Diagnosing and Prioritizing Issues in Automated Order-Taking Systems: A Machine-Assisted Error Discovery Approach

Research Background and Issues

• Identified Problems or Challenges:
  Automated order processing systems, such as IBM watsonx Orders, often encounter errors in Automatic Speech Recognition (ASR), Natural Language Understanding (NLU), and Natural Language Generation (NLG) modules during customer interactions. These issues can lead to conversational misunderstandings, negatively impacting key performance indicators (KPIs) such as automation rate, escalation rate (proportion of orders handled manually), conversational efficiency, and customer experience. Identifying the root causes of these errors requires teams to manually inspect and interpret vast amounts of data, which is costly and time-consuming.

• Importance of the Problem:
  In the quick-service restaurant industry, highly efficient voice interaction systems are critical for enhancing customer experience and business profitability. Even minor errors can significantly affect KPIs, thereby impairing system performance and customer satisfaction.

• Research Motivation and Related Work:
  Existing literature primarily focuses on improving the design and development of conversational AI, with limited attention to error diagnosis and maintenance post-deployment. Current methods (e.g., confidence assessment or repair mechanisms) can only address explicit or real-time errors, lacking effective diagnosis of latent errors in historical data. Furthermore, tools for analyzing large-scale multimodal data often fall short in expressing the overall impact of errors on KPIs.

Solution

• Proposed Method or Solution:
  The authors developed and demonstrated an interactive tool, DTTool, designed to assist in identifying, diagnosing, and prioritizing errors in automated order processing systems. DTTool combines human and machine capabilities to efficiently uncover high-impact errors, significantly reducing the workload of manual analysis.

• Innovations:

  1. Utilization of machine learning algorithms to automatically analyze and annotate conversation logs and voice data, flagging anomalies that may impact KPIs.
  2. Introduction of a graphical user interface (GUI) based on Conversation Analysis, enabling efficient human review of logs and audio for root cause analysis.
  3. Design of a generic, data-pattern-independent data model that can be extended to other conversational AI applications.
  4. Provision of a "top-down" exploration perspective via a Summary Page, helping analysts quickly focus on critical errors.

• Implementation Steps and Key Technologies:

  1. Offline Analysis: Comparing outputs from production models and offline speech recognition models (which have lower error rates but are non-real-time), identifying discrepancies, and performing statistical correlation analysis with KPIs, retaining only anomalies strongly correlated with KPIs.
  2. Data Modeling: Storing conversations and events in a generic data structure, including conversation logs, system-generated annotations, and user annotations.
  3. User Interface: Offering a multi-page user interface to support various tasks, including single-session analysis pages, global views for multi-session analysis, and features for audio spectrogram and conversation visualization.
  4. Result Export and API Support: Integration with Python scripts to enable data scientists to conduct further in-depth analysis.

Research Outcomes

• Specific Results:

  1. The authors implemented DTTool and validated its performance and effectiveness through multiple rounds of user evaluation.
  2. DTTool has been deployed within IBM's Data Triage team, aiding in the discovery of hidden errors and generating tickets for the development team to fix system issues.
  3. The tool successfully identified general errors beyond speech recognition issues (e.g., problems in natural language generation).

• Comparison with Existing Methods and Advantages:
  DTTool significantly reduces the manual effort required to locate hidden errors in historical logs. Its error annotation mechanism and Summary Page greatly enhance the efficiency of error discovery. Moreover, the system can diagnose errors in individual sessions and assess their overall impact on KPIs.

• Experimental and Evaluation Results:

  1. In initial user studies, participants found error annotation and similar segment search highly helpful for error discovery.
  2. In the second iteration, the addition of the Summary Page enabled users to adopt a "top-down" error analysis approach, significantly accelerating error discovery.
  3. Over a period of use, the Data Triage team identified 28 high-impact errors using the tool and prioritized them for resolution.

• Limitations and Future Directions:

  1. The current tool primarily focuses on analyzing speech recognition errors; future work should extend its capabilities to address errors in more general conversational AI models.
  2. Expanding the tool to support additional KPIs, such as order completion time and payment window changes, is necessary.
  3. Exploring the application of the tool to broader conversational AI systems, such as other voice assistants or intelligent customer service platforms, is a promising direction.

In summary, DTTool significantly improves the efficiency of error diagnosis in automated order processing conversational systems, offering substantial value to industry practices and providing new insights and inspiration for the design of human-machine collaboration tools.