The Salesforce AI Breakthrough Explained

Share this article...

Practically everyone in the Salesforce industry is sitting with bated breath waiting for the revolutionary first AI-enabled CRM to materialise. Salesforce announced their first break-through towards that revolution: Text Summarisation.

Although the world is still being kept in suspense, this is an exciting leap in platform capabilities, with this particular area of AI described as ‘one of the most difficult Natural Language Processing (NLP) challenges” – not exactly picking the low hanging fruit for themselves! Still, in true pioneering style typical of Salesforce, they are striving to deliver ground-breaking impact on productivity for their customers – and putting themselves far ahead of the competition.

Text Summarisation: The Basics

As I dropped in the introduction, Text Summarisation was no easy quest, but one of Salesforce’s AI-priorities on their Einstein development roadmap.

The aim of Text Summarisation is to take a block of text and generate a coherent summary. The challenge in the past was that the summaries generated were limited in length and cut-off from context: “because of the way we traditionally train deep learning models, they don’t have a global sense”, explained Socher, a Salesforce AI leading researcher, and as a result, “they were forgetting what was said before”[2]. Now, the AI breakthrough will deliver the world NLP models that will be able to incorporate a ‘global sense’ of the document into their output, giving people summaries that are actually valuable.

Two Existing Methods (and their limitations)

Progression in the field of Text Summarisation has been restricted by limitations in the two ways that the model finds words and compiles the summary. These two methods are:

  • Source language: text strictly from within body of text.
  • Fresh language: uses text from external sources.

The trouble is that neither produces the accuracy required for the summaries to be of any use to the human reader; the source language method isn’t flexible in taking context into account, and the new language often produces nonsense sentences that require untangling.

The team, originally from MetaMind[1], saw the direction they needed to take Text Summarisation in – improving the accuracy of the ‘Fresh’ language method.

3 Things Salesforce are doing differently

Knowing that the key to breakthrough lay in improving the accuracy of ‘fresh’ language, there were three main things the research team did in order to outperform any previous accuracy tests by 12-16%[2] – relatively speaking, a significant achievement!

1. They opened up to more context

Opening the model up to more information (data & text) outside of the document meant the summary could consider the context, and not treat it as a mere standalone document. Contextual information, combined with memory, equals superior summaries.

2. They used ‘reinforcement learning’

Which means, in short, that test was run successively until ‘optimal behaviour’ was reached. How? The model would endeavour to beat its own score each time.

3, They were able to reduce repetitive language

Using some complex method (I don’t know…let’s leave it to the experts!)

Business Cases

The small-scale use cases that we can identify in our working days and customer interactions will actually translate into greater impacts.

Text summaries can be used for[2]:

  • Emails (yes, those ridiculously long essays or convoluted chains…)
  • Customer interaction history (a decade of customer loyalty condensed for your new starter)
  • Meeting documents (let’s face it, we’ve all been in a rush on the way to a meeting)
  • Contracts (the legal padding for the worst case scenario)
  • Online product reviews (was it well received or not, without the life-story)

We’ve all heard the statistics about the shocking proportion of time at work our sales, service, marketing…any colleagues are spending sifting through the web of words people like to express. But most of the text isn’t relevant, and ultimately we are not constructing an efficient way to work. Period.

What’s for sure is that information overload is a chronic plague of the working world, and is responsible for a host of conditions under the umbrella term ‘burn-out’, such as anxiety. We should be moving towards processes and routines of effectiveness, transparency and clarity. In the words of Arianna Huffington: “Ours is a generation bloated with information and starved for wisdom”[3] – something we should pause and consider.

So when’s this coming?

Whoa, not so fast! Although these advances are significant, they won’t appear on the market or baked into a product for some time. But, thankfully, the future of Text Summarisation is coming!


[1]*MetaMind were acquired by Salesforce in April 2014. MetaMind is a platform that provides predictive capabilities in language, vision and database [more].

[2] Forbes, ‘Salesforce Announces AI Breakthrough, Reducing Information Overload’, May 11th, 2017.

[3] Huffington, A., ‘Thrive: The 3rd Metric of Success’

One thought on “The Salesforce AI Breakthrough Explained

Add Comment