COLING 2024 论文列表

Recent explorations with commercial Large Language Models (LLMs) have shown that non-expert users can jailbreak LLMs by simply manipulating their prompts; resulting in degenerate output behavior, privacy and security breaches, offensive outputs, and violations of content regulator policies. Limited studies have been conducted to formalize and analyze these attacks and their mitigations. We bridge this gap by proposing a formalism and a taxonomy of known (and possible) jailbreaks. We survey existing jailbreak methods and their effectiveness on open-source and commercial LLMs (such as GPT-based models, OPT, BLOOM, and FLAN-T5-XXL). We further discuss the challenges of jailbreak detection in terms of their effectiveness against known attacks. For further analysis, we release a dataset of model outputs across 3700 jailbreak prompts over 4 tasks.

Training large language models (LLMs) with open-domain instruction data has yielded remarkable success in aligning to end tasks and human preferences. Extensive research has highlighted the importance of the quality and diversity of instruction data. However, the impact of data complexity, as a crucial metric, remains relatively unexplored from three aspects: (1)where the sustainability of performance improvements with increasing complexity is uncertain; (2)whether the improvement brought by complexity merely comes from introducing more training tokens; and (3)where the potential benefits of incorporating instructions from easy to difficult are not yet fully understood. In this paper, we propose Tree-Instruct to systematically enhance the instruction complexity in a controllable manner. By adding a specified number of nodes to instructions’ semantic trees, this approach not only yields new instruction data from the modified tree but also allows us to control the difficulty level of modified instructions. Our preliminary experiments reveal the following insights: (1)Increasing complexity consistently leads to sustained performance improvements of LLMs. (2)Under the same token budget, a few complex instructions outperform diverse yet simple instructions. (3)Curriculum instruction tuning might not yield the anticipated results; focusing on increasing complexity appears to be the key.

In this study, we address the linguistic challenges posed by Tajiki Persian, a distinct variant of the Persian language that utilizes the Cyrillic script due to historical “Russification”. This distinguishes it from other Persian dialects that adopt the Arabic script. Despite its profound linguistic and cultural significance, Tajiki Persian remains a low-resource language with scant digitized datasets for computational applications. To address this deficiency, we created a parallel corpus using Shahnameh, a seminal Persian epic poem. Employing optical character recognition, we extracted Tajiki Persian verses from primary sources and applied a heuristic method to align them with their Iranian Persian counterparts. We then trained and assessed transliteration models using two prominent sequence-to-sequence architectures: GRU with attention and transformer. Our results underscore the enhanced performance of our models, particularly in contrast to pre-trained large multilingual models like GPT-3.5, emphasizing the value of dedicated datasets in advancing computational approaches for underrepresented languages. With the publication of this work, we are disseminating, for the first time, a vast collection of Persian poetry spanning 1000 years, transcribed in Tajiki scripts for the benefit of the Tajiki-speaking communities. The dataset, along with the model’s code and checkpoints, is accessible at https://github.com/language-ml/Tajiki-Shahname, marking a significant contribution to computational linguistic resources for Tajiki Persian.

Semantic Role Labeling (SRL) is a task in natural language understanding where the goal is to extract semantic roles for a given sentence. English SRL has achieved state-of-the-art performance using Transformer techniques and supervised learning. However, this technique is not a viable choice for smaller languages like Swedish due to the limited amount of training data. In this paper, we present the first effort in building a Transformer-based SRL system for Swedish by exploring multilingual and cross-lingual transfer learning methods and leveraging the Swedish FrameNet resource. We demonstrate that multilingual transfer learning outperforms two different cross-lingual transfer models. We also found some differences between frames in FrameNet that can either hinder or enhance the model’s performance. The resulting end-to-end model is freely available and will be made accessible through Språkbanken Text’s research infrastructure.

Automated Essay Scoring (AES) systems are widely popular in the market as they constitute a cost-effective and time-effective option for grading systems. Nevertheless, many studies have demonstrated that the AES system fails to assign lower grades to irrelevant responses. Thus, detecting the off-topic response in automated essay scoring is crucial in practical tasks where candidates write unrelated text responses to the given task in the question. In this paper, we are proposing an unsupervised technique that jointly scores essays and detects off-topic essays. The proposed Automated Open Essay Scoring (AOES) model uses a novel topic regularization module (TRM), which can be attached on top of a transformer model, and is trained using a proposed hybrid loss function. After training, the AOES model is further used to calculate the Mahalanobis distance score for off-topic essay detection. Our proposed method outperforms the baseline we created and earlier conventional methods on two essay-scoring datasets in off-topic detection as well as on-topic scoring. Experimental evaluation results on different adversarial strategies also show how the suggested method is robust for detecting possible human-level perturbations.

Pre-trained language models have revolutionized the natural language understanding landscape, most notably BERT (Bidirectional Encoder Representations from Transformers). However, a significant challenge remains for low-resource languages, where limited data hinders the effective training of such models. This work presents a novel approach to bridge this gap by transferring BERT capabilities from high-resource to low-resource languages using vocabulary matching. We conduct experiments on the Silesian and Kashubian languages and demonstrate the effectiveness of our approach to improve the performance of BERT models even when the target language has minimal training data. Our results highlight the potential of the proposed technique to effectively train BERT models for low-resource languages, thus democratizing access to advanced language understanding models.

To efficiently train quality estimation of text simplification on a small-scale labeled corpus, we train sentence difficulty estimation prior to fine-tuning the pre-trained language models. Our proposed method improves the quality estimation of text simplification in the framework of transfer fine-tuning, in which pre-trained language models can improve the performance of the target task by additional training on the relevant task prior to fine-tuning. Since the labeled corpus for quality estimation of text simplification is small (600 sentence pairs), an efficient training method is desired. Therefore, we propose a training method for pseudo quality estimation that does not require labels for quality estimation. As a relevant task for quality estimation of text simplification, we train the estimation of sentence difficulty. This is a binary classification task that identifies which sentence is simpler using an existing parallel corpus for text simplification. Experimental results on quality estimation of English text simplification showed that not only the quality estimation performance on simplicity that was trained, but also the quality estimation performance on fluency and meaning preservation could be improved in some cases.

This paper presents a translation-based knowledge geraph embedding method via efficient relation rotation (TransERR), a straightforward yet effective alternative to traditional translation-based knowledge graph embedding models. Different from the previous translation-based models, TransERR encodes knowledge graphs in the hypercomplex-valued space, thus enabling it to possess a higher degree of translation freedom in mining latent information between the head and tail entities. To further minimize the translation distance, TransERR adaptively rotates the head entity and the tail entity with their corresponding unit quaternions, which are learnable in model training. We also provide mathematical proofs to demonstrate the ability of TransERR in modeling various relation patterns, including symmetry, antisymmetry, inversion, composition, and subrelation patterns. The experiments on 10 benchmark datasets validate the effectiveness and the generalization of TransERR. The results also indicate that TransERR can better encode large-scale datasets with fewer parameters than the previous translation-based models. Our code and datasets are available at https://github.com/dellixx/TransERR.

Code pre-trained models (CodePTMs) have recently demonstrated a solid capacity to process various code intelligence tasks, e.g., code clone detection, code translation, and code summarization. The current mainstream method that deploys these models to downstream tasks is to fine-tune them on individual tasks, which is generally costly and needs sufficient data for large models. To tackle the issue, in this paper, we present TransCoder, a unified Transferable fine-tuning strategy for Code representation learning. Inspired by human inherent skills of knowledge generalization, TransCoder drives the model to learn better code-related knowledge like human programmers. Specifically, we employ a tunable prefix encoder to first capture cross-task and cross-language transferable knowledge, subsequently applying the acquired knowledge for optimized downstream adaptation. Besides, our approach confers benefits for tasks with minor training sample sizes and languages with smaller corpora, underscoring versatility and efficacy. Extensive experiments conducted on representative datasets clearly demonstrate that our method can lead to superior performance on various code-related tasks and encourage mutual reinforcement, especially in low-resource scenarios. Our codes are available at https://github.com/QiushiSun/TransCoder.

This paper describes how we train BERT models to carry over a coding system developed on the paragraphs of a Hungarian literary journal to another. The aim of the coding system is to track trends in the perception of literary translation around the political transformation in 1989 in Hungary. To evaluate not only task performance but also the consistence of the annotation, moreover, to get better predictions from an ensemble, we use 10-fold crossvalidation. Extensive hyperparameter tuning is used to obtain the best possible results and fair comparisons. To handle label imbalance, we use loss functions and metrics robust to it. Evaluation of the effect of domain shift is carried out by sampling a test set from the target domain. We establish the sample size by estimating the bootstrapped confidence interval via simulations. This way, we show that our models can carry over one annotation system to the target domain. Comparisons are drawn to provide insights such as learning multilabel correlations and confidence penalty improve resistance to domain shift, and domain adaptation on OCR-ed text on another domain improves performance almost to the same extent as that on the corpus under study. See our code at https://codeberg.org/zsamboki/bert-annotator-ensemble

Over the past two decades, the rise of new technologies and social networks has significantly shaped written language, imbuing it with characteristics akin to the spoken language. This study reports on the ongoing initiative to build the WhAP corpus, a resource featuring WhatsApp conversations in Italian, encompassing both written and spoken messages and totaling at present more than 400.000 tokens, 89 conversations, and 194 participants from diverse age groups and geographical regions of Italy. More specifically, this paper focuses on the practical steps involved in the construction of the resource. Once publicly accessible, the WhAP Corpus will enable in-depth linguistic research on the language used on WhatsApp, which shows unique features such as the blending of written and spoken elements.

Annotation and analysis of corrections in learner corpora have always presented technical challenges, mainly on account of the fact that until now there has not been any standard tool available, and that original and corrected versions of texts have been mostly stored together rather than treated as individual texts. In this paper, we present CJVT Svala 1.0, the Slovene version of the SVALA tool, which was originally used for the annotation of Swedish learner language. The localisation into Slovene resulted in the development of several new features in SVALA such as the support for multiple annotation systems, localisation into other languages, and the support for more complex annotation systems. Adopting the parallel aligned approach to text visualisation and annotation, as well as storing the data, combined with the tool supporting this, i.e. SVALA, are proposed as new standards in Learner Corpus Research.

Conversational Tree Search (Väth et al., 2023) is a recent approach to controllable dialog systems, where domain experts shape the behavior of a Reinforcement Learning agent through a dialog tree. The agent learns to efficiently navigate this tree, while adapting to information needs, e.g., domain familiarity, of different users. However, the need for additional training data hinders deployment in new domains. To address this, we explore approaches to generate this data directly from dialog trees. We improve the original approach, and show that agents trained on synthetic data can achieve comparable dialog success to models trained on human data, both when using a commercial Large Language Model for generation, or when using a smaller open-source model, running on a single GPU. We further demonstrate the scalability of our approach by collecting and testing on two new datasets: ONBOARD, a new domain helping foreign residents moving to a new city, and the medical domain DIAGNOSE, a subset of Wikipedia articles related to scalp and head symptoms. Finally, we perform human testing, where no statistically significant differences were found in either objective or subjective measures between models trained on human and generated data.

This paper analyzes multiple deep-syntactic frameworks with the goal of creating a proposal for a set of universal semantic role labels. The proposal examines various theoretic linguistic perspectives and focuses on Meaning-Text Theory and Functional Generative Description frameworks and PropBank. The research is based on the data from four Indo-European and one Uralic language – Spanish and Catalan (Taulé et al., 2011), Czech (Hajič et al., 2017), English (Hajič et al., 2012), and Finnish (Haverinen et al., 2015). Updated datasets with the new universal semantic role labels are now publicly available as a result of our work. Nevertheless, our proposal is oriented towards Universal Dependencies (UD) (de Marneffe et al., 2021) and our ultimate goal is to apply a subset of the universal labels to the full UD data.

As deep neural models in NLP become more complex, and as a consequence opaque, the necessity to interpret them becomes greater. A burgeoning interest has emerged in rationalizing explanations to provide short and coherent justifications for predictions. In this position paper, we advocate for a formal framework for key concepts and properties about rationalizing explanations to support their evaluation systematically. We also outline one such formal framework, tailored to rationalizing explanations of increasingly complex structures, from free-form explanations to deductive explanations, to argumentative explanations (with the richest structure). Focusing on the automated fact verification task, we provide illustrations of the use and usefulness of our formalization for evaluating explanations, tailored to their varying structures.

We present the first version of a semantic reasoning benchmark for Danish compiled semi-automatically from a number of human-curated lexical-semantic resources, which function as our gold standard. Taken together, the datasets constitute a benchmark for assessing selected language understanding capacities of large language models (LLMs) for Danish. This first version comprises 25 datasets across 6 different tasks and include 3,800 test instances. Although still somewhat limited in size, we go beyond comparative evaluation datasets for Danish by including both negative and contrastive examples as well as low-frequent vocabulary; aspects which tend to challenge current LLMs when based substantially on language transfer. The datasets focus on features such as semantic inference and entailment, similarity, relatedness, and ability to disambiguate words in context. We use ChatGPT to assess to which degree our datasets challenge the ceiling performance of state-of-the-art LLMs, average performance being relatively high with an average accuracy of 0.6 on ChatGPT 3.5 turbo and 0.8 on ChatGPT 4.0.

This paper presents the rationale for a “dedicated” corpus of spoken Maltese, Korpus tal-Malti Mitkellem, KMM, ‘Corpus of Spoken Maltese’, based on the concept of a gold-standard Core collection. The Core collection is designed to cater to as wide a variety of user needs as possible whilst respecting basic principles governing corpus design, such as representativeness and balance, and delivering high quality in terms of both audio quality and annotations. An overview is provided of the composition of the current Core corpus of around 20 hours of data and of the human annotation effort involved. We also carry out a small qualitative analysis of the output of a Maltese ASR system and compare it to the human annotators’ output. Initial results are promising, showing that the ASR is robust enough to generate first-pass texts for annotators to work on, thus reducing the human effort, and consequently, the cost involved.

This paper presents a brief description of some morphosyntactic features of Ancash Quechua, the majority variety of the Central Quechua language family (QI), for the purpose of building a corpus annotated according to the Universal Dependencies (UD) schema. The creation of such a corpus has two objectives: for Quechua linguistics, it opens up the possibility of more systematic linguistic studies and comparisons with other languages. It also enables the development of a syntactic parser, which would be the first NLP tool for a Quechua language of this family. For the UD project, adding Quechua, an agglutinative language with a rich morphology, makes it possible to point out some possible shortcomings of the universal annotation schema, and to fuel the discussion to adapt this schema to the specific features of the languages with a similar typology. The first step towards this work was first to gather and digitise the available linguistic resources, thus creating the first bilingual and sentence-aligned digital corpus in Ancash Quechua and Spanish. After identifying some linguistic features not fully described in the UD schema, we proposed annotation solutions, and built an initial corpus of around twenty sentences, which we are making freely available.

According to the principle of compositional generalization, the meaning of a complex expression can be understood as a function of the meaning of its parts and of how they are combined. This principle is crucial for human language processing and also, arguably, for NLP models in the face of out-of-distribution data. However, many neural network models, including Transformers, have been shown to struggle with compositional generalization. In this paper, we hypothesize that forcing models to in-context learn can provide an inductive bias to promote compositional generalization. To test this hypothesis, we train a causal Transformer in a setting that renders ‘ordinary’ learning very difficult: we present it with different orderings of the training instance and shuffle instance labels. This corresponds to training the model on all possible few-shot learning problems attainable from the dataset. The model can solve the task, however, by utilizing earlier examples to generalize to later ones – i.e., in-context learning. In evaluations on the datasets, SCAN, COGS, and GeoQuery, models trained in this manner indeed show improved compositional generalization. This indicates the usefulness of in-context learning problems as an inductive bias for generalization.