Análisis de Sentimiento para lenguajes de bajos recursos, Dominio: Shipibo-Konibo
Fecha
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Pontificia Universidad Católica del Perú
Acceso al texto completo solo para la Comunidad PUCP
Resumen
Con el objetivo de apoyar a comunidades con bajos recursos digitales en su
integración a la sociedad, se desarrolló un modelo de análisis de sentimiento para
lenguas indígenas, permitiendo la implementación de tecnologías como chatbots y
asistentes virtuales que puedan operar en su lengua materna. Esta propuesta busca
no solo facilitar un mayor acceso a servicios esenciales en áreas como educación,
salud y gobierno, sino también promover la preservación cultural y lingüística de
comunidades históricamente marginadas. La incorporación de herramientas de este
tipo representa una estrategia para reducir la brecha digital y garantizar un acceso
más equitativo a los beneficios de la transformación tecnológica.
Para el idioma Shipibo-Konibo, se utilizaron diversas técnicas de aumento de datos
basadas en errores controlados, incluyendo alteraciones aleatorias, proximidad de
teclado, ambigüedad fonema-grafema y similitud silábica. Estas técnicas
contribuyeron significativamente a incrementar la diversidad y representatividad del
corpus, permitiendo que el modelo entrenado reflejara de manera más realista la
variabilidad natural del lenguaje. Asimismo, se evaluaron modelos de embeddings
multilingües como XLM-Roberta, LaBSE y SIMCSE, seleccionando finalmente el
más adecuado por su capacidad de generalización y desempeño en escenarios
multilingües.
Los experimentos realizados lograron superar el desafío de clasificar oraciones en
categorías positivas, negativas y neutras, incluso en contextos de datos limitados.
Este avance constituye un paso importante hacia la inclusión tecnológica de
comunidades indígenas, ofreciendo herramientas adaptadas a sus necesidades
lingüísticas y fomentando un ecosistema digital más diverso e inclusivo.
With the objective of supporting communities with low digital resources in their integration into society, a sentiment analysis model for indigenous languages was developed, allowing the implementation of technologies such as chatbots and virtual assistants that can operate in their mother tongue. This proposal seeks not only to facilitate greater access to essential services in areas such as education, health, and government, but also to promote the cultural and linguistic preservation of historically marginalized communities. The incorporation of such tools represents a strategy to reduce the digital divide and to guarantee more equitable access to the benefits of technological transformation. For the Shipibo-Konibo language, various data augmentation techniques based on controlled errors were applied, including random alterations, keyboard proximity, phoneme-grapheme ambiguity, and syllabic similarity. These techniques significantly contributed to increasing the diversity and representativeness of the corpus, allowing the trained model to more realistically reflect the natural variability of the language. Likewise, multilingual embedding models such as XLM-Roberta, LaBSE, and SIMCSE were evaluated, ultimately selecting the most suitable one for its generalization capacity and performance in multilingual scenarios. The experiments carried out managed to overcome the challenge of classifying sentences into positive, negative, and neutral categories, even in low-data contexts. This advancement constitutes an important step toward the technological inclusion of indigenous communities, offering tools adapted to their linguistic needs and fostering a more diverse and inclusive digital ecosystem.
With the objective of supporting communities with low digital resources in their integration into society, a sentiment analysis model for indigenous languages was developed, allowing the implementation of technologies such as chatbots and virtual assistants that can operate in their mother tongue. This proposal seeks not only to facilitate greater access to essential services in areas such as education, health, and government, but also to promote the cultural and linguistic preservation of historically marginalized communities. The incorporation of such tools represents a strategy to reduce the digital divide and to guarantee more equitable access to the benefits of technological transformation. For the Shipibo-Konibo language, various data augmentation techniques based on controlled errors were applied, including random alterations, keyboard proximity, phoneme-grapheme ambiguity, and syllabic similarity. These techniques significantly contributed to increasing the diversity and representativeness of the corpus, allowing the trained model to more realistically reflect the natural variability of the language. Likewise, multilingual embedding models such as XLM-Roberta, LaBSE, and SIMCSE were evaluated, ultimately selecting the most suitable one for its generalization capacity and performance in multilingual scenarios. The experiments carried out managed to overcome the challenge of classifying sentences into positive, negative, and neutral categories, even in low-data contexts. This advancement constitutes an important step toward the technological inclusion of indigenous communities, offering tools adapted to their linguistic needs and fostering a more diverse and inclusive digital ecosystem.
Descripción
Palabras clave
Procesamiento en lenguaje natural (Computación), Minería de datos, Inteligencia artificial, Brecha digital, Lenguas indígenas--Perú--(Shipibo-Conibo)
Citación
item.page.endorsement
item.page.review
item.page.supplemented
item.page.referenced
Licencia Creative Commons
Excepto donde se indique lo contrario, la licencia de este ítem se describe como info:eu-repo/semantics/openAccess
