End-to-End Non-Factoid Question Answering with an Interactive Visualization of Neural Attention Weights

Our Contributions

  1. We present an extensible and modular service architecture that enables researchers to transform non-factoid answer selection models into fully-featured end-to-end question answering systems.
  2. Our system includes a user interface to analyze attention mechanisms and allow researchers to compare them interactively side-by-side.

QA System

Our system consists of three webservices that are responsible for different tasks:

They communicate via well-defined HTTP REST APIs and are therefore easily extendable and exchangeable. QA Frontend is responsible for handling user queries, coordinating the other components, and serving our single-page web application for the interactive exploration of attention weights. Candidate Retrieval performs efficient retrieval of candidate answers with ElasticSearch over a large background corpus of answers, e.g., from StackExchange. Candidate Ranking provides an interface to attention-based models and re-ranks a list of candidate answers regarding a query question.

All components are fully configurable and easily extensible, e.g., extending the application with new datasets and models requires little to no adaptations of our source code.

Visualizing Neural Attention Weights

A central component of our application is the single-page web application for the interactive exploration of attention models. Users can select the model for candidate ranking in the user interface and query it interactively with our application. We display all results together with the attention strengths for both the question and the answers:

Due to the system components' loose coupling, we can deploy multiple candidate ranking services with different attention models simultaneously. The user interface leverages this and allows researchers to compare two models side-by-side within the same view interactively:


Advanced attention mechanisms are an important part of successful neural network approaches for non-factoid answer selection because they allow the models to focus on few important segments within rather long answer texts. Analyzing attention mechanisms is thus crucial for understanding strengths and weaknesses of particular models. We present an extensible, highly modular service architecture that enables the transformation of neural network models for non-factoid answer selection into fully featured end-to-end question answering systems. The primary objective of our system is to enable researchers a way to interactively explore and compare attentionbased neural networks for answer selection. Our interactive user interface helps researchers to better understand the capabilities of the different approaches and can aid qualitative analyses. The source-code of our system is publicly available.


  title = {End-to-End Non-Factoid Question Answering with an Interactive Visualization of Neural Attention Weights},
  author = {R{\"u}ckl{\'e}, Andreas and Gurevych, Iryna},
  publisher = {Association for Computational Linguistics},
  booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics-System Demonstrations (ACL 2017)},
  pages = {19--24},
  month = aug,
  year = {2017},
  location = {Vancouver, Canada},
  doi = "10.18653/v1/P17-4004",
  url = "http://aclweb.org/anthology/P17-4004"