LongLM-Eyas

IIE-NLP-Eyas@OutGen: Chinese Outline-guided Story Generation via a Progressive Plot-Event-Story Framework [PDF]

Team Members: Yuqiang Xie, Yunpeng Li, Wei Peng, Ping Guo and Luxi Xing.

Org.: Institute of Information Engineering, Chinese Academy of Sciences, Beijing, China.

Codes are contributed by:

• Yuqiang Xie (Baselines, Data Pre-processing, Event Generation)
• Yunpeng Li (Event Ranking)
• Wei Peng (LCS).

Plot-Event-Story (PES)

Outline:

• Plot (Step 1-2)
• Event (Step 3-4)
• Story (Step 5-8)

A Simple Guide:

Step 1:

convert train/val/test.jsonl into events of each plot

python ./tools/split_kw_sent.py

-> train/val/test_split.jsonl

Step 2:

convert train/val/test_split.jsonl into bart format (source and target)

python ./tools/convert_bartio.py

-> train/val/test.source/target

Step 3:

Train/Eval/Test using LongLM-small model

bash ./longlm/finetune_deepspeed_iie.sh

The best model will be in ./save_model

Step 4:

Generating stories by Top-p sampling:

python ./baselines/generation/gen.py

-> result_of_val/test.txt

Step 5:

convert each event into one line with ‘/t’ splitting

python ./tools/event2data.py

-> result4rank_of_val/test.txt

Step 6:

perform ranking

python ./tools/outline_reranking.py

-> train/val_reranking.jsonl

python ./tools/process_nsp_data.py

-> train/val_nsp.txt

python ./tools/story_nsp.py

-> rerank_test.txt

Step 7:

del repetitive words

python ./tools/data4lcs.py

-> result4lcs.txt

python ./tools/lrc.py

-> final_result.txt

Step 8:

python ./tools/source2jsonl.py

-> submission.jsonl

Parameters for Baselines and Event Generation:

learning rate: 3e-5
epoch: {5, 10}
top-k: K=40
K's temperature: 0.9
batch size: 8

Acknowledgement

Thanks for the baseline model LongLM.