![](https://csdnimg.cn/release/download_crawler_static/13096327/bg4.jpg)
systems to generate an answer (if there is one)
from multiple disconnected passages. There-
fore, it is far more challenging and requires
more sophisticated comprehension skills than
other datasets like SQuAD.
• Narrative QA:
(Kocisk
´
y et al., 2017) is an-
other dataset released by DeepMind and the
University of Oxford, now with descriptive
answers, that tests a system’s ability to cap-
ture the underlying narrative elements to an-
swer questions which cannot be answered by
simple pattern recognition. Question-answer
pairs are created by crowd workers using
stories and plot summaries from books and
movie scripts, without seeing the entire cor-
pus, thus preventing localized context.
Datasets with Multi-choice options
: These
datasets provide multiple-choice questions on the
text and expect the system to choose the correct
option/s. The advantages of these datasets are ob-
jective evaluation and restricted scope, as they only
require correlating options with the questions in-
stead of matching all sentences (as in extractive
datasets). For example, Microsoft’s high-quality
dataset, MCTest (Richardson et al., 2013), consists
of 500 children stories and 2000 multi-choice ques-
tions that are easy and comprehensible and do not
require world knowledge to answer. However, the
dataset size is too small to train a proper neural
model. MCScript (Ostermann et al., 2018), on the
other hand, focuses on questions that need reason-
ing using commonsense knowledge, by providing
multi-choice implicit questions created by crowd
workers on peoples daily activities. The dataset is
large enough for training.
3.2 Techniques
Several end-to-end neural methods, employing dif-
ferent types of attention mechanisms (Bahdanau
et al., 2014; Luong et al., 2015), have become the
default choice for solving the MRC task. Some of
these models are summarized:
• Match-LSTM+Pointer Network
(Wang and
Jiang, 2016): was the first end-to-end neu-
ral architecture proposed for SQuAD. This
model combines the match-LSTM (Wang
and Jiang, 2015), which gets a unidirectional
query-aware representation of passage, and
the Pointer Network (Vinyals et al., 2015),
which constructs an answer with its every to-
ken coming from the input text.
• Bidirectional Attention Flow (BiDAF)
: (Seo
et al., 2016) : uses bi-directional attention
flow, namely, a passage-to-query attention
and a query-to-passage attention, to get a
query-aware passage representation, which is
then modeled together via Bi-LSTM (Hochre-
iter and Schmidhuber, 1997) and DenseNet
(Huang et al., 2016) to generate span.
• Gated-Attention Reader
: (Dhingra et al.,
2016) uses bidirectional Gated Recurrent Unit
(bi-GRU) (Cho et al., 2014) to get contextu-
alized representations from embeddings for
both passage and query. Then, at each stage,
a multiplicative interaction between the query
and the hidden state from the previous stage is
employed in its attention mechanism to gen-
erate contextualized embeddings. It is gen-
erally applied for multi-hop reasoning based
QA over multiple documents.
• Dynamic Coattention Networks (DCN
)
(Xiong et al., 2016): introduces co-attention
mechanism in the encoder to combine
co-dependent representations of query and
document, and a dynamic iteration in the
decoder to avoid being trapped in local
incorrect maxima. The DCN decoder takes
in the output of the co-attention encoder and
generates the final predictions.
• FastQA
: (Weissenborn et al., 2017) achieved
competitive performance with a simple archi-
tecture against its complex QA counterparts
that employ a complex interaction layer to
catch the interactivity between the query and
the context. It makes use of a context/type
matching heuristic to find computable features
on the word level to instill the interaction. Due
to its simplicity, it can serve as a strong neural
baseline for future models.
• ReasoNet
: (Shen et al., 2016) makes use of
RL to dynamically determine the number of
reading and reasoning turns, based on the com-
plexity of queries and passages. The intuition
comes from the fact that the difficulty of dif-
ferent questions can vary in the same dataset
(Chen et al., 2016), and humans usually revisit
important parts of passage and question while
answering.