Towards a resource for multilingual lexicons: an MT assisted and human-in-the-loop multilingual parallel corpus with multi-word expression annotation

arXiv:2011.03783 · cs.CL, cs.AI · Submitted 2026-08-24 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "Towards a resource for multilingual lexicons: an MT assisted and human-in-the-loop multilingual parallel corpus with multi-word expression annotation".

Jane: The paper was written by Lifeng Han, Najet Hadj Mohamed, Malak Rassem, Gareth J. F. Jones, Alan F. Smeaton et al. from Leiden University Medical Center, Leiden University Medical Center, Netherlands and Leiden Institute of Advanced Computer Science, Leiden University, Netherlands and University of Tours, France and University of Sfax, Tunisia and Institute for Natural Language Processing (IMS), University of Stuttgart, Germany and ADAPT Research Centre, Dublin City University, Ireland and Insight Centre for Data Analytics, Dublin City University, Ireland and University of Manchester, The United Kingdom.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Introduction: Tom: We're talking about this paper, "Towards a resource for multilingual lexicons: an MT assisted and human-in-the-loop multilingual parallel corpus with multi-word expression annotation," which is a huge contribution to anyone working on machine translation. The authors are essentially building a massive database that should be helpful for future researchers.

Jane: It’s really important that you mentioned the full title, Tom, because it tells us exactly what they're doing: combining the power of machine translation with human expertise—the "human-in-the-loop" part—to create this resource. It’s not just about letting AI do everything.

Meng: And I find that really interesting because, instead of just taking raw MT output, they are systematically improving it by having human post-editing and annotate the target MWEs across multiple languages. This gives us a clearer picture than what you get from a standard automated translation run.

Lu: The scope of the languages is also impressive; covering Arabic, Chinese, German, Italian, and Polish adds so much to complexity that we usually don't see in single-language studies. It shows an ambitious effort to capture global variation.

Lalam: From a cultural standpoint, that diversity means we are actively collecting data from different regions and dialects—especially the Egyptian and Tunisian variations for Arabic—which allows us to move past generic translations and understand actual human speech patterns.

Tom: That’s right, Lalam, because as they build this "AlphaMWE" corpus, it looks like a massive resource for cross-lingual research. It's not just about one language anymore; it's a whole system of translation and cultural exchange.

Jane: The core idea is to create a parallel corpus where the MWEs are bilingually and multilingually aligned manually, ensuring that the meaning is preserved across different linguistic boundaries, which is vital for accurate communication.

Meng: If we can use this data for training models to handle idiomatic phrases correctly, it could fundamentally change how much work we have to do on post-editing later on the massive amount of content online. The potential impact is huge.

Lu: I think that's exactly the kind of leap forward that will allow us to see the true complexity of human language in a way that was previously impossible with standard automated methods alone.

Lalam: It provides the data needed to bridge cultural gaps, allowing us to access nuances in literature and communication that simply don't translate word for word because they are embedded in those multi-word expressions.

Tom: It sounds like this "AlphaMWE" corpus is doing a lot of heavy lifting, but how does this research point toward future improvements in machine translation? We are really looking at the practical impact here.

Improvements: Jane: The paper points out several specific categories of errors that MT systems make when dealing with MWEs. It goes far beyond just getting a literal translation wrong; it implies a failure to understand the context of why people use those phrases, which is fascinating.

Tom: That’s interesting because it suggests that the AI isn't just struggling with vocabulary; it’s failing at understanding the deeper meaning behind these idiomatic phrases. It’s not just a word replacement issue.

Meng: The paper is suggesting that we need to build systems that are not just sentence-level translators, but model how these multi-word chunks function in the real communication flow, rather than treating them as individual words.

Lu: I think this gives us a blueprint for designing next-generation AI models that understand the structure of language, rather than just predicting the next most likely word based on surface features. It provides a detailed map of failure modes.

Lalam: From a cultural perspective, this means we can move beyond surface translation and capture how an idiom reflects a shared human experience in different societies, which is something that current models almost all'miss.

Tom: It’s not about building one better model, but many distinct improvements tailored to the specific failure modes identified in "Towards a resource for multilingual lexicons." The solutions are specific.

Jane: It offers strategies to prevent us from just relying on the first candidate translation, which is often what happens when we use current MT tools. We now have a way to evaluate and improve that reliance.

Meng: And by focusing on these error types, it gives developers a much better roadmap for how to fix the weaknesses in their training data and model architectures. It's concrete guidance for engineers.

Lu: I think this allows for a deeper level of intervention than just "making the AI smarter." It’s about targeted architectural improvements based on specific error types that we can quantify.

Lalam: By improving our understanding of MWEs, we are improving how we understand human intent across language barriers, which is a profound cultural advancement in communication.

Tom: So, it’s not just about building one better model; it's about many improvements informed by the "AlphaMWE" data and the categorization they provide.

Conclusion: Tom: We’ve covered so much ground—from the initial idea of building this corpus to understanding the specific types of failures in MWE translation. The sheer scope is impressive.

Jane: It’s clear that this project is providing a comprehensive set of tools, combining MT assistance with human expertise, and we have some really promising data points on how it all works, which gives us confidence in the results.

Lu: The sheer volume and detail in "AlphaMWE" will be an asset for decades to come, helping us explore the limits of language processing in ways we haven't been able to before.

Meng: I am particularly interested in how this will allow AI to move from simply predicting the next word to actually understanding the whole phrase as a single unit of meaning. That’s where the real power lies.

Lalam: It’s a gift that allows us to appreciate the depth and richness of global communication, not just its surface level structure, giving us a better view of human expression.

Tom: Before we go, I want to hear final thoughts on "Towards a resource for multilingual lexicons: an MT assisted and human-in-the-loop multilingual parallel corpus with multi-word expression annotation." What is the biggest implication here?

Lu: It’s the kind of work that allows us to see the true complexity of human language in a way that is truly groundbreaking, pushing past what we used to think was possible.

Meng: I think it offers a practical, actionable roadmap for making real machine translation improvements because it provides quantifiable data on exactly where and why systems are failing.

Lalam: It provides the bridge between cultural understanding and technological capability, helping us build tools that respect the nuances of human experience worldwide.

Tom: And I think we can all agree that this "AlphaMWE" corpus is going to be a powerful tool for everyone involved in language tech research.

Final Thoughts: Jane: So, wrapping up our thoughts on this amazing paper, "Towards a resource for multilingual lexicons: an MT assisted and human-in-the-loop multilingual parallel corpus with multi-word expression annotation," it really feels like we have seen a massive leap forward for how AI handles language across cultures.

Tom: Exactly, Jane. What struck me most is how this methodology doesn's only give us data; it gives us a framework for making sure that data is *good* and contextually rich, which is so vital when dealing with tricky idioms and nuances.

Lu: You're right, Jane; the sheer granularity of annotation they’ve achieved—especially tagging those multi-word expressions—that blows open possibilities for AI to handle linguistic ambiguity in ways we barely thought possible before.

Meng: But from an implementation standpoint, Lu, if we can reliably access this level of quality control and structure across so many languages, it means the training data side of things gets fundamentally easier and more robust for commercial deployment.

Lalam: And that robustness has a ripple effect far beyond just better software; it means that the rich cultural tapestry embedded in those unique lexicons can finally be accurately reflected by machines, making global communication truly empathetic.

Tom: I agree with Lalam; it’s about moving past just functional translation toward genuine understanding, which is such a huge step for the field.

Jane: It’s incredibly encouraging to see how the human-in-the-loop aspect is integrated so seamlessly into creating this corpus, rather than being an afterthought.

Lu: Honestly, I think this resource sets a new benchmark; other groups are going to have to match this level of meticulousness if they want their systems to be considered state-of-the-art.

Meng: It definitely shifts the burden of quality from just the raw data collection phase to a structured, repeatable annotation process, which is what we engineers crave.

Lalam: Ultimately, making that multilingual lexicon resource available means that language itself becomes less of a barrier and more of a connective tissue for humanity.

Tom: What an incredible wrap-up; this paper truly gives us so much to think about regarding the future of human-computer interaction through language.

Jane: Thanks so much to all of you for chatting through this today; it was genuinely fascinating.

Lifeng Han, Najet Hadj Mohamed, Malak Rassem, Gareth J. F. Jones, Alan F. Smeaton, Goran Nenadic

Leiden University Medical Center, Leiden University Medical Center, Netherlands · Leiden Institute of Advanced Computer Science, Leiden University, Netherlands · University of Tours, France · University of Sfax, Tunisia · Institute for Natural Language Processing (IMS), University of Stuttgart, Germany · ADAPT Research Centre, Dublin City University, Ireland · Insight Centre for Data Analytics, Dublin City University, Ireland · University of Manchester, The United Kingdom

cs.CL, cs.AI

Submitted: 2026-08-24

Updated: 2026-08-25

Code: https://github.com/aaronlifenghan/AlphaMWE

Project page: https://www.deepl.com/en/translator

Importance score: 94/100

The gist: The paper introduces "AlphaMWE," a multilingual parallel corpus that is constructed using a combination of machine translation (MT) assistance and human-in-the-loop methodology, and which features

Key concepts

Multilingual Parallel Corpus
A large database containing texts in multiple languages that are aligned to each other. This resource is vital for training advanced AI models, ensuring that the meaning and context are preserved when translating between different linguistic boundaries.
Multi-word Expression (MWE)
Phrases or idiomatic chunks of words that cannot be translated by treating them as individual words. The corpus focuses on annotating these MWEs to help AI understand the phrase as a single unit of meaning, rather than just a sequence of words.
Human-in-the-Loop
A methodology where human experts systematically review and improve raw machine translation (MT) output. This process involves human post-editing and annotation, ensuring the data is accurate and contextually rich before it is used for AI training.
AlphaMWE Corpus
The specific name given to the massive resource being built by the authors. It serves as a comprehensive dataset for cross-lingual research, designed to improve machine translation by capturing global variation and cultural nuances.

Terminology

Summary

The paper introduces AlphaMWE, a multilingual parallel corpus that is constructed using a combination of machine translation (MT) assistance and human-in-the-loop methodology, and which features annotations of multi-word expressions (MWEs).

Corpus Scope and Content:

The corpus covers several languages: Arabic, Chinese, English, German, Italian, and Polish. Notably, the Arabic corpus includes both standard and dialectal variations from Egypt and Tunisia. The MWEs included are specifically verbal MWEs (vMWEs) defined in the PARSEME shared task that have a verb as the head of the studied terms.

Corpus Construction Methodology:

The creation process began with an original English source monolingual corpus, which was extracted from the PARSEME shared task in 2018. The methodology involved several steps:

  1. Machine Translation (MT): The source corpus was processed using state-of-the-art MT systems.

  2. Human Post-Editing and Annotation: Following the MT output, human professionals performed post-editing and annotated the corresponding target MWEs.

  3. Quality Control: Strict quality control was applied for error limitation, where each sentence underwent a rigorous process: each MT output sentence received first manual post-editing and annotation plus a second manual quality rechecking till annotators’ consensus is reached.

Evaluation and Findings:

The research encountered challenges in the accurate translation of MWEs, which was reflected through the outcomes of the human-in-the-loop metric known as HOPE. To address this, the authors categorize a categorization of the error types encountered by MT systems in performing MWE-related translation.

To gain a comprehensive view of MT issues, four popular state-of-the-art (SOTA) MT systems were selected for comparison: Microsoft Bing Translator, GoogleMT, Baidu Fanyi, and DeepL MT.

Utility and Conclusion:

The authors conclude that due to the high quality control applied—including noise removal, translation post-editing, and manual annotation—the AlphaMWE data set will be an asset for both monolingual and cross-lingual research, specifically mentioning applications in multi-word term lexicography, MT, and information extraction.

Improvements for AI systems

Based on the analysis of the AlphaMWE corpus and its associated findings, I propose several specific architectural and training improvements to current State-of-the-Art (SOTA) AI translation systems. These enhancements move beyond simple sentence alignment toward deep lexical and contextual comprehension.

The Problem: Current NMT models often decompose MWEs into individual lexemes, losing the idiomatic or functional meaning (e.g., translating cutting capers literally instead of as a single concept).

The Improvement: Implement a Lexical Atomic Unit Layer. This layer treats identified MWEs (especially vMWEs) as single, non-decomposable tokens during the encoding and decoding phases. The system will utilize the AlphaMWE lexicon to ensure that when an input sequence matches a known MWE, the corresponding target translation is retrieved as a single unit.

What the Improved System Can Do:

  • Accurate Idiomatic Translation: Successfully translate idioms (e.g., cutting capers) into their culturally equivalent target-side MWEs (e.g., 欢呼雀跃 in Chinese), achieving functional equivalence rather than literal translation.

  • Semantic Preservation: Maintain the intended meaning of complex phrases regardless of how they are tokenized by ensuring the core semantic unit is preserved across languages.

The Problem: SOTA models operate primarily at the sentence level, leading to Coherence-unaware ambiguity (e.g, failing to translate de-gnoming because it requires knowledge of the preceding context).

The Improvement: Integrate a Document/Paragraph Context Window. The encoder must be expanded to capture not just the immediate source sentence, but a look-ahead and look-back window encompassing several surrounding sentences. This allows the the model to resolve ambiguous MWEs based on local discourse.

What the Improved System Can Do:

  • Resolve Ambiguity: Correctly interpret MWEs like de-gnoming (a literary term) by referencing surrounding context, even if the word itself is not a standard dictionary entry.

  • Maintain Cohesion: Ensure translation choices are consistent across paragraphs, preventing semantic drift or contradictory translations within a single text block.

The Problem: Current models are error-agnostic, failing to learn from specific failure modes like Super Sense or Abstract Phrase errors.

The Improvement: Implement Targeted Error Correction Training (TECT), leveraging the detailed error categories defined in the AlphaMWE corpus and evaluated by the HOPE metric. This involves fine-tuning a specialized loss function that penalizes common MWE failure modes (e.g, literal translation of idioms) far more heavily than standard token mismatch errors.

What the Improved System Can Do:

  • Mitigate Figurative Misinterpretation: Recognize and correct instances where the source text uses figurative language (metaphor/super sense), automatically selecting the corresponding idiomatic target MWE rather than a literal word-for-word translation.

  • Improve Abstract Concept Translation: Successfully translate abstract concepts (a measure of peace, salutatory emptiness) into culturally appropriate target expressions, rather than translating them literally.

The Problem: Models exhibit gender bias (defaulting to male forms in certain languages) and struggle with multi-character names or specific proper nouns.

The Improvement: Incorporate a Gender/Entity Constraint Module. This module requires explicit input on gender, age, or known entity status (e.g., Absalom is a proper name). The model will then apply constraints to ensure the correct corresponding grammatical form is selected in the target language (e.g., choosing przyjaciółka over przyjaciel in Polish).

What the Improved System Can Do:

  • Ensure Gender Accuracy: Correctly select gendered pronouns or nouns when context demands it, eliminating systematic bias inherent in large training datasets.

  • Handle Proper Nouns Robustly: Maintain multi-character names (e.g., Absalom, Herodian) without misspelling them or incorrectly segmenting them into multiple words, ensuring accurate recognition across languages.

The Problem: The system relies on general language models, missing the specialized lexical knowledge of specific linguistic phenomena (like Light Verb Constructions/LVCs).

The Improvement: Integrate a Multilingual MWE Knowledge Graph. This graph stores verified alignments between source-language MWEs and target-language idiomatic equivalents across all languages in the AlphaMWE set. This is not just a dictionary, but a structured relationship map.

What the Improved System Can Do:

  • Cross-Lingual Equivalence: Identify and apply corresponding MWE structures even when translating between languages that share minimal lexical overlap, ensuring the function of the MWE is preserved, not just its meaning.

Sources

Related papers