AI BriefingLG AI Research 374
Key point
LG AI Research's EXAONE model ranked first on KorQuAD 2.0, a benchmark for evaluating Korean machine reading comprehension performance.
Details
LG AI Research's EXAONE Lab is developing large language models (LLMs) for collaboration with LG Group affiliates and partner companies, and has recently achieved first place on the KorQuAD 2.0 benchmark. This follows its first-place result on KorQuAD 1.0 last June.
KorQuAD 2.0 is a Korean machine reading comprehension (MRC) dataset consisting of over 100,000 question-answer pairs. Unlike the previous 1.0 version, it takes Wikipedia pages containing HTML code as input and is a highly challenging task requiring answers in various forms, from short answers to long sentences, tables, and lists.
The existing Span Prediction approach predicts the start and end of an answer, making it fast, but it has the limitation of struggling to handle cases where the answer is not present in the text. In contrast, generative models with an Encoder-Decoder architecture, such as Bart or T5, have the advantage of being able to generate content not found in the text by leveraging the model's parametric knowledge.
The EXAONE model has demonstrated the technical capability to effectively utilize the characteristics of HTML documents and accurately generate answers of various lengths and forms.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.