Imminent Research Grants – Projects Awarded

Rewarding the best projects in language innovation

With the Imminent Research Grants project, each year, Imminent allocates €100,000 to fund five original research projects with grants of €20,000 each to explore the most advanced frontiers in the world of language services. A research grant is assigned to one project in each of the following categories: Language economics – Language data – Machine learning algorithms for translation – Human-computer interaction – Neuroscience of language.
Applications for the 5th edition are now open. You can submit your project here.

2025 – Projects awarded

Building Africa’s First Unified Multilingual Dictionary Platform

Philip Akoda CEO & Co-founder Nkanda Project

Every two weeks, a language dies forever. Across Africa, thousands of languages face extinction due to limited documentation and inaccessible learning resources. While initiatives like the Imminent-funded Yorùbá speech corpus advanced automatic speech recognition (ASR) and text-to-speech (TTS) development, comprehensive lexicographic datasets—the foundational component for natural language understanding (NLU)—remain critically scarce for most African languages. The AFLANG Project (theaflangproject.org) addresses this challenge by creating culturally authentic, interactive digital dictionaries that preserve linguistic heritage while providing structured, machine-readable data essential for NLP advancements. The project is ongoing to expand the Hausa dictionary into Africa’s first unified multilingual dictionary platform with: scalable framework supporting numerous African languages; verified lexicographic data with diacritical marks; native speaker audio recordings capturing authentic pronunciations; rich semantic relationships (synonyms, antonyms, hypernyms, hyponyms, keywords, phonemic transcriptions); cultural and historical context for each entry, embedding language within living heritage; gamified learning features driving intergenerational engagement; offline functionality for connectivity-limited regions.

The First Multilingual Multi-Annotator NLI Benchmark with Ecologically Valid Explanations

Barbara Plank Professor of AI and Computational Linguistics Ludwing Maximilians University Munich

A key to evaluate natural language understanding (NLU) is to consider a model’s ability to perform natural language inference (NLI). However, existing NLI benchmarks suffer from two drawbacks: i) few exists for languages beyond English, ii) most of them contain only labels, no explanations, or only post-hoc explanations. This leaves big gaps opens in truly multilingual NLU that is trustworthy, and can explain why a certain prediction were made or why humans labeled a pair of sentences differently (disagreement or human label variation). This projects wants to build the first multilingual multi-annotator NLI benchmark with ecologically valid explanations: explanations provided by annotators for each label. Such a benchmark can break new grounds not only in multilingual NLU, but also in understanding human label variation holistically, such as allowing to tease apart error from plausible human label variation.

Localizing Multimodal Content Using Generative AI

Simran Khanuja Doctoral Student Carnegie Mellon University

The project revolves around localizing multimodal content using generative AI. For example, an advertisement designed for American audiences may feature symbols, objects, or styles that don’t resonate in Japan or Portugal. Localizing this involves not just translating text, but also identifying and modifying imagery so the content feels natural in a new cultural context. While AI struggles with abstract cultural reasoning (the best AI models only achieve 5-10% accuracy in localizing images for different countries), it excels at following concrete instructions like “change the jaguar to a tiger”. To combine AI’s technical power with human cultural expertise, we’ve developed a human-in-the-loop platform that enables translators to manipulate images with simple text instructions: https://platform.opennlplabs.org/. The translator uploads an image and a text instruction on the changes they want. The system first detects the translator’s intent, runs appropriate models, lets the translator choose the best output, and collects feedback from them, in an iterative fashion. Funding from Imminent would enable us to expand our platform’s capabilities—adding more models, improving functionalities such as replacing multiple objects at once, and scaling to a production-level system. By open-sourcing the platform, we can collect larger datasets, refine our models, and advance research in AI-driven cultural adaptation.

Advancing our Understanding of the Neural Basis of Literal and Metaphorical Language Processing in L2

Francesca Ferroni Post Doctoral Researcher University of Cambridge

Stefana Garello Philosopher of Languege University of Roma Tre

The debate in language studies persists over whether language processing is strictly propositional or also involves a non-propositional dimension (Gallese&Lakoff, 2005). This issue is particularly relevant to second language (L2) comprehension, traditionally viewed as a purely propositional process. However, our recent research challenges this assumption, revealing that the inferior frontal gyrus (IFG) and primary motor cortex (M1) are more engaged when processing both literal and metaphorical sentences in L2 than in the first language (L1) (Garello et al.2024a,b; Ferroni et al.under review). The precise role of these areas remains uncertain—whether their activation is merely epiphenomenal or plays a semantic role in a L2 comprehension. To investigate this, our project employs transcranial magnetic stimulation (TMS) combined with functional magnetic resonance imaging (fMRI) to determine the causal role of these sensorimotor regions in L2 comprehension. We will recruit two groups of participants (n=30each), stimulating or inhibiting IFG and M1. Based on our previous findings, we expect greater involvement of these areas in L2 compared to L1. Furthermore, we hypothesize that modulating these areas through stimulation or inhibition will significantly impact L2 performance, as non-native speakers may rely more on non-propositional processing. This interdisciplinary study, combining Ferroni’s expertise in cognitive neuroscience (PI) and Garello’s linguistic background (Co-PI), aims to produce two scientific papers. Beyond advancing our understanding of the neural basis of literal and metaphorical language processing in L2, this research could inform the development of innovative protocols for AI language tutors and second language learning applications.

Evolving SpeechBrain into a Comprehensive Open-source Framework for Speech and Audio LM

Mirco Ravanelli Assistant Professor Concordia University & Associate Member MILA

SpeechBrain is one of the largest open-source toolkits for speech and audio processing, with over 2.5 million monthly downloads and contributions from more than 150 developers worldwide. Despite its success, the field lacks an open, transparent, and fully replicable speech and audio language model (LM) that can serve as a foundation for a wide range of applications, from automatic speech recognition (ASR) to text-to-speech (TTS), speech enhancement, speech translation. This project aims to evolve SpeechBrain into a comprehensive open-source framework for speech and audio LM, with the idea of creating an “open-source model for everything.” This initiative will further help establish SpeechBrain as the standard open-source framework for speech AI.

Imminent Research Grants

$100,000 to fund language technology innovators

Imminent was founded to help innovators who share the goal of making it easier for everyone living in our multilingual world to understand and be understood by all others. Each year, Imminent allocate $100,000 to fund five original research projects to explore the most advanced frontiers in the world of language services. Topics: Language economics – Language data – Machine learning algorithms for translation – Human-computer interaction – The neuroscience of language.

Apply now

2024 – Projects awarded

CURVATURE-BASED MACHINE TRANSLATION DATASET CURATION

Michalis Korakakis University of Cambridge

Despite recent advances in neural machine translation, data quality continues to play a crucial role in model performance, robustness, and fairness. However, current approaches to curating machine translation datasets rely on domain-specific heuristics, and assume that datasets contain only one specific type of problematic instances, such as noise. Consequently, these methods fail to systematically analyse how various types of training instances—such as noisy, atypical, and underrepresented instances—affect model behaviour.

 To address this the present project proposes to introduce a data curation method that identifies different types of training instances within a dataset by examining the curvature of the loss landscape around an instance—i.e., the magnitude of the eigenvalues of the Hessian of the loss with respect to that instance. Unlike previous approaches,  such proposed method offers a comprehensive framework that provides insights into machine translation datasets independent of model architecture and weight initialisation. Additionally, it is applicable to any language pair and monolingual translation tasks such as text summarisation.

DEVELOPMENT OF MULTILINGUAL MACHINE TRANSLATOR FOR PHILIPPINES LANGUAGES

Charibeth Cheng De La Salle University

The Philippines is an archipelagic country consisting of more than 7,000 islands, and this has contributed to its vast linguistic diversity.  It is home to 175 living, indigenous languages, with Filipino designated as the national language. Within formal education, 28 indigenous languages serve as mediums of instruction, alongside English, which holds official status in business, government, and academia.

The Philippines’ diverse linguistic landscape underscores the need for effective communication bridges.  The present project proposal aims to develop a multilingual machine translation system for at least 7 Philippine languages, aligning with efforts to standardize and preserve indigenous languages. Specifically, this project will focus on the following: First, to collect and curate linguistic data sets in collaboration with linguistic experts and naive speakers to ensure the accuracy and reliability of the translation system. As a second step, to implement machine learning algorithms and natural language processing techniques to train the translation model, considering the low-resource nature of Philippine languages. Finally, to evaluate the efficacy of the developed translation system using standardized metrics and human evaluation.

REALTIME MULTILINGUAL TRANSLATION FROM BRAIN DYNAMICS

Weihao Xia University College London

This project, Realtime Multilingual Translation from Brain Dynamics, is to convert brain waves into multiple natural languages. The goal is to develop a novel brain-computer interface capable of open-vocabulary electroencephalographic (EEG)-to-multilingual translation, facilitating seamless communication. The idea is to align EEG waves with pre-aligned embedding vectors from Multilingual Large Language Models (LLMs). The multi-languages are aligned in the vector space, allowing the model to be trained using only a text corpus in one language. EEG signals are real-time and non-invasive but exhibit significant individual variances. The challenges lie in the EEG-language alignment and across-user generalization. The learned brain representations are then decoded into the desired language using LLMs such as BLOOM that produces coherent text that is almost indistinguishable from text written by humans.

Currently, the primary application targets individuals who are unable to speak or type. However, in the future, as brain signals increasingly serve as the control factor for electrical devices, the potential applications will expand to encompass a broader range of scenarios. 

Read the full research paper

HOW CAN MT AND PE HELP LITERATURE CROSS BORDERS AND REACH WIDER AUDIENCES: A CASE STUDY

Vilelmini Sosoni Ionian University

Researchers studied the usability and creativity of machine translation (MT) in literary texts focusing on translators’ perceptions and readers’ response. But what do authors think? Is post-editing of MT output an answer to having more literature translated especially from lesser-used languages into dominant languages? The study seeks to answer this question by focusing on the book Tango in Blue Nights (2024), a flash story collection about love written by Vassilis Manoussakis, a Greek author, researcher and translator. The book is translated from Greek into English using Translated’s ModernMT system and is then post-edited by 2nd year Modern Greek students at Boston University who are native English speakers and have near native capacity in Greek. They follow detailed PE guidelines developed for literary texts by the researchers.

The author analyses the post-edited version and establishes whether it is fit for publication and how it can be improved. A stylometric analysis is conducted. The study is the first of its kind and wishes to showcase the importance of MT for the dissemination of literature written in lesser-used languages and provide a post-editing protocol for the translation of literary texts.

LANGUAGE MODELS ARE MORE THAN CLASSIFIERS: RETHINKING INTERPRETABILITY IN THE PRESENCE OF INTRINSIC UNCERTAINTY

Julius Cheng University of Cambridge

Language translation is an intrinsically ambiguous task, where one sentence has many possible translations. This fact, combined with the practice of training neural language models (LMs) with large bitext corpora, leads to the well-documented phenomenon that these models allocate probability mass to many semantically similar yet lexically diverse sentences. Consequently, decoding objectives like minimum Bayes risk (MBR), which aggregate information across the entire output distribution, produce higher quality outputs than beam search.

Research on interpretability and explainability for natural language generation (NLG) has thus far almost exclusively focused on generating explanations for a single prediction, yet LMs have many plausible high probability predictions. The proposal aims to adapt interpretability to this context by investigating the question,“ Do similar predictions have similar explanations?” This will be addressed by comparing explanations generated by interpretability methods such as attention-based interpretability, layerwise relevance propagation, and gradient-based attribution across predictions.

The goal of this project is to advance research in interpretability for NLG,  to improve understanding of the generalization capabilities of LMs,  and to develop new methods for MBR decoding.

2023 – Projects awarded

USABILITY OF EXPLAINABLE ERROR DETECTION FOR POST-EDITING NEURAL MACHINE TRANSLATION 

Gabriele Sarti University of Groningen

Predictive uncertainty and other information extracted from MT models provide reasonable estimates of word-level translation quality. However, there is a lack of public studies investigating the impact of error detection methods on post-editing performance in real-world settings. The present project proposal is to conduct a user study with professional translators for two language directions sourced from recent DivEMT dataset. The aim is to assess whether and how error span highlights can improve post-editing productivity while preserving translation quality. There will be a focus on the influence of highlights quality by comparing (un)supervised techniques with best-case estimates using gold human edits, using productivity and enjoyability metrics for evaluation.

Such direction could be relevant to validate the applicability of error detection techniques aimed at improving human-machine collaboration in translation workflows. The  proposal is a reality check for research in interpretability and quality estimation and will likely impact future research in these areas. Moreover, positive outcomes could drive innovation in post-editing practices for the industry.

HUMANITY OF SPEECH

Pauline Larrouy-Maestri Max Planck Institute

Synthetic speech is everywhere, from our living room to the communication channels that connect humans all over the world. Text-to-speech (TTS) tools and AI voice generators aim at creating intelligible and realistic sounds to be understood by humans. Whereas intelligibility is generally accomplished, the voices do not sound natural and lack “humanity,” which impacts users’ engagement in human-computer interaction. 

The present project proposal aims at understanding what a “human” voice is a crucial issue in all domains relative to language, such as computer, psychological, biological, and social sciences. To do so, 1) the timbral and prosodic features that are used by listeners to identify human speech will be investigated, and 2) how “humanness” is categorized and transmitted will be determined. Concretely, a series of online experiments using methods from psychophysics are planned to run. both the speech signal, through extensive acoustic analyses and manipulation of samples, as well as on the cognitive and social processes involved, will be analyzed.

TRACKING INTERHEMISPHERIC INTERACTIONS AND NEURAL PLASTICITY BETWEEN FRONTAL AREAS IN THE BILINGUAL BRAIN

Simone Battaglia University of Bologna

Which is the human brain network that supports excellence in simultaneous spoken-language interpretation? Although there is still no clear answer to this question, recent research in neuroscience has suggested that the dorsolateral prefrontal cortex (dlPFC) is consistently involved in bilingual language use and cognitive control, including working memory (WM), which, in turn, is particularly important for simultaneous interpretation and translation. Importantly, preliminary evidence has shown that functional connectivity between prefrontal regions correlates with the efficient processing of a second language.

The present project proposal aims to characterize space-time features of interhemispheric interactions between left and right dlPFC in bilingual healthy adults divided into two groups of professional simultaneous interpreters and non-expert bilingual individuals. In these two groups, cutting-edge neurophysiological methods are used for testing the dynamics of cortico-cortical connectivity, namely TMS-EEG co-registration, focusing on bilateral dlPFC connectivity. The procedure will allow to non-invasively stimulate the dlPFC and track signal propagation, to characterize the link between different aspects of language processing, executive functions, and bilateral dlPFC connectivity. novel insights into the neural mechanisms of interhemispheric communication in the bilingual brain are provided and characterize the pattern of connectivity associated with proficient simultaneous interpretation.

OPEN-SOURCING A RECENT TEXT TO SPEECH PAPER

Phillip Wang

Open source implementations of scientific papers are one of the essential means by which progress in deep learning is achieved today. Corporate players have no longer open sourced recent text to speech model architectures, often not even trained models. Instead, they tend to publish a scientific paper, sometimes with details in additional material, and an accompanying demo with pre-generated audio snippets. 

The proposed improvement involves implementing a recent TTS paper such as Voicebox, open-sourcing the architecture. In addition, as far as possible, efforts will be made to collect training data, train the model and demonstrate that the open-sourced architecture performs well, for example by illustrating notable features or approximately reproducing some performance results (e.g. CMOS). 

Imminent Research Grants

$100,000 to fund language technology innovators

Imminent was founded to help innovators who share the goal of making it easier for everyone living in our multilingual world to understand and be understood by all others. Each year, Imminent allocate $100,000 to fund five original research projects to explore the most advanced frontiers in the world of language services. Topics: Language economics – Language data – Machine learning algorithms for translation – Human-computer interaction – The neuroscience of language.

Apply now

2022 – Projects awarded

T-INDEX

Economic Complexity research group Centro Ricerca Enrico Fermi

Understanding which countries and languages dominate online sales is a key question for any company wishing to translate its website. The goal of this research project is to complement the T-Index by developing new tools capable of identifying emerging markets and opportunities, thus predicting which languages will become more relevant in the future for a specific product in a specific country. As a first step, the Economic Fitness and Complexity algorithm will be used to identify countries that are expected to undergo significant economic expansion in the coming years. Subsequently, network science and machine learning techniques are used to predict the products and services that growing economies are likely to start importing.

THE NEUROSCIENCE OF TRANSLATION. NOVEL AND DEAD METAPHOR PROCESSING IN NATIVE AND SECOND-LANGUAGE SPEAKERS

Martina Ardizzi and Valentina Cuccio

The NET project aims to investigate the embodied nature of a second language, focusing on a specific linguistic element that merges abstract and concrete conceptual domains: metaphors. The idea behind the project fits within the embodied simulation approach to language, which has been poorly applied in the field of translation despite being widely confirmed in the study of native languages. Specifically, during the project the brain activities of native Italian speakers and second-language Italian speakers will be recorded while they read dead or novel Italian metaphors. It will be expected to show a different involvement of the sensorimotor cortices of the two groups in response to the different types of metaphors. The results of NET may provide new insights on how to improve disembodied AI translations.

COLLECTION OF SPEECH DATA (50 HOURS) IN A CROWDSOURCED VERSION FOR THE YORÙBÁ LANGUAGE

Kọ́lá Túbọ̀sún

Yorùbá is one of the most widely spoken languages in Africa with 46 million first and second language speakers. Yet there is hardly any language technology available in Yorùbá  to help them, especially illiterate or visually impaired people who would benefit most. The present project proposal aims at developing speech technology in Yorùbá in order to make everyone be understood.

As a first action, aligned voice and text resources will be recorded professionally in a quality usable to produce text-to-speech systems. After donating this data under a Creative Commons license to the Mozilla Common Voice repository, further speech data will be collected from volunteers online. To increase the quality of the text, a diacritic restoration engine has already been developed.

INCREMENTAL PARALLEL INFERENCE FOR MACHINE TRANSLATION

Andrea Santilli La Sapienza University

Machine translation works with a de facto standard neural network called Transformer, published in 2017 by a team at Google Brain. The traditional way of producing new sentences from the Transformer is one word at a time, left to right; this is hard to speed up and parallelize. 

A similar problem has been spotted and was  solved in image generation by using “incremental parallel processing”, a technique which refines an image progressively rather than generating it pixel by pixel, yielding speedups of 2-24×.
There is a proposal to port this method to Transformers, using clever linear algebra tricks to make it happen.  This technique and other similar ones could make machine translation less expensive, and therefore accessible to a larger number of use cases, and ultimately people.

INVESTIGATING THE POTENTIAL OF SPEECH TECHNOLOGIES – SYNTHESIS AND RECOGNITION – TO IMPROVE THE QUALITY OF PROFESSIONAL AND TRAINEE TRANSLATORS’ WORK.

Dragoș Ciobanu University of Wien

Translators carry out a cognitively demanding, repetitive task which requires continuous high concentration. When they post-edit neural draft translations, a known source of errors is called the “NMT fluency trap”, where the target sentence sounds very fluent and error-free, but this might hide infidelities or alterations with respect to the source.

Some promising experimental results show that this situation can be helped by reading the source side out loud, using speech synthesis. 

The practicality and cognitive impact of this new modality will be evaluated to ensure that it does not slow down the overall translation process. To do this the translator’s gaze while they work will be tracked, their focus and cognitive load. This idea could make the translator’s work easier and reduce errors.

Imminent Research Grants

$100,000 to fund language technology innovators

Imminent was founded to help innovators who share the goal of making it easier for everyone living in our multilingual world to understand and be understood by all others. Each year, Imminent allocate $100,000 to fund five original research projects to explore the most advanced frontiers in the world of language services. Topics: Language economics – Language data – Machine learning algorithms for translation – Human-computer interaction – The neuroscience of language.

Apply now

Photo credit: Google Deepmind – Unsplash