ACL 2023 Papers

Dialogue models are often enriched with extensive external knowledge to provide informative responses through a retrieval-augmented pipeline. Nevertheless, retrieval-augmented approaches rely on finely annotated retrieval training data and knowledge-grounded response generation data, making it costly to transfer. To tackle this challenge, this paper proposed a retrieval-free approach, KiDG, by automatically turning knowledge documents into simulated multi-turn dialogues through a Multi-Document Traversal algorithm. The simulated knowledge-intensive dialogues constructed by KiDG in one domain can be easily used to train and enhance pre-trained dialogue models’ knowledge w.r.t. this domain without costly annotation. We conduct extensive experiments comparing retrieval-augmented models and a variety of retrieval-free models. We found that dialogue models enhanced with data simulated with KiDG largely outperform state-of-the-art retrieval-free methods, and it achieves comparable performance compared to retrieval-augmented methods while being better, and cheaper at domain transfer.

Instructional videos make learning knowledge more efficient, by providing a detailed multimodal context of each procedure in instruction.A unique challenge posed by instructional videos is key-object degeneracy, where any single modality fails to sufficiently capture the key objects referred to in the procedure. For machine systems, such degeneracy can disturb the performance of a downstream task such as dense video captioning, leading to the generation of incorrect captions omitting key objects. To repair degeneracy, we propose a retrieval-based framework to augment the model representations in the presence of such key-object degeneracy. We validate the effectiveness and generalizability of our proposed framework over baselines using modalities with key-object degeneracy.

Data preparation, also called data wrangling, is considered one of the most expensive and time-consuming steps when performing analytics or building machine learning models. Preparing data typically involves collecting and merging data from complex heterogeneous, and often large-scale data sources, such as data lakes. In this paper, we introduce a novel approach toward automatic data wrangling in an attempt to alleviate the effort of end-users, e.g. data analysts, in structuring dynamic views from data lakes in the form of tabular data. Given a corpus of tables, we propose a retrieval augmented transformer model that is self-trained for the table augmentation tasks of row/column population and data imputation. Our self-learning strategy consists in randomly ablating tables from the corpus and training the retrieval-based model with the objective of reconstructing the partial tables given as input with the original values or headers. We adopt this strategy to first train the dense neural retrieval model encoding portions of tables to vectors, and then the end-to-end model trained to perform table augmentation tasks. We test on EntiTables, the standard benchmark for table augmentation, as well as introduce a new benchmark to advance further research: WebTables. Our model consistently and substantially outperforms both supervised statistical methods and the current state-of-the-art transformer-based models.

Word-level Quality Estimation (QE) of Machine Translation (MT) aims to detect potential translation errors in the translated sentence without reference. Typically, conventional works on word-level QE are usually designed to predict the quality of translated words in terms of the post-editing effort, where the word labels in the dataset, i.e., OK or BAD, are automatically generated by comparing words between MT sentences and the post-edited sentences through a Translation Error Rate (TER) toolkit. While the post-editing effort can be used to measure the translation quality to some extent, we find it usually conflicts with human judgment on whether the word is well or poorly translated. To investigate this conflict, we first create a golden benchmark dataset, namely HJQE (Human Judgement on Quality Estimation), where the source and MT sentences are identical to the original TER-based dataset and the expert translators directly annotate the poorly translated words on their judgments. Based on our analysis, we further propose two tag-correcting strategies which can make the TER-based artificial QE corpus closer to HJQE. We conduct substantial experiments based on the publicly available WMT En-De and En-Zh corpora. The results not only show our proposed dataset is more consistent with human judgment but also confirm the effectiveness of the proposed tag-correcting strategies.For reviewers, the corpora and codes can be found in the attached files.

Language models have been shown to perform better with an increase in scale on a wide variety of tasks via the in-context learning paradigm. In this paper, we investigate the hypothesis that the ability of a large language model to in-context learn-perform a task is not uniformly spread across all of its underlying components. Using a 66 billion parameter language model (OPT-66B) across a diverse set of 14 downstream tasks, we find this is indeed the case: ~70% of the attention heads and ~20% of the feed forward networks can be removed with minimal decline in task performance. We find substantial overlap in the set of attention heads (un)important for in-context learning across tasks and number of in-context examples. We also address our hypothesis through a task-agnostic lens, finding that a small set of attention heads in OPT-66B score highly on their ability to perform primitive induction operations associated with in-context learning, namely, prefix matching and copying. These induction heads overlap with task-specific important heads, reinforcing arguments by Olsson et al. (2022) regarding induction head generality to more sophisticated behaviors associated with in-context learning. Overall, our study provides several insights that indicate large language models may be under-trained for in-context learning and opens up questions on how to pre-train language models to more effectively perform in-context learning.

This paper rethinks translation memory augmented neural machine translation (TM-augmented NMT) from two perspectives, i.e., a probabilistic view of retrieval and the variance-bias decomposition principle. The finding demonstrates that TM-augmented NMT is good at the ability of fitting data (i.e., lower bias) but is more sensitive to the fluctuations in the training data (i.e., higher variance), which provides an explanation to a recently reported contradictory phenomenon on the same translation task: TM-augmented NMT substantially advances NMT without TM under the high resource scenario whereas it fails under the low resource scenario. Then this paper proposes a simple yet effective TM-augmented NMT model to promote the variance and address the contradictory phenomenon. Extensive experiments show that the proposed TM-augmented NMT achieves consistent gains over both conventional NMT and existing TM-augmented NMT under two variance-preferable (low resource and plug-and-play) scenarios as well as the high resource scenario.

Semi-supervised learning (SSL) is a popular setting aiming to effectively utilize unlabelled data to improve model performance in downstream natural language processing (NLP) tasks. Currently, there are two popular approaches to make use of the unlabelled data: Self-training (ST) and Task-adaptive pre-training (TAPT). ST uses a teacher model to assign pseudo-labels to the unlabelled data, while TAPT continues pre-training on the unlabelled data before fine-tuning. To the best of our knowledge, the effectiveness of TAPT in SSL tasks has not been systematically studied, and no previous work has directly compared TAPT and ST in terms of their ability to utilize the pool of unlabelled data. In this paper, we provide an extensive empirical study comparing five state-of-the-art ST approaches and TAPT across various NLP tasks and data sizes, including in- and out-of domain settings. Surprisingly, we find that TAPT is a strong and more robust SSL learner, even when using just a few hundred unlabelled samples or in the presence of domain shifts, compared to more sophisticated ST approaches, and tends to bring greater improvements in SSL than in fully-supervised settings. Our further analysis demonstrates the risks of using ST approaches when the size of labelled or unlabelled data is small or when domain shifts exist, and highlights TAPT as a potential solution.

Automatic evaluation methods for translation often require model training, and thus the availability of parallel corpora limits their applicability to low-resource settings. Round-trip translation is a potential workaround, which can reframe bilingual evaluation into a much simpler monolingual task. Early results from the era of statistical machine translation (SMT) raised fundamental concerns about the utility of this approach, based on poor correlation with human translation quality judgments. In this paper, we revisit this technique with modern neural translation (NMT) and show that round-trip translation does allow for accurate automatic evaluation without the need for reference translations. These opposite findings can be explained through the copy mechanism in SMT that is absent in NMT. We demonstrate that round-trip translation benefits multiple machine translation evaluation tasks: i) predicting forward translation scores; ii) improving the performance of a quality estimation model; and iii) identifying adversarial competitors in shared tasks via cross-system verification.

We revisit the multimodal entity and relation extraction from a translation point of view. Special attention is paid on the misalignment issue in text-image datasets which may mislead the learning. We are motivated by the fact that the cross-modal misalignment is a similar problem of cross-lingual divergence issue in machine translation. The problem can then be transformed and existing solutions can be borrowed by treating a text and its paired image as the translation to each other. We implement a multimodal back-translation using diffusion-based generative models for pseudo-paralleled pairs and a divergence estimator by constructing a high-resource corpora as a bridge for low-resource learners. Fine-grained confidence scores are generated to indicate both types and degrees of alignments with which better representations are obtained. The method has been validated in the experiments by outperforming 14 state-of-the-art methods in both entity and relation extraction tasks. The source code is available at https://github.com/thecharm/TMR.

In this paper, we study Chinese Spelling Correction (CSC) as a joint decision made by two separate models: a language model and an error model. Through empirical analysis, we find that fine-tuning BERT tends to over-fit the error model while under-fit the language model, resulting in poor generalization to out-of-distribution error patterns. Given that BERT is the backbone of most CSC models, this phenomenon has a significant negative impact. To address this issue, we are releasing a multi-domain benchmark LEMON, with higher quality and diversity than existing benchmarks, to allow a comprehensive assessment of the open domain generalization of CSC models. Then, we demonstrate that a very simple strategy – randomly masking 20% non-error tokens from the input sequence during fine-tuning – is sufficient for learning a much better language model without sacrificing the error model. This technique can be applied to any model architecture and achieves new state-of-the-art results on SIGHAN, ECSpell, and LEMON.

Recently, numerous efforts have continued to push up performance boundaries of document-level relation extraction (DocRE) and have claimed significant progress in DocRE. In this paper, we do not aim at proposing a novel model for DocRE. Instead, we take a closer look at the field to see if these performance gains are actually true. By taking a comprehensive literature review and a thorough examination of popular DocRE datasets, we find that these performance gains are achieved upon a strong or even untenable assumption in common: all named entities are perfectly localized, normalized, and typed in advance. Next, we construct four types of entity mention attacks to examine the robustness of typical DocRE models by behavioral probing. We also have a close check on model usability in a more realistic setting. Our findings reveal that most of current DocRE models are vulnerable to entity mention attacks and difficult to be deployed in real-world end-user NLP applications. Our study calls more attentions for future research to stop simplifying problem setups, and to model DocRE in the wild rather than in an unrealistic Utopian world.

We introduce CDBert, a new learning paradigm that enhances the semantics understanding ability of the Chinese PLMs with dictionary knowledge and structure of Chinese characters. We name the two core modules of CDBert as Shuowen and Jiezi, where Shuowen refers to the process of retrieving the most appropriate meaning from Chinese dictionaries and Jiezi refers to the process of enhancing characters’ glyph representations with structure understanding. To facilitate dictionary understanding, we propose three pre-training tasks, i.e.„ Masked Entry Modeling, Contrastive Learning for Synonym and Antonym, and Example Learning. We evaluate our method on both modern Chinese understanding benchmark CLUE and ancient Chinese benchmark CCLUE. Moreover, we propose a new polysemy discrimination task PolyMRC based on the collected dictionary of ancient Chinese. Our paradigm demonstrates consistent improvements on previous Chinese PLMs across all tasks. Moreover, our approach yields significant boosting on few-shot setting of ancient Chinese understanding.

Researchers have traditionally recruited native speakers to provide annotations for the widely used benchmark datasets. But there are languages for which recruiting native speakers is difficult, and it would help to get learners of those languages to annotate the data. In this paper, we investigate whether language learners can contribute annotations to the benchmark datasets. In a carefully controlled annotation experiment, we recruit 36 language learners, provide two types of additional resources (dictionaries and machine-translated sentences), and perform mini-tests to measure their language proficiency. We target three languages, English, Korean, and Indonesian, and four NLP tasks, sentiment analysis, natural language inference, named entity recognition, and machine reading comprehension. We find that language learners, especially those with intermediate or advanced language proficiency, are able to provide fairly accurate labels with the help of additional resources. Moreover, we show that data annotation improves learners’ language proficiency in terms of vocabulary and grammar. The implication of our findings is that broadening the annotation task to include language learners can open up the opportunity to build benchmark datasets for languages for which it is difficult to recruit native speakers.

Different ways of linguistically expressing the same real-world event can lead to different perceptions of what happened. Previous work has shown that different descriptions of gender-based violence (GBV) influence the reader’s perception of who is to blame for the violence, possibly reinforcing stereotypes which see the victim as partly responsible, too. As a contribution to raise awareness on perspective-based writing, and to facilitate access to alternative perspectives, we introduce the novel task of automatically rewriting GBV descriptions as a means to alter the perceived level of blame on the perpetrator. We present a quasi-parallel dataset of sentences with low and high perceived responsibility levels for the perpetrator, and experiment with unsupervised (mBART-based), zero-shot and few-shot (GPT3-based) methods for rewriting sentences. We evaluate our models using a questionnaire study and a suite of automatic metrics.

Previous approaches to turn-taking and response generation in conversational systems have treated it as a two-stage process: First, the end of a turn is detected (based on conversation history), then the system generates an appropriate response. Humans, however, do not take the turn just because it is likely, but also consider whether what they want to say fits the position. In this paper, we present a model (an extension of TurnGPT) that conditions the end-of-turn prediction on both conversation history and what the next speaker wants to say. We found that our model consistently outperforms the baseline model in a variety of metrics. The improvement is most prominent in two scenarios where turn predictions can be ambiguous solely from the conversation history: 1) when the current utterance contains a statement followed by a question; 2) when the end of the current utterance semantically matches the response. Treating the turn-prediction and response-ranking as a one-stage process, our findings suggest that our model can be used as an incremental response ranker, which can be applied in various settings.

Recent advances in language modeling have enabled new conversational systems. In particular, it is often desirable for people to make choices among specified options when using such systems. We address the problem of reference resolution, when people use natural expressions to choose between real world entities. For example, given the choice ‘Should we make a Simnel cake or a Pandan cake¿ a natural response from a non-expert may be indirect: ‘let’s make the green one‘. Reference resolution has been little studied with natural expressions, thus robustly understanding such language has large potential for improving naturalness in dialog, recommendation, and search systems. We create AltEntities (Alternative Entities), a new public dataset of entity pairs and utterances, and develop models for the disambiguation problem. Consisting of 42K indirect referring expressions across three domains, it enables for the first time the study of how large language models can be adapted to this task. We find they achieve 82%-87% accuracy in realistic settings, which while reasonable also invites further advances.

Natural language often contains ambiguities that can lead to misinterpretation and miscommunication. While humans can handle ambiguities effectively by asking clarifying questions and/or relying on contextual cues and common-sense knowledge, resolving ambiguities can be notoriously hard for machines. In this work, we study ambiguities that arise in text-to-image generative models. We curate the Text-to-image Ambiguity Benchmark (TAB) dataset to study different types of ambiguities in text-to-image generative models. We then propose the Text-to-ImagE Disambiguation (TIED) framework to disambiguate the prompts given to the text-to-image generative models by soliciting clarifications from the end user. Through automatic and human evaluations, we show the effectiveness of our framework in generating more faithful images aligned with end user intention in the presence of ambiguities.

Prompt tuning is one of the successful approaches for parameter-efficient tuning of pre-trained language models. Despite being arguably the most parameter-efficient (tuned soft prompts constitute <0.1% of total parameters), it typically performs worse than other efficient tuning methods and is quite sensitive to hyper-parameters. In this work, we introduce Residual Prompt Tuning - a simple and efficient method that significantly improves the performance and stability of prompt tuning. We propose to reparameterize soft prompt embeddings using a shallow network with a residual connection. Our experiments show that Residual Prompt Tuning significantly outperforms prompt tuning across T5-Large, T5-Base and BERT-Base models. Notably, our method reaches +7 points improvement over prompt tuning on SuperGLUE benchmark with T5-Base model and allows to reduce the prompt length by 10 times without hurting performance. In addition, we show that our approach is robust to the choice of learning rate and prompt initialization, and is effective in few-shot settings.

Scientific progress in NLP rests on the reproducibility of researchers’ claims. The *CL conferences created the NLP Reproducibility Checklist in 2020 to be completed by authors at submission to remind them of key information to include. We provide the first analysis of the Checklist by examining 10,405 anonymous responses to it. First, we find evidence of an increase in reporting of information on efficiency, validation performance, summary statistics, and hyperparameters after the Checklist’s introduction. Further, we show acceptance rate grows for submissions with more Yes responses. We find that the 44% of submissions that gather new data are 5% less likely to be accepted than those that did not; the average reviewer-rated reproducibility of these submissions is also 2% lower relative to the rest. We find that only 46% of submissions claim to open-source their code, though submissions that do have 8% higher reproducibility score relative to those that do not, the most for any item. We discuss what can be inferred about the state of reproducibility in NLP, and provide a set of recommendations for future conferences, including: a) allowing submitting code and appendices one week after the deadline, and b) measuring dataset reproducibility by a checklist of data collection practices.