TL;DR
This paper introduces the Knowledge Acquisition from Text (KAT) system, an interactive method for automatically discovering domain-specific concepts and relationships from text to extend existing knowledge bases like WordNet. By focusing on seed concepts and leveraging lexico-syntactic patterns, KAT aims to overcome the limitations of manual knowledge acquisition for large-scale, knowledge-intensive applications.
What Problem Does This Paper Address?
Building comprehensive and up-to-date knowledge bases is a significant bottleneck for many Artificial Natural Language Processing (ANLP) applications, such as question answering and inference. Manually creating these knowledge bases is time-consuming, error-prone, and difficult to scale. Existing Information Extraction systems often rely on hand-coded linguistic rules, which are not easily portable across different domains. The paper addresses the need for an automated and flexible approach to acquire domain-specific knowledge that can augment general-purpose knowledge bases like WordNet, making them more useful for specialized tasks.
Methods
The KAT system operates in a semi-automatic, interactive mode. Its core methodology involves several steps:
- Discover New Concepts:
- Seed Concepts: Users select important domain-specific concepts (e.g., "interest rate," "stock market" in the financial domain).
- Sentence Extraction: Sentences containing these seed concepts are extracted from a corpus (limited to 1000 sentences per seed).
- Parsing: Sentences are POS-tagged and parsed to identify noun phrases (NPs).
- Concept Identification: New concepts are primarily identified from compound nouns and nouns with modifiers within these NPs. Procedures like "WordNet reduction" and "Dictionary reduction" are used to normalize and identify potential concepts.
- User Validation: Crucially, users validate the discovered concepts, as full automation is not yet possible.
- Discover Lexico-Syntactic Patterns:
- The system searches for patterns (often verb-based) that link concepts of interest.
- It prioritizes discovering WordNet's semantic relations (like HYPERNYMY) and also identifies new domain-specific relations (like INFLUENCE, CAUSE, EQUIVALENT).
- Patterns are extracted from sentences where related concepts co-occur.
- Discover New Relationships:
- Using the identified lexico-syntactic patterns, the system searches the corpus for occurrences where one of the concepts is a seed-related concept.
- New relationships are formed between the seed-related concept and any newly identified concept within the pattern.
- User intervention is again required to accept or reject these discovered relationships.
- Knowledge Classification and Integration:
- Discovered concepts and relationships are integrated into a WordNet-like knowledge base.
- This involves classifying new concepts within the existing hierarchy, primarily using the HYPERNYMY relation.
- The process is iterative, ensuring consistency with WordNet's subsumption principles.
- Other discovered relations (INFLUENCE, CAUSE, EQUIVALENT) are then added to the knowledge base.
Datasets
The paper describes experiments using a corpus of 5000 sentences. This corpus was specifically constructed by selecting sentences containing one of five seed concepts from the financial domain: interest rate, stock market, inflation, economic growth, and employment. The system also leverages WordNet and online dictionaries (like OneLook Dictionaries) for concept validation and reduction.
Key Insights and Ideas
- Interactive Acquisition is Key: Full automation of knowledge acquisition is challenging due to linguistic complexity. An interactive approach, where humans validate concepts and relationships, is presented as a practical necessity.
- Leveraging Lexico-Syntactic Patterns: The paper demonstrates the power of lexico-syntactic patterns derived from text to uncover not just concepts but also the semantic relationships between them, going beyond standard WordNet relations.
- Extending Ontologies: The KAT system's goal is to augment existing ontologies (like WordNet) rather than build them from scratch, making it more efficient and leveraging existing structured knowledge.
- Domain-Specific Focus: The methodology is designed to acquire knowledge relevant to specific domains, which is crucial for applications requiring deep understanding beyond general knowledge.
- Compound Nouns as Concept Sources: The focus on compound nouns and nouns with modifiers as sources for new concepts is a practical strategy for identifying terms not yet present in standard lexicons.
Key Results
- The system was tested on five financial domain seed concepts.
- From a 5000-sentence corpus, the system extracted a total of 362 concepts (319 directly related to seeds, 43 via selected relations).
- 22 distinct lexico-syntactic patterns were discovered and used in 63 instances.
- The manual operation for accepting/declining concepts and relationships took approximately 40 minutes per seed for the 5000-sentence corpus.
- The paper shows examples of how new concepts (like "monetary policy") and relationships (like INFLUENCE) are integrated into a WordNet-like structure, connecting previously unrelated concepts.
Limitations and Future Work
- Coreference Resolution and Ambiguity: The paper acknowledges that the system's inability to handle coreference resolution and other complex linguistic phenomena necessitates user intervention.
- Manual Validation Overhead: While semi-automatic, the human validation step still requires significant time, especially for large corpora.
- Future Work: The authors suggest that handling complex linguistic phenomena like coreference resolution and word sense disambiguation is crucial for increasing automation. They also note that the system can be adapted to acquire information for concepts not initially present in the knowledge base by invoking discovery procedures.
Summary updated