1 of 1

QUESTION: What linguistic abstractions support word recognition, beyond surface phonetics?

CASE STUDY: English -er �Three surface-identical sources:

taller, bigger, fancier (comparative, “more X”)
rider, fighter, partier (agentive, “one who Xs”)
whisper, dagger, gossamer (non-morphemic)

�Different phonological distributions:

/-ɚ_AGENTIVE/: rider, partier, manifester, �compliment-receiver → any size base

/-ɚ_COMPARATIVE/: taller✅, fancier✅, *dangerouser❌ → 1-2 syllable base only

APPROACH

We use self-supervised speech models (S3Ms) as interpretable, in-silico proxies for human language processing.

Vector analogy to evaluate representation contents:

COMPARISON: English -z/-s

TAKEAWAYS

Representations optimized for word recognition

sharpen morphological distinctions when they have distributional consequences
discard redundant morpho-phonological information

Models allow us to explore�how task demands shape mental representations in speech perception.

Title:�Subtitle

Word recognition selects for abstract morphological structure in speech models

MODELS TESTED

RESULTS

�

Jon Gauthier & Canaan Breiss

Take a picture to �download the poster

Word model: morpheme category match effect increases

Word model: sensitivity to morphological complexity increases

token of the�same word

token of a�different word

tall : taller :: fight : ____

p_tall = v_tall - v_tall

Average over token-pairs, examine effect of source and target

→

Word model enhances effect of binary morphological complexity (taller and runner vs. whisper), and enhances specific morphological category (taller vs. runner)

Two morphemes: /-z_PLURAL/, /-z_3SG/

Same phonological distributions:

[z] after voiced nonsibilants
[s] after voiceless nonsibilants

Word model: morpheme category match effect decreases

jon@gauthiers.net cbreiss@uchicago.edu

Loss

Wav2Vec word encoding

English Wav2Vec2��contrasts�raw acoustics

Fine-tuned model� contrasts �word types

vs.

( )