TL;DR
This paper presents a methodology for mapping the complex, socially constructed network of the "artworld" by mining Twitter data. By combining seed-based network expansion with Natural Language Processing (NLP) for event detection and Named Entity Recognition (NER), the author successfully identifies thousands of influential actors—including those without social media accounts—and uncovers the community structures that define contemporary art value.
What Problem Does This Paper Address?
The "artworld" isn't just a collection of paintings; it is a network of actors (artists, curators, museums, galleries) whose interactions determine the cultural and economic value of art. Traditionally, mapping this network required cumbersome surveys or interviews. Furthermore, the artworld is in constant flux, making it difficult to identify all relevant players or quantify their influence.
While social media mining has been used to predict elections or track disease outbreaks, it hasn't been deeply explored for artworld analysis. The challenge lies in the "noise": how do you distinguish a meaningful artworld interaction from a generic tweet? How do you account for legendary artists like Picasso who obviously don't have Twitter accounts but remain central to the conversation? This paper seeks to bridge these gaps using computational methods.
Methods
The author proposes a multi-stage pipeline to construct and refine the artworld network:
- Seed Selection: The process begins with a "seed" list of 108 exemplar actors derived from ArtReview’s "Power 100" rankings (2015–2017).
- Network Construction: A directed graph is built where nodes are actors and edges represent interactions: mentions, retweets, and "following" relationships.
- NLP Filtering (Event Detection): To remove non-art noise (like a museum tweeting about a photo contest on Instagram), the author uses TempEx for temporal expression resolution and a probabilistic model to classify tweets. Only tweets referring to art-related processes (e.g., "exhibition," "curated," "retrospective") are retained.
- Named Entity Recognition (NER): To capture actors without Twitter handles or those mentioned by their real names, the author employs T-NER, an NLP tool specifically tuned for the informal language of tweets.
- Graph Analysis:
- Louvain Method: Used for community detection to find clusters of densely connected actors.
- HITS Centrality: Used to evaluate "Authority" (how much an actor is mentioned/followed by others) and "Hub" status (how much an actor mentions/follows others).
Datasets
The primary dataset consists of 57,888 public tweets posted in 2018 by the 108 exemplar actors. The initial seed list was geographically diverse, though weighted toward the USA (31.48%) and UK (13.89%).
Key Insights and Ideas
- The "Elitist" Filter: Simply mapping all interactions leads to a massive, noisy network of over 30,000 actors, including generic entities like "Instagram." By filtering for specific art-related events, the network shrinks to a more "elitist" and relevant core of ~2,200 actors.
- Identifying the "Invisible": The use of NER is a critical "aha moment." It allowed the inclusion of historical figures (Picasso, Warhol) and contemporary artists who avoid social media but are frequently discussed by those who are on it. This significantly improved the accuracy of authority values for these individuals.
- Geographic Clustering: The Louvain method revealed that subcommunities often form along geographic lines. For example, a distinct cluster emerged around the Centre Pompidou, representing the French art scene.
- Hubs vs. Authorities: The study found that commercial galleries (like Gagosian) often act as high-value "hubs"—they follow and mention many others—while major museums and publications act as "authorities" that others look toward.
Key Results
The methodology demonstrated impressive scalability. Starting from just 108 seeds (shown in Figure 2), the system identified over 30,000 potential actors through raw interactions (Figure 3).
After applying the art-event filter, the network was refined to 2,188 high-relevance actors. The addition of NER further expanded this to 6,862 actors, successfully integrating names like Yoko Ono and Gerhard Richter into the network structure. Quantitative shifts were noted: when filtering for events, institutions like the National Gallery of Art saw significant increases in their authority values compared to the raw interaction graph.
Limitations and Future Work
The author acknowledges several hurdles:
- Language Bias: The current NLP pipeline is optimized for English. This resulted in the loss of approximately 3,000 art-related tweets in other languages, potentially underrepresenting non-English speaking actors.
- Platform Specificity: The study relies solely on Twitter. Integrating data from Instagram (visual-heavy) or Facebook could provide a more holistic view.
- Seed Bias: While the ArtReview list is a standard, it has been criticized for a pro-British bias, which may propagate through the network construction.
Future work aims to extend NLP methods to multiple languages and execute the methodology in real-time to track how the artworld transforms as events happen.
Summary updated