OpenTalks.AI - Алексей Бурнаков, Тематическое моделирование новостей на основе детекции цитирований​

Transcript

1. 1.

   NEWS TOPIC MODELLING BASED ON CITATION DETECTION Alexey Burnakov, TASS

2. 2.

   ITAR-TASS www.tass.ru 115th birthday It's been a while…

3. 3.

   News Media Market: A Complex Graph

4. 4.

   Citation Detection

5. 5.

   News Specific Citation Detection Personal data Headline Editorial Cite number

   Cite index Citing media Date News rating Editor rating Board rating Organization rating
6. 6.

   Methods I Cosine similarity Bag of words / tf-idf Generalized

   linear models
7. 7.

   Methods II PageRank Random-Walk Graph Partitioning

8. 8.

   Citation Detection: results precision = 0.89 recall = 0.87 Logistic

   regression output F1 score = 0.88 MCC score = 0.88 AUC: 0.998 We did good at a train dataset
9. 9.

   PageRank: results. TOP-25 of the Russian Mass Media

10. 10.

    NLP Pipeline Raw text Tokenization Who cites TASS Which news

    was cited Topic modelling Customer facing
11. 11.

    Topic Modelling I Motivation: Are there big topics today? Notre-Dame

    de Paris’s on fire : (
12. 12.

    Topic Modelling II Airbus emergency landing Motivation: Are there big

    topics today?
13. 13.

    Topic Modelling III Flood in the Irkutsk Region :( `Losharik`

    Submarine deadly accident :( Motivation: Are there big topics today?
14. 14.

    Topic Report Ex-Kyrgyz president Atambaev seizure by special forces

15. 15.

    Competition Snapshot Which agency did a good job?

16. 16.

    Daily Competition Snapshot Who is the hero of the day?

    Personal data Personal data Personal data Personal data Personal data
17. 17.

    Daily Competition Snapshot https://www.gazeta.ru/politics/2019/08/08_a_12564199.shtml

18. 18.

    THANK YOU!