ACL 2023 Papers

To align conditional text generation model outputs with desired behaviors, there has been an increasing focus on training the model using reinforcement learning (RL) with reward functions learned from human annotations. Under this framework, we identify three common cases where high rewards are incorrectly assigned to undesirable patterns: noise-induced spurious correlation, naturally occurring spurious correlation, and covariate shift. We show that even though learned metrics achieve high performance on the distribution of the data used to train the reward function, the undesirable patterns may be amplified during RL training of the text generation model. While there has been discussion about reward gaming in the RL or safety community, in this discussion piece, we would like to highlight reward gaming in the natural language generation (NLG) community using concrete conditional text generation examples and discuss potential fixes and areas for future work.

Human evaluation is the foundation upon which the evaluation of both summarization systems and automatic metrics rests. However, existing human evaluation studies for summarization either exhibit a low inter-annotator agreement or have insufficient scale, and an in-depth analysis of human evaluation is lacking. Therefore, we address the shortcomings of existing summarization evaluation along the following axes: (1) We propose a modified summarization salience protocol, Atomic Content Units (ACUs), which is based on fine-grained semantic units and allows for a high inter-annotator agreement. (2) We curate the Robust Summarization Evaluation (RoSE) benchmark, a large human evaluation dataset consisting of 22,000 summary-level annotations over 28 top-performing systems on three datasets. (3) We conduct a comparative study of four human evaluation protocols, underscoring potential confounding factors in evaluation setups. (4) We evaluate 50 automatic metrics and their variants using the collected human annotations across evaluation protocols and demonstrate how our benchmark leads to more statistically stable and significant results. The metrics we benchmarked include recent methods based on large language models (LLMs), GPTScore and G-Eval. Furthermore, our findings have important implications for evaluating LLMs, as we show that LLMs adjusted by human feedback (e.g., GPT-3.5) may overfit unconstrained human evaluation, which is affected by the annotators’ prior, input-agnostic preferences, calling for more robust, targeted evaluation methods.

Is the output softmax layer, which is adopted by most language models (LMs), always the best way to compute the next word probability? Given so many attention layers in a modern transformer-based LM, are the pointer networks redundant nowadays? In this study, we discover that the answers to both questions are no. This is because the softmax bottleneck sometimes prevents the LMs from predicting the desired distribution and the pointer networks can be used to break the bottleneck efficiently. Based on the finding, we propose several softmax alternatives by simplifying the pointer networks and accelerating the word-by-word rerankers. In GPT-2, our proposals are significantly better and more efficient than mixture of softmax, a state-of-the-art softmax alternative. In summarization experiments, without very significantly decreasing its training/testing speed, our best method based on T5-Small improves factCC score by 2 points in CNN/DM and XSUM dataset, and improves MAUVE scores by 30% in BookSum paragraph-level dataset.

Recent advancements in high-quality, large-scale English resources have pushed the frontier of English Automatic Text Simplification (ATS) research. However, less work has been done on multilingual text simplification due to the lack of a diverse evaluation benchmark that covers complex-simple sentence pairs in many languages. This paper introduces the MultiSim benchmark, a collection of 27 resources in 12 distinct languages containing over 1.7 million complex-simple sentence pairs. This benchmark will encourage research in developing more effective multilingual text simplification models and evaluation metrics. Our experiments using MultiSim with pre-trained multilingual language models reveal exciting performance improvements from multilingual training in non-English settings. We observe strong performance from Russian in zero-shot cross-lingual transfer to low-resource languages. We further show that few-shot prompting with BLOOM-176b achieves comparable quality to reference simplifications outperforming fine-tuned models in most languages. We validate these findings through human evaluation.

Token dropping is a recently-proposed strategy to speed up the pretraining of masked language models, such as BERT, by skipping the computation of a subset of the input tokens at several middle layers. It can effectively reduce the training time without degrading much performance on downstream tasks. However, we empirically find that token dropping is prone to a semantic loss problem and falls short in handling semantic-intense tasks. Motivated by this, we propose a simple yet effective semantic-consistent learning method (ScTD) to improve the token dropping. ScTD aims to encourage the model to learn how to preserve the semantic information in the representation space. Extensive experiments on 12 tasks show that, with the help of our ScTD, token dropping can achieve consistent and significant performance gains across all task types and model sizes. More encouragingly, ScTD saves up to 57% of pretraining time and brings up to +1.56% average improvement over the vanilla token dropping.

Tasks involving text generation based on multiple input texts, such as multi-document summarization, long-form question answering and contemporary dialogue applications, challenge models for their ability to properly consolidate partly-overlapping multi-text information. However, these tasks entangle the consolidation phase with the often subjective and ill-defined content selection requirement, impeding proper assessment of models’ consolidation capabilities. In this paper, we suggest revisiting the sentence union generation task as an effective well-defined testbed for assessing text consolidation capabilities, decoupling the consolidation challenge from subjective content selection. To support research on this task, we present refined annotation methodology and tools for crowdsourcing sentence union, create the largest union dataset to date and provide an analysis of its rich coverage of various consolidation aspects. We then propose a comprehensive evaluation protocol for union generation, including both human and automatic evaluation. Finally, as baselines, we evaluate state-of-the-art language models on the task, along with a detailed analysis of their capacity to address multi-text consolidation challenges and their limitations.

Knowing exactly how many data points need to be labeled to achieve a certain model performance is a hugely beneficial step towards reducing the overall budgets for annotation. It pertains to both active learning and traditional data annotation, and is particularly beneficial for low resource scenarios. Nevertheless, it remains a largely under-explored area of research in NLP. We therefore explored various techniques for estimating the training sample size necessary to achieve a targeted performance value. We derived a simple yet effective approach to predict the maximum achievable model performance based on small amount of training samples – which serves as an early indicator during data annotation for data quality and sample size determination. We performed ablation studies on four language understanding tasks, and showed that the proposed approach allows us to forecast model performance within a small margin of mean absolute error (~0.9%) with only 10% data.

Relation extraction (RE) is the core NLP task of inferring semantic relationships between entities from text. Standard supervised RE techniques entail training modules to tag tokens comprising entity spans and then predict the relationship between them. Recent work has instead treated the problem as a sequence-to-sequence task, linearizing relations between entities as target strings to be generated conditioned on the input. Here we push the limits of this approach, using larger language models (GPT-3 and Flan-T5 large) than considered in prior work and evaluating their performance on standard RE tasks under varying levels of supervision. We address issues inherent to evaluating generative approaches to RE by doing human evaluations, in lieu of relying on exact matching. Under this refined evaluation, we find that: (1) Few-shot prompting with GPT-3 achieves near SOTA performance, i.e., roughly equivalent to existing fully supervised models; (2) Flan-T5 is not as capable in the few-shot setting, but supervising and fine-tuning it with Chain-of-Thought (CoT) style explanations (generated via GPT-3) yields SOTA results. We release this model as a new baseline for RE tasks.

We revisit the question of why neural models tend to produce high-confidence predictions on inputs that appear nonsensical to humans. Previous work has suggested that the models fail to assign low probabilities to such inputs due to model overconfidence. We evaluate various regularization methods on fact verification benchmarks and find that this problem persists even with well-calibrated or underconfident models, suggesting that overconfidence is not the only underlying cause. We also find that regularizing the models with reduced examples helps improve interpretability but comes with the cost of miscalibration. We show that although these reduced examples are incomprehensible to humans, they can contain valid statistical patterns in the dataset utilized by the model.

In real-world systems, scaling has been critical for improving the translation quality in autoregressive translation (AT), which however has not been well studied for non-autoregressive translation (NAT). In this work, we bridge the gap by systematically studying the impact of scaling on NAT behaviors. Extensive experiments on six WMT benchmarks over two advanced NAT models show that scaling can alleviate the commonly-cited weaknesses of NAT models, resulting in better translation performance. To reduce the side-effect of scaling on decoding speed, we empirically investigate the impact of NAT encoder and decoder on the translation performance. Experimental results on the large-scale WMT20 En-De show that the asymmetric architecture (e.g. bigger encoder and smaller decoder) can achieve comparable performance with the scaling model, while maintaining the superiority of decoding speed with standard NAT models. To this end, we establish a new benchmark by validating scaled NAT models on the scaled dataset, which can be regarded as a strong baseline for future works. We release code and system outputs at https://github.com/DeepLearnXMU/Scaling4NAT.

Event co-occurrences have been proved effective for event extraction (EE) in previous studies, but have not been considered for event argument extraction (EAE) recently. In this paper, we try to fill this gap between EE research and EAE research, by highlighting the question that “Can EAE models learn better when being aware of event co-occurrences?”. To answer this question, we reformulate EAE as a problem of table generation and extend a SOTA prompt-based EAE model into a non-autoregressive generation framework, called TabEAE, which is able to extract the arguments of multiple events in parallel. Under this framework, we experiment with 3 different training-inference schemes on 4 datasets (ACE05, RAMS, WikiEvents and MLEE) and discover that via training the model to extract all events in parallel, it can better distinguish the semantic boundary of each event and its ability to extract single event gets substantially improved. Experimental results show that our method achieves new state-of-the-art performance on the 4 datasets. Our code is avilable at https://github.com/Stardust-hyx/TabEAE.

Most existing cross-lingual summarization (CLS) work constructs CLS corpora by simply and directly translating pre-annotated summaries from one language to another, which can contain errors from both summarization and translation processes. To address this issue, we propose ConvSumX, a cross-lingual conversation summarization benchmark, through a new annotation schema that explicitly considers source input context. ConvSumX consists of 2 sub-tasks under different real-world scenarios, with each covering 3 language directions. We conduct thorough analysis on ConvSumX and 3 widely-used manually annotated CLS corpora and empirically find that ConvSumX is more faithful towards input text. Additionally, based on the same intuition, we propose a 2-Step method, which takes both conversation and summary as input to simulate human annotation process. Experimental results show that 2-Step method surpasses strong baselines on ConvSumX under both automatic and human evaluation. Analysis shows that both source input text and summary are crucial for modeling cross-lingual summaries.

The ability of commonsense reasoning (CR) decides whether a neural machine translation (NMT) model can move beyond pattern recognition. Despite the rapid advancement of NMT and the use of pretraining to enhance NMT models, research on CR in NMT is still in its infancy, leaving much to be explored in terms of effectively training NMT models with high CR abilities and devising accurate automatic evaluation metrics. This paper presents a comprehensive study aimed at expanding the understanding of CR in NMT.For the training, we confirm the effectiveness of incorporating pretrained knowledge into NMT models and subsequently utilizing these models as robust testbeds for investigating CR in NMT. For the evaluation, we propose a novel entity-aware evaluation method that takes into account both the NMT candidate and important entities in the candidate, which is more aligned with human judgement. Based on the strong testbed and evaluation methods, we identify challenges in training NMT models with high CR abilities and suggest directions for further unlabeled data utilization and model design. We hope that our methods and findings will contribute to advancing the research of CR in NMT. Source data, code and scripts are freely available at https://github.com/YutongWang1216/CR-NMT.

Current literature demonstrates that Large Language Models (LLMs) are great few-shot learners, and prompting significantly increases their performance on a range of downstream tasks in a few-shot learning setting. An attempt to automate human-led prompting followed, with some progress achieved. In particular, subsequent work demonstrates that automation can outperform fine-tuning in certain K-shot learning scenarios. In this paper, we revisit techniques for automated prompting on six different downstream tasks and a larger range of K-shot learning settings. We find that automated prompting does not consistently outperform simple manual prompting. Our work suggests that, in addition to fine-tuning, manual prompting should be used as a baseline in this line of research.

We consider the task of few-shot intent detection, which involves training a deep learning model to classify utterances based on their underlying intents using only a small amount of labeled data. The current approach to address this problem is through continual pre-training, i.e., fine-tuning pre-trained language models (PLMs) on external resources (e.g., conversational corpora, public intent detection datasets, or natural language understanding datasets) before using them as utterance encoders for training an intent classifier. In this paper, we show that continual pre-training may not be essential, since the overfitting problem of PLMs on this task may not be as serious as expected. Specifically, we find that directly fine-tuning PLMs on only a handful of labeled examples already yields decent results compared to methods that employ continual pre-training, and the performance gap diminishes rapidly as the number of labeled data increases. To maximize the utilization of the limited available data, we propose a context augmentation method and leverage sequential self-distillation to boost performance. Comprehensive experiments on real-world benchmarks show that given only two or more labeled samples per class, direct fine-tuning outperforms many strong baselines that utilize external data sources for continual pre-training. The code can be found at https://github.com/hdzhang-code/DFTPlus.

Training an effective video-and-language model intuitively requires multiple frames as model inputs. However, it is unclear whether using multiple frames is beneficial to downstream tasks, and if yes, whether the performance gain is worth the drastically-increased computation and memory costs resulting from using more frames. In this work, we explore single-frame models for video-and-language learning. On a diverse set of video-and-language tasks (including text-to-video retrieval and video question answering), we show the surprising result that, with large-scale pre-training and a proper frame ensemble strategy at inference time, a single-frame trained model that does not consider temporal information can achieve better performance than existing methods that use multiple frames for training. This result reveals the existence of a strong “static appearance bias” in popular video-and-language datasets. Therefore, to allow for a more comprehensive evaluation of video-and-language models, we propose two new retrieval tasks based on existing fine-grained action recognition datasets that encourage temporal modeling. Our code is available at https://github.com/jayleicn/singularity.

To better support information retrieval tasks such as web search and open-domain question answering, growing effort is made to develop retrieval-oriented language models, e.g., RetroMAE and many others. Most of the existing works focus on improving the semantic representation capability for the contextualized embedding of the [CLS] token. However, recent study shows that the ordinary tokens besides [CLS] may provide extra information, which help to produce a better representation effect. As such, it’s necessary to extend the current methods where all contextualized embeddings can be jointly pre-trained for the retrieval tasks. In this work, we propose a novel pre-training method called Duplex Masked Auto-Encoder, a.k.a. DupMAE. It is designed to improve the quality of semantic representation where all contextualized embeddings of the pre-trained model can be leveraged. It takes advantage of two complementary auto-encoding tasks: one reconstructs the input sentence on top of the [CLS] embedding; the other one predicts the bag-of-words feature of the input sentence based on the ordinary tokens’ embeddings. The two tasks are jointly conducted to train a unified encoder, where the whole contextualized embeddings are aggregated in a compact way to produce the final semantic representation. DupMAE is simple but empirically competitive: it substantially improves the pre-trained model’s representation capability and transferability, where superior retrieval performances can be achieved on popular benchmarks, like MS MARCO and BEIR. We make our code publicly available at https://github.com/staoxiao/RetroMAE.

We study the problem of cross-lingual transfer learning for event detection (ED) where models trained on a source language are expected to perform well on data for a new target language. Among a few recent works for this problem, the main approaches involve representation matching (e.g., adversarial training) that aims to eliminate language-specific features from the representations to achieve the language-invariant representations. However, due to the mix of language-specific features with event-discriminative context, representation matching methods might also remove important features for event prediction, thus hindering the performance for ED. To address this issue, we introduce a novel approach for cross-lingual ED where representations are augmented with additional context (i.e., not eliminating) to bridge the gap between languages while enriching the contextual information to facilitate ED. At the core of our method involves a retrieval model that retrieves relevant sentences in the target language for an input sentence to compute augmentation representations. Experiments on three languages demonstrate the state-of-the-art performance of our model for cross-lingual ED.

Recent years have witnessed impressive results of pre-trained vision-language models on knowledge-intensive tasks such as visual question answering (VQA). Despite the recent advances in VQA, existing methods mainly adopt a discriminative formulation that predicts answers within a pre-defined label set, leading to easy overfitting on low-resource domains (e.g., medicine) and poor generalization under domain shift to another dataset. To tackle this limitation, we propose a novel generative model enhanced by multimodal prompt retrieval (MPR) that integrates retrieved prompts and multimodal features to generate answers in free text. Our generative model enables rapid zero-shot dataset adaptation to unseen data distributions and open-set answer labels across datasets. Our experiments on medical VQA tasks show that MPR outperforms its non-retrieval counterpart by up to 30% accuracy points in a few-shot domain adaptation setting.