ACL 2023 论文列表

Text detoxification has the potential to mitigate the harms of toxicity by rephrasing text to remove offensive meaning, but subtle toxicity remains challenging to tackle. We introduce MaRCo, a detoxification algorithm that combines controllable generation and text rewriting methods using a Product of Experts with autoencoder language models (LMs). MaRCo uses likelihoods under a non-toxic LM (expert) and a toxic LM (anti-expert) to find candidate words to mask and potentially replace. We evaluate our method on several subtle toxicity and microaggressions datasets, and show that it not only outperforms baselines on automatic metrics, but MaRCo’s rewrites are preferred 2.1 times more in human evaluation. Its applicability to instances of subtle toxicity is especially promising, demonstrating a path forward for addressing increasingly elusive online hate.

In text summarization, extractivity is defined as a measurement of the degree of overlap between a source document and its summary. Previous research has shown that the extractivity level of training data can influence both output extractivity and the amount of factual information (i.e. faithfulness) in outputs for abstractive summarization. However, it remains unclear if and how extractivity impacts the performance of abstractive models. In this work, we investigate the relationship between dataset extractivity and model performance by comparing the performance of trained models under different degrees of extractivity. We find that while low levels of extractivity can improve performance, as extractivity increases, performance is negatively impacted. Furthermore, through an analysis of the model’s copy continuity of content, we discover that higher extractivity leads to a greater tendency for the model to copy text continuously from the source document rather than identifying and summarizing important content that should be covered in the target summary. To address these issues, we propose a simple and effective method to design copy labels for fixing the model’s copying behaviors and train the model with a copy mechanism. The experimental results illustrate the effectiveness of our strategy in alleviating the negative impact on model performance resulting from high dataset extractivity, and that our method outperforms several competitive baselines.

While the problem of hallucinations in neural machine translation has long been recognized, so far the progress on its alleviation is very little. Indeed, recently it turned out that without artificially encouraging models to hallucinate, previously existing methods fall short and even the standard sequence log-probability is more informative. It means that internal characteristics of the model can give much more information than we expect, and before using external models and measures, we first need to ask: how far can we go if we use nothing but the translation model itself ? We propose to use a method that evaluates the percentage of the source contribution to a generated translation. Intuitively, hallucinations are translations “detached” from the source, hence they can be identified by low source contribution. This method improves detection accuracy for the most severe hallucinations by a factor of 2 and is able to alleviate hallucinations at test time on par with the previous best approach that relies on external models. Next, if we move away from internal model characteristics and allow external tools, we show that using sentence similarity from cross-lingual embeddings further improves these results. We release the code of our experiments.

Recent model editing techniques promise to mitigate the problem of memorizing false or outdated associations during LLM training. However, we show that these techniques can introduce large unwanted side effects which are not detected by existing specificity benchmarks. We extend the existing CounterFact benchmark to include a dynamic component and dub our benchmark CounterFact+. Additionally, we extend the metrics used for measuring specificity by a principled KL divergence-based metric. We use this improved benchmark to evaluate recent model editing techniques and find that they suffer from low specificity. Our findings highlight the need for improved specificity benchmarks that identify and prevent unwanted side effects.

The COVID-19 pandemic created a deluge of questionable and contradictory scientific claims about drug efficacy – an “infodemic” with lasting consequences for science and society. In this work, we argue that NLP models can help domain experts distill and understand the literature in this complex, high-stakes area. Our task is to automatically identify contradictory claims about COVID-19 drug efficacy. We frame this as a natural language inference problem and offer a new NLI dataset created by domain experts. The NLI framing allows us to create curricula combining existing datasets and our own. The resulting models are useful investigative tools. We provide a case study of how these models help a domain expert summarize and assess evidence concerning remdisivir and hydroxychloroquine.

Deep neural networks (DNNs) have been proven to be sensitive towards perturbations on input samples, and previous works highlight that adversarial samples are even more vulnerable than normal ones. In this work, this phenomenon is illustrated frWe first show that adversarial samples locate in steep and narrow local minima of the loss landscape (high sharpness) while normal samples, which differs distinctly from adversarial ones, reside in the loss surface that is more flatter (low sharpness).om the perspective of sharpness via visualizing the input loss landscape of models. Based on this, we propose a simple and effective sharpness-based detector to distinct adversarial samples by maximizing the loss increment within the region where the inference sample is located. Considering that the notion of sharpness of a loss landscape is relative, we further propose an adaptive optimization strategy in an attempt to fairly compare the relative sharpness among different samples. Experimental results show that our approach can outperform previous detection methods by large margins (average +6.6 F1 score) for four advanced attack strategies considered in this paper across three text classification tasks.

Masked language models (MLM) do not explicitly define a distribution over language, i.e., they are not language models per se. However, recent work has implicitly treated them as such for the purposes of generation and scoring. This paper studies methods for deriving explicit joint distributions from MLMs, focusing on distributions over two tokens, which makes it possible to calculate exact distributional properties. We find that an approach based on identifying joints whose conditionals are closest to those of the MLM works well and outperforms existing Markov random field-based approaches. We further find that this derived model’s conditionals can even occasionally outperform the original MLM’s conditionals.

Using psycholinguistic and computational experiments we compare the ability of humans and several pre-trained masked language models to correctly identify control dependencies in Spanish sentences such as ‘José le prometió/ordenó a María ser ordenado/a’ (‘Joseph promised/ordered Mary to be tidy’). These structures underlie complex anaphoric and agreement relations at the interface of syntax and semantics, allowing us to study lexically-guided antecedent retrieval processes. Our results show that while humans correctly identify the (un)acceptability of the strings, language models often fail to identify the correct antecedent in non-adjacent dependencies, showing their reliance on linearity. Additional experiments on Galician reinforce these conclusions. Our findings are equally valuable for the evaluation of language models’ ability to capture linguistic generalizations, as well as for psycholinguistic theories of anaphor resolution.

ATOMIC is a large-scale commonsense knowledge graph (CSKG) containing everyday if-then knowledge triplets, i.e., head event, relation, tail event. The one-hop annotation manner made ATOMIC a set of independent bipartite graphs, which ignored the numerous links between events in different bipartite graphs and consequently caused shortages in knowledge coverage and multi-hop paths. In this work, we aim to construct Dense-ATOMIC with high knowledge coverage and massive multi-hop paths. The events in ATOMIC are normalized to a consistent pattern at first. We then propose a CSKG completion method called Rel-CSKGC to predict the relation given the head event and the tail event of a triplet, and train a CSKG completion model based on existing triplets in ATOMIC. We finally utilize the model to complete the missing links in ATOMIC and accordingly construct Dense-ATOMIC. Both automatic and human evaluation on an annotated subgraph of ATOMIC demonstrate the advantage of Rel-CSKGC over strong baselines. We further conduct extensive evaluations on Dense-ATOMIC in terms of statistics, human evaluation, and simple downstream tasks, all proving Dense-ATOMIC’s advantages in Knowledge Coverage and Multi-hop Paths. Both the source code of Rel-CSKGC and Dense-ATOMIC are publicly available on https://github.com/NUSTM/Dense-ATOMIC.

Recently, the task of distantly supervised (DS) ultra-fine entity typing has received significant attention. However, DS data is noisy and often suffers from missing or wrong labeling issues resulting in low precision and low recall. This paper proposes a novel ultra-fine entity typing model with denoising capability. Specifically, we build a noise model to estimate the unknown labeling noise distribution over input contexts and noisy type labels. With the noise model, more trustworthy labels can be recovered by subtracting the estimated noise from the input. Furthermore, we propose an entity typing model, which adopts a bi-encoder architecture, is trained on the denoised data. Finally, the noise model and entity typing model are trained iteratively to enhance each other. We conduct extensive experiments on the Ultra-Fine entity typing dataset as well as OntoNotes dataset and demonstrate that our approach significantly outperforms other baseline methods.

Video multimodal fusion aims to integrate multimodal signals in videos, such as visual, audio and text, to make a complementary prediction with multiple modalities contents. However, unlike other image-text multimodal tasks, video has longer multimodal sequences with more redundancy and noise in both visual and audio modalities. Prior denoising methods like forget gate are coarse in the granularity of noise filtering. They often suppress the redundant and noisy information at the risk of losing critical information. Therefore, we propose a denoising bottleneck fusion (DBF) model for fine-grained video multimodal fusion. On the one hand, we employ a bottleneck mechanism to filter out noise and redundancy with a restrained receptive field. On the other hand, we use a mutual information maximization module to regulate the filter-out module to preserve key information within different modalities. Our DBF model achieves significant improvement over current state-of-the-art baselines on multiple benchmarks covering multimodal sentiment analysis and multimodal summarization tasks. It proves that our model can effectively capture salient features from noisy and redundant video, audio, and text inputs. The code for this paper will be publicly available at https://github.com/WSXRHFG/DBF

Contrastive Language-Image Pre-training (CLIP) formulates image classification as an image-to-text matching task, i.e., matching images to the corresponding natural language descriptions instead of discrete category IDs. This allows for open-vocabulary visual recognition, where the model can recognize images from an open class set (also known as an open vocabulary) in a zero-shot manner. However, evaluating the openness of CLIP-like models is challenging, as the models are open to arbitrary vocabulary in theory, but their accuracy varies in practice. To address this, we resort to an incremental perspective to assess the openness through vocabulary expansions, and define extensibility to measure a model’s ability to handle novel classes. Our evaluation shows that CLIP-like models are not truly open, and their performance deteriorates as the vocabulary expands. We further dissect the feature space of CLIP from the perspectives of representation alignment and uniformity. Our investigation reveals that the overestimation of openness is due to confusion among competing text features, rather than a failure to capture the similarity between image features and text features of novel classes. We hope that our investigation and analysis will facilitate future research on the CLIP openness issue.

Recent work studies the cognitive capabilities of language models through psychological tests designed for humans. While these studies are helpful for understanding the general capabilities of these models, there is no guarantee that a model possessing sufficient capabilities to pass those tests would actually use those capabilities in performing real-life tasks. In this work, we formulate task-oriented cognitive capabilities, which are human-like cognitive capabilities that language models leverage to perform tasks. These capabilities are (i) the ability to quickly generate good candidate utterances (the search capability) (ii) the ability to predict how a listener interprets those utterances and choose the most appropriate one (the pragmatic capability). We design an evaluation scheme for comparing these capabilities of a language model with those of a human. Applying this scheme to examine various models in a navigation instruction generation problem, we find that their pragmatic capability is severely lacking. This insight leads us to augment them with better models of the listener and obtain a significant boost of 11% in success rate in guiding real humans. Our work advocates for having a principled procedure for aligning language models with humans that involves (i) formulating task-oriented capabilities, (ii) devising a method to quantify their deficiency, and (iii) iteratively improving them.

Textual backdoor attack, as a novel attack model, has been shown to be effective in adding a backdoor to the model during training. Defending against such backdoor attacks has become urgent and important. In this paper, we propose AttDef, an efficient attribution-based pipeline to defend against two insertion-based poisoning attacks, BadNL and InSent. Specifically, we regard the tokens with larger attribution scores as potential triggers since larger attribution words contribute more to the false prediction results and therefore are more likely to be poison triggers. Additionally, we further utilize an external pre-trained language model to distinguish whether input is poisoned or not. We show that our proposed method can generalize sufficiently well in two common attack scenarios (poisoning training data and testing data), which consistently improves previous methods. For instance, AttDef can successfully mitigate both attacks with an average accuracy of 79.97% (56.59% up) and 48.34% (3.99% up) under pre-training and post-training attack defense respectively, achieving the new state-of-the-art performance on prediction recovery over four benchmark datasets.

Recent advancements in multimodal foundation models (e.g., CLIP) have excelled in zero-shot generalization. Prompt tuning involved in the knowledge transfer from foundation models to downstream tasks has gained significant attention recently. Existing prompt-tuning methods in cross-modal learning, however, either solely focus on language branch, or learn vision-language interaction in a shallow mechanism. In this context, we propose a Deeply coupled Cross-modal Prompt learning (DCP) method based on CLIP. DCP flexibly accommodates the interplay between vision and language with a Cross-Modal Prompt Attention (CMPA) mechanism, which enables the mutual exchange of respective representation through a well-connected multi-head attention progressively and strongly. We then conduct comprehensive few-shot learning experiments on 11 image classification datasets and analyze the robustness to domain shift as well. Thorough experimental analysis evidently demonstrates the superb few-shot generalization and compelling domain adaption capacity of a well-executed DCP.

Span-based models are one of the most straightforward methods for named entity recognition (NER). Existing span-based NER systems shallowly aggregate the token representations to span representations. However, this typically results in significant ineffectiveness for long entities, a coupling between the representations of overlapping spans, and ultimately a performance degradation. In this study, we propose DSpERT (Deep Span Encoder Representations from Transformers), which comprises a standard Transformer and a span Transformer. The latter uses low-layered span representations as queries, and aggregates the token representations as keys and values, layer by layer from bottom to top. Thus, DSpERT produces span representations of deep semantics. With weight initialization from pretrained language models, DSpERT achieves performance higher than or competitive with recent state-of-the-art systems on six NER benchmarks. Experimental results verify the importance of the depth for span representations, and show that DSpERT performs particularly well on long-span entities and nested structures. Further, the deep span representations are well structured and easily separable in the feature space.

Natural language understanding (NLU) tasks face a non-trivial amount of ambiguous samples where veracity of their labels is debatable among annotators. NLU models should thus account for such ambiguity, but they approximate the human opinion distributions quite poorly and tend to produce over-confident predictions. To address this problem, we must consider how to exactly capture the degree of relationship between each sample and its candidate classes. In this work, we propose a novel method with deep model compression and show how such relationship can be accounted for. We see that more reasonably represented relationships can be discovered in the lower layers and that validation accuracies are converging at these layers, which naturally leads to layer pruning. We also see that distilling the relationship knowledge from a lower layer helps models produce better distribution. Experimental results demonstrate that our method makes substantial improvement on quantifying ambiguity without gold distribution labels. As positive side-effects, our method is found to reduce the model size significantly and improve latency, both attractive aspects of NLU products.

In this work, we argue that non-autoregressive (NAR) sequence generative models can equivalently be regarded as an iterative refinement process towards the target sequence, implying an underlying dynamical system of NAR model: z = f (z, x) → y. In such a way, the optimal prediction of a NAR model should be the equilibrium state of its dynamics if given infinitely many iterations. However, this is infeasible in practice due to limited computational and memory budgets. To this end, we propose DEQNAR to directly solve for the equilibrium state of NAR models based on deep equilibrium networks (Bai et al., 2019) with black-box root-finding solvers and back-propagate through the equilibrium point via implicit differentiation with constant memory. We conduct extensive experiments on four WMT machine translation benchmarks. Our main findings show that DEQNAR can indeed converge to a more accurate prediction and is a general-purpose framework that consistently helps yield substantial improvement for several strong NAR backbones.

Building a system for morphological processing is a challenging task in morphologically complex languages like Arabic. Although there are some deep learning based models that achieve successful results, these models rely on a large amount of annotated data. Building such datasets, specially for some of the lower-resource Arabic dialects, is very difficult, time-consuming, and expensive. In addition, some parts of the annotated data do not contain useful information for training machine learning models. Active learning strategies allow the learner algorithm to select the most informative samples for annotation. There has been little research that focuses on applying active learning for morphological inflection and morphophonological processing. In this paper, we have proposed a deep active learning method for this task. Our experiments on Egyptian Arabic show that with only about 30% of annotated data, we achieve the same results as does the state-of-the-art model on the whole dataset.