Introduction
Diathesis alternations—the capacity of verbs to express the same semantic content through varying syntactic argument structures—are a cornerstone of linguistic theory. A classic example is the dative alternation, where "John offered shares to his employees" (prepositional) can be expressed as "John offered his employees shares" (double-object). While manual classifications, such as those by Beth Levin, have provided foundational taxonomies, they often rely on linguistic intuition rather than large-scale empirical data. This study presents an automated pipeline to identify these alternations within massive text corpora, aiming to quantify linguistic concepts like "productivity" and "typicality."
Methodology
The research utilizes a multi-stage pipeline to extract and validate verb frames from the 100-million-word British National Corpus (BNC).
- Shallow Parsing and Heuristics: The tool Gsearch was employed to scan for syntactic patterns like
V NP1 NP2. To mitigate "noise" inherent in automated parsing, linguistic heuristics were applied. For example, adjacent proper names were filtered out to avoid misidentifying single entities as distinct arguments. - Compound Noun Detection: To prevent mistaking compound nouns (e.g., "bank manager") for separate arguments, the system integrated WordNet and calculated log-likelihood ratios (G-scores) to assess lexical association.
- PP Attachment and Semantic Filtering: For prepositional phrases, lexical association scores determined whether a PP attached to the verb or a preceding noun. Benefactive PPs (e.g., "for her") were identified by verifying the semantic primitive of the prepositional object in WordNet (e.g., "person" or "group").
- Statistical Filtering: Verbs with low relative frequencies were discarded to ensure the resulting frames were not artifacts of tagging errors.
Results and Evaluation
The automated system demonstrated high reliability across several extraction tasks. The linguistic heuristics achieved a precision of 96.9% for rejecting false matches, while compound noun detection reached near-perfect precision (98.9% to 99.1%).
The pipeline successfully acquired 162 verbs for the double-object frame and identified 10 alternating verbs not previously listed in Levin’s manual classification, such as cause, deliver, and hand.
Discussion: Productivity vs. Intuition
A key contribution of this work is the distinction between Productivity (how many verbs in a class alternate) and Typicality (how balanced a verb's usage is between frames). The corpus data revealed significant discrepancies between manual intuition and empirical usage.
As shown in Figure 1, while "Throwing" verbs are prominent in manual lists, "Future Having" verbs proved more frequent and productive in actual usage. Furthermore, the "GIVE" class was identified as the most typical for dative alternations, whereas "DRIVE" verbs (e.g., ferry) appeared as peripheral, rarely utilizing the double-object format.
Conclusion and Limitations
While the system is highly effective for dative alternations, it faced challenges with "for-PP" attachment, achieving only 36% precision due to the high polysemy of the preposition "for." Future work will explore how these productivity and typicality estimates generalize across different corpora and genres, providing a more robust empirical foundation for lexical semantics.
Summary updated