Classification and clustering in media monitoring: from knowledge engineering to deep learning
1 like197 views
This PhD thesis examines classification and clustering techniques for media monitoring, including news grouping, multi-label text classification, and business polarity detection. It focuses on applying these methods to the PULS media monitoring system, which collects over 10,000 news articles daily. The thesis contributes novel algorithms and datasets for grouping news into stories based on named entity salience, large-scale multi-label text classification balancing training sets, and the first dataset and methods for entity-level business polarity detection.
More Related Content
Similar to Classification and clustering in media monitoring: from knowledge engineering to deep learning (20)
Classification and clustering in media monitoring: from knowledge engineering to deep learning
- 1. Lidia Pivovarova Classification and clustering in media monitoring: from knowledge engineering to deep learning PhD Thesis Supervised by Dr Roman Yangarber December 21st 2018
- 2. 2 Outline ● PULS media monitoring system ● News grouping ● Multi-label text classification ● Business polarity detection ● Conclusion
- 3. 3 Outline ● PULS media monitoring system ● News grouping ● Multi-label text classification ● Business polarity detection ● Conclusion
- 4. 4 PULS media monitoring system ● puls.cs.helsinki.fi ● Applied to several domains, e.g., Cross-Border Security and Epidemic Surveillance ● Thesis focuses on most recent work on Bussiness Intelligence
- 5. 5 PULS business monitoring ● Collects daily ~10,000 articles from more than 1000 news sources ● Pipeline of natural language processing modules: – Named entity recognition (NER) – Multi-label topic and industry classifiers – Polarity detection – Grouping of articles into stories
- 6. 6
- 7. 7
- 10. 10 Named Entity Recognition Grouping Polarity Detection
- 11. 11 Named Entity Recognition Grouping Polarity Detection Multi-label Text Classification
- 12. 12 Named Entity Recognition Grouping Polarity Detection Multi-label Text Classification
- 13. 13 Outline ● PULS media monitoring system ● News grouping ● Multi-label text classification ● Business polarity detection ● Conclusion
- 14. 14
- 15. 15
- 16. 16
- 17. 17 Grouping into stories ● Different from topical text clustering: – fine-grained – named entities are crucial – group size distribution is skewed ● The dataset: manually annotated one ”typical" day ~4000 documents
- 18. 18
- 19. 19 NE salience ● General nature of news articles: – most salient NEs are mentioned early in the text and then repeated – less salient NEs are mentioned in the later paragraphs and are less frequent ● Hierarchical clustering: word-based and name- based distances above a certain threshold
- 20. 20 Outline ● PULS media monitoring system ● News grouping ● Multi-label text classification ● Business polarity detection ● Conclusion
- 21. 21 Topic label Top-level industry sector 2nd -level industry sector
- 22. 22 Topic label Top-level industry sector
- 23. 23 Topic label Top-level industry sector 2nd -level industry sector
- 24. 24 Challenges ● Complex label hierarchy ● Significant class imbalance ● Data drawn from multiple sources over significant time periods
- 25. 25 Ensemble of SVM classifiers ● Single binary classifier trained for each label ● Key idea: – training set balancing for stable performance
- 26. 26 Rote classifier ● Key idea: use company label distribution collected by PULS IE
- 29. 29 Outline ● PULS media monitoring system ● News grouping ● Multi-label text classification ● Business polarity detection ● Conclusion
- 30. 30
- 31. 31 Task ● Similar to (entity-level) sentiment analysis ● But business news: – contain genre-specific word usages – typically do not express emotions or subjectivity ● Cannot use resources developed for more general sentiment analysis
- 32. 32 Data ● Novel dataset: – ~17,000 documents, ~20,000 company names ● Larger than any existing dataset for the task ● Still smaller than datasets usually used for deep learning ● Using much larger data annotated for event classification: – manual mapping – unsupervised feature transfer
- 33. 33
- 34. 34 Outline ● PULS media monitoring system ● News grouping ● Multi-label text classification ● Business polarity detection ● Conclusion
- 35. 35 Contributions ● Grouping: – novel dataset based on real data – novel algorithm based on NE salience – The best method uses combination of salience with domain-specific embeddings
- 36. 36 Contributions ● Large-scale multi-label text classification: – Balancing training set ensures stable performance – NEs combined with keywords yield better performance than keyword or name features alone – Proposed CNN yields higher performance than previously reported methods – NEs are important for industry sector classification, less important for topic classification
- 37. 37 Contributions ● Entity-level polarity detection – novel dataset, largest for the task – unsupervised knowledge transfer outperforms manual mapping – result consistent for two different CNNs
- 38. 38 Lessons learnt ● data pre-processing is important: much effort spent on data clean-up, reorganization and manual annotation ● precise linguistic analysis is important: pattern-based IE engine provides features for ML components ● NEs are important: most tasks in media monitoring gain advantage from special treatment of NEs
- 39. 39 Lidia Pivovarova Classification and clustering in media monitoring: from knowledge engineering to deep learning PhD Thesis Supervised by Dr Roman Yangarber December 21st 2018