The Acquisition of Conceptual Structure for the Lexicon

James Pustejovsky, Sabine BerglerAAAI 1987

Introduction

In the late 1980s, Natural Language Processing (NLP) faced a significant challenge known as the "lexical bottleneck." While machine-readable dictionaries (MRDs) provided a vast repository of data, they often lacked the structured semantic information necessary for deep language understanding. This paper introduces TULLY, an interactive machine learning system designed to bridge this gap by transforming unstructured dictionary definitions into formal semantic representations.

The Extended Aspect Calculus (EAC)

The core of the TULLY system is the Extended Aspect Calculus (EAC), a formal semantic theory that provides a structured "bias" for the learning process. The EAC decomposes verb meanings into three primary fields: causal, spatial, and aspectual.

The aspectual field is particularly crucial, as it categorizes verbs based on their temporal properties. This categorization follows a hierarchy of event types, distinguishing between states, processes, and accomplishments.

The hierarchy of event types in the Extended Aspect Calculus.

By using this hierarchy, TULLY can systematically narrow down a verb's meaning through a series of targeted questions to a human "critic," such as whether an action has a logical endpoint or is a continuous process.

The TULLY System and Thematic Mapping

TULLY utilizes a core data structure called the Thematic Mapping Index (TMI) to represent the relationship between a verb's syntactic arguments and its semantic roles. The TMI maps subjects and objects to roles like Causer, Theme, or Goal across specific temporal intervals.

An example of a Thematic Mapping Index (TMI) for the verb 'give'.

The learning algorithm proceeds in several stages:

  1. Primitive Selection: The user identifies core concepts (e.g., cause, move).
  2. Candidate Search: TULLY scans dictionary definitions (such as those from the American Heritage Dictionary) for these primitives using specialized "head-finding" algorithms.
  3. Interactive Credit Assignment: The system generates hypotheses and asks the user clarifying questions (e.g., "Is the subject animate?") to refine the TMI.
  4. Generalization: The refined TMI is integrated into a "tangled hierarchy" of concepts, allowing for property inheritance from multiple parent concepts.

Results and Insights

The authors demonstrated TULLY's capabilities by having it learn verbs like propel and throw. The system successfully identified propel as a process involving a causer and a moving object. By interacting with a human user, it correctly classified the verb within the aspectual hierarchy. Furthermore, the system showed how specialized sub-routines, or "thematic specialists," could identify incorporated instruments, such as "the arm" in the definition of throw.

This approach highlights the power of using linguistic theory as a learning bias. Instead of learning from scratch, the system fills in a pre-defined template, making the acquisition of complex lexical knowledge significantly more efficient and manageable for a computer.

Conclusion

TULLY represents a middle ground between manual hand-coding and fully automatic lexical acquisition. By combining formal semantic models with human-in-the-loop interaction, the system can build a lexicon that is richer and more structured than standard digital dictionaries. Future work aims to scale this approach to larger MRDs like the Longman Dictionary of Contemporary English (LDOCE) and improve automated word sense disambiguation.

Summary updated

BibTeX
@inproceedings{pustejovsky1987acquisition,
  author = {James Pustejovsky and Sabine Bergler},
  title = {{The Acquisition of Conceptual Structure for the Lexicon}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1987},
}
Figure