Related Experiment Videos
Is deeper always better? Replacing linear mappings with deep learning networks in the Discriminative Lexicon Model
Maria Heitmeier1, Valeria Schmidt1, Hendrik P A Lensch1
1Eberhard Karls Universität Tübingen, Tübingen, Germany.
Linguistics Vanguard : Multimodal Online Journal
|July 24, 2026
Summary
Deep discriminative learning (DDL) improves cognitive models of language by creating more accurate mappings than linear methods for large datasets. However, frequency-informed deep learning (FIDDL) is needed to outperform linear models on reaction times.
Area of Science:
- Cognitive Science
- Computational Linguistics
- Artificial Intelligence
Background:
- Deep learning models are increasingly applied to cognitive modeling of language.
- The study investigates if deep learning surpasses linear methods in understanding the speaker's learning problem.
Purpose of the Study:
- To compare deep discriminative learning (DDL) with linear discriminative learning (LDL) in cognitive language modeling.
- To assess the performance of DDL and frequency-informed deep learning (FIDDL) against linear models.
Main Methods:
- Utilized the Discriminative Lexicon Model, replacing linear mappings with deep dense neural networks (DDL).
- Compared DDL and LDL on English, Dutch, Estonian, and Taiwan Mandarin datasets.
- Evaluated frequency-informed linear mappings (FIL) and FIDDL using average reaction times.
Main Results:
- DDL achieved more accurate mappings for large, diverse English and Dutch datasets, outperforming LDL for pseudo-morphological words.
- FIDDL substantially outperformed FIL, while DDL alone was outperformed by FIL on reaction times.
- Linear mappings are more effective for trial-to-trial incremental lexical learning updates than deep mappings.
Conclusions:
- Deep learning models offer advancements in cognitive language modeling, particularly with frequency-informed training.
- Both linear and deep learning approaches provide valuable insights into language processing and learning.
