Frequency Balanced Datasets Lead to Better Language Models (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing evidence that high-frequency tokens in pretraining data might bias learning, causing undesired effects, is not clear. |
| Approach: | They propose a sampling algorithm that iteratively assesses token frequencies and removes sentences that contain still high-frequency tokens, resulting in a balanced dataset. |
| Outcome: | The proposed method reduces the amount of pre-training data required for training attention-based transformer language models by up to three times. |
Similar Papers
Stable Language Model Pre-training by Reducing Embedding Variability (2024.emnlp-main)
Copied to clipboard
| Challenge: | Stable pre-training is essential for achieving better-performing language models, but tracking pre-train stability is impractical due to high computational costs. |
| Approach: | They propose to use Token Embedding Variability as a proxy to estimate pre-training stability. |
| Outcome: | The proposed method improves stability and lowers perplexities even at deeper layer counts. |
Token-level Adaptive Training for Neural Machine Translation (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing token imbalance phenomenon in natural language as different tokens appear with different frequencies, which leads to different learning difficulties for tokens in Neural machine translation (NMT). |
| Approach: | They propose to assign tokens with different frequencies to target tokens during training to encourage the model to pay more attention to low-frequency tokens. |
| Outcome: | The proposed model yields consistent improvements on ZH-EN, EN-RO, and EN-DE translation tasks, especially on sentences that contain more low-frequency tokens. |
Exploring Quantization for Efficient Pre-Training of Transformer Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Quantization has proven to be effective after pre-training and during fine-tuning, but its effects on pre-trainer performance have remained unexplored. |
| Approach: | They propose a linear quantization strategy to be applied during the pre-training of Transformers to improve model efficiency and stability. |
| Outcome: | The proposed method improves model efficiency, stability, and performance while maintaining language modeling ability. |
Your fairness may vary: Pretrained language model fairness in toxic text classification (2022.findings-acl)
Copied to clipboard
| Challenge: | Pre-trained, bidirectional language models have revolutionized natural language processing research . authors show that focusing on accuracy measures alone can lead to models with wide variation in fairness characteristics . |
| Approach: | They propose to use two post-processing methods to improve model fairness without retraining . they use pretrained language models of varying sizes on two toxic text classification tasks . |
| Outcome: | The proposed methods improve model fairness without retraining . the results show that the fairness variation is more than just accuracy . |
Selective Prefix Tuning for Pre-trained Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for fine-tuning pre-trained models are time-consuming and memory-inefficient. |
| Approach: | They propose a method that inserts learnable vectors into each Transformer layer . they propose SL to encourage diversity in prefix tokens . |
| Outcome: | Extensive experiments validate the effectiveness of Prefix Tuning in sentence and token classification tasks. |
How much pretraining data do language models need to learn syntax? (2021.emnlp-main)
Copied to clipboard
| Challenge: | Pretraining methods are convenient, but expensive in terms of time and resources. |
| Approach: | They investigate the impact of pretraining data size on the syntactic capabilities of RoBERTa by using syntaktic structural probes to determine whether models pretrained on more data encode a higher amount of syntastic information. |
| Outcome: | The proposed models perform better on part-of-speech tagging, dependency parsing and paraphrase identification. |
Mini But Mighty: Efficient Multilingual Pretraining with Linguistically-Informed Data Selection (2023.findings-eacl)
Copied to clipboard
| Challenge: | AfriBERTa shows that training transformer models from scratch on 1GB of data from many unrelated African languages outperforms massively multilingual models on downstream NLP tasks. |
| Approach: | They propose that training on smaller amounts of data but from related languages could match the performance of models trained on large, unrelated data. |
| Outcome: | The proposed model outperforms models trained on large, unrelated datasets on downstream NLP tasks. |
Teacher and Student Models of Offensive Language in Social Media (2023.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to identify offensive language online use large pre-trained transformer models. however, the inference time, disk, and memory requirements of these models are prohibitively large. |
| Approach: | They propose to transfer knowledge from large transformer models to much smaller neural models to make predictions at the token- and post-level. |
| Outcome: | The proposed model performs 100 times better than transformer models but with 100 times less parameters and much less memory usage. |
Exploring Versatile Generative Language Model Via Parameter-Efficient Transfer Learning (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Large-scale language models can be fine-tuned to learn highly transferable embedding, but they are expensive and require multiple model parameters. |
| Approach: | They propose a way to fine-tune multiple down-stream generation tasks simultaneously using a single, large pretrained model. |
| Outcome: | The proposed model can maintain or improve the performance of fine-tuning the whole model. |
Adding Instructions during Pretraining: Effective way of Controlling Toxicity in Language Models (2023.eacl-main)
Copied to clipboard
| Challenge: | Pretrained large language models generate harmful language encompassing hate speech, abusive language, social biases, and threats. |
| Approach: | They propose two strategies that augment pretraining data to reduce model toxicity . MEDA adds raw toxicity score as meta-data and INST adds instructions indicating toxicity to pretraining samples. |
| Outcome: | The proposed strategies reduce toxicity probability up to 61% while preserving accuracy on five benchmark NLP tasks and improving AUC scores on bias detection tasks by 1.3%. |