Abstract
Translating grammatical tense and mood between English and German involves complex, non-linear mappings that extend beyond local syntactic rules. This paper presents a large-scale corpus-based contrastive analysis to identify the pragmatic factors—such as genre, register, and speaker intent—that drive human translation choices. By examining over 6 million parallel sentences, the authors argue that current machine translation (MT) architectures must integrate document-level context to overcome the limitations of sentence-level sequence mapping.
Introduction
In Natural Language Processing, translation is frequently modeled as a sequence-to-sequence problem. However, tense and mood are "slippery" categories where a single source form can map to multiple target forms depending on the broader context. For example, the English present perfect "I have read" may translate into various German forms based on whether the text is a news report, a medical document, or a transcript of spoken dialogue. This study addresses the lack of systematic descriptions for these transformations, positing that high-quality MT requires "pragmatic information" that human translators naturally utilize.
Methodology
The researchers developed a robust annotation and alignment pipeline to analyze human translation patterns. Key to this approach was the TMV annotator, which automatically labels tense, mood, and voice. To identify corresponding "Verb Complexes" (VCs) across languages, the study employed Giza++ for word alignment and the Mate parser for dependency parsing. This allowed for a granular comparison of how specific grammatical structures in English are realized in German.
Datasets
The analysis utilized a massive collection of English-German parallel data across four distinct registers:
- News: 272k sentences of journalistic text.
- Europarl: 1.9 million sentences of parliamentary proceedings.
- Crawl: 2.4 million sentences of mixed-domain web documents.
- Pattr: 1.8 million sentences of medical documents.
Results and Analysis
The study reveals a "many-to-many" reality in translation. A primary challenge is the English progressive aspect (e.g., "is reading"), which lacks a direct equivalent in German. The data shows that while approximately 77% of non-progressive English present perfect forms map to German past tenses, only 56% of progressive forms do, indicating that aspect significantly influences tense selection.
Furthermore, the choice of tense is highly domain-specific. In medical texts (Pattr), the German Präsens (present tense) accounts for 97% of usage. In contrast, news and parliamentary debates show a more varied distribution, with the Präteritum (simple past) and Perfekt (present perfect) often serving different stylistic functions depending on the medium—Perfekt for spoken discourse and Präteritum for written narration.
The "Subjunctive Gap" also poses a significant modeling challenge. German's robust subjunctive system (Konjunktiv I and II) is frequently used for reported speech and counterfactuals, whereas English often relies on modals or quasi-subjunctive forms. Additionally, English non-finite clauses (e.g., gerunds) often require translation into finite German clauses, forcing MT systems to "invent" appropriate tense and mood markers that are not explicitly present in the source.
Conclusion and Future Work
The findings suggest that MT systems should move beyond local features toward document-level context. Future research should focus on integrating pragmatic features—such as Aktionsart or temporal ordering—into neural models. Providing users with control over the desired register (e.g., formal vs. narrative) could also significantly improve the naturalness and accuracy of machine-translated text.
Summary updated