5. 何が研究者達をText-to-imageに駆り立てている?
“The additional information from these descriptions could be used to simplify the
image modelling task.” [Mansimov+, ICLR2016]
“Automatic synthesis of realistic images from text would be interesting and useful,
but current AI systems are still far from this goal.” [Reed+, ICML2016]
“Automatic generation of realistic images from text descriptions has numerous
potential applications, for instance in image editing, in computer gaming or in law
enforcement.” [Sharma+, ICLR2018workshop]
“Automatically generating images according to natural language descriptions is a
fundamental problem in many applications, such as art generation and
computer-aided design” [Xu+, CVPR2018]
“Due to its significant potential in a number of applications...” [Qiao+, CVPR2019]
39. 言語モデルによる条件付き生成
先頭に特殊なトークンやキーとなる単語、文をつけて生成していくものが多い
ここ最近流れが来ていそう
・On Extractive and Abstractive Neural Document Summarization with Transformer Language Models
https://arxiv.org/abs/1909.03186
・GPT-based Generation for Classical Chinese Poetry
https://arxiv.org/pdf/1907.00151.pdf
・Transforming Delete, Retrieve, Generate Approach for Controlled Text Style Transfer (EMNLP 2019)
https://arxiv.org/abs/1908.09368
40. ・On Extractive and Abstractive Neural Document Summarization with
Transformer Language Models
重要文をPointerNetworkで抜き出した後、導入、重要文、要約、本文の順に並べた文書生成を自己注意機構ベース言語モデルで学習。推論
時は導入、重要文で条件付して要約を生成 (岡之原氏のツイートより)
https://twitter.com/hillbig/status/1171585503990140928?s=21
既存のseq2seqよりも良い
F-1 ROUGE scores
GTP-2 の220M パラメータ数のバージョンを使用
41. GPT-based Generation for Classical Chinese Poetry
https://arxiv.org/pdf/1907.00151.pdf
・古典中国詩を生成
・先頭に形式、タイトルで条件づけ
・専門の詩人によって書かれたものに近い出力
・中国語のニュースコーパスでpre-training
・publicly available classical Chinese poemsでfine-tuning
42. Transforming Delete, Retrieve, Generate Approach for
Controlled Text Style Transfer
Code: https://github.com/agaralabs/transformer-drg-style-transfer
1. Content と 2. Attributes で文が成り立っていると仮定
AttributesをContentの先頭につけて生成している
例
削除および生成を使用したコンテンツからの否定的な感情文(中立)の生成
Content: The food was and the service was
Output: The food tasteless and the service was horrible.
Delete、Retrieve、Generateを使用した、コンテンツからの否定的な感情文(中立)の生成
Content: The food was and the service was
Attributes: blend, slow
Output: The food was blend and the service was slow.
43. 言語モデルの対話への利用
・Large-Scale Transfer Learning for Natural Language Generation
https://aclweb.org/anthology/P19-1608
・ConvAI2
・NeurIPSで開かれている対話のコンペ
・personachatという、対話データとユーザーの情報から対話を生成
・GPTベースの転移学習が優勝
・優勝者が実装と解説をブログに載せている
・https://medium.com/huggingface/how-to-build-a-state-of-the-art-conversational-ai-with-transfer-learning-2d818ac26313