Public Sphere 2.0
Targeted Commenting in Online News Media
DSLAB PR @ 2019.07.11
이준범 ( jun@beomi.net)
Abstract
- News consumption => Max(ad revenue)
- Reader Engagement 👍 => *Comments*
- Traditional view: Comments from the full article
- Present Landscape:
Comments from only particular sections of the article - Build Neural Net to find 'comment ~ article section'
1. Introduction
Background
- Paradigm shift consuming news 📰
- Online media > Offline media
- Comment & Share ideas on articles
- Comments = Most effective tool for user engagement
- Online news acts as the facilitator of public debates
User patterns
- F-shaped pattern
- User's attention is focused on initial texts
- News website = New public sphere
2. Dataset & Motiv
Dataset
- 1352 Guardian
- 1020 NYTimes
- with Comments
Dataset
- 60%+ comments in >20 paragraphs length
- More Longer Article = More comments
Dataset
- Label data with relevance score 1(irrelevant)~5(relevant)
- Judged by presence and absence of
- common words
- common thoughts - 2 Annotators => Cohen Kappa 0.71
Overview
- 42.7% = Relevent to the *whole* article
- 48.9%/48.8% = Relevent to 2-3 paragraphs
- More releveant comments
=> Beginning paragraph
3. Linking
Comments to Paragraphs
Approach
- Baseline: Traditional ML
ex) NB,DT,RF,K-NN,R-SVM,AdaBoosts,LR - DeepLearning Method:
- LSTM
- GRU 👈 Best Score!
Bi-LSTM & GRU
- Pretrained 300-dim Google News Vectors
- Text(article/comment) 👉 Embedding Vector
(OOV 👉 zero vector) - Text Vector = avg(word vectors)
- LSTM / GRU 👉 150 dim vector
- Merge Article vec + Comment vec
👉 FC layer 👉 Softmax
👉 5 classes 🔥
Eval
- 10-fold cross-validation
(DL Model: val each epoch, total 5 epochs)
Related works
- Comment Ranking
- Comment Recommendation
- Comment Analysis
Conclusion
- Traditional comment UI needs revamp
- Comments are more related to particular sections
- DeepLearning, GRU performs better
(Could BERT/XLNet perform better?)