Chatbot Evaluation and Maintenance

Exploratory data analysis and interactive model-understanding and evaluation tool for chatbot training data and feedback.

The tool is built to answer questions such as:

• What examples does my model perform poorly on?
  • In terms of classification?
  • In terms of user feedback?
• Can user feedback be attributed to adversarial behaviour?
• Is there mislabeled text in the training set?

Features

The tool runs through a browser-based dashboard. Standard features include:

• A visualization of the low-dimensional representation of the embeddings for both user queries as well as any stored data.
• Aggregate analysis on chatbot metrics:
  • language
  • feedback type (upvote, downvote, something else, none)
  • chatbot FAQ ID
  • confidence of top intent
  • outlier scores for training data
  • novelty scores for feedback data
• Future metrics to be added include:
  • ranking of delivered content from bot API
  • visible ranking when presented to users
  • timestamp
  • website
  • IP address
  • session ID
  • attach policies (set of rules governing a chatbot)
  • user annotated FAQ ID
  • distance of query from FAQ ID

Feedback Workflows

Workflow templates provide standard solutions for chatbot performance evaluation and maintenance.

• Confirmation that upvote user feedback agrees with chatbot predictions.
  • If user labels agree with chatbot labels, then feedback is likely genuine and can be quickly added to training data.
  • If user labels are in many different categories, then the classifier performed poorly, but user says the result is correct. The examples should be checked before being added to training data.
  • If user labels are in another category, then feedback can be attributed to adversarial behaviour.
• Confirmation that downvote user feedback agrees with chatbot predictions.
  • If user labels agree with chatbot labels, then feedback can be attributed to adversarial behaviour.
  • If user labels are in many different categories, then the classifier performed poorly and the user confirms this. The examples should be checked and reviewed.
  • If user labels are in another category, then feedback is likely genuine, but needs confirmation. The examples should be checked before being added to training data.
• Confirmation that "something else" button presses agree with chatbot predictions.
  • When a user chooses something else, confidences of previously returned chatbot intents should be low.
  • Examples should be checked before deciding whether to add data to an existing category or a new category.