Salesforce’s new AI project aims to organize emails and documents

salesforce

Salesforce.com Inc. believes that artificial intelligence could hold the key to tackling one of the biggest time sinks in the modern workplace: checking email.

On Thursday, the company shared details about an internal project aimed at using deep learning to automatically summarize messages and documents so that they can be read more quickly. Salesforce is not the first tech firm that is trying to apply the technology in such a way, with IBM Corp., Google Inc. and several others pursuing similar initiatives. But its researchers claim that their approach has the potential to provide much more pronounced benefits than older methods.

Salesforce’s project is based on what is known as a recurrent neural network, a type of deep learning system that lends itself well to processing large data volumes. There are many variations, but those designed to handle text usually follow the same basic principles.

The average model works by generating a mathematical representation of each phrase in a document and using it to produce a new word sequence, which can be a summary, a translation or something else. The process is then repeated with the new sequence serving as the input. This approach enables the neural network to keep improving the result until it meets its pre-programmed accuracy requirements. In a blog post, Salesforce detailed that it expanded upon the formula by mixing contextual information into the loop.

As part of the initial text evaluation, the company’s model reads phrases not only from left to right but also the other way around in a bid to better understand their meaning. It keeps consulting the source material during the summary generation stage to ensure that the output is up the standard. Moreover, results produced early in the process are retained as well to be used as reference once the final version starts taking shape. 

According to Salesforce, the model broke several industry records for automated summarization during an internal test. It was shown to be 12 to 16 percent more effective depending on the difficulty of the sample content. The company’s researchers expect that there will be opportunities to make it even more accurate in the future as the quality standards used to guide AI projects improve. 

The project is part of a broader effort by Salesforce that focuses on using deep learning to improve productivity of workers who rely on its cloud services. The company’s plans revolve around Einstein, a set of automation features designed to streamline tasks such as engaging leads and handling support requests.

Image: Salesforce