Multipurpose Intelligent Process Automation via Conversational Assistant

Alena Moiseeva, Dietrich Trautmann, Hinrich SchützearXiv 2020 · OpenReview

Abstract

Intelligent Process Automation (IPA) is an emerging technology with a primary goal to assist the knowledge worker by taking care of repetitive, routine and low-cognitive tasks. Conversational agents that can interact with users in a natural language are potential application for IPA systems. Such intelligent agents can assist the user by answering specific questions and executing routine tasks that are ordinarily performed in a natural language (i.e., customer support). In this work, we tackle a challenge of implementing an IPA conversational assistant in a real-world industrial setting with a lack of structured training data. Our proposed system brings two significant benefits: First, it reduces repetitive and time-consuming activities and, therefore, allows workers to focus on more intelligent processes. Second, by interacting with users, it augments the resources with structured and to some extent labeled training data. We showcase the usage of the latter by re-implementing several components of our system with Transfer Learning (TL) methods.

Introduction

While Robotic Process Automation (RPA) has successfully automated many "click-based" workflows, it remains inherently brittle. Minor changes in a user interface or workflow often cause these bots to fail. Intelligent Process Automation (IPA) seeks to overcome these limitations by integrating artificial intelligence to handle natural language and complex reasoning.

However, deploying IPA conversational assistants in niche industrial domains—such as the German e-learning platform OMB+—presents a significant "cold start" problem. Deep learning models typically require vast amounts of labeled dialogue data, which is rarely available for specialized applications. This paper details a strategy to implement a robust assistant starting from zero training data, eventually transitioning to a flexible, neural-network-based system.

Methodology

The researchers implemented a two-phase "bootstrap" approach to build the assistant:

Phase 1: Rule-Based Foundation

Initially, the system relied on handcrafted rules, Regular Expressions (RegEx), and the Elasticsearch API. For Natural Language Understanding (NLU), keyword-based intent classification and fuzzy string matching were used to extract entities like chapter numbers or exercise IDs, even in the presence of user typos. Dialogue management was handled by a finite state machine governed by "Mutually Exclusive Rules" (MER), which prevented redundant questioning by understanding the hierarchical structure of the e-learning content.

Phase 2: Neural Upgrade via Transfer Learning

As the rule-based system interacted with students, it automatically logged structured interaction "triples" (User Request, Next Action, Response). This generated a domain-specific dataset that was subsequently used to fine-tune Bidirectional Encoder Representations from Transformers (BERT) models for two critical tasks:

  1. Named Entity Recognition (NER): Identifying math topics and exercise identifiers.
  2. Next Action Prediction (NAP): Determining the system's next step based on the conversation history.

Dataset and Evaluation

The resulting OMB+ dataset comprises 300 structured dialogues (approximately 1,740 utterances) in German. This dataset includes labels for 7 entity types and 13 unique system actions.

Dataset statistics for the OMB+ conversational assistant.

The dataset was partitioned into 200 dialogues for training, with 50 each for evaluation and testing. Despite the relatively small size of this "toy dataset," it proved sufficient for effective transfer learning.

Results and Discussion

The initial rule-based system demonstrated high reliability, passing 100% of 130 manually annotated test scenarios. However, the transition to neural models provided greater flexibility and performance.

BERT Model Performance

The fine-tuned BERT model achieved a macro F1 score of 0.93 for NER, indicating that transfer learning is highly effective even with limited domain-specific data. For the Next Action Prediction (NAP) task, the model achieved an accuracy of 0.81 in its "extended" configuration.

F1 scores for Next Action Prediction (NAP) comparing German (GER) and Multilingual (Mult) BERT models.

A key finding was the importance of context. By including the previous system action as an additional input feature (the "extended" setting), the model's ability to resolve ambiguous conversational flows improved significantly—by roughly 60% in specific ambiguous cases. Qualitative feedback from human tutors also indicated a marked reduction in repetitive data-gathering tasks, allowing them to focus on complex pedagogical support.

Conclusion

This work demonstrates that conversational agents can serve as their own data labelers. By using a rule-based "gatekeeper" to structure initial interactions, organizations can overcome the cold-start problem in IPA. Future work will focus on expanding the dataset to further improve NAP performance and exploring the generalizability of the system to other industrial domains.

Summary updated

Abstract

Intelligent Process Automation (IPA) is an emerging technology with a primary goal to assist the knowledge worker by taking care of repetitive, routine and low-cognitive tasks. Conversational agents that can interact with users in a natural language are potential application for IPA systems. Such intelligent agents can assist the user by answering specific questions and executing routine tasks that are ordinarily performed in a natural language (i.e., customer support). In this work, we tackle a challenge of implementing an IPA conversational assistant in a real-world industrial setting with a lack of structured training data. Our proposed system brings two significant benefits: First, it reduces repetitive and time-consuming activities and, therefore, allows workers to focus on more intelligent processes. Second, by interacting with users, it augments the resources with structured and to some extent labeled training data. We showcase the usage of the latter by re-implementing several components of our system with Transfer Learning (TL) methods.
BibTeX
@misc{moiseeva2020multipurpose,
  author = {Alena Moiseeva and Dietrich Trautmann and Hinrich Schütze},
  title = {{Multipurpose Intelligent Process Automation via Conversational Assistant}},
  year = {2020},
  eprint = {2001.02284},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2001.02284},
}
Figure