In 2017, the average person is expected to spend 12 hours and 7 minutes every day consuming some form of media and that number will only go up from here. Connected devices are ubiquitous with everything from Alexa and our thermostats to our cars and cell phones creating new information at a dizzying pace. Today the struggle is not getting access to information, it is keeping up with the influx of news, social media, email, and texts. But what if you could get accurate and succinct summaries of the key points of every email you receive or news article you scroll by?

Salesforce Research is tackling this exact challenge and today we’re excited to announce two new breakthroughs in natural language processing towards the goal of automatically summarizing a long text and serving up coherent, digestible highlights that help you stay informed in a fraction of the time. Text summarization is a very tough challenge, especially for longer texts such as news articles, and the work we are doing at Salesforce Research is pushing the state of the art. I’m honored to work with Caiming Xiong and Richard Socher to introduce a more contextual word generation model and a new way of training summarization models with reinforcement learning (RL). Together, these AI improvements enable algorithms to create easy-to-read, highly accurate multi-sentence summaries of long text, significantly improving on previous results. See here for more info on how our models work.

Illustration of our model generating a multi-sentence summary from a news article. For each generated word, the model pays attention to specific words of the input and the previously generated output.

Reaching New Levels of Productivity

Improved natural language text summaries present great potential for people and businesses of all industries to boost productivity. Take sales for example. The average B2B selling cycle may take weeks or months, and often generates lengthy and complex email threads. With our improved text summarization, reps could quickly get recaps of all emails over the course of an open opportunity, understand the important highlights and easily identify the appropriate next step, rather than spending hours manually retracing the conversation. This could allow a new sales rep to come on board and quickly get up to speed and frees up reps’ time, allowing them to focus on building smarter, more impactful relationships with customers. Or let’s say a sales rep is about to walk into a customer meeting and they get a push notification from The New York Times announcing that this same customer has acquired a company. Instead of having to race through reading a long article, our deep reinforced model for summarization could surface up the key takeaways in an easy-to-read summary so they can walk into their meeting more informed. Marketers can also benefit from this research by unlocking valuable insights from pages and pages of customer product reviews. With text summaries, marketers can easily decipher which customers need to be put into a nurture campaign versus which ones should get a loyalty coupon or discount so they can personalize marketing messages for each customer.

Business is moving faster than ever and everyone is looking for ways to become more productive and efficient. At Salesforce Research, our mission is to help advanced AI technologies and help our engineering and product teams apply these advances to real-world business problems. This research is a big step forward in tackling the hard problem of summarization. We're excited to see the impact this model can have on academics and our customers.