Abstract
Building comprehensive semantic lexicons for natural language processing (NLP) has historically been a labor-intensive manual task, creating a significant bottleneck in artificial intelligence development. This paper presents PETRARCA, a system designed to semi-automatically acquire and interpret word meanings by analyzing lexical co-occurrences within large text corpora. By mapping surface-level syntactic structures to deep conceptual representations using Conceptual Graphs, PETRARCA bridges the gap between raw text and structured knowledge, offering a scalable alternative to traditional manual lexicon construction.
1. Introduction
In the late 1980s, the primary challenge facing natural language processing was the "knowledge acquisition bottleneck." For an NLP system to be functionally useful, it requires a semantic lexicon containing thousands of terms, each defined with specific world knowledge. Traditionally, these lexicons were built manually—a process that is not only slow and expensive but also subjective and prone to inconsistency.
While on-line dictionaries were initially viewed as a potential solution, they often suffered from circular definitions and lacked the nuanced, real-world usage data found in actual documents. This paper proposes a "collocative" approach, arguing that deep conceptual meaning can be derived by observing and interpreting surface-level word associations in raw text.
2. Methodology: The PETRARCA System
PETRARCA utilizes an algorithmic approach to build a case-based semantic lexicon through a five-stage pipeline:
- Parsing: The system performs a morphosyntactic analysis of every sentence in a corpus containing a target word ($W$).
- Syntactic Attachment: It identifies valid syntactic co-occurrences, such as determining which noun phrase serves as the direct object of a specific verb.
- Semantic Interpretation: Using a catalog of "Syntax-to-Semantic" (SS) rules and "Selectional Restriction" (CR) rules, the system maps syntactic patterns to conceptual relations. For example, a prepositional phrase might be mapped to a
PARTICIPANTorLOCATIONrole based on the governing verb. - Generalization: The system analyzes multiple instances of a word to find the most general valid constraint. If a verb is consistently associated with "company," "organization," and "president," PETRARCA generalizes the requirement to a
HUMAN_ENTITYorLEGAL_ENTITYwithin its type hierarchy. - Manual Supervision: A human reviewer performs a final check on the derived entries to ensure accuracy and resolve ambiguities the system could not handle.
The system employs Conceptual Graphs as its underlying semantic language, providing a structured, formal way to represent the participants and actions within a sentence.
3. Conceptual Framework
Surface vs. Deep Semantics
A core tenet of this research is that while "deep" conceptual meaning is often subjective, surface-level word associations are objective, observable phenomena. By capturing how words are used in everyday language, PETRARCA approximates the "deep sense" of a word more reliably than abstract manual definitions.
Syntax as a Semantic Filter
The authors observe that syntax alone often allows for semantically nonsensical associations. PETRARCA uses selectional restrictions to filter these out. For instance, if the system encounters the phrase "The party decided," it evaluates the different senses of "party." Since a "FEAST" cannot perform the action of "DECIDING," the system rejects that interpretation in favor of "POLITICAL_PARTY."
4. Experimental Results
The system was evaluated using a primary corpus of approximately 100,000 words from press agency releases focused on finance and economics, supported by a base morphosyntactic lexicon of 7,000 lemmata. To test domain portability, the system was also applied to the text of the Italian Constitution.
Key performance metrics included:
- Coverage: The morphosyntactic analyzer achieved 100% coverage on the target corpus, while the syntactic parser successfully processed 80% of sentences.
- Efficiency: Running on an IBM mainframe, the system parsed complex sentences (containing 3-4 prepositional phrases) in 1–2 seconds.
- Functional Utility: When integrated into the DANTE system, a lexicon built via these methods enabled the computer to successfully answer over 90% of natural language queries regarding the source texts.
- Scalability: The effort required to assign supertypes to 10,000 words was found to be comparable to building a standard morphological lexicon, suggesting that large-scale semantic processing is feasible.
5. Limitations and Future Work
Despite its successes, PETRARCA requires human intervention for the final validation of entries. The system currently struggles with non-literal language, such as metonymies and idioms, which often violate standard semantic constraints. While the system can detect these violations, it cannot yet automatically reinterpret them.
Future research directions include moving toward semi-automatic word clustering based on fuzzy set theory. This would further reduce the manual effort required to organize and maintain the conceptual hierarchies that underpin the system’s generalization capabilities.
Summary updated