Emotions are the drivers of almost all human actions. Advertising companies have recognised this for ages but up till now have had no reliable way to measure emotions in text. To fill this need, AI Mining provides emotional analysis of text. You can ask it any word or phrase and receive an emotional breakdown. For example “great web page” is 55% love, 21% laughter, 16% surprise, 3% sad and 5% angry. The name of this product is Emowatch.
Below are the primary features and uses of our Emowatch analysis:
- Emotional analysis – love 💟, laughter 😝, surprise 😵, sadness 😢 and anger 😡
- Uses machine learning
- Can recognise 3,000,000 words in sequences 50 words long
- Updated regularly
- All updates are tested against actual user reactions
- Emotions from reader’s point of view
- Marketing companies – use our analysis to talk to your customers emotionally
- Financial companies – predict how users will respond to breaking news about companies
- Social analytics companies – expand your analysis from sentiment to full emotional analysis
- NLP and ML companies – improve your models by extracting emotional features from text data
- Everyone – write better headlines and ads, improve your digital communications
Emotional Analysis vs Sentiment Analysis
The emotions we analyse are the same as those suggested and used by Facebook: love, haha, wow, sad and angry. This is different from the six or seven basic emotions proposed by psychologists and from those used by other providers such as IBM Alchemy Tone Analyzer (which both include fear and disgust). We feel that the Facebook emotions are more balanced between positive and negative, and using them aligns us much more closely with modern trends in social media.
Emotional analysis can be seen as an extension of the fast growing field of sentiment analysis. Sentiment analysis tries to determine if a piece of text is positive or negative. This is useful, and is already a core part of many brand analytics products. But emotional analysis can provide so much more.
For example, our own research has shown that a news headline which causes readers to feel haha is more likely to be shared than one which produces just love. In customer service, prioritising angry over sad customers can help avoid cancellations. Sentiment analysis would not be able to provide this differentiation.
Machine Learning vs Lexical Analysis
Many other companies in this field use lexical analysis to establish the sentiment of a document, such as Linguistic Inquiry and Word Count. This technique basically counts how many positive and negative words there are in a piece of text, does various calculations and produces a score. The lexicon (subset of a dictionary) used for this approach contains just a few thousand keywords.
Compare this to our approach using machine learning. We have trained a multi-layer recurrent neural network which accepts sequences of words as input and outputs 5 percentage scores for our 5 emotions. Our network has a working vocabulary of approximately 3 million individual words and multi-word phrases, and it operates on sequences of up to 50 words. This gives it the capacity to differentiate between millions and millions of separate word sequences.
Our network has essentially learned for itself which words and combinations of words lead to different emotions. In lexical analysis, the programmer or psychologist must first say that the word e.g. “adore” is 75% love. In machine learning, the machine learns this percentage on its own, by being presented with thousands or millions of example uses of the word “adore”. This means it can learn any word or sequence, not just a few thousand. The programmer’s job changes from labelling words to finding data, setting network parameters, training a network and testing it. We have spent several months doing exactly that. Emowatch can out-perform lexical analysis because it finds subtle patterns that evade humans and dictionaries.
Updated vs Static
Our Emowatch translator is regularly updated. Our data comes from major web news source and social networks. We download new data every few minutes and retrain our underlying neural network every 1 to 7 days. Compare this to the LIWC lexicon which was last updated in 2015, and before that in 2007.
Provable vs Not Provable
After every update we retest our neural network against actual user reactions to news. So if a model is trained Tuesday morning, we wait until Tuesday evening and test it on Tuesday’s afternoon’s news. We can then statistically prove that our model is behaving as it should. See our technical demo for how we analyse this.
Reader vs Writer
Emowatch is subtly different from most other sentiment and emotional analysis products, because the emotions are from the reader’s point of view. In the academic literature, the reader of a piece of text is referred to as the target. Most sentiment analysis products operate from the writer’s point of view, also known as the holder or source. It is a small but important difference. For example, when someone reads a newspaper headline which says “Yawning Almost Killed Man” (from the Brighton Evening Argus), they will probably feel like laughing. In this case, the emotion of the man in question or the journalist who wrote the headline may well be completely different.
This approach has many useful real-world applications. For example, in May 2013, the massive American retailer JCPenney produced a kettle which bore a resemblance to Adolf Hitler. Lexical analysis alone would produce a negative score for this headline (try Stanford’s sentiment analysis demo) based mainly on the word “Hitler”. Emowatch on the other hand estimates 72% laughter and 14% surprise, a good recipe for sharing. And this turned out to be the case. The kettle went viral, sold out in minutes and had no impact on JCPenney’s stock price which was $17.26 at the beginning of the month and $17.58 at the end. This shows the importance of reader’s emotions as opposed to the writer. When composing headlines and ads, it is the reader’s emotion which must be considered, not the writer. Which is exactly what Emowatch does.
We have created two use cases of Emowatch. The first estimates how people are responding to actual news headlines. This brand analysis use case is inspired by Thomson Reuter’s Eikon financial analysis tool, which provides sentiment analysis of news articles as they are published, allowing investors and traders to track businesses. We use Emowatch to do the same thing with emotions. In other words, we predict how readers will respond to a press release.
Our second demo is to assist ad and headline writers. Readers are more likely to share things which make them laugh. They comment on things which surprise them or make them angry. Furthermore Improving the click through rate of an ad campaign by a few points can have a significant effect on profitability. Put those two things together to write better headlines and ads.