Abstract
This research addresses the translation of compound words within a transfer-based Machine Translation (MT) environment. By integrating contemporary linguistic theory—specifically lexicalist generative morphology—the study characterizes, represents, and translates productive and compositional compounds. The proposed strategies are designed for stratificational linguistic frameworks capable of emulating feature value percolation. Through a reductionist approach, the research isolates a set of compositionally translatable compound types, providing a robust alternative to the heuristic-based methods common in early MT systems.
Introduction
Published in 1996, this research explores the complex "grey area" where word structure (morphology) meets sentence structure (syntax). In the context of Machine Translation, compounding represents a significant challenge because different languages utilize vastly different naming conventions to describe the same concept. The paper seeks to move beyond the ad-hoc heuristics of the era, proposing an elegant linguistic solution that can be applied across Romance, Germanic, and Greek language families.
The Problem: Cross-Linguistic Inconsistency
Compounding is notoriously difficult for Natural Language Processing (NLP) because languages "package" concepts differently. For example:
- English often uses multi-word phrases: safety belt.
- Greek may use a head noun modified by a genitive case: zwnj asfaleias.
- German or Greek may fuse these concepts into a single word: kapnokalliergeia (tobacco cultivation).
Traditional MT systems often struggled with "productive" compounds—new words created on the fly by speakers. Without a formal linguistic framework, these systems could not reliably decompose a source compound and reconstruct it accurately in a target language with different structural rules.
Methodology and Framework
The research is situated within the Eurotra project, a major European transfer-based MT system. The author applies principles from lexicalist generative morphology (drawing on the work of Selkirk, Williams, and Lieber) to the computational environment.
The Stratificational Approach
The system processes language through distinct levels, such as Normalized Text, Morphological Structure, and Relational Structure. A key innovation proposed is the Multiword Compound (EMCS) stage. This stage bridges the gap between the morphological level (which typically analyzes single words) and the syntactic level (which analyzes sentence structure).
Syntactic Atomicity and Feature Percolation
To prevent the sentence-level parser from misinterpreting compounds, the system treats them as "syntactic atoms." This designation informs the parser that the entire phrase should be treated as a single noun, preventing the insertion of adjectives or pronouns within the compound structure.
Furthermore, the system utilizes Feature Percolation based on the "Righthand Head Rule." This allows the grammatical properties of the "head" of a compound (e.g., the "belt" in "safety belt") to move up to the top node of the word's internal tree structure, ensuring the compound functions correctly within the wider sentence.
Key Linguistic Insights
- Orthography vs. Linguistics: The paper argues that spaces or hyphens are unreliable indicators of a compound. A "word" is a linguistic concept, not merely a string of characters between spaces.
- Syntactic Opacity: Compounds are distinct from mini-sentences because they are "opaque" to certain syntactic rules. For instance, while one can say "a very industrial city" (a phrase), one cannot say "*a very industrial revolution" (a compound). This opacity is a primary marker for identifying compounds.
- Hierarchy and Binary Branching: Rather than "flattening" a compound into a single label, the author advocates for a binary branching tree. This preserves the internal relationship between components, allowing the system to recognize that "tobacco cultivation" is, fundamentally, a type of "cultivation."
Results and Implementation
The research produced a formal set of Interface Structure (IS) building rules for the Eurotra environment. Key outcomes include:
- Isolation of Translatable Types: The study successfully identified a set of compositionally translatable productive compound types.
- Hierarchical Mapping: The system demonstrates how to map a hierarchical structure in a source language (like Greek) to a matching structure in a target language (like English) without losing internal semantic relationships.
- Solving Locality Issues: The framework ensures that morphology and syntax remain distinct while sharing the necessary information to translate complex terms accurately.
Limitations and Future Directions
Despite its successes, the author identifies several remaining hurdles:
- Semantic Ambiguity: In 1996, the Eurotra environment lacked deep semantic processing. Resolving whether an "apple pie" is a pie made of apples or a pie for apples requires world knowledge that was not yet integrated.
- The Double Analysis Problem: Distinguishing between a true linguistic compound and a "nonce sequence" (a random string of nouns) remains a challenge for automated systems.
- Synthesis Challenges: While the paper focuses on analysis and transfer, determining when a target language prefers a single-word compound versus a multi-word phrase remains a complex task for future MT development.
Summary updated