FPI: Failure Point Isolation in Large-scale Conversational Assistants

Abstract

Large-scale conversational assistants such as Cortana, Alexa, Google Assistant and Siri process requests through a series of modules for wake word detection, speech recognition, language understanding and response generation. An error in one of these modules can cascade through the system. Given the large traffic volumes in these assistants, it is infeasible to manually analyze the data, identify requests with processing errors and isolate the source of error. We present a machine learning system to address this challenge. First, we embed the incoming request and context, such as system response and subsequent turns, using pretrained transformer models. Then, we combine these embeddings with encodings of additional metadata features (such as confidence scores from different modules in the online system) using a "mixing-encoder" to output the failure point predictions. Our system obtains 92.2% of human performance on this task while scaling to analyze the entire traffic in 8 different languages of a large-scale conversational assistant. We present detailed ablation studies analyzing the impact of different modeling choices.

A Mixed-Methods Approach to Understanding User Trust after Voice Assistant Failures

“…Such similarities motivate us to draw parallels between the NLP robustness literature and HCI perspectives of system failures. By understanding how different types of failures affect trust in voice assistants overall, we can then try to pinpoint the underlying NLP components that are the root cause of the most critical failures that erode trust [30]. Technical solutions can then be leveraged to improve the robustness of the most critical parts of the system in order to increase user trust and long-term engagement most efficiently.…”

Improving Large-Scale Conversational Assistants using Model Interpretation based Training Sample Selection

“…the survey in (Hedderich et al, 2020)). A number of works identify utterances with processing errors through offline analysis (Sethi et al, 2021;Gupta et al, 2021;Chada et al, 2021;Khaziev et al, 2022). These approaches however still need human annotation in an active learning loop to improve production models.…”