TL;DR:
- IBM explores generative AI, specifically large language models, for automating business processes.
- SNAP framework generates predictions for the next actions in a business process based on historical events.
- Semantic Stories for Next Activity Prediction (SNAP) leverages language models like GPT-3 to create coherent narratives from data.
- SNAP’s three-step process involves template creation, narrative building, and LLM training.
- Experiments show SNAP, powered by models like GPT-3 and BERT, outperforms traditional AI in next-action prediction.
- Coherent semantic stories significantly improve prediction accuracy compared to mere attribute strings.
- Generative AI unlocks insights from vast process data, especially in domains with complex categorical features.
- SNAP’s effectiveness relies on the presence of semantic information in datasets.
- Market implications: Generative AI like SNAP can enhance automation and decision-making in various industries, especially where unstructured data is abundant.
Main AI News:
In the realm of enterprise operations, the pursuit of automation for enhanced efficiency remains a paramount objective. The eminent technology conglomerate, IBM, has embarked on a novel exploration within its recent research endeavors, questioning whether generative artificial intelligence (AI) and, specifically, large language models (LLMs) can serve as a pivotal catalyst in the journey toward automation.
Dubbed as “SNAP,” IBM’s innovative software framework is designed to train a large language model to generate predictions for the subsequent actions within a business process, drawing insights from the entirety of preceding events. These predictive insights, in turn, emerge as invaluable recommendations for businesses to consider.
Alon Oved and his team at IBM Research have elucidated the potential of SNAP in a recent publication titled “SNAP: Semantic Stories for Next Activity Prediction,” hosted on the arXiv pre-print server. Their work demonstrates that SNAP can significantly enhance the predictive performance of next activity sequences across a spectrum of business process management (BPM) datasets.
Notably, IBM’s research contributes to a growing trend of employing LLMs to forecast forthcoming events or actions within a series. While previous scholars focused on analyzing time series data to discern trends, IBM’s approach deviates by concentrating on the concept of sequential events and their anticipated outcomes.
SNAP, an acronym for “semantic stories for the next activity prediction,” encompasses a pivotal addition to this framework. It harnesses the linguistic richness inherent in programs such as GPT-3, enabling a more comprehensive comprehension of business processes. LLMs can capture intricate details and translate them into a coherent and articulate “story” presented in natural language, surpassing the capabilities of traditional AI programs.
Unlike conventional AI programs that primarily rely on the sequence of activities for predictions, LLMs can extract a multitude of details from a database and seamlessly weave them into a natural language narrative. For instance, in a loan application scenario, an LLM can synthesize data points, such as loan amount and request start date, into a narrative format that provides a comprehensive overview of the process.
The SNAP system operates in three sequential stages. First, it establishes a template for a narrative using attributes like loan amount. In the second step, this template is employed to construct a complete narrative, which is then filled out by the language model. In the final phase, numerous such narratives are utilized to train the LLM to predict forthcoming events, constituting the “ground truth” training examples.
In their extensive research, Oved and the team evaluated whether SNAP outperforms older AI systems in next-action prediction. They conducted experiments using four publicly available datasets, including a database of IT incidents from Volvo, environmental permitting records, and fictional human resources cases. Their analysis incorporated three distinct “language foundational models”: OpenAI’s GPT-3, Google’s BERT, and Microsoft’s DeBERTa. Notably, all three exhibited superior outcomes compared to established benchmarks.
Remarkably, despite GPT-3’s superior computational power, its performance in these tests displayed relative moderation. The authors concluded that even smaller open-source language foundational models like BERT delivered robust results in the context of SNAP.
Furthermore, the authors underscored the significance of coherent and grammatically correct semantic stories derived from business process logs, emphasizing that this element is pivotal in the SNAP algorithm. Comparative analysis revealed that the narrative approach, utilizing complete and grammatical sentences, substantially outperformed the alternative strategy of combining information into a single, protracted text string.
Ultimately, the authors assert that generative AI serves as an invaluable tool in unearthing insights from vast troves of process data that traditional AI systems often struggle to capture. This holds particular promise in domains characterized by extensive categorical feature spaces, such as user-generated utterances and other unstructured text attributes.
However, it is worth noting that the efficacy of SNAP diminishes when applied to datasets that lack substantial semantic information or written detail. The authors posit that as newer technologies like robotic process automation continue to evolve, there may be opportunities to enrich datasets with richer semantic information, thus further enhancing the accuracy of predictive models.
Conclusion:
IBM’s SNAP represents a significant advancement in automating business processes using generative AI. Its ability to extract meaningful insights from complex data has the potential to revolutionize automation across various industries, making it a promising technology for businesses seeking enhanced efficiency and decision-making capabilities in an increasingly data-driven world.
