1 / 20

From bursty patterns to bursty facts: The effectiveness of temporal text mining for news

Explore the effectiveness of temporal text mining methods through a comprehensive evaluation framework. This study delves into different methods like bursty words, group representation, and combo representation, showcasing strengths and weaknesses. The approach includes cross-method evaluation and query generation to improve results, with a focus on precision-oriented measures and named entities. Results suggest that combo representation methods are more robust, especially in precision-oriented tasks. The study also highlights the importance of standardized, diverse datasets for future work and suggests exploring sources of bias in method evaluation.

ngilliland
Télécharger la présentation

From bursty patterns to bursty facts: The effectiveness of temporal text mining for news

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. From bursty patterns to bursty facts:The effectiveness of temporal text mining for news Ilija Subašić & Bettina Berendt K.U. Leuven, Belgium

  2. Agenda • The first problem: temporal text mining • Solution methods • The second problem: evaluation of 2) • Our approach: cross-evaluation framework • Case study: evaluation of 3 methods

  3. Temporal text mining (TTM): The problem time • What happened? • What were the important new developments in a time period?

  4. TTM: keyword representation methods [Kleinberg, 2002] “bursty“ ~ more frequent in 1985-94 than in the whole analysed time

  5. TTM:group representation methods [Mei & Zhai, 2005]

  6. TTM: combo representation methods– STORIES: graphical summary [Subašić, Berendt, 2008+]

  7. Demo at ECML/PKDD STORIES: story tracking and exploration

  8. Evaluations so far • Standardized tasks and competitions • DUC update task & ROUGE framework: summarization • TREC Novelty track (2002-04): novel-sentence retrieval • Disadvantages: • # documents too small: 10 for DUC and 25 per topic for TREC • Output is textual  not possible to compare all TTM methods • TTM: evaluations limited to the respective method and corpora

  9. Retrieved sentences Patterns Queries Cross-evaluation: Our approach (1) Sentence retrieval Query-likelihood retrieval QL More : https://sites.google.com/site/subasicilija/ttm-evaluation Query generation Generic / method-specific (top bursty elements & combinations) shave hair britney_spears Last Friday, pop star Britney shaved her head, parting with her long hair.

  10. Our approach (2):Sentences‘ “precision/recall“ Sentence retrieval Query-likelihood retrieval QL Query generation Generic / method-specific (top bursty elements & combinations) Retrieved sentences Patterns Queries IR-style evaluation ROUGE2, ROUGE.SU4, aggregate measures, Friedman and Tukey‘s multiple comparison test Ground-truth sentences

  11. Our approach (3):“Recall-oriented“ aggregate measure maxMR Retrieved sentences t Ground- truth sentences t Best fit (ROUGE) Normalize by max. possible best fit All sentences t

  12. Our approach (4):“Precision-oriented“ aggregate measure maxMP Retrieved sentences method I Ground- truth sentences  Method II has a better chance of good matches  Scale maxMR by Retrieved sentences method II

  13. Case study experiment: Data & settings • Corpus 1: Crime case • 21 weeks, 306 documents, 31 ground-truth sentences • Corpus 2: Celebrity reporting • 8 weeks, 3000 documents, 19 ground-truth sentences • Corpora available at https://sites.google.com/site/subasicilija/ttm-evaluation • M1: a keyword representation method: Kleinberg‘s bursty words • M2: a group repr.method: Mei & Zhai‘s temporal text mining • M3: a combo representation method: STORIES

  14. Results: Top group method comparison maxMR maxMP

  15. Results: Top group method comparison maxMR maxMP

  16. maxMP maxMR Results: Query generation comparison

  17. maxMP maxMR Results: Query generation comparison

  18. Summary • First cross-methods evaluation framework for Temporal Text Mining methods with different patterns • Experimental investigation of 3 TTM types • Results: • different methods – different strengths and weaknesses • M3/named entities: most robust method over settings • M3 variants > M1, M2 in “precision-oriented” measures • specific query generation improves “precision-oriented” results, especially for M1 and M2 • corpus dependence

  19. Future work • Standardized, bigger, more varied datasets • Establish a baseline (ROUGE originally for longer text sequences) • Explore possible sources of bias for/against specific methods • User studies (in progress)

  20. STORIES: graphical summary, textual summary, documents

More Related