This repo contains our research summary for Story Ending Generation (SEG), and we also provide the codes and generated results of our work on SEG.

Contents:

1. SEG Paper List
2. SHGN (DASFAA'22)

Story ending generation is the task of generating an ending sentence of a story given a story context. For example, given the story context:

Today is Halloween. 
Jack is so excited to go trick or treating tonight.
He is going to dress up like a monster.
The costume is real scary.

We hope the SEG model could generate a reasonable ending for the above story, such as:

He hopes to get a lot of candy.

Existing SEG works all utilize ROCStories Corpus to evaluate performances of SEG model. Specifically, the ROCStories Corpus contains 98,162 five-sentence stories, in which the first four sentences is used as story context while the last one is regarded as story ending sentence.

Paper	Conference/Journal	Results (BLEU-1/2)	Evaluation Tools	Code	Tags
From Plots to Endings: A Reinforced Pointer Generator for Story Ending Generation	NLPCC 2018	28.51/11.92	nlg-eval	SEG	arch-LSTM、train-MLE
Generating Reasonable and Diversified Story Ending Using Sequence to Sequence Model with Adversarial Training	COLING 2018	-	-	arch-LSTM、train-GAN、train-MLE
WriterForcing: Generating more interesting story endings	ACL 2019 @ Storytelling	-	-	WriterForcing	arch-GRU、info-Keywords、train-MLE、train-ITF
Learning to Control the Fine-grained Sentiment for Story Ending Generation	ACL 2019 Short	19.8/6.7	-	sentimental-story-ending	arch-LSTM、info-Sentiment、train-MLE
Story Ending Generation with Incremental Encoding and Commonsense Knowledge	AAAI 2019	26.82/9.36	-	StoryEndGen	arch-LSTM、info-knowledge、train-MLE
Generating Diverse Story Continuations with Controllable Semantics	EMNLP 2019 @ NGT	-	-	-	arch-LSTM、info-Controllable、train-MLE
Toward a Better Story End: Collecting Human Evaluation with Reasons	INLG 2019	-	-	SEG_HumanEvaluationReasons	task-Metric
Story Ending Generation with Multi-Level Graph Convolutional Networks over Dependency Trees	AAAI 2021	24.6/8.6	-	MLGCN-DP	arch-LSTM、arch-GCN、info-DP、train-MLE
Incorporating sentimental trend into gated mechanism based transformer network for story ending generation	Neurocomputing 2021	27.03/7.62	-	-	arch-Transformer、info-Sentiment、train-MLE
Incorporating Commonsense Knowledge into Story Ending Generation via Heterogeneous Graph Networks	DASFAA 2022	25.6/9.4	nlg-eval & py-rouge	SHGN	arch-Transformer、arch-gat、info-Sentiment、info-knowledge
CLseg: Contrastive Learning of Story Ending Generation	ICASSP 2022			CLSEG

The concepts used in Tags are illustrated as follows:

• arch：The architecture of the model, includes arch-LSTM、arch-GRU、arch-Transformer and arch-GCN tags.
• train：The training strategy of the model, includes train-MLE、train-GAN and train-ITF tags.
• info：The additional infomation used in SEG, includes info-Keywords、info-Sentiment、info-knowledge、info-DP (Dependency Parsing) and info-Controllable tags.
• task：task-Metric tag indicates the evaluation work.

We provide the codes and generated results of the DASFAA 2022 paper Incorporating Commonsense Knowledge into Story Ending Generation via Heterogeneous Graph Networks.

Please refer to the SHGN directory.

The generated results of our SHGN are available at SHGN.txt. To reproduce the evaluation scores of our paper, please use nlg-eval and py-rouge toolkits to calculate BLEU and ROUGE scores, respectively.