ACL 2023 Papers

Despite their impressive performance on diverse tasks, large language models (LMs) still struggle with tasks requiring rich world knowledge, implying the difficulty of encoding a wealth of world knowledge in their parameters. This paper aims to understand LMs’ strengths and limitations in memorizing factual knowledge, by conducting large-scale knowledge probing experiments on two open-domain entity-centric QA datasets: PopQA, our new dataset with 14k questions about long-tail entities, and EntityQuestions, a widely used open-domain QA dataset. We find that LMs struggle with less popular factual knowledge, and that retrieval augmentation helps significantly in these cases. Scaling, on the other hand, mainly improves memorization of popular knowledge, and fails to appreciably improve memorization of factual knowledge in the tail. Based on those findings, we devise a new method for retrieval-augmentation that improves performance and reduces inference costs by only retrieving non-parametric memories when necessary.

Large pre-trained language models (PLMs) have garnered significant attention for their versatility and potential for solving a wide spectrum of natural language processing (NLP) tasks. However, the cost of running these PLMs may be prohibitive. Furthermore, PLMs may not be open-sourced due to commercial considerations and potential risks of misuse, such as GPT-3. The parameters and gradients of PLMs are unavailable in this scenario. To solve the issue, black-box tuning has been proposed, which utilizes derivative-free optimization (DFO), instead of gradient descent, for training task-specific continuous prompts. However, these gradient-free methods still exhibit a significant gap compared to gradient-based methods. In this paper, we introduce gradient descent into black-box tuning scenario through knowledge distillation. Furthermore, we propose a novel method GDFO, which integrates gradient descent and derivative-free optimization to optimize task-specific continuous prompts in a harmonized manner. Experimental results show that GDFO can achieve significant performance gains over previous state-of-the-art methods.

Although proper handling of discourse significantly contributes to the quality of machine translation (MT), these improvements are not adequately measured in common translation quality metrics. Recent works in context-aware MT attempt to target a small set of discourse phenomena during evaluation, however not in a fully systematic way. In this paper, we develop the Multilingual Discourse-Aware (MuDA) benchmark, a series of taggers that identify and evaluate model performance on discourse phenomena in any given dataset. The choice of phenomena is inspired by a novel methodology to systematically identify translations that require context. This methodology confirms the difficulty of previously studied phenomena while uncovering others which were not previously addressed. We find that commonly studied context-aware MT models make only marginal improvements over context-agnostic models, which suggests these models do not handle these ambiguities effectively. We release code and data for 14 language pairs to encourage the MT community to focus on accurately capturing discourse phenomena. Code available at https://github.com/neulab/contextual-mt

Multi-task learning (MTL) aims at achieving a better model by leveraging data and knowledge from multiple tasks. However, MTL does not always work – sometimes negative transfer occurs between tasks, especially when aggregating loosely related skills, leaving it an open question when MTL works. Previous studies show that MTL performance can be improved by algorithmic tricks. However, what tasks and skills should be included is less well explored. In this work, we conduct a case study in Financial NLP where multiple datasets exist for skills relevant to the domain, such as numeric reasoning and sentiment analysis. Due to the task difficulty and data scarcity in the Financial NLP domain, we explore when aggregating such diverse skills from multiple datasets with MTL can work. Our findings suggest that the key to MTL success lies in skill diversity, relatedness between tasks, and choice of aggregation size and shared capacity. Specifically, MTL works well when tasks are diverse but related, and when the size of the task aggregation and the shared capacity of the model are balanced to avoid overwhelming certain tasks.

In the last five years, there has been a significant focus in Natural Language Processing (NLP) on developing larger Pretrained Language Models (PLMs) and introducing benchmarks such as SuperGLUE and SQuAD to measure their abilities in language understanding, reasoning, and reading comprehension. These PLMs have achieved impressive results on these benchmarks, even surpassing human performance in some cases. This has led to claims of superhuman capabilities and the provocative idea that certain tasks have been solved. In this position paper, we take a critical look at these claims and ask whether PLMs truly have superhuman abilities and what the current benchmarks are really evaluating. We show that these benchmarks have serious limitations affecting the comparison between humans and PLMs and provide recommendations for fairer and more transparent benchmarks.

This paper presents a method for building a personalized open-domain dialogue system to address the WWH (WHAT, WHEN, and HOW) problem for natural response generation in a commercial setting, where personalized dialogue responses are heavily interleaved with casual response turns. The proposed approach involves weighted dataset blending, negative persona information augmentation methods, and the design of personalized conversation datasets to address the challenges of WWH in personalized, open-domain dialogue systems. Our work effectively balances dialogue fluency and tendency to ground, while also introducing a response-type label to improve the controllability and explainability of the grounded responses. The combination of these methods leads to more fluent conversations, as evidenced by subjective human evaluations as well as objective evaluations.

Mental health disorders (MHD) are increasingly prevalent worldwide and constitute one of the greatest challenges facing our healthcare systems and modern societies in general. In response to this societal challenge, there has been a surge in digital mental health research geared towards the development of new techniques for unobtrusive and efficient automatic detection of MHD. Within this area of research, natural language processing techniques are playing an increasingly important role, showing promising detection results from a variety of textual data. Recently, there has been a growing interest in improving mental illness detection from textual data by way of leveraging emotions: ‘Emotion fusion’ refers to the process of integrating emotion information with general textual information to obtain enhanced information for decision-making. However, while the available research has shown that MHD prediction can be improved through a variety of different fusion strategies, previous works have been confined to a particular fusion strategy applied to a specific dataset, and so is limited by the lack of meaningful comparability. In this work, we integrate and extend this research by conducting extensive experiments with three types of deep learning-based fusion strategies: (i) feature-level fusion, where a pre-trained masked language model for mental health detection (MentalRoBERTa) was infused with a comprehensive set of engineered features, (ii) model fusion, where the MentalRoBERTa model was infused with hidden representations of other language models and (iii) task fusion, where a multi-task framework was leveraged to learn the features for auxiliary tasks. In addition to exploring the role of different fusion strategies, we expand on previous work by broadening the information infusion to include a second domain related to mental health, namely personality. We evaluate algorithm performance on data from two benchmark datasets, encompassing five mental health conditions: attention deficit hyperactivity disorder, anxiety, bipolar disorder, depression and psychological stress.

Diffusion models are a milestone in text-to-image generation, but they remain poorly understood, lacking interpretability analyses. In this paper, we perform a text-image attribution analysis on Stable Diffusion, a recently open-sourced model. To produce attribution maps, we upscale and aggregate cross-attention maps in the denoising module, naming our method DAAM. We validate it by testing its segmentation ability on nouns, as well as its generalized attribution quality on all parts of speech, rated by humans. On two generated datasets, we attain a competitive 58.8-64.8 mIoU on noun segmentation and fair to good mean opinion scores (3.4-4.2) on generalized attribution. Then, we apply DAAM to study the role of syntax in the pixel space across head–dependent heat map interaction patterns for ten common dependency relations. We show that, for some relations, the head map consistently subsumes the dependent, while the opposite is true for others. Finally, we study several semantic phenomena, focusing on feature entanglement; we find that the presence of cohyponyms worsens generation quality by 9%, and descriptive adjectives attend too broadly. We are the first to interpret large diffusion models from a visuolinguistic perspective, which enables future research. Our code is at https://github.com/castorini/daam.

Much research has sought to evaluate the degree to which large language models reflect social biases. We complement such work with an approach to elucidating the connections between language model predictions and people’s social attitudes. We show how word preferences in a large language model reflect social attitudes about gender, using two datasets from human experiments that found differences in gendered or gender neutral word choices by participants with differing views on gender (progressive, moderate, or conservative). We find that the language model BERT takes into account factors that shape human lexical choice of such language, but may not weigh those factors in the same way people do. Moreover, we show that BERT’s predictions most resemble responses from participants with moderate to conservative views on gender. Such findings illuminate how a language model: (1) may differ from people in how it deploys words that signal gender, and (2) may prioritize some social attitudes over others.

This paper aims to benchmark recent progress in language understanding models that output contextualised representations at the character level. Many such modelling architectures and methods to train those architectures have been proposed, but it is currently unclear what the relative contributions of the architecture vs. the pretraining objective are to final model performance. We explore the design space of such models, comparing architectural innovations (Clark et al., 2022, Jaegle et al., 2022, Tay et al., 2021) and a variety of different pretraining objectives on a suite of evaluation tasks with a fixed training procedure in order to find the currently optimal way to build and train character-level BERT-like models. We find that our best performing character-level model exceeds the performance of a token-based model trained with the same settings on the same data, suggesting that character-level models are ready for more widespread adoption. Unfortunately, the best method to train character-level models still relies on a subword-level tokeniser during pretraining, and final model performance is highly dependent on tokeniser quality. We believe our results demonstrate the readiness of character-level models for multilingual language representation, and encourage NLP practitioners to try them as drop-in replacements for token-based models.

Asking and answering questions are inseparable parts of human social life. The primary purposes of asking questions are to gain knowledge or request help which has been the subject of question-answering studies. However, questions can also reflect negative intentions and include implicit offenses, such as highlighting one’s lack of knowledge or bolstering an alleged superior knowledge, which can lead to conflict in conversations; yet has been scarcely researched. This paper is the first study to introduce a dataset (Question Intention Dataset) that includes questions with positive/neutral and negative intentions and the underlying intention categories within each group. We further conduct a meta-analysis to highlight tacit and apparent intents. We also propose a classification method using Transformers augmented by TF-IDF-based features and report the results of several models for classifying the main intention categories. We aim to highlight the importance of taking intentions into account, especially implicit and negative ones, to gain insight into conflict-evoking questions and better understand human-human communication on the web for NLP applications.

Humans can effortlessly understand the coordinate structure of sentences such as “Niels Bohr and Kurt Cobain were born in Copenhagen and Seattle, *respectively*”. In the context of natural language inference (NLI), we examine how language models (LMs) reason with respective readings (Gawron and Kehler, 2004) from two perspectives: syntactic-semantic and commonsense-world knowledge. We propose a controlled synthetic dataset WikiResNLI and a naturally occurring dataset NatResNLI to encompass various explicit and implicit realizations of “respectively”. We show that fine-tuned NLI models struggle with understanding such readings without explicit supervision. While few-shot learning is easy in the presence of explicit cues, longer training is required when the reading is evoked implicitly, leaving models to rely on common sense inferences. Furthermore, our fine-grained analysis indicates models fail to generalize across different constructions. To conclude, we demonstrate that LMs still lag behind humans in generalizing to the long tail of linguistic constructions.

Moral rhetoric influences our judgement. Although social scientists recognize moral expression as domain specific, there are no systematic methods for analyzing whether a text classifier learns the domain-specific expression of moral language or not. We propose Tomea, a method to compare a supervised classifier’s representation of moral rhetoric across domains. Tomea enables quantitative and qualitative comparisons of moral rhetoric via an interpretable exploration of similarities and differences across moral concepts and domains. We apply Tomea on moral narratives in thirty-five thousand tweets from seven domains. We extensively evaluate the method via a crowd study, a series of cross-domain moral classification comparisons, and a qualitative analysis of cross-domain moral expression.

Summarization models often generate text that is poorly calibrated to quality metrics because they are trained to maximize the likelihood of a single reference (MLE). To address this, recent work has added a calibration step, which exposes a model to its own ranked outputs to improve relevance or, in a separate line of work, contrasts positive and negative sets to improve faithfulness. While effective, much of this work has focused on how to generate and optimize these sets. Less is known about why one setup is more effective than another. In this work, we uncover the underlying characteristics of effective sets. For each training instance, we form a large, diverse pool of candidates and systematically vary the subsets used for calibration fine-tuning. Each selection strategy targets distinct aspects of the sets, such as lexical diversity or the size of the gap between positive and negatives. On three diverse scientific long-form summarization datasets (spanning biomedical, clinical, and chemical domains), we find, among others, that faithfulness calibration is optimal when the negative sets are extractive and more likely to be generated, whereas for relevance calibration, the metric margin between candidates should be maximized and surprise–the disagreement between model and metric defined candidate rankings–minimized.

As 3rd-person pronoun usage shifts to include novel forms, e.g., neopronouns, we need more research on identity-inclusive NLP. Exclusion is particularly harmful in one of the most popular NLP applications, machine translation (MT). Wrong pronoun translations can discriminate against marginalized groups, e.g., non-binary individuals (Dev et al., 2021). In this “reality check”, we study how three commercial MT systems translate 3rd-person pronouns. Concretely, we compare the translations of gendered vs. gender-neutral pronouns from English to five other languages (Danish, Farsi, French, German, Italian), and vice versa, from Danish to English.Our error analysis shows that the presence of a gender-neutral pronoun often leads to grammatical and semantic translation errors. Similarly, gender neutrality is often not preserved. By surveying the opinions of affected native speakers from diverse languages, we provide recommendations to address the issue in future MT research.

Current methods for prompt learning in zero-shot scenarios widely rely on a development set with sufficient human-annotated data to select the best-performing prompt template a posteriori. This is not ideal because in a real-world zero-shot scenario of practical relevance, no labelled data is available. Thus, we propose a simple yet effective method for screening reasonable prompt templates in zero-shot text classification: Perplexity Selection (Perplection). We hypothesize that language discrepancy can be used to measure the efficacy of prompt templates, and thereby develop a substantiated perplexity-based scheme allowing for forecasting the performance of prompt templates in advance. Experiments show that our method leads to improved prediction performance in a realistic zero-shot setting, eliminating the need for any labelled examples.

kNN-MT presents a new paradigm for domain adaptation by building an external datastore, which usually saves all target language token occurrences in the parallel corpus. As a result, the constructed datastore is usually large and possibly redundant. In this paper, we investigate the interpretability issue of this approach: what knowledge does the NMT model need? We propose the notion of local correctness (LAC) as a new angle, which describes the potential translation correctness for a single entry and for a given neighborhood. Empirical study shows that our investigation successfully finds the conditions where the NMT model could easily fail and need related knowledge. Experiments on six diverse target domains and two language-pairs show that pruning according to local correctness brings a light and more explainable memory for kNN-MT domain adaptation.

The EAE task extracts a structured event record from an event text. Most existing approaches train the EAE model on each dataset independently and ignore the overlap knowledge across datasets. However, insufficient event records in a single dataset often prevent the existing model from achieving better performance. In this paper, we clearly define the overlap knowledge across datasets and split the knowledge of the EAE task into overlap knowledge across datasets and specific knowledge of the target dataset. We propose APE model to learn the two parts of knowledge in two serial learning phases without causing catastrophic forgetting. In addition, we formulate both learning phases as conditional generation tasks and design Stressing Entity Type Prompt to close the gap between the two phases. The experiments show APE achieves new state-of-the-art with a large margin in the EAE task. When only ten records are available in the target dataset, our model dramatically outperforms the baseline model with average 27.27% F1 gain.

Large language models (LLMs) exploit in-context learning (ICL) to solve tasks with only a few demonstrations, but its mechanisms are not yet well-understood. Some works suggest that LLMs only recall already learned concepts from pre-training, while others hint that ICL performs implicit learning over demonstrations. We characterize two ways through which ICL leverages demonstrations. Task recognition (TR) captures the extent to which LLMs can recognize a task through demonstrations – even without ground-truth labels – and apply their pre-trained priors, whereas task learning (TL) is the ability to capture new input-label mappings unseen in pre-training. Using a wide range of classification datasets and three LLM families (GPT-3, LLaMA and OPT), we design controlled experiments to disentangle the roles of TR and TL in ICL. We show that (1) models can achieve non-trivial performance with only TR, and TR does not further improve with larger models or more demonstrations; (2) LLMs acquire TL as the model scales, and TL’s performance consistently improves with more demonstrations in context. Our findings unravel two different forces behind ICL and we advocate for discriminating them in future ICL research due to their distinct nature.