You are on page 1of 1

Question Answering System for SQuAD

Zhongjie Li, Xuandong Lei, Lu Bian

Introduction Baseline Model Co-attention Model


Question answering (QA) is a system that can pull
answers from an unstructured collection of natural
language documents. Its a crucial task that requires both The model includes an
natural language understanding and processing. encoder that understand
We are using Stanford Question Answering Dataset the meaning of question
(SQuAD), which predicts an answer span as a starting and paragraph, and a
decoder to predict the
and ending indices in context. Results are evaluated with
answers starting and
softmax loss, F1 score and EM score. ending indices based on
the understanding.
Context: .... in adverse weather conditions , ice and
snow clearing equipment can be used to improve T h e b a s e l i n e mo d e l
traction on the landing strip... implements a sequence
Question: in adverse weather conditions , what can be attention mix model.
used to improve traction on the landing strip ?
Answer: ice and snow clearing equipment

Dataset
Glove Vectors for Word
Bi-Directional Attention Model
Representation Experiment Evaluation
Training: 81387 triplets of
(Question, Context, Predictions: p(answer_start), p(answer_end)
Answers) Loss: CE(y_s, y_s) + CE(y_e, y_e)
Bi-directional attention works best, it has
Validation: 4285 triplets of more hidden layers
(Question, Context,
Answers)

Max length of context: 766

Max length of question: 60

99.9% length of context:


449
99.9% length of question: Future Goals
30
Dynamic Pointing Decoder: There may exist several answer spans, each corresponding to a local
If want to fix the length of maxima. Using dynamic pointing decoder which alternates between estimating the start and end of the
paragraph and question in Similarity Matrix: S = PWQT, use W to learn the similarity between P and Q answer span, enables model to recover from initial local maxima.
the model, we can pick Paragraph-to-Question Attention: which question word are most relevant to each paragraph Encoder: Use natural language sentence matching (NLSM) to assess the relevance between
450 and 30 as their word, P = softmax(S)P paragraphs and questions
lengths Question-to-Paragraph Attention: which paragraph words have most similarity to each query
word and can be used to answer the question [1]Xiong C. et. al. Dynamic coattention networks for question answering. arXiv preprint 2016.
Q = QTsoftmax(maxcol(S)) [21] Wang. Z. et. al. Bilateral multi-perspective matching for natural language sentences. arXiv preprint
2017.

You might also like