Traditional sentiment analysis operates on a simple premise: if a writer uses "happy" words, the text is positive. However, this approach focuses exclusively on expressed sentiment—the internal state or intent of the author. It fails to capture induced sentiment, which is the actual emotional response triggered in the audience.
A news report about a tragedy might be written in a perfectly neutral, objective tone, yet it induces profound sadness or anger in its readers. Conversely, a horror movie character might be screaming in terror (expressed fear), but the audience feels a sense of thrill or excitement (induced arousal). Researchers are shifting the focus from the "speaker" to the "listener" to better understand social impact, media effectiveness, and implicit communication.
The Divergence Between Intent and Reaction
The core motivation for this research direction is the frequent mismatch between what a creator intends and how a crowd reacts. Kevin Hsin-Yih Lin and Hsin-Hsi Chen argue that reader reactions often diverge significantly from the literal content of a text. For instance, an author's attempt at humor might be perceived as condescending, or a neutral disaster report might be the most "emotional" text a person reads all day.
This divergence makes "perceived" or "induced" emotion a more actionable metric for fields like disaster management. Cornelia Caragea, Shrey Desai, and Junyi Jessy Li found that during hurricanes, the audience's interpretation of a message is often more critical for relief coordination than the author's original internal state. Similarly, in literary analysis, Steffen Eger suggests that grounding annotations in the reader's experience provides a more stable and observable target for AI models than trying to reconstruct a long-dead poet's original intent.
Modeling Emotions as Distributions
Because a single text can evoke different feelings in different people, researchers have moved away from single-label classification (e.g., "This text is Sad") toward emotional distribution regression.
Kevin Hsin-Yih Lin and Hsin-Hsi Chen pioneered frameworks that treat reader emotions as a ranked distribution, acknowledging that an article might be 60% "angry," 30% "worried," and 10% "sympathetic." Xin Li and colleagues furthered this by using a Reader Perspective Weighted Model, which systematically captures the collective response of a crowd to objective information. This allows models to predict "social emotions"—the aggregate reaction of a population to a trending topic.
Using Comments as a Mirror
One of the most effective ways to train models on induced sentiment is to use the audience's own words as a "mirror." Instead of asking experts to guess how people feel, researchers look at the comments section.
- Micro-videos: Qi Jia and Zhenhua Guo developed the MVIndEmo dataset, which uses millions of user comments to label the emotions induced by short-form videos.
- Social Impact: Jing Li and Keyang Ding introduced the HEC-Dataset, showing that incorporating user comments as context allows models to outperform individual humans in predicting how a broad population will react to a topic.
- Detecting Condescension: Jiaxun Yang and Bin Li demonstrated that "soft" toxic speech, like patronizing language, is often invisible if you only look at the creator. By using comments as a contextual signal, their CPCLDetector can identify when an audience feels talked down to, even if the video itself appears polite.
Simulating the Reader
The latest frontier in this area involves moving from passive observation to active simulation. If the "implied reader" is the key to understanding a text's emotion, why not build one?
Jian Liao 0005 and Yu Feng introduced RAPPIE, a framework that uses Large Language Models (LLMs) to act as "reader agents." These agents simulate how different personas would react to a post. By analyzing the propagation of these simulated emotions through a graph neural network, the model can identify subtle, implicit emotions that traditional text-centric models miss. This confirms a growing consensus: the perspective of the reader is not just a secondary metric, but a primary lens for decoding human communication.
Multimodal Evocation in Media
In video and film, the gap between expressed and induced emotion is even wider. Ha Thi Phuong Thao, Dorien Herremans, and Gemma Roig developed AttendAffectNet to predict viewer arousal and valence. Their research highlights that viewer affect is driven by a complex synthesis of cinematic elements—like audio cues and optical flow—rather than just the facial expressions of the actors on screen. This approach has been extended to advertising, where models like TSAM can predict how a crowd will respond to a commercial by training on millions of human annotations, allowing brands to "decode" viewer emotions before an ad even airs.