Conceptual Association for Compound Noun Analysis

Mark LauerACL 1994

TL;DR

This paper tackles the challenge of automatically interpreting compound nouns (like "pottery coffee mug") by using statistical associations between thesaurus categories. They developed a method that achieves 75% accuracy in disambiguating the syntactic structure of compound nouns, paving the way for deeper semantic understanding.

What Problem Does This Paper Address?

Understanding compound nouns (CNs) is a surprisingly tricky problem in Natural Language Processing (NLP). While simple CNs like "coffee mug" are straightforward, longer ones like "pottery coffee mug holder" can have multiple possible syntactic structures, leading to different meanings. For example, is it a "holder for pottery coffee mugs" or a "coffee mug holder made of pottery"? Traditional syntactic parsers struggle with this ambiguity because the attachments aren't always dictated by strict grammatical rules. This paper aims to automatically derive the most probable syntactic structure for any given compound noun, even without explicit contextual clues.

Methods

The core of this research lies in leveraging conceptual association derived from corpus statistics. They start by extracting a large number of compound nouns from an online encyclopedia. To overcome the "acquisition problem" (getting enough linguistic knowledge), they use a freely available version of Roget's Thesaurus to group words into concepts.

The key steps are:

  1. Extraction: Identifying sequences of nouns from a large corpus (7.8 million words).
  2. Thesaurus Categorization: Tagging each noun in the extracted compounds with its relevant thesaurus categories.
  3. Conceptual Association (CA) Calculation: For pairs of thesaurus categories (t1, t2), they calculate a measure of mutual information. This CA(t1, t2) quantifies how strongly one category predicts the other, essentially measuring how likely it is that a word from category t1 modifies a word from category t2.
  4. Syntactic Disambiguation: For ambiguous compound nouns (especially those with three or more nouns), they use the calculated conceptual associations to choose the most probable binary bracketing. The principle is that the modifier and the modified noun should have the strongest conceptual association. For compounds longer than three nouns, they select the bracketing that maximizes the product of these associations.

Datasets

  • Training Data: Approximately 35,000 noun sequences extracted from Grolier's multimedia on-line encyclopedia. This was filtered down to 24,285 pairs of nouns after removing sequences with nouns not found in Roget's Thesaurus.
  • Evaluation Data: A test set of 308 ambiguous compound nouns (specifically, triples) extracted from the Grolier's corpus, where all constituent nouns were present in Roget's Thesaurus.

Key Insights and Ideas

  • Thesaurus Categories as Proxies for Meaning: Instead of relying solely on word-to-word associations, the paper demonstrates the power of using broader thesaurus categories. This helps generalize associations and requires less data.
  • Conceptual Association for Syntactic Disambiguation: The core idea that strong conceptual links between categories predict likely syntactic attachments is a powerful heuristic for resolving ambiguity.
  • Independence of Syntactic Ambiguity and Lexical Preferences: The authors operate under the assumption that the tendency for words to associate conceptually is relatively independent of whether the syntactic structure is ambiguous. This allows them to build associations from unambiguous cases and apply them to ambiguous ones.

Key Results

  • The system achieved 75% accuracy in syntactically disambiguating a test set of 244 compound noun examples.
  • The results showed a higher success rate with left-branching attachments (e.g., (pottery (coffee mug))) compared to right-branching ones.
  • Out of the 308 test set triples, 163 were correctly identified as left-branching, and 81 as right-branching. 35 were deemed semantically indeterminate, and 29 were errors.

Limitations and Future Work

  • Thesaurus Limitations: The use of the 1911 version of Roget's Thesaurus is a limitation, as it may not fully capture modern vocabulary or nuanced meanings. A more comprehensive and consistent thesaurus is desired.
  • Extraction Noise: The extraction method, while reasonably effective, can still include non-compound noun sequences, introducing some noise.
  • Context-Free Analysis: The current analysis is performed without considering the surrounding text. Future work includes an experiment with human subjects to establish a performance target for context-free analysis.
  • Semantic Role Assignment: Syntactic disambiguation is only the first step. The ultimate goal is to assign semantic roles (e.g., "material," "purpose") to fully interpret the compound noun, which is the next major area of research.
  • Exploring Other Association Schemes: The authors plan to investigate alternative statistical measures for association beyond mutual information.

Summary updated

BibTeX
@inproceedings{lauer1994conceptual,
  author = {Mark Lauer},
  title = {{Conceptual Association for Compound Noun Analysis}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1994},
}
Figure