COLING 2024 论文列表

Deep neural networks can be vulnerable to adversarial attacks, even for the mainstream Transformer-based models. Although several robustness enhancement approaches have been proposed, they usually focus on some certain type of perturbation. As the types of attack can be various and unpredictable in practical scenarios, a general and strong defense method is urgently in require. We notice that most well-trained models can be weakly robust in the perturbation space, i.e., only a small ratio of adversarial examples exist. Inspired by the weak robust property, this paper presents a novel ensemble method for enhancing robustness. We propose a lightweight framework PAD to save computational resources in realizing an ensemble. Instead of training multiple models, a plugin module is designed to perturb the parameters of a base model which can achieve the effect of multiple models. Then, to diversify adversarial example distributions among different models, we promote each model to have different attention patterns via optimizing a diversity measure we defined. Experiments on various widely-used datasets and target models show that PAD can consistently improve the defense ability against many types of adversarial attacks while maintaining accuracy on clean data. Besides, PAD also presents good interpretability via visualizing diverse attention patterns.

In an era characterized by the rapid proliferation of information, the pervasive issues of misinformation and disinformation have significantly impacted numerous individuals. Consequently, the evaluation of information’s truthfulness and accuracy has garnered substantial attention among researchers. In this work, we present a novel fact-checking framework called PACAR, fact-checking based on planning and customized action reasoning using LLMs. It comprises four modules: a claim decomposer with self-reflection, an LLM-centric planner module, an executor for carrying out planned actions, and a verifier module that assesses veracity and generates explanations based on the overall reasoning process. Unlike previous work that employs single-path decision-making and single-step verdict prediction, PACAR focuses on the use of LLMs in dynamic planning and execution of actions. Furthermore, in contrast to previous work that relied primarily on general reasoning, we introduce tailored actions such as numerical reasoning and entity disambiguation to effectively address potential challenges in fact-checking. Our PACAR framework, incorporating LLM-centric planning along with customized action reasoning, significantly outperforms baseline methods across three datasets from different domains and with varying complexity levels. Additional experiments, including multidimensional and sliced observations, demonstrate the effectiveness of PACAR and offer valuable insights for the advancement of automated fact-checking.

Out-of-Domain (OOD) intent detection is vital for practical dialogue systems, and it usually requires considering multi-turn dialogue contexts. However, most previous OOD intent detection approaches are limited to single dialogue turns. In this paper, we introduce a context-aware OOD intent detection (Caro) framework to model multi-turn contexts in OOD intent detection tasks. Specifically, we follow the information bottleneck principle to extract robust representations from multi-turn dialogue contexts. Two different views are constructed for each input sample and the superfluous information not related to intent detection is removed using a multi-view information bottleneck loss. Moreover, we also explore utilizing unlabeled data in Caro. A two-stage training process is introduced to mine OOD samples from these unlabeled data, and these OOD samples are used to train the resulting model with a bootstrapping approach. Comprehensive experiments demonstrate that Caro establishes state-of-the-art performances on multi-turn OOD detection tasks by improving the F1-OOD score of over 29% compared to the previous best method.

This paper presents GePaDe_SpkAtt , a new corpus for speaker attribution in German parliamentary debates, with more than 7,700 manually annotated events of speech, thought and writing. Our role inventory includes the sources, addressees, messages and topics of the speech event and also two additional roles, medium and evidence. We report baseline results for the automatic prediction of speech events and their roles, with high scores for both, event triggers and roles. Then we apply our model to predict speech events in 20 years of parliamentary debates and investigate the use of factives in the rhetoric of MPs.

Large language models (LLMs) use pretraining to predict the subsequent word; however, their expansion requires significant computing resources. Numerous big tech companies and research institutes have developed multilingual LLMs (MLLMs) to meet current demands, overlooking less-resourced languages (LRLs). This study proposed three strategies to enhance the performance of LRLs based on the publicly available MLLMs. First, the MLLM vocabularies of LRLs were expanded to enhance expressiveness. Second, bilingual data were used for pretraining to align the high- and less-resourced languages. Third, a high-quality small-scale instruction dataset was constructed and instruction-tuning was performed to augment the LRL. The experiments employed the Llama2 model and Korean was used as the LRL, which was quantitatively evaluated against other developed LLMs across eight tasks. Furthermore, a qualitative assessment was performed based on human evaluation and GPT4. Experimental results showed that our proposed Bllossom model exhibited superior performance in qualitative analyses compared to previously proposed Korean monolingual models.

As in the existing opinion summary data set, more than 70% are positive texts, the current opinion summarization approaches are reluctant to generate the negative opinion summary given the input of negative opinions. To address such sentiment bias, two approaches are proposed through two perspectives: model-specific and model-agnostic. For the model-specific approach, a variational autoencoder is proposed to disentangle the input representation into sentiment-relevant and sentiment-irrelevant components through adversarial loss. Therefore, the sentiment information in the input is kept and employed for the following decoding which avoids interference of content information with emotional signals. To further avoid relying on some specific opinion summarization frameworks, a model-agnostic approach based on counterfactual data augmentation is proposed. A dataset with a more balanced emotional polarity distribution is constructed using a large pre-trained language model based on some pairwise and mini-edited principles. Experimental results show that the sentiment consistency of the generated summaries is significantly improved using the proposed approaches, while their semantics quality is unaffected.

Opinion mining is an important task in natural language processing. The MPQA Opinion Corpus is a fine-grained and comprehensive dataset of private states (i.e., the condition of a source who has an attitude which may be directed toward a target) based on context. Although this dataset was released years ago, because of its complex definition of annotations and hard-to-read data format, almost all existing research works have only focused on a small subset of the dataset. In this paper, we present a comprehensive study of the entire MPQA 2.0 dataset. In order to achieve this goal, we first provide a clean version of MPQA 2.0 in a more interpretable format. Then, we propose two novel approaches for opinion mining, establishing new high baselines for future work. We use two pre-trained large language models, BERT and T5, to automatically identify the type, polarity, and intensity of private states expressed in phrases, and we use T5 to detect opinion expressions and their agents (i.e., sources).

We develop and evaluate multilingual scientific documents similarity measurement models in this work. Such models can be used to find related papers in different languages, which can help multilingual researchers find and explore papers more efficiently. We propose the first multilingual scientific documents dataset, Open-access Multilingual Scientific Documents (OpenMSD), which has 74M papers in 103 languages and 778M citation pairs. With OpenMSD, we develop multilingual SDSM models by adjusting and extending the state-of-the-art methods designed for English SDSM tasks. We find that: (i)Some highly successful methods in English SDSM yield significantly worse performance in multilingual SDSM. (ii)Our best model, which enriches the non-English papers with English summaries, outperforms strong baselines by 7% (in mean average precision) on multilingual SDSM tasks, without compromising the performance on English SDSM tasks.

Modern Transformers achieved impressive results on various Natural Language Processing tasks over the last few years. The one downside of this success is the size of these models. Huge capacity, which sometimes surpasses billions of parameters, improves generalization abilities, but makes it difficult to employ. Developing field of model compression seeks to reduce the model size and inference latency. This research focuses on one of the compression techniques — Post-Training Quantization. We present a methodology to effectively quantize at least 95% of Transformer weights and corresponding activations to INT8 without any access to task-specific data so the drop in performance does not exceed 0.02%. Furthermore, we provide intriguing observations that reflect cross-domain nature of some of the quantization properties.

The superior performance of supervised classification methods in the information extraction (IE) area heavily relies on a large amount of gold standard data. Recent zero-shot classification methods converted the task to other NLP tasks (e.g., textual entailment) and used off-the-shelf models of these NLP tasks to directly perform inference on the test data without using a large amount of IE annotation data. A potentially valuable by-product of these methods is the large-scale silver standard data, i.e., pseudo-labeled data by the off-the-shelf models of other NLP tasks. However, there is no further investigation into the use of these data. In this paper, we propose a new framework, Clean-LaVe, which aims to utilize silver standard data to enhance the zero-shot performance. Clean-LaVe includes four phases: (1) Obtaining silver data; (2) Identifying relatively clean data from silver data; (3) Finetuning the off-the-shelf model using clean data; (4) Inference on the test data. The experimental results show that Clean-LaVe can outperform the baseline by 5% and 6% on TACRED and Wiki80 dataset in the zero-shot relation classification task, and by 3% ~7 % on Smile (Korean and Polish) in the zero-shot cross-lingual relation classification task, and by 8% on ACE05-E+ in the zero-shot event argument classification task.

Language models have long been shown to embed geographical information in their hidden representations. This line of work has recently been revisited by extending this result to Large Language Models (LLMs). In this paper, we propose to fill the gap between well-established and recent literature by observing how geographical knowledge evolves when scaling language models. We show that geographical knowledge is observable even for tiny models, and that it scales consistently as we increase the model size. Notably, we observe that larger language models cannot mitigate the geographical bias that is inherent to the training data.

Prompt Tuning is a popular parameter-efficient finetuning method for pre-trained large language models (PLMs). Based on experiments with RoBERTa, it has been suggested that Prompt Tuning activates specific neurons in the transformer’s feed-forward networks, that are highly predictive and selective for the given task. In this paper, we study the robustness of Prompt Tuning in relation to these “skill neurons”, using RoBERTa and T5. We show that prompts tuned for a specific task are transferable to tasks of the same type but are not very robust to adversarial data. While prompts tuned for RoBERTa yield below-chance performance on adversarial data, prompts tuned for T5 are slightly more robust and retain above-chance performance in two out of three cases. At the same time, we replicate the finding that skill neurons exist in RoBERTa and further show that skill neurons also exist in T5. Interestingly, the skill neurons of T5 determined on non-adversarial data are also among the most predictive neurons on the adversarial data, which is not the case for RoBERTa. We conclude that higher adversarial robustness may be related to a model’s ability to consistently activate the relevant skill neurons on adversarial data.

One of the prominent issues stifling the current generation of large language models is their limited context length. Recent proprietary models such as GPT-4 and Claude 2 have introduced longer context lengths, 8k/32k and 100k, respectively; however, despite the efforts in the community, most common models, such as LLama-2, have a context length of 4k or less. Unlimiformer (Bertsch et al., 2023) is a recently popular vector-retrieval augmentation method that offloads cross-attention computations to a kNN index. However, its main limitation is incompatibility with decoder-only transformers out of the box. In this work, we explore practical considerations of adapting Unlimiformer to decoder-only transformers and introduce a series of modifications to overcome this limitation. Moreover, we expand the original experimental setup on summarization to include a new task (i.e., free-form Q&A) and an instruction-tuned model (i.e., a custom 6.7B GPT model). Our results showcase the effectiveness of these modifications on summarization, performing on par with a model with 2x the context length. Moreover, we discuss limitations and future directions for free-form Q&A and instruction-tuned models.

Citations using URL (URL citations) that appear in scholarly papers can be used as an information source for the research resource search engines. In particular, the information about the types of cited resources and reasons for their citation is crucial to describe the resources and their relations in the search services. To obtain this information, previous studies proposed some methods for classifying URL citations. However, their methods trained the model using a simple fine-tuning strategy and exhibited insufficient performance. We propose a classification method using a novel intermediate task. Our method trains the model on our intermediate task of identifying whether sample pairs belong to the same class before being fine-tuned on the target task. In the experiment, our method outperformed previous methods using the simple fine-tuning strategy with higher macro F-scores for different model sizes and architectures. Our analysis results indicate that the model learns the class boundaries of the target task by training our intermediate task. Our intermediate task also demonstrated higher performance and computational efficiency than an alternative intermediate task using triplet loss. Finally, we applied our method to other text classification tasks and confirmed the effectiveness when a simple fine-tuning strategy does not stably work.

With the emergence of numerous Large Language Models (LLM), the usage of such models in various Natural Language Processing (NLP) applications is increasing extensively. Counterspeech generation is one such key task where efforts are made to develop generative models by fine-tuning LLMs with hatespeech - counterspeech pairs, but none of these attempts explores the intrinsic properties of large language models in zero-shot settings. In this work, we present a comprehensive analysis of the performances of four LLMs namely GPT-2, DialoGPT, ChatGPT and FlanT5 in zero-shot settings for counterspeech generation, which is the first of its kind. For GPT-2 and DialoGPT, we further investigate the deviation in performance with respect to the sizes (small, medium, large) of the models. On the other hand, we propose three different prompting strategies for generating different types of counterspeech and analyse the impact of such strategies on the performance of the models. Our analysis shows that there is an improvement in generation quality for two datasets (17%), however the toxicity increase (25%) with increase in model size. Considering type of model, GPT-2 and FlanT5 models are significantly better in terms of counterspeech quality but also have high toxicity as compared to DialoGPT. ChatGPT are much better at generating counter speech than other models across all metrics. In terms of prompting, we find that our proposed strategies help in improving counter speech generation across all the models.

In this article we look at how two different standards for lexical resources, TEI and OntoLex, deal with corpus citations in lexicons. We will focus on how corpus citations in retrodigitised dictionaries can be modelled using each of the two standards since this provides us with a suitably challenging use case. After looking at the structure of an example entry from a legacy dictionary, we examine the two approaches offered by the two different standards by outlining an encoding for the example entry using both of them (note that this article features the first extended discussion of how the Frequency Attestation and Corpus (FrAC) module of OntoLex deals with citations). After comparing the two approaches and looking at the advantages and disadvantages of both, we argue for a combination of both. In the last part of the article we discuss different ways of doing this, giving our preference for a strategy which makes use of RDFa.

This study addresses the application of encoder-only Pre-trained Language Models (PLMs) in keyphrase generation (KPG) amidst the broader availability of domain-tailored encoder-only models compared to encoder-decoder models. We investigate three core inquiries: (1) the efficacy of encoder-only PLMs in KPG, (2) optimal architectural decisions for employing encoder-only PLMs in KPG, and (3) a performance comparison between in-domain encoder-only and encoder-decoder PLMs across varied resource settings. Our findings, derived from extensive experimentation in two domains reveal that with encoder-only PLMs, although keyphrase extraction with Conditional Random Fields slightly excels in identifying present keyphrases, the KPG formulation renders a broader spectrum of keyphrase predictions. Additionally, prefix-LM fine-tuning of encoder-only PLMs emerges as a strong and data-efficient strategy for KPG, outperforming general-domain seq2seq PLMs. We also identify a favorable parameter allocation towards model depth rather than width when employing encoder-decoder architectures initialized with encoder-only PLMs. The study sheds light on the potential of utilizing encoder-only PLMs for advancing KPG systems and provides a groundwork for future KPG methods. Our code and pre-trained checkpoints are released at https://github.com/uclanlp/DeepKPG.

Pretrained language models can be applied for various downstream tasks but are susceptible to subtle perturbations. Most adversarial defense methods often introduce adversarial training during the fine-tuning phase to enhance empirical robustness. However, the repeated execution of adversarial training hinders training efficiency when transitioning to different tasks. In this paper, we explore the transferability of robustness within subnetworks and leverage this insight to introduce a novel adversarial defense method ORTicket, eliminating the need for separate adversarial training across diverse downstream tasks. Specifically, (i) pruning the full model using the MLM task (the same task employed for BERT pretraining) yields a task-agnostic robust subnetwork(i.e., winning ticket in Lottery Ticket Hypothesis); and (ii) fine-tuning this subnetwork for downstream tasks. Extensive experiments demonstrate that our approach achieves comparable robustness to other defense methods while retaining the efficiency of traditional fine-tuning.This also confirms the significance of selecting MLM task for identifying the transferable robust subnetwork. Furthermore, our method is orthogonal to other adversarial training approaches, indicating the potential for further enhancement of model robustness.

We focus on morphological inflection in out-of-vocabulary (OOV) conditions, an under-researched subtask in which state-of-the-art systems usually are less effective. We developed three systems: a retrograde model and two sequence-to-sequence (seq2seq) models based on LSTM and Transformer. For testing in OOV conditions, we automatically extracted a large dataset of nouns in the morphologically rich Czech language, with lemma-disjoint data splits, and we further manually annotated a real-world OOV dataset of neologisms. In the standard OOV conditions, Transformer achieves the best results, with increasing performance in ensemble with LSTM, the retrograde model and SIGMORPHON baselines. On the real-world OOV dataset of neologisms, the retrograde model outperforms all neural models. Finally, our seq2seq models achieve state-of-the-art results in 9 out of 16 languages from SIGMORPHON 2022 shared task data in the OOV evaluation (feature overlap) in the large data condition. We release the Czech OOV Inflection Dataset for rigorous evaluation in OOV conditions. Further, we release the inflection system with the seq2seq models as a ready-to-use Python library.