Cross-Lingual Explicitation of Implicit Discourse Relations

49 concepts24 papers1 proposal

When we speak or write, we rarely state every logical connection explicitly. We might say, "It rained; I stayed home," leaving the reader to infer a causal link. In linguistics, these "hidden" connections are known as implicit discourse relations.

While humans navigate these gaps effortlessly, they present a fundamental challenge for translation. Different languages have vastly different "densities" of explicit markers. Translating between them requires more than just swapping words; it requires surfacing the underlying logic—a process known as explicitation.

The Discourse Density Gap

The core of the problem lies in the distinction between parataxis and hypotaxis. As Zhiyuan Liu explains in English and Chinese Thought Patterns and Their Impact on Translation Teaching, analytical languages like English favor hypotaxis, using explicit logical connectors (like "because" or "although") to signal relationships. In contrast, synthetical languages like Chinese often rely on parataxis, where logic is implied through the mere sequence of ideas.

This is not a minor stylistic preference. Yuping Zhou et al. found in PDTB-style Discourse Annotation of Chinese Text that over 80% of rhetorical relations in Chinese lack explicit markers. When a language omits connectives in nearly three-quarters of instances, a translation system that only looks at surface-level words will inevitably miss the "connective tissue" of the text.

The Explicitation Hypothesis

Human translators naturally bridge this gap. According to the Explicitation Hypothesis, translators tend to make discourse relations clearer in the target text than they were in the source. Frances Yung notes in Towards a discourse relation-aware approach for Chinese-English machine translation that approximately 25% to 30% of implicit relations in a source text are rendered as explicit connectives in the target language to ensure naturalness.

This process isn't just about adding words; it's about specification. Frances Yung et al. demonstrate in Investigating Explicitation of Discourse Connectives in Translation using Automatic Annotations that translators often replace vague markers (like "and") with precise terms (like "furthermore" or "but") to reduce semantic ambiguity. This "sharpening" is particularly common in contrastive relations, where failing to signal an opposition could lead to reader confusion.

Why Machines Struggle with Hidden Logic

Current machine translation (MT) systems often fail this test of coherence. As Jingjing Feng et al. argue in Research on Common Errors and Post Translation Editing Strategies in Chinese Japanese Machine Translation, MT models frequently preserve the omission of conjunctions found in the source. The result is a series of "disconnected clauses" that are grammatically correct but logically disjointed.

The difficulty is twofold:

  1. Inference: The model must first "see" a relation that isn't linguistically present.
  2. Non-Projectability: As shown by Sárka Zikánová et al. in DiscoGeM 2.0: A Parallel Corpus of English, German, French and Czech Implicit Discourse Relations, discourse relations are not always preserved across languages. A temporal link in one language might be reinterpreted as a causal link in another, meaning the rhetorical structure itself can shift during translation.

Modeling the "Connective Tissue"

To solve this, researchers are moving beyond sentence-to-sentence mapping. One approach is source-side explicitation: manually or automatically inserting missing connectives into the source text before translation. While Kevin Duh et al. found in Crosslingual Annotation and Analysis of Implicit Discourse Connectives for Machine Translation that simple text augmentation provides only modest gains, it has opened the door to more sophisticated architectural changes.

In Towards Discourse-Aware Document-Level Neural Machine Translation, Xin Tan et al. propose Disco2NMT, which models Rhetorical Structure Theory (RST) trees as dependency graphs. By using Graph Convolutional Networks (GCNs), the model learns the hierarchical logic of a document, allowing it to generate the appropriate "connective tissue" in the target language.

Furthermore, the human tendency to explicitate provides a clever source of training data. Wei Shi et al. suggest in Using Explicit Discourse Connectives in Translation for Implicit Discourse Relation Classification that we can automatically label implicit relations by looking at where human translators added connectives in parallel corpora. By back-translating these explicit markers, we can teach models to recognize the "hidden" logic that humans naturally perceive.

Key Foundational Work

The Automatic Translation of Discourse Structures (2000) — contributed 1 concept elements to this direction.
Disambiguating temporal-contrastive connectives for machine translation (2011) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 24 papers
  2. Version 1 · 2026-10-04 · 17 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_implicit_discourse_relations_8de1d6b3,
  title = {implicit discourse relations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-discourse-relations/rdr_8de1d6b31e894f4f8be2e1cb610fc1f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_implicit_discourse_relations_8de1d6b3_v2,
  title = {implicit discourse relations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/implicit-discourse-relations/rdr_8de1d6b31e894f4f8be2e1cb610fc1f5/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake