Sentiment analysis with Umigon 2.0

Umigon is the name of the function performing sentiment analysis, which I initially developed in 2012.

Umigon just got an upgrade that I will develop in this blog post, but let’s first describe quickly what it does:

  • sentiment analysis with three “classes” of sentiments: positive, neutral, negative
  • works on texts in English or French
  • focus on sentiment, not negative factuals, which is important
  • for social media: short texts such as tweets, instagram comments, facebook posts. Handles hashtags, emojis and emoticons, slang and mispellings: natural language in all its forms.
  • created for non coders: click and point, just send a text and it returns the result
  • free without restrictions, and available here: https://nocodefunctions.com/umigon/sentiment_analysis_tool.html
  • evaluated as best in class, compared to 23 other solutions (commercial or not): https://epjdatascience.springeropen.com/articles/10.1140/epjds/s13688-016-0085-1

Umigon: a new heuristics engine with MVEL

As described in this paper, Umigon’s performance is based on using lists of terms which are markers of sentiment. Obviously though, terms alone would not suffice, as in this example:

I like this brand [positive sentiment]

Reading a book is really like travelling but without taking a trip [neutral sentiment]

I really don’t like this taste of ice cream [negative sentiment]

How can we distinguish between each case?

The solution consists in taking the context into account. How is “like” used in the sentence? For each word, Umigon can query and take into account whether:

  • the preceding words in the sentence are negative, positive, or some specific words
  • the following words in the sentence are negative, positive, or some specific words
  • the word itself is written in caps, or appears at the beginning or end of the sentence
  • etc.

These rules can of course be combined, which leads to some very precise handling of the semantics of a text.

Until now, expressing the combination of these rules was done with a simple syntax I had developed (see pages 3 and 4 of this paper).

However, it was crude in the sense that it could not accomodate cases where the logic was a bit involved, such as:

“I would like this word to be a positive marker of sentiment IF it is preceded by this term AND NOT followed by this term. ELSE, consider it neutral except IF etc…”

First I thought I had to develop such a thing myself, but actually a Java project already provides it out of the box. It is called MVEL, check it out!

MVEL can handle all the boolean logic I need and as a bonus, the conditions are expressed in a language which is very natural, so that I can re-read myself easily. One of the most complex heuristics for a term looks like:

if(A && B && C){12} else if (C == false) {11}

Where A, B and C are true / false values that result from checks on the context of the term, and “11” and “12” are the codes for “positive marker” and “negative” marker.

Bug fixes: minor but big impact

Umigon has been put back online in the Spring of 2021, and it will be maintained for the long term. It is part of a (free) web app which makes it easy for users to test their texts and report any misclassifications with just one click. I have received about 30 of such reports, which were as many tests on edge cases which I had missed. Thanks to them, I could spot and fix these issues:

  • negations were not taken into account if they were one word apart from the term under examination. I don't really like was considered neutral because don't was not immediately before like. FIXED
  • emojis following each other 🥰⛷️ were considered as one big emoji, so they escaped detection. FIXED
  • the list of terms that I call “amplifiers” (seriously, ultra, ridiculously, freaking, etc.) did not load properly so it was never put in action in the heuristics where it played a role 🤦‍♂️. FIXED
  • apostrophs were breaking term detection, which is annoying given that they appear in contexts that carry rich semantics (can't, I'm … and same in French!). FIXED
  • exagerations like looooove were not properly handled, hence if the term was carrying a mark of sentiment (love) it was ignored. FIXED

Addition of new terms

Language is constantly evolving, especially the type of oral expressions and slang which finds a written expression on social media. Umigon will never capture all of it, but it covers a pretty large scope (try it!). With new user reports, no new addition for slang but “classic” terms which were not included yet in the “negative” category - always the richest 😀:

  • grotesque
  • infuriating
  • heart breaking
  • mortifying
  • on what planet
  • how can

All these changes have been integrated and you should see a great improvement in the results. By the way, if you conduct benchmarks on sentiment analysis, I would love to hear about it!


Try Umigon here, it is free and without any registration: https://nocodefunctions.com/umigon/sentiment_analysis_tool.html

And discover all the other functions: https://nocodefunctions.com/

 Date: September 20, 2021

Previous
⏪ Pointwise mutual information and tf idf, when to use them

Next
Just released - Converter from gexf (used by Gephi) to VOSviewer json file format ⏩