TL;DR
This paper investigates how humans process the internal structure of words, specifically whether they parse morphemes into hierarchical structures like sentences or treat them as linear strings. Using computational models and human reaction time data, they found that models incorporating hierarchical structures (PCFG) best explain human behavior, suggesting that morphological processing is more complex than previously thought.
What Problem Does This Paper Address?
While sentences are widely understood to have hierarchical syntactic structures, the internal structure of words (morphology) has been a subject of debate. Some theories propose that words, especially derived ones, also have hierarchical structures, similar to syntax. However, many computational models of morphological processing treat words as simple linear sequences of morphemes, akin to finite-state automata. This approach, often termed "amorphous," overlooks the potential for hierarchical parsing within words. This paper aims to bridge this gap by computationally modeling morphological processing and comparing different representational assumptions (linear vs. hierarchical) against human behavioral data. The core question is whether human morphological processing mirrors sentence processing in its reliance on hierarchical structures.
Methods
The researchers employed five computational models with varying representational assumptions:
- "Amorphous" Models:
- Letter Markov Model: Models transitions between letters.
- Syllable Markov Model: Models transitions between syllables.
- "Morphous" Models:
- Morpheme Markov Model: Models transitions between morphemes.
- Hidden Markov Model (HMM): Models transitions between hidden Part-of-Speech (POS) tags and emissions of morphemes.
- Probabilistic Context-Free Grammar (PCFG): Models hierarchical syntactic structures of morphemes.
These models were trained on morphological tree structures from the CELEX corpus. To evaluate their performance, the researchers used a linking hypothesis of "cumulative surprisal," which posits that processing difficulty is proportional to the accumulated surprisal of morphemes encountered sequentially. This metric was then used to predict human reaction times (RTs) from the English Lexicon Project (ELP) using ordinary linear regression. Control predictors like word length, syllable count, neighborhood density, and surface frequency were also included in the regression models.
Datasets
- CELEX: Used for annotating morphological tree structures to train the computational models. The revised CELEX (O'Donnell, 2015) was used, focusing on morphologically complex derived and monomorphemic words.
- English Lexicon Project (ELP): Used to obtain human reaction times (RTs) from visual lexical decision experiments. A restricted version of ELP was used, ensuring words had available RTs and control predictors.
The simulation data was created by intersecting CELEX and ELP, resulting in 13,244 morphologically complex derived and monomorphemic words. Monte Carlo cross-validation (MCCV) was used for training and testing splits.
Key Insights and Ideas
- Morphemes Matter: "Morphous" models (those that explicitly consider morphemes) significantly outperformed "amorphous" models (which treat words as sequences of letters or syllables). This strongly suggests that humans do indeed process words by breaking them down into their constituent morphemes.
- Hierarchy is Key: The PCFG model, which explicitly models hierarchical syntactic structures within words, achieved the highest psychological accuracy. This indicates that human morphological processing is not just about linear sequences of morphemes but involves parsing them into more complex, hierarchical arrangements, similar to sentence processing.
- Syntactic Granularity Maps to Processing: The study found a strong correlation between linguistic accuracy (how well a model predicts grammatical structure) and psychological accuracy (how well it predicts human reaction times). This suggests that the level of structural representation (hierarchical vs. linear) directly impacts processing efficiency.
- Incremental Processing: The success of the PCFG model, particularly when combined with a probabilistic Earley parser (which processes incrementally), suggests that morphological parsing might also occur incrementally from left to right, despite the inherently non-incremental nature of visual word recognition.
- Storage and Computation Coexist: The results showed that both surface frequency (often interpreted as evidence for storage of whole words) and cumulative surprisal (reflecting computation from morphemes) were significant predictors of reaction times. This supports models where both storage and computation play a role in morphological processing.
Key Results
- Morpheme-based models were psychologically more accurate than amorphous models. All "morphous" models were statistically significant, while only one of two "amorphous" models reached significance.
- PCFG was the most psychologically accurate model. Its accuracy was significantly higher than HMM and Morpheme Markov Model.
- PCFG was also the most linguistically accurate model.
- Surface frequency effects were robustly observed across models, but PCFG explained unique variance not accounted for by surface frequency alone.
Limitations and Future Work
- Limited to Bimorphemic Words: The study primarily used bimorphemic words for testing. The authors acknowledge that bimorphemic words may not be sufficient to distinguish between hierarchical structures and linear strings, and suggest that future work should evaluate models on trimorphemic words to ensure generalizability.
- Investigating Other Models: The paper mentions that the Cohort Model might harmonize with the idea of probabilistic parsing and incremental processing, suggesting this as a future avenue for investigation.
- Distinguishing Storage vs. Computation: While the study found evidence for both storage (surface frequency) and computation (surprisal), the authors note that simple RTs are an "end-point" measure and cannot definitively distinguish between dual-route models of morphological processing. Further research might be needed to disentangle these processes more clearly.
Summary updated