TL;DR
This paper introduces the MILE corpus, a structured English dataset designed to aid in the creation of machine translation systems for "Less Commonly Taught Languages" (LCTLs). By systematically exploring linguistic meanings that impact morphology and syntax, MILE aims to overcome the data scarcity challenge for LCTLs, enabling better MT training.
What Problem Does This Paper Address?
Building effective machine translation (MT) systems, especially statistical or example-based ones, relies heavily on large amounts of parallel text data. However, for the vast majority of the world's 6,000 languages – the "Less Commonly Taught Languages" (LCTLs) – such resources are scarce or non-existent. This data bottleneck severely hinders the development of MT and other language technologies for these languages. The MILE corpus aims to address this by providing a structured, systematically generated English dataset that can be translated into LCTLs, creating valuable parallel corpora for training MT systems.
Methods
The core of the MILE corpus creation lies in a "Feature Specification" that defines linguistic features (like tense, person, number) and their possible values. This specification is then used to automatically generate "Feature Structure Sets," which represent combinations of these features. These feature structures are the foundation, and English sentences are subsequently written to correspond to them. Crucially, the process prioritizes meanings known to affect morphology and syntax across languages, even if they aren't explicitly obvious in English. The paper also highlights the development of tools to facilitate this process, including an XML schema, a feature structure generator, and a browser.
Datasets
The paper describes the creation of the MILE corpus, which consists of 12,875 English sentences, totaling 76,202 word tokens. This corpus is designed to be translated into multiple LCTLs, forming parallel corpora. The paper focuses on the English side of this corpus and does not detail any specific LCTL datasets used for evaluation in this publication.
Key Insights and Ideas
- Systematic Meaning Exploration: Instead of relying solely on naturally occurring text, MILE systematically enumerates meanings known to influence language structure. This is crucial for LCTLs where such phenomena might be rare in natural data.
- Feature Structures as the Foundation: The corpus is built from the "ground up" using feature structures that codify meaning, rather than starting with English sentences. This allows for greater flexibility in generating parallel data for multiple LCTLs from a single semantic representation.
- Context Field for Nuance: A "context field" is used to convey meanings not explicitly present in the English sentence itself, such as speaker-hearer relationships, which can be critical for accurate translation in LCTLs.
- Tooling for Scalability: The development of reusable tools for corpus creation is a significant aspect, aiming to make the process of building similar corpora for other LCTLs more efficient.
Key Results
The paper describes the construction of the MILE corpus and the methodology behind it. While quantitative results on MT performance are not presented in this paper, the authors mention small-scale experiments that showed "consistent, accurate translations" for Hebrew and Japanese, indicating the potential effectiveness of this approach. The corpus itself is a key output, providing a structured resource for LCTL research.
Limitations and Future Work
The authors acknowledge that due to sampling and focusing on crucial features, the corpus might miss some important feature combinations in certain languages. They also note that a corpus containing all possible feature combinations would be prohibitively large. Future work includes building a "Corpus Navigation System" to dynamically explore the feature space and using machine learning to identify redundant or essential feature combinations. Additionally, they plan to transition from human sentence writing to using natural language generation systems for corpus creation.
Summary updated