A Refined Notion of Memory Usage for Minimalist Parsing

Thomas Graf, Brigitta Fodor, James Monette, Gianpaul Rachiele, Aunika Warren, Chong Zhang2015 · DOI

Abstract

Recently there has been a lot of interest in testing the processing predictions of a specific top-down parser for Minimalist grammars. Most of this work relies on memory-based difficulty metrics that relate the shape of the parse tree to processing behavior. We show that none of the difficulty metrics proposed so far can explain why subject relative clauses are more easily processed than object relative clauses in Chinese, Korean, and Japanese. However, a minor tweak to how memory load is determined is sufficient to fully capture the data. This result thus lends further support to the hypothesis that very simple notions of resource usage are powerful enough to explain a variety of processing phenomena.

Introduction: The Subject Preference Paradox

In psycholinguistics, the Subject Preference is a well-documented phenomenon: Subject Relative Clauses (SRCs), such as "the reporter [who attacked the senator]," are generally easier for humans to process than Object Relative Clauses (ORCs), such as "the reporter [who the senator attacked]." In English, this is often explained by the linear distance between the filler and the gap. However, in prenominal languages like Chinese, Japanese, and Korean, the relative clause precedes the head noun. In these cases, string-based models often predict an ORC advantage, yet experimental data consistently shows that humans still find SRCs easier to process.

This paper investigates whether a computational model based on Minimalist Grammars (MGs) can resolve this paradox by looking at structural resource usage rather than linear distance.

Minimalist Grammars and Top-Down Parsing

Minimalist Grammars formalize Chomskyan syntax using feature-driven operations like Merge and Move. The authors utilize a top-down parser that maintains a memory queue of predicted nodes. The complexity of a parse is determined by how many items the parser must keep in memory and for how long.

Figure 1: A Minimalist Grammar parse tree illustrating the complex structural dependencies and feature-driven movement (dashed lines) that a top-down parser must track.

Two primary syntactic analyses of relative clauses are considered:

  1. Wh-movement: A silent operator moves from within the clause to its periphery.
  2. Promotion: The head noun itself originates inside the clause and moves to the head position.

Evaluating Memory Metrics

Standard metrics for measuring parsing difficulty include Tenure (how long a node stays in memory), Payload (the number of nodes in memory), and Max (the peak tenure value). In prenominal languages, the parser encounters the relative clause before the head noun, creating a structural "bottleneck." High-level nodes enter memory early but cannot be cleared until the entire clause is built, masking the subtle differences between SRCs and ORCs.

Figure 3: Comparison of Subject (left) and Object (right) Relative Clause structures under a Wh-movement analysis. The structural asymmetry between the subject and object positions is key to the proposed refined metric.

The Refined "Gap" Metric

The authors argue that memory load is not just a function of time or quantity, but also the "size" of memory items. Specifically, a node carrying a pending movement feature (a filler-gap dependency) imposes a higher cognitive load. They introduce the Gap metric, which sums the distances (in parse steps) of all moving phrases. Because subjects occupy a higher structural position than objects, the "gap" for an object is deeper in the tree, requiring the parser to carry movement features through more steps. This increased cost explains why ORCs are harder, even in prenominal languages.

Results and Discussion

The study demonstrates that standard metrics (Max, Box, or Sum) fail to consistently predict the SRC preference across Chinese, Japanese, and Korean. However, when the Gap metric is used as a tie-breaker, the model correctly SRCPredicts the SRC advantage across all tested languages and syntactic theories. Furthermore, the results suggest that processing data can help adjudicate between syntactic theories; for instance, the "Box" metric only aligned with human data under the "Promotion" analysis.

Conclusion

By refining the notion of memory usage to account for the cost of tracking movement features, the authors show that simple resource-based metrics are powerful enough to explain complex cross-linguistic processing phenomena. This result reinforces the idea that the human parser's limitations are intrinsically linked to the formal structural properties of language.

Summary updated

Abstract

Recently there has been a lot of interest in testing the processing predictions of a specific top-down parser for Minimalist grammars Most of this work relies on memory-based difficulty metrics that relate the shape of the parse tree to processing behavior. We show that none of the difficulty metrics proposed so far can explain why subject relative clauses are more easily processed than object relative clauses in Chinese, Korean, and Japanese. However, a minor tweak to how memory load is determined is sufficient to fully capture the data. This result thus lends further support to the hypothesis that very simple notions of resource usage are powerful enough to explain a variety of processing phenomena.
BibTeX
@misc{graf2015refined,
  author = {Thomas Graf and Brigitta Fodor and James Monette and Gianpaul Rachiele and Aunika Warren and Chong Zhang},
  title = {{A Refined Notion of Memory Usage for Minimalist Parsing}},
  year = {2015},
  doi = {10.3115/v1/w15-2301},
}
Figure