- generic and easy method using knowledge graph
- improve information, diversity, appropriateness
- reform decoding method and objective function
| Type | Sentence |
|---|---|
| Post | so instead of adding a simple key combination, they make you use their backwards mouse . |
| Reference | no there is a dedicated button for it on the keyboard , or you can just keep your hand on your input device . |
| Beam Search | you 're right . i 'm not a smart guy , but i do n't think it 's necessary to be a mouse |
| + Proposed | i do n't know if you can use a mouse pad , but it 's not easy to control the mouse . |
| Method | BLEU-1 | BLEU-2 | DIST-1 | DIST-2 |
|---|---|---|---|---|
| Beam Search | 14.545 | 2.827 | 3.328 | 12.872 |
| + Proposed | 14.285 | 2.771 | 3.934 | 17.619 |
- use dataset for CCM (Commonsense Knowledge Aware Conversation Generation with Graph Attention)
- Commonsense Conversation Dataset (English)
- embedding vector is GloVe
- dialog data from Reddit (not Twitter)
- commonsense is ConceptNet
- Windows10 (coding and run), Ubuntu18.04 (run)
- python 3.6.8
- pytorch 1.1.0
- CUDA 10.1
- download CCM dataset
- create folders (./data and ./log and ./model)
- dataset into ./data folder
- edit param.py
- $ python train.py
- (stop learning with 'Ctrl+C')
- $ python test.py
- $ pip install rouge (for ROUGE eval)
- $ nltk.download('wordnet') (for METEOR eval)
- $ python eval.py
- prepare data
- create dictionary
- GRU encoder
- GRU decoder
- attention
- greedy search
- sampling
- top-k sampling
- top-p sampling
- beam search
- diverse beam search
- maximum mutual information
- + with knowledge graph
- softmax cross-entropy loss
- inverse token frequency loss
- inverse N-gram frequency loss
- + with knowledge graph
- beam score
- entity score
- breakdown possibility
- auto evaluation
- Length
- BLEU
- NIST
- ROUGE
- DIST
- Repeat
- METEOR
- Entity score
- length normalization
- repetitive suppression
- IDF
- http://coai.cs.tsinghua.edu.cn/hml/dataset/#commonsense
- https://conceptnet.io/
- https://huggingface.co/blog/how-to-generate
- https://miyanetdev.com/archives/1308
- https://qiita.com/m__k/items/646044788c5f94eadc8d
- https://takoroy-ai.hatenadiary.jp/entry/2018/07/02/224216
- jojonki/arXivNotes#159
- https://stanford.edu/~shervine/l/ja/teaching/cs-230/cheatsheet-recurrent-neural-networks
- https://ksksksks2.hatenadiary.jp/entry/20191202/1575212640
- http://unicorn.ike.tottori-u.ac.jp/2010/s072046/paper/graduation-thesis/node32.html
- https://www.nltk.org/api/nltk.translate.html
- http://unicorn.ike.tottori-u.ac.jp/2012/s092013/paper/graduation-thesis/node31.html
- http://yagami12.hatenablog.com/entry/2017/12/30/175113
- https://medium.com/@aiii/pytorch-e64c248ab428
- https://qdata.github.io/deep2Read//talks2019/Extra19s/TkachStochasticBeamSearch.pdf
- http://coai.cs.tsinghua.edu.cn/hml/media/files/2018_commonsense_ZhouHao_3_TYVQ7Iq.pdf
- https://www.anlp.jp/proceedings/annual_meeting/2020/pdf_dir/P6-12.pdf
- https://arxiv.org/pdf/1911.02707.pdf
- https://www.anlp.jp/proceedings/annual_meeting/2019/pdf_dir/F5-2.pdf
- https://www.anlp.jp/proceedings/annual_meeting/2020/pdf_dir/P5-11.pdf
- https://www.anlp.jp/proceedings/annual_meeting/2019/pdf_dir/P3-34.pdf
- https://arxiv.org/pdf/1904.09751.pdf
- https://ahcweb01.naist.jp/papers/conference/2019/201906_SIGNL_sara-as/201906_SIGNL_sara-as.paper.pdf
- https://www.jstage.jst.go.jp/article/jnlp/21/3/21_421/_pdf
- https://arxiv.org/pdf/1911.03587.pdf
- https://arxiv.org/pdf/1611.08562.pdf
- https://www.anlp.jp/proceedings/annual_meeting/2006/pdf_dir/B4-4.pdf
- https://anlp.jp/proceedings/annual_meeting/2018/pdf_dir/A4-2.pdf