An Application of Lexical Semantics to Knowledge Acquisition from Corpora

Peter Anick, James PustejovskyCOLING 1990

TL;DR

This paper introduces a novel approach to automatically acquiring lexical knowledge for information retrieval (IR) systems by leveraging a computational theory of lexical semantics. Instead of relying on manual knowledge engineering or purely statistical methods, the authors propose using a "generative lexicon" and its "qualia structure" to infer semantic relationships between words directly from large text corpora. This method aims to build richer, more structured knowledge bases for IR than traditional approaches.

What Problem Does This Paper Address?

The paper tackles the challenge of improving information retrieval (IR) systems. While full-text indexing is common, its effectiveness is limited by precision and recall issues. Domain knowledge can enhance IR, but manually creating this knowledge is time-consuming and often incomplete. Existing automated methods, like statistical collocational analysis, lack the depth to capture nuanced semantic relationships. The authors propose a way to automatically acquire this deeper semantic knowledge, specifically focusing on nominals (nouns), which are crucial for query formulation in IR.

Methods

The core of the paper lies in applying a computational lexical semantic theory, specifically Pustejovsky's "generative lexicon," to knowledge acquisition from corpora. This theory posits that lexical items have an internal "qualia structure" that defines their meaning along four dimensions: Constituent (parts), Form (structure), Telic (purpose/function), and Agentive (origin).

The authors' knowledge acquisition procedure involves several steps:

  1. Morphological Analysis: Assigning morphological paradigms to words.
  2. Noun Compound Bracketing: Identifying and structuring noun compounds (e.g., "[database management] system").
  3. Collecting Prepositional Phrases: Gathering noun phrases related by prepositions from "collocational systems" (sets of phrases that systematically co-occur with a lexical item).
  4. Hypothesizing Subclass Relationships: Inferring "ISA" (is-a) relationships between nouns based on their co-occurrence patterns in compounds and phrases.
  5. Distributional Verification: Seeking further evidence to support hypothesized subclass relationships.
  6. Semantic Classification: Attempting to classify words into known lexical categories by matching their syntactic constructions to diagnostic sets.
  7. Clustering Predicates: Grouping verbs based on the semantic roles they play with respect to a noun's qualia structure (e.g., verbs related to the Telic role of "tape").

A key innovation is the use of "projective conclusion spaces" derived from the qualia structure. These spaces allow for the dynamic creation of related concepts through transformations (like negation, temporal precedence), which helps in identifying semantic relationships and potential metonymic extensions.

Datasets

The authors describe experiments conducted on a Digital Equipment Corporation (DEC) on-line corpus of 3000 articles containing VMS troubleshooting information. This corpus was used to test their knowledge acquisition procedure. No new benchmarks were introduced; rather, existing corpora were utilized.

Key Insights and Ideas

  • Lexical Semantics as a Bridge: The paper demonstrates that a formal theory of lexical semantics can serve as a powerful framework for guiding automated knowledge acquisition from text.
  • Qualia Structure's Predictive Power: The qualia structure of nouns is shown to correlate with observable collocational systems in corpora, meaning that the inherent semantic properties of a noun predict how it will be used in sentences.
  • Metonymy as a Clue: Contrary to some views, the authors argue that motivated metonymic structures (e.g., using "tape" to refer to the information on it) are valuable clues for semantic analysis, not obstacles.
  • Generative Approach to Lexicon: The idea of a "generative lexicon" where meanings are built through rules and devices, rather than just a static list of primitives, is central to their approach.
  • Nominals are Paramount: The focus on nominals is justified by their importance in IR queries and the authors' belief that their semantic contribution has been historically under-appreciated.

Key Results

The paper presents qualitative results demonstrating the application of their procedure to the nouns "tape" and "mouse." For "tape," they derive a qualia structure that includes its constituent (information, file), form (mount, dismount), telic (read, write, copy), and agentive (artifact) roles. Similarly, for "mouse" (in its computer artifact sense), they derive roles related to its function as an instrument.

The authors highlight that this acquired information can be used for tasks like homonym disambiguation and query reformulation in IR. They also note that the procedure can help discriminate between different semantic roles (e.g., Telic vs. Formal) of a noun by analyzing verb intersections.

Limitations and Future Work

The authors acknowledge several limitations:

  • Susceptibility to False Conclusions: The algorithm can be "too susceptible to jumping to false conclusions," especially with ambiguous prepositions and spurious verb-object combinations.
  • Missed Cases: The current heuristics may miss some valid relationships between terms.
  • Preliminary Nature: The experiments are described as preliminary, and it's too early to definitively assess how well the approach scales to larger corpora with more diverse senses of words.

Future work includes:

  • Improving Precision and Recall: Refining discovery procedures to reduce misses and false alarms, potentially by incorporating information-theoretic statistics.
  • Extending Coverage: Testing the techniques on a wider range of terms and senses.
  • Empirical Evaluation: Applying the acquired lexical knowledge within an experimental IR system to test its effectiveness for indexing and retrieval.

Summary updated

BibTeX
@inproceedings{anick1990application,
  author = {Peter Anick and James Pustejovsky},
  title = {{An Application of Lexical Semantics to Knowledge Acquisition from Corpora}},
  booktitle = {International Conference on Computational Linguistics},
  year = {1990},
}
Figure