Formal Ontology Learning on Factual IS-A Corpus in English using Description Logics

Sourish Dasgupta, Ankur Padia, Kushal Shah, Prasenjit Majumder

TL;DR

This paper introduces DLOLIS-A, a novel framework for formal ontology learning from factual "IS-A" sentences in English. It tackles the challenge of accurately capturing the nuanced semantics of these sentences, often missed by lighter approaches, by leveraging Description Logics (DL) and a deep linguistic analysis. The authors demonstrate significant improvements over existing methods, particularly for complex and non-trivial IS-A constructions.

What Problem Does This Paper Address?

Ontology Learning (OL) aims to automatically build knowledge bases from unstructured text. Existing methods, often termed "Light-Weight OL," primarily rely on statistical and lexico-syntactic patterns. While these approaches are prevalent, they struggle with the inherent linguistic nuances of natural language, leading to less accurate learning. This paper focuses on "Formal OL," a less explored but more principled approach that uses deep linguistic analysis to generate formal axioms. Specifically, it addresses the challenge of learning from factual "IS-A" sentences (e.g., "A cat is an animal"), which, despite their seemingly simple structure, can exhibit complex semantic variations that are difficult to formalize. The authors argue that these "non-trivial" IS-A sentences are crucial for building accurate ontologies and require specialized study.

Methods

The core of the proposed approach is a formal ontology learning framework based on Description Logics (DL). The system, named DLOLIS-A, processes English IS-A sentences through a pipeline:

  1. Characterization: A novel "characterization template" is used to parse IS-A sentences, identifying subject, predicate (IS-A variations), and object. This template is designed to be more robust than traditional full parsing for this specific task.
  2. Triple Extractor: This component simplifies complex and compound sentences into a set of simpler ones, facilitating easier translation.
  3. Singularizer: Plural forms of subjects, objects, and predicates are lemmatized to their singular versions to avoid redundant concepts in DL.
  4. Normalizer: Lexical variations of IS-A, quantifiers, and clausal tokens are normalized to standard forms.
  5. Characterizer: A Part-Of-Speech (POS) tagger is used to identify linguistic categories and extract subject/object phrases, along with their modifiers and quantifiers.
  6. DL Translator: The characterized sentences are translated into DL expressions, forming T-Box (concept definitions) and A-Box (instance assertions) rules. This translation handles various trivial and non-trivial IS-A sentence structures, including those with quantifiers, modifiers, and complex semantic nuances.
  7. Ontology Generation & Consistency Checking: A DL reasoner (FACT++) is used to generate the final ontology and check its consistency.

The authors emphasize the use of DL as a formal language due to its decidability, compactness, and one-to-one correspondence with OWL.

Datasets

  • WCL v.1.1 dataset: A large, domain-independent, factual dataset primarily extracted from Wikipedia, rich in definitional and IS-A sentences.
  • Community representative dataset: A newly created dataset specifically designed to include a comprehensive set of trivial and non-trivial IS-A sentences, addressing cases not well-represented in existing benchmarks.

Key Insights and Ideas

  • Non-trivial IS-A sentences are common and challenging: The paper highlights that seemingly simple IS-A sentences can have complex semantic interpretations that are not directly derivable from grammatical structure alone. These "non-trivial" cases are crucial for accurate ontology learning.
  • Description Logics are well-suited for formal IS-A learning: DL's expressive power and formal semantics provide a robust foundation for translating nuanced linguistic structures into machine-readable axioms.
  • Deep linguistic analysis is necessary: Statistical and lexico-syntactic methods are insufficient for capturing the full semantics of IS-A sentences. A principled approach involving detailed linguistic analysis is required.
  • Characterization templates simplify parsing: For the specific task of IS-A sentence analysis, a specialized "characterization template" is proposed as a more efficient alternative to full syntactic parsing.
  • Normalization and singularization are key preprocessing steps: These steps help manage linguistic variations and reduce redundancy in the DL representation.

Key Results

  • Significant improvement over benchmarks: DLOLIS-A demonstrated considerable improvements in ontology learning accuracy compared to light-weight OL tool Text2Onto and formal OL tool FRED on the WCL dataset.
  • High Characterization Accuracy: The DLOLIS-A characterizer achieved 100% Characterization Precision (CP) and 99% Characterization Recall (CR) on the WCL dataset.
  • Superior performance on custom datasets: On their custom-created datasets for trivial and non-trivial IS-A sentences, DLOLIS-A showed significantly better performance than FRED in terms of correctly classifying sentences, with FRED exhibiting a higher rate of wrong or partial semantic constructions.
  • Runtime efficiency: The paper presents runtime analysis, showing that DLOLIS-A's DL translation phase is efficient, with overall process times presented for different dataset sizes.

Limitations and Future Work

  • Non-monotonic reasoning details omitted: The paper mentions that the details of non-monotonic reasoning are kept for future work.
  • DLOLFULL is a work-in-progress: The current tool DLOLIS-A focuses on IS-A sentences. A more comprehensive tool, DLOLFULL, is under development to handle non-IS-A sentences.
  • Anaphoric ambiguity resolution: The current scope does not include resolving anaphoric ambiguity in subject instances.
  • Temporal modal constraints and spatial clauses: The representation of temporal modal constraints on "is a" and the handling of spatial clauses are left for future work.
  • Backward modification not implemented: The DLOLIS-A tool does not currently implement backward modification of modifiers.
  • DL representation of temporal clauses: The DL representation of temporal clauses is left as future work.
  • Purely IS-A sentences: The paper notes that counter-factual and exemplificative clauses, as well as some spatial clauses, may not be purely IS-A in nature, requiring further investigation.
  • Ontological validity of past/future tense: DLOLIS-A takes a "play safe" attitude for past and future tense ambiguities, not fully solving the ontological ambiguity.
  • DL representation of "is sometimes" ambiguity: The DL representation for the "is sometimes" variation of "is a" is presented, but its full implications are not detailed.

Summary updated

Abstract

Ontology Learning (OL) is the computational task of generating a knowledge base in the form of an ontology given an unstructured corpus whose content is in natural language (NL). Several works can be found in this area most of which are limited to statistical and lexico-syntactic pattern matching based techniques Light-Weight OL. These techniques do not lead to very accurate learning mostly because of several linguistic nuances in NL. Formal OL is an alternative (less explored) methodology were deep linguistics analysis is made using theory and tools found in computational linguistics to generate formal axioms and definitions instead simply inducing a taxonomy. In this paper we propose "Description Logic (DL)" based formal OL framework for learning factual IS-A type sentences in English. We claim that semantic construction of IS-A sentences is non trivial. Hence, we also claim that such sentences requires special studies in the context of OL before any truly formal OL can be proposed. We introduce a learner tool, called DLOL_IS-A, that generated such ontologies in the owl format. We have adopted "Gold Standard" based OL evaluation on IS-A rich WCL v.1.1 dataset and our own Community representative IS-A dataset. We observed significant improvement of DLOL_IS-A when compared to the light-weight OL tool Text2Onto and formal OL tool FRED.
BibTeX
@misc{dasgupta2013formal,
  author = {Sourish Dasgupta and Ankur Padia and Kushal Shah and Prasenjit Majumder},
  title = {{Formal Ontology Learning on Factual IS-A Corpus in English using Description Logics}},
  year = {2013},
  eprint = {1312.6947},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1312.6947},
}
Figure