Semantic Interpretation of Nominalizations

AAAI 1996

TL;DR

This paper presents a computational framework for semantically interpreting nominalizations, which are nouns derived from verbs (e.g., "decoration" from "decorate"). The authors tackle the challenge of disambiguating the intended meaning of these words and correctly identifying their arguments, crucial for deeper natural language understanding. Their approach leverages existing verb knowledge and introduces novel algorithms to handle the unique syntactic structures of nominalizations.

What Problem Does This Paper Address?

Natural language often uses nominalizations to express actions, events, or concepts. For example, "The king sent another fleet to break the Muslims' control over spice" uses "control" as a nominalization. Understanding such sentences requires more than just recognizing the words; it involves grasping the underlying meaning. This paper focuses on two key challenges:

  1. Disambiguating Nominalizations: Many nominalizations have both verbal (action-oriented) and non-verbal (object-oriented) senses. For instance, "decoration" can refer to the act of decorating or an ornamental object. Deciding which sense is intended is critical.
  2. Interpreting Nominalization Modifiers: Nominalizations often appear in noun phrases (NPs) with various modifiers (genitives, adjectives, prepositional phrases). These modifiers can represent the arguments (like subject or object) of the original verb. The paper aims to correctly map these modifiers to the thematic roles of the underlying verbal concept.

The authors highlight that existing work on word sense disambiguation often doesn't specifically address the nuances of nominalizations, particularly how their modifiers relate to the original verb's semantics.

Methods

The core of the paper is a computational framework for semantic interpretation, with specific algorithms designed for nominalizations. The approach is built upon:

  • Leveraging Verb Knowledge: The system reuses existing knowledge about verbs (represented as "verbal concepts" or "VM rules") to interpret their nominalized forms. This significantly reduces the need for new knowledge specific to each nominalization.
  • Three Key Algorithms:
    1. Nominalization Noun Phrase Algorithm: This algorithm is triggered when a noun is identified as a potential nominalization. Its goals are to determine the underlying verbal concept and identify which nouns/adjectives within the NP fill its thematic roles. It uses heuristics and mapping rules to handle how verb arguments (subject, object) are expressed as genitives, adjectives, or prepositions in NPs.
  • Prepositional Attachment and Meaning Determination Algorithm: This algorithm handles prepositional phrases (PPs) that follow a nominalization. It determines if the PP attaches to the nominalization and what its semantic role is, using the verbal concept's rules and selectional restrictions.
  • End-of-Clause Algorithm: This algorithm is a fallback mechanism. It attempts to determine the verbal concept if it's still unknown at the end of a clause. It also re-evaluates modifiers to ensure they've been interpreted. It employs heuristics to distinguish between verbal and non-verbal senses when other methods fail.
  • WordNet Integration: The system uses WordNet to identify potential nominalizations by checking if noun senses are hypernyms of actions or events. It also uses WordNet's collocation lists to help disambiguate nominalizations.

Datasets

The algorithms were tested on a corpus of biographical articles from the World Book Encyclopedia. This domain was chosen because it is used in an ongoing project to acquire historical knowledge from encyclopedic texts. The authors selected ten nominalizations that appeared at least 20 times in 5000 articles for their evaluation. No new benchmarks were introduced; rather, existing text was used for evaluation.

Key Insights and Ideas

  • Reusing Verb Knowledge is Key: The most significant insight is that a substantial amount of knowledge about verbs can be directly reused for nominalizations, greatly simplifying the process. This is achieved by mapping nominalization modifiers to the verbal concept's thematic roles.
  • Modifiers Reveal Meaning: The presence and type of modifiers (genitives, PPs) are crucial clues for disambiguating both the nominalization's sense and the underlying verb's meaning. For example, "her defense in court" vs. "the defense of Richmond."
  • Obligatory Roles for Verbal Sense: For nominalizations with both verbal and non-verbal senses, requiring certain thematic roles (like "theme" for "defense") to be present helps identify the verbal sense.
  • Handling Syntactic Differences: Nominalizations don't take direct objects like verbs do; their arguments are expressed differently (e.g., genitives for subjects/objects). The algorithms explicitly address these syntactic transformations.

Key Results

The algorithms were tested on ten nominalizations, achieving the following average results across 1247 occurrences:

  • Nominalization Sense Disambiguation: 88% accuracy. The worst performance was 72% for "control," which has 11 WordNet senses.
  • Genitive Role Interpretation: 93% accuracy.
  • NP Element Interpretation (excluding genitives): 71% accuracy. This was noted as a limitation, particularly for adjectives describing manner rather than filling thematic roles.
  • Prepositional Phrase Attachment and Meaning: 96% accuracy. This indicates that the verb's PP handling mechanisms are effectively transferable.

Limitations and Future Work

  • Limitations:
    • Adjective Interpretation: The interpretation of adjectives that describe manner or are not directly related to thematic roles remains a challenge (71% accuracy for NP elements).
  • Ambiguous PP Heads: Failures in PP interpretation sometimes occurred due to ambiguous PP heads or when nominalizations used prepositions different from their root verbs.
  • Hand-Crafted Knowledge: The system relies on hand-crafted representations of VM rules, verbal concepts, and a general ontology.
  • Future Work:
    • Integrating WordNet: The authors are working on a parallel project to integrate WordNet's lexical knowledge base, which they believe will simplify the definition of VM rules and verbal concepts by interfacing with WordNet's noun ontology and verb hierarchy.
  • Handling More Complex Structures: While the paper focuses on nominalizations as head nouns, future work could extend to nominalizations in other positions.

Summary updated

BibTeX
@inproceedings{1996semantic,
  title = {{Semantic Interpretation of Nominalizations}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1996},
}
Figure