Published on February 2, 2024, 6:04 am

A significant focus in enterprise work today is the automation of human tasks for increased efficiency. IBM is exploring the potential of generative artificial intelligence (AI), specifically large language models (LLMs), as a means to automate processes. Their proposed software framework, called “SNAP,” trains an LLM to generate predictions of the next actions in a business process based on past events. These predictions can then serve as suggestions for businesses.

IBM’s research paper, titled “SNAP: Semantic Stories for Next Activity Prediction,” highlights how SNAP can enhance the performance of predicting the next activity across various business process management datasets. While other scholars have been using time series data to predict future events or actions, IBM’s work focuses on events in sequence and probable outcomes.

SNAP stands for “semantic stories for the next activity prediction.” The concept of predicting the next activity (NAP) is not new; it has been an area of systems research for decades. Typically, NAP utilizes older forms of AI to forecast what will happen next based on previously inputted steps from business logs, commonly known as “process mining.” IBM’s contribution to this framework lies in incorporating semantic stories using language models like GPT-3. By harnessing the power of these models, more intricate details about a business process can be captured and transformed into coherent narratives using natural language.

Compared to older AI programs that only consider activity sequences without accounting for additional numerical and categorical attributes, LLMs excel at extracting finer details and presenting them as stories. For example, when analyzing a loan application process with various associated data points, an LLM can effortlessly convert those data items into a meaningful narrative: “The requested loan amount was $20,000 and was made by the customer. The activity ‘Register Application’ took place on turn 6, which occurred 12 days after the case started…”

The SNAP system involves three steps. First, a story template is created, followed by filling out the template to develop a complete narrative. Lastly, these stories are used to train the LLM in predicting the next event in similar narratives. By transforming business process attributes into generative AI narratives, IBM’s SNAP enables businesses to anticipate future developments.

During their research, Oved and his team tested whether SNAP outperformed older AI programs in next-action prediction. They utilized four publicly available datasets, including car-maker Volvo’s IT incidents database, environmental permitting process records, and a collection of imaginary human resources cases. The authors evaluated three different language foundational models: OpenAI’s GPT-3, Google’s BERT, and Microsoft’s DeBERTa. Ultimately, they found that all three models surpassed established benchmarks in terms of predictive outcomes. Interestingly, while GPT-3 is more powerful than the other two models overall, its performance on tests was relatively modest. This leads to the conclusion that “even relatively small open-source LFMs like BERT have solid SNAP results compared to large models.”

Furthermore, the researchers discovered that the use of full sentences in language models significantly impacted performance. The coherent and grammatically correct structure of semantic stories from business process logs played a crucial role in the success of the SNAP algorithm. A comparison between full sentences generated by GPT-3 and other models versus a concatenated text string approach revealed that the former approach yielded far greater accuracy.

Generative AI proves its value in extracting insights from data that traditional AI struggles to capture fully. In particular, it excels when dealing with vast categorical feature spaces like user utterances and free-text attributes. However, SNAP’s advantages diminish when working with datasets lacking semantic information or written details.

Looking ahead, advancements in technologies such as robotic process automation hold promise for enhancing datasets with richer semantic information derived from user and system utterances. These enhancements can potentially improve the accuracy of predictions made by approaches like SNAP.

IBM’s innovative work with SNAP demonstrates the potential of generative AI to automate processes, predict future actions, and leverage the richness of language models to extract insights from complex business processes. As AI technologies continue to evolve, businesses can harness these advancements to drive efficiency and make informed decisions based on precise predictions.


Comments are closed.