[Paper Review] Just ASK: Building an Architecture for Extensible Self-Service Spoken Language Understanding

Anjishnu Kumar, Arpit Gupta+10 more|arXiv (Cornell University)|Nov 1, 2017

Topic Modeling31 references56 citations

TL;DR

Describes the design and components of the Alexa Skills Kit (ASK) SLU architecture that enables extensible, self-service development of voice skills with rapid iteration and model transferability.

ABSTRACT

This paper presents the design of the machine learning architecture that underlies the Alexa Skills Kit (ASK) a large scale Spoken Language Understanding (SLU) Software Development Kit (SDK) that enables developers to extend the capabilities of Amazon's virtual assistant, Alexa. At Amazon, the infrastructure powers over 25,000 skills deployed through the ASK, as well as AWS's Amazon Lex SLU Service. The ASK emphasizes flexibility, predictability and a rapid iteration cycle for third party developers. It imposes inductive biases that allow it to learn robust SLU models from extremely small and sparse datasets and, in doing so, removes significant barriers to entry for software developers and dialogue systems researchers.

Motivation & Objective

• Motivate the need for a self-service SLU architecture to empower third-party developers and accelerate Alexa skill deployment.
• Present design considerations that balance flexibility, user experience, rapid iteration, and model deployability.
• Describe the Alexa Meaning Representation Language and how built-in intents/slot types support reuse of data resources.
• Introduce the BlueFlow infrastructure to couple research experimentation with production deployment.
• Outline the workflow from skill definition to query processing and dialogue management within ASK.

Proposed method

• Propose a modular, decoupled SLU architecture with an API-facing interface based on intents and slots.
• Use AlexaMRL to define reusable actions, entities, and properties that map to skill ontologies.
• Represent ASR and NLU with weighted finite-state transducers (WFST) generated from developer grammars and sample utterances.
• Employ transfer learning and low-resource learning via data/resources, representation spaces, and feature-level techniques to enable small data regimes.
• Adopt BlueFlow as a Python-based DSL to manage data flow, experiments, and production deployment with CLI-to-Server guarantees.
• Incorporate knowledge-base features, model optimization for deployment, and a hybrid NLU pipeline (deterministic + stochastic components).

Experimental results

Research questions

• RQ1 How can a self-service SLU framework be designed to support thousands of third-party skills while preserving first-party experience quality?
• RQ2 What representations and tooling enable rapid iteration and deployment of SLU models in a production-scale system?
• RQ3 How can built-in resources (intent schemas, slot types, AlexaMRL) be reused to reduce data requirements for new skills?
• RQ4 What infrastructure and data-flow abstractions best bridge research experimentation and production deployment?
• RQ5 How can knowledge injection and transfer learning improve SLU performance in low-data regimes?

Key findings

• The ASK architecture powers over 25,000 skills deployed via the ASK platform (as of 2017).
• A modular, decoupled design enables rapid iteration and independent updates of components without harming user experience.
• AlexaMRL enables reuse of ontology-backed actions and entities to derive semantic roles and reduce data needs for new skills.
• WFST-based ASR and NLU representations support deterministic grammar-based recognition alongside probabilistic models, enabling predictable yet flexible parsing.
• BlueFlow provides a shared framework for researchers and production to experiment and deploy models efficiently, with artifacts stored in S3 and runtime data in DynamoDB.