NAACL 2024 Papers

Existing grammatical error correction tools do not provide natural language explanations of the errors that they correct in user-written text. However, such explanations are essential for helping users learn the language by gaining a deeper understanding of its grammatical rules (DeKeyser, 2003; Ellis et al., 2006).To address this gap, we propose the task of grammar error explanation, where a system needs to provide one-sentence explanations for each grammatical error in a pair of erroneous and corrected sentences. The task is not easily solved by prompting LLMs: we find that, using one-shot prompting, GPT-4 only explains 40.6% of the errors and does not even attempt to explain 39.8% of the errors.Since LLMs struggle to identify grammar errors, we develop a two-step pipeline that leverages fine-tuned and prompted large language models to perform structured atomic token edit extraction, followed by prompting GPT-4 to explain each edit. We evaluate our pipeline on German, Chinese, and English grammar error correction data. Our atomic edit extraction achieves an F1 of 0.93 on German, 0.91 on Chinese, and 0.891 on English. Human evaluation of generated explanations reveals that 93.9% of German errors, 96.4% of Chinese errors, and 92.20% of English errors are correctly detected and explained. To encourage further research, we open-source our data and code.

Predicting unseen relations that cannot be observed during the training phase is a challenging task in relation extraction. Previous works have made progress by matching the semantics between input instances and label descriptions. However, fine-grained matching often requires laborious manual annotation, and rich interactions between instances and label descriptions come with significant computational overhead. In this work, we propose an efficient multi-grained matching approach that uses virtual entity matching to reduce manual annotation cost, and fuses coarse-grained recall and fine-grained classification for rich interactions with guaranteed inference speed.Experimental results show that our approach outperforms the previous State Of The Art (SOTA) methods, and achieves a balance between inference efficiency and prediction accuracy in zero-shot relation extraction tasks.Our code is available at https://github.com/longls777/EMMA.

ChatGPT has recently emerged as a powerful NLP tool that can carry out a variety of tasks. However, the range of languages ChatGPT can handle remains largely a mystery. To uncover which languages ChatGPT ‘knows’, we investigate its language identification (LID) abilities. For this purpose, we compile Babel-670, a benchmark comprising 670 languages representing 23 language families spoken in five continents. Languages in Babel-670 run the gamut from the very high-resource to the very low-resource. We then study ChatGPT’s (both GPT-3.5 and GPT-4) ability to (i) identify language names and language codes (ii) under zero- and few-shot conditions (iii) with and without provision of a label set. When compared to smaller finetuned LID tools, we find that ChatGPT lags behind. For example, it has poor performance on African languages. We conclude that current large language models would benefit from further development before they can sufficiently serve diverse communities.

Language models are often at risk of diverse backdoor attacks, especially data poisoning. Thus, it is important to investigate defense solutions for addressing them. Existing backdoor defense methods mainly focus on backdoor attacks with explicit triggers, leaving a universal defense against various backdoor attacks with diverse triggers largely unexplored. In this paper, we propose an end-to-end ensemble-based backdoor defense framework, DPoE (Denoised Product-of-Experts), which is inspired by the shortcut nature of backdoor attacks, to defend various backdoor attacks. DPoE consists of two models: a shallow model that captures the backdoor shortcuts and a main model that is prevented from learning the shortcuts. To address the label flip caused by backdoor attackers, DPoE incorporates a denoising design. Experiments on three NLP tasks show that DPoE significantly improves the defense performance against various types of backdoor triggers including word-level, sentence-level, and syntactic triggers. Furthermore, DPoE is also effective under a more challenging but practical setting that mixes multiple types of triggers.

In the realm of Large Language Models (LLMs), the balance between instruction data quality and quantity is a focal point. Recognizing this, we introduce a self-guided methodology for LLMs to autonomously discern and select cherry samples from open-source datasets, effectively minimizing manual curation and potential cost for instruction tuning an LLM. Our key innovation, the Instruction-Following Difficulty (IFD) metric, emerges as a pivotal metric to identify discrepancies between a model’s expected responses and its intrinsic generation capability. Through the application of IFD, cherry samples can be pinpointed, leading to a marked uptick in model training efficiency. Empirical validations on datasets like Alpaca and WizardLM underpin our findings; with a mere 10% of original data input, our strategy showcases improved results. This synthesis of self-guided cherry-picking and the IFD metric signifies a transformative leap in the instruction tuning of LLMs, promising both efficiency and resource-conscious advancements. Codes, data, and models are available.

Large Language Models (LLMs) have achieved remarkable success, where instruction tuning is the critical step in aligning LLMs with user intentions. In this work, we investigate how the instruction tuning adjusts pre-trained models with a focus on intrinsic changes. Specifically, we first develop several local and global explanation methods, including a gradient-based method for input-output attribution, and techniques for interpreting patterns and concepts in self-attention and feed-forward layers. The impact of instruction tuning is then studied by comparing the explanations derived from the pre-trained and instruction-tuned models. This approach provides an internal perspective of the model shifts on a human-comprehensible level. Our findings reveal three significant impacts of instruction tuning: 1) It empowers LLMs to recognize the instruction parts of user prompts, and promotes the response generation constantly conditioned on the instructions. 2) It encourages the self-attention heads to capture more word-word relationships about instruction verbs. 3) It encourages the feed-forward networks to rotate their pre-trained knowledge toward user-oriented tasks. These insights contribute to a more comprehensive understanding of instruction tuning and lay the groundwork for future work that aims at explaining and optimizing LLMs for various applications. Our code and data are publicly available at https://github.com/JacksonWuxs/Interpret_Instruction_Tuning_LLMs.

Large language models (LLMs) exhibit positional bias in how they use context, which especially affects listwise ranking. To address this, we propose permutation self-consistency, a form of self-consistency over the ranking list outputs of black-box LLMs. Our key idea is to marginalize out different list orders in the prompt to produce an order-independent ranking with less positional bias. First, given some input prompt, we repeatedly shuffle the list in the prompt and pass it through the LLM while holding the instructions the same. Next, we aggregate the resulting sample of rankings by computing the central ranking closest in distance to all of them, marginalizing out prompt order biases in the process. Theoretically, we prove the robustness of our method, showing convergence to the true ranking under random perturbations.Empirically, on five datasets in sorting and passage reranking, our approach improves scores from conventional inference by up to 34-52% for Mistral, 7-18% for GPT-3.5, 8-16% for LLaMA v2 (70B). Our code is at https://github.com/castorini/perm-sc.

The widespread adoption of large language models (LLMs) across various regions underscores the urgent need to evaluate their alignment with human values. Current benchmarks, however, fall short of effectively uncovering safety vulnerabilities in LLMs. Despite numerous models achieving high scores and ‘topping the chart’ in these evaluations, there is still a significant gap in LLMs’ deeper alignment with human values and achieving genuine harmlessness. To this end, this paper proposes a value alignment benchmark named Flames, which encompasses both common harmlessness principles and a unique morality dimension that integrates specific Chinese values such as harmony. Accordingly, we carefully design adversarial prompts that incorporate complex scenarios and jailbreaking methods, mostly with implicit malice. By prompting 17 mainstream LLMs, we obtain model responses and rigorously annotate them for detailed evaluation. Our findings indicate that all the evaluated LLMs demonstrate relatively poor performance on Flames, particularly in the safety and fairness dimensions. We also develop a lightweight specified scorer capable of scoring LLMs across multiple dimensions to efficiently evaluate new models on the benchmark. The complexity of Flames has far exceeded existing benchmarks, setting a new challenge for contemporary LLMs and highlighting the need for further alignment of LLMs. Our benchmark is publicly available at https://github.com/AIFlames/Flames.

In this paper we study the fine-tuning of pre-trained large high-resource language models (LLMs) into many-to-one multilingual machine translators for extremely-low-resource languages such as endangered Indigenous languages. We explore those issues using datasets created from pseudo-parallel translations to English of The Bible written in 39 Brazilian Indigenous languages using mBART50 and WMT19 as pre-trained models and multiple translation metrics. We examine bilingual and multilingual models and show that, according to machine translation metrics, same-linguistic family models tend to perform best. However, we also found that many-to-one multilingual systems have a tendency to learn a “rogue” strategy of storing output strings from the training data in the LLM structure and retrieving them instead of performing actual translations. We show that rephrasing the output of the training samples seems to solve the problem.

Many NLP researchers are experiencing an existential crisis triggered by the astonishing success of ChatGPT and other systems based on large language models (LLMs). After such a disruptive change to our understanding of the field, what is left to do? Taking a historical lens, we look for guidance from the first era of LLMs, which began in 2005 with large n-gram models for machine translation (MT). We identify durable lessons from the first era, and more importantly, we identify evergreen problems where NLP researchers can continue to make meaningful contributions in areas where LLMs are ascendant. We argue that disparities in scale are transient and researchers can work to reduce them; that data, rather than hardware, is still a bottleneck for many applications; that meaningful realistic evaluation is still an open problem; and that there is still room for speculative approaches.

In machine translation, the problem of ambiguously gendered input has been pointed out, where the gender of an entity is not available in the source sentence. To address this ambiguity issue, the task of controlled translation that takes the gender of the ambiguous entity as additional input have been proposed. However, most existing works have only considered a simplified setup of one target gender for input. In this paper, we tackle controlled translation in a more realistic setting of inputs with multiple entities and propose Gender-of-Entity (GoE) prompting method for LLMs. Our proposed method instructs the model with fine-grained entity-level gender information to translate with correct gender inflections. By utilizing four evaluation benchmarks, we investigate the controlled translation capability of LLMs in multiple dimensions and find that LLMs reach state-of-the-art performance in controlled translation. Furthermore, we discover an emergence of gender interference phenomenon when controlling the gender of multiple entities. Finally, we address the limitations of existing gender accuracy evaluation metrics and propose leveraging LLMs as an evaluator for gender inflection in machine translation.

Reinforcement learning from human feedback (RLHF) has emerged as an effective approach to aligning large language models (LLMs) to human preferences.RLHF contains three steps, i.e., human preference collecting, reward learning, and policy optimization, which are usually performed serially.Despite its popularity, however, (fixed) reward models may suffer from inaccurate off-distribution, since policy optimization continuously shifts LLMs’ data distribution.Repeatedly collecting new preference data from the latest LLMs may alleviate this issue, which unfortunately makes the resulting system more complicated and difficult to optimize.In this paper, we propose reward learning on policy (RLP), an unsupervised framework that refines a reward model using policy samples to keep it on-distribution.Specifically, an unsupervised multi-view learning method is introduced to learn robust representations of policy samples.Meanwhile, a synthetic preference generation approach is developed to simulate high-quality preference data with policy outputs.Extensive experiments on three benchmark datasets show that RLP consistently outperforms the state-of-the-art.Our code is available at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/rlp.

Reliable human evaluation is critical to the development of successful natural language generation models, but achieving it is notoriously difficult. Stability is a crucial requirement when ranking systems by quality: consistent ranking of systems across repeated evaluations is not just desirable, but essential. Without it, there is no reliable foundation for hill-climbing or product launch decisions. In this paper, we use machine translation and its state-of-the-art human evaluation framework, MQM, as a case study to understand how to set up reliable human evaluations that yield stable conclusions. We investigate the optimal configurations for item allocation to raters, number of ratings per item, and score normalization. Our study on two language pairs provides concrete recommendations for designing replicable human evaluation studies. We also collect and release the largest publicly available dataset of multi-segment translations rated by multiple professional translators, consisting of nearly 140,000 segment annotations across two language pairs.

In the financial industry, identifying the location of parties involved in payments is a major challenge in the context of Anti-Money Laundering transaction monitoring. For this purpose address parsing entails extracting fields such as street, postal code, or country from free text message attributes. While payment processing platforms are updating their standards with more structured formats such as SWIFT with ISO 20022, address parsing remains essential for a considerable volume of messages. With the emergence of Transformers and Generative Large Language Models (LLM), we explore the performance of state-of-the-art solutions given the constraint of processing a vast amount of daily data. This paper also aims to show the need for training robust models capable of dealing with real-world noisy transactional data. Our results suggest that a well fine-tuned Transformer model using early-stopping significantly outperforms other approaches. Nevertheless, generative LLMs demonstrate strong zero_shot performance and warrant further investigations.

Few-shot Knowledge Graph (KG) Relational Reasoning aims to predict unseen triplets (i.e., query triplets) for rare relations in KGs, given only several triplets of these relations as references (i.e., support triplets). This task has gained significant traction due to the widespread use of knowledge graphs in various natural language processing applications. Previous approaches have utilized meta-training methods and manually constructed meta-relation sets to tackle this task. Recent efforts have focused on edge-mask-based methods, which exploit the structure of the contextualized graphs of target triplets (i.e., a subgraph containing relevant triplets in the KG). However, existing edge-mask-based methods have limitations in extracting insufficient information from KG and are highly influenced by spurious information in KG. To overcome these challenges, we propose SAFER (Subgraph Adaptation for Few-shot Relational Reasoning), a novel approach that effectively adapts the information in contextualized graphs to various subgraphs generated from support and query triplets to perform the prediction. Specifically, SAFER enables the extraction of more comprehensive information from support triplets while minimizing the impact of spurious information when predicting query triplets. Experimental results on three prevalent datasets demonstrate the superiority of our proposed framework SAFER.

Scientific texts are distinctive from ordinary texts in quite a few aspects like their vocabulary and discourse structure. Consequently, Information Extraction (IE) tasks for scientific texts come with their own set of challenges. The classical definition of Named Entities restricts the inclusion of all scientific terms under its hood, which is why previous works have used the terms Named Entities and Keyphrases interchangeably. We suggest the rechristening of Named Entities for the scientific domain as Typed Keyphrases (TK), broadening their scope. We advocate for exploring this task in the few-shot domain due to the scarcity of labeled scientific IE data. Currently, no dataset exists for few-shot scientific Typed Keyphrase Recognition. To address this gap, we develop an annotation schema and present Few-TK, a dataset in the AI/ML field that includes scientific Typed Keyphrase annotations on abstracts of 500 research papers. To the best of our knowledge, this is the introductory few-shot Typed Keyphrase recognition dataset and only the second dataset structured specifically for few-shot NER, after Few-NERD. We report the results of several few-shot sequence-labelling models applied to our dataset. The data and code are available at https://github.com/AvishekLahiri/Few_TK.git

Federated Multilingual Modeling (FMM) plays a crucial role in the applications of natural language processing due to the increasing diversity of languages and the growing demand for data privacy. However, FMM faces limitations stemming from (1) the substantial communication costs in networking and (2) the conflicts arising from parameter interference between different languages. To address these challenges, we introduce a communication-efficient federated learning framework with low-rank adaptation and language family clustering for Multilingual Modeling (MM). In this framework, we maintain the weights of the base model, exclusively updating the lightweight Low-rank adaptation (LoRA) parameters to minimize communication costs. Additionally, we mitigate parameter conflicts by grouping languages based on their language family affiliations, as opposed to aggregating all LoRA parameters. Experiments demonstrate that our proposed model not only surpasses the baseline models in performance but also reduces the communication overhead. Our code is available at https://github.com/zhihan-guo/FedLFC.

We present FastFit, a Python package designed to provide fast and accurate few-shot classification, especially for scenarios with many semantically similar classes. FastFit utilizes a novel approach integrating batch contrastive learning and token-level similarity score. Compared to existing few-shot learning packages, such as SetFit, Transformers, or few-shot prompting of large language models via API calls, FastFit significantly improves multi-class classification performance in speed and accuracy across various English and Multilingual datasets. FastFit demonstrates a 3-20x improvement in training speed, completing training in just a few seconds. The FastFit package is now available on GitHub, presenting a user-friendly solution for NLP practitioners.

The growing awareness of safety concerns in large language models (LLMs) has sparked considerable interest in the evaluation of safety. This study investigates an under-explored issue about the evaluation of LLMs, namely the substantial discrepancy in performance between multiple-choice questions and open-ended questions. Inspired by research on jailbreak attack patterns, we argue this is caused by mismatched generalization. That is, LLM only remembers the answer style for open-ended safety questions, which makes it unable to solve other forms of safety tests. We refer to this phenomenon as fake alignment and construct a comparative benchmark to empirically verify its existence in LLMs. We introduce a Fake alIgNment Evaluation (FINE) framework and two novel metrics——Consistency Score (CS) and Consistent Safety Score (CSS), which jointly assess two complementary forms of evaluation to quantify fake alignment and obtain corrected performance estimation. Applying FINE to 14 widely-used LLMs reveals several models with purported safety are poorly aligned in practice. Subsequently, we found that multiple-choice format data can also be used as high-quality contrast distillation-based fine-tuning data, which can strongly improve the alignment consistency of LLMs with minimal fine-tuning overhead. For data and code, see https://github.com/AIFlames/Fake-Alignment.