Retrieving Top K% Relevant Patterns for Relation Extraction in Bangla Using Distant Supervision

Information extraction constitutes a piv-otal domain within natural language processing (NLP) aiming to distill structured insights-such as entity types and their relationships-from diverse textual sources. Relation extraction, a core task within this field, automates identifying and classifying relation-ships between entities mentioned in unstructured text. This paper explores distant supervision (DS) methods tailored for Bangla, addressing the linguistic nuances and challenges of data scarcity. DS leverages knowl-edge bases for automatic corpus labeling, although this can introduce noise. Our location mnemonics help reduce some noisy patterns related to location-based relations. Additionally, our proposed method retrieves the top K% relevant patterns using conflict scores which is further refined with probability scores. Setting K to 80% results in an impressive F1 score of 91 %, demonstrating the efficacy of our approach.