Skip to main content

CURACION Y EXPLOTACION DE RECURSOS HETEROGENEOS PARA LA TRADUCCION DE LENGUAS CON POCOS RECURSOS

PID2024-158157OB-C31

•
Funding agency name Agencia Estatal de Investigación
Funding agency acronym AEI
Program Programa Estatal para la Investigación y el Desarrollo Experimental
Subprogram Subprograma Estatal de Generación de Conocimiento Científico-Técnico y Desarrollo Experimental
Call Proyectos de I+D+i (Generación de Conocimiento y Retos Investigación)
Call year 2024
State Plan 2024-2027 Plan Estatal de Investigación Científica y Técnica y de Innovación
Beneficiary centre UNIVERSIDAD DE ALICANTE

Resultados relacionados

Resultados totales (Incluyendo duplicados): 5
Encontrada(s) 1 página(s)

RUA. Repositorio Institucional de la Universidad de Alicante
Publication Otros (Other)

Data-centric strategies for machine translation: efficient acquisition, cleaning, and exploitation of parallel resources

  • García Romero, Cristian
  • García Romero, Cristian|||/items/4fad4894-8a55-40f8-852a-d21c46a9f92a
High-quality parallel corpora are an essential resource for crosslingual natural language processing. Particularly in machine translation (MT), dedicated systems heavily rely on massive amounts of this data for training, while large language models (LLMs) require less parallel data to be adapted for translation through fine-tuning and in-context learning, where quality is prioritized over quantity. However, acquiring high-quality parallel corpora from the web presents significant efficiency and quality challenges. This thesis proposes a comprehensive framework for the acquisition, cleaning, and exploitation of parallel data. First, we tackle the inefficiency of traditional web crawling with a novel smart bilingual focused crawler. By utilizing URL-based models to infer language and document parallelness, our approach reduces wasted bandwidth and accelerates the discovery of parallel content. Second, to ensure data quality, we address the growing proliferation of synthetic, machine-translated text on the web, which can cause model degradation. We introduce an automatic machine translation detection method that leverages the internal representations of a surrogate multilingual MT model to accurately filter out synthetic data, outperforming state-of-the-art techniques, with particularly strong results for non-English language pairs. Finally, we explore the optimal exploitation of high-quality parallel data for MT using LLMs. We propose an in-context learning approach driven by reinforcement learning to dynamically select the most effective parallel examples, maximizing the final translation quality.
RUA. Repositorio Institucional de la Universidad de Alicante
Publication Otros (Other)

Automatic Machine Translation Detection Using a Surrogate Multilingual Translation Model

  • García Romero, Cristian
  • Esplà-Gomis, Miquel
  • Sánchez-Martínez, Felipe
  • García Romero, Cristian|||/items/4fad4894-8a55-40f8-852a-d21c46a9f92a
  • Esplà-Gomis, Miquel|||/items/6614b776-cf96-43c9-b300-de4c8cf17c43
  • Sánchez-Martínez, Felipe|||/items/117378d0-8f7d-40ef-9b4a-5646b1af1999
Modern machine translation (MT) systems depend on large parallel corpora, often collected from the Internet. However, recent evidence indicates that (i) a substantial portion of these texts are machine-generated translations, and (ii) an overreliance on such synthetic content in training data can significantly degrade translation quality. As a result, filtering out non-human translations is becoming an essential pre-processing step in building high-quality MT systems. In this work, we propose a novel approach that directly exploits the internal representations of a surrogate multilingual MT model to distinguish between human and machine-translated sentences. Experimental results show that our method outperforms current state-of-the-art techniques, particularly for non-English language pairs, achieving gains of at least 5 percentage points of accuracy.
RUA. Repositorio Institucional de la Universidad de Alicante
Publication Otros (Other)

Multi-Hypothesis Distillation of Multilingual Neural Translation Models for Low-Resource Languages

  • Galiano Jiménez, Aarón|||/items/ff07f91f-bf9b-4ef1-b2f1-8457faff7914
  • Pérez-Ortiz, Juan Antonio|||/items/a13b860a-9e21-4163-a290-337d45dce50e
  • Sánchez-Martínez, Felipe|||/items/117378d0-8f7d-40ef-9b4a-5646b1af1999
  • Sánchez-Cartagena, Víctor M.|||/items/782c1de3-587b-40f4-87f0-d214fe789104
  • Sánchez-Cartagena, Víctor M.
This paper explores knowledge distillation (KD) of multilingual pre-trained encoder-decoder and large language models for machine translation (MT). We argue that the teacher model's output distribution holds valuable insights for the student, beyond the approximated mode obtained through beam search (the standard decoding method for MT), and present Multi-Hypothesis Distillation (MHD), a sequence-level KD method that generates multiple translations for each source sentence. This provides a larger representation of the teacher model distribution and exposes the student model to a wider range of target-side prefixes. We explore n-best lists from beam search to guide the student's learning and examine alternative decoding methods to address issues like low variability and the under-representation of infrequent tokens. For low-resource languages, our research shows that while sampling methods may slightly compromise translation quality compared to beam search based approaches, they enhance the generated corpora with greater variability and lexical richness. This ultimately improves student model performance and mitigates the gender bias amplification often associated with KD. Our experiments show that MHD improves over existing KD techniques both when training encoder-decoder students from scratch and when distilling into pre-trained decoder-only models.
RUA. Repositorio Institucional de la Universidad de Alicante
Publication Otros (Other)

AI-TraLow: AI-Driven Translation for Low-Resource Languages and Cultures

  • Oliver, Antoni
  • Melero, Maite
  • Sánchez-Martínez, Felipe
  • Sánchez-Cartagena, Víctor M.
  • Sánchez-Martínez, Felipe|||/items/117378d0-8f7d-40ef-9b4a-5646b1af1999
  • Sánchez-Cartagena, Víctor M.|||/items/782c1de3-587b-40f4-87f0-d214fe789104
In this paper, we present AI-TraLow, a project dedicated to advancing AI-driven translation for low-resource languages and cultures. The research is structured around three primary objectives: firstly, the development of advanced data curation techniques designed to refine parallel corpora and detect machine-generated content; secondly, the exploration of integrating linguistic resources—such as dictionaries and grammatical rules—directly into model prompts and fine-tuning techniques to enhance translation precision; and thirdly, the mitigation of hardware constraints through knowledge distillation to produce efficient models viable for standard desktop environments. By targeting specific linguistic groups, including Iberian varieties (Aranese, Aragonese, Asturian and Eonavian), Mayan languages, and languages of vulnerable migrant communities, AI-TraLow seeks to foster linguistic diversity and digital inclusion. Ultimately, this initiative delivers open-source tools and models that ensure cultural heritage is both preserved and accessible within the contemporary digital landscape.
RUA. Repositorio Institucional de la Universidad de Alicante
Publication Otros (Other)

When Translations Surprise: Human Awareness of Predictability in Translations

  • García Romero, Cristian
  • Esplà-Gomis, Miquel
  • Sánchez-Martínez, Felipe
  • García Romero, Cristian|||/items/4fad4894-8a55-40f8-852a-d21c46a9f92a
  • Esplà-Gomis, Miquel|||/items/6614b776-cf96-43c9-b300-de4c8cf17c43
  • Sánchez-Martínez, Felipe|||/items/117378d0-8f7d-40ef-9b4a-5646b1af1999
Machine translation (MT) has achieved near-human quality for some language pairs, yet its output remains distinct from human translation, primarily in its predictability. While MT systems generate low-perplexity text, humans produce less predictable outputs. This raises the question of whether humans can intuitively use this difference in predictability to distinguish between human- and machine-translated text. We report on a study with 30 native Spanish speakers tasked with identifying the origin of English-to-Spanish translations. We compared their performance against two perplexity-based baselines: a large language model capturing fluency, and a neural MT model, conditioned on the source text, capturing both fluency and adequacy. Our findings reveal that human judgments correlate with fluency-based perplexity, but show no correlation with the perplexity that also accounts for adequacy. This suggests that annotators’ decisions are driven by the target text’s fluency. Consequently, a simple computational baseline using source-aware perplexity significantly outperforms human annotators. This work contributes to a deeper understanding of human perception of MT, highlighting a potential bias in current evaluation protocols toward fluency over adequacy. This bias may lead to an overestimation of the capabilities of highly fluent systems and underscores the need for evaluation methods ensuring translation adequacy is not overlooked.