NAACL 2022 Papers

Recently, a boom of papers has shown extraordinary progress in zero-shot and few-shot learning with various prompt-based models. It is commonly argued that prompts help models to learn faster in the same way that humans learn faster when provided with task instructions expressed in natural language. In this study, we experiment with over 30 prompts manually written for natural language inference (NLI). We find that models can learn just as fast with many prompts that are intentionally irrelevant or even pathologically misleading as they do with instructively “good” prompts. Further, such patterns hold even for models as large as 175 billion parameters (Brown et al., 2020) as well as the recently proposed instruction-tuned models which are trained on hundreds of prompts (Sanh et al., 2021). That is, instruction-tuned models often produce good predictions with irrelevant and misleading prompts even at zero shots. In sum, notwithstanding prompt-based models’ impressive improvement, we find evidence of serious limitations that question the degree to which such improvement is derived from models understanding task instructions in ways analogous to humans’ use of task instructions.

Deep Learning (DL) techniques have been increasingly adopted for Automatic Text Scoring in education. However, these techniques often suffer from their inabilities to explain and justify how a prediction is made, which, unavoidably, decreases their trustworthiness and hinders educators from embracing them in practice. This study aimed to investigate whether (and to what extent) DL-based graders align with human graders regarding the important words they identify when marking short answer questions. To this end, we first conducted a user study to ask human graders to manually annotate important words in assessing answer quality and then measured the overlap between these human-annotated words and those identified by DL-based graders (i.e., those receiving large attention weights). Furthermore, we ran a randomized controlled experiment to explore the impact of highlighting important words detected by DL-based graders on human grading. The results showed that: (i) DL-based graders, to a certain degree, displayed alignment with human graders no matter whether DL-based graders and human graders agreed on the quality of an answer; and (ii) it is possible to facilitate human grading by highlighting those DL-detected important words, though further investigations are necessary to understand how human graders exploit such highlighted words.

Generative dialogue models suffer badly from the generic response problem, limiting their applications to a few toy scenarios. Recently, an interesting approach, namely negative training, has been proposed to alleviate this problem by reminding the model not to generate high-frequency responses during training. However, its performance is hindered by two issues, ignoring low-frequency but generic responses and bringing low-frequency but meaningless responses. In this paper, we propose a novel negative training paradigm, called negative distillation, to keep the model away from the undesirable generic responses while avoiding the above problems. First, we introduce a negative teacher model that can produce query-wise generic responses, and then the student model is required to maximize the distance with multi-level negative knowledge. Empirical results show that our method outperforms previous negative training methods significantly.

Product aspect extraction from reviews is a critical task for e-commerce services to understand customer preferences and pain points. While aspect phrases extraction and sentiment analysis have received a lot of attention, clustering of aspect phrases and assigning human readable names to clusters in e-commerce reviews is an extremely important and challenging problem due to the scale of the reviews that makes human review infeasible. In this paper, we propose fully automated methods for clustering aspect words and generating human readable names for the clusters without any manually labeled data. We train transformer based sentence embeddings that are aware of unique e-commerce language characteristics (eg. incomplete sentences, spelling and grammar errors, vernacular etc.). We also train transformer based sequence to sequence models to generate human readable aspect names from clusters. Both the models are trained using heuristic based distant supervision. Additionally, the models are used to improve each other. Extensive empirical testing showed that the clustering model improves the Silhouette Score by 64% when compared to the state-of-the-art baseline and the aspect naming model achieves a high ROUGE-L score of 0.79.

In this paper, we explore the task of determining indirect answers to yes-no questions in real conversations. We work with transcripts of phone conversations in the Switchboard Dialog Act (SwDA) corpus and create SwDA-IndirectAnswers (SwDA-IA), a subset of SwDA consisting of all conversations containing a yes-no question with an indirect answer. We annotate the underlying direct answers to the yes-no questions (yes, probably yes, middle, probably no, or no). We show that doing so requires taking into account conversation context: the indirect answer alone is insufficient to determine the ground truth. Experimental results also show that taking into account context is beneficial. More importantly, our results demonstrate that existing corpora with synthetic indirect answers to yes-no questions are not beneficial when working with real conversations. Our best models outperform the majority baseline by a substantial margin, but the task remains a challenge (F1: 0.46).

The emergence of language between artificial agents is a recent focus of computational linguistics, as it offers a synthetic substrate for reasoning about human language evolution. From the perspective of cognitive science, sophisticated categorization in humans is thought to enable reasoning about novel observations, and thus compose old information to describe new phenomena. Unfortunately, the literature to date has not managed to isolate the effect of categorization power in artificial agents on their inter-communication ability, particularly on novel, unseen objects. In this work, we propose the use of disentangled representations from representation learning to quantify the categorization power of agents, enabling a differential analysis between combinations of heterogeneous systems, e.g., pairs of agents which learn to communicate despite mismatched concept realization. Through this approach, we observe that agent heterogeneity can cut signaling accuracy by up to 40%, despite encouraging compositionality in the artificial language. We conclude that the reasoning process of agents plays a key role in their communication, with unexpected benefits arising from their mixing, such as better language compositionality.

Building models to detect vaccine attitudes on social media is challenging because of the composite, often intricate aspects involved, and the limited availability of annotated data. Existing approaches have relied heavily on supervised training that requires abundant annotations and pre-defined aspect categories. Instead, with the aim of leveraging the large amount of unannotated data now available on vaccination, we propose a novel semi-supervised approach for vaccine attitude detection, called VADet. A variational autoencoding architecture based on language models is employed to learn from unlabelled data the topical information of the domain. Then, the model is fine-tuned with a few manually annotated examples of user attitudes. We validate the effectiveness of VADet on our annotated data and also on an existing vaccination corpus annotated with opinions on vaccines. Our results show that VADet is able to learn disentangled stance and aspect topics, and outperforms existing aspect-based sentiment analysis models on both stance detection and tweet clustering.

Action in video usually involves the interaction of human with objects. Action labels are typically composed of various combinations of verbs and nouns, but we may not have training data for all possible combinations. In this paper, we aim to improve the generalization ability of the compositional action recognition model to novel verbs or novel nouns that are unseen during training time, by leveraging the power of knowledge graphs. Previous work utilizes verb-noun compositional action nodes in the knowledge graph, making it inefficient to scale since the number of compositional action nodes grows quadratically with respect to the number of verbs and nouns. To address this issue, we propose our approach: Disentangled Action Recognition with Knowledge-bases (DARK), which leverages the inherent compositionality of actions. DARK trains a factorized model by first extracting disentangled feature representations for verbs and nouns, and then predicting classification weights using relations in external knowledge graphs. The type constraint between verb and noun is extracted from external knowledge bases and finally applied when composing actions. DARK has better scalability in the number of objects and verbs, and achieves state-of-the-art performance on the Charades dataset. We further propose a new benchmark split based on the Epic-kitchen dataset which is an order of magnitude bigger in the numbers of classes and samples, and benchmark various models on this benchmark.

We propose DiffCSE, an unsupervised contrastive learning framework for learning sentence embeddings. DiffCSE learns sentence embeddings that are sensitive to the difference between the original sentence and an edited sentence, where the edited sentence is obtained by stochastically masking out the original sentence and then sampling from a masked language model. We show that DiffSCE is an instance of equivariant contrastive learning, which generalizes contrastive learning and learns representations that are insensitive to certain types of augmentations and sensitive to other “harmful” types of augmentations. Our experiments show that DiffCSE achieves state-of-the-art results among unsupervised sentence representation learning methods, outperforming unsupervised SimCSE by 2.3 absolute points on semantic textual similarity tasks.

Task-oriented dialogue generation is challenging since the underlying knowledge is often dynamic and effectively incorporating knowledge into the learning process is hard. It is particularly challenging to generate both human-like and informative responses in this setting. Recent research primarily focused on various knowledge distillation methods where the underlying relationship between the facts in a knowledge base is not effectively captured. In this paper, we go one step further and demonstrate how the structural information of a knowledge graph can improve the system’s inference capabilities. Specifically, we propose DialoKG, a novel task-oriented dialogue system that effectively incorporates knowledge into a language model. Our proposed system views relational knowledge as a knowledge graph and introduces (1) a structure-aware knowledge embedding technique, and (2) a knowledge graph-weighted attention masking strategy to facilitate the system selecting relevant information during the dialogue generation. An empirical evaluation demonstrates the effectiveness of DialoKG over state-of-the-art methods on several standard benchmark datasets.

Dialogue summarization is receiving increasing attention from researchers due to its extraordinary difficulty and unique application value. We observe that current dialogue summarization models have flaws that may not be well exposed by frequently used metrics such as ROUGE. In our paper, we re-evaluate 18 categories of metrics in terms of four dimensions: coherence, consistency, fluency and relevance, as well as a unified human evaluation of various models for the first time. Some noteworthy trends which are different from the conventional summarization tasks are identified. We will release DialSummEval, a multi-faceted dataset of human judgments containing the outputs of 14 models on SAMSum.

Vision-and-language navigation (VLN) is a multimodal task where an agent follows natural language instructions and navigates in visual environments. Multiple setups have been proposed, and researchers apply new model architectures or training techniques to boost navigation performance. However, there still exist non-negligible gaps between machines’ performance and human benchmarks. Moreover, the agents’ inner mechanisms for navigation decisions remain unclear. To the best of our knowledge, how the agents perceive the multimodal input is under-studied and needs investigation. In this work, we conduct a series of diagnostic experiments to unveil agents’ focus during navigation. Results show that indoor navigation agents refer to both object and direction tokens when making decisions. In contrast, outdoor navigation agents heavily rely on direction tokens and poorly understand the object tokens. Transformer-based agents acquire a better cross-modal understanding of objects and display strong numerical reasoning ability than non-Transformer-based agents. When it comes to vision-and-language alignments, many models claim that they can align object tokens with specific visual targets. We find unbalanced attention on the vision and text input and doubt the reliability of such cross-modal alignments.

For agents at a contact centre receiving calls, the most important piece of information is the reason for a given call. An agent cannot provide support on a call if they do not know why a customer is calling. In this paper we describe our implementation of a commercial system to detect Purpose of Call statements in English business call transcripts in real time. We present a detailed analysis of types of Purpose of Call statements and language patterns related to them, discuss an approach to collect rich training data by bootstrapping from a set of rules to a neural model, and describe a hybrid model which consists of a transformer-based classifier and a set of rules by leveraging insights from the analysis of call transcripts. The model achieved 88.6 F1 on average in various types of business calls when tested on real life data and has low inference time. We reflect on the challenges and design decisions when developing and deploying the system.

Automatic extraction of narrative elements from text, combining narrative theories with computational models, has been receiving increasing attention over the last few years. Previous works have utilized the oral narrative theory by Labov and Waletzky to identify various narrative elements in personal stories texts. Instead, we direct our focus to informational texts, specifically news stories. We introduce NEAT (Narrative Elements AnnoTation) – a novel NLP task for detecting narrative elements in raw text. For this purpose, we designed a new multi-label narrative annotation scheme, better suited for informational text (e.g. news media), by adapting elements from the narrative theory of Labov and Waletzky (Complication and Resolution) and adding a new narrative element of our own (Success). We then used this scheme to annotate a new dataset of 2,209 sentences, compiled from 46 news articles from various category domains. We trained a number of supervised models in several different setups over the annotated dataset to identify the different narrative elements, achieving an average F1 score of up to 0.77. The results demonstrate the holistic nature of our annotation scheme as well as its robustness to domain category.

Massive false rumors emerging along with breaking news or trending topics severely hinder the truth. Existing rumor detection approaches achieve promising performance on the yesterday’s news, since there is enough corpus collected from the same domain for model training. However, they are poor at detecting rumors about unforeseen events especially those propagated in minority languages due to the lack of training data and prior knowledge (i.e., low-resource regimes). In this paper, we propose an adversarial contrastive learning framework to detect rumors by adapting the features learned from well-resourced rumor data to that of the low-resourced. Our model explicitly overcomes the restriction of domain and/or language usage via language alignment and a novel supervised contrastive training paradigm. Moreover, we develop an adversarial augmentation mechanism to further enhance the robustness of low-resource rumor representation. Extensive experiments conducted on two low-resource datasets collected from real-world microblog platforms demonstrate that our framework achieves much better performance than state-of-the-art methods and exhibits a superior capacity for detecting rumors at early stages.

Many users turn to document retrieval systems (e.g. search engines) to seek answers to controversial or open-ended questions. However, classical document retrieval systems fall short at delivering users a set of direct and diverse responses in such cases, which requires identifying responses within web documents in the context of the query, and aggregating the responses based on their different perspectives. The goal of this work is to survey and study the user information needs for building a multi-perspective search engine of such. We examine the challenges of synthesizing such language understanding objectives with document retrieval, and study a new perspective-oriented document retrieval paradigm. We discuss and assess the inherent natural language understanding challenges one needs to address in order to achieve the goal. Following the design challenges and principles, we propose and evaluate a practical prototype pipeline system. We use the prototype system to conduct a user survey in order to assess the utility of our paradigm, as well as understanding the user information needs when issuing controversial and open-ended queries to a search engine.

Leveraging the dependency tree of the input sentence is able to improve the model performance for relation extraction. A challenging issue is how to remove confusions from the tree. Efforts have been made to utilize the dependency connections between words to selectively emphasize target-relevant information. However, these approaches are limited in focusing on exploiting dependency types. In this paper, we propose dependency position encoding (DPE), an efficient way of incorporating both dependency connections and dependency types into the self-attention mechanism to distinguish the importance of different word dependencies for the task. In contrast to previous studies that process input sentence and dependency information in separate streams, DPE can be seamlessly incorporated into the Transformer and makes it possible to use an one-stream scheme to extract relations between entity pairs. Extensive experiments show that models with our DPE significantly outperform the previous methods on SemEval 2010 Task 8, KBP37, and TACRED.

News recommendation is different from movie or e-commercial recommendation as people usually do not grade the news. Therefore, user feedback for news is always implicit (click behavior, reading time, etc). Inevitably, there are noises in implicit feedback. On one hand, the user may exit immediately after clicking the news as he dislikes the news content, leaving the noise in his positive implicit feedback; on the other hand, the user may be recommended multiple interesting news at the same time and only click one of them, producing the noise in his negative implicit feedback. Opposite implicit feedback could construct more integrated user preferences and help each other to minimize the noise influence. Previous works on news recommendation only used positive implicit feedback and suffered from the noise impact. In this paper, we propose a denoising neural network for news recommendation with positive and negative implicit feedback, named DRPN. DRPN utilizes both feedback for recommendation with a module to denoise both positive and negative implicit feedback to further enhance the performance. Experiments on the real-world large-scale dataset demonstrate the state-of-the-art performance of DRPN.

Recent years have witnessed the improving performance of Chinese Named Entity Recognition (NER) from proposing new frameworks or incorporating word lexicons. However, the inner composition of entity mentions in character-level Chinese NER has been rarely studied. Actually, most mentions of regular types have strong name regularity. For example, entities end with indicator words such as “公司 (company) ” or “银行 (bank)” usually belong to organization. In this paper, we propose a simple but effective method for investigating the regularity of entity spans in Chinese NER, dubbed as Regularity-Inspired reCOgnition Network (RICON). Specifically, the proposed model consists of two branches: a regularity-aware module and a regularity-agnostic module. The regularity-aware module captures the internal regularity of each span for better entity type prediction, while the regularity-agnostic module is employed to locate the boundary of entities and relieve the excessive attention to span regularity. An orthogonality space is further constructed to encourage two modules to extract different aspects of regularity features. To verify the effectiveness of our method, we conduct extensive experiments on three benchmark datasets and a practical medical dataset. The experimental results show that our RICON significantly outperforms previous state-of-the-art methods, including various lexicon-based methods.