Enforcing Lexical Consistency in Document Translation

64 concepts53 papers1 proposal

When humans read or translate a document, we maintain a mental map of the entities, technical terms, and core concepts introduced along the way. If a character is named "Mr. Green" in Chapter 1, we do not expect them to randomly become "Mr. Vert" in Chapter 2. If a patent application refers to a "locking mechanism" in the first paragraph, switching to "fastening device" later on could invalidate the entire legal claim.

Yet, standard Natural Language Processing (NLP) models traditionally operate with a form of "tunnel vision." By treating documents as disconnected bags of individual sentences, these systems frequently suffer from lexical inconsistency—translating or segmenting the exact same source term into different target words across a single text.

Building systems that maintain terminological stability requires moving beyond local sentence boundaries to capture global document context.

The Blind Spot of Sentence-Level NLP

The root cause of lexical inconsistency lies in the independence assumptions of standard sequence-to-sequence models. When translating a sentence, a model maximizes the probability of the target words given only the current source sentence.

As Longyue Wang et al. observed in Exploiting Cross-Sentence Context for Neural Machine Translation, this isolated processing means that a word with multiple valid synonyms may be translated differently from one sentence to the next based purely on minor local syntactic variations.

This failure mode spans multiple domains:

  • Literary Fiction: In long-form narratives, context-blind models break reader immersion by alternating character names, titles, or unique fictional items across chapters, as detailed by Lifeng Liu et al. in CloudSheep System for WMT24 Discourse-Level Literary Translation.
  • Technical and Legal Writing: In highly specialized domains like patent processing, uniform terminology is mandatory. Lekang Jiang et al. showed in Patent-CR: A Dataset for Patent Claim Revision that general-purpose language models frequently struggle with strict consistency, introducing ambiguities that can create severe legal vulnerabilities.
  • Word Segmentation: The issue even impacts low-level tokenization. Hwee Tou Ng et al. demonstrated in A Maximum Entropy Approach to Chinese Word Segmentation that segmentation models can inconsistently split the same character sequence within a single document, treating it as a single word in one instance and multiple words in another.

This problem long went unnoticed by automated benchmarks because traditional evaluation metrics like BLEU evaluate sentences in isolation. To expose this gap, Samuel Läubli et al. pioneered document-level evaluation frameworks in Has Machine Translation Achieved Human Parity? A Case for Document-level Evaluation. Their suprasentential analysis proved that while machines might achieve parity with humans on a sentence-by-sentence basis, they fail drastically at maintaining a stable terminology set across a full document.

Furthermore, Yirong Sun et al. established in Fine-Grained and Multi-Dimensional Metrics for Document-Level Machine Translation that traditional metrics can actively penalize superior, cohesive document-level translations because they fail to reward cross-sentence consistency.

Architectural Memory and Document-Wide Decoding

To resolve this, researchers initially sought to equip translation models with an explicit "memory" of past lexical choices.

One foundational approach was the integration of cache mechanisms. Zhengxian Gong et al. developed a framework in Cache-based Document-level Statistical Machine Translation that tracks recently generated target words, dynamically boosting their probability if the corresponding source terms reappear later in the document. To capture deeper semantic relationships beyond exact word matches, Cristina España-Bonet et al. integrated bilingual word embeddings into the decoding process in Document-Level Machine Translation with Word Vector Models, ensuring semantic uniformity even when words have multiple valid meanings.

As the field transitioned to deep learning, Sameen Maruf scaled this intuition to neural architectures in Document Context Neural Machine Translation with Memory Networks, using memory networks to store and query global context from both the source and target documents during an iterative decoding process.

Sentence-by-Sentence Decoding (Tunnel Vision):
Source S1 ──> [ NMT Model ] ──> Target T1 (Term translated as "Synonym A")
Source S2 ──> [ NMT Model ] ──> Target T2 (Term translated as "Synonym B") ❌ Inconsistent

Document-Wide Decoding (Global Search):
Source S1 ──┐
            ├─> [ Document Decoder ] ──> Optimized Document Text 
Source S2 ──┘   (Rewards uniform reuse of "Synonym A" across boundaries)  µ Consistent

An alternative to changing model architectures is changing how we search for translations. Instead of decoding sentence-by-sentence, Christian Hardmeier et al. introduced document-wide decoding in Document-Wide Decoding for Phrase-Based Statistical Machine Translation. This approach translates an entire document at once, using a stochastic local search (hill climbing) to optimize a global scoring function. By incorporating metrics that explicitly reward topical consistency and penalize unnecessary vocabulary diversity across sentence boundaries, the decoder is guided toward globally cohesive lexical choices (Jörg Tiedemann et al. in Statistical Machine Translation with Readability Constraints).

From Hard Constraints to Soft Preferences

While forcing a model to reuse the exact same word ensures consistency, it can sometimes destroy local grammatical fluency. A term might need to change its form or synonym based on the local syntactic context of a new sentence.

To balance local fluency with global cohesion, Xinglin Lyu et al. proposed modeling consistency as a soft preference rather than a rigid constraint in Modeling Consistency Preference via Lexical Chains for Document-level Neural Machine Translation.

The authors achieved this by constructing lexical chains—sequences of semantically related words that span across the document—and feeding them into a Conditional Variational Autoencoder (CVAE). The CVAE learns a latent distribution representing the document's overall stylistic and terminological preference. Instead of forcing an exact word match, the model dynamically encourages the decoder to select target words that align with the established lexical chain, significantly reducing reader confusion while preserving natural sentence structure.

The Modern Era: Multi-Agent Proofreading

With the advent of large language models (LLMs), the strategy for enforcing consistency has shifted from modifying decoding algorithms to designing multi-agent orchestration pipelines.

Modern long-form translation systems mimic the workflows of professional human translation houses, where a draft is first generated and then passed to a specialized editor to ensure terminological stability (Ayush Maheshwari et al. in UDAAN - Machine Learning based Post-Editing tool for Document Translation).

For example, in the Final Submission of SJTULoveFiction to Literary Task, Haoxiang Sun et al. developed a multi-agent pipeline for translating long chapters of fiction. The system uses a primary translation agent to generate initial text chunks, but routes the output through a dedicated Term Proofreader Agent. This agent maintains a running glossary of characters, locations, and unique items across the entire narrative, automatically scanning the text and correcting any lexical drift before final output.

By separating the creative act of translation from the structured task of terminology enforcement, modern frameworks successfully deliver long-form translations that are both locally fluent and globally coherent.

Key Foundational Work

A Support Tool for Writing Multilingual Instructions (1995) — contributed 1 concept elements to this direction.
A Maximum Entropy Approach to Chinese Word Segmentation (2005) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 53 papers
  2. Version 1 · 2026-10-04 · 55 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_lexical_inconsistency_602f13ad,
  title = {lexical inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexical-inconsistency/rdr_602f13ad7b3840f18b15722ecabdf08b},
  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_lexical_inconsistency_602f13ad_v2,
  title = {lexical inconsistency},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/lexical-inconsistency/rdr_602f13ad7b3840f18b15722ecabdf08b/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