TL;DR
This paper presents a machine translation system (GBT) that leverages Chomsky's Government and Binding (GB) theory to translate between English and Spanish. By defining a universal grammar with language-specific parameters, the system aims to reduce the complexity of translation and enable easier addition of new languages.
What Problem Does This Paper Address?
Traditional machine translation systems often require defining entirely separate grammars for each language pair, leading to significant redundancy and complexity. The authors address this by proposing a system based on Government and Binding (GB) theory, which posits universal grammatical principles underlying all languages, with variations accounted for by parameters. This approach aims to create a more modular and scalable translation system, where adding a new language primarily involves defining its specific parameter settings rather than building a new grammar from scratch. The core challenge is to demonstrate the feasibility of using a theoretical linguistic framework like GB for practical machine translation.
Methods
The GB-Translator (GBT) system is built around a Universal Grammar (UG) component and two language-specific components (English and Spanish). The UG is composed of:
- X-bar syntax: For phrase structure rules.
- Transformational component: Including rules like "Move Affix" (affix-hopping) and the general "Move Alpha" (which allows for movement of constituents, subject to constraints like subjacency).
- Well-formedness component: Enforcing constraints on surface representations, such as the Doubly-filled COMP Filter, WH-Filter, Case Filter, Empty Category Principle (ECP), and Binding Conditions.
The system operates in several stages:
- Parsing: The source sentence is parsed to build an S-structure representation.
- Reverse Transformations: The S-structure is converted to a D-structure (deep structure) by undoing movement transformations.
- Lexical Transfer: Words are translated based on their entries in the lexicon, which include syntactic categories and transfer values.
- Transformations: Movement transformations are applied to the target D-structure to generate one or more S-structures.
- Validation: Generated S-structures are checked for well-formedness against the UG constraints.
- Generation: Only well-formed target sentences are "flattened" and displayed.
The system uses Prolog for implementation and represents dictionary entries as Prolog unit clauses, including features for subcategorization, transfer values, and theoretical properties like "theta" (thematic role assignment) and "sdel" (S-bar deletion).
Datasets
Not discussed in the paper. The system is demonstrated with example sentences provided by the user.
Key Insights and Ideas
- Parametric Grammar for Translation: The core idea is that by abstracting universal grammatical principles and parametrizing language-specific variations, a more efficient and extensible machine translation system can be built.
- GB Theory as a Foundation: Demonstrates that a theoretical linguistic framework like GB, with its focus on underlying structures and constraints, can be practically applied to machine translation.
- D-structure for Lexical Transfer: Using D-structures for lexical transfer simplifies the process by ensuring that thematic elements are in their canonical positions, avoiding the complexities of translating between surface or LF representations.
- Parameter-driven Language Differences: The system highlights how parameters, such as the "pro-drop" parameter or "proper governor" settings, directly influence syntactic phenomena like subject omission and preposition-stranding, leading to distinct sentence structures across languages.
Key Results
The paper demonstrates the system's ability to translate between English and Spanish in both directions. Examples show the system generating multiple valid translations for a single input sentence, reflecting different syntactic possibilities allowed by the grammar. For instance, translating "Creo que has dicho la verdad" (Spanish) into English yields three possibilities: "I believe that you have told the truth," "I believe you have told the truth," and "I believe you to have told the truth." The system also illustrates how parameter settings affect grammatical well-formedness, such as the ECP's role in allowing preposition-stranding in English but not in Spanish.
Limitations and Future Work
- Syntactic Focus: The current GBT system is purely syntactic and does not incorporate semantic processing, pragmatics, or discourse. The authors acknowledge that a complete translation system would need to address these aspects, ideally by extending the parametric approach to these domains.
- Ambiguity and Structural Alteration: Problems of semantic ambiguity and structural alteration during translation are not addressed, as they are considered issues of translation theory rather than GB.
- Head-Initial Assumption: The current phrase structure rules are designed for head-initial grammars, and accommodating head-final grammars (like German) would require modifications and introduce parsing challenges for Prolog.
- Future Work: The authors suggest that the feasibility of a generalized syntactic component has been illustrated, paving the way for more comprehensive language processing devices that incorporate parametric principles across various linguistic levels.
Summary updated