TL;DR
This paper challenges the long-held view that Japanese verbs are simply memorized, demonstrating through computational modeling and a novel bidirectional "wug" test that Japanese speakers can productively inflect verbs. The study reveals that the direction of inflection (present-to-past vs. past-to-present) significantly impacts both computational complexity and human accuracy, suggesting a nuanced, direction-dependent learning process.
What Problem Does This Paper Address?
For decades, "wug" tests, which assess a speaker's ability to apply morphological rules to novel words, have suggested that Japanese verbs are primarily stored in the mental lexicon rather than being generated through productive rules. This conclusion stems from the observation that Japanese speakers struggle to inflect novel present-tense verbs into past-tense forms. However, this research questions the implicit assumption that present-tense forms are the "base" for inflection, particularly in morphologically rich languages like Japanese where both present and past tense forms are complex. The paper investigates whether past-tense forms might serve as the base, and whether inflection can occur productively in either direction (present-to-past or past-to-present).
Methods
The core of this research lies in two main components:
- Minimal Generalization Learner (MGL): A rule-based computational model was used to simulate verbal inflection in Japanese. The MGL was trained on a lexicon of existing Japanese verb pairs in two directions: Present → Past and Past → Present. The model then generated rules based on minimal generalization. Model complexity was evaluated using three metrics: mean confidence score, mean confidence margin, and mean confidence entropy.
- Bidirectional "Wug" Test: A novel experimental design was created to test human intuition. Participants were presented with novel verb roots and asked to inflect them into either past tense (from present) or present tense (from past). This bidirectional approach allowed for a direct comparison of performance across the two inflection directions. Human accuracy was measured by comparing their responses to the MGL's "winning" outputs.
Datasets
- Training Data: The MGL was trained on a lexicon of 1269 existent Japanese verb pairs. V-V compounds and light verb constructions were excluded to avoid inflating the number of particular inflections.
- Stimuli for "Wug" Test: Novel bisyllabic verb roots ("wug" roots) were created by randomly combining two open syllables. These syllables were extracted from the same lexicon used for training. The "wug" roots were designed to avoid resemblance to actual Japanese words or containing problematic phonetic sequences.
Key Insights and Ideas
- Japanese Verbs are Productive: Contrary to previous findings, this study strongly suggests that Japanese verbs can be computed online via productive generalizations, not just memorized.
- Direction Matters: The direction of morphological inflection (Past → Present vs. Present → Past) is not arbitrary. It significantly impacts both the computational complexity for a model and the accuracy of human performance.
- Past-to-Present is "Easier": The Past → Present direction appears to be computationally less complex for the MGL and experimentally more accurate for humans. This suggests that past-tense forms might indeed serve as a more natural base for inflection in Japanese.
- Rule vs. Analogy Hypothesis: The strong correlation between the model and humans in the Past → Present direction, contrasted with a weaker correlation in the Present → Past direction, leads to a hypothesis: the Past → Present direction might be more "rule-based," while Present → Past might lean more towards "analogy."
Key Results
- Model Complexity: The Past → Present direction showed higher mean confidence and margin scores, and lower mean entropy, indicating it was computationally less complex for the MGL.
- Human Accuracy: The current experiment achieved 72% accuracy for the Past → Present direction, significantly higher than the 48% for the Present → Past direction. This Past → Present accuracy is also notably higher than previous "wug" test results in the literature.
- Correlation: A strong correlation (r = 0.534) was found between model confidence scores and human production probabilities in the Past → Present direction, while the correlation was weak (r = 0.2448) in the Present → Past direction.
Limitations and Future Work
- Stimuli Creation: The study acknowledges that different "wug" roots were used for the two directions, which could influence results. Future work should control for this.
- Training Data Realism: The use of complete verb pairs as training data is noted as unrealistic, as human language learning often involves incomplete paradigms (Zipf's law). Future research could explore using child-directed speech (CDS) as training data.
- Source-Oriented Generalizations: The MGL's "product-oriented" learning (mapping forms without explicit roots) is a limitation. Future work should incorporate morphological decomposition to retrieve roots and test source-oriented generalizations.
- Rule vs. Analogy: The paper proposes a hypothesis that Past → Present is rule-based and Present → Past is analogy-based, predicting corresponding ERP components (LAN and N400). This prediction warrants further experimental investigation.
- Machine Learning Models: The authors suggest comparing their rule-based MGL with contemporary analogy-based models like Recurrent Neural Networks (RNNs) to further explore the nature of generalizations.
Summary updated