koalaonbed/news_consumption

SJA

How to train SJA

1. Scrutinize screenshots to view what kind of unique features is associated to the event you are interested in
2. use labelImg(https://github.com/heartexlabs/labelImg) to annotate image
3. use YOLO to train

How to build SJA to infer

1. Use YOLO to predict object & OCR to infer texts and their boundaries
2. Apply condition expression to identify event and split media units
3. Each media unit has its text, calculate their similarity to judge if they are the same media units or not
4. Determine the media unit is news or not by searching if it contains the keyword from the whitelist of news organizations
5. output the CSV

How to implement

1. first_inference_0912.py
2. OCR_size.py
3. third_split_compare.py

Rules of Facebook screenshots

• Rules

Events or split	Rules
Split posts	The upper boundary of "three_dots" (upper right at every post)
opening external link	labeled 6 kinds of external link's headers
viewing comments	上方的 ">" 和 "讚" & 下方的 "留言框"

• Labels in Facebook https://docs.google.com/presentation/d/1qUElYN8_q6ycYiinwAJWs9vJpD1zuEvrSzLonbqFJO4/edit#slide=id.g11da0161188_0_31

Reference

• https://cloud.google.com/vision/docs/ocr
• https://github.com/AlexeyAB/darknet/tree/darknet_yolo_v3_optimal