Introducing AI Sentiment Analysis in the Sonar Ecosystem

Here at Sonar, we’re constantly innovating and engineering new solutions to make crypto investing more accessible and safe for everyone, regardless of their experience or expertise. We’re proud to be releasing our newest AI feature, Sentiment Analysis & Analytics.

Sonar platform users will now be able to study both major blockchains such as Ethereum and smaller tokens such as PING, with insights into both AI market sentiment, follower/media activity, and much more.

We scrape, clean, and analyze a plethora of data from Reddit, YouTube, Twitter, and Telegram for an unprecedented number of tokens, providing one of the most cutting-edge systems in the world to help you stay on top of the market’s feelings.

What is Sentiment Analysis? Why is it useful?

Sentiment analysis is a subfield of Natural Language Processing (NLP), the study of how computers can learn to interact with and understand human language. Sentiment analysis aims to model the polarity of utterances, a task very important for problems such as the analysis of customer opinions, where one would like to understand what makes clients happy with their purchases, or the monitoring of online forums to gauge public perception of presidential candidates. Depending on how these models are architectured, they may produce boolean (positive/negative), or more specific outputs (happy, sad, jealous, etc.)

Naturally, sentiment analysis lends itself well to understanding financial markets insofar as they can monitor how the public feels about specific stocks, companies, and even cryptocurrencies! Knowing whether the community is positive or negative about a specific token can help users make more educated guesses on both short-term and long-term financial outcomes.

Analyzing sentiment scores of tokens alongside price charts can assist investors to determine why exactly a rise or fall in price is taking place by helping differentiate overall market conditions from token-specific good or bad news.

Sudden drops or spikes in sentiment may signal something major that’s about to happen to the price of a token, a pattern that can sometimes be observed when looking back on historical price/sentiment data.

How does Sentiment Analysis Work?

There are two main categories of algorithms that are used in sentiment analysis, namely rule-based and Machine Learning algorithms.

Rule-based algorithms are explicit, user-defined functions that are created ahead of time. A couple examples of problem domains and corresponding rules may be:

● Movie reviews | Count the number of words in a pre-defined bin (“great”, “fantastic”, etc.) that appear next to actor names. If this number is great than X, mark this review as positive. Else, move on to the next rule.

● Cryptocurrency markets | Check if the word “bullish” is used next to the name of a token in a pre-defined list. If so, then check that phrases “just kidding” or “sike” are not used. If these conditions are both met, mark this comment as positive. Else, mark it as negative.

This method can do the job in some circumstances, particularly those in which the problem domain is very limited. However, it is very easy to see the drawbacks.

There are a limitless number of possible ways to express a thought, and it’s impossible to separate and create rules for every way imaginable. Even within one small connotation of a sentence, there are countless ways to word something.

Machines, though, may find patterns in data that humans can’t, which leads us to the method pervasive throughout the industry today:

Machine Learning algorithms allow computers to perform tasks and make predictions without being explicitly programmed to. These methods typically consist of the following steps (a large simplification):

1. Data collection — Extracting a large amount of data that is representative of a wide range of opinions, not too biased one way or the other
2. Data cleaning — Ensuring that aspects of the data unnecessary to the algorithm are cleaned up (URLs, misspellings, erroneous labels)
3. Algorithm selection — Depending on the problem and data, different algorithms have pros and cons (Naive Bayes, Support Vector Machines, Linear Discriminant Analysis, Neural Networks, etc.)
4. Model evaluation and deployment — Extrapolating the accuracy of the model in the real world and deploying it to customers.

Through finding patterns in the data and learning representations of meaning as a whole instead of explicit hard-coded rules, these algorithms are able to predict the sentiment of utterances, impervious to aspects such as word order, use of synonyms, etc.

What makes our Sentiment Analysis cutting-edge?

We make use of the latest and greatest ML models, called transformers. No, we’re not talking about Optimus Prime or Bumblebee, but rather the latest evolution of artificial neural networks: machine learning algorithms that mimic how the human brain works.

No. Not this kind of Transformer.

Transformers are the state-of-the-art in industry technology that power the Artificial Intelligence behind systems such as AlphaGo and GPT-3, even powering common-sense reasoning for Google Search results! We utilize transformer-based models such as BERT (Bidirectional Encoder Representations from Transformers) to output sentiment predictions of utterances regardless of word choice and length.

We mitigate word choice bias by representing them in our models as word vectors in 300-dimensional space. Length bias is mitigated via bidirectional encodings.

Extra care is taken to ensure data quality before performing analysis — it is paramount to not have poor quality data (as the saying goes in AI, “garbage in, garbage out”). Other platforms that also have “sentiment analysis” are often using extremely poor quality data (some explicitly show their data points), useless for any accurate predictions.

All of our outputs are intelligently normalized and take into account sample sizes. Possible FUDing attacks are detected and filtered out to keep the sentiment scores useful.

More importantly, as opposed to almost all other services on the market, we offer data not just for major blockchains on Twitter, but also for any token in the world with an active Telegram server. We are the only service to offer analysis of this many tokens with this quality of data and analysis.

So what does all this mean for you, the user?

We’re introducing sentiment analysis with all its power to the Sonar Platform in the form of a brand new application called Sentiment Lab as well as by activating the long-awaited Intelligence tab and sentiment chart within Token Studio.

Sentiment Lab

A new application under the Sonar Platform umbrella that acts as a global market overview of sentiment trends and their influences.

Sentiment Lab features all the top coins and tokens and their sentiment performance as well as dedicated tabs for the most positive and most negative projects on the market.

There’s a section on the tokens and coins experiencing the largest changes in mood, aptly named “Mood Swingers” and a brief written market overview at the top of the page.

Next to every project and its sentiment chart for the past week is a list of news articles shedding more light on what is driving the project’s current growth or decline in sentiment.

New features in Token Studio

First, each token page is enriched with the same sentiment chart you find in sentiment labs, in an expanded form.

Then, we added a series of new indicators that in addition to showing you the current sentiment for both the community and admins of a specific project, also display valuable social metrics detailing engagement and growth across major social media channels. Used together, these tools begin to provide truly indispensable insight into any project’s progress.

• Community Mood — The overall market’s current attitude toward the specific project being looked at.
• Admins Mood — (Something completely unique to Sonar) The measure of the current attitude a project’s leaders have within their community and the tone of their communication — achieved by analyzing community chatrooms and forums!

Thank you for reading!

We hope you enjoy these new features — there’s much more to come, with AI features that are unprecedented in the Cryptocurrency Analysis space. To learn more about Artificial Intelligence, keep an eye out for educational articles from us, or check out some recommended resources from our AI team:

1) https://work.caltech.edu/telecourse

2) http://web.stanford.edu/class/cs224n/

3) https://www.statlearning.com/

4) https://ocw.mit.edu/courses/mathematics/18-06-linear-algebra-spring-2010/ (Image source: https://miro.medium.com/max/876/0*ViwaI3Vvbnd-CJSQ.png)