2 April 2026 in IEEE Access
The challenge of knowledge graph construction from unstructured text is a fundamental challenge in knowledge representation, especially in cases of informal text and narrative, such as that found in movie dialogue. Existing knowledge graph construction and information extraction approaches often struggle with linguistic variations, limited understanding of relevant contextual information, and have required considerable manual effort. This study introduces a fully automated pipeline for domain-adapted knowledge graph construction from unstructured text using large language models (LLMs), specifically fine-tuned DeBERTa-v3-base, to implement both Named Entity Recognition (NER) and Relation Extraction (RE). We built a novel, custom-labeled dataset of English movie subtitles and additional types of entities in order to extract richer semantic features from the movie dialogue including Action, Emotion, Role, and Spirit. Annotating the dataset involved generative-based prompts for labeling the entities using both GPT-3.5 and GPT-4o and a hybrid form of labeling movie subtitles using a singular text by human annotation. This method produced high-quality BIO-tagged data to model document generation models. We utilized our dataset to build fine-tuned NER and RE models. Our results indicated considerable improvement over pre-trained models and baseline models showcasing the importance of contextual model fine-tuning for domain-specific works. Our approach reduces reliance on manual knowledge engineering by automatically extracting semantically dense, and emotionally expressive and context-based knowledge triples. The resulting knowledge graph represented as ⟨head entity, relation, tail entity⟩ triplets demonstrate the feasibility of a large-scale knowledge graph from informal text using modern deep learning architectures. This generated knowledge graph serves as a structured representation that supports downstream tasks such as semantic exploration, narrative analysis, and visualization. This research provides a scalable and accurate method for knowledge extraction in narrative contexts with implications for intelligent narrative dialogue systems, emotion-aware artificial intelligence, and dynamic knowledge graphs.