Stochastically sampling word segmentations from a subword tokeniser, also called subword regularisation, is a known way to increase robustness of language models to out-of-distribution inputs, such as text containing spelling errors. Recent work has observed that usual augmentations that make popular deterministic subword tokenisers stochastic still cause only a handful of all possible segmentations to be sampled. It has been proposed to uniformly sample across these instead, through rejection sampling of paths in an unweighted segmentation graph. In this paper, we argue that uniformly random segmentation in turn skews the distributions of certain segmentational properties (e.g. token lengths and amount of tokens produced) away from uniformity, which still ends up hiding meaningfully diverse tokenisations. We propose an alternative uniform sampler using the same segmentation graph, but weighted by counting the paths through it. Our sampling algorithm, GRaMPa, provides hyperparameters allowing sampled tokenisations to skew towards fewer, longer tokens. Furthermore, GRaMPa is single-pass, guaranteeing significantly better computational complexity than previous approaches relying on rejection sampling. We show experimentally that language models trained with GRaMPa outperform existing regularising tokenisers in a data-scarce setting on token-level tasks such as dependency parsing, especially with spelling errors present.

Pixel-based language models have emerged as a compelling alternative to subword-based language modelling, particularly because they can represent virtually any script. PIXEL, a canonical example of such a model, is a vision transformer that has been pre-trained on rendered text. While PIXEL has shown promising cross-script transfer abilities and robustness to orthographic perturbations, it falls short of outperforming monolingual subword counterparts like BERT in most other contexts. This discrepancy raises questions about the amount of linguistic knowledge learnt by these models and whether their performance in language tasks stems more from their visual capabilities than their linguistic ones. To explore this, we probe PIXEL using a variety of linguistic and visual tasks to assess its position on the vision-to-language spectrum. Our findings reveal a substantial gap between the model’s visual and linguistic understanding. The lower layers of PIXEL predominantly capture superficial visual features, whereas the higher layers gradually learn more syntactic and semantic abstractions. Additionally, we examine variants of PIXEL trained with different text rendering strategies, discovering that introducing certain orthographic constraints at the input level can facilitate earlier learning of surface-level features. With this study, we hope to provide insights that aid the further development of pixel-based language models.

Byte-pair encoding (BPE) has become the default subword tokeniser in language models (LMs), allowing the representation of an infinite space of text with a finite set of units. Yet, BPE training is unsupervised, receiving no explicit information about a language’s morphology. This results in a subword vocabulary wherein many units are a concatenation of partial morphemes, preventing their formation as tokens. This, in turn, causes consistent intra-word patterns to be displayed inconsistently to downstream models, and bloats the vocabulary, hence requiring unnecessary embedding storage. In this paper, we address this issue by identifying blameworthy BPE merges and removing the resulting subwords from the BPE vocabulary, without impeding further use of merges that relied on them. We find that our method, BPE-knockout, is effective at making BPE’s segmentation positions adhere better to derivational and compound boundaries in English, Dutch and German, and improves token-based tasks in Dutch RoBERTa models, indicating that a tokeniser’s adherence to morphology impacts downstream models. We demonstrate the latter not only by training LMs from scratch, but also by continuing the pre-training of existing LMs. This proves promising, showing that suboptimal tokenisers can be remedied whilst salvaging training cost of downstream LMs.