Abstract
Inspiration moves individuals to see new possibilities and transforms their perception of personal potential. Despite its psychological significance, inspiration has received limited attention in Natural Language Processing (NLP). This work represents the first computational study of inspiration, aiming to automatically detect inspiring content in social media data. We analyze social media posts to identify the characteristics and topics that define inspiration. We introduce a dataset of 11,600 Reddit posts (5,800 inspiring and 5,800 non-inspiring) and employ linguistic heuristics and machine learning models to classify content. Our findings reveal that inspiration is a complex construct distinct from simple positive sentiment, often involving themes of rising above adversity.
Introduction
While social media is frequently criticized for negativity, it also serves as a vast repository of content that motivates and elevates users. However, "inspiration" is a complex psychological construct that transcends basic positive sentiment. Psychologically, it is defined by three pillars: evocation (it is triggered externally), transcendence (orienting toward something greater than oneself), and approach motivation (an urge to act).
Historically, NLP research has focused on sentiment analysis or basic emotions like joy and anger. This study fills a critical gap by defining and detecting "inspiring" content, a necessary step for developing recommendation systems that prioritize user well-being over mere engagement.
Methodology
The research team developed a multi-stage pipeline to transition from raw social media data to a functional classifier:
- Heuristic Filtering: Linguistic heuristics were used to identify candidate inspiring posts on Reddit (e.g., posts containing "inspir*" or "uplift," or responses to specific "AskReddit" prompts).
- Crowdsourcing: Annotators on Amazon Mechanical Turk labeled these candidates. The survey design required participants to identify the influence of the post (e.g., did it improve their mood or motivate action?) and the specific emotions it evoked.
- Classification: Three standard NLP architectures were evaluated: fastText, BERT, and RoBERTa.
- Analysis: To understand the drivers of inspiration, the authors utilized TF-IDF for keyword importance, k-means clustering on GloVe embeddings for topic discovery, and the GoEmotions taxonomy for emotional mapping.
Datasets
The study introduces a new benchmark dataset:
- pR (Reddit): A primary dataset consisting of 5,796 inspiring and 5,796 non-inspiring post IDs, all human-annotated.
- S (Facebook): A larger, weakly-labeled dataset of over 600,000 posts used for initial training and transfer learning experiments.
Key Insights and Ideas
Inspiration vs. Pure Positivity
A significant finding is that inspiring content is not synonymous with pure positivity. Inspiring posts often incorporate "negative" emotions, reflecting the theme of "rising above adversity." Word clusters identified in the data include terms like misery, adversity, refugee, and debt alongside thrive and generosity.
Two Facets of Inspiration
The data suggests two primary types of inspiration: posts that make people "feel good" (shifting mood or perspective) and those that "motivate to act" (encouraging new behaviors). In the Reddit dataset, "feeling good" was the more prevalent response.
Emotional Profile
While "admiration" and "optimism" were the strongest predictors of inspiration, "curiosity" emerged as a significant factor. This aligns with the "transcendence" aspect of inspiration—seeing new possibilities beyond one's current state.
Results and Performance
- Model Accuracy: RoBERTa outperformed other models, achieving an 87.24% accuracy and an 88.08 F1-score on the Reddit test set.
- Human Agreement: Annotators reached a Fleiss Kappa score of 0.26 ("fair agreement"), highlighting the subjective nature of inspiration while confirming that consistent linguistic patterns exist.
- Sentiment Complexity: Most inspiring posts scored in the positive range (0 to 0.5), but a substantial portion were neutral or negative, confirming that inspiration is more nuanced than a simple "happy" label.
Conclusion and Future Work
This work establishes a foundation for the computational study of inspiration. Future research should address the heuristic bias inherent in keyword-based data collection and expand to multi-modal content, including images and video. Additionally, further analysis of linguistic style (e.g., narrative structure and metaphor) could provide deeper insights into how text triggers inspirational experiences.
Summary updated