Michael
Färber and Achim Rettinger
Finding Novel Facts in Unstructured Text
Abstract: Users often have a
clearly defined information need like monitoring natural-language texts
for novel statements about certain entities. However, common
keyword-based search technologies are focused on finding relevant
documents and cannot judge the novelty of statements contained in the
text. In this work, we propose a new approach where statements that are both novel and relevant are retrieved from natural-language texts. Novelty detection is triggered by end user queries, while relevance is ensured by the compatibility of facts with a background knowledge base representing the user's domain of interest. An evaluation performed on English Wikipedia texts, DBpedia as background knowledge, and our new publicly available gold standard demonstrates the challenges addressed by and the effectiveness of our approach.
Input:
- Knowledge Base
- Document corpus
- Queries
Output:
- Output of our Novelty Detection System for Wikipedia Top 50 Articles
- Output of ReVerb-baseline for Wikipedia Top 50 Articles
Evaluation results (incl. gold standard):
- All
evaluation results
- Overview of gold
standard and evaluation (PDF)
- Preprocessed Wikipedia articles
Demo:
- Screencast ( YouTube-Link
):
- Demo (optimized for 1680x1050 px; Please note that the performance of the server is low.)
(c) 2015 Michael Färber, Institute AIFB, KIT
