Adapting large language models (LLMs) to new languages typically involves continual pre-training (CT) followed by supervised fine-tuning (SFT). However, this CT-then-SFT approach struggles with limited data in the context of low-resource languages, failing to balance language modeling and task-solving capabilities. We thus propose a new model merging solution as an alternative for low-resource languages, combining models with distinct capabilities into a single model without additional training. We use model merging to develop task-solving LLMs for low-resource languages without SFT data in the target languages. Our experiments based on Llama-2-7B demonstrate that model merging effectively endows LLMs for low-resource languages with task-solving abilities, outperforming CT-then-SFT in scenarios with extremely scarce data. Observing performance saturation in model merging with increasingly more training tokens, we further analyze the merging process and introduce a slack variable to the model merging algorithm to mitigate the loss of important parameters, thereby enhancing model performance. We hope that model merging can benefit more human languages suffering from data scarcity with its higher data efficiency.

Current large language models demonstrate deficiencies in understanding low-resource languages, particularly the minority languages in China. This limitation stems from the scarcity of available pre-training data. To address this accessibility challenge, we present MC², a Multilingual Corpus of Minority Languages in China, which is the largest open-source corpus of its kind so far. MC² includes four underrepresented languages: Tibetan, Uyghur, Kazakh, and Mongolian. Notably, we focus on the less common writing systems of Kazakh and Mongolian, i.e., Kazakh Arabic script and traditional Mongolian script, respectively, which have been long neglected in previous corpus construction efforts. Recognizing the prevalence of language contamination within existing corpora, we adopt a quality-centric solution for collecting MC², prioritizing accuracy while enhancing diversity. Furthermore, we underscore the importance of attending to the multiplicity of writing systems, which is closely related to the cultural awareness of the resulting models. The MC² corpus and related models are made public to the community.

In this paper, we introduce a novel dynamic expert selection framework for Mixture of Experts (MoE) models, aiming to enhance computational efficiency and model performance by adjusting the number of activated experts based on input difficulty. Unlike existing MoE approaches that rely on fixed TopK Routing, which activates a predetermined number of experts regardless of the input’s complexity, our method dynamically allocates experts based on the confidence level in expert selection for each input. This allows for more efficient utilization of computational resources, activating more experts for complex tasks requiring advanced reasoning and fewer for simpler tasks. Through extensive evaluations, our dynamic routing method demonstrates substantial improvements over Top2 Routing across various benchmarks, achieving an average improvement of 0.7% with less than 90% activated parameters. Further analysis shows our model dispatches more experts to tasks requiring complex reasoning skills, like BBH, confirming its ability to dynamically allocate computational resources in alignment with the input’s complexity.Our findings also highlight a variation in the number of experts needed across different layers of the transformer model, offering insights into the potential for designing heterogeneous MoE frameworks. The code and models are available at https://github.com/ZhenweiAn/Dynamic_MoE.

The advancement of Multimodal Large Language Models (MLLMs) has greatly accelerated the development of applications in understanding integrated texts and images. Recent works leverage image-caption datasets to train MLLMs, achieving state-of-the-art performance on image-to-text tasks. However, there are few studies exploring which layers of MLLMs make the most effort to the global image information, which plays vital roles in multimodal comprehension and generation. In this study, we find that the intermediate layers of models can encode more global semantic information, whose representation vectors perform better on visual-language entailment tasks, rather than the topmost layers. We further probe models regarding local semantic representations through object recognition tasks. We find that the topmost layers may excessively focus on local information, leading to a diminished ability to encode global information. Our code and data are released via https://github.com/kobayashikanna01/probing_MLLM_rep.

Event temporal relation extraction (ETRE) is usually formulated as a multi-label classification task, where each type of relation is simply treated as a one-hot label. This formulation ignores the meaning of relations and wipes out their intrinsic dependency. After examining the relation definitions in various ETRE tasks, we observe that all relations can be interpreted using the start and end time points of events. For example, relation Includes could be interpreted as event 1 starting no later than event 2 and ending no earlier than event 2. In this paper, we propose a unified event temporal relation extraction framework, which transforms temporal relations into logical expressions of time points and completes the ETRE by predicting the relations between certain time point pairs. Experiments on TB-Dense and MATRES show significant improvements over a strong baseline and outperform the state-of-the-art model by 0.3% on both datasets. By representing all relations in a unified framework, we can leverage the relations with sufficient data to assist the learning of other relations, thus achieving stable improvement in low-data scenarios. When the relation definitions are changed, our method can quickly adapt to the new ones by simply modifying the logic expressions that map time points to new event relations. The code is released at https://github.com/AndrewZhe/A-Unified-Framework-for-ETRE

Pre-trained language models greatly improve the performance of various tasks but at a cost of high computation overhead. To facilitate practical applications, there are mainly two lines of research to accelerate model inference: model compression and dynamic computation (e.g., dynamic token pruning). Existing works either adopt these methods individually or simply apply dynamic computation approaches upon a compressed small language model. We argue that they are sub-optimal since the two approaches are separately designed so the compressed model may not be tailored for dynamic computation. To tackle this problem and make compressed small language models faster, we propose Length-Adaptive Distillation, a two-stage knowledge distillation framework that aims to produce a customized small language model for dynamic token pruning. In the general distillation stage, we enforce the student to mimic and reconstruct the teacher’s output based on the dynamically pruned representations. Then in the task-specific distillation stage, the student is further accustomed to token pruning while absorbing the task-specific knowledge. Experimental results on GLUE benchmark demonstrate that our method can make the small language model more customized for dynamic token pruning and achieve better speed-performance trade-off.

Document-level Event Argument Extraction (EAE) requires the model to extract arguments of multiple events from a single document. Considering the underlying dependencies between these events, recent efforts leverage the idea of “memory”, where the results of already predicted events are cached and can be retrieved to help the prediction of upcoming events. These methods extract events according to their appearance order in the document, however, the event that appears in the first sentence does not mean that it is the easiest to extract. Existing methods might introduce noise to the extraction of upcoming events if they rely on an incorrect prediction of previous events. In order to provide more reliable memory, we propose a simple-to-complex progressive framework for document-level EAE. Specifically, we first calculate the difficulty of each event and then, we conduct the extraction following a simple-to-complex order. In this way, the memory will store the most certain results, and the model could use these reliable sources to help the prediction of more difficult events. Experiments on WikiEvents show that our model outperforms SOTA by 1.4% in F1, indicating the proposed simple-to-complex framework is useful in the EAE task.

Multi-hop Knowledge Base Question Answering(KBQA) aims to find the answer entity in a knowledge graph (KG), which requires multiple steps of reasoning. Existing retrieval-based approaches solve this task by concentrating on the specific relation at different hops and predicting the intermediate entity within the reasoning path. However, these models fail to utilize information from head-tail entities and the semantic connection between relations to enhance the current relation representation, which undermines the information capturing of relations in KGs. To address this issue, we construct a dual relation graph where each node denotes a relation in the original KG (primal entity graph) and edges are constructed between relations sharing same head or tail entities. Then we iteratively do primal entity graph reasoning, dual relation graph information propagation, and interaction between these two graphs. In this way, the interaction between entity and relation is enhanced, and we derive better entity and relation representations. Experiments on two public datasets, WebQSP and CWQ, show that our approach achieves a significant performance gain over the prior state-of-the-art.

DocRED is a widely used dataset for document-level relation extraction. In the large-scale annotation, a recommend-revise scheme is adopted to reduce the workload. Within this scheme, annotators are provided with candidate relation instances from distant supervision, and they then manually supplement and remove relational facts based on the recommendations. However, when comparing DocRED with a subset relabeled from scratch, we find that this scheme results in a considerable amount of false negative samples and an obvious bias towards popular entities and relations. Furthermore, we observe that the models trained on DocRED have low recall on our relabeled dataset and inherit the same bias in the training data. Through the analysis of annotators’ behaviors, we figure out the underlying reason for the problems above: the scheme actually discourages annotators from supplementing adequate instances in the revision phase. We appeal to future research to take into consideration the issues with the recommend-revise scheme when designing new models and annotation schemes. The relabeled dataset is released at https://github.com/AndrewZhe/Revisit-DocRED, to serve as a more reliable test set of document RE models.

Knowledge distillation has been proven effective when customizing small language models for specific tasks. Here, a corpus as ‘textbook’ plays an indispensable role, only through which the teacher can teach the student. Prevailing methods adopt a two-stage distillation paradigm: general distillation first with task-agnostic general corpus and task-specific distillation next with augmented task-specific corpus. We argue that such a paradigm may not be optimal. In general distillation, it’s extravagant to let the diverse but desultory general knowledge overwhelms the limited model capacity of the student. While in task-specific distillation, the task corpus is usually limited and narrow, preventing the student from learning enough knowledge. To mitigate the issues in the two gapped corpora, we present a better textbook for the student to learn: contextualized corpus that contextualizes task corpus with large-scale general corpus through relevance-based text retrieval. Experimental results on GLUE benchmark demonstrate that contextualized corpus is the better textbook compared with jointly using general corpus and augmented task-specific corpus. Surprisingly, it enables task-specific distillation from scratch without general distillation while maintaining comparable performance, making it more flexible to customize the student model with desired model size under various computation constraints.

The charge prediction task aims to predict the charge for a case given its fact description. Recent models have already achieved impressive accuracy in this task, however, little is understood about the mechanisms they use to perform the judgment.For practical applications, a charge prediction model should conform to the certain legal theory in civil law countries, as under the framework of civil law, all cases are judged according to certain local legal theories. In China, for example, nearly all criminal judges make decisions based on the Four Elements Theory (FET).In this paper, we argue that trustworthy charge prediction models should take legal theories into consideration, and standing on prior studies in model interpretation, we propose three principles for trustworthy models should follow in this task, which are sensitive, selective, and presumption of innocence.We further design a new framework to evaluate whether existing charge prediction models learn legal theories. Our findings indicate that, while existing charge prediction models meet the selective principle on a benchmark dataset, most of them are still not sensitive enough and do not satisfy the presumption of innocence. Our code and dataset are released at https://github.com/ZhenweiAn/EXP_LJP.

Recent studies strive to incorporate various human rationales into neural networks to improve model performance, but few pay attention to the quality of the rationales. Most existing methods distribute their models’ focus to distantly-labeled rationale words entirely and equally, while ignoring the potential important non-rationale words and not distinguishing the importance of different rationale words. In this paper, we propose two novel auxiliary loss functions to make better use of distantly-labeled rationales, which encourage models to maintain their focus on important words beyond labeled rationales (PINs) and alleviate redundant training on non-helpful rationales (NoIRs). Experiments on two representative classification tasks show that our proposed methods can push a classification model to effectively learn crucial clues from non-perfect rationales while maintaining the ability to spread its focus to other unlabeled important words, thus significantly outperform existing methods.

Document-level Relation Extraction (RE) is a more challenging task than sentence RE as it often requires reasoning over multiple sentences. Yet, human annotators usually use a small number of sentences to identify the relationship between a given entity pair. In this paper, we present an embarrassingly simple but effective method to heuristically select evidence sentences for document-level RE, which can be easily combined with BiLSTM to achieve good performance on benchmark datasets, even better than fancy graph neural network based methods. We have released our code at https://github.com/AndrewZhe/Three-Sentences-Are-All-You-Need.

Code comments are vital for software maintenance and comprehension, but many software projects suffer from the lack of meaningful and up-to-date comments in practice. This paper presents a novel approach to automatically generate code comments at a function level by targeting object-oriented programming languages. Unlike prior work that only uses information locally available within the target function, our approach leverages broader contextual information by considering all other functions of the same class. To propagate and integrate information beyond the scope of the target function, we design a novel learning framework based on the bidirectional gated recurrent unit and a graph attention network with a pointer mechanism. We apply our approach to produce code comments for Java methods and compare it against four strong baseline methods. Experimental results show that our approach outperforms most methods by a large margin and achieves a comparable result with the state-of-the-art method.