Keywords

1 Introduction

The modern art market, especially the online sector, is plagued by a lack of transparency in pricing. While wealthy collectors can pay for highly priced art consultants to suggest works to them and give them a sense of security about prices, these are inaccessible for most consumers. This has created obstacles in the growth of the online art sales environment [28].

Although existing collectors are used to secrecy and non-transparency when it comes to pricing, this is an aspect which clearly doesn’t sit comfortably with new buyers. In this year’s survey, 90% of new buyers and 92% of small spenders said that price transparency was a key consideration when buying art online. [28]

This problem is exacerbated by the uncertain nature of the art market and the challenge of explaining it using traditional economic models [7]. The number of hours of labor, or the cost of the materials is not the primary driver of an artwork’s price [7]. If auctions are studied, that adds another layer of complexity due to the prices being influenced by interactions between bidders [11]. In [25] they wrote when finding that the shifts of the art market were difficult to model:

This may be because the fundamental value of art, as defined before, is hard to grasp. Combined with the impossibility of short-selling, this uncertainty implies a potential role for art buyer sentiment, which could be defined as unjustified optimism (or pessimism) about future resale values.

In [30], the author argues that prices are not just about commodity exchange but are a means of conveying cultural information. Prices are bound to narratives that give information about how the artist, buyer and dealer see themselves [30].

This problem is relevant due to the increasingly large role that online art sales play in the global economy. The online art market was valued in 2017 at 4.22 billion USD [28]. This number, which only represents 29% of total gallery sales, is an increase from 1.5 billion USD in online art sales during 2013 [28].

A consumer making a purchase of a work of art online on a site such as Artfinder.com [3] or Saatchi.com [26] only has a limited amount of information available on the work. Without doing extensive research elsewhere, a consumer would only have information that the artist themselves or someone representing that artist posted. One of the questions that this work explores is how much does that information reflect the price of the artwork. This information is likely restricted to photographs and descriptions of the work, information about the artist, such as their education, and reviews. How can an artist describe themselves and their work to make consumers feel comfortable in making purchases? How can a consumer that is not an expert in the art market use that information to make good purchasing decisions?

This work proposes utilizing text analytics to develop new features to be used to create a model for artworks pricing. Some of these features use the Distributed Paragraph Vector Memory Model, or Paragraph2Vec, to cluster similar artist information. Other features, such as using the artist’s presence on social media and the length of their descriptions and biographies, are also examined. Lastly, features are developed using sentiment analysis on the artist description, biography and artwork title. This work is an extension of [22], which was presented at IJCRS 2019. This work adds new experiments, greater context, and provides further explanation of unexpected results in the original work.

2 Methods

2.1 Dataset Acquisition and Design

In this work, Artfinder.com was scraped to create a dataset [3]. Artfinder possesses a large variety of artists and works across many countries. Artworks posted on Artfinder.com are listed by the artist’s themselves or their representatives rather than being resold. In this work, the focus is solely on the primary art market. The intricacies of the secondary art market are beyond the scope of this project. The data was extracted using a webscraper written in Python. Artfinder makes use of Javascript, so Apache Selenium [27] was used and Beautiful Soup [6] was used for parsing the relevant text. The resulting dataset contains approximately 160,000 individual works from over 2,000 artists from more than 60 countries.

The listed prices for works in this dataset span from less than 20 USD up to 1,000,000 USD. This distribution is not even and is heavily skewed toward lower prices. 0.26% of the works are valued at more than 10,000 USD and 85.66% are valued at 1000 USD or less. See Fig. 1 for more detail. Works valued at greater than 10,000 USD are omitted in Fig. 1 due to scale. Collectors are becoming more confident in large online purchases, but as of 2018 only 25% of buyers surveyed pay more than 5,000 USD per purchase online [28]. For this work, the prices have been reduced to a set of discrete intervals. The prices were discretized by examining the dataset for areas with very few works. Cuts were then placed at these sparse points. However, the number of intervals was very large and their range was very small as prices cluster around the 50 and 100 values. Small sets of intervals were combined to create the following discretization: (0−105), (105−205), (205−405), (405−605), (605−810), (810−1030), (1030−1445), (1445−1825), (1825−2455), (2455−3855), (3855−5000), (5000−10,000), (>10000). The represented prices are all in USD, but this is often converted from another currency.

A number of useful features were extracted from the artist’s about pages. All artists on Artfinder.com have pages where they can post their biographical information, links to their social media and personal websites, and information about their education, awards or events. In addition to this information, many artists have reviews from past sales visible on their profiles. These reviews were used to construct features, however, they are of limited value when assessing an artist’s work. Reviews can only be posted by past customers, which means that a significant number of artists do not have any associated reviews. In this dataset, approximately 50% of the artist’s have one or more reviews. Reviews are on a 5 point star scale and generally include comments, however these are unstructured and may or may not specify which work was purchased. Additionally, of the visible reviews the average score is 4.899 out of 5. These factors diminish the predictive power of review score alone.

Other features used were extracted from the artwork pages themselves. Each artwork page has at least one photograph of the work as well as information about that work. Artfinder.com allows customers to search by medium or subject or even size. Artists can tag their works with this information in order to make it easier for potential buyers to find. In most cases, artist’s provide short text descriptions of their works. These descriptions will be the primary focus of Sect. 2.2.

To use for comparisons, a set of basic features was selected. The attributes focused on in this work are extensions of this set of features. All of these features are discrete or were discretized. The features developed are influenced by the features used in [20].

  • artistID—A unique identifier for an artist. It may or may not be their legal name and serves as their username on the site.

  • artistCountry—The artist’s current country of residence as listed on their profile

  • artwork_height—The height of the artwork in inches.

  • artwork_width—The width of the artwork in inches.

  • authentication—Artist provided method of authenticating the work.

  • percent_five_stars—The percentage of five star reviews out of the total number of reviews.

  • percent_four_stars—The percentage of four star reviews out of the total number of reviews.

  • percent_three_stars—The percentage of three star reviews out of the total number of reviews.

  • percent_two_stars—The percentage of two star reviews out of the total number of reviews.

  • percent_one_star—The percentage of one star reviews out of the total number of reviews.

  • medium—Artist provided medium of the artwork.

  • style—Artist provided style of the artwork.

  • subject—Artist provided subject of the artwork.

Fig. 1
figure 1

Price distribution

Full size image

2.2 Product Description

On Artfinder, as with many other sales sites, artists write descriptions of their products. Product descriptions can be used to capture customer interest, but they are more factual in nature than advertising [29]. In the Shotfarm Product Information Report, they found that,

Ninety-five percent of those surveyed say product information is important when making a purchase decision, with nearly four in five indicating that it is very important. [1]

However, there is little certainty about what is considered a complete or informative description.

The ideal length of a description and its impact on sales is an open question. In [23], the researchers determined that between 40 to 55 words should be focused on product description for eBay sellers. They also found that the use of words denoting uncertainty, such as “probably” or jargon harmed sales [23]. However, that length for descriptions is far from universally agreed upon. In [19], retailers are advised to keep product descriptions between 350 and 400 words.

As can be seen in Table 3, the number of words used in the document does have some utility as a predictive feature. However, the artwork descriptions are dominated by simple functional words. In Figs. 2 and 3, word clouds were generated from 25,000 randomly selected documents. Figure 2 uses documents that reference artworks priced at less than 1000 USD and Fig. 3 uses documents that reference artworks priced at more than 1000 USD. It is notable how similar the word clouds are. The most frequently used words reference shipping and sizing rather than describing the artwork. Factual terms such as ‘canvas’, ‘x’ (which is used to indicate the dimensions of the work), ‘original’, ‘signed’, and ‘shipping’ are heavily emphasized. More emotionally evocative words, such as ‘beautiful’, ‘love’, ‘inspired’ and ‘feel’ are much less frequent. Interestingly, the word ‘please’ appears quite frequently. This lack of emphasis on description may be due to the domination of the artwork image on the product page.

Fig. 2
figure 2

Frequently appearing words in descriptions of works priced at less than 1000 USD

Full size image
Fig. 3
figure 3

Frequently appearing words in descriptions of works priced at greater than 1000 USD

Full size image

2.3 Social Media and Sales

Social media has a strong role in driving sales of contemporary art. In [28], 32% of the art buyers surveyed in 2018 said ‘social media had an increasing impact on their decision to buy art’, which is an increase from 2017. They also found that among art buyers under 35, Instagram was used by 79% of those surveyed to seek out artists and 82% use it to follow artists [28]. Instagram is the most popular platform according to Hiscox [28].

What differentiates Instagram from other social media platforms is the pronounced lack of words; it is overwhelmingly visual and experiential. While on one level, words are essential to describing, explaining and understanding art, on another level, tons of people out there are not big fans of how they intrude on the pure art experience. [4]

In this work, the social media section of artist’s Artfinder about pages was collected. The words ‘Facebook’, ‘Twitter’, and ‘Instagram’ were searched for in the collected text and used as a Boolean attribute. In this dataset, approximately 85% of the artists have Facebook pages listed, 42% have links to Twitter and 54% have posted Instagram links.

2.4 Determining Text Similarity Using Vectors

This work explores the application of clustering text using vector representations of that text. These representations are created by using the “Distributed Memory Model of Paragraph Vectors (PV-DM)” which is also called Paragraph Vector or Paragraph2Vec [8, 14]. Paragraph2Vec is an unsupervised model that creates vector representations with a fixed length from unstructured pieces of text [14]. In spite of the name, Paragraph Vector can handle documents of any length [14]. The concept for Paragraph Vector was derived from ‘Word2Vec’ [8].

Word2Vec represents words as vectors [8, 17]. It makes use of the Skip-Gram model which was created in [16]. This model uses a neural network to predict a word given other words in a sequence of words [8, 16, 17]. “The training objective of the Skip-gram model is to find word representations that are useful for predicting the surrounding words in a sentence or a document.” [17] This model was then used in [17] to create Word2Vec. It can be used to find words that are used in similar contexts and is more efficient than other methods developed at the time [17]. Another interesting feature is that vectors can be combined mathematically with simple vector addition to find words near the sum of two terms [17].

Paragraph Vector expands the Word2Vec so that vector representations can be created for sentences or lengthy documents [14]. Each paragraph is represented as a unique vector in a matrix which is concatenated with the vectors for each word in that matrix [14]. The identifier for the paragraph “remembers” the subject of the document but otherwise functions as another word [14].

Paragraph Vector and Word2Vec have been applied to a number of research problems. Word2Vec combined with decision trees was used in [31]. In [5], this method is extended into Item2Vec. Product descriptions are categorized using Paragraph Vector in [15], and it was used for determining the similarity of pieces of text in [10].

In this work, to create the vectors, the Gensim [24] implementation of Paragraph Vector, which is termed Doc2Vec, was trained on text provided by the artist. The vectors were calculated for the text of the artist’s biographies, the text of the artist provided description of the artwork, the title of the artwork, the artist provided description of their education, the artist provided description of awards received, and the artist provided description of events they have held. The biography, awards section, education section and events section were found on the artist’s about page. Gensim includes pre-built packages for basic text preprocessing which were used on the text before it was used in the creation of the model. Then, the resulting model was given the original text for each piece of training text. This created a set of 100 term vectors. They were then clustered using K-Means, implemented with Python’s Sci-Kit Learn Library [21].

2.5 Sentiment Analysis

Sentiment analysis, along with opinion mining, is an important tool for classifying textual input [2]. As said in [18],

As regards sentiment analysis (also called opinion mining, review mining or appraisal extraction, attitude analysis), in an attempt to enclose the whole body of work that has been carried out in the field, we can define it as the task of detecting, extracting and classifying opinions, sentiments and attitudes concerning different topics, as expressed in textual input.

A common topic in the field of sentiment analysis is the attempt to classify a piece of text as either positively inclined, negatively inclined or in some cases neutral [18]. This often relies on the use of an ‘opinion lexicon’, which is a list of words or phrases that are useful when assessing the sentiment of a piece of text [2]. In this work, the opinion lexicon developed in [13], known as the ‘Valence Aware Dictionary for sEntiment Reasoning’ or VADER was used to classify text as positive, negative or neutral.

VADER uses a combination of a lexicon and a set of 5 main rules [13]. The researchers had over 9,000 candidate terms, including acronyms and emojis, rated by human raters on a scale from –4, indicating extreme negativity, to +4 which indicates extreme positivity [13]. Then, they developed a set of five rules that take into account factors such as negation, intensifiers and capitalization [13].

Another set of features was created by performing basic sentiment analysis on the artwork descriptions, artist biographies and artwork titles. Three numeric scores, positive, negative and neutral, were calculated for each piece of text. All scores for a piece of text sum to one. Score of one means that this piece of text solely represents that sentiment. VADER was used as part of the Natural Language Toolkit [9, 13].

Figure 4 is the level of positive sentiment in the descriptions of 150,000 pieces of text, while Figs. 5 and 6 represent the level of negative sentiment and neutral sentiment respectively. The description scores are often only slightly positive, with almost no negativity. The title displays a very similar pattern, with almost total neutrality with only minor instances of sentiment. The biography scores, which can be seen in Figs. 7, 8 and 9, are quite similar, but there are more positive sentiments and a small number of negative statements present.

Fig. 4
figure 4

Positive description score

Full size image
Fig. 5
figure 5

Negative description score

Full size image
Fig. 6
figure 6

Neutral description score

Full size image
Fig. 7
figure 7

Positive biography score

Full size image
Fig. 8
figure 8

Negative biography score

Full size image
Fig. 9
figure 9

Neutral description score

Full size image

3 Results

Randomly selected subsets of 150,000 works were used to obtain the following set of experimental results. All experiments were preformed using Orange [12], and 10 fold cross validation was used to check the resulting models.

3.1 Base Features

The results of testing the base set of features without extension are shown in Table 1. Experiments were conducted using the base set, see Sect. 2.1, with k-Nearest Neighbors, with a k of 5, Logistic Regression, Support Vector Machines and Random Forest, with 100 trees, as classifiers. The random forest classifier had the best results so it was used for all further tests.

Table 1 Results with base features
Full size table

3.2 Social Media Presence

The features constructed based on the presence of the artist’s social media links in the social media section of the profile do not have an apparent impact on the accuracy of the classifier. The complete results can be found in Table 2.

Table 2 Results with social media
Full size table

These results were unexpected, as social media is widely assumed to be a driver of sales. However, this can be easily understood when examining Figs. 10, 11 and 12. Figure 10 represents the relative frequency of listed Facebook pages at the different prices. Facebook is widely used across all price levels, which makes it of limited value as a predictive feature.

Fig. 10
figure 10

Relative frequency of linked Facebook pages

Full size image
Fig. 11
figure 11

Relative frequency of linked Twitter accounts

Full size image
Fig. 12
figure 12

Relative frequency of linked Instagram accounts

Full size image
Table 3 Results with word counts
Full size table

3.3 Word Count Results

The number of words used in the biography and in the description of artworks appear to influence the price of a work of art. The results found when testing this impact are shown in Table 3, and the distribution of word counts across price levels are shown in Figs. 13, 14 and 15. As can be seen from the table, the word count of the description has a more notable impact on the accuracy of the classifier than the word count of either the title or the artist’s biography. However, the word count of the biography combined with the word count of the description has a slight increase in the accuracy of the classifier. Adding the word count of the title to these two features does not have an apparent impact on the accuracy of the classifier (Relative frequencies shown in Figs. 13, 14 and 15).

Fig. 13
figure 13

Relative frequency of biography word counts

Full size image
Fig. 14
figure 14

Relative frequency of description word counts

Full size image
Fig. 15
figure 15

Relative frequency of title word counts

Full size image

3.4 Document Vector Clusters

In Tables 4, 5, and 6 the results of testing the quality of the classifier extracted from the dataset using the base feature set extended by the document vector based clusters are shown. In Table 4, each piece of text is placed in one of 10 clusters, in Table 5, 25 clusters are used, and in Table 6, 50 clusters are used. Extending the feature set with the awards cluster, biography cluster, education cluster, or events cluster in some combination has the highest positive impact. There are small changes in the accuracy of the classifier as the number of clusters changes, but it is not apparent how notable these changes are to the accuracy of the classifier.

Table 4 Results with 10 clusters
Full size table
Table 5 Results with 25 clusters
Full size table
Table 6 Results with 50 clusters
Full size table

3.5 Results with Sentiment

The sentiment of the text, in particular the sentiment of the artwork description has an interesting impact on the accuracy of the classifier. The combination of the scores for positive sentiment for all three examined features had a comparatively strong impact on the classifier (Table 7).

Table 7 Results with sentiment features
Full size table

3.6 Combined Features

Table 8 shows the results of extending the dataset with different sets of combined features. The word count features have a strong impact on the accuracy of the classifier when combined with the social media features. These features also combine with the text clustering features for an overall positive effect. In keeping with the results when testing the cluster features alone, the largest gain appeared when the awards, biography, education or events clusters were used. The sentiment features also have a visible positive effect when combined with the other features, and have a positive impact on the accuracy of the classifier when added to the others.

Table 8 Results with combined features
Full size table

4 Conclusions

This work addresses how the text that an artist chooses to share with the public reflects artwork prices. What an artist chooses to share, how much they share and their tone all reflect aspects of the artist that appear in their prices.

The presence or absent of features derived based on the artist’s presence on social media has no notable impact on the accuracy of the classifier. This may be due to the relative ubiquity of these pages. Facebook pages are present, at a roughly equal proportion, across all price levels. This makes them less useful as a price determinant. Other features, such as the follower count, number of likes per post, or similar metrics may have greater value as predictive features.

The word count features, on the other hand, do show interesting variations across different price levels. Just a cursory examination of the word count frequency figures shows that higher priced works have a tendency to have short descriptions and be associated with long biographies. This make intuitive sense as the length of the biography may be a reflection on the achievements of the artist. However, the associations between the length of biographies and descriptions and prices is not fully explained and may be explored in a later work.

Clustering text using vectors provides an unsupervised method for grouping similar content. The most notable positive impact on the classifier came from the clusters for the education, awards, events and biography sections. There are approximately 2000 of each of these, so it may be 10, 25 or 50 clusters which is too few to adequately describe the descriptions and titles leading to meaningless clusters.

Taking the sentiment of the text had a visible increase in the accuracy of the classifier. In that case, the level of positive sentiment in the description had a very strong impact, as did all three description features combined. The positivity and neutrality of all three features together was also notable. Measuring the sentiment of the title has almost no impact. This may be due to the relative lack of non-neutral titles.

Sentiment, even in factual text, can be a reflection of underlying traits that influence prices. Text clustering using vectors and sentiment analysis can be combined to improve the accuracy of classifiers for predicting the prices of contemporary artworks.