TL;DR
This paper tackles the challenge of translating anaphoric dependencies, like pronoun references, across languages using unification-based grammar formalisms. They propose a novel constraint-based approach that leverages underspecification and target language knowledge to overcome limitations of existing "structural" transfer methods, particularly for complex cases like relative clauses with differing grammatical constraints.
What Problem Does This Paper Address?
Translating sentences with anaphoric dependencies (e.g., "The man who I saw...") across languages is tricky. In linguistics, these dependencies are often represented using "re-entrance" in feature structures, where multiple parts of a structure point to the same value. While effective for analysis within a single language, this re-entrance poses problems when translating between languages, especially when grammatical rules for handling these dependencies differ. Existing "structural" transfer methods, which break down source structures and build target ones, struggle with "non-local" re-entrances and cases where languages handle these phenomena differently (like preposition stranding in relative clauses). This paper aims to provide a more flexible and robust method for translating these complex dependencies.
Methods
The authors propose a constraint-based approach that builds upon the equational constraint language of Lexical-Functional Grammar (LFG). Instead of directly mapping entire source structures to target ones, this method involves:
- Underspecification: Only specifying some of the correspondences between source and target structures.
- Bilingual Constraints: Using LFG's equational language to state constraints that link elements across source and target feature structures via mapping functions (τ and τ').
- Inferential Procedure: Combining the underspecified bilingual constraints with monolingual target language knowledge (e.g., functional uncertainty equations) to infer the complete target structure.
- Focus on One Path: For re-entrances, the approach often involves stating a correspondence for only one of the paths involved in the source re-entrance, allowing the target grammar to resolve the other.
This contrasts with "structural" approaches that attempt to translate both ends of a re-entrance, leading to inflexibility.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Re-entrances are a Translation Bottleneck: The paper highlights that while re-entrance is a powerful tool for representing anaphora within a language, it creates significant challenges for cross-lingual transfer.
- Structural Transfer is Too Rigid: Traditional methods that decompose source structures and build target ones struggle with non-local dependencies and differing cross-linguistic constraints, leading to incorrect translations.
- Constraint-Based is More Flexible: By using underspecification and allowing source and target constraints to interact, the proposed method offers greater flexibility in handling linguistic variations.
- Target Knowledge is Crucial: The approach effectively leverages monolingual target grammar constraints (like functional uncertainty) to resolve ambiguities and ensure correct target structures, especially when source and target grammatical behaviors diverge.
- Underspecification Enables Strategy Mismatches: The ability to underspecify correspondences is key to handling cases where languages employ different strategies for anaphoric dependencies (e.g., pied-piping vs. preposition stranding).
Key Results
The paper demonstrates the proposed approach's ability to correctly translate cases that existing methods fail on. Specifically, it shows how to correctly translate the English sentence "The man who I have answered" into French "La lettre à laquelle j'ai répondu" (correct) rather than the ungrammatical "La lettre laquelle j'ai répondu à" (incorrect) produced by the Kaplan et al. approach. This is achieved by correctly handling the differing requirements for preposition stranding vs. pied-piping in relative clauses. The paper also suggests the approach can handle other "strategy mismatches" like gap vs. resumptive pronoun translations.
Limitations and Future Work
- Limitations of the Kaplan et al. Approach: The paper identifies that the Kaplan et al. constraint-based approach, while an improvement, still has limitations. It can incorrectly translate cases where relativised positions differ between languages, as the source grammar dictates what will be re-entrant on the target side.
- Future Work: The authors suggest that their approach can be extended to handle other "strategy mismatches" beyond preposition stranding, such as translating gaps in one language to resumptive pronouns in another. They also imply that further exploration of how source-derived equations interact with target constraints is warranted.
Summary updated