Signal in Noise: Exploring Meaning Encoded in Random Character Sequences with Character-Aware Language Models

Mark Chu, Bhargav Srinivasa Desikan, Ethan O. Nadler, Donald Ruggiero Lo Sardo, Elise Darragh-Ford, Douglas GuilbeaultACL 2022 · DOI

Introduction

Modern Natural Language Processing (NLP) predominantly relies on the distributional hypothesis, which posits that a word's meaning is derived from its context. While effective for known vocabulary, this paradigm struggles with out-of-vocabulary (OOV) terms or nonsensical strings lacking context. Standard models like BERT utilize sub-word tokenization to mitigate this, yet they often fail to capture the primitive information inherent in character sequences. This study investigates how character-aware models, specifically CharacterBERT, represent the building blocks of meaning in the absence of context by analyzing "garble" (random strings), pseudowords, and extant language.

Methodology

The researchers utilized CharacterBERT, a variant of BERT that employs a Character-level Convolutional Neural Network (CharacterCNN) instead of standard tokenization. This architecture enables the generation of a single embedding for any character sequence. To analyze the 512-dimensional embedding space, the authors employed:

  • UMAP (Uniform Manifold Approximation and Projection): For high-dimensional visualization.
  • The Information Axis: A specific dimension identified to capture the variance between random noise and real words.
  • Hyperplane Classifiers: Support Vector Machines (SVM) were trained to distinguish noise from language.
  • Statistical Comparison: Results were compared against a Variable Order Markov Model to determine if the transformer learns patterns beyond simple character frequencies.

The datasets comprised 40,000 common English words (extant words), 40,000 randomly generated character sequences (garble), and 20,000 phonologically plausible non-existent words (pseudowords).

The Linguistic Gradient and the Information Axis

A primary finding is the existence of a "linguistic gradient" within the embedding space. Rather than a binary distinction between noise and language, the model exhibits a smooth transition from pure garble to pseudowords, and finally to extant words.

Figure 1: UMAP visualization of the embedding space showing the transition from random n-grams (black) through pseudowords (pink) to extant words (blue) along the identified Information Axis.

The SVM classifier achieved a 98.9% accuracy in distinguishing unseen extant words from garble, confirming that the "information axis" is a robust internal feature. Interestingly, pseudowords act as a structural bridge, occupying the space between noise and meaningful language.

Morphology and Part-of-Speech Distribution

The model's internal representations naturally segregate parts of speech along the information axis. This suggests that CharacterBERT utilizes repetitive morphological patterns—such as suffixes and prefixes—as primary signals for "languageness."

Figure 3: UMAP and probability density plots illustrating the distribution of different parts of speech along the information axis. Adverbs, characterized by distinct suffixes like "-ly", show the highest separation from noise.

Adverbs are positioned furthest from random noise, while nouns and verbs sit closer to the center of the linguistic spectrum. Statistical tests (Kolmogorov-Smirnov) confirmed these distributions are distinct. Furthermore, a specific sub-cluster of strings ending in "s" was observed; the model pulls these closer to "real word" territory due to the high frequency of pluralization in its training data.

Structural vs. Semantic Information

A significant discovery is the near-orthogonality (86.6°) between the "Information Axis" (structural validity) and the "Concreteness Axis" (semantic meaning). This implies that the model's architecture inherently separates primitive structural information from high-level semantic concepts. CharacterBERT captures a "vibe" of language—a structural coherence—that significantly outperforms traditional statistical Markov models.

Conclusion and Future Work

While this study provides deep insights into character-level representations, it is limited to the English language and the final embedding layer. Future research could investigate how these signals develop across intermediate attention heads and compare model representations with human psychological judgments of "word-likeness." Extending this analysis to diverse media, such as rap lyrics or graffiti, may further illuminate how models distinguish structured signal from noise.

Summary updated

BibTeX
@inproceedings{chu2022signal,
  author = {Mark Chu and Bhargav Srinivasa Desikan and Ethan O. Nadler and Donald Ruggiero Lo Sardo and Elise Darragh-Ford and Douglas Guilbeault},
  title = {{Signal in Noise: Exploring Meaning Encoded in Random Character Sequences with Character-Aware Language Models}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.acl-long.492},
}
Figure