Construction of Conceptual Graph Representation of Texts

Svetlana HensmanNAACL 2004

TL;DR

This paper presents a novel two-step system for automatically converting natural language text into conceptual graphs (CGs), a powerful knowledge representation formalism. By combining linguistic resources like VerbNet and WordNet with domain-specific knowledge, the system aims to bridge the gap between human language and machine-understandable knowledge for applications like information retrieval and question answering.

What Problem Does This Paper Address?

The automatic acquisition of knowledge from text is a long-standing challenge in natural language processing. While linguists have explored this for decades, creating machine-readable representations that capture the nuances of language remains difficult. Specifically, this paper tackles the problem of transforming unstructured text into conceptual graphs (CGs), a formalism that offers greater expressive power than first-order logic, particularly for temporal and non-monotonic reasoning. Existing approaches often rely heavily on manually crafted grammars and lexicons, limiting their scalability and applicability to specific domains. This work seeks to develop a more automated and robust method for CG generation.

Methods

The proposed system employs a two-step approach:

  1. Semantic Role Identification: This module identifies the semantic roles (e.g., Agent, Theme, Source) within a sentence. It leverages:
  2. VerbNet: A computational verb lexicon that provides syntactic and semantic information for verbs, including thematic roles and how they are realized in sentences.
  3. WordNet: A lexical database that organizes words into synsets and provides hierarchical relationships, used here to check selectional restrictions for semantic roles.
  4. Syntactic Parsing: The system first parses sentences to identify verb-clause structures and then matches these structures against VerbNet's semantic frames, considering selectional constraints.
  5. Conceptual Graph Construction: Once semantic roles are identified, this module constructs the CG. The process involves:
  6. Phrase-level CGs: Each phrase in a sentence is recursively converted into a CG. For noun phrases, this involves analyzing common syntactic patterns and creating concepts for nouns and adjectives, linked by relations like "Attribute." For prepositional phrases, the preposition is tracked to mark relationships.
  7. Attaching Constituents to the Verb: All phrase-level CGs are then attached to the CG representing the main verb of the clause, forming a single graph for the sentence. Generic relations are used initially if specific semantic roles are not identified.
  8. Resolving Unknown Relations: A database of "relation correction rules" is used to resolve generic relations into more specific ones. These rules are semi-automatically compiled by scanning the corpus for common syntactic patterns and manually identifying the corresponding semantic relations.

Datasets

The system was tested on two corpora:

  • Reuters-21578 Text Categorization Collection: A publicly available dataset for text categorization.
  • Irish Air Accident Investigation Unit (AAIU) Reports: A collection of aviation incident reports.

The paper does not introduce new benchmarks but uses these existing corpora to evaluate the system's performance.

Key Insights and Ideas

  • Leveraging VerbNet and WordNet Synergy: The core innovation lies in the synergistic use of VerbNet for semantic role information and WordNet for constraint checking. This combination allows for a more principled and data-driven approach to semantic role labeling compared to purely manual methods.
  • Two-Step Automation: The clear separation into semantic role identification and CG construction provides a modular and understandable pipeline for automated knowledge extraction.
  • Semi-Automated Relation Resolution: The relation correction rules database is a practical approach to bridge the gap between generic CG structures and more specific, domain-relevant relations, acknowledging that complete automation is challenging.
  • Handling Syntactic Variety: The system's ability to process various syntactic patterns for noun phrases, prepositional phrases, and subordinate clauses demonstrates its robustness in handling diverse sentence structures.

Key Results

The paper reports preliminary results for the semantic role annotation module:

  • VerbNet Coverage: The coverage of verbs described in VerbNet was relatively low (66% for Reuters, 53% for AAIU).
  • Semantic Role Accuracy: When verbs were covered by VerbNet, the semantic roles were correctly identified in 39% of cases for Reuters and 35% for AAIU. This translates to 59% and 66% accuracy, respectively, relative to the verbs present in VerbNet.

The authors acknowledge that these initial accuracies are influenced by the limited VerbNet coverage and the lack of semantic frames for certain verb senses.

Limitations and Future Work

  • VerbNet Coverage: A significant limitation is the low coverage of verbs in the corpora by VerbNet. The authors plan to extend this coverage by manually identifying and adding frames for verbs not present in VerbNet.
  • Verb Sense Disambiguation: The current system does not explicitly handle verb sense disambiguation, which can lead to incorrect frame selection. Implementing a verb sense disambiguation component is a key area for future work.
  • Domain-Specific Knowledge: While the system uses some domain-specific knowledge, its integration and impact could be further explored.
  • Relation Correction Database: Building and maintaining the relation correction rules database is a challenging, semi-automated task. Further research into automating this process is implied.
  • Testing and Tuning: The authors state they are "in the process of testing and tuning" the system, suggesting that the reported results are preliminary and further optimization is expected.

Summary updated

BibTeX
@inproceedings{hensman2004construction,
  author = {Svetlana Hensman},
  title = {{Construction of Conceptual Graph Representation of Texts}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2004},
}
Figure