SentimentArcs
A self-supervised ensemble method for diachronic sentiment analysis of narrative. It introduced dynamic time warping to diachronic sentiment analysis of narrative, and reported that state-of-the-art transformers can struggle to find narrative arcs.
- Full title. SentimentArcs: A Novel Method for Self-Supervised Sentiment Analysis of Time Series Shows SOTA Transformers Can Struggle Finding Narrative Arcs
- Author. Jon A. Chun
- Year. 2021 (arXiv, 18 October 2021)
- Identifier. arXiv:2110.09454
- Code. github.com/jon-chun/sentiment_arcs, maintained since 2019
- Access. Open (arXiv preprint)
Summary
Sentiment analysis of a full-length narrative is unreliable when it rests on a single model. SentimentArcs passes dozens of sentiment models over the same text and weighs them against one another, treating smoothing and model selection as interpretive choices rather than neutral preprocessing. The resulting arcs are standardized, smoothed with a 10% moving average, and reduced by LTTB downsampling to a common number of points that preserves peaks, valleys, and endpoints. Dynamic time warping then computes the distance between two whole arcs while absorbing the temporal shifts and stretches that separate otherwise similar story shapes, and those distances drive hierarchical clustering. Across corpora running from roughly 1,400 to 13,000 data points per novel, this is what makes arcs of unequal length comparable by distance.
What this work contributed
- An ensemble, not a single trusted model. Dozens of sentiment models are weighed against one another, and the disagreement between them is treated as information.
- Dynamic time warping for narrative. DTW was introduced to diachronic sentiment analysis of narrative, giving a computable distance between two whole arcs of unequal length. DTW itself is a general time-series technique that long predates this work; the contribution is its application here, paired with LTTB downsampling and hierarchical clustering.
- A negative result. State-of-the-art transformers can struggle to find narrative arcs. This is a limit of scale, not of tuning.
- Working infrastructure. The open-source implementation, maintained since 2019, is the technical foundation of The Shapes of Stories (Cambridge University Press, 2022).
Relationship to earlier and later work
It follows the 2019 “middle reading” paper, which asked whether sentiment methods survive nonlinear narrative. It is extended by eXplainable AI with GPT4 (2023), which adds sentence-level trajectories and comparability on shared coordinates, and by MultiSentimentArcs (2024), which carries the method across modalities. The technical lineage is set out on Narrative trajectories & DTW; scholarly uptake is traced on Reception.
Cite
Plain text. Chun, Jon. “SentimentArcs: A Novel Method for Self-Supervised Sentiment Analysis of Time Series Shows SOTA Transformers Can Struggle Finding Narrative Arcs.” arXiv:2110.09454 (2021).
BibTeX.
@article{chun2021sentimentarcs,
title = {SentimentArcs: A Novel Method for Self-Supervised Sentiment Analysis
of Time Series Shows SOTA Transformers Can Struggle Finding Narrative Arcs},
author = {Chun, Jon},
journal= {arXiv preprint arXiv:2110.09454},
year = {2021}
}