Evaluation and Improvement of Chatbot Text Classification Data Quality Using Plausible Negative Examples

Kit Kuksenok, Andriy MartynivACL 2019

TL;DR

This paper introduces nex-cv, a novel metric for evaluating chatbot text classification that uses "plausible negative examples" from low-population classes. This metric is designed to be actionable for non-developers, less optimistic than traditional metrics, and model-agnostic, enabling better data quality improvements for chatbot development.

What Problem Does This Paper Address?

Chatbots, especially those used in business domains like recruitment, rely heavily on Natural Language Processing (NLP) and Machine Learning (ML) to understand user queries and provide relevant answers. The performance of these chatbots is directly tied to the quality of their training data, which often consists of "intents" (categories of user queries). However, maintaining and evaluating these datasets can be challenging, especially for non-developer staff like recruiters who are often responsible for dataset curation. Existing evaluation metrics can be overly optimistic, making it difficult to identify genuine areas for improvement, and may not be easily understood by all stakeholders. This paper addresses the need for an evaluation metric that is actionable, interpretable, and helps improve the quality of chatbot training data, particularly for unbalanced datasets with many low-frequency classes.

Methods

The core contribution of this paper is the nex-cv (negative example cross-validation) metric. Unlike traditional cross-validation, nex-cv incorporates "plausible negative examples" by strategically using low-population classes from the dataset. The idea is that these rare or emergent topics, while not always the primary focus, are crucial for a chatbot to handle gracefully.

The nex-cv metric works by:

  1. Partitioning data: It splits the dataset into training and testing sets.
  2. Leveraging low-population classes: It identifies classes with few examples. These classes are then used as negative examples during the evaluation.
  3. Two selection strategies:
    • Cutoff-based selection (K parameter): Classes with fewer than K occurrences are designated as negative examples.
  4. Proportional selection (P parameter): A proportion P of the least frequent classes are selected as negative examples.
  5. Model-agnostic evaluation: The metric is designed to work with any text classification model, allowing for comparisons between different NLP engines.

The authors also propose an "inspection-explanation-refinement" process for using nex-cv, where the metric helps identify "confusions" between classes, guiding conceptual review and manual dataset improvement.

Datasets

The paper validates the nex-cv metric using seven real-world datasets from the recruitment domain, collected over a year. These datasets are in English and German and vary in size, number of questions, and number of categories (classes). The authors also mention using an "aggregate anonymous benchmark dataset" for comparison purposes. No new datasets were introduced in this paper; instead, existing recruitment-domain datasets were used for evaluation.

Key Insights and Ideas

  • The value of "plausible negative examples": Treating low-population classes as negative examples is crucial for identifying subtle errors and improving a chatbot's robustness, especially when dealing with real-world, often unpredictable user queries.
  • Bridging the gap between developers and domain experts: nex-cv is designed to be actionable by non-developers (like recruiters), allowing them to directly contribute to data quality improvements without needing deep ML expertise.
  • Overcoming metric optimism: Traditional metrics can be misleadingly optimistic. nex-cv aims to provide a more realistic assessment of performance, highlighting areas where a chatbot might fail in practice.
  • Actionable insights from "confusions": Identifying pairs of classes that the classifier frequently confuses provides concrete, understandable feedback for dataset curation and improvement.

Key Results

  • Improved data quality: The paper demonstrates that using nex-cv led to significant data quality improvements in real-world projects, resulting in an increase in the automated response rate from as low as 40% to a stable 70-75%.
  • Less optimistic evaluation: Figures 4 and 5 show that nex-cv (particularly the (0;0.15) setting) provides a less optimistic accuracy estimate compared to traditional metrics, aligning better with human ratings and revealing areas for improvement.
  • Model-agnostic comparison: Figure 6 shows that nex-cv can effectively compare different NLP engines, demonstrating that jobpal's domain-specific classifier outperformed general-purpose platforms in accuracy and "carefulness" (the ability to avoid incorrect answers due to low confidence).

Limitations and Future Work

  • Limitations: The authors acknowledge that while nex-cv is effective for identifying conceptual overlaps and improving data quality, other sources of classification error exist. The metric's applicability to domains significantly different from recruitment is not extensively explored.
  • Future Work: The authors plan to explore other text classifier and chatbot evaluation strategies, with a continued focus on understandability and transparency for multi-stakeholder design and maintenance processes. They also aim to further refine the "carefulness" score and explore other evaluation strategies.

Summary updated

BibTeX
@inproceedings{kuksenok2019evaluation,
  author = {Kit Kuksenok and Andriy Martyniv},
  title = {{Evaluation and Improvement of Chatbot Text Classification Data Quality Using Plausible Negative Examples}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure