A Computational Framework for Non-Lexicalist Semantics

Jimmy LinWorkshops @ NAACL 2004 · OpenReview

Abstract

Under a lexicalist approach to semantics, a verb completely encodes its syntactic and semantic structures, along with the relevant syntax-to-semantics mapping; polysemy is typically attributed to the existence of different lexical entries. A lexicon organized in this fashion contains much redundant information and is unable to capture cross-categorial morphological derivations. The solution is to spread the "semantic load" of lexical entries to other morphemes not typically taken to bear semantic content. This approach follows current trends in linguistic theory, and more perspicuously accounts for alternations in argument structure. I demonstrate how such a framework can be computationally realized with a feature-based, agenda-driven chart parser for the Minimalist Program.

Introduction

Traditional computational linguistics often relies on a "lexicalist" approach where verbs serve as the primary locus of semantic information. Under this paradigm, polysemous verbs like break require multiple distinct lexical entries to account for causative ("John broke the window"), inchoative ("The window broke"), and instrumental ("John broke the window with a hammer") uses. This leads to significant redundancy and a failure to capture cross-categorial morphological relationships, such as the link between the adjective flat and the verb flatten.

This paper proposes a shift toward a compositional framework based on Distributed Morphology (DM). By spreading the semantic load across morphemes and functional heads, we can derive complex meanings dynamically through syntax rather than relying on exhaustive lexical lookups.

Methodology: Distributed Morphology and the Minimalist Program

The proposed framework decomposes events into primitive predicates: DO (activities), BECOME (changes of state), and BE (static states). These are introduced by "verbalizing heads" (light verbs like $v_{DO}$ or $v_{BE}$) that assign thematic roles based on their position in the syntactic tree.

The implementation utilizes a feature-based, agenda-driven chart parser for the Minimalist Program. The core operation is MERGE, which combines items with matching features (e.g., a determiner seeking a noun) to simultaneously construct the syntactic structure and the corresponding logical semantic expression.

Syntactic and semantic derivation of an activity event.

Key Insights: Syntax as Semantics

In this framework, the structure of the sentence is the meaning. There is no separate mapping step; the logical form is a direct byproduct of the syntactic derivation. This allows the system to handle "template augmentation," where simple events are combined into complex ones (e.g., "sweeping the floor clean") without requiring specialized lexical entries.

Furthermore, by treating derivational affixes like -en as active syntactic participants, the model can derive meanings for words it has not explicitly seen before, provided the root and suffix are known. This process often involves "head movement," where a root moves to attach to its suffix during derivation.

Results: Proof-of-Concept Derivations

The author demonstrates the framework's efficacy through a "toy parser." A primary success is the step-by-step derivation of sentences involving change-of-state morphology.

Derivation of the change-of-state event "The tire flattened," showing the integration of the root "flat" and the suffix "-en".

As shown in Figure 5, the system correctly composes the meaning from individual morphemes (tire, flat, and -en). This demonstrates lexical efficiency: a small set of light verbs and features can account for a wide range of argument alternations that would otherwise bloat a traditional lexicon.

Limitations and Future Directions

While theoretically elegant, the current implementation is a proof-of-concept. Scaling this approach to real-world text requires high-quality, fine-grained lexicons, which are currently labor-intensive to build. Future work focuses on lexical induction—automatically learning morphological and syntactic features from raw data to bridge the gap between theoretical linguistics and large-scale statistical NLP.

Summary updated

Abstract

Under a lexicalist approach to semantics, a verb completely encodes its syntactic and semantic structures, along with the relevant syntax-to-semantics mapping; polysemy is typically attributed to the existence of different lexical entries. A lexicon organized in this fashion contains much redundant information and is unable to capture cross-categorial morphological derivations. The solution is to spread the "semantic load" of lexical entries to other morphemes not typically taken to bear semantic content. This approach follows current trends in linguistic theory, and more perspicuously accounts for alternations in argument structure. I demonstrate how such a framework can be computationally realized with a feature-based, agenda-driven chart parser for the Minimalist Program.
BibTeX
@inproceedings{lin2004computational,
  author = {Jimmy Lin},
  title = {{A Computational Framework for Non-Lexicalist Semantics}},
  booktitle = {Workshops @ NAACL},
  year = {2004},
  url = {https://openreview.net/forum?id=B1NDCzbOWS},
}
Figure