Cross-Lingual Mapping of Tense and Aspect

53 concepts32 papers1 proposal

When translating a sentence like "I ate rice" into a "tenseless" language like Chinese, the verb itself does not change form. Conversely, when translating from Chinese back into English, a machine translation system must reconstruct grammatical inflections (such as the suffix -ed) that do not exist in the source text.

This fundamental mismatch is known as cross-linguistic temporal asymmetry. Different languages use entirely distinct mechanisms to convey time: some rely on explicit morphological inflections (tense), others on lexical modifiers like adverbs or aspectual particles, and others on rich pairs of verbs that distinguish ongoing from completed actions. Because standard natural language processing models are often "tense-blind," mapping these disparate systems remains a major hurdle for translation, dialogue, and temporal reasoning systems.

The Challenge of Asymmetric Encoding

To understand why standard models fail, we must separate tense (when an event occurs relative to the present) from aspect (how the action unfolds over time, such as whether it is a state, an ongoing process, or a completed event).

When translating between languages with different grammatical systems, temporal information that is implicit in the source must be made explicit in the target. As Yiming Ai et al. demonstrate, even state-of-the-art neural machine translation systems fail to predict the correct tense roughly 20% of the time. This failure distorts the chronological sequence of events, even when the general meaning of the sentence remains recognizable.

[Source: Tenseless Language] ---> "Yesterday he eat rice" (Implicit Tense)
                                        |
                                 [MT Architecture] 
                                        |
[Target: Inflected Language] <--- "Yesterday he eats rice" (Incorrect Explicit Tense)

This problem is compounded by structural and category shifts. As Yun Sun et al. observe through the lens of Translation Shifts Theory, moving between linguistically distant languages frequently requires shifting from grammatical inflections to lexical markers (such as adding specific adverbs to denote completed actions). Furthermore, the lack of direct one-to-one mappings between tenses means that nuances in duration and time reference require complex compositional transformations to preserve meaning across language boundaries, as shown by Lisette Appelo.

Aspectual Divergence and Lexical Semantics

A core reason translation systems stumble is aspectual divergence—a concept formalized in a single-author paper by Bonnie J. Dorr. Aspectual divergence occurs when a single verb in one language maps to multiple potential verbs or grammatical structures in another based entirely on completion or duration.

To resolve this ambiguity, a system must understand the internal temporal structure of an action, specifically its telicity (whether the action has an inherent endpoint). For example, knowing whether a sentence contains a destination allows a system to correctly choose between telic (completed) and atelic (ongoing) verb forms in the target language. As Bonnie J. Dorr et al. later demonstrated, deriving these aspectual properties at both the individual verb level and the full sentence level prevents the loss of factual accuracy during generation.

To achieve this, models must be able to classify the underlying nature of a predicate. The single-author work by Eric V. Siegel utilized machine learning to automatically classify verbs as either steady states (e.g., "seems") or dynamic events (e.g., "ran"). Identifying this distinction is a necessary prerequisite for selecting the appropriate grammatical tense in an inflectional target language. Even simple temporal prepositions can cause failures; as Manabu Sasayama et al. note, prepositions that share a single surface translation may imply completely different temporal boundaries, requiring deeper semantic normalization to preserve the intended timeline.

Reconstructing Latent Timelines

When the source language lacks explicit morphological markers, how can an automated system accurately infer the intended timeline? Researchers have developed two primary strategies: latent label projection and contextual sequence modeling.

Latent Label Projection

One powerful intuition is to leverage parallel corpora to "project" temporal information from a language that requires it to one that does not. Feifan Liu et al. introduced a method that aligns English-Chinese parallel text, projects the explicit English tense inflections onto the corresponding tenseless Chinese verbs, and uses this bootstrapped data to train a standalone tense predictor.

[English Parallel Text] ---> "She walked [PAST] to the store"
                                    |  (Project Tense Label)
                                    v
[Chinese Parallel Text] ---> "她 [PAST] 走路 去 商店" ---> Train Tense Predictor

Contextual Sequence Modeling

Because tense is rarely a localized phenomenon, systems must track the flow of time across sentences. Zhengxian Gong et al. and Min Zhang et al. developed n-gram and classifier-based tense models that track tense sequences both within a sentence (intra-tense) and across an entire document (inter-tense). By leveraging surrounding contextual clues, discourse markers, and time adverbs, these frameworks allow machine translation pipelines to automatically infer the temporal location of situations in tenseless text, a roadmap originally laid out by Nianwen Xue.

The Evaluation Blindspot and Modern Benchmarks

Why does this problem persist despite massive advances in deep learning? A major culprit is the way we evaluate language models.

Standard evaluation metrics like BLEU and COMET are fundamentally "tense-blind." As Yiming Ai et al. point out, these metrics prioritize surface-level word overlap and general semantic similarity over precise grammatical conjugation. A model can completely flip a past-tense verb into the future tense, yet receive a high BLEU score because the surrounding nouns and adjectives match.

Furthermore, early temporal annotation standards like ISO-TimeML were heavily optimized for English syntax, failing to capture the nuances of languages with complex inflected modals or rich aspectual pairs (André Bittar et al.). This English-centric bias extends to information extraction; traditional temporal relation extraction (TEMPREL) and normalization pipelines frequently rely on rule-based scripts tailored to English date formats, making cross-lingual generalization difficult (Sen Yang et al.; Haoyu Wang et al.). Temporal expressions are also deeply rooted in cultural contexts, meaning a model trained on one culture's schedule may fail to accurately reason about appropriate timings in another (Yuiko Tsunomori et al.).

To fix this blindspot, recent work focuses on building explicit, cross-linguistic aspectual resources and benchmarks:

  • Slavic Aspectual Pairs: Ranka Stanković et al. developed datasets mapping explicit verbal aspectual pairs for Bosnian, Croatian, and Serbian, ensuring that the subtle distinctions between ongoing and completed actions are preserved in downstream NLP tasks.
  • Multimodal Temporal Reasoning: The challenge is not restricted to text alone. The authors of the Perfect Times dataset (Deep Temporal Reasoning in Video Language Models) evaluated state-of-the-art video-language models across English, Italian, Russian, and Japanese. They discovered that even advanced models like GPT-4o and Gemini struggle to distinguish between completed and ongoing actions when forced to navigate diverse cross-linguistic encodings of time.

By shifting focus from surface-level vocabulary matching to the underlying logic of tense and aspect, this research direction aims to build models that truly understand when and how an action takes place, regardless of the language in which it is expressed.

Key Foundational Work

The Specification of Time Meaning for Machine Translation (1985) — contributed 4 concept elements to this direction.
The Transfer of Finite Verb Forms in a Machine Translation System (1982) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_cross_linguistic_aspectual_mapping_93e33c53,
  title = {cross-linguistic aspectual mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-linguistic-aspectual-mapping/rdr_93e33c53d3f94c09a97b6076f06b63fe},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_cross_linguistic_aspectual_mapping_93e33c53_v1,
  title = {cross-linguistic aspectual mapping},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/cross-linguistic-aspectual-mapping/rdr_93e33c53d3f94c09a97b6076f06b63fe/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake