AI Briefing
KO

EMNLP 2023

Key point

Upstage proved its Korean NLP capabilities by getting 2 papers accepted at EMNLP 2023.

Details

Upstage reaffirmed its global technical prowess by presenting 2 papers at EMNLP 2023, the most prestigious conference in the natural language processing field. EMNLP 2023 will be held in Singapore from December 6-10, and last year's EMNLP 2022 recorded 3,242 submissions and 715 acceptances (22% acceptance rate).

This achievement is Korean NLP research resulting from a collaboration between Upstage's Tech Lead Chanjun Park and Professor Heuiseok Lim's research team at Korea University. The company explained that this acceptance demonstrates its competitiveness on the international academic stage, competing against Google, Apple, Amazon, Baidu, and others.

  • KEBAP: Korean Error Explainable Benchmark Dataset for ASR and Post-processing

    • This paper builds a new benchmark dataset for Korean ASR (speech recognition) post-processing.
    • It points out that existing evaluations fail to sufficiently explain errors at the speech and text level, and proposes an evaluation methodology that enhances explainability.
    • The dataset breaks errors down into 37 speech-level types and 13 text-level error types reflecting background noise and speaker characteristics, and applies the methodology to commercial systems such as Google Cloud Speech Recognition and CLOVA for analysis.
  • CHEF in the Language Kitchen: A Generative Data Augmentation Leveraging Korean Morpheme Ingredients

    • This paper proposes CHEF, a new data augmentation technique that reflects the morpheme characteristics of Korean.
    • It notes the limitation that, because sentence meaning in Korean changes depending on morpheme combinations, random augmentation can distort meaning or produce unnatural sentences.
    • It is designed so that a generative language model varies morpheme combinations in diverse ways to produce more natural Korean sentences.

Upstage also presented 7 papers at ICML 2023-DMLR last June, the highest achievement among Korean companies, and stated that it has presented 100 papers in AI domestically and internationally in the 3 years since its founding. It also emphasized that, among the top 7 conferences in the NLP field according to Google Scholar — ACL, EMNLP, NAACL, TACL, COLING, LREC, and WMT — it has achieved paper acceptances at all of them except the journal TACL.

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.