LTUHH@STSS - Applying Coreference to Literary Scene Segmentation.


Cited 0|Views1
No score
In this work, we describe a system for scene segmentation that, relying on character constellations as one of the defining characteristics of scenes, employs a state-of-the-art coreference system. Conceptually building on one of the presented baseline systems, we use a transformer model, enhanced with additional coreference-based features, to identify scene boundaries on the basis of sentence pairs. Finding one of our system’s core weaknesses to lie in its local decision making, we adapt an equidistance constraint, avoiding the common error of predicting very short scenes that in many cases only cover a single sentence. We show that coreference is a suitable feature for scene segmentation and experiment with dynamic programming approaches for non-local decisions. This work is a submission for the shared task scene segmentation (STSS) held at KONVENS 2021, where task participants were asked to, given annotated training data, build systems that split novels into scenes: segments narrating a coherent action in one location with the same characters. Our system ranks 4/4 and 4/5 in Track 1 and Track 2, respectively.
Translated text
AI Read Science
Must-Reading Tree
Generate MRT to find the research sequence of this paper
Chat Paper
Summary is being generated by the instructions you defined