Introduction

Crypto art is limited-edition digital art, cryptographically registered with a non-fungible token on a blockchain (Bailey 2017; Yaga et al. 2018). Tokens represent a transparent, auditable origin for a piece of digital art. Blockchain technologies allow tokens to be held and securely traded without the involvement of third parties.

Chiara Braidotti, one of the earliest crypto art curators, in the forward of the catalog of the exhibition The Truth Is by crypto artist Hackatao, says (translated by the author):

“The real potential of the emerging crypto art current is to give a digital image the dignity of a true work of art, made unique, eternal and collectible through blockchain technology. Born almost for fun, crypto art animations are now protagonists of the growing digital art market associated with tokens, cryptographic codes that prove their authenticity. In relation to the infinite reproducibility of a virtual image, these codes are the equivalent of the artist’s signature, the unequivocal authentication that identifies the original work, eternally stored in the blockchain’s decentralized ledger”.

Sergio Scalet from Hackatao, a first mover into crypto art territory, during an interview for Coiners, speaks about this new trend as follows (translated by the author):

“Crypto art is one of the few artistic movements of this millennium that speaks the language of its time; it allows new digital expression languages to have a fast, pulsing, tailor-made market on the new generation of collectors. If until some time ago an artist who created digital artworks had very little chance to put them on the market, for the obvious reason of their infinite reproducibility, today thanks to blockchain technology these digital artworks are created in a certified small number of editions and hence acquire an exchangeable value. Crypto artworks remain infinitely reproducible and visible to all, but only the purchasing collector owns what the artist will call the original, the unique token of the work. Obviously, the collector must accept the perspective of owning a digital work and not a physical object to hang on the wall. A concept much more understandable to the Millenials than to the generation of Boomers still linked to the material, weighty palpability of a work of art”.

Crypto art draws its origins from conceptual art (Lippard 1973): sharing the immaterial and distributive nature of artworks, the tight blending of artworks with currency, and the rejection of conventional art markets and institutions. Notable examples are Duchamp’s Tzanck Check (1919) and Seth Siegelaub’s Artist’s Contract (1971). Many characteristics that were endemic to Duchamp’s practice, and later conceptual art in the 1960s and 1970s, are now visible in the immaterial and distributive logic of crypto art. See Finucane (2018) for a parallelism between conceptual and crypto art and Franceschet et al. (2020) for a crypto art position paper written by artists, collectors, gallerists, art historians and data scientists.

One of the major crypto art galleries is SuperRare. Born in April 2018, at the writing time the gallery features 7125 artworks, more than half sold at least one time with an average resale value of +895%, for a value of 446,626$ earned by more than 200 artists. The typical workflow of crypto art on SuperRare is as follows:

  1. an artist creates a digital artwork (an image or animation) and uploads it to the gallery. The author specifies the title, description, a list of tag words and possibly a price;
  2. the smart contract of the gallery creates a non-fungible token on the Ethereum blockchain associated with the artwork and transfers the token to the artist’s wallet;
  3. the gallery distributes the artwork file over the IPFS peer-to-peer network; hence neither the token nor the artwork are on any central server;
  4. collectors can place bids on the artwork by transferring the bidden amount to the smart contract of the gallery (the collector can withdraw bids at any time);
  5. eventually the artist accepts one of the bids: the smart contract of the gallery transfers the artwork’s token to the collector’s wallet and the agreed cryptocurrency to the artist’s wallet;
  6. the artwork remains tradable on the market. Each re-sale in the secondary market of the gallery keeps rewarding the original artist.

For example, EthGirl, a collaboration between traditional painter Trevor Jones and digital artist Alotta Money, was tokenized on December 14, 2019, and immediately received a long series of significant bids from diverse collectors (technically, a bid war). It was acquired by collector moderats the day after for 59.5 Ether (plus 12.6 Ether of gallery fees).

As observed in Franceschet and Colavizza (2019), the crypto art market configures as a timed stream of events involving the creation and acquisition of artworks. This flow produces data of different varieties, including:

In this paper, we focus on the unstructured, textual information that accompanies each artwork. We propose a text mining analysis (Aggarwal and Zhai 2012) on the textual metadata (title, description and tag words) of artworks present in SuperRare gallery. In particular, we perform the following analyses:

In this work we use tidyverse R packages for data science and tidytext R package for text mining (Wickham and Grolemund 2017; Silge and Robinson 2017).

Sentiment Analysis

We made a sentiment analysis on the text in the metadata (title, description and tags) of artworks using the lexicon nrc by Saif Mohammad and Peter Turney. The nrc lexicon is a list of English words and their associations with eight basic emotions (anger, fear, anticipation, trust, surprise, sadness, joy, and disgust) and two sentiment poles (negative and positive).

For each artwork and emotion, we computed the number of words used in the artwork metadata that match the emotion, the total number of words used in the artwork metadata and found in the nrc lexicon, and the ratio between the two. For each emotion, we sorted the artworks by the emotion ratio. In the next table we show the artworks leading the rankings for the different emotions and the links to their gallery display on SuperRare gallery. In order to consider only artworks with a long enough description, we included only tokens with metadata containing at least 60 words in the lexicon. Following the table, we briefly describe the artworks leading each emotion ranking.

The artworks leading the rankings for the different emotions and the links to their gallery display on SuperRare gallery (we included only tokens with metadata containing at least 60 words in the lexicon).
tokenId emotion emotion words total words ratio tokenLink
5230 positive 34 97 0.3505155 https://superrare.co/artwork-v2/5230
2197 negative 15 65 0.2307692 https://superrare.co/artwork/2197
4285 trust 16 77 0.2077922 https://superrare.co/artwork/4285
2062 joy 20 109 0.1834862 https://superrare.co/artwork/2062
2815 anticipation 14 79 0.1772152 https://superrare.co/artwork/2815
5092 fear 11 64 0.1718750 https://superrare.co/artwork-v2/5092
2197 sadness 10 65 0.1538462 https://superrare.co/artwork/2197
2197 anger 9 65 0.1384615 https://superrare.co/artwork/2197
5092 disgust 8 64 0.1250000 https://superrare.co/artwork-v2/5092
4757 surprise 11 105 0.1047619 https://superrare.co/artwork-v2/4757
  1. the most positive artwork is Discover Your Hidden Talent! by Saito Kareshi. It is an encouraging piece claiming that by opening one’s mind to multiple possibilities one can best evaluate their probable talent;
  2. the most negative artwork is Spleen by hex6c, dedicated to the homonymous poem by Charles Baudelaire. Notably, this is also leading the rankings of sadness and anger emotions;
  3. the leader of trust emotion is Golden Ganesha, Guan Yin & Guan Yu by Saito Kareshi, a piece reassuring the observer that as long as there is mental peace, inner harmony and a strong will to not fear problems, one will get past all obstacles and mental barriers, succeeding in any endeavor by focusing on the goal;
  4. the leader of joy emotion is Silly Daddy Merry-Go-Round by Joe Chiappetta. This is a brief but significant abstract of the description that the author makes about the piece: “Since the circle represents unity, and since unity is what all reasonable families genuinely desire, I set out to make some animated digital art that rotates in a circle, represents the family, and communicates a happy, fun unity”;
  5. the leader of anticipation emotion is ALEPH-2 by Yura Miron, a controversial piece describing a psychedelic trip after assumption of drugs;
  6. the leader of fear and disgust emotions is Burn In Hell by Saito Kareshi. The piece is dedicated to hell, here is an emblematic abstract of the description made by the author: “In religion and folklore, Hell is an afterlife location in which evil souls are subjected to punitive suffering, often torture as eternal punishment after death”;
  7. finally, the leader of surprise emotion is Sugar: The Silent Killer by Saito Kareshi again (the artist seems to have a clear talent for expressing emotions through art). As the author explains, the piece highlights the feeling we have when we eat something which we should not indulge into as often: while it feels very tasty, we know with a sense of guilt, that no matter what we tell ourselves, our body keeps the score and sooner or later we shall see what we are ‘sowing’.

Emotions and Market Success

In this part we correlated emotions with quantitative indicators of market success for artworks. We want to investigate if artworks expressing a clear emotion in their in metadata are also more successful among collectors.

To this end, we filtered artworks with at least 10 metadata words in the lexicon. Then, using Pearson method, we correlated the sale variable (a Boolean variable where 0 is unsold and 1 is sold) with each sentiment and emotion variable over the selected artworks. Each sentiment/emotion variable is a numeric variable giving the number of words with that sentiment/emotion used in the metadata for the artwork.

We also made a finer analysis using TokenRank metric for artworks. TokenRank is an overall rating for artworks on SuperRare considering the following facets of success:

  1. market: the priced amount of all sales and bids made by the artwork (on the primary and the secondary markets of the gallery). The amount is expressed in fiat money using the exchange rate (according to coinmarketcap.com) of the time of the event (sale or bid);
  2. artist: the importance of the artist that created the artwork, measured as the overall amount of sales made by the artist. Again, the amount is expressed in fiat money using the exchange rates of the sale times;
  3. popularity: the number of likes and views collected by the artwork on the gallery website;
  4. speed: the speed of the first incoming bid that the artwork received, relative to the artwork creation time.

In particular, the metric TokenRank is obtained as follows:

  1. for each facet (market, artist, popularity and speed) we extract the percentile of the token facet within the gallery facet distribution. The percentile is a number between 0 and 1, the higher the better for the token. For instance, a percentile of 0.75 for a facet of a given token means that 75% of the tokens in the gallery have a lower value for that facet. We used a percentile approach since we observed that the facet distributions are highly right-skewed;
  2. the final rating is then defined as the weighted mean of the 4 percentile facets using the following weights: 0.4 for market, 0.3 for artist, 0.2 for popularity, and 0.1 for speed.

First of all, we noticed that 52% of the artworks in the gallery have been sold at least once. However, the sale variable is not significantly correlated with any of the sentiment and emotion variables (significance level = 0.01). On the other hand, TokenRank is significantly correlated to artworks expressing trust, and this is the higher correlation among emotions, and the only emotion that has a significant correlation.

The emotion variables cluster in 3 main groups (we used the hierarchical clustering method):

  1. a positive group containing joy, trust and anticipation emotions;
  2. a negative group containing sadness, disgust, angry and fear emotions;
  3. the surprise emotion, which is in-between. In fact, surprise emotion is more correlated to positive emotions (in particular to anticipation) rather than negative ones.

The Temporal Evolution of the Emotional Spectrum

Finally, we analyse the temporal evolution of the emotions expressed by artworks in the SuperRare gallery during its entire (short) history. We divide the gallery history, from April 2018 to December 2019, into time intervals of 30 days. For each time interval, we compute the relative frequency of words used in the artwork metadata matching the emotions and sentiment poles of the nrc lexicon. We observe the following facts:

  1. after an initial period of emotional instability, characterized by chaotic oscillations of sentiments and emotions, the emotional spectrum chills down, getting more regular and even;
  2. positive sentiments dominate over negative ones; in the last time interval (December 2019), for example, 67% of metadata words are positive and the remaining 33% are negative;
  3. emotions joy, trust and anticipation lead the emotional spectrum. They are followed, at a distance, by fear and sadness. Disgust is the least represented emotion.

Topic modelling

The goal here is to use the textual information about the artwork to automatically extract popular artistic topics for the SuperRare gallery. We first explore the word space of the gallery through a word frequency analysis of the text used in the metadata of artworks. The most popular words are abstract, animation, ai, glitch and generative. Some popular words are clearly related, like blockchain, bitcoin, and ethereum, or ai and gan, or painting, illustration, portrait.

Next, we apply a more involving technique of text mining known as topic modeling. In text mining, we often have collections of documents, such as blog posts or news articles, that we wish to divide into natural groups so that we can understand them separately. Topic modeling is a method for unsupervised classification of such documents, similar to clustering on numeric data, which finds natural groups of items even when we are not sure what we are looking for. Latent Dirichlet allocation (LDA) is a particularly popular method for fitting a topic model. It treats each document as a mixture of topics, and each topic as a mixture of words. This allows documents to “overlap” each other in terms of content, rather than being separated into discrete groups, in a way that mirrors typical use of natural language (Silge and Robinson 2017).

We used artworks, more precisely the metadata of artworks, as documents and we asked the LDA to classify the artworks into 4 topics. The graphics below depicts the 3 terms that are most common within each topic:

  1. the first topic is clearly self-referential, dominated by the words bitcoin, blockchain and ethereum. A representative of this topic is artwork Block 1920000 - The DAO fork, telling the story of the infamous The Decentralized Autonomous Organization (known as The DAO), that was meant to operate like a venture capital fund for the crypto space;
  2. the most common words in the second topic are glitch, collage and 3D. An emblematic artwork of this topic is Tokenized Cloud Sphere Twelve (ATL to YYZ), a computer-aided photo collage which was created from photo series taken on passenger flights across North America;
  3. the third topic features AI-oriented animated artworks. For example, Ugly T-Shirt, generated with Ganbreeder app;
  4. the fourth and last topic discovered is less clear, featuring terms like abstract, painting, digital art as well as color. An archetype of this topic is Water Lilies after Claude Monet, a piece created upon studying “Water Lilies”, an oil on canvas by Claude Monet from 1906.

Conclusion

The crypto art dataset is rich and variegated. It offers to the greedy data scientist a continuous real-time stream of structured (like market transactions) and unstructured (like text and images) data. In this contribution, we focused on the textual data accompanying an artwork, including title, description and tag words. Using text mining techniques, we monitored the sentiment expressed by the artworks exhibited on the SuperRare gallery, the main venue for crypto artists and collectors at the moment. Moreover, we identified artistic topics that are trendy in the gallery. This analysis is informative for all actors in the crypto art system, in particular for artists seeking for style in composition as well as for collectors and art investors looking for rewarding art to acquire. Furthermore, the gallery administrators can benefit from it to better present and highlight the artworks on the gallery website and mobile app. In the future, we are planning to combine this investigation with that on images and animations representing the artworks themselves. Some open questions include: What are the sentiments communicated by the artwork images? Can these artwork images be grouped into meaningful art topics and trends? Does the image analysis match the text investigation? Can we use this information to predict the success of new coming artworks and artists?

References

Aggarwal, Charu C., and Cheng Xiang Zhai. 2012. Mining Text Data. Springer Publishing Company, Incorporated.

Bailey, Jason. 2017. “The Blockchain Art Market Is Here.” https://www.artnome.com/news/2017/12/22/the-blockchain-art-market-is-here.

Finucane, Blake Patricia. 2018. “Creating with Blockchain Technology: The ‘Provably Rare’ Possibilities of Crypto Art.” Master’s thesis, The University of British Columbia; https://open.library.ubc.ca/cIRcle/collections/ubctheses/24/items/1.0370991.

Franceschet, Massimo, and Giovanni Colavizza. 2019. “Art Metrics.” CoRR abs/1907.07758. http://arxiv.org/abs/1907.07758.

Franceschet, Massimo, Giovanni Colavizza, T’ai Smith, Blake Finucane, Martin Lukas Ostachowski, Sergio Scalet, Jonathan Perkins, James Morgan, and Sebástian Hernández. 2020. “Crypto Art: A Decentralized View.” Leonardo. MIT Press. http://arxiv.org/abs/1906.03263.

Lippard, Lucy. 1973. Six Years: The Dematerialization of the Art Object from 1966 to 1972. New York: Praeger.

Silge, Julia, and David Robinson. 2017. Text Mining with R: A Tidy Approach. 1st ed. O’Reilly Media, Inc.

Wickham, Hadley, and Garrett Grolemund. 2017. R for Data Science: Import, Tidy, Transform, Visualize, and Model Data. 1st ed. O’Reilly Media, Inc.

Yaga, Dylan, Peter Mell, Nik Roby, and Karen Scarfone. 2018. “Blockchain Technology Overview.” National Institute of Standards; Technology; https://doi.org/10.6028/NIST.IR.8202.