Toward Intelligent Music Service

By Johnson Wu@LINE TECHPULSE 2019 https://techpulse.line.me/

Transcript

2. Toward Intelligent 
 Music Service > Johnson Wu / LINE Taiwan Data Dev
3. Self-Introduction • Johnson Wu • 2019.03 Join as Data Scientist / Machine learning Engineer @ LINE TW data dev team • Providing data-driven solutions for local projects • LINE Fact-Checker, LINE MUSIC TW, User Tagging, General dictionary services …etc.
4. Agenda > AI technology on LINE MUSIC TW > Module introduction > Summary and Future works
5. AI Technology on LINE MUSIC TW How AI can be applied ? A basic start. > Name Entity Recognition > Link prediction > Word Embedding Knowledge Graph With User Query Extraction of Knowledge From Unstructured Data > Link prediction Auto-Complete Recommendation > SmartText AutoComplete > A trie-based structure model for popularity and correction on Chinese words.
6. AutoComplete Auto-complete user query with popularity search query •the source to start knowledge extraction From 20M+ user queries, 4M+ music dataset … E S O N R T H A Ea… Eason ? Earth song MJ ?
7. Module Introduction How to do auto-complete with popularity? Trie-based search solution trained on 20M+ user query logs. homophones are trained for spell correction as well 武越添（wu yue tien)→五⽉月天 AutoComplete on Recent popular queries First word Complete 31.84% acc Saving input % 34% Middle word Complete 89.93% acc auto-complete ranking #(Eagles) > #(ear) > #(earth)
 Search Count: 60 E G L E R T H A S Search Count: 50 Search Count: 100
8. AI Technology on LINE MUSIC TW How AI can be applied ? Could we understand user’s query? > Name Entity Recognition > Link prediction > Word Embedding Knowledge Graph With User Query Auto-Complete Recommendation > SmartText AutoComplete > A trie-based structure model for popularity and correction on Chinese words. Extraction of Knowledge From Unstructured Data > Link prediction
9. 蔡依林林 Womxnly 蔡依林林 Ugly Beauty 蔡依林林 Lady in Red 蔡依林林 Life Sucks Muse ⽇日不落落 阿信 ⾦金金多蝦 剃⼑刀蔣 夜半三更更 … Link Prediction (1/2) Auto-suggesting with knowledgeable results Jolin Tsai sing_song Ugly Beauty Lady in Red Life Sucks Wom xnly lyrics_written 阿信 剃⼑刀蔣 composed 夜半三更更 compose ⾦金金多蝦 write_lyrics 瘋狂世 界候⿃鳥 ⼈人⽣生有 限公司 same_album Muse ⽇日不落落 Prediction of relationship, entities on the user queries or extracted triples. sing_song eng_name eng_name
10. Link Prediction (2/2) Auto-suggesting with knowledgeable results Related Singers ? Jolin Tsai sing_song Ugly Beauty 我很怕⿊黑 感覺你 的存在 Jay Chou eng_name eng_name Knight sprint Pirate Say love you sing_song: Songs ranked by popularity compose 說好不哭 告⽩白氣球 不該 Knight Spirit Pirate Say Love You Popcorn’s Flavor This Is Love 說好不哭 (by Jay Chou) 告⽩白氣球 (by Jay Chou） 不該（by Jay Chou） … Prediction of relationship, entities on the user queries or extracted triples.
11. Find a very far-distance relation easily with KG. A singer, composer, album and song network Link Prediction (3/3) Relation can be found by Knowledge graph
12. Module Introduction How to do knowledge extraction? GraphSpace: a StarSpace-based solution of word embedding model on entity/relation prediction from a query/ triple input. Trained on 4M entities and 36M relation triples. Example of word analogy: female_singer chinese_named_song #relation: sing_song
 蔡依林林 倒帶 #sing_song 安室奈美惠 嘻哈時尚女王 #sing_song 江蕙 家後 #sing_song
13. AI Technology on LINE MUSIC TW How AI can be applied ? > Name Entity Recognition > Link prediction > Word Embedding Knowledge Graph With User Query Auto-Complete Recommendation > SmartText AutoComplete > A trie-based structure model for popularity and correction on Chinese words. Extraction of Knowledge From Unstructured Data > Link prediction
14. Extract name entities from unstructured articles; compliment related search. Name Entity Recognition Explore knowledge from our data! Randy Merrill Lady Gaga 黃婷 Singer of the Song Detected Entities Related to This Song GJ蔣卓嘉 ⼀一個巨星 的誕⽣生 Album Intro YOU & I 布拉格管 弦樂團 Coco Lee
15. Module Introduction(3/3) How to do name entity recognition? A fine-tuned BERT with bi-lstm based decoder model. Costum tags like “Singer”, “Song ”, “Album”, “Composer”, “Lyric writer”, “Album producer” 
 Model NER F1-score on costum tags Bert+Bilstm 0.63 Bilstm 0.21 Original text NER results 48,190 album articles 54,762 NER valid detections
16. Summary Takeaway 1. A convenient autocomplete function for search service 2. Knowledge graph solution for more intelligent auto-suggest 3. A potential way to extract knowledge from unstructured data.
17. Future Works Making the model better 1. Higher accuracy on NER recognition. 2. More contents to recognize from unstructured articles. •Singer style •Song music Style 3. Human Evaluation designs. 4. Better efficiency for application service.
18. Thank You