With custom AI, these graphs tell you all about themselves — in natural language

November 20, 2023

Tags: AI & Machine Learning, Data Intelligence, Natural Language Processing

There’s a lot you can learn from graphing your data, but Adobe Customer Journey Analytics users can go a step further. With updates developed by Adobe Research, a graph can now tell you about itself — in natural language. The feature is called Intelligent Captions, and the latest updates are already available for Customer Journey Analytics users.

Intelligent Captions were designed for marketers and other analysts who want to glean new information from the graphs in Customer Journey Analytics, and be able to share what they learn — in clear, simple language — with their colleagues and clients. With a click, the feature automatically generates key takeaways, so teams get more insights, more quickly. For example, if a marketer is analyzing product sales over the course of a summer, Intelligent Captions might tell them that sales of one particular product peaked in a given week, that a second product had a seasonal sales pattern with peaks and valleys on particular days of the week, and that another had abnormal trends during some parts of the summer.

Intelligent captions are powered by custom machine learning and artificial intelligence technology designed by Adobe Research.

“Developing a novel, AI-intensive feature capable of translating charts into natural language insights was challenging,” says Eunyee Koh, Principal Scientist at Adobe Research.

“Generated captions often sound robotic or repetitive — there’s just a lack of naturalness. So we distilled and fine-tuned our large language models to include a custom insight paraphrasing model that creates more natural-sounding captions. We also had to balance between accuracy and model size, which was critical for fast inference time and cost optimization.”

The Adobe Research paraphrasing model includes custom guardrails that prevent it from creating inaccurate information, referred to as “hallucinating” language. “Our experiments show that our custom paraphrasing model is much less likely to hallucinate than common large-language models,” says Koh. In testing, the guardrails reduced hallucinations of false number-related facts from 3 percent to 0 percent.

According to Prithvi Bhutani, Product Manager for Customer Journey Analytics, the team is already getting positive feedback from customers about the latest updates. Many have used the feature multiple times. They report that they’ve copied and pasted Intelligent Captions insights directly into email messages to their stakeholders, and some are already predicting that the feature will save analysts hours of time understanding and describing data.

The first release of Intelligent Captions was developed with contributions from members of the Adobe Research team, including Fan Du, Eunyee Koh, Shiv Kumar Saini, Sunav Choudhary, Ryan Rossi, Gromit Chan, Shunan Guo, Jane Hoffswell, Sana Malik, Sungchul Kim, and Xin Qian. The newly released version with generative AI technology was developed by members of Adobe Research, including Victor Soares Bursztyn, Shunan Guo, and Eunyee Koh, and Customer Journey Analytics engineer, Wei Zhang.

Wondering what else is happening inside Adobe Research? Check out our latest news here.

Related Posts