Large Language Models (LLMs) have significantly advanced Natural Language Processing (NLP), particularly in Natural Language Understanding (NLU) tasks. As we progress toward an agentic world where LLM-based agents autonomously handle specialized tasks, it becomes crucial for these models to adapt to new tasks without forgetting previously learned information - a challenge known as catastrophic forgetting. This study evaluates the continual fine-tuning of various open-source LLMs with different parameter sizes (specifically models under 10 billion parameters) on key NLU tasks from the GLUE benchmark, including SST-2, MRPC, CoLA, and MNLI. By employing prompt engineering and task-specific adjustments, we assess and compare the models' abilities to retain prior knowledge while learning new tasks. Our results indicate that models such as Phi-3.5-mini exhibit minimal forgetting while maintaining strong learning capabilities, making them well-suited for continual learning environments. Additionally, models like Orca-2-7b and Qwen2.5-7B demonstrate impressive learning abilities and overall performance after fine-tuning. This work contributes to understanding catastrophic forgetting in LLMs and highlights prompting engineering to optimize model performance for continual learning scenarios.
This paper investigates the efficacy of RWKV, a novel language model architecture known for its linear attention mechanism, for generating sentence embeddings in a zero-shot setting. I conduct a layer-wise analysis to evaluate the semantic similarity captured by embeddings from different hidden layers of a pre-trained RWKV model. The performance is assessed on the Microsoft Research Paraphrase Corpus (MRPC) dataset using Spearman correlation and compared against a GloVe-based baseline. My results indicate that while RWKV embeddings capture some semantic relatedness, they underperform compared to the GloVe baseline in terms of Spearman correlation. I also analyze the inference time and GPU memory usage, highlighting the computational trade-offs associated with RWKV embeddings. The findings suggest that while RWKV offers potential advantages in terms of linear scaling, its zero-shot sentence embedding quality for semantic similarity tasks requires further investigation and potential task-specific fine-tuning to match or exceed simpler baselines.
A major limitation of prompt tuning is its dependence on large labeled training datasets. Under few-shot learning settings, prompt tuning lags far behind full-model fine-tuning, limiting its scope of application. In this paper, we leverage the powerful LLMs to synthesize task-specific labeled data for training the soft prompts. We first introduce a distribution-aligned weighted generator tuning (DawGen) method to encourage generating in-distribution data that aligns with the few-shot real data. Then, we train soft prompts on both synthetic and real datasets using a gradient surgery approach, which eliminates the conflicting gradients from different data sources. Experiments on seven sentence-pair classification datasets demonstrate the effectiveness of our proposed method for boosting prompt tuning in few-shot learning settings. Results on QQP, MRPC, and SICK datasets are even comparable to the performance of transfer learning from large real-world datasets, showing the promise of synthetic data as an alternative for enhancing soft prompt tuning.
Low-rank Adaption (LoRA) has been the de-facto parameter-efficient fine-tuning technique for large language models. We present HeteroLoRA, a light-weight search algorithm that leverages zero-cost proxies to allocate the limited LoRA trainable parameters across the model for better fine-tuned performance. In addition to the allocation for the standard LoRA-adapted models, we also demonstrate the efficacy of HeteroLoRA by performing the allocation in a more challenging search space that includes LoRA modules and LoRA-adapted shortcut connections. Experiments show that HeteroLoRA enables improvements in model performance given the same parameter budge. For example, on MRPC, we see an improvement of 1.6% in accuracy with similar training parameter budget. We will open-source our algorithm once the paper is accepted.
Sentence semantic matching is a research hotspot in natural language processing, which is considerably significant in various key scenarios, such as community question answering, searching, chatbot, and recommendation. Since most of the advanced models directly model the semantic relevance among words between two sentences while neglecting the \textit{keywords} and \textit{intents} concepts of them, DC-Match is proposed to disentangle keywords from intents and utilizes them to optimize the matching performance. Although DC-Match is a simple yet effective method for semantic matching, it highly depends on the external NER techniques to identify the keywords of sentences, which limits the performance of semantic matching for minor languages since satisfactory NER tools are usually hard to obtain. In this paper, we propose to generally and flexibly resolve the text into multi concepts for multilingual semantic matching to liberate the model from the reliance on NER models. To this end, we devise a \underline{M}ulti-\underline{C}oncept \underline{P}arsed \underline{S}emantic \underline{M}atching framework based on the pre-trained language models, abbreviated as \textbf{MCP-SM}, to extract various concepts and infuse them into the classification tokens. We conduct comprehensive experiments on English datasets QQP and MRPC, and Chinese dataset Medical-SM. Besides, we experiment on Arabic datasets MQ2Q and XNLI, the outstanding performance further prove MCP-SM's applicability in low-resource languages.
Fine-tuning large pre-trained language models for downstream tasks remains a critical challenge in natural language processing. This paper presents an empirical analysis comparing two efficient fine-tuning methods - BitFit and adapter modules - to standard full model fine-tuning. Experiments conducted on GLUE benchmark datasets (MRPC, COLA, STS-B) reveal several key insights. The BitFit approach, which trains only bias terms and task heads, matches full fine-tuning performance across varying amounts of training data and time constraints. It demonstrates remarkable stability even with only 30\% of data, outperforming full fine-tuning at intermediate data levels. Adapter modules exhibit high variability, with inconsistent gains over default models. The findings indicate BitFit offers an attractive balance between performance and parameter efficiency. Our work provides valuable perspectives on model tuning, emphasizing robustness and highlighting BitFit as a promising alternative for resource-constrained or streaming task settings. The analysis offers actionable guidelines for efficient adaptation of large pre-trained models, while illustrating open challenges in stabilizing techniques like adapter modules.
Building on the cost-efficient pretraining advancements brought about by Crammed BERT, we enhance its performance and interpretability further by introducing a novel pretrained model Dependency Agreement Crammed BERT (DACBERT) and its two-stage pretraining framework - Dependency Agreement Pretraining. This framework, grounded by linguistic theories, seamlessly weaves syntax and semantic information into the pretraining process. The first stage employs four dedicated submodels to capture representative dependency agreements at the chunk level, effectively converting these agreements into embeddings. The second stage uses these refined embeddings, in tandem with conventional BERT embeddings, to guide the pretraining of the rest of the model. Evaluated on the GLUE benchmark, our DACBERT demonstrates notable improvement across various tasks, surpassing Crammed BERT by 3.13% in the RTE task and by 2.26% in the MRPC task. Furthermore, our method boosts the average GLUE score by 0.83%, underscoring its significant potential. The pretraining process can be efficiently executed on a single GPU within a 24-hour cycle, necessitating no supplementary computational resources or extending the pretraining duration compared with the Crammed BERT. Extensive studies further illuminate our approach's instrumental role in bolstering the interpretability of pretrained language models for natural language understanding tasks.
Adapter tuning, which updates only a few parameters, has become a mainstream method for fine-tuning pretrained language models to downstream tasks. However, it often yields subpar results in few-shot learning. AdapterFusion, which assembles pretrained adapters using composition layers tailored to specific tasks, is a possible solution but significantly increases trainable parameters and deployment costs. Despite this, our preliminary study reveals that even single adapters can outperform Adapterfusion in few-shot learning, urging us to propose \textbf{\texttt{Merging Pretrained Adapters}} (MerA) that efficiently incorporates pretrained adapters to a single model through model fusion. Extensive experiments on two PLMs demonstrate that MerA achieves substantial improvements compared to both single adapters and AdapterFusion. To further enhance the capacity of MerA, we also introduce a simple yet effective technique, referred to as the "\textit{same-track}" setting, that merges adapters from the same track of pretraining tasks. With the implementation of the "\textit{same-track}" setting, we observe even more impressive gains, surpassing the performance of both full fine-tuning and adapter tuning by a substantial margin, e.g., 3.5\% in MRPC and 5.0\% in MNLI.
In this paper, we study the problem of generating obstinate (over-stability) adversarial examples by word substitution in NLP, where input text is meaningfully changed but the model's prediction does not, even though it should. Previous word substitution approaches have predominantly focused on manually designed antonym-based strategies for generating obstinate adversarial examples, which hinders its application as these strategies can only find a subset of obstinate adversarial examples and require human efforts. To address this issue, in this paper, we introduce a novel word substitution method named GradObstinate, a gradient-based approach that automatically generates obstinate adversarial examples without any constraints on the search space or the need for manual design principles. To empirically evaluate the efficacy of GradObstinate, we conduct comprehensive experiments on five representative models (Electra, ALBERT, Roberta, DistillBERT, and CLIP) finetuned on four NLP benchmarks (SST-2, MRPC, SNLI, and SQuAD) and a language-grounding benchmark (MSCOCO). Extensive experiments show that our proposed GradObstinate generates more powerful obstinate adversarial examples, exhibiting a higher attack success rate compared to antonym-based methods. Furthermore, to show the transferability of obstinate word substitutions found by GradObstinate, we replace the words in four representative NLP benchmarks with their obstinate substitutions. Notably, obstinate substitutions exhibit a high success rate when transferred to other models in black-box settings, including even GPT-3 and ChatGPT. Examples of obstinate adversarial examples found by GradObstinate are available at https://huggingface.co/spaces/anonauthors/SecretLanguage.
Combining large language models with logical reasoning enhances their capacity to address problems in a robust and reliable manner. Nevertheless, the intricate nature of logical reasoning poses challenges when gathering reliable data from the web to build comprehensive training datasets, subsequently affecting performance on downstream tasks. To address this, we introduce a novel logic-driven data augmentation approach, AMR-LDA. AMR-LDA converts the original text into an Abstract Meaning Representation (AMR) graph, a structured semantic representation that encapsulates the logical structure of the sentence, upon which operations are performed to generate logically modified AMR graphs. The modified AMR graphs are subsequently converted back into text to create augmented data. Notably, our methodology is architecture-agnostic and enhances both generative large language models, such as GPT-3.5 and GPT-4, through prompt augmentation, and discriminative large language models through contrastive learning with logic-driven data augmentation. Empirical evidence underscores the efficacy of our proposed method with improvement in performance across seven downstream tasks, such as reading comprehension requiring logical reasoning, textual entailment, and natural language inference. Furthermore, our method leads on the ReClor leaderboard at https://eval.ai/web/challenges/challenge-page/503/leaderboard/1347. The source code and data are publicly available at https://github.com/Strong-AI-Lab/Logical-Equivalence-driven-AMR-Data-Augmentation-for-Representation-Learning.
Through exploiting a high level of parallelism enabled by graphics processing units, transformer architectures have enabled tremendous strides forward in the field of natural language processing. In a traditional masked language model, special MASK tokens are used to prompt our model to gather contextual information from surrounding words to restore originally hidden information. In this paper, we explore a task-specific masking framework for pre-trained large language models that enables superior performance on particular downstream tasks on the datasets in the GLUE benchmark. We develop our own masking algorithm, Typhoon, based on token input gradients, and compare this with other standard baselines. We find that Typhoon offers performance competitive with whole-word masking on the MRPC dataset. Our implementation can be found in a public Github Repository.
Recently, there has been a surge in the use of generated data to enhance the performance of downstream models, largely due to the advancements in pre-trained language models. However, most prevailing methods trained generative and discriminative models in isolation, which left them unable to adapt to changes in each other. These approaches lead to generative models that are prone to deviating from the true data distribution and providing limited benefits to discriminative models. While some works have proposed jointly training generative and discriminative language models, their methods remain challenging due to the non-differentiable nature of discrete data. To overcome these issues, we introduce a \textit{self-consistent learning} framework in the text field that involves training a discriminator and generator cooperatively in a closed-loop manner until a scoring consensus is reached. By learning directly from selected samples, our framework are able to mitigate training instabilities such as mode collapse and non-convergence. Extensive experiments on four downstream benchmarks, including AFQMC, CHIP-STS, QQP, and MRPC, demonstrate the efficacy of the proposed framework.
Recently, neural language representation models pre-trained on large corpus can capture rich co-occurrence information and be fine-tuned in downstream tasks to improve the performance. As a result, they have achieved state-of-the-art results in a large range of language tasks. However, there exists other valuable semantic information such as similar, opposite, or other possible meanings in external knowledge graphs (KGs). We argue that entities in KGs could be used to enhance the correct semantic meaning of language sentences. In this paper, we propose a new method CKG: Dynamic Representation Based on \textbf{C}ontext and \textbf{K}nowledge \textbf{G}raph. On the one side, CKG can extract rich semantic information of large corpus. On the other side, it can make full use of inside information such as co-occurrence in large corpus and outside information such as similar entities in KGs. We conduct extensive experiments on a wide range of tasks, including QQP, MRPC, SST-5, SQuAD, CoNLL 2003, and SNLI. The experiment results show that CKG achieves SOTA 89.2 on SQuAD compared with SAN (84.4), ELMo (85.8), and BERT$_{Base}$ (88.5).
With the yearning for deep learning democratization, there are increasing demands to implement Transformer-based natural language processing (NLP) models on resource-constrained devices for low-latency and high accuracy. Existing BERT pruning methods require domain experts to heuristically handcraft hyperparameters to strike a balance among model size, latency, and accuracy. In this work, we propose AE-BERT, an automatic and efficient BERT pruning framework with efficient evaluation to select a "good" sub-network candidate (with high accuracy) given the overall pruning ratio constraints. Our proposed method requires no human experts experience and achieves a better accuracy performance on many NLP tasks. Our experimental results on General Language Understanding Evaluation (GLUE) benchmark show that AE-BERT outperforms the state-of-the-art (SOTA) hand-crafted pruning methods on BERT$_{\mathrm{BASE}}$. On QNLI and RTE, we obtain 75\% and 42.8\% more overall pruning ratio while achieving higher accuracy. On MRPC, we obtain a 4.6 higher score than the SOTA at the same overall pruning ratio of 0.5. On STS-B, we can achieve a 40\% higher pruning ratio with a very small loss in Spearman correlation compared to SOTA hand-crafted pruning methods. Experimental results also show that after model compression, the inference time of a single BERT$_{\mathrm{BASE}}$ encoder on Xilinx Alveo U200 FPGA board has a 1.83$\times$ speedup compared to Intel(R) Xeon(R) Gold 5218 (2.30GHz) CPU, which shows the reasonableness of deploying the proposed method generated subnets of BERT$_{\mathrm{BASE}}$ model on computation restricted devices.
While contrastive learning is proven to be an effective training strategy in computer vision, Natural Language Processing (NLP) is only recently adopting it as a self-supervised alternative to Masked Language Modeling (MLM) for improving sequence representations. This paper introduces SupCL-Seq, which extends the supervised contrastive learning from computer vision to the optimization of sequence representations in NLP. By altering the dropout mask probability in standard Transformer architectures, for every representation (anchor), we generate augmented altered views. A supervised contrastive loss is then utilized to maximize the system's capability of pulling together similar samples (e.g., anchors and their altered views) and pushing apart the samples belonging to the other classes. Despite its simplicity, SupCLSeq leads to large gains in many sequence classification tasks on the GLUE benchmark compared to a standard BERTbase, including 6% absolute improvement on CoLA, 5.4% on MRPC, 4.7% on RTE and 2.6% on STSB. We also show consistent gains over self supervised contrastively learned representations, especially in non-semantic tasks. Finally we show that these gains are not solely due to augmentation, but rather to a downstream optimized sequence representation. Code: https://github.com/hooman650/SupCL-Seq
Although pretrained language models can be fine-tuned to produce state-of-the-art results for a very wide range of language understanding tasks, the dynamics of this process are not well understood, especially in the low data regime. Why can we use relatively vanilla gradient descent algorithms (e.g., without strong regularization) to tune a model with hundreds of millions of parameters on datasets with only hundreds or thousands of labeled examples? In this paper, we argue that analyzing fine-tuning through the lens of intrinsic dimension provides us with empirical and theoretical intuitions to explain this remarkable phenomenon. We empirically show that common pre-trained models have a very low intrinsic dimension; in other words, there exists a low dimension reparameterization that is as effective for fine-tuning as the full parameter space. For example, by optimizing only 200 trainable parameters randomly projected back into the full space, we can tune a RoBERTa model to achieve 90\% of the full parameter performance levels on MRPC. Furthermore, we empirically show that pre-training implicitly minimizes intrinsic dimension and, perhaps surprisingly, larger models tend to have lower intrinsic dimension after a fixed number of pre-training updates, at least in part explaining their extreme effectiveness. Lastly, we connect intrinsic dimensionality with low dimensional task representations and compression based generalization bounds to provide intrinsic-dimension-based generalization bounds that are independent of the full parameter count.
Large pre-trained language representation models (LMs) have recently collected a huge number of successes in many NLP tasks. In 2018 BERT, and later its successors (e.g. RoBERTa), obtained state-of-the-art results in classical benchmark tasks, such as GLUE benchmark. After that, works about adversarial attacks have been published to test their generalization proprieties and robustness. In this work, we design Evolutionary Fooling Sentences Generator (EFSG), a model- and task-agnostic adversarial attack algorithm built using an evolutionary approach to generate false-positive sentences for binary classification tasks. We successfully apply EFSG to CoLA and MRPC tasks, on BERT and RoBERTa, comparing performances. Results prove the presence of weak spots in state-of-the-art LMs. We finally test adversarial training as a data augmentation defence approach against EFSG, obtaining stronger improved models with no loss of accuracy when tested on the original datasets.
Neural Architecture Search (NAS) methods, which automatically learn entire neural model or individual neural cell architectures, have recently achieved competitive or state-of-the-art (SOTA) performance on variety of natural language processing and computer vision tasks, including language modeling, natural language inference, and image classification. In this work, we explore the applicability of a SOTA NAS algorithm, Efficient Neural Architecture Search (ENAS) (Pham et al., 2018) to two sentence pair tasks, paraphrase detection and semantic textual similarity. We use ENAS to perform a micro-level search and learn a task-optimized RNN cell architecture as a drop-in replacement for an LSTM. We explore the effectiveness of ENAS through experiments on three datasets (MRPC, SICK, STS-B), with two different models (ESIM, BiLSTM-Max), and two sets of embeddings (Glove, BERT). In contrast to prior work applying ENAS to NLP tasks, our results are mixed -- we find that ENAS architectures sometimes, but not always, outperform LSTMs and perform similarly to random architecture search.
Newly-introduced deep learning architectures, namely BERT, XLNet, RoBERTa and ALBERT, have been proved to be robust on several NLP tasks. However, the datasets trained on these architectures are fixed in terms of size and generalizability. To relieve this issue, we apply one of the most inexpensive solutions to update these datasets. We call this approach BET by which we analyze the backtranslation data augmentation on the transformer-based architectures. Using the Google Translate API with ten intermediary languages from ten different language families, we externally evaluate the results in the context of automatic paraphrase identification in a transformer-based framework. Our findings suggest that BET improves the paraphrase identification performance on the Microsoft Research Paraphrase Corpus (MRPC) to more than 3% on both accuracy and F1 score. We also analyze the augmentation in the low-data regime with downsampled versions of MRPC, Twitter Paraphrase Corpus (TPC) and Quora Question Pairs. In many low-data cases, we observe a switch from a failing model on the test set to reasonable performances. The results demonstrate that BET is a highly promising data augmentation technique: to push the current state-of-the-art of existing datasets and to bootstrap the utilization of deep learning architectures in the low-data regime of a hundred samples.
We derive a new margin-based regularization formulation, termed multi-margin regularization (MMR), for deep neural networks (DNNs). The MMR is inspired by principles that were applied in margin analysis of shallow linear classifiers, e.g., support vector machine (SVM). Unlike SVM, MMR is continuously scaled by the radius of the bounding sphere (i.e., the maximal norm of the feature vector in the data), which is constantly changing during training. We empirically demonstrate that by a simple supplement to the loss function, our method achieves better results on various classification tasks across domains. Using the same concept, we also derive a selective sampling scheme and demonstrate accelerated training of DNNs by selecting samples according to a minimal margin score (MMS). This score measures the minimal amount of displacement an input should undergo until its predicted classification is switched. We evaluate our proposed methods on three image classification tasks and six language text classification tasks. Specifically, we show improved empirical results on CIFAR10, CIFAR100 and ImageNet using state-of-the-art convolutional neural networks (CNNs) and BERT-BASE architecture for the MNLI, QQP, QNLI, MRPC, SST-2 and RTE benchmarks.