Action Power Co., Ltd. consecutively publishes papers in three world-renowned academic societies... Recognized for global AI technology excellence

by Jung Jin

Published 22 May.2023 10:32(KST)

[Example of Image Data Generation Results for Training VQA Models]

Multimodal AI specialist company ActionPower Co., Ltd. (co-CEOs Hongshik Cho and Jihwa Lee) announced on the 22nd that its research team’s papers were consecutively published in the world’s most prestigious conferences: INTERSPEECH (International Speech Communication Association), ICASSP (International Conference on Acoustics, Speech, and Signal Processing), and ACL (Association for Computational Linguistics). Through this achievement, ActionPower has proven that despite being a startup, it possesses world-class artificial intelligence technology capabilities.

INTERSPEECH and ICASSP are international conferences in the fields of speech, acoustics, and signal processing, attracting thousands of leading AI experts worldwide every year to share the latest research results, boasting the largest scale globally. ACL, held for the 61st year this year, is also a world-renowned academic conference where computer science experts, including those in natural language processing (NLP), share the latest research and technologies.

ActionPower has pursued both research on and service development of Korea’s top-level NLP and ASR foundational technologies centered around its AI knowledge management app, Dagglo. Recently expanding into the vision domain, the company has published seven papers over the past three years in top international conferences including INTERSPEECH. Based on these research outcomes, it holds 21 domestic patents and 2 overseas patents. Additionally, there are 18 domestic and 11 overseas patents pending, actively strengthening its core competitive capabilities as a technology company while conducting vigorous research and development projects.

Seongmin Park, head of the NLP research team, presented two papers related to natural language processing at INTERSPEECH 2023 and ACL 2023 this year.

At INTERSPEECH 2023, a research result was presented that can perform the process of judging and distinguishing sentence similarity ten times faster than before. By applying hyperdimensional vectors (vectors with more than 10,000 dimensions) to the existing ‘topic segmentation’ method?which analyzes long texts sentence by sentence to separate paragraphs?each sentence is given unique characteristics, adding a clear distinction step. This enables language models to more efficiently improve the readability and accuracy of generated texts.

At ACL 2023, a technology for automatically labeling topics in text data was introduced. For machine learning use, labeling is required to indicate the correct topic of a text, but labeling countless texts individually is impractical. To solve this, the developed ‘Pseudo-labeling’ technique adds a sentence such as “This text is about 00” at the end of a paragraph or the entire text, naturally connecting the actual content with the added sentence’s explanation, thereby labeling the text as “about 00.” This technology can significantly improve the speed of machine learning.

Kyungho Kim, an NLP researcher, presented research at ICASSP 2023 that innovates the training process of VQA (Visual Question and Answering) models using generative AI. Training VQA models requires images, questions about the images, and answers to those questions, but securing image data has been a challenge. This research devised a method to generate multiple images using questions and answers as prompts, allowing the training input to include not only original but also numerous generated images. This can assist many researchers who have struggled to obtain image data. Actively utilizing this technology is expected to greatly accelerate the advancement of AI related to image data.

Jihwa Lee, co-CEO and CTO of ActionPower, said, “This research has elevated the level of natural language processing and image processing and greatly improved the efficiency of the training process, marking a new step forward in ActionPower’s growth. These research results will lead to innovative services in various fields provided not only through our Dagglo service but also in B2C and B2B sectors.”

He added, “We will continue to strive to provide services that enable everyday innovation based on our world-recognized proprietary technologies. In particular, we plan to invest generously to continuously recruit top-tier talent and further raise the level of our research.”