COLING 2024 Papers

The accurate identification of loanwords within a given text holds significant potential as a valuable tool for addressing data augmentation and mitigating data sparsity issues. Such identification can improve the performance of various natural language processing tasks, particularly in the context of low-resource languages that lack standardized spelling conventions.This research proposes a supervised method to identify loanwords in Emakhuwa, borrowed from Portuguese. Our methodology encompasses a two-fold approach. Firstly, we employ traditional machine learning algorithms incorporating handcrafted features, including language-specific and similarity-based features. We build upon prior studies to extract similarity features and propose utilizing two external resources: a Sequence-to-Sequence model and a dictionary. This innovative approach allows us to identify loanwords solely by analyzing the target word without prior knowledge about its donor counterpart. Furthermore, we fine-tune the pre-trained CANINE model for the downstream task of loanword detection, which culminates in the impressive achievement of the F1-score of 93%. To the best of our knowledge, this study is the first of its kind focusing on Emakhuwa, and the preliminary results are promising as they pave the way to further advancements.

Impact assessment is an evolving area of research that aims at measuring and predicting the potential effects of projects or programs. Measuring the impact of scientific research is a vibrant subdomain, closely intertwined with impact assessment. A recurring obstacle pertains to the absence of an efficient framework which can facilitate the analysis of lengthy reports and text labeling. To address this issue, we propose a framework for automatically assessing the impact of scientific research projects by identifying pertinent sections in project reports that indicate the potential impacts. We leverage a mixed-method approach, combining manual annotations with supervised machine learning, to extract these passages from project reports. We experiment with different machine learning algorithms, including traditional statistical models as well as pre-trained transformer language models. Our experiments show that our proposed method achieves accuracy scores up to 0.81, and that our method is generalizable to scientific research from different domains and different languages.

We explore a strategy to handle controversial topics in LLM-based chatbots based on Wikipedia’s Neutral Point of View (NPOV) principle: acknowledge the absence of a single true answer and surface multiple perspectives. We frame this as retrieval augmented generation, where perspectives are retrieved from a knowledge base and the LLM is tasked with generating a fluent and faithful response from the given perspectives. As a starting point, we use a deterministic retrieval system and then focus on common LLM failure modes that arise during this approach to text generation, namely hallucination and coverage errors. We propose and evaluate three methods to detect such errors based on (1) word-overlap, (2) salience, and (3) LLM-based classifiers. Our results demonstrate that LLM-based classifiers, even when trained only on synthetic errors, achieve high error detection performance, with ROC AUC scores of 95.3% for hallucination and 90.5% for coverage error detection on unambiguous error cases. We show that when no training data is available, our other methods still yield good results on hallucination (84.0%) and coverage error (85.2%) detection.

Cybercrime is a serious and growing threat affecting millions of people worldwide. Detecting cybercrimes from text messages is challenging, as it requires understanding the linguistic and cultural nuances of different languages and regions. Roman Urdu is a widely used language in Pakistan and other South Asian countries, however, it lacks sufficient resources and tools for natural language processing and cybercrime detection. To address this problem, we make three main contributions in this paper. (1) We create and release CRU, a benchmark dataset for text-based cybercrime detection in Roman Urdu, which covers a number of cybercrimes as defined by the Prevention of Electronic Crimes Act (PECA) of Pakistan. This dataset is annotated by experts following a standardized procedure based on Pakistan’s legal framework. (2) We perform experiments on four pre-trained language models (PLMs) for cybercrime text classification in Roman Urdu. Our results show that xlm-roberta-base is the best model for this task, achieving the highest performance on all metrics. (3) We explore the utility of prompt engineering techniques, namely prefix and cloze prompts, for enhancing the performance of PLMs for low-resource languages such as Roman Urdu. We analyze the impact of different prompt shapes and k-shot settings on the performance of xlm-roberta-base and bert-base-multilingual-cased. We find that prefix prompts are more effective than cloze prompts for Roman Urdu classification tasks, as they provide more contextually relevant completions for the models. Our work provides useful insights and resources for future research on cybercrime detection and text classification in low-resource languages.

Recent machine translation (MT) systems have overcome language barriers for a wide range of users, yet they still carry the risk of critical meaning deviation. Critical error detection (CED) is a task that identifies an inherent risk of catastrophic meaning distortions in the machine translation output. With the importance of reflecting cultural elements in detecting critical errors, we introduce the culture-aware “Politeness” type in detecting English-Korean critical translation errors. Besides, we facilitate two tasks by providing multiclass labels: critical error detection and critical error type classification (CETC). Empirical evaluations reveal that our introduced data augmentation approach using a newly presented perturber significantly outperforms existing baselines in both tasks. Further analysis highlights the significance of multiclass labeling by demonstrating its superior effectiveness compared to binary labels.

We show a novel approach to detecting noun abstraction within a large language model (LLM). Starting from a psychologically motivated set of noun pairs in taxonomic relationships, we instantiate surface patterns indicating hypernymy and analyze the attention matrices produced by BERT. We compare the results to two sets of counterfactuals and show that we can detect hypernymy in the abstraction mechanism, which cannot solely be related to the distributional similarity of noun pairs. Our findings are a first step towards the explainability of conceptual abstraction in LLMs.

Underpinning much of the recent progress in deep learning is the transformer architecture, which takes as input a sequence of embeddings E and emits an updated sequence of embeddings E’. A special [CLS] embedding is often included in this sequence, serving as a description of the sequence once processed and used as the basis for subsequent sequence-level tasks. The processed [CLS] embedding loses utility, however, when the model is presented with a multi-entity sequence and asked to perform an entity-specific task. When processing a multi-speaker dialogue, for example, the [CLS] embedding describes the entire dialogue, not any individual utterance/speaker. Existing methods toward entity-specific prediction involve redundant computation or post-processing outside of the transformer. We present a novel methodology for deriving entity-specific embeddings from a multi-entity sequence completely within the transformer, with a loose definition of entity amenable to many problem spaces. To show the generic applicability of our method, we apply it to widely different tasks: emotion recognition in conversation and player performance projection in baseball and show that it can be used to achieve SOTA in both. Code can be found at https://github.com/c-heat16/EntitySpecificEmbeddings.

Language Models pretrained on large textual data have been shown to encode different types of knowledge simultaneously. Traditionally, only the features from the last layer are used when adapting to new tasks or data. We put forward that, when using or finetuning deep pretrained models, intermediate layer features that may be relevant to the downstream task are buried too deep to be used efficiently in terms of needed samples or steps. To test this, we propose a new layer fusion method: Depth-Wise Attention (DWAtt), to help re-surface signals from non-final layers. We compare DWAtt to a basic concatenation-based layer fusion method (Concat), and compare both to a deeper model baseline—all kept within a similar parameter budget. Our findings show that DWAtt and Concat are more step- and sample-efficient than the baseline, especially in the few-shot setting. DWAtt outperforms Concat on larger data sizes. On CoNLL-03 NER, layer fusion shows 3.68 − 9.73% F1 gain at different few-shot sizes. The layer fusion models presented significantly outperform the baseline in various training scenarios with different data sizes, architectures, and training constraints.

Continual learning is an emerging area of machine learning that deals with the issue where models adapt well to the latest data but lose the ability to remember past data due to changes in the data source. A widely adopted solution is by keeping a small memory of previous learned data that use replay. Most of the previous studies on continual learning focused on classification tasks, such as image classification and text classification, where the model needs only to categorize the input data. Inspired by the human ability to incrementally learn knowledge and solve different problems using learned knowledge, we considered a more pratical scenario, knowledge based quesiton answering about continual learning. In this scenario, each single question is different from others(means different fact trippes to answer them) while classification tasks only need to find feature boundaries of different categories, which are the curves or surfaces that separate different categories in the feature space. To address this issue, we proposed a depth aware hierarchical replay framework which include a tree structure classfier to have a sense of knowledge distribution and fill the gap between text classfication tasks and question-answering tasks for continual learning, a local sampler to grasp these critical samples and a depth aware learning network to reconstructe the feature space of a single learning round. In our experiments, we have demonstrated that our proposed model outperforms previous continual learning methods in mitigating the issue of catastrophic forgetting.

Purpose: Based on the examples of English and German, we investigate to what extent parsers trained on modern variants of these languages can be transferred to older language levels without loss. Methods: We developed a treebank called DoTT (https://github.com/texttechnologylab/DoTT) which covers, roughly, the time period from 1800 until today, in conjunction with the further development of the annotation tool DependencyAnnotator. DoTT consists of a collection of diachronic corpora enriched with dependency annotations using 3 parsers, 6 pre-trained language models, 5 newly trained models for German, and two tag sets (TIGER and Universal Dependencies). To assess how the different parsers perform on texts from different time periods, we created a gold standard sample as a benchmark. Results: We found that the parsers/models perform quite well on modern texts (document-level LAS ranging from 82.89 to 88.54) and slightly worse on older texts, as expected (average document-level LAS 84.60 vs. 86.14), but not significantly. For German texts, the (German) TIGER scheme achieved slightly better results than UD. Conclusion: Overall, this result speaks for the transferability of parsers to past language levels, at least dating back until around 1800. This very transferability, it is however argued, means that studies of language change in the field of dependency syntax can draw on dependency distance but miss out on some grammatical phenomena.

In table-text open-domain question answering, a retriever system retrieves relevant evidence from tables and text to answer questions. Previous studies in table-text open-domain question answering have two common challenges: firstly, their retrievers can be affected by false-positive labels in training datasets; secondly, they may struggle to provide appropriate evidence for questions that require reasoning across the table. To address these issues, we propose Denoised Table-Text Retriever (DoTTeR). Our approach involves utilizing a denoised training dataset with fewer false positive labels by discarding instances with lower question-relevance scores measured through a false positive detection model. Subsequently, we integrate table-level ranking information into the retriever to assist in finding evidence for questions that demand reasoning across the table. To encode this ranking information, we fine-tune a rank-aware column encoder to identify minimum and maximum values within a column. Experimental results demonstrate that DoTTeR significantly outperforms strong baselines on both retrieval recall and downstream QA tasks. Our code is available at https://github.com/deokhk/DoTTeR.

Noisily labeled textual data is ample on internet platforms that allow user-created content. Training models, such as offensive language detection models for comment moderation, on such data may prove difficult as the noise in the labels prevents the model to converge. In this work, we propose to use active learning methods for the purposes of denoising training data for model training. The goal is to sample examples the most informative examples with noisy labels with active learning and send them to the oracle for reannotation thus reducing the overall cost of reannotation. In this setting we tested three existing active learning methods, namely DBAL, Variance of Gradients (VoG) and BADGE. The proposed approach to data denoising is tested on the problem of offensive language detection. We observe that active learning can be effectively used for the purposes of data denoising, however care should be taken when choosing the algorithm for this purpose.

We tackle Event Argument Extraction (EAE) in the manner of template-based generation. Based on our exploration of generative EAE, it suffers from several issues, such as multiple arguments of one role, generating words out of context and inconsistency with prescribed format. We attribute it to the weakness of following complex input prompts. To address these problems, we propose the demonstration retrieval-augmented generative EAE (DRAGEAE), containing two components: event knowledge-injected generator (EKG) and demonstration retriever (DR). EKG employs event knowledge prompts to capture role dependencies and semantics. DR aims to search informative demonstrations from training data, facilitating the conditional generation of EKG. To train DR, we use the probability-based rankings from large language models (LLMs) as supervised signals. Experimental results on ACE-2005, RAMS and WIKIEVENTS demonstrate that our method outperforms all strong baselines and it can be generalized to various datasets. Further analysis is conducted to discuss the impact of diverse LLMs and prove that our model alleviates the above issues.

Deep Reinforcement learning (DRL) has been successfully applied to the dialogue policy of task-oriented dialogue systems. However, one challenge in the existing DRL-based dialogue policy methods is their unstructured state-action representations without the ability to learn the relationship between dialogue states and actions. To alleviate this problem, we propose a graph-structured dialogue policy framework for task-oriented dialogue systems. More specifically, we use an unsupervised approach to construct two different bipartite graphs. Then, we generate the user-related and knowledge-related subgraphs based on the matching dialogue sub-states with bipartite graph nodes. A variant of graph convolutional network is employed to encode dialogue subgraphs. After that, we use a bidirectional gated cycle unit (BGRU) and self-attention mechanism to obtain the high-level historical state representations and employ a neural network for the high-level current state representations. The two state representations are joined to learn the action value of dialogue policy. Experiments implemented with different DRL algorithms demonstrate that the proposed framework significantly improves the effectiveness and stability of dialogue policies.

Traditionally, approximate dynamic programming is employed in dialogue generation with greedy policy improvement through action sampling, as the natural language action space is vast. However, this practice is inefficient for reinforcement learning (RL) due to the sparsity of eligible responses with high action values, which leads to weak improvement sustained by random sampling. This paper presents theoretical analysis and experiments that reveal the performance of the dialogue policy is positively correlated with the sampling size. To overcome this limitation, we introduce a novel dual-granularity Q-function that explores the most promising response category to intervene in the sampling process. Our approach extracts actions based on a grained hierarchy, thereby achieving the optimum with fewer policy iterations. Additionally, we use offline RL and learn from multiple reward functions designed to capture emotional nuances in human interactions. Empirical studies demonstrate that our algorithm outperforms baselines across automatic metrics and human evaluations. Further testing reveals that our algorithm exhibits both explainability and controllability, as well as generates responses with higher expected rewards.

Food touches our lives through various endeavors, including flavor, nourishment, health, and sustainability. Recipes are cultural capsules transmitted across generations via unstructured text. Automated protocols for recognizing named entities, the building blocks of recipe text, are of immense value for various applications ranging from information extraction to novel recipe generation. Named entity recognition is a technique for extracting information from unstructured or semi-structured data with known labels. Starting with manually-annotated data of 6,611 ingredient phrases, we created an augmented dataset of 26,445 phrases cumulatively. Simultaneously, we systematically cleaned and analyzed ingredient phrases from RecipeDB, the gold-standard recipe data repository, and annotated them using the Stanford NER. Based on the analysis, we sampled a subset of 88,526 phrases using a clustering-based approach while preserving the diversity to create the machine-annotated dataset. A thorough investigation of NER approaches on these three datasets involving statistical, fine-tuning of deep learning-based language models and few-shot prompting on large language models (LLMs) provides deep insights. We conclude that few-shot prompting on LLMs has abysmal performance, whereas the fine-tuned spaCy-transformer emerges as the best model with macro-F1 scores of 95.9%, 96.04%, and 95.71% for the manually-annotated, augmented, and machine-annotated datasets, respectively.

The ability of large language models (LLMs) to “learn in context” based on the provided prompt has led to an explosive growth in their use, culminating in the proliferation of AI assistants such as ChatGPT, Claude, and Bard. These AI assistants are known to be robust to minor prompt modifications, mostly due to alignment techniques that use human feedback. In contrast, the underlying pre-trained LLMs they use as a backbone are known to be brittle in this respect. Building high-quality backbone models remains a core challenge, and a common approach to assessing their quality is to conduct few-shot evaluation. Such evaluation is notorious for being highly sensitive to minor prompt modifications, as well as the choice of specific in-context examples. Prior work has examined how modifying different elements of the prompt can affect model performance. However, these earlier studies tended to concentrate on a limited number of specific prompt attributes and often produced contradictory results. Additionally, previous research either focused on models with fewer than 15 billion parameters or exclusively examined black-box models like GPT-3 or PaLM, making replication challenging. In the present study, we decompose the entire prompt into four components: task description, demonstration inputs, labels, and inline instructions provided for each demonstration. We investigate the effects of structural and semantic corruptions of these elements on model performance. We study models ranging from 1.5B to 70B in size, using ten datasets covering classification and generation tasks. We find that repeating text within the prompt boosts model performance, and bigger models (≥30B) are more sensitive to the semantics of the prompt. Finally, we observe that adding task and inline instructions to the demonstrations enhances model performance even when the instructions are semantically corrupted. The code is available at this URL.

Multi-modal Named Entity Recognition, a fundamental task for multi-modal knowledge graph construction, requires integrating multi-modal information to extract named entities from text. Previous research has explored the integration of multi-modal representations at different granularities. However, they struggle to integrate all these multi-modal representations to provide rich contextual information to improve multi-modal named entity recognition. In this paper, we propose DPE-MNER, which is an iterative reasoning framework that dynamically incorporates all the diverse multi-modal representations following the strategy of “decompose, prioritize, and eliminate”. Within the framework, the fusion of diverse multi-modal representations is decomposed into hierarchically connected fusion layers that are easier to handle. The incorporation of multi-modal information prioritizes transitioning from “easy-to-hard” and “coarse-to-fine”. The explicit modeling of cross-modal relevance eliminate the irrelevances that will mislead the MNER prediction. Extensive experiments on two public datasets have demonstrated the effectiveness of our approach.

Large language models have demonstrated exceptional capability in natural language understanding and generation. However, their generation speed is limited by the inherently sequential nature of their decoding process, posing challenges for real-time applications. This paper introduces Lexical Unit Decoding (LUD), a novel decoding methodology implemented in a data-driven manner, accelerating the decoding process without sacrificing output quality. The core of our approach is the observation that a pre-trained language model can confidently predict multiple contiguous tokens, forming the basis for a lexical unit, in which these contiguous tokens could be decoded in parallel. Extensive experiments validate that our method substantially reduces decoding time while maintaining generation quality, i.e., 33% speed up on natural language generation with no quality loss, and 30% speed up on code generation with a negligible quality loss of 3%. Distinctively, LUD requires no auxiliary models and does not require changes to existing architectures. It can also be integrated with other decoding acceleration methods, thus achieving an even more pronounced inference efficiency boost. We posit that the foundational principles of LUD could define a new decoding paradigm for future language models, enhancing their applicability for a broader spectrum of applications. All codes are be publicly available at https://github.com/tjunlp-lab/Lexical-Unit-Decoding-LUD-.