PaperlayerView exact revision on arXiv

Source-linked AI summary

How Do Language Models Represent and Use Phonological Information for Allomorph Selection?

Sangwoo Kim, Sangah Lee

arXiv:2609.04708v1cs.CL

TL;DR

Language models trained on tokenized text can produce phonologically conditioned forms, but it is unclear whether they memorize item-specific patterns or generalize phonologically and how that process works. This paper combines embedding analysis, token-level wug tests, and mechanistic probing to study English a/an and related phenomena. It finds a causal, forecast-mediated phonological mechanism while distinguishing generation-time behavior from explicit phonological judgment.

  • Problem

    Because tokenization obscures sound structure, it remains unclear whether language models’ phonologically conditioned allomorph choices reflect memorization or rule-like generalization, and how such generalization is implemented.

  • Method

    The study analyzes trigger-token embeddings, applies token-level wug tests and causal steering, and uses FutureLens to trace forecasts underlying article selection across languages and tasks.

  • Results

    The phonological condition for English a/an is encoded in a single embedding direction that causally shapes article generation through an internal forecast of the upcoming trigger token, with analogous effects beyond English and a dissociation from explicit sound judgments.

  • Takeaways & Limitations

    The findings provide a mechanistic account of phonologically conditioned allomorph selection in language models and separate generation-time ability from explicit metalinguistic judgment.

  • Takeaways & Limitations

    The mechanism is inferred through an auxiliary FutureLens probe, so the study does not identify the exact computations implementing it.

Abstract

from arXiv · show

Language models are trained on tokenized text that obscures the sound structure of words, yet they reliably produce morphemes whose form is phonologically conditioned. It remains unclear whether they rely on item-specific memorization or rule-like generalization and, if the latter, how that generalization is implemented. We therefore ask whether this phonological condition is represented within language models and how it is causally used for allomorph selection. For the English indefinite article a/an, we show that the phonological condition is encoded along a single linear direction in trigger-token embeddings, that this direction causally drives article selection in token-level wug tests, and that, at the article-prediction position, the model forecasts the upcoming trigger token and uses the forecasted trigger's phonological feature to choose the article. We then ask whether this rule-like generalization extends beyond English article selection, both to allomorph selection in other languages and to explicit phonological judgment. Together, these results provide a mechanistic account of phonologically conditioned allomorph selection in language models, and dissociate this generation-time ability from explicit metalinguistic judgments.

1 Introduction

The paper asks how language models represent and use phonological information despite training on tokenized text, distinguishing memorization from rule-like generalization. It develops a mechanistic account of phonologically conditioned allomorph selection and examines its extension beyond English a/an.

  • Motivation: LLMs trained on raw tokenized text lack explicit access to word sounds, yet produce phonologically conditioned morphemes.This creates uncertainty about whether their behavior reflects item-specific memorization or rule-like generalization.
  • Research question: The study tests whether models represent a phonological condition shared across tokens and use it for allomorph selection.
  • Contributions: For English a/an, the relevant condition is encoded along a single direction in trigger-token embedding space.
  • Contributions: Token-level wug tests assess whether this direction causally controls article selection for novel trigger tokens.
  • Contributions: The analysis traces article choice to forecasts of upcoming trigger tokens and extends the investigation to other languages and explicit phonological judgments.It reports analogous causal effects in Korean, Turkish, Italian, and French allomorphy, alongside a dissociation from explicit sound judgments.

2 Related Work

Prior work has shown that language models can encode sub-token linguistic information or answer phonological questions, but it has not established how standard models use phonological conditions during ordinary generation. This study combines representation analysis with causal intervention to address that gap.

  • Sub-token structure: Standard language models receive tokens that do not explicitly expose orthographic, phonological, or morphological structure.
  • Prior approaches: Some studies modify inputs or training to provide bytes, characters, phonemes, IPA, or phonological features.These approaches show that neural models can access and sometimes represent phonological structure when it is made available.
  • Prior approaches: Metalinguistic benchmarks test whether models answer direct questions about linguistic form, leaving ordinary generation mechanisms unclear.
  • Present contribution: This study identifies a phonological direction in token embeddings and tests whether it causally contributes to allomorph selection.

3 Preliminaries

The paper frames phonologically conditioned allomorphy as productive variation in morpheme form determined by surrounding sound structure. It studies English a/an in decoder-only Transformers, whose next-token distributions are computed from final-layer hidden states through an unembedding map.

  • Phonologically Conditioned Allomorphy: Phonologically conditioned allomorphy occurs when a morpheme’s surface form is selected by surrounding sound structure and can apply to novel words.
  • English a/an: English a appears before consonant-initial words and an before vowel-initial words, including cases where spelling and pronunciation conflict.The following word-start token whose initial sound conditions the choice is called the trigger token.
  • Notation and Setup: The study analyzes decoder-only Transformers with L transformer layers, hidden dimension d, and token embeddings in R^|V|×d.
  • Notation and Setup: At timestep t, the model computes a next-token distribution P_t by applying an unembedding map to the final transformer-layer state.
  • Analysis plan: The English analysis identifies a phonological direction, tests its causal use in article prediction, and examines how the information is routed through three instruction-tuned models.

4 A Phonological Direction in Token Embedding Space

The paper tests whether phonological information for English a/an is represented in token embeddings as a rule-like distinction rather than orthographic or item-specific information. Across exception-token tests and dimensionality analyses, the distinction generalizes strongly and is concentrated in one dominant direction.

  • The study asks whether a/an-relevant phonological information is represented in token embedding space and how that representation is organized.
  • Identifying a Phonological Direction: The classifier is trained on tokens where orthographic and phonological labels agree, then tested on exception tokens where spelling predicts the wrong allomorph.Examples include hour, honor, and heir versus one, unique, and university.
  • Identifying a Phonological Direction: The learned direction exceeds 99.8% cross-validation accuracy across all three models and remains highly accurate on held-out orthography–phonology exceptions.Because spelling predicts the wrong article for these exceptions, the result indicates a phonological rather than merely orthographic distinction.
  • Testing the Dimensionality of the Representation: After the first direction is projected out, a second classifier should remain predictive only if the phonological information is distributed across additional directions.The residual test uses the same train/test split and evaluates whether information remains beyond d1.
  • Testing the Dimensionality of the Representation: Across all three models, residual classifiers and a nonlinear MLP probe remain at or below chance after d1 is removed, indicating a one-dimensional encoding.The class imbalance means chance is model-specific rather than fixed at 50%.

5 Causal Use in Token-Level Wug Tests

The token-level wug test varies only a nonce trigger token’s component along the phonological direction d, testing whether that direction causally controls a/an selection rather than merely being decodable. Across models, moving the nonce embedding from consonant-start toward vowel-start values shifts article preference, with a stronger preference for a than an.

  • Test design: The test varies nonce trigger embeddings only along d, isolating the phonological feature while keeping controlled variants of each token otherwise fixed.Nonce bases are sampled from ordinary trigger-token embeddings, their d component is removed, and a chosen amount is added back before insertion into a reserved nonce token.
  • Test design: Nonce tokens are inserted into a Chinese-to-English translation template, and the model’s next-token probabilities for a and an are measured after the English prefix.The template leaves the article choice unspecified by the Chinese source sentence.
  • Results: The nonce-token sweep strongly shifts article prediction in all three models, from a preference near the consonant-start anchor toward an near the vowel-start anchor.The plotted quantity is P(an)−P(a), averaged over 500 nonce bases; α is normalized from the consonant-start class mean to the vowel-start class mean.
  • Results: Because the nonce token is held fixed except for its d component, the shift shows that the phonological direction causally affects allomorph choice.The controlled embedding intervention distinguishes causal use from mere decodability.
  • Results: The effect reflects productive phonological generalization rather than lexical recall because the nonce token never appeared in training.The sweep also shows a stronger preference for a than an, consistent with English’s consonant-initial base-rate skew.

6 Forecasting the Trigger During Article Prediction

FutureLens tests whether the article-prediction state forecasts the upcoming trigger token, then maps that forecasted phonological information back to test its causal role in choosing a or an. The probe predicts triggers strongly in late layers, and steering the mapped direction flips most article choices.

  • 6.1 Forecasting the Trigger from the Prediction Position: FutureLens maps the article-prediction state to a forecast of the upcoming trigger token and compares its decoded distribution with the article-position output.The method fits layer-specific maps from intermediate source states to future states and evaluates trigger-token predictions.
  • 6.1 Forecasting the Trigger from the Prediction Position: The probe is substantially more predictive for a/an examples than matched random-token controls, reaching around 90% peak top-5 agreement across models.This pattern occurs in late layers, where the article-prediction state contains a forecast of the following trigger token.
  • 6.2 Steering the Forecasted Phonological Feature: The FutureLens-transposed phonological direction maps the trigger-token feature back to the article-prediction position for causal intervention.Orthogonality makes the linear map invertible through its transpose, while weight tying permits direct use of the embedding-space phonological direction.
  • 6.2 Steering the Forecasted Phonological Feature: Steering baseline-a examples toward the an class mean and baseline-an examples toward the a class mean flips roughly 84–90% of article choices at peak layers.The intervention is evaluated on held-out examples selected by the model’s unmodified top-1 article prediction.
  • 6.2 Steering the Forecasted Phonological Feature: Together, the results indicate that the model forecasts the upcoming trigger and uses its phonological feature to determine the appropriate article.The causal steering result links the forecasted phonological direction to article selection rather than only to representational decoding.

7 Generalization Beyond English Article Selection

The learned causal direction generalizes across four non-English allomorph systems, while its effect is weak and inconsistent in explicit sound judgments. FutureLens analyses also find trigger forecasting for Italian and French allomorph selection.

  • 7.1 Cross-Linguistic Allomorph Selection: The study tests Korean, Turkish, Italian, and French allomorph systems spanning different language families, morphological types, trigger positions, and surface-form inventories.The conditioning feature is binary, although Turkish and French have more than two surface forms.
  • 7.1 Cross-Linguistic Allomorph Selection: Across all four languages, the learned embedding direction produces a clear top-1 shift toward the corresponding allomorph despite language-specific baseline preferences.The intervention follows the token-level wug protocol and varies only the nonce trigger’s component along the learned direction.
  • 7.1 Cross-Linguistic Allomorph Selection: For Italian and French, FutureLens analysis finds that the model forecasts the upcoming trigger and uses its phonological feature to select the allomorph.These cases test the forecasting mechanism when the trigger follows the allomorph.
  • 7.2 Explicit Sound Judgments: Explicit sound judgments show much weaker and less consistent effects of the English phonological direction, without a corresponding monotonic shift toward the expected answer.The prompts ask whether nonce nouns begin with consonant or vowel sounds while sweeping the same direction component used for article interventions.
  • 7.2 Explicit Sound Judgments: The results support a dissociation between phonological information used for allomorph generation and metalinguistic reports about sound.The direction reliably affects generation tasks but is not reliably exposed through explicit sound judgments.

8 Conclusion

The paper concludes that language models represent and causally use phonological information for allomorph selection through an internal trigger forecast. This generalization appears across languages and model families, while its relationship to explicit sound judgments remains open.

  • Conclusion: The phonological feature distinguishing English a and an is encoded as a one-dimensional embedding direction that causally shapes article generation.The model deploys this direction through an internal forecast of the upcoming trigger token.
  • Conclusion: The same rule-like generalization appears across multiple model families and languages, extending beyond English a/an.The paper presents this as a generalization for phonologically conditioned allomorphy rather than a language-specific effect.
  • Conclusion: The token-level wug protocol and inverse FutureLens mapping provide extensible tools for probing and causally testing latent linguistic features.The proposed extensions include higher-dimensional subspaces, other nonce constructions, and features decoded from future forecasts.
  • Conclusion: The connection and dissociation between generation-time phonological information and explicit sound judgments require more detailed analysis in future work.The conclusion identifies this relationship as an open direction rather than resolving it fully.

Limitations

The study is limited to conventional decoder-only language models and relies on corpus-derived phonological labels. Its forecasting mechanism is inferred through an auxiliary probe rather than identified at the circuit level.

  • Model and tokenizer scope: The analysis covers decoder-only models with conventional tokenizers, excluding BLT-style architectures and multimodal models processing speech with text.The authors identify extending the approach to these architectures as future work.
  • Label construction: Phonological labels are derived from empirical corpus patterns rather than observed pronunciations, so retained items may imperfectly represent phonological classes.Strict filtering does not eliminate possible noise in the corpus-derived labels.
  • Mechanistic scope: The proposed forecasting mechanism is inferred with an auxiliary FutureLens probe and does not identify the exact computations implementing it.The authors describe a more circuit-level account as an important direction for future work.

A Experimental Details for Section 4

Section 4 constructs phonological labels from article-following corpus tokens, fits a direction in embedding space, and tests its stability and residual information. Across models, the direction captures the phonological distinction while residual probes fall to chance-like performance.

  • Corpus and token filters: Tokens were counted after single-token a or an forms in English Wikipedia, then filtered by frequency, article-association purity, and lowercase alphabetic word-start criteria.Orthographic labels use the first stripped character, with a, e, i, o, u treated as vowel letters.
  • Direction fitting: A logistic classifier fit to language-model output-embedding rows defines direction d1, with positive scores corresponding to vowel-initial tokens.The classifier uses L-BFGS optimization with C = 1.0 and up to 1,000 iterations.
  • Stability analysis: Direction estimates are refit across 200 bootstrap training resamples and 50 redrawn five-fold splits to assess sensitivity to token and fold sampling.The procedure tests dependence on training-set composition and cross-validation partitioning.
  • Residual directions: Residual classifiers are trained after projecting out d1, with the search continuing for up to 20 residual directions and stopping after five-fold accuracy falls below the criterion.The residual procedure repeats projection before fitting each subsequent direction.
  • Chance calibration: Residual chance levels are estimated by label permutation because the training classes are imbalanced, yielding model-specific null accuracies rather than assuming 50%.The null cross-validation accuracies are 61.2±0.6% for Llama, 59.4±0.6% for Qwen, and 56.9±0.6% for Gemma.
  • Results: Cross-validation accuracy exceeds 99.8% in every model, while exception-set accuracy remains high and residual classifiers fall below their entire permutation null distributions.An MLP probe on residual embeddings also stays near chance, supporting concentration of the feature in one direction.

B Experimental Details for Section 5

Section 5 validates the direction with nonce embeddings and token-level wug tests, then connects article decisions to forecasted trigger tokens through FutureLens and counterfactual steering. The experiments include strict matched splits, orthogonal-map checks, and a larger untied model.

  • Token-level wug test: The wug test sweeps nonce embeddings across a normalized direction whose endpoints represent consonant-start and vowel-start class means.Rates are computed over 500 nonce bases in a Chinese-to-English translation template.
  • Nonce validation: Nonce bases match the trigger-token distribution in centroid distance while showing lower nearest-vocabulary cosine, indicating proximity without collapse onto existing lexical items.The validation compares sampled nonce bases with real trigger tokens using both measures.
  • Token-level wug test: Near the consonant-start anchor, almost all nonce bases select a; near the vowel-start anchor, most select an, while other top-1 tokens remain rare.This categorical shift shows article choices track the steered embedding direction.
  • Forecasting setup: Strict FutureLens splits retain examples where the target and following tokens are top-1 predictions, using 90,000 training and 10,000 held-out examples per condition.The a/an split is balanced before splitting, with 50,000 examples per article form.
  • Orthogonal FutureLens: Orthogonal FutureLens maps preserve late-layer top-5 forecasting accuracy, so the orthogonality constraint does not explain away the forecasting effect.The comparison is made on matched strict a/an runs across Llama, Qwen, and Gemma.
  • Counterfactual steering: Steering toward the an class mean increases P(an) and decreases P(a), with the reverse probability shift when steering toward the a class mean.The probability changes occur in the late layers where top-1 flips are most frequent.
  • Larger untied model: In Llama-3.3-70B-Instruct, trigger forecasting peaks at 81.8% top-5 agreement versus 45.1% for the random-token control, while steering flips 86.1% of held-out article choices at layer 78.The effects reappear in a model more than twenty times larger than those in the main analysis and do not require weight tying.

F The Forecasting Mechanism in French and Italian

In French and Italian, where the allomorph precedes its trigger, FutureLens detects the upcoming trigger and steering along its phonological direction changes allomorph choices. The results support a mechanism analogous to English in these two languages.

  • Motivation and setup: French and Italian are tested because their triggers follow the allomorph, whereas Korean and Turkish provide no upcoming trigger to forecast.The study repeats the English forecasting and steering experiments in French and Italian.
  • Forecasting results: French reaches 94.8% top-5 trigger agreement versus 57.0% for its control, while Italian reaches 91.4% versus 36.9%.French peaks in the final layer and Italian at layer 31.
  • Steering results: Counterfactual steering flips 74.9% of held-out French allomorph choices and 66.3% of Italian choices in the final layer.For French, directional rates are 71.1% from le/la to l’ and 78.7% in the reverse direction.
  • Layerwise pattern: In both languages, forecasting is accurate by middle layers, while the steering effect grows sharply only in the final layers.This timing parallels the reported English mechanism.
  • Conclusion: The allomorph-prediction position carries a forecast of the upcoming trigger, and a single direction within that forecast causally affects the allomorph decision.The authors interpret these findings as evidence for a similar mechanism beyond English a/an.
Loading 2609.04708v1…