TACL 2024 论文列表

Widely used learned metrics for machine translation evaluation, such as COMET and BLEURT, estimate the quality of a translation hypothesis by providing a single sentence-level score. As such, they offer little insight into translation errors (e.g., what are the errors and what is their severity). On the other hand, generative large language models (LLMs) are amplifying the adoption of more granular strategies to evaluation, attempting to detail and categorize translation errors. In this work, we introduce xCOMET, an open-source learned metric designed to bridge the gap between these approaches. xCOMET integrates both sentence-level evaluation and error span detection capabilities, exhibiting state-of-the-art performance across all types of evaluation (sentence-level, system-level, and error span detection). Moreover, it does so while highlighting and categorizing error spans, thus enriching the quality assessment. We also provide a robustness analysis with stress tests, and show that xCOMET is largely capable of identifying localized critical errors and hallucinations.

Creoles represent an under-explored and marginalized group of languages, with few available resources for NLP research. While the genealogical ties between Creoles and a number of highly resourced languages imply a significant potential for transfer learning, this potential is hampered due to this lack of annotated data. In this work we present CreoleVal, a collection of benchmark datasets spanning 8 different NLP tasks, covering up to 28 Creole languages; it is an aggregate of novel development datasets for reading comprehension relation classification, and machine translation for Creoles, in addition to a practical gateway to a handful of preexisting benchmarks. For each benchmark, we conduct baseline experiments in a zero-shot setting in order to further ascertain the capabilities and limitations of transfer learning for Creoles. Ultimately, we see CreoleVal as an opportunity to empower research on Creoles in NLP and computational linguistics, and in general, a step towards more equitable language technology around the globe.

Recent advances in LLMs have led to an abundance of evaluation benchmarks, which typically rely on a single instruction template per task. We create a large-scale collection of instruction paraphrases and comprehensively analyze the brittleness introduced by single-prompt evaluations across 6.5M instances, involving 20 different LLMs and 39 tasks from 3 benchmarks. We find that different instruction templates lead to very different performance, both absolute and relative. Instead, we propose a set of diverse metrics on multiple instruction paraphrases, specifically tailored for different use cases (e.g., LLM vs. downstream development), ensuring a more reliable and meaningful assessment of LLM capabilities. We show that our metrics provide new insights into the strengths and limitations of current LLMs.

Compositional Natural Language Inference (NLI) has been explored to assess the true abilities of neural models to perform NLI. Yet, current evaluations assume models to have full access to all primitive inferences in advance, in contrast to humans that continuously acquire inference knowledge. In this paper, we introduce the Continual Compositional Generalization in Inference (C2Gen NLI) challenge, where a model continuously acquires knowledge of constituting primitive inference tasks as a basis for compositional inferences. We explore how continual learning affects compositional generalization in NLI, by designing a continual learning setup for compositional NLI inference tasks. Our experiments demonstrate that models fail to compositionally generalize in a continual scenario. To address this problem, we first benchmark various continual learning algorithms and verify their efficacy. We then further analyze C2Gen, focusing on how to order primitives and compositional inference types, and examining correlations between subtasks. Our analyses show that by learning subtasks continuously while observing their dependencies and increasing degrees of difficulty, continual learning can enhance composition generalization ability.1

We describe a class of tasks called decision-oriented dialogues, in which AI assistants such as large language models (LMs) must collaborate with one or more humans via natural language to help them make complex decisions. We formalize three domains in which users face everyday decisions: (1) choosing an assignment of reviewers to conference papers, (2) planning a multi-step itinerary in a city, and (3) negotiating travel plans for a group of friends. In each of these settings, AI assistants and users have disparate abilities that they must combine to arrive at the best decision: Assistants can access and process large amounts of information, while users have preferences and constraints external to the system. For each task, we build a dialogue environment where agents receive a reward based on the quality of the final decision they reach. We evaluate LMs in self-play and in collaboration with humans and find that they fall short compared to human assistants, achieving much lower rewards despite engaging in longer dialogues. We highlight a number of challenges models face in decision-oriented dialogues, ranging from goal-directed behavior to reasoning and optimization, and release our environments as a testbed for future work.

Authorship verification is the task of determining if two distinct writing samples share the same author and is typically concerned with the attribution of written text. In this paper, we explore the attribution of transcribed speech, which poses novel challenges. The main challenge is that many stylistic features, such as punctuation and capitalization, are not informative in this setting. On the other hand, transcribed speech exhibits other patterns, such as filler words and backchannels (e.g., um, uh-huh), which may be characteristic of different speakers. We propose a new benchmark for speaker attribution focused on human-transcribed conversational speech transcripts. To limit spurious associations of speakers with topic, we employ both conversation prompts and speakers participating in the same conversation to construct verification trials of varying difficulties. We establish the state of the art on this new benchmark by comparing a suite of neural and non-neural baselines, finding that although written text attribution models achieve surprisingly good performance in certain settings, they perform markedly worse as conversational topic is increasingly controlled. We present analyses of the impact of transcription style on performance as well as the ability of fine-tuning on speech transcripts to improve performance.1

Despite significant improvements in enhancing the quality of translation, context-aware machine translation (MT) models underperform in many cases. One of the main reasons is that they fail to utilize the correct features from context when the context is too long or their models are overly complex. This can lead to the explain-away effect, wherein the models only consider features easier to explain predictions, resulting in inaccurate translations. To address this issue, we propose a model that explains the decisions made for translation by predicting coreference features in the input. We construct a model for input coreference by exploiting contextual features from both the input and translation output representations on top of an existing MT model. We evaluate and analyze our method in the WMT document-level translation task of English-German dataset, the English-Russian dataset, and the multilingual TED talk dataset, demonstrating an improvement of over 1.0 BLEU score when compared with other context-aware models.

Metrics are the foundation for automatic evaluation in grammatical error correction (GEC), with their evaluation of the metrics (meta-evaluation) relying on their correlation with human judgments. However, conventional meta-evaluations in English GEC encounter several challenges, including biases caused by inconsistencies in evaluation granularity and an outdated setup using classical systems. These problems can lead to misinterpretation of metrics and potentially hinder the applicability of GEC techniques. To address these issues, this paper proposes SEEDA, a new dataset for GEC meta-evaluation. SEEDA consists of corrections with human ratings along two different granularities: edit-based and sentence-based, covering 12 state-of-the-art systems including large language models, and two human corrections with different focuses. The results of improved correlations by aligning the granularity in the sentence-level meta-evaluation suggest that edit-based metrics may have been underestimated in existing studies. Furthermore, correlations of most metrics decrease when changing from classical to neural systems, indicating that traditional metrics are relatively poor at evaluating fluently corrected sentences with many edits.

Classification systems are evaluated in a countless number of papers. However, we find that evaluation practice is often nebulous. Frequently, metrics are selected without arguments, and blurry terminology invites misconceptions. For instance, many works use so-called ‘macro’ metrics to rank systems (e.g., ‘macro F1’) but do not clearly specify what they would expect from such a ‘macro’ metric. This is problematic, since picking a metric can affect research findings and thus any clarity in the process should be maximized. Starting from the intuitive concepts of bias and prevalence, we perform an analysis of common evaluation metrics. The analysis helps us understand the metrics’ underlying properties, and how they align with expectations as found expressed in papers. Then we reflect on the practical situation in the field, and survey evaluation practice in recent shared tasks. We find that metric selection is often not supported with convincing arguments, an issue that can make a system ranking seem arbitrary. Our work aims at providing overview and guidance for more informed and transparent metric selection, fostering meaningful evaluation.

We address a growing debate about the extent to which large language models (LLMs) produce behavior consistent with Theory of Mind (ToM) in humans. We present EPITOME: a battery of six experiments that tap diverse ToM capacities, including belief attribution, emotional inference, and pragmatic reasoning. We elicit a performance baseline from human participants for each task. We use the dataset to ask whether distributional linguistic information learned by LLMs is sufficient to explain ToM in humans. We compare performance of five LLMs to a baseline of responses from human comprehenders. Results are mixed. LLMs display considerable sensitivity to mental states and match human performance in several tasks. Yet, they commit systematic errors in others, especially those requiring pragmatic reasoning on the basis of mental state information. Such uneven performance indicates that human-level ToM may require resources beyond distributional information.

Answering factual questions from heterogenous sources, such as graphs and text, is a key capacity of intelligent systems. Current approaches either (i) perform question answering over text and structured sources as separate pipelines followed by a merge step or (ii) provide an early integration, giving up the strengths of particular information sources. To solve this problem, we present “HumanIQ”, a method that teaches language models to dynamically combine retrieved information by imitating how humans use retrieval tools. Our approach couples a generic method for gathering human demonstrations of tool use with adaptive few-shot learning for tool augmented models. We show that HumanIQ confers significant benefits, including i) reducing the error rate of our strongest baseline (GPT-4) by over 50% across 3 benchmarks, (ii) improving human preference over responses from vanilla GPT-4 (45.3% wins, 46.7% ties, 8.0% loss), and (iii) outperforming numerous task-specific baselines.

Recent studies show that instruction tuning (IT) and reinforcement learning from human feedback (RLHF) improve the abilities of large language models (LMs) dramatically. While these tuning methods can help align models with human objectives and generate high-quality text, not much is known about their potential adverse effects. In this work, we investigate the effect of IT and RLHF on decision making and reasoning in LMs, focusing on three cognitive biases—the decoy effect, the certainty effect, and the belief bias—all of which are known to influence human decision-making and reasoning. Our findings highlight the presence of these biases in various models from the GPT-3, Mistral, and T5 families. Notably, we find a stronger presence of biases in models that have undergone instruction tuning, such as Flan-T5, Mistral-Instruct, GPT3.5, and GPT4. Our work constitutes a step toward comprehending cognitive biases in instruction-tuned LMs, which is crucial for the development of more reliable and unbiased language models.1

When children learn new words, they employ constraints such as the mutual exclusivity (ME) bias: A novel word is mapped to a novel object rather than a familiar one. This bias has been studied computationally, but only in models that use discrete word representations as input, ignoring the high variability of spoken words. We investigate the ME bias in the context of visually grounded speech models that learn from natural images and continuous speech audio. Concretely, we train a model on familiar words and test its ME bias by asking it to select between a novel and a familiar object when queried with a novel word. To simulate prior acoustic and visual knowledge, we experiment with several initialization strategies using pretrained speech and vision networks. Our findings reveal the ME bias across the different initialization approaches, with a stronger bias in models with more prior (in particular, visual) knowledge. Additional tests confirm the robustness of our results, even when different loss functions are considered. Based on detailed analyses to piece out the model’s representation space, we attribute the ME bias to how familiar and novel classes are distinctly separated in the resulting space.

Sentences containing multiple semantic operators with overlapping scope often create ambiguities in interpretation, known as scope ambiguities. These ambiguities offer rich insights into the interaction between semantic structure and world knowledge in language processing. Despite this, there has been little research into how modern large language models treat them. In this paper, we investigate how different versions of certain autoregressive language models—GPT-2, GPT-3/3.5, Llama 2, and GPT-4—treat scope ambiguous sentences, and compare this with human judgments. We introduce novel datasets that contain a joint total of almost 1,000 unique scope-ambiguous sentences, containing interactions between a range of semantic operators, and annotated for human judgments. Using these datasets, we find evidence that several models (i) are sensitive to the meaning ambiguity in these sentences, in a way that patterns well with human judgments, and (ii) can successfully identify human-preferred readings at a high level of accuracy (over 90% in some cases).1

Previous unsupervised domain adaptation (UDA) methods for question answering (QA) require access to source domain data while fine-tuning the model for the target domain. Source domain data may, however, contain sensitive information and should be protected. In this study, we investigate a more challenging setting, source-free UDA, in which we have only the pretrained source model and target domain data, without access to source domain data. We propose a novel self-training approach to QA models that integrates a specially designed mask module for domain adaptation. The mask is auto-adjusted to extract key domain knowledge when trained on the source domain. To maintain previously learned domain knowledge, certain mask weights are frozen during adaptation, while other weights are adjusted to mitigate domain shifts with pseudo-labeled samples generated in the target domain. Our empirical results on four benchmark datasets suggest that our approach significantly enhances the performance of pretrained QA models on the target domain, and even outperforms models that have access to the source data during adaptation.

The introduction of large public legal datasets has brought about a renaissance in legal NLP. Many of these datasets are composed of legal judgments—the product of judges deciding cases. Since ML algorithms learn to model the data they are trained on, several legal NLP models are models of judges. While some have argued for the automation of judges, in this position piece, we argue that automating the role of the judge raises difficult ethical challenges, in particular for common law legal systems. Our argument follows from the social role of the judge in actively shaping the law, rather than merely applying it. Since current NLP models are too far away from having the facilities necessary for this task, they should not be used to automate judges. Furthermore, even in the case that the models could achieve human-level capabilities, there would still be remaining ethical concerns inherent in the automation of the legal process.

Instruction-following models are attractive alternatives to fine-tuned approaches for question answering (QA). By simply prepending relevant documents and an instruction to their input, these models can be adapted to various information domains and tasks without additional training. However, these models tend to produce verbose responses with supplementary information, which makes traditional QA metrics like exact match (EM) and F1 unreliable for accurately quantifying model performance. In this work, we evaluate instruction-following models along two fronts: 1) how well they satisfy user’s information need (correctness), and 2) whether they disseminate information supported by the provided knowledge (faithfulness). Guided by human evaluation and analysis, we highlight the shortcomings of traditional metrics for both correctness and faithfulness and propose simple token-overlap metrics that correlate highly with human judgments. Our analysis reveals that for correctness, instruction-following models perform comparably to models specifically fine-tuned for that task. However, they struggle to accurately judge the relevance of the provided knowledge and often hallucinate in their responses. We hope our work encourages more holistic evaluation of instruction-following models for QA. Our code and human annotation data is available at https://github.com/McGill-NLP/instruct-qa.

Previous work in prompt engineering for large language models has introduced different gradient-free probability-based prompt selection methods that aim to choose the optimal prompt among the candidates for a given task but have failed to provide a comprehensive and fair comparison between each other. In this paper, we propose a unified framework to interpret and evaluate the existing probability-based prompt selection methods by performing extensive experiments on 13 common and diverse NLP tasks. We find that each of the existing methods can be interpreted as some variant of the method that maximizes mutual information between the input and the predicted output (MI). Utilizing this finding, we develop several other combinatorial variants of MI and increase the effectiveness of the oracle prompt selection method from 87.79% to 94.98%, measured as the ratio of the performance of the selected prompt to that of the optimal oracle prompt. Furthermore, considering that all the methods rely on the output probability distribution of the model that might be biased, we propose a novel calibration method called Calibration by Marginalization (CBM) that is orthogonal to the existing methods and helps increase the prompt selection effectiveness of the best method to 96.85%, achieving 99.44% of the oracle prompt F1 without calibration.1

This paper revisits a classical, yet fundamental, discussion of theoretical computational linguistics: the computational complexity of natural languages. Past studies have revealed that syntax, as observed in Swiss-German, is not weakly context-free. Concerning morphology, Culy (1985) employed a construction in Bambara to show that morphology is not weakly context-free; however, Manaster-Ramer (1988) pointed out that the Bambara case can be problematic because the wordhood of the construction is reliant on special tonal behaviors, and it is ambiguous whether the behaviors belong to the morphological domain. This raises doubts about whether the case can be considered a genuine morphological phenomenon. In this paper, we argue that Classical Ainu, a language we examine, also defies weak context-freeness at the morphological level. The construction we introduce is unambiguously morphological because this language’s valency-sensitive structure and valency-changing operations, such as noun incorporation, preclude its grammatical interpretation as syntactic.