Can we improve narrative closure with bookending, i.e., relating the last sentence back to the first sentence? We present RENarGen, a paradigm for both LMs and LLMs that generates narratives with related endpoint sentences. Accepted at #NAACL2024

Excited to present our work “Returning to the Start: Generating Narratives with Related Endpoints” at #NAACL2024 during Poster Session 2 on 17 June @ 2 PM! Paper: Poster:

@mark_riedl @snigdhac25 Thank you for sharing this relevant work with us. We will add a reference to it in our camera-ready version

Feel free to reach out if you have any questions Paper: Code:

Automatic and human evaluations indicate RENarGen generated narratives have better related endpoint sentences, have more closure/coherency, and are more preferred than narratives generated by the baselines.

RENarGen for LLMs uses 6 different techniques for bookending including semantic relatedness, erotetic relatedness, and “matching ending” relatedness

RENarGen for LMs uses semantic relatedness to generate a related last sentence. Then it completes the rest of the narrative with a dynamic infilling method