Mitigating Frequency Bias and Anisotropy in Language Model Pre-Training with Syntactic Smoothing (2024.emnlp-main)
Copied to clipboard
| Challenge: | Language models rely on frequency information because they maximize the likelihood of tokens during training. |
| Approach: | They propose a method for quantifying the frequency bias of a language model by assessing sentence-level perplexity with respect to token-level frequency. |
| Outcome: | The proposed method reduces the frequency bias of a language model by inducing a syntactic prior over token representations during pre-training. |
Similar Papers
A Natural Bias for Language Generation Models (2023.acl-short)
Copied to clipboard
| Challenge: | a standard probabilistic model for language generation has likely not yet learnt many semantic or syntactic rules of natural language, making it difficult to estimate the probability distribution over next tokens. |
| Approach: | They propose to initialise bias terms in a model's final linear layer with the log-unigram distribution and use it to output the unigram frequency statistics as prior knowledge. |
| Outcome: | The proposed method improves learning efficiency and improves overall performance. |
Token-level and sequence-level loss smoothing for RNN language models (P18-1)
Copied to clipboard
| Challenge: | Maximum likelihood estimation treats all sentences that do not match the ground truth as equally poor, ignoring the structure of the output space. |
| Approach: | They propose to extend the reward augmented maximum likelihood approach to token-level loss smoothing by using token-based approaches to improve the model's performance. |
| Outcome: | The proposed model improves on image captioning and machine translation tasks and treats all sentences that do not match the ground truth as poor . |
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. |
Debiasing Pre-trained Contextualised Embeddings (2021.eacl-main)
Copied to clipboard
| Challenge: | a study of contextualised word embeddings shows discriminative biases are encoded in contextualised embeddables. |
| Approach: | They propose a fine-tuning method that can be applied at token- or sentence-levels to debias pre-trained contextualised embeddings. |
| Outcome: | The proposed method can be applied at token- or sentence-levels to debias pre-trained models without requiring retrains. |
Overestimation of Syntactic Representation in Neural Language Models (2020.acl-main)
Copied to clipboard
| Challenge: | Several testing methodologies have been developed to probe models’ syntactic representations. |
| Approach: | They propose a method to determine syntactic structure by training a model on strings generated according to a template and testing its ability to distinguish between similar ones with different syntax. |
| Outcome: | The proposed method reproduces positive results with two non-syntactic baseline language models: an n-gram model and an LSTM model trained on scrambled inputs. |
An Empirical Analysis of Parameter-Efficient Methods for Debiasing Pre-Trained Language Models (2023.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models inherit more human-like biases from the training corpora, causing computationally expensive problems. |
| Approach: | They propose parameter-efficient methods in combination with counterfactual data augmentation for bias mitigation. |
| Outcome: | The proposed methods are effective in mitigating gender bias, prompt tuning is more suitable for GPT-2 than BERT, and less effective when it comes to racial and religious bias. |
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 Self-Evolution Training for Sequence-to-Sequence Learning (2023.acl-short)
Copied to clipboard
| Challenge: | Adaptive training approaches do not consider the variation of learning difficulty in different training steps, making the learning deterministic and sub-optimal. |
| Approach: | They propose a dynamic token-level self-evolution training method that reweighs the training losses of different target tokens based on priors. |
| Outcome: | Empirically, the proposed method yields significant improvements on three translation tasks. |
Debiasing Large Language Models with Structured Knowledge (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods to reduce biases in pre-training models are hampered by their performance. |
| Approach: | They propose a method that utilizes structured knowledge to mitigate bias in LLMs . their method obviates the need for training from scratch, thus offering enhanced scalability . |
| Outcome: | The proposed method outperforms state-of-the-art (SOTA) baselines in the debiasing ability. |
How to Plant Trees in Language Models: Data and Architectural Effects on the Emergence of Syntactic Inductive Biases (2023.acl-long)
Copied to clipboard
| Challenge: | a recent study found that pre-training can teach language models to rely on hierarchical syntactic features . aaron ramirez: we find that pretraining on simpler language induces a hierarchic bias . |
| Approach: | They find that pre-training can teach language models to rely on hierarchical syntactic features . authors: this suggests that in cognitively plausible language acquisition settings, models may be more data-efficient . |
| Outcome: | a recent study shows that pre-training can teach language models to rely on hierarchical features . the findings suggest that in plausible language acquisition settings, language models may be more data-efficient than previously thought . |