Abstract
This paper presents a unified framework for the processing of nominalizations—nouns derived from verbs—within the PUNDIT (Prolog UNDerstands and Integrates Text) system. By leveraging semantic parity between verbs and their nominal counterparts, the system efficiently extracts meaning and resolves references in technical domains. The approach is demonstrated using Navy casualty reports (CASREPS), showing how a two-pass role-filling algorithm and temporal anchoring can overcome the syntactic and contextual challenges inherent in nominalized expressions.
Introduction
In natural language processing, nominalizations represent a significant bridge between syntax and semantics. A sentence like "The pump failed" (a clause) and the phrase "The failure of the pump" (a nominalization) convey nearly identical information, yet they present vastly different challenges for computational systems. While verbs have well-defined syntactic slots for subjects and objects, nominalizations are more fluid, often omitting essential information or obscuring the relationship between the action and its participants.
This work details the integration of nominalization processing into PUNDIT, a system designed to process and understand technical prose. The core thesis is that by treating nominalizations and verbs as semantically identical at a deep level, a system can achieve greater robustness and efficiency in information extraction.
The Problem of Nominalization
Nominalizations present three primary hurdles for NLP systems:
- Syntactic Mapping: Unlike clauses, where the subject typically performs the action, nominalizations allow the "actor" to appear in various forms, such as possessives ("the pump's failure") or prepositional phrases ("failure of the pump").
- Missing Tense: Verbs are inherently tensed, providing a temporal anchor for the event. Nouns lack this feature, requiring the system to infer the timing of a "failure" from the surrounding context.
- Implicit Arguments: Nominalizations often omit participants. In the phrase "Investigation revealed metal contamination," the agent (who investigated) and the specific object (what was contaminated) are left implicit. For a system to fully understand the event, it must recover these missing pieces.
Prior to the PUNDIT approach, many systems treated nominalizations as simple noun-noun compounds, which discarded the rich "who-did-what-to-whom" structure inherent in the related verb.
Methodology: The PUNDIT Framework
PUNDIT addresses these challenges through three integrated strategies:
Unified Predicate Decomposition
The system utilizes a shared semantic template for both a verb and its related nominalization. For example, both the verb "fail" and the noun "failure" map to the same underlying structure: become(inoperative(patient(P))). This ensures that the system's reasoning engine treats the event identically regardless of its grammatical form.
Two-Pass Role Filling
Because nominalizations are syntactically flexible, PUNDIT employs a two-pass algorithm to fill semantic roles:
- Pass 1 (Syntactic Cues): The system first attempts to fill roles using explicit syntactic markers, such as "of" or "in" phrases.
- Pass 2 (Reference Resolution): If roles remain empty, the system uses reference resolution to look for previously mentioned entities. If a prior sentence mentioned "The pump failed," a subsequent mention of "the failure" allows the system to automatically inherit "the pump" as the patient of the event.
Temporal Anchoring
To compensate for the lack of tense in nouns, PUNDIT "borrows" time from the matrix verb or temporal connectives. In the sentence "Investigation revealed...", the investigation is anchored to the past tense of the verb "revealed." Similarly, connectives like "after" or "during" provide the necessary temporal context to place the nominalized event on a timeline.
Evaluation and Results
The system was developed and tested using CASREPS (Casualty Reports), which are technical messages used by the U.S. Navy to report equipment failures. These reports are characterized by telegraphic, jargon-heavy prose, providing a rigorous testbed for the system's robustness.
PUNDIT successfully demonstrated a uniform treatment of diverse syntactic forms. Whether a report utilized a noun-noun compound ("SAC oil pressure") or prepositional phrases ("pressure of oil in SAC"), the system generated the same semantic representation: pressureP(theme(oil), location(sac)).
Key qualitative results include:
- Cross-sentential Linking: The ability to link events across sentences by resolving implicit arguments.
- Compound Handling: Treating noun-noun compounds as a tractable subset of nominalizations.
- Pragmatic Inference: Using domain-specific pragmatics to fill essential roles, such as identifying that a "request for replacement" implies the replacement of a previously identified broken component.
Discussion and Limitations
While the PUNDIT approach offers a powerful method for handling nominalizations, certain limitations remain. The system currently struggles with highly nested modifiers (e.g., "impellor blade tip erosion"), as it assumes modifiers attach directly to the head noun rather than forming complex sub-units. Additionally, the criteria for "essential" roles—such as the requirement that an investigation involves a damaged component—are currently hard-coded for the Navy maintenance domain. Future work will focus on generalizing these criteria and improving the handling of complex modifier structures to enhance the system's portability across different technical domains.
Summary updated