Use this link to cite:
https://hdl.handle.net/2183/49406 Enfoques de traducción automática con modelos de lenguaje en obras wuxia
Loading...
Identifiers
Publication date
Authors
Silvosa Cuervo, Hugo
Other responsabilities
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] Traducir literatura web china de los géneros wuxia y xianxia al inglés puede resultar una tarea compleja debido a su fuerte carga idiomática y cultural. En este trabajo contrastamos la eficacia de los sistemas de traducción automática neuronal (NMT, por sus siglas en inglés) frente a los grandes modelos de lenguaje (LLM, por sus siglas en inglés) mediante el uso de un corpus paralelo ad hoc específico al dominio. Para adaptar los modelos a este dominio, entrenamos los modelos NMT mediante ajuste fino convencional. Por su parte, en los LLMs se experimentó con diferentes técnicas, como la ingeniería de instrucciones y el aprendizaje en contexto, hasta culminar en su ajuste con QLoRA. Tras la evaluación, observamos que, si bien la NMT lidera en métricas de solapamiento léxico, los LLM logran una equivalencia semántica comparable. Este análisis se complementa con una comparativa con sistemas comerciales y una evaluación cualitativa en la que, brevemente, se revisan las principales problématicas encontradas en las traducciones. De esta manera, los resultados demuestran que, frente a la rigidez y las limitaciones contextuales de la NMT, los LLM ajustados al dominio logran priorizar el ritmo narrativo y el contexto cultural, actuando de una manera mucho más cercana a la de un traductor humano profesional en textos de dominios muy específicos.
{Abstract] Translating Chinese web literature of the wuxia and xianxia genres into English can be a complex task due to its heavy idiomatic and cultural load. In this work, we contrast the effectiveness of neural machine translation (NMT) systems against large language models (LLMs) using an ad hoc, domain-specific parallel corpus. To adapt the models to this domain, we trained the NMT models through conventional fine-tuning. Meanwhile, for the LLMs, we experimented with various techniques, such as prompt engineering and in-context learning, culminating in their fine-tuning using QLoRA. Upon evaluation, we observed that while NMT leads in lexical overlap metrics, LLMs achieve comparable semantic equivalence. This análisis is complemented by a comparison with commercial systems and a qualitative evaluation that briefly reviews the main issues encountered in the translations. Thus, the results demonstrate that, compared to the rigidity and contextual limitations of NMT, domain-adapted LLMs manage to prioritize narrative pacing and cultural context, performing in a manner much closer to that of a professional human translator when dealing with highly complex texts.
{Abstract] Translating Chinese web literature of the wuxia and xianxia genres into English can be a complex task due to its heavy idiomatic and cultural load. In this work, we contrast the effectiveness of neural machine translation (NMT) systems against large language models (LLMs) using an ad hoc, domain-specific parallel corpus. To adapt the models to this domain, we trained the NMT models through conventional fine-tuning. Meanwhile, for the LLMs, we experimented with various techniques, such as prompt engineering and in-context learning, culminating in their fine-tuning using QLoRA. Upon evaluation, we observed that while NMT leads in lexical overlap metrics, LLMs achieve comparable semantic equivalence. This análisis is complemented by a comparison with commercial systems and a qualitative evaluation that briefly reviews the main issues encountered in the translations. Thus, the results demonstrate that, compared to the rigidity and contextual limitations of NMT, domain-adapted LLMs manage to prioritize narrative pacing and cultural context, performing in a manner much closer to that of a professional human translator when dealing with highly complex texts.
Description
Keywords
Procesamiento del lenguaje natural Traducción automática neurona Grandes modelos de lenguaje Literatura fantástica china (wuxia/xianxia) Adaptación del dominio Aprendizaje en contexto Natural Language Processing Neural Machine Translation Large language models Chinese fantasy literature (wuxia/xianxia) Domain adaptation In-Context Learning
Editor version
Rights
Attribution 4.0 International








