CLEF 2019

Conference Link: http://clef2019.clef-initiative.eu/
Paper Link: http://ceur-ws.org/Vol-2380/paper_98.pdf
Submissions of CLEF 2019 Check-That

• Core approach:

  • We have used a hybrid approach in which we use MART algorithm with 421 linguistic based features. Next, we apply a rule-based approach to decrease the ranks of sentences with phrases like “thank you”. We set the number of trees to 50 and the number of leaves to 2.

• Important/interesting/novel novel representations used (features, embeddings...):

  • Primary: Our features include:
    • the existence of named entities identified by Stanford NLP tool [1].
    • the topic of sentences identified by IBM Watson’s NLP tool [2].
    • POS tags, bi-grams that appear at least 50 times in only check-worthy sentences or only not-check-worthy sentences in the training set.
    • whether the sentence is a question sentence or not.
  • Contrastive 1: In addition to features used in our primary method, we also used the speaker of the statements as features.
  • Contrastive 2: We used the same features as Contrastive 1 but we used logistic regression instead of MART.

• Important/interesting/novel tools used:

  • We used IBM-Watson, Stanford-NLP tool for feature extraction and Ranklib and sci-kit libraries for MART and logistic regression models, respectively.

• Significant data pre/post-processing:

  • We have not applied any preprocessing step.

• Other data used (outside of the provided):

  • We have not used any external data. We have just used the provided training data.

• References (if applicable):

  • https://nlp.stanford.edu/
  • https://www.ibm.com/watson/services/natural-language-understanding/
  • https://sourceforge.net/p/lemur/wiki/RankLib/