ACL 2022 论文列表

Recent advances in NLP often stem from large transformer-based pre-trained models, which rapidly grow in size and use more and more training data. Such models are often released to the public so that end users can fine-tune them on a task dataset. While it is common to treat pre-training data as public, it may still contain personally identifiable information (PII), such as names, phone numbers, and copyrighted material. Recent findings show that the capacity of these models allows them to memorize parts of the training data, and suggest differentially private (DP) training as a potential mitigation. While there is recent work on DP fine-tuning of NLP models, the effects of DP pre-training are less well understood: it is not clear how downstream performance is affected by DP pre-training, and whether DP pre-training mitigates some of the memorization concerns. We focus on T5 and show that by using recent advances in JAX and XLA we can train models with DP that do not suffer a large drop in pre-training utility, nor in training speed, and can still be fine-tuned to high accuracies on downstream tasks (e.g. GLUE). Moreover, we show that T5’s span corruption is a good defense against data memorization.

Pre-trained language models (e.g. BART) have shown impressive results when fine-tuned on large summarization datasets. However, little is understood about this fine-tuning process, including what knowledge is retained from pre-training time or how content selection and generation strategies are learnt across iterations. In this work, we analyze the training dynamics for generation models, focusing on summarization. Across different datasets (CNN/DM, XSum, MediaSum) and summary properties, such as abstractiveness and hallucination, we study what the model learns at different stages of its fine-tuning process. We find that a propensity to copy the input is learned early in the training process consistently across all datasets studied. On the other hand, factual errors, such as hallucination of unsupported facts, are learnt in the later stages, though this behavior is more varied across domains. Based on these observations, we explore complementary approaches for modifying training: first, disregarding high-loss tokens that are challenging to learn and second, disregarding low-loss tokens that are learnt very quickly in the latter stages of the training process. We show that these simple training modifications allow us to configure our model to achieve different goals, such as improving factuality or improving abstractiveness.

Retrieval-based methods have been shown to be effective in NLP tasks via introducing external knowledge. However, the indexing and retrieving of large-scale corpora bring considerable computational cost. Surprisingly, we found that REtrieving from the traINing datA (REINA) only can lead to significant gains on multiple NLG and NLU tasks. We retrieve the labeled training instances most similar to the input text and then concatenate them with the input to feed into the model to generate the output. Experimental results show that this simple method can achieve significantly better performance on a variety of NLU and NLG tasks, including summarization, machine translation, language modeling, and question answering tasks. For instance, our proposed method achieved state-of-the-art results on XSum, BigPatent, and CommonsenseQA. Our code is released, https://github.com/microsoft/REINA .

Machine reading comprehension is a heavily-studied research and test field for evaluating new pre-trained language models (PrLMs) and fine-tuning strategies, and recent studies have enriched the pre-trained language models with syntactic, semantic and other linguistic information to improve the performance of the models. In this paper, we imitate the human reading process in connecting the anaphoric expressions and explicitly leverage the coreference information of the entities to enhance the word embeddings from the pre-trained language model, in order to highlight the coreference mentions of the entities that must be identified for coreference-intensive question answering in QUOREF, a relatively new dataset that is specifically designed to evaluate the coreference-related performance of a model. We use two strategies to fine-tune a pre-trained language model, namely, placing an additional encoder layer after a pre-trained language model to focus on the coreference mentions or constructing a relational graph convolutional network to model the coreference relations. We demonstrate that the explicit incorporation of coreference information in the fine-tuning stage performs better than the incorporation of the coreference information in pre-training a language model.

Toxic language detection systems often falsely flag text that contains minority group mentions as toxic, as those groups are often the targets of online hate. Such over-reliance on spurious correlations also causes systems to struggle with detecting implicitly toxic language. To help mitigate these issues, we create ToxiGen, a new large-scale and machine-generated dataset of 274k toxic and benign statements about 13 minority groups. We develop a demonstration-based prompting framework and an adversarial classifier-in-the-loop decoding method to generate subtly toxic and benign text with a massive pretrained language model. Controlling machine generation in this way allows ToxiGen to cover implicitly toxic text at a larger scale, and about more demographic groups, than previous resources of human-written text. We conduct a human evaluation on a challenging subset of ToxiGen and find that annotators struggle to distinguish machine-generated text from human-written language. We also find that 94.5% of toxic examples are labeled as hate speech by human annotators. Using three publicly-available datasets, we show that finetuning a toxicity classifier on our data improves its performance on human-written data substantially. We also demonstrate that ToxiGen can be used to fight machine-generated toxicity as finetuning improves the classifier significantly on our evaluation subset.

The aspect-based sentiment analysis (ABSA) is a fine-grained task that aims to determine the sentiment polarity towards targeted aspect terms occurring in the sentence. The development of the ABSA task is very much hindered by the lack of annotated data. To tackle this, the prior works have studied the possibility of utilizing the sentiment analysis (SA) datasets to assist in training the ABSA model, primarily via pretraining or multi-task learning. In this article, we follow this line, and for the first time, we manage to apply the Pseudo-Label (PL) method to merge the two homogeneous tasks. While it seems straightforward to use generated pseudo labels to handle this case of label granularity unification for two highly related tasks, we identify its major challenge in this paper and propose a novel framework, dubbed as Dual-granularity Pseudo Labeling (DPL). Further, similar to PL, we regard the DPL as a general framework capable of combining other prior methods in the literature. Through extensive experiments, DPL has achieved state-of-the-art performance on standard benchmarks surpassing the prior work significantly.

Existing studies on semantic parsing focus on mapping a natural-language utterance to a logical form (LF) in one turn. However, because natural language may contain ambiguity and variability, this is a difficult challenge. In this work, we investigate an interactive semantic parsing framework that explains the predicted LF step by step in natural language and enables the user to make corrections through natural-language feedback for individual steps. We focus on question answering over knowledge bases (KBQA) as an instantiation of our framework, aiming to increase the transparency of the parsing process and help the user trust the final answer. We construct INSPIRED, a crowdsourced dialogue dataset derived from the ComplexWebQuestions dataset. Our experiments show that this framework has the potential to greatly improve overall parse accuracy. Furthermore, we develop a pipeline for dialogue simulation to evaluate our framework w.r.t. a variety of state-of-the-art KBQA models without further crowdsourcing effort. The results demonstrate that our framework promises to be effective across such models.

The robustness of Text-to-SQL parsers against adversarial perturbations plays a crucial role in delivering highly reliable applications. Previous studies along this line primarily focused on perturbations in the natural language question side, neglecting the variability of tables. Motivated by this, we propose the Adversarial Table Perturbation (ATP) as a new attacking paradigm to measure robustness of Text-to-SQL models. Following this proposition, we curate ADVETA, the first robustness evaluation benchmark featuring natural and realistic ATPs. All tested state-of-the-art models experience dramatic performance drops on ADVETA, revealing significant room of improvement. To defense against ATP, we build a systematic adversarial training example generation framework tailored for better contextualization of tabular data. Experiments show that our approach brings models best robustness improvement against ATP, while also substantially boost model robustness against NL-side perturbations. We will release ADVETA and code to facilitate future research.

When building NLP models, there is a tendency to aim for broader coverage, often overlooking cultural and (socio)linguistic nuance. In this position paper, we make the case for care and attention to such nuances, particularly in dataset annotation, as well as the inclusion of cultural and linguistic expertise in the process. We present a playbook for responsible dataset creation for polyglossic, multidialectal languages. This work is informed by a study on Arabic annotation of social media content.

This paper demonstrates that multilingual pretraining and multilingual fine-tuning are both critical for facilitating cross-lingual transfer in zero-shot translation, where the neural machine translation (NMT) model is tested on source languages unseen during supervised training. Following this idea, we present SixT+, a strong many-to-English NMT model that supports 100 source languages but is trained with a parallel dataset in only six source languages. SixT+ initializes the decoder embedding and the full encoder with XLM-R large and then trains the encoder and decoder layers with a simple two-stage training strategy. SixT+ achieves impressive performance on many-to-English translation. It significantly outperforms CRISS and m2m-100, two strong multilingual NMT systems, with an average gain of 7.2 and 5.0 BLEU respectively. Additionally, SixT+ offers a set of model parameters that can be further fine-tuned to other unsupervised tasks. We demonstrate that adding SixT+ initialization outperforms state-of-the-art explicitly designed unsupervised NMT models on Si<->En and Ne<->En by over 1.2 average BLEU. When applied to zero-shot cross-lingual abstractive summarization, it produces an average performance gain of 12.3 ROUGE-L over mBART-ft. We conduct detailed analyses to understand the key ingredients of SixT+, including multilinguality of the auxiliary parallel data, positional disentangled encoder, and the cross-lingual transferability of its encoder.

Understanding the functional (dis)-similarity of source code is significant for code modeling tasks such as software vulnerability and code clone detection. We present DISCO (DIS-similarity of COde), a novel self-supervised model focusing on identifying (dis)similar functionalities of source code. Different from existing works, our approach does not require a huge amount of randomly collected datasets. Rather, we design structure-guided code transformation algorithms to generate synthetic code clones and inject real-world security bugs, augmenting the collected datasets in a targeted way. We propose to pre-train the Transformer model with such automatically generated program contrasts to better identify similar code in the wild and differentiate vulnerable programs from benign ones. To better capture the structural features of source code, we propose a new cloze objective to encode the local tree-based context (e.g., parents or sibling nodes). We pre-train our model with a much smaller dataset, the size of which is only 5% of the state-of-the-art models’ training datasets, to illustrate the effectiveness of our data augmentation and the pre-training approach. The evaluation shows that, even with much less data, DISCO can still outperform the state-of-the-art models in vulnerability and code clone detection tasks.

Users interacting with voice assistants today need to phrase their requests in a very specific manner to elicit an appropriate response. This limits the user experience, and is partly due to the lack of reasoning capabilities of dialogue platforms and the hand-crafted rules that require extensive labor. One possible solution to improve user experience and relieve the manual efforts of designers is to build an end-to-end dialogue system that can do reasoning itself while perceiving user’s utterances. In this work, we propose a novel method to incorporate the knowledge reasoning capability into dialog systems in a more scalable and generalizable manner. Our proposed method allows a single transformer model to directly walk on a large-scale knowledge graph to generate responses. To the best of our knowledge, this is the first work to have transformer models generate responses by reasoning over differentiable knowledge graphs. We investigate the reasoning abilities of the proposed method on both task-oriented and domain-specific chit-chat dialogues. Empirical results show that this method can effectively and efficiently incorporate a knowledge graph into a dialogue system with fully-interpretable reasoning paths.

Entity recognition is a fundamental task in understanding document images. Traditional sequence labeling frameworks treat the entity types as class IDs and rely on extensive data and high-quality annotations to learn semantics which are typically expensive in practice. In this paper, we aim to build an entity recognition model requiring only a few shots of annotated document images. To overcome the data limitation, we propose to leverage the label surface names to better inform the model of the target entity type semantics and also embed the labels into the spatial embedding space to capture the spatial correspondence between regions and labels. Specifically, we go beyond sequence labeling and develop a novel label-aware seq2seq framework, LASER. The proposed model follows a new labeling scheme that generates the label surface names word-by-word explicitly after generating the entities. During training, LASER refines the label semantics by updating the label surface name representations and also strengthens the label-region correlation. In this way, LASER recognizes the entities from document images through both semantic and layout correspondence. Extensive experiments on two benchmark datasets demonstrate the superiority of LASER under the few-shot setting.

Dialogue State Tracking (DST) is primarily evaluated using Joint Goal Accuracy (JGA) defined as the fraction of turns where the ground-truth dialogue state exactly matches the prediction. Generally in DST, the dialogue state or belief state for a given turn contain all the intents shown by the user till that turn. Due to this cumulative nature of the belief state, it is difficult to get a correct prediction once a misprediction has occurred. Thus, although being a useful metric, it can be harsh at times and underestimate the true potential of a DST model. Moreover, an improvement in JGA can sometimes decrease the performance of turn-level or non-cumulative belief state prediction due to inconsistency in annotations. So, using JGA as the only metric for model selection may not be ideal for all scenarios. In this work, we discuss various evaluation metrics used for DST along with their shortcomings. To address the existing issues, we propose a new evaluation metric named Flexible Goal Accuracy (FGA). FGA is a generalized version of JGA. But unlike JGA, it tries to give penalized rewards to mispredictions that are locally correct i.e. the root cause of the error is an earlier turn. By doing so, FGA considers the performance of both cumulative and turn-level prediction flexibly and provides a better insight than the existing metrics. We also show that FGA is a better discriminator of DST model performance.

We consider the task of document-level entity linking (EL), where it is important to make consistent decisions for entity mentions over the full document jointly. We aim to leverage explicit “connections” among mentions within the document itself: we propose to join EL and coreference resolution (coref) in a single structured prediction task over directed trees and use a globally normalized model to solve it. This contrasts with related works where two separate models are trained for each of the tasks and additional logic is required to merge the outputs. Experimental results on two datasets show a boost of up to +5% F1-score on both coref and EL tasks, compared to their standalone counterparts. For a subset of hard cases, with individual mentions lacking the correct EL in their candidate entity list, we obtain a +50% increase in accuracy.

Predicting the approval chance of a patent application is a challenging problem involving multiple facets. The most crucial facet is arguably the novelty — 35 U.S. Code § 102 rejects more recent applications that have very similar prior arts. Such novelty evaluations differ the patent approval prediction from conventional document classification — Successful patent applications may share similar writing patterns; however, too-similar newer applications would receive the opposite label, thus confusing standard document classifiers (e.g., BERT). To address this issue, we propose a novel framework that unifies the document classifier with handcrafted features, particularly time-dependent novelty scores. Specifically, we formulate the novelty scores by comparing each application with millions of prior arts using a hybrid of efficient filters and a neural bi-encoder. Moreover, we impose a new regularization term into the classification objective to enforce the monotonic change of approval prediction w.r.t. novelty scores. From extensive experiments on a large-scale USPTO dataset, we find that standard BERT fine-tuning can partially learn the correct relationship between novelty and approvals from inconsistent data. However, our time-dependent novelty features offer a boost on top of it. Also, our monotonic regularization, while shrinking the search space, can drive the optimizer to better local optima, yielding a further small performance gain.

Recent work in task-independent graph semantic parsing has shifted from grammar-based symbolic approaches to neural models, showing strong performance on different types of meaning representations. However, it is still unclear that what are the limitations of these neural parsers, and whether these limitations can be compensated by incorporating symbolic knowledge into model inference. In this paper, we address these questions by taking English Resource Grammar (ERG) parsing as a case study. Specifically, we first develop a state-of-the-art, T5-based neural ERG parser, and conduct detail analyses of parser performance within fine-grained linguistic categories. The neural parser attains superior performance on in-distribution test set, but degrades significantly on long-tail situations, while the symbolic parser performs more robustly. To address this, we further propose a simple yet principled collaborative framework for neural-symbolic semantic parsing, by designing a decision criterion for beam search that incorporates the prior knowledge from a symbolic parser and accounts for model uncertainty. Experimental results show that the proposed framework yields comprehensive improvement over neural baseline across long-tail categories, yielding the best known Smatch score (97.01) on the well-studied DeepBank benchmark.

To effectively characterize the nature of paraphrase pairs without expert human annotation, we proposes two new metrics: word position deviation (WPD) and lexical deviation (LD). WPD measures the degree of structural alteration, while LD measures the difference in vocabulary used. We apply these metrics to better understand the commonly-used MRPC dataset and study how it differs from PAWS, another paraphrase identification dataset. We also perform a detailed study on MRPC and propose improvements to the dataset, showing that it improves generalizability of models trained on the dataset. Lastly, we apply our metrics to filter the output of a paraphrase generation model and show how it can be used to generate specific forms of paraphrases for data augmentation or robustness testing of NLP models.

Aligning with ACL 2022 special Theme on “Language Diversity: from Low Resource to Endangered Languages”, we discuss the major linguistic and sociopolitical challenges facing development of NLP technologies for African languages. Situating African languages in a typological framework, we discuss how the particulars of these languages can be harnessed. To facilitate future research, we also highlight current efforts, communities, venues, datasets, and tools. Our main objective is to motivate and advocate for an Afrocentric approach to technology development. With this in mind, we recommend what technologies to build and how to build, evaluate, and deploy them based on the needs of local African communities.