Blockchain technology popularity, particularly that of the Ethereum platform and smart contracts, keeps increasing over time and is one of the most exciting trends in computer science, economy, and finance. The importance of this technology is also witnessed by the fact that not only developers and researchers are interested in understanding smart contracts, but also generic users are turning their attention to blockchain. Notably, the focus of general users is on Non-Fungible-Tokens (NFT), whose popularity spread over the last three years and has become one of the most popular trends regarding blockchain technology. The users' interest started to apply also on social media (social networks, forums, blogs etc.), arising potentially interesting discussions about NFT and, in general, SCs programming. This work proposes an analysis of smart contract topics on Reddit and Twitter through a topic modeling approach to spot relevant arguments in users' discussions. Starting from a dataset containing subreddits and tweets (which were analysed separately), we built a transformer model to perform our topic modeling. Our results show that Reddit has several exciting topics related to smart contracts programming, such as games, Initial Coin Ofering (ICO), Crowdsales, and complex Decentralized applications building. Twitter has mainly posts related to NFT giveaways and, generally, on NFTs promotion.
The Ethereum platform was oficially launched in 2015, introducing a significant improvement
and exciting innovation: Turing complete smart contracts (SC) programming. Over time, the
Solidity programming language improved remarkably, allowing the development of increasingly
sophisticated programs [
Fungible Tokens. • Bank: programs that implement functions to keep ETH safe (acting as a bank). • Ether Lock Time Constraints (ELTC): programs that work as banks, but these particular contracts can keep ETH safe for a limited time. • Bid: programs that implement Initial Coin Ofering (ICO) and Crowdsales. • Game: programs that implement role-playing games. With these particular SCs, buying items with spending tokens is possible. • Gambling: programs that implement dice games, roulettes, blackjack, and other gambling games. • Wallet: programs that implement wallet-like operations. • Chain Management: contracts that implement functions to help users to interact easily with the blockchain.
• Money Investment: contracts that allow users to invest their money to gain interest tax.
Our findings confirm that the most popular categories are Token and CNFT programs in deployed
SCs of the Ethereum main net [
The Ethereum platform, particularly smart contracts, has caught the interest of researchers since the spread of this technology in 2017. Several studies have been conducted on SCs analysis [12] and classification. Apart from the first-ever taxonomy conceived [ 13], which nowadays is quite limited for reasons bound to the limits of the Ethereum platform in its early years, several studies analyse SCs’ design patterns [14, 15] and other aspects related to classification. For instance, several works try to detect the design pattern of a specific contract by exploring its transactions network [16], and others perform analyses based on the source code and the bytecode [17] [18]. Moreover, other works focus their aim on spotting vulnerabilities inside SCs [19] [20]. Still, in the current literature, there is not much previous work about Ethereum programs topics on social media, except for research discussing SCs issues [21]. Therefore, we decided to extend our research from the programs deployed on the Ethereum blockchain to users’ discussions about SCs on social media. This kind of analysis could help to enrich the understanding of the opinion and the interest of two disjointed groups of users: one includes researchers and developers, the other is composed by generic users more interested in these programs’ economic aspects and on the possibility of easy earnings. Moreover, this work could provide also hints from a statistical point of view, highlighting exciting topics about smart contracts and mostly confirming our findings of the massive SCs categorization, looking for possible matches on most popular programs categories and most popular topics discussed by users.
The data collection process is a crucial step of this work and must be done carefully. Notably, we are studying the topics related to SCs discussions on social networks and forums. Therefore, the provenance of our corpus should come from those two types of platforms. We mainly decided to retrieve our data from Reddit and Twitter 2, respectively, one of the most popular forums and social networks. Potentially, Reddit 3 could be a precious source of data since it is a platform used by most diferent users; indeed, it is possible to find both SCs repositories [22] and subreddits dedicated to SCs programming. On the one hand, from Reddit, we expect to highlight exciting topics related to SCs, such as how to program specific types of contracts, build decentralized applications, information about Non-Fungible Tokens and related games, and other popular arguments related to Ethereum (diferences between standards, ICO, Crodwdsale, and others). On the other hand, on Twitter, we expect to find mainly NFTs giveaways, or at most, we hope to spot sponsored crowdfunding platforms. Indeed, it is highly improbable to find any smart contracts programming topic on Twitter. To build our dataset, we took advantage of Phantom Buster 4. This cloud-based platform allows developers to extract social media, forums, and e-mail data thanks to ’phantoms,’ which are programs focused on a specific text-scraping task. Other use cases of Phantom Buster include data and e-mail enrichment, lead extraction, and social media automation, providing solutions for a wide range of social network analysis tasks. Specifically, we used the Reddit subreddit post extractor and the Twitter hashtag search export phantoms. The first one expects as input the subreddits in which we are interested (in our case, smart contracts), and as output, the phantom returns titles, links, the number of comments, the creation date, and other info for each extracted post. The second one expects an array of hashtags; then, the phantom will pull the tweets containing those hashtags, returning the tweet, the hashtags related to that tweet, the creation date, and other info. However, the Phantom settings must be done carefully since the research and the data extraction will be performed very specifically. Indeed, any specific Phantom Buster 2https://twitter.com/ 3https://www.reddit.com/ 4https://phantombuster.com/ solution includes instructions on what the tool needs as input and what returns as output, including a tutorial. Remarkably, in our case, the Twitter hashtag search export phantom needs to be run with diferent parameters to perform exhaustive research because it will look for tweets with the same array of hashtags given as input. Therefore, if we insert as hashtags ’#smartcontracts#ico#crowdsale#nft#nfts#smartcontractgames#eth#token’, the phantom will look exclusively for tweets containing this list of hashtags, which could be pretty limiting. Indeed, our analysis on Twitter was performed by looking for tweets containing a single hashtag and then running the algorithm cyclically combining several hashtags to ensure variety in our data. Concerning the time interval of the subreddits and tweets, we did not look only for a specific one since we thought that any particular year of the Ethereum lifecycle could be interesting for design patterns development and smart contracts evolution. Instead, we generally looked for every post from the oldest to the newest. At the end of the procedure, we retrieved a sample of subreddits from 2016 to 2022 and tweets posted from 2017 to 2022. The idea is to perform two disjointed analyses: the first consists of training a transformer model, giving as input the corpus of subreddits; during the second one, instead, the input text will be composed of tweets. The reason behind this analysis lies in the dissimilarities of the two selected platforms; indeed, we expect to find various topics between the two platforms and, specifically for Twitter, only a small subset of the arguments found on Reddit. The next section illustrates the research methodology we followed to deliver our analysis.
We already discussed the research methodology’s first step; the data collection process, was explained in the previous section, but further background information on the phantom settings to scrape data are helpful. The phantom’s options for Reddit data extraction are trivial since we only have to give as input the subreddit or subreddits in which we are interested. Indeed, we are interested only in subreddits discussing SCs. The settings for Twitter are pretty diferent since, in this case, the phantom needs a list of hashtags to perform the scraping task. So, since it could be rather constraining to use only ’#smartcontracts’ as a hashtag, we also set the phantom with other terms related to SCs such as #ico, #crowdsale, #nft, #solidity, #solidityprogramming, and others. The idea is that we probably have more chances to find topic variety by adding these key terms to our search. Once we have our data collected, the next step involves choosing the topic modeling technique to manage and analyze our data. One of the first topic modeling choices is usually the Latent Dirichlet Allocation Model [23](LDA). Nonetheless, an LDA analysis depends on the hyperparameters setting and does not consider the text’s contextual nature. So, we opted for BERT [24], a method that takes advantage of transformers [25] and a class based TF-IDF (c-TF-IDF), which is used to calculate words interesting to each topic such to produce easily interpretable topics. Moreover, BERT leverages contextual embeddings and can capture the text’s contextual nature. The BERTopic model will return the number of topics (associated with documents) and include keywords for each class, allowing us to deliver an accurate analysis of the arguments discussed by users on social. First, the corpus should be converted to numerical data, creating embeddings. Notably, we took advantage of ’sentence-transformers,’ which include a set of models optimized to build vector representations considering semantic similarity. The next step consists of dimensionality reduction since, generally, most clustering algorithms do not perform well when dealing with high dimensionality. We delivered dimensionality reduction using the Uniform Manifold Approximation and Projection (UMAP) technique [26], which approximates data samples to a lower dimension, assuming they are evenly distributed over a topological space. We decided to cluster documents to group documents with similar topics and spot them within these clusters taking advantage of the hierarchical density-based clustering algorithm (HDBSCAN) [27], which extends the DBSCAN [28] algorithm but, as the name suggests, it converts the algorithm itself into a hierarchical one. After dimensionality reduction (if needed) and document clustering, we can build our topic representation using BERT and eventually reduce the topics if the number returned by the model is enormous. Once the model produces the topics, they must be interpreted, which should not be dificult since BERT creates easily interpretable topics. The whole process should be applied both for the Reddit corpus and the Twitter one since we want to keep the topics related to each platform disjointed. Once the topics’ probabilities are associated with our documents, we can deliver relevant analysis, such as: • The most and the less relevant topics • The diferences between the discussions of Reddit and Twitter users • Discover if there are any matches between SCs categories in the Ethereum chain and topics discussed by social network users
As a result of the Reddit corpus analysis, we highlighted 14 diferent topics inside the users’ discussions, though many documents were classified as outliers and not associated with any specific category. The colored bar chart on the y-axis represents the number of clusters and the color associated with each group. To compute the similarity between documents, we used the cosine measure. Therefore, we converted the similarity matrix to a distance measure to plot the distance between the clusters represented by the x and y axis of Figure 1. As highlighted in 1, Reddit’s most discussed smart contracts topics concern Non-Fungible-Tokens and tokens. This result confirms the NFT category as the most popular SC trend and our previous work ifndings. It is exciting to observe the variety of topics found inside the corpus. Indeed, except for classic arguments and discussions such as the need for help, advice, and general questions about SCs programming and development, there are pretty exciting topics. Remarkably, games, smart contracts hacking, and audits are the most relevant. The ’Game’ design pattern has been widespread, especially in the early years of the Ethereum lifecycle, when the development of gambling games could be a possible way of gaining ETH. Over time, developers start to build Role-Playing-Games where it is possible to buy particular digital assets by spending tokens. The topics spotted inside the Reddit corpus witness the further evolution of game programs; indeed, the users’ focus is more aimed at those games where it is possible to collect NFTs, which is another proof of their popularity. The audit topic is exciting, too, because, as a term of economics science, an audit should verify the fairness of the financial statement data and the correctness of business procedures. In terms of smart contracts, it is unclear what an audit should provide since, manually checking the subreddits, the only information available is relative to programs delivering an ’audit’. According to the available tools, a SC audit should provide quality code checking and look for any bugs with gas consumption. Therefore, a tool performing an audit
Topic Ethereum
NFT on SCs could be matched with a general smart contract checker and vulnerabilities detector. The last relevant topic concerns smart contract hacking and, in particular, information about how to avoid specific attacks on SCs, explanations of how famous hacks were delivered against precise contracts, and how to prevent the hijack of ETH transfer calls by malicious users. So, notably, the interest is focused on how to avoid SCs hacking and not the contrary.
In terms of smart contracts, Twitter does not explicitly provide posts related to SCs but provides tweets that discuss the Ethereum blockchain in general. Others promote NTF giveaways and platforms that purposefully allow users to collect tokens, but most documents are separate from any specific topic. The main diference concerning Reddit is that the Twitter analysis returned only two clusters and, therefore, much less variety of topics 2. The lack of topics on Twitter concerning Reddit is not a surprise since the last one is more suitable for discussions, while the ifrst is ideal for disclosing information. The reasons behind the few topics lie in the diferent categories of users using the two platforms; indeed, the Twitter platform is operated by a wide range of users with general interests, and apart from a few accounts, such as the Ethereum one, is pretty unusual to find a user promoting or managing information relative to smart contracts. Reddit, instead, has a subreddit dedicated to smart contracts and is more suitable for retrieving the information we are interested in since developers and researchers take advantage of this platform.
The most exciting platform for a number and variety of topics between Reddit and Twitter is
the first one. Notably, we saw how Twitter highlights arguments of interest as smart contracts
auditing, games purposefully aimed to collect Non Fungible Tokens, and also critical topics
such SCs vulnerabilities and hacking. Nevertheless, Twitter confirms the great importance
and widespread of NFTs since many of the documents’ arguments concern token and NFTs
giveaways. Moreover, we found matches with our previous work [
Learning research 3 (2003) 993–1022. [24] M. Grootendorst, Bertopic: Neural topic modeling with a class-based tf-idf procedure, arXiv preprint arXiv:2203.05794 (2022). [25] T. Lin, Y. Wang, X. Liu, X. Qiu, A survey of transformers, arXiv preprint arXiv:2106.04554 (2021). [26] L. McInnes, J. Healy, J. Melville, Umap: Uniform manifold approximation and projection for dimension reduction, arXiv preprint arXiv:1802.03426 (2018). [27] L. McInnes, J. Healy, S. Astels, hdbscan: Hierarchical density based clustering., J. Open
Source Softw. 2 (2017) 205. [28] M. Ester, H.-P. Kriegel, J. Sander, X. Xu, et al., A density-based algorithm for discovering clusters in large spatial databases with noise., in: kdd, volume 96, 1996, pp. 226–231.