Mitigating Frequency Bias and Anisotropy in Language Model Pre-Training with Syntactic Smoothing

Abstract: LLMs strongly rely on frequency information because they maximize the likelihood of tokens during pre-training. As a consequence, LLMs tend to not generalize well to tokens that are seldom seen during training. Moreover, maximum likelihood training has been discovered to give rise to anisotropy: representations of tokens in a model tend to cluster tightly in a high-dimensional cone, rather than spreading out over their representational capacity. Our work introduces a method for quantifying the frequency bias of a LLM by assessing sentence-level perplexity with respect to token-level frequency. We then present a method for reducing the frequency bias of a LLM by inducing a syntactic prior over token representations during pre-training. Our Syntactic Smoothing method adjusts the maximum likelihood objective function to distribute the learning signal to syntactically similar tokens. This approach results in better performance on infrequent English tokens and a decrease in anisotropy. We empirically show that the degree of anisotropy in a model correlates with its frequency bias.

• The paper introduces Syntactic Smoothing, a novel pre-training method that successfully mitigates frequency bias and representational anisotropy in language models.
• Evaluations show Syntactic Smoothing significantly reduces frequency bias and anisotropy, enabling competitive performance on various NLP tasks.
• The method offers a path to more equitable language model training, improving performance on rare tokens without requiring larger datasets or architectures.

Mitigating Frequency Bias and Anisotropy in LLM Pre-Training with Syntactic Smoothing

The paper "Mitigating Frequency Bias and Anisotropy in LLM Pre-Training with Syntactic Smoothing" addresses significant challenges in LLM training, specifically frequency bias and representational anisotropy. LLMs often exhibit frequency bias due to the over-representation of frequent tokens in training data, following the Zipfian distribution. This bias adversely impacts the model's ability to generalize, especially to infrequent tokens. Additionally, the anisotropy in LLM representations, where hidden states cluster in a limited subspace, hampers the representational efficacy.

The research introduces a novel approach termed 'Syntactic Smoothing' as a solution to these issues. By integrating a syntactic prior using a modified loss function during pre-training, the method distributes learning signals across syntactically similar tokens. This process allows infrequent tokens to benefit from the frequent updates of syntactically analogous but more common tokens. The efficacy of this method was evaluated via a new metric for quantifying frequency bias in LLMs and demonstrated a reduction in both frequency bias and anisotropy.

In quantitative evaluations, the method showed that the implementation of Syntactic Smoothing significantly reduces frequency bias, nearly negating it in certain models. The approach also resulted in a lower degree of anisotropy, yielding a more evenly distributed representation space. These improvements did not compromise the model’s overall language understanding and enabled it to achieve competitive performance on various NLP tasks, comparable to existing baselines.

The implications of this research are noteworthy, offering pathways to more equitable LLM training without the need for excessively larger datasets or model architectures. Practically, this could lead to more robust models with better performance in processing rare and low-frequency lexical items. Theoretically, it opens up discussions regarding the representation needs of LLMs and challenges the prevailing assumptions about model scaling as a universal solution to representation issues.

Future developments could explore the scalability of Syntactic Smoothing in larger models and diverse languages other than English. Additionally, the exploration of alternative linguistic priors that could inform and optimize model training processes presents an intriguing avenue for extending this research. The findings indubitably contribute to the ongoing AI discourse on optimizing LLMs for more nuanced and balanced language comprehension.