EMNLP 2022 Papers

With a handful of demonstration examples, large-scale language models demonstrate strong capability to perform various tasks by in-context learning from these examples, without any fine-tuning. We demonstrate that in-context learning performance can be highly unstable across samples of examples, indicating the idiosyncrasies of how language models acquire information. We formulate example selection for in-context learning as a sequential decision problem, and propose a reinforcement learning algorithm for identifying generalizable policies to select demonstration examples. For GPT-2, our learned policies demonstrate strong abilities of generalizing to unseen tasks in training, with a 5.8% improvement on average. Examples selected from our learned policies can even achieve a small improvement on GPT-3 Ada. However, the improvement diminishes on larger GPT-3 models, suggesting emerging capabilities of large language models.

The role of the attention mechanism in encoding linguistic knowledge has received special interest in NLP. However, the ability of the attention heads to judge the grammatical acceptability of a sentence has been underexplored. This paper approaches the paradigm of acceptability judgments with topological data analysis (TDA), showing that the geometric properties of the attention graph can be efficiently exploited for two standard practices in linguistics: binary judgments and linguistic minimal pairs. Topological features enhance the BERT-based acceptability classifier scores by 8%-24% on CoLA in three languages (English, Italian, and Swedish). By revealing the topological discrepancy between attention maps of minimal pairs, we achieve the human-level performance on the BLiMP benchmark, outperforming nine statistical and Transformer LM baselines. At the same time, TDA provides the foundation for analyzing the linguistic functions of attention heads and interpreting the correspondence between the graph features and grammatical phenomena. We publicly release the code and other materials used in the experiments.

Medical literature is a vast and constantly expanding source of information about diseases, their diagnoses and treatments. One of the ways to extract insights from this type of data is through mining association rules between such entities. However, existing solutions do not take into account the semantics of sentences from which entity co-occurrences are extracted. We propose a scalable solution for the automated discovery of semantic associations between different entities such as diseases and their symptoms. Our approach employs the UMLS semantic network and a binary relation classification model trained with distant supervision to validate and help ranking the most likely entity associations pairs extracted with frequency-based association rule mining algorithms. We evaluate the proposed system on the task of extracting disease-symptom associations from a collection of over 14M PubMed abstracts and validate our results against a publicly available known list of disease-symptom pairs.

Novel inverted index-based learned sparse ranking models provide more effective, but less efficient, retrieval performance compared to traditional ranking models like BM25. In this paper, we introduce a technique we call postings clipping to improve the query efficiency of learned representations. Our technique amplifies the benefit of dynamic pruning query processing techniques by accounting for changes in term importance distributions of learned ranking models. The new clipping mechanism accelerates top-k retrieval by up to 9.6X without any loss in effectiveness.

Sequential abstractive neural summarizers often do not use the underlying structure in the input article or dependencies between the input sentences. This structure is essential to integrate and consolidate information from different parts of the text. To address this shortcoming, we propose a hierarchy-aware graph neural network (HierGNN) which captures such dependencies through three main steps: 1) learning a hierarchical document structure through a latent structure tree learned by a sparse matrix-tree computation; 2) propagating sentence information over this structure using a novel message-passing node propagation mechanism to identify salient information; 3) using graph-level attention to concentrate the decoder on salient information. Experiments confirm HierGNN improves strong sequence models such as BART, with a 0.55 and 0.75 margin in average ROUGE-1/2/L for CNN/DM and XSum. Further human evaluation demonstrates that summaries produced by our model are more relevant and less redundant than the baselines, into which HierGNN is incorporated. We also find HierGNN synthesizes summaries by fusing multiple source sentences more, rather than compressing a single source sentence, and that it processes long inputs more effectively.

We introduce KiloGram, a resource for studying abstract visual reasoning in humans and machines. Drawing on the history of tangram puzzles as stimuli in cognitive science, we build a richly annotated dataset that, with >1k distinct stimuli, is orders of magnitude larger and more diverse than prior resources. It is both visually and linguistically richer, moving beyond whole shape descriptions to include segmentation maps and part labels. We use this resource to evaluate the abstract visual reasoning capacities of recent multi-modal models. We observe that pre-trained weights demonstrate limited abstract reasoning, which dramatically improves with fine-tuning. We also observe that explicitly describing parts aids abstract reasoning for both humans and models, especially when jointly encoding the linguistic and visual inputs.

Aspect Based Sentiment Analysis is a dominant research area with potential applications in social media analytics, business, finance, and health. Prior works in this area are primarily based on supervised methods, with a few techniques using weak supervision limited to predicting a single aspect category per review sentence. In this paper, we present an extremely weakly supervised multi-label Aspect Category Sentiment Analysis framework which does not use any labelled data. We only rely on a single word per class as an initial indicative information. We further propose an automatic word selection technique to choose these seed categories and sentiment words. We explore unsupervised language model post-training to improve the overall performance, and propose a multi-label generator model to generate multiple aspect category-sentiment pairs per review sentence. Experiments conducted on four benchmark datasets showcase our method to outperform other weakly supervised baselines by a significant margin.

This work introduces a new multi-task, parameter-efficient language model (LM) tuning method that learns to transfer knowledge across different tasks via a mixture of soft prompts—small prefix embedding vectors pre-trained for different tasks. Our method, called ATTEMPT (ATTEntional Mixtures of Prompt Tuning), obtains source prompts as encodings of large-scale source tasks into a small number of parameters and trains an attention module to interpolate the source prompts and a newly initialized target prompt for every instance in the target task. During training, only the target task prompt and the attention weights, which are shared between tasks in multi-task training, are updated, while the original LM and source prompts are intact. ATTEMPT is highly parameter-efficient (e.g., updates 2,300 times fewer parameters than full fine-tuning), while it overcomes instability of prompt tuning and achieves high task performance using learned knowledge from high-resource tasks. Moreover, it is modular using pre-trained soft prompts, and can flexibly add or remove source prompts for effective knowledge transfer. Our experimental results across 21 diverse NLP datasets show that ATTEMPT significantly outperforms prompt tuning and outperforms or matches fully fine-tuned or other parameter-efficient tuning approaches that use 10 times more parameters. Finally, ATTEMPT outperforms previous work in few-shot learning settings.

Recent progress on open domain factoid question answering (QA) does not easily transfer to the task of long-form QA, where the goal is to answer questions that require in-depth explanations. The hurdles include a lack of high-quality data and the absence of a well-defined notion of an answer’s quality. In this work, we address these problems by releasing a novel dataset and a task that we call ASQA (Answer Summaries for Questions which are Ambiguous); and proposing a reliable metric for measuring performance on ASQA. Our task focuses on ambiguous factoid questions which have different correct answers depending on the interpretation. Answers to ambiguous questions should combine factual information from multiple sources into a coherent long-form summary that resolves the ambiguity. In contrast to existing long-form QA tasks (such as ELI5), ASQA admits a clear notion of correctness: a user faced with a good summary should be able to answer different interpretations of the original ambiguous question. Our analysis demonstrates an agreement between this metric and human judgments, and reveals a considerable gap between human performance and strong baselines.

Data-to-text generation is challenging due to the great variety of the input data in terms of domains (e.g., finance vs sports) or schemata (e.g., diverse predicates). Recent end-to-end neural methods thus require substantial training examples to learn to disambiguate and describe the data. Yet, real-world data-to-text problems often suffer from various data-scarce issues: one may have access to only a handful of or no training examples, and/or have to rely on examples in a different domain or schema. To fill this gap, we propose Any-Shot Data-to-Text (ASDOT), a new approach flexibly applicable to diverse settings by making efficient use of any given (or no) examples. ASDOT consists of two steps, data disambiguation and sentence fusion, both of which are amenable to be solved with off-the-shelf pretrained language models (LMs) with optional finetuning. In the data disambiguation stage, we employ the prompted GPT-3 model to understand possibly ambiguous triples from the input data and convert each into a short sentence with reduced ambiguity. The sentence fusion stage then uses an LM like T5 to fuse all the resulting sentences into a coherent paragraph as the final description. We evaluate extensively on various datasets in different scenarios, including the zero-/few-/full-shot settings, and generalization to unseen predicates and out-of-domain data. Experimental results show that ASDOT consistently achieves significant improvement over baselines, e.g., a 30.81 BLEU gain on the DART dataset under the zero-shot setting.

Text-to-Image Synthesis (TIS) is a popular task to convert natural language texts into realistic images. Recently, transformer-based TIS models (such as DALL-E) have been proposed using the encoder-decoder architectures. Yet, these billion-scale TIS models are difficult to tune and deploy in resource-constrained environments. In addition, there is a lack of language-specific TIS benchmarks for Chinese, together with high-performing models with moderate sizes. In this work, we present ARTIST, A tRansformer-based Chinese Text-to-Image SynThesizer for high-resolution image generation. In ARTIST, the rich linguistic and relational knowledge facts are injected into the model to ensure better model performance without the usage of ultra-large models. We further establish a large-scale Chinese TIS benchmark with the re-production results of state-of-the-art transformer-based TIS models.Results show ARTIST outperforms previous approaches.

In hate speech detection, developing training and evaluation datasets across various domains is the critical issue. Whereas, major approaches crawl social media texts and hire crowd-workers to annotate the data. Following this convention often restricts the scope of pejorative expressions to a single domain lacking generalization. Sometimes domain overlap between training corpus and evaluation set overestimate the prediction performance when pretraining language models on low-data language. To alleviate these problems in Korean, we propose APEACH that asks unspecified users to generate hate speech examples followed by minimal post-labeling. We find that APEACH can collect useful datasets that are less sensitive to the lexical overlaps between the pretraining corpus and the evaluation set, thereby properly measuring the model performance.

NLP has advanced greatly together with the proliferation of Transformer-based pre-trained language models. To adapt to a downstream task, the pre-trained language models need to be fine-tuned with a sufficient supply of annotated examples. In recent years, Adapter-based fine-tuning methods have expanded the applicability of pre-trained language models by substantially lowering the required amount of annotated examples. However, existing Adapter-based methods still fail to yield meaningful results in the few-shot regime where only a few annotated examples are provided. In this study, we present a meta-learning-driven low-rank adapter pooling method, called AMAL, for leveraging pre-trained language models even with just a few data points. We evaluate our method on five text classification benchmark datasets. The results show that AMAL significantly outperforms previous few-shot learning methods and achieves a new state-of-the-art.

Embodied Vision and Language Task Completion requires an embodied agent to interpret natural language instructions and egocentric visual observations to navigate through and interact with environments. In this work, we examine ALFRED, a challenging benchmark for embodied task completion, with the goal of gaining insight into how effectively models utilize language. We find evidence that sequence-to-sequence and transformer-based models trained on this benchmark are not sufficiently sensitive to changes in input language instructions. Next, we construct a new test split – ALFRED-L to test whether ALFRED models can generalize to task structures not seen during training that intuitively require the same types of language understanding required in ALFRED. Evaluation of existing models on ALFRED-L suggests that (a) models are overly reliant on the sequence in which objects are visited in typical ALFRED trajectories and fail to adapt to modifications of this sequence and (b) models trained with additional augmented trajectories are able to adapt relatively better to such changes in input language instructions.

Argument generation is an important but challenging task in computational argumentation.Existing studies have mainly focused on generating individual short arguments, while research on generating long and coherent argumentative essays is still under-explored.In this paper, we propose a new task, Argumentative Essay Generation (AEG).Given a writing prompt, the goal of AEG is to automatically generate an argumentative essay with strong persuasiveness.We construct a large-scale dataset, ArgEssay, for this new task and establish a strong model based on a dual-decoder Transformer architecture.Our proposed model contains two decoders, a planning decoder (PD) and a writing decoder (WD), where PD is used to generate a sequence for essay content planning and WD incorporates the planning information to write an essay.Further, we pre-train this model on a large news dataset to enhance the plan-and-write paradigm.Automatic and human evaluation results show that our model can generate more coherent and persuasive essays with higher diversity and less repetition compared to several baselines.

Adversarial Examples Detection (AED) is a crucial defense technique against adversarial attacks and has drawn increasing attention from the Natural Language Processing (NLP) community. Despite the surge of new AED methods, our studies show that existing methods heavily rely on a shortcut to achieve good performance. In other words, current search-based adversarial attacks in NLP stop once model predictions change, and thus most adversarial examples generated by those attacks are located near model decision boundaries. To surpass this shortcut and fairly evaluate AED methods, we propose to test AED methods with Far Boundary (FB) adversarial examples. Existing methods show worse than random guess performance under this scenario. To overcome this limitation, we propose a new technique, ADDMU, adversary detection with data and model uncertainty, which combines two types of uncertainty estimation for both regular and FB adversarial example detection. Our new method outperforms previous methods by 3.6 and 6.0 AUC points under each scenario. Finally, our analysis shows that the two types of uncertainty provided by ADDMU can be leveraged to characterize adversarialexamples and identify the ones that contribute most to model’s robustness in adversarial training.

Augmenting pre-trained language models (PLMs) with knowledge graphs (KGs) has demonstrated superior performance on commonsense reasoning. Given a commonsense based QA context (question and multiple choices), existing approaches usually estimate the plausibility of candidate choices separately based on their respective retrieved KGs, without considering the interference among different choices. In this paper, we propose an Attention guided Commonsense rEasoning Network (ACENet) to endow the neural network with the capability of integrating hybrid knowledge. Specifically, our model applies the multi-layer interaction of answer choices to continually strengthen correct choice information and guide the message passing of GNN. In addition, we also design a mix attention mechanism of nodes and edges to iteratively select supporting evidence on hybrid knowledge graph. Experimental results demonstrate the effectiveness of our proposed model through considerable performance gains across CommonsenseQA and OpenbookQA datasets.

Document-level relation extraction (RE) aims to identify relations between entities across multiple sentences. Most previous methods focused on document-level RE under full supervision. However, in real-world scenario, it is expensive and difficult to completely label all relations in a document because the number of entity pairs in document-level RE grows quadratically with the number of entities. To solve the common incomplete labeling problem, we propose a unified positive-unlabeled learning framework - shift and squared ranking loss positive-unlabeled (SSR-PU) learning. We use positive-unlabeled (PU) learning on document-level RE for the first time. Considering that labeled data of a dataset may lead to prior shift of unlabeled data, we introduce a PU learning under prior shift of training data. Also, using none-class score as an adaptive threshold, we propose squared ranking loss and prove its Bayesian consistency with multi-label ranking metrics. Extensive experiments demonstrate that our method achieves an improvement of about 14 F1 points relative to the previous baseline with incomplete labeling. In addition, it outperforms previous state-of-the-art results under both fully supervised and extremely unlabeled settings as well.

Readability assessment is a basic research task in the field of education. Traditional methods mainly employ machine learning classifiers with hundreds of linguistic features. Although the deep learning model has become the prominent approach for almost all NLP tasks, it is less explored for readability assessment. In this paper, we propose a BERT-based model with feature projection and length-balanced loss (BERT-FP-LBL) to determine the difficulty level of a given text. First, we introduce topic features guided by difficulty knowledge to complement the traditional linguistic features. From the linguistic features, we extract really useful orthogonal features to supplement BERT representations by means of projection filtering. Furthermore, we design a length-balanced loss to handle the greatly varying length distribution of the readability data. We conduct experiments on three English benchmark datasets and one Chinese dataset, and the experimental results show that our proposed model achieves significant improvements over baseline models. Interestingly, our proposed model achieves comparable results with human experts in consistency test.