Dr. Vivek Kumar
Senior Researcher
Actions
Dr. Vivek Kumar is a Senior Researcher at the chair of Open Source Intelligence at the University of Federal Armed Forces under Federal Ministry of Defence, Germany. He coordinates STELAR, an EU's Horizon 2020 program targeting "Artificial Intelligence, Data, and Robotics" in agri-food. Dr. Kumar won the Marie Sklodowska-Curie fellowship in 2019 for the EU's Horizon 2020 ITN network program PhilHumans. He worked as a Marie Curie Researcher with Philips Research, Netherlands, and the University of Cagliari, Italy, and received the doctorate in 2023. His current research focuses on NLP, Knowledge Graphs, LLMs, AI Fairness & Bias, Mental Health, and Domain Adaptation.
Links
Area of Expertise
Topics
Unlocking LLMs: From Forgiving Errors to Ethical Considerations in Domain Adaptation
Synthetic data generation has helped mitigate the Bias in machine learning (ML) pipelines and scarce data challenges. Recently, large language models (LLMs) have been readily used for this purpose, but “Inherent Bias,” “Data Privacy & Confidentiality,” “Hallucinations,” & “Stochastic Parrot” nature limit the LLM’s reliability for direct & unsupervised use. Therefore, this talk undertakes a real-world use case to address the hazards in food products, shows intelligent prompting to generate the first-ever expert-annotated hazard dataset & presents the dangers of forgiving errors and cherry-picking in LLMs when dealing with sensitive domains.
Knowledge Augmenting Practices for Domain Adaptation
Although Language Models (LM) have successfully imparted human-like abilities, available training data size and the complex domain-specific context are two significant constraints that jeopardize LM's optimal and reliable performance. A potential solution to these limitations is to equip the LMs with domain knowledge. While the available research works use Knowledge Graphs Embeddings (KGEs) to inject domain knowledge, this work provides a Knowledge-based LLM to use the Resource Description Framework (RDF) triples directly at the input level. The proposed LLM works at the crossroads of Generative Pretrained Transformer (GPT-2) and Bidirectional Encoder Representations from Transformers (BERT). It uses a novel pipeline to select, categorize, and filter the RDF triples and introduces heuristic methods to inject domain-specific Knowledge into Knowledge LM. It is one of the few works that directly use KGs to overcome domain adaptation and scarce challenges at once at a fine-grained level. The proposed LM is field tested in domains (healthcare, scholarly), and the results show that our proposed Knowledge LM has significantly outperformed the existing LM (BERT for each KG). The findings of this work also conclude the importance of the relevance of KG and quantify the knowledge injection of RDF triples, demonstrating that Knowledge LM is a potential choice for domain adaptation to solve knowledge-driven problems for academia and industry, both.
Data Makers Fest 2024 Sessionize Event Upcoming
Data Makers Fest Sessionize Event
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top