Neural Question Answering (QA) System

A Neural Question Answering (QA) System is a QA System that implements a neural QA algorithm and uses a Deep Neural Network System.

• AKA: DNN QA System.
• Context:
  • ...
  • It can solve a Deep QA Task by implementing a Deep QA Algorithm.
• Example(s):
  • an EAGLi System,
  • a Deep Similarity Neural Network QA System,
  • a QA-CNN System,
  • a QA-RNN System,
  • a QA-LSTM System.
  • …
• Counter-Examples:
  • a ELIZA System,
  • a Baseball System,
  • a LUNAR System.
• See: Deep Learning, Natural Language Processing, Question Answering System, QA Dataset, Artificial Neural Network, Convolutional Neural Network, Recurrent Neural Network, Long Short-Term Memory.

---

---

References

2016a

• (Stroh & Mathur, 2016) ⇒ Eylon Stroh, and Priyank Mathur (2016). "Question Answering Using Deep Learning".
  • QUOTE: (...) GRU and LSTM units allow recurrent neural networks (RNNs) to handle the longer texts required for QA. Further improvements – such as attention mechanisms and memory networks – allow the network to focus on the most relevant facts. Such networks provide the current state-of-the-art performance for deep-learning-based QA.

    In this project, we study the application of several deep learning models to the question answering task. After describing two RNN-based baselines, we focus our attention on end-to-end memory networks, which have provided state-of-the-art results on some QA tasks while being relatively fast to train.

2016b

• (Tan et al., 2016) ⇒ Ming Tan, Cicero dos Santos, Bing Xiang, and Bowen Zhou. (2016). “LSTM-based Deep Learning Models for Non-factoid Answer Selection.” In: Proceedings of ICLR 2016 Workshop. eprint arXiv:1511.04108
  • QUOTE: In this paper, we propose a deep learning framework for answer selection which does not require any feature engineering, linguistic tools, or external resources. This framework is based on building bidirectional long short term memory (biLSTM) models on both questions and answers respectively, connecting with a pooling layer and utilizing a similarity metric to measure the matching degree. We improve this basic model from two perspectives. Firstly, a simple pooling layer may suffer from the incapability of keeping the local linguistic information. In order to obtain better embeddings for the questions and answers, we build a convolutional neural network (CNN) structure on top of biLSTM. Secondly, in order to better distinguish candidate answers according to the question, we introduce a simple but efficient attention model to this framework for the answer embedding generation according to the question context.

2015a

• (Feng et al., 2015) ⇒ Minwei Feng, Bing Xiang, Michael R. Glass, Lidan Wang, and Bowen Zhou. (2015). “Applying Deep Learning to Answer Selection: A Study and An Open Task.” In: Proceedings of 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU 2015). ISBN:978-1-4799-7291-3 doi:10.1109/ASRU.2015.7404872
  • QUOTE: Our work serves as solid evidence that deep learning based QA is an encouraging research direction. The scientific contributions can be summarized as follows: (1) creating a new QA task in the insurance domain and releasing a new corpus so that different methods can be fairly compared; (2) proposing a general deep learning framework with several variants for the QA task and comparison experiments have been conducted; (3) utilizing novel techniques that bring improvements: multi-layer CNN with layer-wise supervision, augmented CNN with discontinuous convolution and novel similarity metric that combine both L2-norm and inner product information; (4) the best scores in this paper are very promising: for this challenging task (select one answer from a pool with size 500), the top one accuracy of test corpus can reach up to 65.3%; (5) for researchers who want to proceed with this task, this paper provides valuable guidance: a shared layer structure should be adopted; no need to append a hidden layer after the CNN; two levels of CNN with layerwise training improves accuracy; discontinuous convolution sometimes can help; the similarity metric plays a crucial role and the proposed metric is preferred and finally increasing the filter quantity brings improvement.

2015b

• (Gobeill et al., 2015) ⇒ Julien Gobeill, Arnaud Gaudinat, Emilie Pasche, Dina Vishnyakova, Pascale Gaudet, Amos Bairoch, and Patrick Ruch. (2015). “Deep Question Answering for Protein Annotation.” In: Database Journal, 2015. doi:10.1093/database/bav081
  • QUOTE: Deep QA is the ability of a QA engine to propose answers found in no corpus. Deep QA is needed to answer questions such as ‘What molecular functions are associated with protein X ?’. Such questions are simple regarding their structure, they are basically ‘What’ questions but traditional factoid QA systems, which are based on redundancy and dictionaries, cannot find the relevant answers. Figure 2 illustrates this difference between standard QA, for which explicit answers are found in retrieved documents, and deep QA, for which implicit answers are found in the output of a supervised classifier applied a posteriori on these retrieved documents.

     

    

    Figure 2. Deep QA. In standard QA, answers are extracted from some retrieved documents. In Deep QA, curated data are exploited to build a supervised classification model, which is then used to generate answers.

---