TL;DR
This paper introduces the Layered Domain Class (LDC) system, a transportable natural language processor designed to query structured but unnormalized data. LDC's key innovation lies in its ability to learn and process user-defined modifiers, allowing for flexible and semantically rich queries without requiring extensive pre-programming for each new data domain.
What Problem Does This Paper Address?
Natural language processing (NLP) systems often struggle with flexibility and adaptability. Building a system that can understand queries about diverse, user-defined data structures is a significant challenge. Traditional approaches require extensive manual effort to define the vocabulary, syntax, and semantics for each new domain. This paper addresses the need for a "transportable" NLP system that can be customized for new data domains by learning from the user, specifically focusing on how to acquire and utilize user-defined modifiers (like adjectives, verbs, and prepositional phrases) to interpret complex natural language queries.
Methods
The LDC system is composed of two main parts: a knowledge acquisition component ("Prep") and a User-Phase Processor.
- Knowledge Acquisition (Prep): This module interactively learns about a new domain from the user.
- Stage 1: Identifies entities (e.g., "student," "course") and their basic properties (synonyms, types, subtypes, relationships to other entities).
- Stage 2: Learns morphological variants of words (plurals, verb tenses, comparative/superlative forms).
- Stage 3: Acquires the semantics of user-defined modifiers. This is done by defining modifiers using a set of primitives (like
id,val,num,sum,aug,pct) and specifying a "path" through the domain's hierarchical structure to connect the modifier to the relevant data. For example, defining "desirable instructor" might involve specifying the path from "instructor" to "grade" and then applying a predicate (e.g., "gave a good grade to more than half his students"). - User-Phase Processor: This component handles user queries. It involves:
- Scanner: Reads the input and looks up tokens in a dictionary.
- Parser: Determines the syntactic structure of the input using an augmented phrase-structure grammar.
- Translator: Converts the parsed input into a formal query.
- Retrieval: Executes the query against the data.
A key aspect is the use of domain-specific files (Pattern File, Dictionary, Compat File, Macro File) generated by Prep, which allow the system to be domain-independent. The Macro File stores the acquired modifier meanings in an executable format.
Datasets
Not discussed in the paper. The paper uses examples drawn from two complex domains: a "final grades" domain for academic departments and a "building organization" domain. Several simpler domains are also mentioned, including employee information and stock market quotations.
Key Insights and Ideas
- User-Defined Modifiers are Key: The paper argues that much of the semantic sophistication in natural language queries comes from noun phrase modifiers, including relative clauses. LDC's ability to learn these is crucial for flexibility.
- Hierarchical Domain Structure is Sufficient: LDC restricts itself to domains with hierarchical decompositions, simplifying the knowledge acquisition process. This assumption allows for more manageable semantic learning.
- Separation of Syntax and Semantics Acquisition: Prep acquires both syntactic (morphology) and semantic information separately, but the system's overall processing integrates them.
- Executable Modifier Meanings: Storing modifier meanings in a "macro file" that can be directly executed by the retrieval module simplifies the translation from parse structures to formal queries and allows for deriving meanings of related modifiers (e.g., "largest" from "large").
- Handling Negation: The paper notes that scope ambiguities related to negation are a partially unsolved problem, but suggests that inserting a "Not" command at an appropriate point in the macro body can often resolve meaningful interpretations.
Key Results
The paper demonstrates LDC's capability to process queries involving user-defined modifiers, including:
- Syntactic and Semantic Implementation: Table 1 shows that LDC implements syntax for most modifier types and semantics for many, with ongoing work on others.
- Examples of Learned Modifiers: The paper provides concrete examples of how LDC learns modifiers like "desirable instructor" and "good grade" by specifying paths and predicates.
- Transportability: LDC has been customized for several complex and simple domains, showcasing its transportable nature.
Limitations and Future Work
- Negation Scope Ambiguity: The paper acknowledges that resolving the scope of negation is a partially unsolved problem, and the current system takes a literal reading, which might not always align with user intent (e.g., presupposition).
- Complexity of Semantic Acquisition: While LDC simplifies semantic acquisition by relying on hierarchical domains and primitives, acquiring complex semantics from untrained users remains a challenge.
- Further Work: The paper implies ongoing work on implementing semantics for all modifier types and refining the handling of complex linguistic phenomena like negation scope. The "Compat File" also suggests future work in predicting conceptual relationships based on heuristics.
Summary updated