The Future Of Sentiment Analysis Is Here

94.5% Accuracy, 100+ Languages

 

Sentiment analysis – understanding how people feel about different topics by analyzing articles, conversations, social media posts and other forms of textual content – is helping individuals, organizations and even governments make important decisions. But progress so far is slow, and the accuracy and value of sentiment analysis to date is distinctly underwhelming.

However, by taking a different approach to sentiment analysis, we at HIPSTO have achieved an exciting breakthrough that is set to be a gamechanger. With 94% accuracy outperforming every other commercially available sentiment analysis engine and platform, consistent coverage of over 100 languages and true Natural Language Understanding, we believe the future of sentiment analysis has arrived.

 

 

In this article, we’re going to take a deeper dive into what sentiment analysis is and why it is important, how HIPSTO’s approach differs from legacy methodologies and some of the unique advantages of our offering.

Please note that we’ve written this for a non-technical audience, so if you already know your RNNs and your LSTMs, please bear with us!

 

What is sentiment analysis?

 

Sentiment analysis is an increasingly important approach that analyzes large amounts of complex, structured and unstructured text from multiple sources to understand people’s opinions about a variety of topics. It identifies and extracts human sentiment across a wide range of content types including news articles, web pages, discussions, conversations, product reviews, blogs, social media updates, reports, survey responses, voice and video transcripts and more. Basically, sentiment analysis can be run over any text.

A significant proportion of sentiment analysis engines are still based on statistical methods, although more advanced approaches utilize Natural Language Processing (NLP), textual analysis and other AI techniques to analyze the subjective text of whatever content and sources you choose to feed into the sentiment analysis engine. Most popular sentiment analysis offerings will generally classify the sentiment of an element of text as either positive, negative or neutral, and then aggregate these scores to give you an overall sentiment, such as at an article level.

But sentiment analysis is not easy; it’s very rarely a case of a person simply saying something is either “good” or “bad”. Consider a review in a newspaper about a restaurant, a customer review on a site or a discussion thread about the restaurant. In order to understand the sentiment of a piece of text with all the nuances of language, usage of slang, subtle sarcasm, the multiple opinions expressed just within one review or discussion and other contexts, you need human judgement and comprehension. Sentiment analysis requires a very sophisticated approach to replicate that level of understanding.

 

Why is sentiment analysis important?

 

Opinions are important as they impact both everything we do and the decisions we make. From brands wanting to make better products for their customers, to ascertaining sentiment about the performance of stocks to enable the right trading decisions, to governments wishing to improve the lives of citizens, the successful ability to track sentiment gives us enormous power to make the key decisions we need to make.

Given the value that accurate sentiment analysis can bring, it’s no surprise that an entire industry of products and services has grown around analyzing the sentiment derived from different collections of data and content.

But there’s a big problem. Most sentiment analysis delivered by established products or frameworks, even by major providers like Amazon Comprehend and Google Cloud, is simply not as good as it should be. It does not result in the value we expect, with outputs having caveats and disclaimers. Trend analysis doesn’t reflect the true complexity of sentiment that occurs across content, and it’s certainly not actionable. The levels of accuracy are, to be blunt, distinctly mediocre. The whole industry is characterized by a series of ifs and buts.

At HIPSTO, we’ve developed an innovative approach that delivers the kind of ground-breaking sentiment analysis that many in the industry have been waiting for. By using a combination of the right neural network architecture and additional proprietary mathematical weightings, we’ve been able to analyze content for sentiment that is much closer to true human understanding.

 

What are the use cases for advanced sentiment analysis?

 

You can start to see why strong sentiment analysis is so important when you consider the wide set of use cases for it, including:

• Brand management: Understanding how the public feels about your brand across web pages, social media posts and conversations, enhancing and protecting your buyers’ view of your brand and beyond…
• Marketing and PR: Gaining a deeper understanding of how customers feel about different topics and products in order to develop deep marketing intelligence.
• E-commerce: Understanding product reviews and customer sentiment to both inform buying decisions and improve products.
• Financial services: Mining text from news, blogs, audio feeds, regulatory filings and even conversations on Reddit to understand the sentiment on the performance of stocks to support trading and investment decisions.
• Government: Understanding public opinion in relation to different topics, issues and initiatives to improve decisions that impact the lives of citizens.
• Projects and initiatives: Understanding sentiment in real-time about specific projects, initiatives or issues to inform better decision-making on the ground.
• And many more!

 

Why has sentiment analysis been underwhelming up to now?

 

Sentiment analysis is a complex and academic topic that has, arguably, been around since the fifties in one form or another, and is still evolving. Over the last decade, advances in sentiment analysis have accelerated, but have still principally relied on a set of statistical methods. Early sentiment analysis tended to depend on a statistical technique known as “bag of words”, based on the occurrence of words appearing within a text.

Increasingly, advanced sentiment analysis has leveraged AI, but here there has been considerable evolution, moving from Natural Language Processing (NLP) to Natural Language Understanding (NLU). Neural networks (Recurrent Neural Networks, Long-Short Term Memory) were utilized, although these methods have their drawbacks. More recently, advanced sentiment analysis is moving towards using “Transformers” – a more advanced neural network architecture that better understands the context of a piece of content and the opinions within it. However, even here, some popular Transformers like BERT (developed by Google) are starting to get out of date.

As already noted, sentiment analysis is not easy because human judgement is sometimes required to properly interpret the sentiment of a piece of content. For example, sarcasm and dry humour are not always easy for a person to interpret, let alone an algorithm. Content and language themselves change and evolve, particularly in discussions that include slang. If something is “sick”, does that mean it is good or bad? The answer may well depend on the age of the person you ask. There is also an increasing use of emojis and abbreviations in describing how we feel about something.

A further complicating factor is the multi-language nature of certain content. If you are tracking employee sentiment across online discussions, for example, then you may have to analyze conversations in hundreds of different languages.

Another problem with the often “simplistic” approach of current sentiment analysis solutions is that they aggregate the sentiment of a piece of content based on the balance between positive, neutral and negative words or elements of a text. For example, a review of a restaurant might say it loves the ambience and service, feels indifferent about the main course, but absolutely hates the dessert. Supposing an article or a conversation contained equally strong positive and negative sentiment, this might then be deemed by a solution to be of neutral sentiment, with the positive and negative cancelling each other out.

But that’s not really going to tell you much; it would be far more useful to be aware of more of the detail – a body of text can be both positive and negative. You’d like to know how people feel about the service and the food distinctly.

The problem with many of the ways in which existing solutions determine sentiment analysis is that they are based on Natural Language Processing, but really require Natural Language Understanding. We need AI that can replicate the human judgement required to actually interpret text and understand the sentiment; this is where HIPSTO’s Apex platform differs from other solutions, delivering the Natural Language Understanding that is needed at the centre of effective sentiment analysis.

 

Why HIPSTO’s APEX solution is state of the art

 

Let’s explore in more detail why HIPSTO’s approach is unique. We bypass traditional statistical approaches that dominate current solutions, and leverage state-of-the-art neural networks and best-in-class Transformers (mT5 rather than BERT) to deliver true Natural Language Understanding. Effectively, we take the best of what has gone before, and build on it to deliver exceptional results. This includes adding extra processing layers and mathematical weighting formulae that accurately define sentiment composition at the “article” or “item” level.

In more detail, our process works something like this:

1. The input text is broken into units and fed into our proprietary neural network architecture that is based on best-of-breed Transformers, effective across over 100 major languages.
2. The sentiment of the text is determined based on Natural Language Understanding, taking into account the full context of the text such as the order of words or when meaning changes due to the presence of another word.
3. We run a further set of proprietary weightings and algorithms that can aggregate the sentiment for the whole article or item, remove the neutrality bias that is characteristic of other providers, particularly relating to long-form content, and simulate how a human would judge sentiment.
4. We present the findings, accurately highlighting the proportion of positive, negative and neutral sentiment in a format that is then actionable based on your needs.

 

 

Over 94% accuracy

 

A popular measure of the success of the neural network architecture is something called an F1 score – a metric commonly used in machine learning circles to demonstrate the accuracy of a model, taking into account both precision (the proportion of results that are accurate) and recall (the proportion of relevant results that are identified). The closer to 1, the better, and HIPSTO’s Apex Solution currently scores 0.9446 – a score derived and verified using a consistent verification method that includes human evaluation. We have continually used these tests in our commitment to developing and improving our solution, benchmarking ourselves against major platforms and using human evaluation to consistently verify our F1 score.

Subsequently, our offering outperforms every other provider we have tested it against for accuracy, including Amazon Comprehend and Google Cloud’s Sentiment Analysis engine. With 94.5% accuracy, HIPSTO Apex is truly state-of-the-art and considerably ahead of others, where scores between 70 and 80% are the norm and where 80 to 85% is regarded as industry-leading.

 

 

Why HIPSTO’s approach is unique

 

However, it’s not just our impressive F1 score that we believe makes our approach stand out. There are a number of other unique approaches we utilise, including:

• Multi-lingual processing capability
• Overcoming neutrality bias
• Coverage of all content forms
• End-to-end automation

 

Multi-lingual processing capability

 

One of the most exciting elements of HIPSTO’s approach is the ability of our proprietary neural network architecture to detect sentiment delivered in over 100 languages, with a consistent level of accuracy across whatever language the content is delivered in. In terms of sentiment analysis, this is highly significant as it means you can track public opinion across social media, the global press and even unstructured internet pages. For global organizations looking to track brand sentiment and protect their reputation, the ability to get a sense of how your global customer base truly regards you across all major languages is both liberating and even daunting for those who have not accessed those insights before.

 

Overcoming neutrality bias

 

One of the shortcomings of existing sentiment analysis solutions is that they tend to exaggerate neutral sentiment compared to how a human would judge a piece of text. For example, Amazon Comprehend adds content to a fourth sentiment bucket beyond positive-negative-neutral called “mixed”, but the result is that too much content ends us being added to this category, making it not useful for analysis. Meanwhile, Google Cloud’s AI services tend to let positive and negative sentiment cancel each other out, with more content
being declared neutral.

The aggregation weighting within HIPSTO’s Apex solution gives more prominence to the positive and negative sentiment, mimicking human judgement and reporting on the true complexity of sentiment inherent in a piece of text through percentage scores for positive, negative and neutral sentiment.

 

Coverage of all content forms

 

HIPSTO’s approach covers all forms of content, whether it is long-form like an analyst report or short-form like a comment on a discussion forum. Some Transformers have had constraints relating to longer-form content, but the HIPSTO approach has overcome these.

Through our AI platform, FalconV, we can cover high-frequency content types like news articles, as well as low-frequency content such as analyst reports. This enables us to run industry-leading, accurate sentiment analysis on all content from regulatory filings such as SEC filings, to Reddit groups, to web pages, to news articles, to Twitter feeds, to product reviews and beyond. Combined with our multi-language approach, there are no constraints on what our sentiment analysis can cover.

 

End-to-end automation

 

Because sentiment analysis runs as part of our overall AI platform, it is a real-time, full-cycle service with 100% end-to-end automation. This means that within one solution, we can scrape the data from the sources you need right across the internet, process and analyze the data, report on it and even generate data in a format that can potentially be used to create custom reports or be processed within other systems.

In short, the HIPSTO platform can extract, label, categorize, detect duplicates, analyze and then deliver data without the need for human intervention, so can always be operational and monitoring for sentiment. An advanced sentiment API is also available, which means the platform can plug into your existing technical stack – this can also be used in conjunction with our proprietary web scraping technology.

 

 

What does the future hold?

 

We are excited about the future. We will continue to evolve, improve and evaluate our sentiment analysis across our offerings and for different use cases, leveraging deep knowledge in AI and NLP fields, as well as experience in how our services can be applied to different sectors such as financial services. We are also working on building knowledge graphs based on sentiment. We’ll be covering some of these developments in more detail on our blog.

We’d also like to think that our superior results and accuracy will be a wake-up call to some of the larger providers like Amazon Comprehend and Google Cloud to recalibrate their sentiment analysis services; in this way, everyone will benefit.

Above all, we hope our sentiment analysis will enable the kind of decision-making that helps improve organizations and even people’s lives. When a change is brought about that helps turn a negative sentiment into a positive one, it kickstarts real change and real value.