Rules versus statistics: insights from a highly inflected language.

Area of Science:

• Computational Linguistics
• Psycholinguistics
• Morphology

Background:

• Inflectional morphology is often viewed as rule-governed, particularly based on English's simple system.
• Serbian possesses a complex inflectional system encoding number, gender, and case for nouns, challenging rule-based explanations.
• Previous research has primarily focused on simpler inflectional systems, limiting generalizability.

Purpose of the Study:

• To investigate the representation of inflectional morphology in Serbian, a typologically diverse language.
• To determine if a connectionist network can model the quasi-regular nature of Serbian inflection.
• To explore common computational mechanisms underlying inflectional morphology across languages.

Main Methods:

• Analysis of a large corpus of 3,244 Serbian nouns to identify patterns in inflectional morphology.
• Development and training of a simple connectionist network to predict inflected forms from lemma, gender, number, and case.
• A word-naming experiment with native Serbian speakers to compare model and human performance.

Main Results:

• Serbian inflectional morphology exhibits quasi-regularity with partial regularities and varying neighborhood sizes.
• The connectionist model successfully learned and generalized Serbian noun inflections, mirroring human performance.
• Model performance was influenced by surface/lemma frequency and inflectional neighborhood size, a novel measure also affecting human participants.

Conclusions:

• Generating inflected forms involves satisfying probabilistic constraints relating meaning and form, applicable to both Serbian and English.
• Quasi-regularity and statistical learning are key to understanding inflectional morphology in diverse languages.
• Common computational mechanisms likely underpin inflectional processing across typologically different languages.