TL;DR
This paper argues that transformational grammars, often dismissed for efficiency concerns in natural language understanding, can actually simplify query processing for database systems. By using a transformational parser to modify query structures before semantic interpretation, the system avoids the need for data-base-specific modifications, leading to more elegant and maintainable code.
What Problem Does This Paper Address?
The prevailing view in artificial intelligence at the time was that transformational grammars were too inefficient for practical natural language understanding systems, especially when interacting with databases. Researchers like Woods (1975) argued against their use due to these perceived efficiency issues. The challenge was to reconcile the linguistic power of transformational grammars with the practical demands of database query systems, which often require specialized handling for different data fields or query types. For instance, a query about "number of employees" versus "gross income" for a company might require different processing logic.
Methods
The paper proposes using a transformational parser not just for initial linguistic analysis but also as a pre-processing step for semantic interpretation. Instead of modifying the semantic interpreter or the database functions to handle specific query types (e.g., distinguishing between "EMPLOYEE" and "GROSS" data fields), the transformational parser is employed to alter the underlying query structure itself. This is achieved through "string transformations," a set of rules that operate on "lexical trees." These transformations can modify the structure so that a single semantic interpreter and database access mechanism can handle a wider variety of queries. The key innovation is using the parser to "insulate" the semantic interpreter and database functions from the specifics of the data organization.
Datasets
Not discussed in the paper. The paper uses hypothetical examples of company databases and city planning files.
Key Insights and Ideas
- Transformational Grammars for Efficiency: The paper challenges the assumption that transformational grammars are inherently inefficient for database interaction.
- Structural Modification as a Solution: Instead of creating specialized code for each data field or query type, the system modifies the query's underlying structure using transformational rules.
- Insulating the System: This approach effectively insulates the semantic interpreter and database access logic from the specifics of the data base schema, leading to a more generalized system.
- Reduced Special Cases: By handling variations through structural transformations, the need for numerous special-case modifications in the semantic interpreter or database functions is significantly reduced.
- Linguistic Elegance and Practical Benefits: The authors found that a linguistically motivated approach (using transformational grammars) unexpectedly yielded practical engineering benefits.
Key Results
The paper doesn't present quantitative results in terms of accuracy or speedup. Instead, it focuses on the qualitative benefits of the approach. The authors demonstrate how their system can handle queries like "How many employees does GE have?" and "What were GE's 1970 earnings?" by transforming the underlying query structure. For example, a query about employees would result in a structure identifying "EMPLOYEE" rather than "GROSS," allowing a single semantic function to process it. The authors state that the cost of writing new transformational rules was much less than developing new programs for each special case.
Limitations and Future Work
The authors acknowledge that their system, in its initial form, might have required modifications to the SIZEOF function to handle different data fields. However, they argue that their later approach using transformational parsing to modify structures is more elegant. They also note that while they haven't yet encountered a class of structural changes requiring more than one rule, it's possible such situations could arise. The paper implicitly suggests that further exploration of string transformations could lead to even more generalized and robust natural language understanding systems.
Summary updated