カンファレンスセッションの選択傾向を知り たい / Let’s study trends of entry to conference sessions Kanazawa.rb meetup #84 Takayuki Takagi 

Who am I? ● Takayuki Takagi (高木貴之 / ニボシーニョ) ● @TAKAyuki_atkwsk / takayukiatkwsk ● Freelance programmer ● Remote work ● Scala, Ruby, Python, AWS, Docker, etc. ● Like beer and gyoza 

Today’s topic ● I want to know trends of entry to conference sessions ○ Extract characteristic words from their title and description with NLP(Natural Language Processing) ○ But I’m NOT familiar with NLP, so I want to use as easy tools as possible ○ Easy tools - Cloud APIs 

Cloud APIs for NLP ● Amazon Comprehend API (AWS) ● Cloud Natural Language API (GCP) ○ Syntactic analysis ● Text Analytics API (Azure) ○ Key-phrase extraction API -> These APIs are directly available in Japanese!! 

Make input data ● Copy session titles and descriptions to spreadsheet ○ Japan Container Days 2018 (no descriptions) ○ Scala Kansai Summit 2018 ○ JAWS DAYS 2019 ○ Scala Matsuri 2019 ○ Google Cloud Next Tokyo 2019 ● Export as CSV (script input) ○ id, title ○ id, title + description 

Analysis methods 1. Extract key-phrases with Text Analytics API 2. Analyze syntax with Cloud Natural Language API 3. Analyze syntax with MeCab + NEologd (for comparison) ● Source code ○ https://github.com/TAKAyukiatkwsk/session_analytics_sample 

Extract key-phrases with Text Analytics API 

Key-phrase frequency ● [title] find technical words but they are the low frequency(Max=4, mostly 1) ● [title + description] the high frequency(Max=9) but they are not technical words ● “Kubernetes” is 4 in title, but is 3 in title + description 

Analyze syntax with Cloud Natural Language API 

N-gram Frequency ● [title unigram] More general topics (ex. Scala, Kubernetes, サービス, コンテナ, Cloud) ● Trends: Scala, Kubernetes, Akka, 機械学習, サーバーレス, Cloud Spanner, マイクロサービス 

N-gram Frequency ● [title + desc bigram] more understandable words than title bigram ● “分散トレーシング” is a characteristic phrase 

Analyze syntax with MeCab 

N-gram Frequency ● “型” is tokenized as a noun (as an affix with Natural Language API) ● “機械学習” and “サーバーレス” are tokenzied as one word ● “関数型” is a characteristic phrase 

N-gram Frequency ● There are not abstract words like “よう” “こと” “ため” ● “GraphQL” and “マイクロサービス” are tokenzied as one word (not in chart) ● “分散トレーシング” is a characteristic phrase 

Results ● Trends: Kubernetes, Serverless, Scala, 分散トレーシング, 関数型, Akka, Cloud Spanner ○ The frequency depends on title and description quality ● Cloud APIs are useful ● Key-phrase is not enough, using N-gram too is better ● MeCab + NEologd can analyze better than Native Language API (in Japanese/specific category?) 

Results ● If you are familiar with NLP, please teach me NLP and analyzing methods!!! 

References ● Cloud Natural Language | Cloud Natural Language API | Google Cloud ○ https://cloud.google.com/natural-language/?hl=ja ● Text Analytics API とは - 機能 - - Azure Cognitive Services | Microsoft Docs ○ https://docs.microsoft.com/ja-jp/azure/cognitive-services/text-analytics/overview ● Amazon Comprehend（テキストのインサイトや関係性を検出） | AWS ○ https://aws.amazon.com/jp/comprehend/ ● TF-IDFで見る評価の高いラーメン屋の口コミ傾向（自然言語処理 , TF-IDF, Mecab, wordcloud, 形態素 解析、分かち書き） - ギークなエンジニアを目指す男 ○ https://www.takapy.work/entry/2019/01/14/142128 ● N-gramモデルを利用したテキスト分析　 ―インデックスページ― ○ http://www.shuiren.org/chuden/teach/n-gram/index-j.html