ACL 2023 论文列表

When acquiring syntax, children consistently choose hierarchical rules over competing non-hierarchical possibilities. Is this preference due to a learning bias for hierarchical structure, or due to more general biases that interact with hierarchical cues in children’s linguistic input? We explore these possibilities by training LSTMs and Transformers - two types of neural networks without a hierarchical bias - on data similar in quantity and content to children’s linguistic input: text from the CHILDES corpus. We then evaluate what these models have learned about English yes/no questions, a phenomenon for which hierarchical structure is crucial. We find that, though they perform well at capturing the surface statistics of child-directed speech (as measured by perplexity), both model types generalize in a way more consistent with an incorrect linear rule than the correct hierarchical rule. These results suggest that human-like generalization from text alone requires stronger biases than the general sequence-processing biases of standard neural network architectures.

This paper presents an investigation into the differences between processing monolingual input and code-switching (CSW) input in the context of machine translation (MT). Specifically, we compare the performance of three MT systems (Google, mBART-50 and M2M-100-big) in terms of their ability to translate monolingual Vietnamese, a low-resource language, and Vietnamese-English CSW respectively. To our knowledge, this is the first study to systematically analyse what might happen when multilingual MT systems are exposed to CSW data using both automatic and human metrics. We find that state-of-the-art neural translation systems not only achieve higher scores on automatic metrics when processing CSW input (compared to monolingual input), but also produce translations that are consistently rated as more semantically faithful by humans. We further suggest that automatic evaluation alone is insufficient for evaluating the translation of CSW input. Our findings establish a new benchmark that offers insights into the relationship between MT and CSW.

Masked language modeling (MLM) is a widely used self-supervised pretraining objective, where a model needs to predict an original token that is replaced with a mask given contexts. Although simpler and computationally efficient pretraining objectives, e.g., predicting the first character of a masked token, have recently shown comparable results to MLM, no objectives with a masking scheme actually outperform it in downstream tasks. Motivated by the assumption that their lack of complexity plays a vital role in the degradation, we validate whether more complex masked objectives can achieve better results and investigate how much complexity they should have to perform comparably to MLM. Our results using GLUE, SQuAD, and Universal Dependencies benchmarks demonstrate that more complicated objectives tend to show better downstream results with at least half of the MLM complexity needed to perform comparably to MLM. Finally, we discuss how we should pretrain a model using a masked objective from the task complexity perspective.

Syntactic parsing is the task of assigning a syntactic structure to a sentence. There are two popular syntactic parsing methods: constituency and dependency parsing. Recent works have used syntactic embeddings based on constituency trees, incremental top-down parsing, and other word syntactic features for brain activity prediction given the text stimuli to study how the syntax structure is represented in the brain’s language network. However, the effectiveness of dependency parse trees or the relative predictive power of the various syntax parsers across brain areas, especially for the listening task, is yet unexplored. In this study, we investigate the predictive power of the brain encoding models in three settings: (i) individual performance of the constituency and dependency syntactic parsing based embedding methods, (ii) efficacy of these syntactic parsing based embedding methods when controlling for basic syntactic signals, (iii) relative effectiveness of each of the syntactic embedding methods when controlling for the other. Further, we explore the relative importance of syntactic information (from these syntactic embedding methods) versus semantic information using BERT embeddings. We find that constituency parsers help explain activations in the temporal lobe and middle-frontal gyrus, while dependency parsers better encode syntactic structure in the angular gyrus and posterior cingulate cortex. Although semantic signals from BERT are more effective compared to any of the syntactic features or embedding methods, syntactic embedding methods explain additional variance for a few brain regions.

Natural Language Processing (NLP) models based on Machine Learning (ML) are susceptible to adversarial attacks – malicious algorithms that imperceptibly modify input text to force models into making incorrect predictions. However, evaluations of these attacks ignore the property of imperceptibility or study it under limited settings. This entails that adversarial perturbations would not pass any human quality gate and do not represent real threats to human-checked NLP systems. To bypass this limitation and enable proper assessment (and later, improvement) of NLP model robustness, we have surveyed 378 human participants about the perceptibility of text adversarial examples produced by state-of-the-art methods. Our results underline that existing text attacks are impractical in real-world scenarios where humans are involved. This contrasts with previous smaller-scale human studies, which reported overly optimistic conclusions regarding attack success. Through our work, we hope to position human perceptibility as a first-class success criterion for text attacks, and provide guidance for research to build effective attack algorithms and, in turn, design appropriate defence mechanisms.

Motivated by the question of the extent to which large language models “understand” social intelligence, we investigate the ability of such models to generate correct responses to questions involving descriptions of faux pas situations. The faux pas test is a test used in clinical psychology, which is known to be more challenging for children than individual tests of theory-of-mind or social intelligence. Our results demonstrate that, while the models seem to sometimes offer correct responses, they in fact struggle with this task, and that many of the seemingly correct responses can be attributed to over-interpretation by the human reader (“the ELIZA effect”). An additional phenomenon observed is the failure of most models to generate a correct response to presupposition questions. Finally, in an experiment in which the models are tasked with generating original faux pas stories, we find that while some models are capable of generating novel faux pas stories, the stories are all explicit, as the models are limited in their abilities to describe situations in an implicit manner.

Incomplete utterance rewriting (IUR) aims to restore the incomplete utterance with sufficient context information for comprehension. This paper introduces a simple yet efficient IUR method. Different from prior studies, we first employ only one-layer MLP architecture to mine latent semantic information between joint utterances for IUR task (MIUR). After that, we conduct a joint feature matrix to predict the token type and thus restore the incomplete utterance. The well-designed network and simple architecture make our method significantly superior to existing methods in terms of quality and inference speedOur code is available at https://github.com/IMU-MachineLearningSXD/MIUR.

The multi-answer phenomenon, where a question may have multiple answers scattered in the document, can be well handled by humans but is challenging enough for machine reading comprehension (MRC) systems. Despite recent progress in multi-answer MRC, there lacks a systematic analysis of how this phenomenon arises and how to better address it. In this work, we design a taxonomy to categorize commonly-seen multi-answer MRC instances, with which we inspect three multi-answer datasets and analyze where the multi-answer challenge comes from. We further analyze how well different paradigms of current multi-answer MRC models deal with different types of multi-answer instances. We find that some paradigms capture well the key information in the questions while others better model the relation between questions and contexts. We thus explore strategies to make the best of the strengths of different paradigms. Experiments show that generation models can be a promising platform to incorporate different paradigms. Our annotations and code are released for further research.

Compositional generalization–understanding unseen combinations of seen primitives–is an essential reasoning capability in human intelligence. The AI community mainly studies this capability by fine-tuning neural networks on lots of training samples, while it is still unclear whether and how in-context learning–the prevailing few-shot paradigm based on large language models–exhibits compositional generalization. In this paper, we present CoFe, a test suite to investigate in-context compositional generalization. We find that the compositional generalization performance can be easily affected by the selection of in-context examples, thus raising the research question what the key factors are to make good in-context examples for compositional generalization. We study three potential factors: similarity, diversity and complexity. Our systematic experiments indicate that in-context examples should be structurally similar to the test case, diverse from each other, and individually simple. Furthermore, two strong limitations are observed: in-context compositional generalization on fictional words is much weaker than that on commonly used ones; it is still critical that the in-context examples should cover required linguistic structures, even though the backbone model has been pre-trained on large corpus. We hope our analysis would facilitate the understanding and utilization of in-context learning paradigm.

Hedging is a strategy for softening the impact of a statement in conversation. In reducing the strength of an expression, it may help to avoid embarrassment (more technically, “face threat”) to one’s listener. For this reason, it is often found in contexts of instruction, such as tutoring. In this work, we develop a model of hedge generation based on i) fine-tuning state-of-the-art language models trained on human-human tutoring data, followed by ii) reranking to select the candidate that best matches the expected hedging strategy within a candidate pool using a hedge classifier. We apply this method to a natural peer-tutoring corpus containing a significant number of disfluencies, repetitions, and repairs. The results show that generation in this noisy environment is feasible with reranking. By conducting an error analysis for both approaches, we reveal the challenges faced by systems attempting to accomplish both social and task-oriented goals in conversation.

In recent years, language models have drastically grown in size, and the abilities of these models have been shown to improve with scale. The majority of recent scaling laws studies focused on high-compute high-parameter count settings, leaving the question of when these abilities begin to emerge largely unanswered. In this paper, we investigate whether the effects of pre-training can be observed when the problem size is reduced, modeling a smaller, reduced-vocabulary language. We show the benefits of pre-training with masked language modeling (MLM) objective in models as small as 1.25M parameters, and establish a strong correlation between pre-training perplexity and downstream performance (GLUE benchmark). We examine downscaling effects, extending scaling laws to models as small as ~1M parameters. At this scale, we observe a break of the power law for compute-optimal models and show that the MLM loss does not scale smoothly with compute-cost (FLOPs) below 2.2 × 1015 FLOPs. We also find that adding layers does not always benefit downstream performance.Our filtered pre-training data, reduced English vocabulary, and code are available at https://github.com/text-machine-lab/mini_bertgithub.com/text-machine-lab/mini_bert

Current methods for generating attractive headlines often learn directly from data, which bases attractiveness on the number of user clicks and views. Although clicks or views do reflect user interest, they can fail to reveal how much interest is raised by the writing style and how much is due to the event or topic itself. Also, such approaches can lead to harmful inventions by over-exaggerating the content, aggravating the spread of false information. In this work, we propose HonestBait, a novel framework for solving these issues from another aspect: generating headlines using forward references (FRs), a writing technique often used for clickbait. A self-verification process is included during training to avoid spurious inventions. We begin with a preliminary user study to understand how FRs affect user interest, after which we present PANCO, an innovative dataset containing pairs of fake news with verified news for attractive but faithful news headline generation. Auto matic metrics and human evaluations show that our framework yields more attractive results (+11.25% compared to human-written verified news headlines) while maintaining high veracity, which helps promote real information to fight against fake news.

We propose a method for formulating CCG as a recursive composition in a continuous vector space. Recent CCG supertagging and parsing models generally demonstrate high performance, yet rely on black-box neural architectures to implicitly model phrase structure dependencies. Instead, we leverage the method of holographic embeddings as a compositional operator to explicitly model the dependencies between words and phrase structures in the embedding space. Experimental results revealed that holographic composition effectively improves the supertagging accuracy to achieve state-of-the-art parsing performance when using a C&C parser. The proposed span-based parsing algorithm using holographic composition achieves performance comparable to state-of-the-art neural parsing with Transformers. Furthermore, our model can semantically and syntactically infill text at the phrase level due to the decomposability of holographic composition.

Graph-based prediction is essential in NLP tasks such as temporal knowledge graph completion. A cardinal question in this field is, how to predict the future links, nodes, and attributes of a time-evolving attributed graph? Unfortunately, existing techniques assume that each link, node, and attribute prediction is independent, and fall short of predicting the appearance of new nodes that were not observed in the past. In this paper, we address two interrelated questions; (1) can we exploit task interdependence to improve prediction accuracy? and (2) can we predict new nodes with their attributes? We propose a unified framework that predicts node attributes and topology changes such as the appearance and disappearance of links and the emergence and loss of nodes. This frame-work comprises components for independent and interactive prediction and for predicting new nodes. Our experimental study using real-world data confirms that our interdependent prediction framework achieves higher accuracy than methods based on independent prediction.

Temporal knowledge graph (TKG) completion models typically rely on having access to the entire graph during training. However, in real-world scenarios, TKG data is often received incrementally as events unfold, leading to a dynamic non-stationary data distribution over time. While one could incorporate fine-tuning to existing methods to allow them to adapt to evolving TKG data, this can lead to forgetting previously learned patterns. Alternatively, retraining the model with the entire updated TKG can mitigate forgetting but is computationally burdensome. To address these challenges, we propose a general continual training framework that is applicable to any TKG completion method, and leverages two key ideas: (i) a temporal regularization that encourages repurposing of less important model parameters for learning new knowledge, and (ii) a clustering-based experience replay that reinforces the past knowledge by selectively preserving only a small portion of the past data. Our experimental results on widely used event-centric TKG datasets demonstrate the effectiveness of our proposed continual training framework in adapting to new events while reducing catastrophic forgetting. Further, we perform ablation studies to show the effectiveness of each component of our proposed framework. Finally, we investigate the relation between the memory dedicated to experience replay and the benefit gained from our clustering-based sampling strategy.

Context information modeling is an important task in conversational KBQA. However, existing methods usually assume the independence of utterances and model them in isolation. In this paper, we propose a History Semantic Graph Enhanced KBQA model (HSGE) that is able to effectively model long-range semantic dependencies in conversation history while maintaining low computational cost. The framework incorporates a context-aware encoder, which employs a dynamic memory decay mechanism and models context at different levels of granularity. We evaluate HSGE on a widely used benchmark dataset for complex sequential question answering. Experimental results demonstrate that it outperforms existing baselines averaged on all question types.

Despite the extensive applications of relation extraction (RE) tasks in various domains, little has been explored in the historical context, which contains promising data across hundreds and thousands of years. To promote the historical RE research, we present HistRED constructed from Yeonhaengnok. Yeonhaengnok is a collection of records originally written in Hanja, the classical Chinese writing, which has later been translated into Korean. HistRED provides bilingual annotations such that RE can be performed on Korean and Hanja texts. In addition, HistRED supports various self-contained subtexts with different lengths, from a sentence level to a document level, supporting diverse context settings for researchers to evaluate the robustness of their RE models. To demonstrate the usefulness of our dataset, we propose a bilingual RE model that leverages both Korean and Hanja contexts to predict relations between entities. Our model outperforms monolingual baselines on HistRED, showing that employing multiple language contexts supplements the RE predictions. The dataset is publicly available at: https://huggingface.co/datasets/Soyoung/HistRED under CC BY-NC-ND 4.0 license.

Language Models (LM) are becoming more and more useful for providing representations upon which to train Natural Language Processing applications. However, there is now clear evidence that attention-based transformers require a critical amount of language data to produce good enough LMs. The question we have addressed in this paper is to what extent the critical amount of data varies for languages of different morphological typology, in particular those that have a rich inflectional morphology, and whether the tokenization method to preprocess the data can make a difference. These details can be important for low-resourced languages that need to plan the production of datasets. We evaluated intrinsically and extrinsically the differences of five different languages with different pretraining dataset sizes and three different tokenization methods for each. The results confirm that the size of the vocabulary due to morphological characteristics is directly correlated with both the LM perplexity and the performance of two typical downstream tasks such as NER identification and POS labeling. The experiments also provide new evidence that a canonical tokenizer can reduce perplexity by more than a half for a polysynthetic language like Quechua as well as raising F1 from 0.8 to more than 0.9 in both downstream tasks with a LM trained with only 6M tokens.

Due to the complex label hierarchy and intensive labeling cost in practice, the hierarchical text classification (HTC) suffers a poor performance especially when low-resource or few-shot settings are considered. Recently, there is a growing trend of applying prompts on pre-trained language models (PLMs), which has exhibited effectiveness in the few-shot flat text classification tasks. However, limited work has studied the paradigm of prompt-based learning in the HTC problem when the training data is extremely scarce. In this work, we define a path-based few-shot setting and establish a strict path-based evaluation metric to further explore few-shot HTC tasks. To address the issue, we propose the hierarchical verbalizer (“HierVerb”), a multi-verbalizer framework treating HTC as a single- or multi-label classification problem at multiple layers and learning vectors as verbalizers constrained by hierarchical structure and hierarchical contrastive learning. In this manner, HierVerb fuses label hierarchy knowledge into verbalizers and remarkably outperforms those who inject hierarchy through graph encoders, maximizing the benefits of PLMs. Extensive experiments on three popular HTC datasets under the few-shot settings demonstrate that prompt with HierVerb significantly boosts the HTC performance, meanwhile indicating an elegant way to bridge the gap between the large pre-trained model and downstream hierarchical classification tasks.