Abstract
The Right Association (RA) parsing principle suggests that modifiers attach to the nearest possible phrase. While highly effective for simple structures, RA often fails in complex sentences. This paper investigates whether RA is a fundamental cognitive constraint or a byproduct of linguistic "heaviness." By analyzing the Penn Treebank, we demonstrate that the reliability of RA depends on the complexity of the modifier and the preceding argument. Our findings suggest that speakers avoid placing light modifiers after heavy phrases, creating a distributional pattern that parsers exploit as a shortcut.
Introduction
In computational linguistics, Right Association (RA) is a classic heuristic: when a parser encounters a modifier, such as an adverb, it should attach it to the lowest (most recent) possible node in the syntactic tree. For instance, in the sentence "John said that Bill left yesterday," the adverb "yesterday" is typically interpreted as modifying "left" (low attachment) rather than "said" (high attachment).
However, this rule is notoriously inconsistent. While it works for simple adverbs, it frequently leads to incorrect interpretations in more complex structures. This study, authored by Michael Niv, explores whether RA is a hard-wired parsing strategy or an emergent property of the "principle of end weight"—the tendency in English to place longer, information-dense phrases at the end of a sentence.
Methodology
The study utilized a corpus-based approach, examining approximately 1 million words from the Wall Street Journal portion of the Penn Treebank. To test the influence of "heaviness," Niv defined a constituent as heavy if it contained an internal clause (a subject-verb structure).
The analysis focused on sentences where an adverbial could potentially attach to two different sites (low vs. high). Adverbials were categorized into two groups: "light" (single-word adverbs like recently or often) and "heavy" (clausal adverbials like because the economy remains guided...). The researchers then measured the frequency of these adverbials in various positions relative to heavy and light arguments.
The Heavy Barrier and Distributional Patterns
The data reveals a significant interaction between the weight of an argument and the placement of subsequent modifiers. As shown in the table below, single-word adverbials are rarely placed after heavy arguments.
Specifically, when a verb has a heavy argument, a single-word adverb follows it only about 1.8% of the time (5 out of 272 cases). Instead, writers tend to move the adverb to a pre-argument position. This suggests a "Heavy Barrier": the linguistic system avoids placing light elements after heavy ones to prevent "garden path" sentences that are difficult to process.
Key Results: RA Accuracy and Complexity
The effectiveness of Right Association varies dramatically depending on the type of adverbial involved. For single-word adverbs, RA is an exceptionally reliable heuristic, achieving over 99% accuracy. However, for clausal adverbials, the failure rate increases significantly.
As indicated in the results, clausal adverbials attached "high" (defying RA) in 19.2% of cases. This nearly 20-fold increase in the failure rate suggests that RA is not a universal parsing rule but a shortcut that works best for simple, light constituents.
Discussion: The Conspiracy of Generation
The paper posits that the success of RA is a result of a "conspiracy" between speakers and listeners. Writers and speakers naturally organize sentences to minimize processing difficulty, adhering to the principle of end weight. Because they rarely place light words after heavy phrases, the parser can safely assume that any light word it encounters should be attached locally.
Furthermore, "heaviness" may serve as a proxy for discourse "givenness." We tend to use short expressions for information already established in the context and longer, more complex expressions for new information. By ending sentences with "new" (heavy) content, speakers signal structure to the listener.
Conclusion and Future Work
This research demonstrates that Right Association is a byproduct of how sentences are generated rather than an innate parsing constraint. While the study provides a compelling explanation for RA's performance, it is limited by its reliance on formal written prose (WSJ). Future research should investigate these patterns in spontaneous speech and explore more nuanced, information-theoretic definitions of heaviness. Ultimately, these insights could lead to more sophisticated NLP parsers that dynamically adjust attachment preferences based on constituent weight and discourse context.
Summary updated