Triplétoile: Extraction of knowledge from microblogging text

Area of Science:

• Natural Language Processing
• Data Science
• Artificial Intelligence

Background:

• Existing knowledge graph extraction methods struggle with open-domain text sources like social media.
• Micro-blogging posts contain unique entities and relations not easily modeled by current pipelines.

Purpose of the Study:

• To develop an enhanced information extraction pipeline for knowledge graph construction from micro-blogging data.
• To address the challenge of modeling open-domain entities and relations in social media text.

Main Methods:

• Leveraged dependency parsing for enhanced information extraction.
• Employed unsupervised hierarchical clustering over word embeddings for relation classification.
• Applied the pipeline to a corpus of 100,000 tweets on digital transformation.

Main Results:

• Achieved over 95% precision in extracting semantic triples.
• Outperformed similar pipelines by approximately 5% in precision.
• Generated a higher number of triples compared to existing methods.

Conclusions:

• The proposed pipeline effectively extracts knowledge graphs from social media.
• The system demonstrates superior performance and efficiency for open-domain information extraction.
• The generated knowledge graph and methodology are publicly released for further research.