Shotaro Ishihara (Nikkei Inc.), and Yasufumi Nakama [email protected] IEEE BigData 2022, Industry and Government Program Does Text Length matter? Analysis and Estimation of News Article Reading Time with Multimodal Machine Learning

Research Overview 2 ● text length ● headline / body text ● thumbnail image ● others like genre ● past reading history reading time

Summary 1: Dataset 3 ● text length ● headline / body text ● thumbnail image ● others like genre ● past reading history reading time ✅ Real-world content and access log of Nikkei

Summary 2: Text length 4 ● text length ● headline / body text ● thumbnail image ● others like genre ● past reading history reading time ✅ Doesn’t strongly correlate with reading time

Summary 3: Multimodal 5 ● text length ● headline / body text ● thumbnail image ● others like genre ● past reading history reading time ✅ Boosted performance

Outline 6 ● Introduction ● Problem Formulation ● Proposed Method ● Experiments ● Conclusion and Future Work

Reading time estimation helps: 7 ● Push notifications [1] ● Recommendation [2, 4-6] ● User decision support [3, 7] ● Clickbait analysis [22-23]

How can we estimate reading time? 8 ● text length ● headline / body text ● thumbnail image ● others like genre ● past reading history reading time

Research questions 9 1. How much does text length correlate with reading time? 2. How much do features other than text length improve the performance of reading time estimation?

Reading time dataset 10 ● A large dataset that includes reading time from Japanese financial news from the Nikkei. ○ About 1,000 articles a day, 800,000 paid subscribers (and data infrastructure) ○ Larger and more scalable than some existing data on recording eye movements [8] [9] and brain activity [10]

Dataset details 11 100,000 sessions * 3 ● train: 21-12-01 ● val: 21-12-08 ● test: 21-12-15

RQ1: text length (x) & reading time (y) 12 Correlation coefficient is 0.04 (and 0.31)

13 ● Architecture corresponding to the specific data ● E2E fine-tuning Proposed Method

Experiments: Features & Models 14 1. The model was fixed to LightGBM [16] and the features were explored. 2. We fixed the features and observed differences. a. Ridge regression b. MLP c. Proposed method (w/wo E2E fine-tuning)

Experiments: Features 15 Additional features improved the metric. ●

1. mean reading time 2. text length 3. minimum reading time 4. embedding of body text (dimension 193) 5. embedding of thumbnail image (dimension 88) Important features by LightGBM 16

Experiments: Models 17 ● LightGBM worked better in the same feature. ● Proposed method outperformed LightGBM by adding LSTM, and e2e fine-tuning.

Multimodal training tips 18 ● Different learning rate: 2e-5 for BERT, 1e-4 for Swin Transformer, and 1e-2 for the others ● CosineAnnealingLR: For training stability

Conclusion 19 ● We highlighted the importance of reading time and evaluated the implementation. ● Our analysis revealed reading time does not strongly correlate with text length. ● Our experiments showed a multimodal machine learning approach led to a more accurate estimation than simply using text length.

Future Work 20 ● Offline evaluation => Online operation ● Further feature & model exploration ● Clickbait analysis