Third outstanding paper presentation, by Elizabeth Clark: Neural Text Generation in Stories Using Entity Representations as Context, work with Yangfeng Ji, Noah A. Smith #naacl2018
Given the context "All of a sudden, Emily walked towards the dragon", and the current sentence "Seth yelled at her to get back but ______" - what are the next words that should be generated?
Can the current state of entities that have been mentioned before be used to improve text generation for stories?
The model has to make decisions, e.g. should the next word refer to an entity? If so, which entity? One that has been mentioned? A new entity?
History, previous sentence, and entity representations are all taken into account to build current context representation. Another setup is "entity-unaware" and only uses previous state and previous sentence. Another setup uses history and entity, but not previous sentence.
Data: Toronto book corpus: 390 books split into 42,000 segments; 43m tokens, 35,000 types; annotations from StanfordNLP to get coreference info.
Evaluation: Mention generation. Passages from corpus are taken with entity mentions deleted; system goes through and selects an entity to fill each entity slot.
Evaluation: Sentence selection. Given a context, the correct next sentence, and a distrator sentence, pick the correct sentence
Human evaluation: MTurkers are asked to choose between two possible continuation sentences, and to explain their choice. What did they pay attention to?
Humans did *not* explain their choice with coreference. When they did mention it, it was not a reason for choosing one sentence, but a reason for rejecting the other sentence
Humans reject sentences e.g. because "she" occurs in the sentence, and the context contains no (obviously) female entities.
Other reasons for human choices include tone or other social cues that seemed out of place wrt given context.
Future directions: Deeper entity knowledge - social commonsense, modeling inter-entity relationships. Structure in story generation: discourse structure, semantics, story structure. New domains: news articles, recipes.
The paper is available at aclanthology.coli.uni-saarland.de/papers/N18-120…
See also: Thread by @ErikaVaris
See also: Thread by @cjmay4754
