TL;DR
This paper proposes the concept of a "lexical knowledge base" (LKB) as a comprehensive computational repository of word meanings and their relationships, going beyond traditional dictionaries. The authors argue that LKBs are crucial for advancing natural language understanding (NLU) by integrating various knowledge sources to capture the nuances of human world knowledge.
What Problem Does This Paper Address?
Traditional dictionaries, while useful, offer a limited view of word meaning. They provide definitions that evoke concepts for humans but are insufficient for computational systems to truly understand and reason about the world. Natural language processing (NLP) systems, such as parsers and generators, require a deeper, more structured understanding of lexical concepts and their interrelationships to perform complex tasks. The challenge lies in how to computationally represent and integrate the vast and varied knowledge associated with words, including not just definitions but also their usage, associations, and even visual representations.
Methods
The paper outlines a vision for constructing LKBs by integrating information from multiple sources:
- Machine-Readable Dictionaries: These are the primary source, providing definitions and structured relationships. The authors highlight the utility of ISA (is-a) hierarchies derived from dictionary definitions (e.g., "crossbow" IS-A "weapon") and subject classifications (e.g., "crossbow" belongs to the "military history" domain).
- Statistical Analyses of Text Usages: While not detailed in this excerpt, this suggests using corpus data to understand how words are actually used.
- Manual Human World Knowledge: This acknowledges the need to incorporate knowledge that isn't explicitly in dictionaries.
- Pictorial Information: The paper emphasizes the importance of encoding information from illustrations and diagrams, which can convey spatial, temporal, and relational knowledge not easily captured in text alone. This includes static images of objects and sequential images of actions.
The core idea is to assemble these disparate pieces of information into a unified database, rather than building specialized knowledge for individual NLP applications.
Datasets
The paper mentions the following data sources:
- Merriam-Webster Pocket Dictionary: Used to derive a "tangled" ISA hierarchy of 27,000 nominal and 12,000 verbal concepts.
- Longman Dictionary of Contemporary English: Used to extract subject codes for approximately 17,000 concepts.
- 'Duden' series of pictorial dictionaries: Cited as an example of resources containing line drawings and illustrations of tangible objects and scenes.
No new datasets were introduced, but existing resources were leveraged for their rich lexical and conceptual information.
Key Insights and Ideas
- Dictionaries as Indices, Not Meanings: Dictionary definitions are seen as pointers to concepts, not the complete meaning itself. LKBs aim to build upon this foundation.
- Beyond Definitions: The paper stresses the need to go beyond simple definitions to include information about usage levels, geographic/chronological associations, and semantic/syntactic restrictions.
- Implicit Knowledge Extraction: Analyzing collections of definitions can reveal implicit knowledge (e.g., analyzing verbs of motion to understand "move" more deeply).
- The Importance of Visual and Procedural Knowledge: LKBs must incorporate procedural knowledge (how to do things) and visual imagery, which are poorly represented in traditional dictionaries. This is crucial for understanding concepts like size, position, and orientation.
- LKB as a General Repository: The goal is to create a versatile resource that can serve multiple NLP applications, rather than task-specific knowledge bases.
Key Results
The paper is primarily conceptual and outlines a framework rather than presenting empirical results. The "results" are the identification of existing components that can be assembled into an LKB and the articulation of the need for further integration of visual and procedural knowledge.
Limitations and Future Work
- Encoding Visual Information: A significant challenge identified is developing methods to encode pictorial information (illustrations, scenes) for computational reasoning. This requires bridging the gap between natural language understanding and computer vision research.
- Inferrability of Information: The authors propose a guiding principle for adding new information to an LKB: it should be added only if it is demonstrably needed for NLU problems and cannot be inferred from existing material.
- Integration Complexity: While components are identified, the paper doesn't detail the complex task of integrating these diverse knowledge sources into a cohesive and efficient database.
The authors suggest that the fundamental question for future work is determining what information is necessary for NLU and how to acquire and represent it effectively.
Summary updated