We introduce DavIR, a model-based data selection method for post-training Large Language Models. DavIR generalizes Reducible Holdout Loss to core-set selection problem of causal language modeling, and quantifies the learnability of a given datum with respect to a pre-trained LLM based on relative reduction in loss during fine-tuning, a metric we show to be closely related to the implicit reward model described in Direct Preference Optimization (DPO). We show that 6% of Alpaca dataset selected with DavIR can steer both the LLaMA and Gemma model family to produce superior performance compared to the same models trained on the full 52K dataset. We also show that Alpaca dataset compressed with DavIR can be combined with GSM8K dataset to effectively balance open-domain freeform QA and mathematical reasoning capabilities. Finally, we apply the DavIR objective to DPO and develop a normalized DavIR-DPO objective which improves alignment performance of Zephyr-7B-SFT model by 8% (relative) on AlpacaEval, compared against training on vanilla DPO objective.

Fine-tuning pre-trained Large Language Models (LLMs) for specialized tasks incurs substantial computational and data costs. While model merging offers a training-free solution to integrate multiple task-specific models, existing methods suffer from safety-utility conflicts where enhanced general capabilities degrade safety safeguards. We identify two root causes: neuron misidentification due to simplistic parameter magnitude-based selection, and cross-task neuron interference during merging.To address these challenges, we propose LED-Merging, a three-stage framework that Locates task-specific neurons via gradient-based attribution, dynamically Elects critical neurons through multi-model importance fusion, and Disjoints conflicting updates through parameter isolation.Extensive experiments on Llama-3-8B, Mistral-7B, and Llama2-13B demonstrate that LED-Merging effectively reduces harmful response rates, showing a 31.4% decrease on Llama-3-8B-Instruct on HarmBench, while simultaneously preserving 95% of utility performance, such as achieving 52.39% accuracy on GSM8K.LED-Merging resolves safety-utility conflicts and provides a lightweight, training-free paradigm for constructing reliable multi-task LLMs.Code is available at https://github.com/MqLeet/LED-Merging

Hierarchical Text Classification (HTC) is a challenging task which aims to extract the labels in a tree structure corresponding to a given text. Discriminative methods usually incorporate the hierarchical structure information into the encoding process, while generative methods decode the features according to it. However, the data distribution varies widely among different categories of samples, but current methods ignore the data imbalance, making the predictions biased and susceptible to error propagation. In this paper, we propose an **IM**plicitly **A**ugmented **G**enerativ **E** framework with distribution modification for hierarchical text classification (**IMAGE**). Specifically, we translate the distributions of original samples along various directions through implicit augmentation to get more diverse data. Furthermore, given the scarcity of the samples of tail classes, we adjust their distributions by transferring knowledge from other classes in label space. In this way, the generative framework learns a better beginning of the feature sequence without a prediction bias and avoids being misled by its wrong predictions for head classes. Experimental results show that **IMAGE** obtains competitive results compared with state-of-the-art methods and prove its superiority on unbalanced data.

Incremental Learning (IL) has been a long-standing problem in both vision and Natural Language Processing (NLP) communities.In recent years, as Pre-trained Language Models (PLMs) have achieved remarkable progress in various NLP downstream tasks, utilizing PLMs as backbones has become a common practice in recent research of IL in NLP.Most assume that catastrophic forgetting is the biggest obstacle to achieving superior IL performance and propose various techniques to overcome this issue.However, we find that this assumption is problematic.Specifically, we revisit more than 20 methods on four classification tasks (Text Classification, Intent Classification, Relation Extraction, and Named Entity Recognition) under the two most popular IL settings (Class-Incremental and Task-Incremental) and reveal that most of them severely underestimate the inherent anti-forgetting ability of PLMs.Based on the observation, we propose a frustratingly easy method called SEQ* for IL with PLMs.The results show that SEQ* has competitive or superior performance compared with state-of-the-art (SOTA) IL methods yet requires considerably less trainable parameters and training time.These findings urge us to revisit the IL with PLMs and encourage future studies to have a fundamental understanding of the catastrophic forgetting in PLMs.

Metaphor Detection aims to identify the metaphorical meaning of words in the sentence. Most existing work is discriminant models, which use the contextual semantic information extracted by transformers for classifications directly. Due to insufficient training data and corresponding paraphrases, recent methods focus on how to get external resources and utilize them to introduce more knowledge. Currently, contextual modeling and external data are two key issues in the field. In this paper, we propose **A**n **A**uto-**A**ugmented **S**tructure-aware generative model (**AAAS**) for metaphor detection, which transforms the classification task into a keywords-extraction task. Specifically, we propose the task of structure information extraction to allow the model to use the ‘structural language’ to describe the whole sentence. Furthermore, without any other external resources, we design a simple but effective auto-augmented method to expand the limited datasets. Experimental results show that **AAAS** obtains competitive results compared with state-of-the-art methods.

Cross-domain NER is a practical yet challenging problem since the data scarcity in the real-world scenario. A common practice is first to learn a NER model in a rich-resource general domain and then adapt the model to specific domains. Due to the mismatch problem between entity types across domains, the wide knowledge in the general domain can not effectively transfer to the target domain NER model. To this end, we model the label relationship as a probability distribution and construct label graphs in both source and target label spaces. To enhance the contextual representation with label structures, we fuse the label graph into the word embedding output by BERT. By representing label relationships as graphs, we formulate cross-domain NER as a graph matching problem. Furthermore, the proposed method has good applicability with pre-training methods and is potentially capable of other cross-domain prediction tasks. Empirical results on four datasets show that our method outperforms a series of transfer learning, multi-task learning, and few-shot learning methods.

Emotion-cause pair extraction (ECPE) aims to extract emotion clauses and corresponding cause clauses, which have recently received growing attention. Previous methods sequentially encode features with a specified order. They first encode the emotion and cause features for clause extraction and then combine them for pair extraction. This lead to an imbalance in inter-task feature interaction where features extracted later have no direct contact with the former. To address this issue, we propose a novel **P**air-**B**ased **J**oint **E**ncoding (**PBJE**) network, which generates pairs and clauses features simultaneously in a joint feature encoding manner to model the causal relationship in clauses. PBJE can balance the information flow among emotion clauses, cause clauses and pairs. From a multi-relational perspective, we construct a heterogeneous undirected graph and apply the Relational Graph Convolutional Network (RGCN) to capture the multiplex relationship between clauses and the relationship between pairs and clauses. Experimental results show that PBJE achieves state-of-the-art performance on the Chinese benchmark corpus.

Continual Learning for Named Entity Recognition (CL-NER) aims to learn a growing number of entity types over time from a stream of data. However, simply learning Other-Class in the same way as new entity types amplifies the catastrophic forgetting and leads to a substantial performance drop. The main cause behind this is that Other-Class samples usually contain old entity types, and the old knowledge in these Other-Class samples is not preserved properly. Thanks to the causal inference, we identify that the forgetting is caused by the missing causal effect from the old data.To this end, we propose a unified causal framework to retrieve the causality from both new entity types and Other-Class.Furthermore, we apply curriculum learning to mitigate the impact of label noise and introduce a self-adaptive weight for balancing the causal effects between new entity types and Other-Class. Experimental results on three benchmark datasets show that our method outperforms the state-of-the-art method by a large margin. Moreover, our method can be combined with the existing state-of-the-art methods to improve the performance in CL-NER.

Metaphors are ubiquitous in natural language, and detecting them requires contextual reasoning about whether a semantic incongruence actually exists. Most existing work addresses this problem using pre-trained contextualized models. Despite their success, these models require a large amount of labeled data and are not linguistically-based. In this paper, we proposed a ContrAstive pre-Trained modEl (CATE) for metaphor detection with semi-supervised learning. Our model first uses a pre-trained model to obtain a contextual representation of target words and employs a contrastive objective to promote an increased distance between target words’ literal and metaphorical senses based on linguistic theories. Furthermore, we propose a simple strategy to collect large-scale candidate instances from the general corpus and generalize the model via self-training. Extensive experiments show that CATE achieves better performance against state-of-the-art baselines on several benchmark datasets.

Sequential sentence classification aims to classify each sentence in the document based on the context in which sentences appear. Most existing work addresses this problem using a hierarchical sequence labeling network. However, they ignore considering the latent segment structure of the document, in which contiguous sentences often have coherent semantics. In this paper, we proposed a span-based dynamic local attention model that could explicitly capture the structural information by the proposed supervised dynamic local attention. We further introduce an auxiliary task called span-based classification to explore the span-level representations. Extensive experiments show that our model achieves better or competitive performance against state-of-the-art baselines on two benchmark datasets.

Hierarchical text classification is an important yet challenging task due to the complex structure of the label hierarchy. Existing methods ignore the semantic relationship between text and labels, so they cannot make full use of the hierarchical information. To this end, we formulate the text-label semantics relationship as a semantic matching problem and thus propose a hierarchy-aware label semantics matching network (HiMatch). First, we project text semantics and label semantics into a joint embedding space. We then introduce a joint embedding loss and a matching learning loss to model the matching relationship between the text semantics and the label semantics. Our model captures the text-label semantics matching relationship among coarse-grained labels and fine-grained labels in a hierarchy-aware manner. The experimental results on various benchmark datasets verify that our model achieves state-of-the-art results.

The central problem of sentence classification is to extract multi-scale n-gram features for understanding the semantic meaning of sentences. Most existing models tackle this problem by stacking CNN and RNN models, which easily leads to feature redundancy and overfitting because of relatively limited datasets. In this paper, we propose a simple yet effective model called Multi-scale Orthogonal inDependEnt LSTM (MODE-LSTM), which not only has effective parameters and good generalization ability, but also considers multiscale n-gram features. We disentangle the hidden state of the LSTM into several independently updated small hidden states and apply an orthogonal constraint on their recurrent matrices. We then equip this structure with sliding windows of different sizes for extracting multi-scale n-gram features. Extensive experiments demonstrate that our model achieves better or competitive performance against state-of-the-art baselines on eight benchmark datasets. We also combine our model with BERT to further boost the generalization performance.