Nlp

We explore the use of large language models (LLMs) for ontology population tasks in finance. The findings reveal that LLMs-based generative approaches, especially when provided with positive and negative examples in the prompts, outperform span-based methods, achieving a 61.05% F1 score.

We introduce a methodology for event extraction, particularly aimed at low-resource languages like Thai. The proposed framework simplifies the traditional annotation process, enhancing efficiency in extracting events from text.

This paper presents a cross-document coreference resolution approach tailored for low-resource languages, specifically focusing on Thai. The authors adapt an existing English model that uses agglomerative clustering to identify and group coreferent entities across documents. The study also compares manual and automatic span detection methods, finding that a fine-tuned longformer model provides the best performance, achieving a CoNLL F1 score of 72.87. This research offers a framework that could be extended to other low-resource languages.

The paper presents a progressive approach to building Thai spelling correction systems by integrating both neural-based and dictionary-based methods. It introduces an Extendable Neural Contextual Corrector (XNCC), which allows for vocabulary expansion post-training, demonstrating that effective spelling correction can be achieved with minimal annotated data and minor enhancements to traditional methods.

We adopted multi-hop KGQA for Thai language data, using a manually-constructed knowledge graph based on banking products. The technique scored well in tests, but potentially proned to over-fitting when the answers are within 1-hop from the starting node.