TL;DR
This paper introduces an ambiguity-preserving machine translation (MT) approach that leverages packed f-structure representations. By avoiding premature disambiguation and transferring these underspecified structures, the system can generate target language utterances that maximally reflect the original ambiguities, leading to more nuanced translations.
What Problem Does This Paper Address?
Machine translation systems, especially those with broad coverage grammars, grapple with the inherent ambiguity of natural language. Resolving all ambiguities is often impossible without vast amounts of world knowledge, leading to a phenomenon known as the "AI completeness problem." This paper addresses the challenge of translating ambiguous utterances without losing the ambiguity in the target language. The authors argue that for many ambiguous source language (SL) utterances, equivalent ambiguous target language (TL) phrases exist. Their goal is to build an MT architecture that can preserve these "preservable ambiguities" rather than forcing a premature, potentially incorrect, disambiguation.
Methods
The core of the proposed method is an "ambiguity-preserving MT approach" that works with packed f-structure representations. F-structures, a formalism from Lexical-Functional Grammar (LFG), are hierarchical representations of sentence structure. "Packed" f-structures represent multiple possible interpretations (e.g., different attachment ambiguities) within a single structure, avoiding the exponential explosion of enumerating all possible parse trees.
The authors' approach involves:
- Conversion to Flat Term Sets: Hierarchical f-structures are converted into a flat set of Prolog predicates. This transformation is crucial for applying transfer rules, which are described as a graph rewriting process.
- Shake-and-Bake like Transfer: The paper extends a "Shake-and-Bake" style transfer approach (or a semantic-based transfer) to handle these packed f-structures. The transfer rules operate on the flat term sets, mapping SL predicates and grammatical functions to their TL counterparts.
- Ambiguity Preservation: The key innovation is that the transfer rules are designed to preserve local disjunctions within the packed representations. This means that if an ambiguity exists in the SL f-structure, it is carried over to the TL f-structure without being resolved.
- Local Disambiguation (when necessary): In cases where full ambiguity preservation is not possible (e.g., translating to a language with different grammatical constraints), the system can perform local disambiguation. This is achieved by integrating selectional restrictions into the transfer rules, allowing for a partial resolution of ambiguities.
Datasets
Not discussed in the paper. The paper focuses on the methodology and uses illustrative examples rather than empirical evaluation on specific datasets.
Key Insights and Ideas
- Preservable Ambiguity: The concept that many ambiguities in the source language can be directly mapped to equivalent ambiguities in the target language is central. This shifts the focus from resolving ambiguity to preserving it.
- Packed Representations as a Solution: Using packed f-structures is presented as an efficient way to represent multiple interpretations without combinatorial explosion. This allows the MT system to handle ambiguity more gracefully.
- Flat Term Sets for Transfer: Converting hierarchical f-structures into flat Prolog terms is a practical step that makes them amenable to existing transfer mechanisms, like graph rewriting.
- Extending Shake-and-Bake: The paper demonstrates that established transfer paradigms can be adapted to work with underspecified, packed representations, extending their applicability.
- Orthogonality of Syntactic and Semantic Ambiguity: The authors suggest that their approach to handling syntactic ambiguity with packed representations is orthogonal to the treatment of semantic ambiguities, implying a modular design.
Key Results
The paper demonstrates the feasibility of their approach through detailed examples. Specifically, they show how a PP attachment ambiguity in a German sentence ("wir treffen die Kollegen in Berlin") can be preserved in an English translation. The packed f-structure for the German sentence, containing a local disjunction for the PP attachment, is transferred to a corresponding packed f-structure in English. This packed representation is then passed to a generator, which can produce utterances that reflect the original ambiguity. The paper also illustrates how local disambiguation can be applied when necessary, leading to a more controlled translation process.
Limitations and Future Work
- Local Disjunction Notation: The paper notes that the notation for local disjunctions used for illustration is simplified. The actual implementation relies on "contexted constraints" as developed in the XLE system.
- Human Intervention: For cases where ambiguity cannot be fully resolved by the system, the authors suggest that the output can be presented to a human "negotiator" (e.g., in an interactive translation system) to make the final disambiguation choice.
- Semantic Ambiguities: While the paper focuses on syntactic ambiguities (like PP attachment), it mentions that the approach could be extended to handle underspecified semantic representations, suggesting future work in integrating both types of ambiguity.
- Application Domain: The current implementation is being extended for translating instruction manuals, indicating a specific but potentially broad application area.
Summary updated