions or logical forms (SPARQL). In sum, these results demonstrate that multi-hop reasoning does not emerge naturally in generative QA models, but can be encouraged by advances in training or modeling techniques. Code is available at https://github.com/jzbjyb/multihop.

Attention mechanism has been used as an important component across Vision-and-Language(VL) tasks in order to bridge the semantic gap between visual and textual features. While attention has been widely used in VL tasks, it has not been examined the capability of different attention alignment calculation in bridging the semantic gap between visual and textual clues. In this research, we conduct a comprehensive analysis on understanding the role of attention alignment by looking into the attention score calculation methods and check how it actually represents the visual region’s and textual token’s significance for the global assessment. We also analyse the conditions which attention score calculation mechanism would be more (or less) interpretable, and which may impact the model performance on three different VL tasks, including visual question answering, text-to-image generation, text-and-image matching (both sentence and image retrieval). Our analysis is the first of its kind and provides useful insights of the importance of each attention alignment score calculation when applied at the training phase of VL tasks, commonly ignored in attention-based cross modal models, and/or pretrained models. Our code is available at: https://github.com/adlnlp/Attention_VL

The widespread of fake news has detrimental societal effects. Recent works model information propagation as graph structure and aggregate structural features from user interactions for fake news detection. However, they usually neglect a broader propagation uncertainty issue, caused by some missing and unreliable interactions during actual spreading, and suffer from learning accurate and diverse structural properties. In this paper, we propose a novel dual graph-based model, Uncertainty-aware Propagation Structure Reconstruction (UPSR) for improving fake news detection. Specifically, after the original propagation modeling, we introduce propagation structure reconstruction to fully explore latent interactions in the actual propagation. We design a novel Gaussian Propagation Estimation to refine the original deterministic node representation by multiple Gaussian distributions and arise latent interactions with KL divergence between distributions in a multi-facet manner. Extensive experiments on two real-world datasets demonstrate the effectiveness and superiority of our model.