The research team of Professor Koo Myoung-wan, Department of Computer Science & Engineering,

Paper Accepted at 'EACL 2024', the Most Prestigious Conference in Natural Language Processing

▲ (From left) Professor Koo Myoung-wan, Department of Computer Science & Engineering; Kim Min-ju, master's student, Graduate Department of Artificial Intelligence; Yeon Hee-yeon, MSc

The paper 'Towards Context-Based Violence Detection: A Korean Crime Dialogue Dataset', presented by the Intelligent Spoken Dialogue Systems (ISDS) research team supervised by Professor Koo Myoung-wan of the Department of Computer Science & Engineering, has been accepted as an EACL Findings Paper at EACL 2024 (The 18th Conference of the European Chapter of the Association for Computational Linguistics).

Master's student Kim Min-ju of the Graduate Department of Artificial Intelligence (joint first author), Yeon Hee-yeon, MSc (joint first author), and Professor Koo Myoung-wan (corresponding author) will present a dataset, protocol and baseline for detecting violence and hate speech with contextual awareness in offline settings at EACL 2024 (listed at BK IF 2 in natural language processing).

▲ The harmful dialogue classification model architecture proposed in the paper

Datasets and methodologies for automatically detecting hate speech and biased expression have drawn attention in natural language processing for some time. Many datasets for identifying harmful data have accordingly been released, but they have been limited in that they largely assume an online environment and do not take context into account. Building on this research landscape, the team identified the absence of datasets for detecting violence and hate situations with dialogue context in offline settings, and released a dataset for classifying violent dialogue based on international standard crime classification criteria, developed in collaboration with legal experts.

The team also released an effective protocol for building such a dataset. They proposed the 'Legal Expert Collaborative Data Building Process', a data generation protocol that maintains data quality in settings where crowd workers must write dialogue data themselves rather than simply annotating it. They further released 'Relationship Aware BERT' as a baseline for the released dataset. This model improves performance by modelling the distinctive characteristics of dialogue text: the team added a 'speaker token' to the dialogue text and trained the model to classify the whole dialogue while also classifying whether each speaker is a perpetrator, a victim or an ordinary person.

The master's students who took part in this research are DHE (Digital Human Entertainment) scholarship holders sponsored by Smilegate, from the first and second cohorts respectively. They said: "We were able to achieve this result thanks to the research support of the department and our professor and the help of our seniors and juniors," adding: "We will keep working to develop technology that can help the world."

Professor Koo said: "I will continue to take a steady interest and provide guidance so that students can achieve research results."

EACL is the European chapter, established in 1982, of ACL, the most prestigious conference in natural language processing. EACL 2024, the eighteenth such conference, will be held in St Julian's, Malta, from 18 to 20 March.

Source: News – Research Achievements https://sogang.ac.kr/ko/story/media-center?tab=3