Upgrade to Pro — share decks privately, control downloads, hide ads and more …

進化?迷走?CasualConc ファミリーアプリの現在地 @ 英語コーパス学会 2026

Avatar for Yasu Imao Yasu Imao
September 02, 2026

進化?迷走?CasualConc ファミリーアプリの現在地 @ 英語コーパス学会 2026

Avatar for Yasu Imao

Yasu Imao

September 02, 2026

More Decks by Yasu Imao

Other Decks in Research

Transcript

  1. アプリ開発の現状 Vibe Coding 生成 AI に自然言語で指示しながら コードを生成・修正する開発スタイル Codex / Claude

    Code / Gemini Code Assist すべてのコードを自分で書かなくてもアプリケーション開発が可能
  2. would can Harry Potter 2 may 0.3 could 400 Harry

    Potter 3 Oz Narnia LOR HP Oz Narnia LOR HP Oz Narnia LOR HP felt asked place of upon stood Harry Potter 7 time these little once am thought shall than make good much don't this but for have room hand byman 0.1 0.0 -0.1 ** Book 3 - THE TWO TOWERS 1 5 - THE RETURN OF THE KING 1 Book ** Book 2 - THE FELLOWSHIP OF THE RING 2 ** * saw nothing many before -0.6 Harry Potter 6 -0.4 -0.2 w l eyes face harry's mark arts 0 2 0.0 0.2 seen ever never since heart dark long year last 4 time every night Value sprout mcgonagall trelawney flitwick lord Book 6 - The Silver Chair Book 5 - The Voyage of the Dawn Treader Book 7 - The Last Battle Book 4 - Prince Caspian Book 3 - The Horse and His Boy Book 1 - The Magician's Nephew Book 2 - The Lion, the Witch and the Wardrobe Book 1 - THE FELLOWSHIP OF THE RING 1 Book 6 - THE RETURN OF THE KING 2 Book 4 - THE TWO TOWERS 2 Book 3 - THE TWO TOWERS 1 Book 2 - THE FELLOWSHIP OF THE RING 2 Book 5 - THE RETURN OF THE KING 1 Harry Potter 7 Harry Potter 6 Harry Potter 5 Harry Potter 1 Harry Potter 3 Harry Potter 4 Harry Potter 2 THE LOST PRINCESS OF OZ THE MAGIC OF OZ GLINDA OF OZ THE WONDERFUL WIZARD OF OZ RINKITINK IN OZ THE MARVELOUS LAND OF OZ THE ROAD TO OZ THE PATCHWORK GIRL OF OZ THE TIN WOODMAN OF OZ DOROTHY AND THE WIZARD IN OZ THE SCARECROW OF OZ THE EMERALD CITY OF OZ TIK-TOK OF OZ OZMA OF OZ black voldemort professor oh potter norris right now years even just looks tower minutes later ten great hall entrance eater eatersdeath opened behinddoor front told took muttered though tried knew quickly yelled thought felt shouted got heard like moment pulled turned found noticed gryffindor five thing get back toward go let anyone else everyone way came come doors another one better hermioneron sat table ago albus watchedwent whisperedlooking asked saw glanced stood looked stared harry around staff first pomfrey sirius okay well umbridge lucius dumbledore ginnylupin snape vernon's draco moody hagrid tonks madam george malfoy vernon fred uncle nothing neville know luna slughorn ollivander percy said good tell mr sir crouch us fudge yeah quietly weasley mrs yes look ami me my the here while upon as head with man made good sked so little rothy why who to her she make can agic izard eople all when after much found other place other one which this be do these we us our are let has is must no some any find way take here ever go by king own your their them they were for and will shall then stood great came but in efore away far long old not many may yet of side than its went come last there or if even now more that have from left again frodo dark emed took black urned heard ehind ough still hand mself he him his rough eyes mr over yes did ron harry ione essor oking told edore wand didn't agrid door round room get ause i'm an don't want just it's tell know could thing ould three knew how you well might once oh first time two see thing very a never down feet too felt on about ought saw it like ment say only voice been was had face said right what thing look think into at oked going back off got out up 0.4 nick 1 (44.57%) Dimension magicaleye headless nearly creatures o -2 * Book 6 - THE RETURN OF THE KING 2 Book Book 4 - THE TWO TOWERS 2 1 - THE FELLOWSHIP OF THE RING 1 Color Key -4 * ** * 7 - The Last Battle Book -0.4 find here would from after Book 6 - The Silver Chair Book 1 - The Magician's Nephew Book 5 - The Voyage of the Dawn Treader Book 2 - The Lion, the Witch and the Wardrobe Book 3 - The Horse and His Boy Book 4 - Prince Caspian -0.2 which had seemed who has dark Harry Potter 5Harry Potter 6 Harry Potter 2Harry Potter 7 Harry Potter Harry 4 Potter 1 -0.3 could own there Dimension 2 (18.3%) some it's what its wand not you yes think great know two i that old ** *THE *OZ RINKITINK IN ** MARVELOUS LAND OF OZ * * * * * * can when Harry Potter 5 tell an even them THE TIN WOODMAN OF OZ THE PATCHWORK GIRL OF OZ THE SCARECROW OF OZ TIK-TOK OF OZ Harry Potter 3 came left way how other where must why or weknew again are is oh do may looking Oz she no moment your behind Narnia well right king head just LOR yet dumbledore did so want more 0 were it as he they and said voice to if might feet their about like people frodo one first i'm told look get got going been the her though Harry Potter 1 hagrid a didn't face mr me into off be ron harryall on up in our see black down looked him side 100 his three say only out never back us was let through ever around willat Harry Potter 4 HP freq then turned now over went 200 far come professor heard any wizard my go thing made too take while magic another because eyes found long hermione something door very away still with 300 took himself * * * * * *** 0.2 dorothy last THE ROAD TO OZ OZMA OF OZ THE EMERALD CITY OF OZ GLINDA OF OZ THE WONDERFUL WIZARD OFDOROTHY OZ AND THE WIZARD IN OZ THE MAGIC OF OZ THE LOST PRINCESS OF OZ anything say
  3. その後の開発 • Version 4.0 マルチメディアコーパス作成 • Version 4.1 単語埋め込み (word

    embedding) (Word2Vec/FastText) • Version 4.2 BERTopic を利用したトピックモデル分析 • Version 3.5 - MALLET を利用したトピックモデル分析
  4. 開発の方向性 • 開発における生成 AI の利用 ✓CasualConc ファミリーアプリは RubyMotion で開発 ‣

    リソースが少ないためか動くまでに試行錯誤が必要 ✓NLP あたりのアプリは Python で動くものが多い ‣ Python スクリプトはリソースが豊富 ‣ Chat で指示するだけで動くものができてくる
  5. 開発の方向性 • 生成 AI の利用状況 ✓マルチメディアコーパス構築 ‣ 生成 AI の成果物

    (Whisper) を利用 ✓トピックモデル分析 ‣ Python での処理を生成 AI に任せる ✓単語埋め込み分析 ‣ Unix アプリを動かすところで補助的に使用 ‣ Python での計算処理を任せる
  6. OpenAI Whisper • Automatic Speech Recognition (ASR) system ✓自動音声認識 •

    • オープンソース ハルシネーションの問題あり • • WhisperKit 68万時間の多言語・多タスクの教師ありデータに基づく ✓macOS の Core ML を使って Apple Silicon に最適化 ✓蒸留したモデル (英語のみ) で高速処理が可能 (M2 以降) https://openai.com/index/whisper/
  7. マルチメディアコーパス構築 CasualTranscriber での自動書き起こし Intel Mac Whisper のインストールが必要 Homebrew (CLT) Python

    ffmpeg Apple Silicon Mac openai̲whisper-tiny WhisperKit は同梱 モデルも同梱 M1 Mac では 使えるモデルに 制限あり モデルの入手に Command Line Tools は必要
  8. WhisperKit モデル比較 25分程度の英語音声ファイルの書き起こし (M2) (Whisper の large-turbo モデルで実時間程度) モデル 処理時間

    openai̲whisper-large-v3 12 分 44 秒 openai̲whisper-large-v3̲turbo 7 分 29 秒 distil-whisper̲distil-large-v3 2 分 00 秒 distil-whisper̲distil-large-v3̲turbo 1 分 35 秒
  9. • トピックモデル LDA (MALLET/Java) ✓文書中の単語の出現パターンを利用 ‣ 潜在的なトピックを推定 • ✓事前にトピック数の指定が必要 BERTopic

    (Python) ✓事前学習済みの埋め込みモデルを使って文書をベクトルに変換 ‣ 意味の近い文書をまとめてトピックを推定 ✓トピック数は自動決定 (どのトピックにも属さない文書を許容)
  10. • トピックモデル 文書の前処理 (CasualConc / CasualTagger) ✓文書の分割 (プレインテキスト) ✓ストップワードの処理 ✓頻度情報などでの単語の選別

    (CasualConc) ✓POS タグを利用した単語の選別 ‣ TreeTagger, MeCab (CasualConc) ‣ SpaCy (CasualTagger) ✓BERTopic 用 JSON ファイルの出力 (CasualTagger)
  11. トピックモデル (LDA) • CasualConc ✓Java アプリケーションの MALLET を利用 (同梱済み) ✓JDK

    (Java Development Kit) のインストールが必要 (無料) ✓用意する文書ファイル ‣ 1 文書 (ファイル) あたり 500‒2000 語程度 ‣ 文書長はある程度揃っていることが望ましい ‣ CasualConc/CasualTagger でテキストファイルの前処理可能
  12. トピックモデル (BERTopic) • CasualConc ✓Python ライブラリの BERTopic を利用 ✓Python /

    BERTopic のインストールが必要 (venv) ‣ Transformer モデルをダウンロード ✓用意する文書ファイル ‣ 1 文書 (ファイル) あたり 50‒500 語程度 ‣ 文書長を揃える必要はあまりない ‣ CasualTagger でテキストファイルを前処理し JSON で出力
  13. • CasualConc 単語埋め込み ✓Word2Vec / FastText による単語埋め込みモデルを作成 ‣ Word2Vec /

    FastText は同梱 - FastText は未知語や表記の揺れに強い ‣ 類似語の検索 ‣ 単語間の類似度の比較 ‣ コーパス間で単語の意味・用法の違いを比較 - 直交プロクラステス法によるアラインメント (Python)
  14. 単語埋め込み 60 t-SNE: Dim 1 vs. Dim 2 fluency proficiency

    * competence 0 -20 -40 * * input task * -60 output * -80 Dimension 2 20 40 accuracy complexity * * * * feedback * interaction -60 -40 -20 Dimension 1 0 20 40
  15. CasualTagger • タグ付け補助アプリケーション ✓POS タグ付け ‣ EngTagger, TreeTagger, MeCab, Stanford

    CoreNLP, macOS built-in tagger, SpaCy, Stanza ✓構文解析 (dependency) ‣ Stanford CoreNLP, Stanza, SpaCy - 現在 CasualConc で扱えるのは CoreNLP の出力のみ
  16. まとめ (まとまる?) • どこへ向かっているのか結局わからない 色々詰め込み放題 • • ✓何をしていいかわからない場合は色々できることが大切 今後やりたいこと ✓Stanza/SpaCy

    の依存文法タグを CasualConc で使いたい ✓Biber の MDA っぽいこともできたりするのではなかろうか • バグは色々残っているので報告お願いします