Graph-Based Methods for Hot Event Discovery, Long Text Matching, and Ontology Construction in Natural Language Processing

The presentation begins with an overview of natural language processing (NLP) history, from bag‑of‑words to large pre‑trained language models, and argues that graph‑structured representations combined with graph neural networks (GNNs) are the next promising direction for NLP.

It introduces the Story Forest system, which organizes news articles into hierarchical story trees using a two‑stage graph‑based clustering pipeline (keyword community detection followed by document graph community detection) to reduce redundancy and reveal event evolution.

The talk then describes the GIANT framework for ontology creation and user interest modeling. By constructing a search‑click graph from query‑document interactions, the system extracts concepts, events, and topics, inserts them into a knowledge graph, and tags user interests for personalized recommendation.

For long‑text matching, a divide‑and‑conquer strategy is proposed: texts are split into sub‑topics, aligned locally, and then aggregated globally via a GCN that performs message passing on a concept interaction graph, overcoming the limitations of traditional encoder‑ or interaction‑based models.

Experimental results on two long‑text datasets show that the graph‑based models (SimNet, CIG‑Siam, CIG‑Siam‑GCN) outperform baseline methods, with GCN aggregation improving accuracy by over 10%.

The session concludes with a Q&A covering story forecasting advantages, keyword extraction tools, clustering pipelines, and the benefits of community‑detection clustering over density‑based methods.