💬 NLP

Translating the Untranslatable: An Operationalizable Ontology for Untranslatability

This paper introduces a structured ontology and taxonomy of untranslatability to operationalize the concept for NLP, resulting in a new multilingual dataset and initial findings that demonstrate translation quality improves when models employ explanatory compensation strategies.

Original authors: Jacob Bremerman, Brihi Joshi, Hirona Arai, Xiang Ren, Jonathan May

Published 2026-06-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jacob Bremerman, Brihi Joshi, Hirona Arai, Xiang Ren, Jonathan May

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: When Translation Hits a Wall

Imagine you are trying to move a piece of furniture from your house to a new one. Usually, you just pick it up and carry it over. That's how most computer translation works today: it moves words from Language A to Language B.

But sometimes, the furniture is shaped weirdly, or the doorways in the new house are too small. You can't just carry it through; you have to take it apart, wrap it in bubble wrap, or even leave a note explaining why it looks different.

This paper is about those "weirdly shaped" moments in language. The authors call this untranslatability. It happens when a word or phrase in one language (like Japanese or Spanish) carries a meaning, a joke, or a cultural feeling that simply doesn't exist in English. If you try to translate it word-for-word, you lose the soul of the message.

The Problem: Computers Are Too Literal

Current AI translators are like very strict librarians. They are great at finding the exact dictionary definition of a word. But they struggle when a word is more like a feeling, a pun, or a cultural inside joke.

For example, the paper mentions the Japanese internet slang "wwww" (which looks like grass and means "lol"). An AI might try to translate it as "grass," which makes no sense to an English speaker. Or it might translate a Spanish word for "time spent chatting after a meal" (sobremesa) as just "after dinner," losing the cozy, social vibe of the original word.

The authors argue that we need to stop treating these as "bugs" or rare mistakes. Instead, we should treat them as a specific category of problem that needs a special toolkit.

The Solution: A New Map and a Toolkit

To fix this, the researchers built two main things:

1. The Map (The Ontology)
They created a structured map to categorize why something is untranslatable. They broke it down into three main neighborhoods:

Linguistic: The grammar or sounds are different (like a tongue-twister that doesn't work in another language).
Figurative: It's about jokes, puns, or idioms (like saying "it's raining cats and dogs").
Cultural: It's about history, religion, or customs (like a greeting that involves kissing on the cheek, which might be weird in another culture).

2. The Toolkit (Compensation Strategies)
Since you can't just "carry" the meaning over, you need a strategy to get the message across. The authors defined six ways to handle this, like different tools in a toolbox:

Annotation (The "Note" Strategy): Translate the word, but add a little explanation in parentheses. Example: "We enjoyed the sobremesa (a long chat after dinner)."
Adaptation (The "Swap" Strategy): Replace the original joke with a different joke that makes the audience laugh in the same way, even if the words are totally different.
Borrowing (The "Keep It" Strategy): Just keep the original foreign word and hope people learn it. Example: "We had a nice feng-shui."
Paraphrase (The "Explain It" Strategy): Throw away the literal words and just explain the meaning in plain English.
Options (The "Maybe" Strategy): If the original sentence is vague, give the reader multiple choices. Example: "He/She went home."
Calque (The "Literal Copy" Strategy): Translate the words exactly as they are, even if it sounds weird, so the reader can guess the meaning.

The Experiment: Testing the Tools

The researchers didn't just talk about this; they built a massive dataset. They took 1,300 tricky sentences from Spanish and Japanese and generated 18,200 different English translations, each using a different strategy from their toolkit.

Then, they asked real humans to act as judges. They showed people the original sentence and the different translations, asking: "Which one do you like better?"

What They Found

The results were surprising and important:

People love the "Note" strategy: The most popular translation was usually the one that included an explanation (Annotation). People preferred knowing why a word was used, even if it made the sentence longer.
Context matters: If the translation was for a textbook (where learning is key), people loved the explanations. If it was for a movie (where flow is key), they wanted it shorter, but they still preferred explanations over confusing literal translations.
One size does not fit all: A strategy that works for a pun might fail for a cultural custom. The best translation depends on what is being translated and where it will be read.

The Takeaway

The paper concludes that machine translation is currently trying to force a square peg into a round hole. By recognizing that some things cannot be translated directly, and by giving computers a structured list of strategies (like adding notes or swapping jokes) to handle those moments, we can make translations that actually make sense to humans.

They aren't just fixing errors; they are teaching computers how to be more like human translators: flexible, creative, and willing to add a little extra context to make sure the message is truly understood.

Technical Summary: Translating the Untranslatable

Problem Statement

While machine translation (MT) and large language models (LLMs) have achieved high performance on standard benchmarks, significant challenges remain in "long-tail" cases where meaning cannot be fully preserved across languages. These cases, termed untranslatability, arise from linguistic, cultural, or stylistic mismatches where no single target sentence can capture all aspects of the source meaning. Current NLP approaches often treat translation as a one-to-one mapping between semantically equivalent texts, failing to address scenarios where direct translation is impossible. Existing resources in linguistics and translation studies provide theoretical definitions but lack large-scale, operationalizable frameworks suitable for computational experimentation.

Methodology

1. Structured Framework

The authors introduce a formal framework to decompose untranslatability into two core components:

Untranslatability Types (uTypes): A structured ontology categorizing the source of the mismatch into three domains:
- Linguistic: Novel Words, Specificity, Syntax, Verb Forms, Other Inflections, Phonetics, Orthography.
- Figurative: Puns, Idioms.
- Cultural: Religion, History, Customs.
Compensation Strategies (cStrats): A taxonomy of six techniques used to convey meaning when direct translation fails:
- Adaptation (AD): Replacing source elements with target equivalents to preserve effect (e.g., humor).
- Annotation (AN): Adding explanatory context or footnotes.
- Borrowing (BO): Retaining the source-language term.
- Calque (CA): Literal translation to preserve structure.
- Options (OP): Providing multiple valid translations for underspecified sources.
- Paraphrase (PA): Re-expressing meaning in plain target-language wording.

The authors propose a reformulation of the translation probability $P(e|f)$ to include the identification of the uType ( $u$ ), the context ( $m$ ), and the selection of a compensation strategy ( $c$ ), rather than treating translation as a simple token-level prediction.

2. Dataset Construction

To operationalize this framework, the authors constructed a multilingual dataset involving Spanish and Japanese as source languages and English as the target.

Generation Process: An iterative pipeline combining human expert seed examples with LLM (GPT-4o) generation. Human experts provided initial seeds for each uType and cStrat, which were used to prompt the LLM to generate additional examples and translations.
Scale: The dataset contains 1,300 source sentences per language (100 examples for each of the 13 uTypes). Each source sentence is paired with translations generated using the six cStrats plus a default (unprompted) translation, resulting in 18,200 total translations.
Validation: Human language experts verified a sample of the data, finding approximately 96% of Spanish and 95% of Japanese examples to be valid instances of their assigned uTypes.

3. Human Preference Study

The authors conducted a study to evaluate how compensation strategies affect perceived translation quality.

Setup: Bilingual annotators ranked seven candidate translations (one per strategy plus a default) for various source sentences.
Variables: The study examined preferences across different uTypes, translation contexts (general, textbook, movie), and source languages (Spanish vs. Japanese).
Metrics: Mean Reciprocal Rank (MRR) was used to aggregate annotator rankings.

Key Results

Strategy Preference Varies by Untranslatability Type:
- Annotation (AN) was the most preferred strategy overall. Annotators consistently favored outputs that included additional explanatory context, suggesting a preference for clarity and completeness over strict adherence to conventional translation formats in untranslatable cases.
- Adaptation (AD) performed well for linguistic mismatches (e.g., Novel Words, Verb Forms) but poorly for cultural/figurative mismatches where altering content distorted meaning.
- Borrowing (BO) was generally the least preferred, though it performed relatively better for culturally grounded cases (Customs).
- Options (OP) was highly effective for Specificity mismatches where multiple interpretations exist.
Context Influences Strategy:
- In Textbook contexts (prioritizing accurate understanding), Annotation was significantly more preferred.
- In Movie contexts (prioritizing naturalness and brevity), Annotation was less preferred, though it still outperformed the default translation.
- The default translation (no explicit strategy) was consistently less preferred in context-dependent scenarios, indicating that explicit strategy application better aligns with user expectations.
Source Language Differences:
- While Annotation was the top choice for both Spanish and Japanese, the preference was stronger for Japanese. This reflects the greater linguistic and cultural distance between Japanese and English, necessitating more explicit explanation.
- Adaptation was much more preferred for Spanish than for Japanese, suggesting that structural and cultural differences make adaptation more difficult for Japanese.
Downstream Utility:
- A preliminary experiment demonstrated that incorporating examples from the dataset improved the accuracy of a classifier tasked with detecting untranslatable sentences compared to a zero-shot baseline (0.761 vs. 0.704 accuracy).

Significance and Contributions

The paper claims to establish the first structured framework for studying untranslatability in NLP, moving the field beyond treating these cases as isolated edge cases. The specific contributions are:

A structured ontology of untranslatability types for machine translation.
A taxonomy of compensation strategies.
A multilingual dataset (18,200 examples) operationalizing this framework.
Empirical evidence showing that translation quality depends on the strategy used, with a consistent human preference for explanatory context (Annotation).
A formulation of strategy-informed machine translation, proposing that translation should be viewed as a structured process involving mismatch identification and strategy selection rather than simple semantic equivalence.

The authors conclude that their work provides a foundation for modeling translation as a process of negotiating meaning across linguistic and cultural boundaries, highlighting a gap between current MT objectives and human expectations.