TL;DR
This paper tackles the challenge of accurately translating aspectual verb forms in Slavic languages (Russian and Polish) for machine translation (MT) systems. It proposes a novel "uniqueness-based" approach, arguing that aspect choice hinges on whether an event-situation is presented as unique or not, rather than solely on its boundedness or countability. This framework offers a more nuanced way to handle aspectual distinctions, improving translation quality.
What Problem Does This Paper Address?
Machine translation (MT) systems often struggle with languages that have rich aspectual systems, like Russian and Polish. These languages mark verbs for whether an action is completed (perfective) or ongoing/habitual (imperfective), a distinction not directly present in English or Swedish. Human translators intuitively choose the correct aspectual form based on context, but formalizing this for MT is difficult. Traditional approaches often rely on a "count vs. mass" analogy for events, which fails to explain certain "untypical" aspect uses, such as imperfective verbs for completed actions in specific question contexts. The paper aims to bridge this gap by proposing a more robust linguistic principle for aspect choice that can be implemented in MT.
Methods
The paper introduces a "uniqueness-based" approach to aspect choice. Instead of focusing on whether an event is "countable" or "mass-like," it posits that the core distinction lies in whether the event-situation is presented as unique within the discourse or not.
The authors describe a PROLOG implementation for aspect choice within an MT transfer stage. This system uses a "Predicate Driven Referent Grammar" (PDRG) formalism. The process involves:
- Parsing: Input sentences (Swedish/English) are parsed into three representations:
- Categorial Representation (c_rep): Language-specific syntactic information.
- Functional Representation (f_rep): Traditional grammatical roles (subject, object, etc.).
- Semantic Representation (s_rep): Abstract semantic roles (actor, patient, etc.).
- Uniqueness Cue Identification: The system then analyzes the semantic representation to determine the "uniqueness value" of the event using three types of rules:
- Type 1 Rules: Look for explicit "uniqueness indices" within the functional and semantic representations (e.g., indefinite frequency or durativity adverbials, progressive tenses).
- Type 2 Rules: Compare the current predication with stored "proto_events" (representations of typical event-situations associated with a verb) in a knowledge base.
- Type 3 Rules: Analyze the context and infer the probability of aspect change.
- Aspect Assignment: Based on the identified uniqueness value, the system assigns the appropriate perfective or imperfective aspect to the target language verb.
Datasets
Not explicitly discussed in the paper. The authors mention using a "sample text" from the preface of H.A. Gleason's "An Introduction to Descriptive Linguistics" for illustration and analysis. The system is described as being implemented for translation between Swedish/English and Russian/Polish.
Key Insights and Ideas
- Uniqueness as the Core Principle: The paper argues that the perfective aspect marks event-situations as unique, while the imperfective marks them as non-unique or backgrounded. This is a more powerful generalization than the count/mass analogy.
- Explaining "Untypical" Cases: The uniqueness approach successfully explains why imperfective verbs are used in questions like "Have you already had breakfast?" (where the event is time-limited but the focus is on the state of hunger, not a unique meal).
- Interplay of Cues: Aspect choice isn't based on a single cue but on an interplay of adverbials, verb semantics, and discourse context.
- Language-Specific vs. Universal: The overt aspect marking is language-specific (e.g., prefixes in Russian/Polish), but the underlying conceptual distinction (uniqueness) is proposed as universal.
- MT Transfer Stage: The proposed system operates as an intermediate "transfer" stage in MT, bridging the gap between source language parsing and target language generation.
Key Results
The paper does not present quantitative results from a fully implemented MT system. Instead, it focuses on demonstrating the linguistic principles and the logic of the proposed aspect choice rules through qualitative analysis and examples. The key "result" is the conceptual framework itself and its potential to improve MT accuracy by providing a more linguistically grounded method for aspect assignment. Examples illustrate how the uniqueness rules correctly predict aspect choices in various scenarios.
Limitations and Future Work
- Limited Scope of Implementation: The authors acknowledge that the implemented program is a simplified version, with only a restricted number of cues for aspect change implemented.
- Complexity of Type 3 Rules: Rules for inferring aspect change based on context (Type 3) are noted as the most complicated and require further development.
- Formalizing Proto-Events: Describing "typical events" (proto_events) concisely and coherently is identified as a challenge for future research.
- Infinitival Constructions: Further investigation is needed for aspect choice in constructions with infinitives not preceded by aspectual verbs (like "start" or "finish").
- Further Research: The paper calls for further research into designing appropriate rule hierarchies and describing conditions for aspect change.
The authors believe their "uniqueness-based" approach, while a compromise between different linguistic theories, offers a useful and promising direction for MT systems dealing with aspect.
Summary updated