Did Aristotle Use a Laptop?
A Question Answering Benchmark with Implicit Reasoning Strategies
Mor Geva1,2, Daniel Khashabi2, Elad Segal1, Tushar Khot2,
Dan Roth3, Jonathan Berant1,2
2Allen Institute for AI
3宾夕法尼亚大学
1Tel Aviv University
morgeva@mail.tau.ac.il, {danielk,tushark}@allenai.org,
elad.segal@gmail.com, danroth@seas.upenn.edu
joberant@cs.tau.ac.il
抽象的
在这个
challenge
A key limitation in current datasets for multi-
hop reasoning is that
the required steps
for answering the question are mentioned
in it explicitly. 在这项工作中, we introduce
STRATEGYQA, a question answering (QA)
benchmark where the required reasoning
steps are implicit
in the question, 和
should be inferred using a strategy. A
基本的
setup is
how to elicit such creative questions from
crowdsourcing workers, while covering a
broad range of potential strategies. We propose
a data collection procedure that combines
term-based priming to inspire annotators,
careful control over the annotator population,
and adversarial
eliminating
reasoning shortcuts. 而且, we annotate
each question with (1) a decomposition into
reasoning steps for answering it, 和 (2)
Wikipedia paragraphs that contain the answers
to each step. 全面的, STRATEGYQA includes
2,780 examples, each consisting of a strategy
its decomposition, and evidence
问题,
段落. Analysis shows that questions
in STRATEGYQA are short, topic-diverse, 和
cover a wide range of strategies. Empirically,
we show that humans perform well (87%) 在
this task, while our best baseline reaches an
accuracy of ∼ 66%.
filtering for
1 介绍
Developing models
successfully reason
那
over multiple parts of their input has attracted
substantial attention recently,
leading to the
creation of many multi-step reasoning Question
Answering (QA) benchmarks (Welbl et al., 2018;
Talmor and Berant, 2018; Khashabi et al., 2018;
杨等人。, 2018; Dua et al., 2019; Suhr et al.,
2019).
346
Commonly, the language of questions in such
benchmarks explicitly describes the process for
deriving the answer. 例如 (数字 1, Q2),
the question Was Aristotle alive when the laptop
被发明? explicitly specifies the required
reasoning steps. 然而, in real-life questions,
reasoning is often implicit. 例如,
这
question Did Aristotle use a laptop? (Q1) 能
be answered using the same steps, 但
这
model must infer the strategy for answering the
question–temporal comparison, 在这种情况下.
retrieving the context
Answering implicit questions poses several
challenges compared to answering their explicit
是
同行. 第一的,
difficult as there is little overlap between the
question and its context (数字 1, Q1 and ‘E’).
而且, questions tend to be short, lowering
the possibility of the model exploiting shortcuts
in the language of the question. 在这项工作中, 我们
introduce STRATEGYQA, a Boolean QA benchmark
focusing on implicit multi-hop reasoning for
strategy questions, where a strategy is
这
ability to infer from a question its atomic sub-
问题. In contrast to previous benchmarks
(Khot et al., 2020A; 杨等人。, 2018), 问题
in STRATEGYQA are not
limited to predefined
decomposition patterns and cover a wide range
of strategies that humans apply when answering
问题.
Eliciting strategy questions using crowdsourc-
ing is non-trivial. 第一的, authoring such questions
requires creativity. Past work often collected
multi-hop questions by showing workers an entire
语境, which led to limited creativity and high
lexical overlap between questions and contexts
and consequently to reasoning shortcuts (Khot
等人。, 2020A; 杨等人。, 2018). An alter-
native approach, applied in Natural Questions
计算语言学协会会刊, 卷. 9, PP. 346–361, 2021. https://doi.org/10.1162/tacl 00370
动作编辑器: Kristina Toutanova. 提交批次: 10/2020; 修改批次: 12/2020; 已发表 4/2021.
C(西德:13) 2021 计算语言学协会. 根据 CC-BY 分发 4.0 执照.
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
t
A
C
我
/
我
A
r
t
我
C
e
–
p
d
F
/
d
哦
我
/
.
1
0
1
1
6
2
/
t
我
A
C
_
A
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
我
A
C
_
A
_
0
0
3
7
0
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
那
我们的
分析
STRATEGYQA
节目
necessitates reasoning on a wide variety of
knowledge domains (物理, 地理, ETC。)
and logical operations (例如, number comparison).
而且, experiments show that STRATEGYQA
poses a combined challenge of retrieval and
QA, and while humans perform well on these
问题, even strong systems struggle to answer
他们.
数字 1: Questions in STRATEGYQA (Q1) require
implicit decomposition into reasoning steps (D),
for which we annotate supporting evidence from
维基百科 (乙). This is in contrast
to multi-step
questions that explicitly specify the reasoning process
(Q2).
(Kwiatkowski et al., 2019) and MS-MARCO
(Nguyen et al., 2016), overcomes this by col-
lecting real user questions. 然而, can we elicit
creative questions independently of the context
and without access to users?
第二, an important property in STRATEGYQA
is that questions entail diverse strategies. 尽管
the example in Figure 1 necessitates temporal
推理,
there are many possible strategies
for answering questions (桌子 1). We want
a benchmark that exposes a broad range of
策略. But crowdsourcing workers often use
repetitive patterns, which may limit question
diversity.
To overcome these difficulties, 我们使用
following techniques in our pipeline for eliciting
strategy questions: (A) we prime crowd workers
with random Wikipedia terms that serve as a
minimal context to inspire their imagination and
increase their creativity; (乙) we use a large set of
annotators to increase question diversity, 限制
the number of questions a single annotator can
写; 和 (C) we continuously train adversarial
models during data collection, slowly increasing
the difficulty in question writing and preventing
recurring patterns (Bartolo et al., 2020).
Beyond the questions, as part of STRATEGYQA,
we annotated: (A) question decompositions: A
sequence of steps sufficient for answering the
问题 (‘D’ in Figure 1), 和 (乙) 证据
段落: Wikipedia paragraphs that contain
the answer to each decomposition step (‘E’ in
数字 1). STRATEGYQA is the first QA dataset to
provide decompositions and evidence annotations
for each individual step of the reasoning process.
347
总之, the contributions of this work are:
1. Defining strategy questions—a class of
implicit multi-step
requiring
问题
推理.
2. STRATEGYQA, the first benchmark for implicit
that covers a diverse set
multi-step QA,
of reasoning skills. STRATEGYQA consists
的 2,780 问题, annotated with their
decomposition and per-step evidence.
3. A novel annotation pipeline designed to
elicit quality strategy questions, with minimal
context for priming workers.
The dataset and codebase are publicly available at
https://allenai.org/data/strategyqa.
2 Strategy Questions
2.1 Desiderata
We define strategy questions by characterizing
their desired properties. Some properties, 例如
whether the question is answerable, also depend
on the context used for answering the question.
在这项工作中, we assume this context is a corpus
of documents, 具体来说, 维基百科, 我们
assume provides correct content.
Multi-step Strategy questions are multi-step
问题, 那是, they comprise a sequence of
single-step questions. A single-step question is
任何一个 (A) a question that can be answered from
a short text fragment in the corpus (例如, 脚步
1 和 2 图中 1), 或者 (乙) a logical operation
over answers from previous steps (例如, step 3 在
数字 1). A strategy question should have at least
two steps for deriving the answer. Example multi-
and single- step questions are provided in Table 2.
We define the reasoning process structure in §2.2.
Feasible Questions should be answerable from
paragraphs in the corpus. 具体来说, for each
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
t
A
C
我
/
我
A
r
t
我
C
e
–
p
d
F
/
d
哦
我
/
.
1
0
1
1
6
2
/
t
我
A
C
_
A
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
我
A
C
_
A
_
0
0
3
7
0
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
问题
Can one spot helium? (不)
Would Hades and Osiris hypothetically
compete for real estate in the Underworld?
(是的)
Would a monocle be appropriate for a cyclop?
(是的)
Should a finished website have lorem ipsum
段落? (不)
Is it normal to find parsley in multiple sections
of the grocery store? (是的)
Implicit facts
Helium is a gas, Helium is odorless, Helium is
tasteless, Helium has no color.
Hades was the Greek god of death and the
Underworld. Osiris was the Egyptian god of the
Underworld.
Cyclops have one eye. A monocle helps one eye at
一次.
Lorem Ipsum paragraphs are meant to be temporary.
Web designers always remove lorem ipsum
paragraphs before launch.
Parsley is available in both fresh and dry forms.
Fresh parsley must be kept cool. Dry parsley is a
shelf stable product.
桌子 1: Example strategy questions and the implicit facts needed for answering them.
问题
Was Barack Obama born in
美国? (是的)
Do cars use drinking water
to power their engine? (不)
Are sharks faster than crabs?
(是的)
Was Tom Cruise married to
the female star of Inland
帝国? (不)
Are more watermelons
grown in Texas than in
南极洲? (是的)
Would someone with a
nosebleed benefit from
Coca? (是的)
MS
IM Explanation
The question explicitly states the required information for
the answer–the birth place of Barack Obama. The answer
is likely to be found in a single text fragment in Wikipedia.
The question explicitly states the required information for
the answer–the liquid used to power car engines. 这
answer is likely to be found in a single text fragment in
维基百科.
The question explicitly states the required reasoning steps:
1) How fast are sharks? 2) How fast are crabs? 3) Is #1
faster than #2?
The question explicitly states the required reasoning steps:
1) Who is the female star of Inland Empire? 2) Was Tom
Cruise married to #2?
X
X
X X The answer can be derived through geographical/botanical
reasoning that the climate in Antarctica does not support
growth of watermelons.
X X The answer can be derived through biological reasoning
that Coca constricts blood vessels, 因此, serves to
stop bleeding.
桌子 2: Example questions demonstrating the multi-step (MS) and implicit (IM) properties of strategy
问题.
reasoning step in the sequence, there should be
sufficient evidence from the corpus to answer the
问题. 例如, the answer to the question
Would a monocle be appropriate for a cyclop? 能
be derived from paragraphs stating that cyclops
have one eye and that a monocle is used by one
eye at the time. This information is found in
our corpus, 维基百科, and thus the question is
可行的. 相比之下, the question Does Justin
Beiber own a Zune? is not feasible, 因为
answering it requires going through Beiber’s
348
belongings, and this information is unlikely to
be found in Wikipedia.
Implicit A key property distinguishing strategy
questions from prior multi-hop questions is
their implicit nature. In explicit questions, each
step in the reasoning process can be inferred
from the language of the question directly. 为了
two questions
例子,
are explicitly stated, one in the main clause
and one in the adverbial clause. 反过来,
图中 1,
the first
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
t
A
C
我
/
我
A
r
t
我
C
e
–
p
d
F
/
d
哦
我
/
.
1
0
1
1
6
2
/
t
我
A
C
_
A
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
我
A
C
_
A
_
0
0
3
7
0
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
reasoning steps in strategy questions require
going beyond the language of the question. Due
to language variability, a precise definition of
implicit questions based on lexical overlap is
elusive, but a good rule-of-thumb is the following:
If the question decomposition can be written with
a vocabulary limited to words from the questions,
their inflections, and function words, then it is
an explicit question. If new content words must
be introduced to describe the reasoning process,
the question is implicit. Examples for implicit and
explicit questions are in Table 2.
Definite A type of questions we wish to
avoid are non-definitive questions, such as Are
hamburgers considered a sandwich? and Does
chocolate taste better than vanilla? for which
there is no clear answer. We would like to
collect questions where the answer is definitive
或者, 至少, very likely, based on the corpus.
例如, consider the question Does wood
conduct electricity?. Although it is possible that a
damp wood will conduct electricity, the answer is
generally no.
总结一下, strategy questions are multi-
step questions with implicit reasoning (a strategy)
and a definitive answer that can be reached given
a corpus. We limit ourselves to Boolean yes/no
问题, which limits the output space, but lets us
focus on the complexity of the questions, 这是
the key contribution. Example strategy questions
are in Table 1, and examples that demonstrate the
mentioned properties are in Table 2. 下一个 (§2.2),
we describe additional structures annotated during
data collection.
2.2 Decomposing Strategy Questions
Strategy questions involve complex reasoning that
leads to a yes/no answer. To guide and evaluate
the QA process, we annotate every example with
a description of the expected reasoning process.
Prior work used rationales or supporting facts,
即, text snippets extracted from the context
(DeYoung et al., 2020; 杨等人。, 2018;
Kwiatkowski et al., 2019; Khot et al., 2020A) 作为
evidence for an answer. 然而, reasoning can
rely on elements that are not explicitly expressed
in the context. 而且, answering a question
based on relevant context does not imply that
the model performs reasoning properly (Jiang and
Bansal, 2019).
问题
Did the Battle
of Peleliu or
the Seven
Days Battles
last longer?
Can the
President of
Mexico vote
在
New Mexico
primaries?
Can a
microwave
melt a Toyota
Prius battery?
Would it be
common to
find a penguin
in Miami?
Decomposition
(1) How long did the Battle of
Peleliu last?
(2) How long did the Seven
Days Battle last?
(3) Which is longer of #1, #2?
(1) What is the citizenship
requirement for voting
in New Mexico?
(2) What is the citizenship
requirement of any
President of Mexico?
(3) Is #2 the same as #1?
(1) What kind of battery does
a Toyota Prius use?
(2) What type of material is
#1 made out of?
(3) What is the melting point
的 #2?
(4) Can a microwave’s
temperature reach at least
#3?
(1) Where is a typical
penguin’s natural habitat?
(2) What conditions make #1
suitable for penguins?
(3) Are all of #2 present in
Miami?
桌子 3: Explicit (排 1) and strategy (rows 2–4)
question decompositions. We mark words that
are explicit (italic) or implicit in the input (bold).
Inspired by recent work (Wolfson et al.,
2020), we associate every question-answer pair
with a strategy question decomposition. A
decomposition of a question q is a sequence of n
steps hs(1), s(2), . . . , s(n)i required for computing
the answer to q. Each step s(我) corresponds to
a single-step question and may include special
参考, which are placeholders
referring
to the result of a previous step s(j). 这
last decomposition step (IE。, s(n)) returns the
final answer to the question. 桌子 3 节目
decomposition examples.
Wolfson et al. (2020) targeted explicit multi-
step questions (first row in Table 3), 哪里的
decomposition is restricted to a small vocabulary
derived almost entirely from the original question.
反过来, decomposing strategy questions
349
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
t
A
C
我
/
我
A
r
t
我
C
e
–
p
d
F
/
d
哦
我
/
.
1
0
1
1
6
2
/
t
我
A
C
_
A
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
我
A
C
_
A
_
0
0
3
7
0
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
数字 2: Overview of the data collection pipeline. 第一的 (CQW, §3.1), a worker is presented with a term (时间) 和
an expected answer (A) and writes a question (问) and the facts (F1, F2) required to answer it. 下一个, 问题是
decomposed (SQD, §3.2) into steps (S1, S2) along with Wikipedia page titles (P1, P2) that the worker expects to
find the answer in. 最后的 (EVM, §3.3), decomposition steps are matched with evidence from Wikipedia (E1, E2).
requires using implicit knowledge, 因此
decompositions can include any token that
是
needed for describing the implicit reasoning (rows
2–4 in Table 3). This makes the decomposition
task significantly harder for strategy questions.
在这项工作中, we distinguish between two types
of required actions for executing a step. Retrieval,
a step that requires retrieval from the corpus,
and operation, a logical function over answers to
previous steps. In the second row of Table 3, 这
first two steps are retrieval steps, and the last step
is an operation. A decomposition step can require
both retrieval and an operation (see last row in
桌子 3).
To verify that steps are valid single-step
questions that can be answered using the corpus
(维基百科), we collect supporting evidence for
each retrieval step and annotate operation steps.
A supporting evidence is one or more paragraphs
that provide an answer to the retrieval step.
总之, each example in our dataset
contains a) a strategy question, 乙) the strategy
question decomposition,
supporting
evidence per decomposition step. Collecting
strategy questions and their annotations is the
main challenge of this work, and we turn to this
下一个.
and c)
3 Data Collection Pipeline
是
Our goal
to establish a procedure for
collecting strategy questions and their annotations
at scale. 为此, we build a multi-step
crowdsourcing1 pipeline designed for encouraging
worker creativity, while preventing biases in the
数据.
1We use Amazon Mechanical Turk as our framework.
We break the data collection into three tasks:
question writing (§3.1), question decomposition
(§3.2), and evidence matching (§3.3). 此外,
we implement mechanisms for quality assurance
(§3.4). An overview of the data collection pipeline
is in Figure 2.
3.1 Creative Question Writing (CQW)
Generating natural language annotations through
是
crowdsourcing (例如, question generation)
known to suffer
from several shortcomings.
第一的, when annotators generate many instances,
they use recurring patterns that lead to biases
in the data. (Gururangan et al., 2018; Geva
等人。, 2019). 第二, when language is generated
conditioned on a long context, such as a paragraph,
annotators use similar language (Kwiatkowski
等人。, 2019),
leading to high lexical overlap
因此, inadvertently, to an easier problem.
而且, a unique property of our setup is that
we wish to cover a broad and diverse set of
策略. 因此, we must discourage repeated use
of the same strategy.
We tackle these challenges on multiple fronts.
第一的, rather than using a long paragraph as
语境, we prime workers to write questions
given single terms from Wikipedia, reducing the
overlap with the context to a minimum. 第二,
to encourage diversity, we control the population
of annotators, making sure a large number of
annotators contribute to the dataset. 第三, 我们用
model-in-the-loop adversarial annotations (Dua
等人。, 2019; Khot et al., 2020A; Bartolo et al., 2020)
to filter our questions, and only accept questions
that fool our models. While some model-in-the-
loop approaches use fixed pre-trained models
to eliminate ‘‘easy’’ questions, we continuously
350
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
t
A
C
我
/
我
A
r
t
我
C
e
–
p
d
F
/
d
哦
我
/
.
1
0
1
1
6
2
/
t
我
A
C
_
A
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
我
A
C
_
A
_
0
0
3
7
0
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
update the models during data collection to combat
the use of repeated patterns or strategies.
We now provide a description of the task, 和
elaborate on these methods (数字 2, upper row).
Task description Given a term (例如, silk), A
description of the term, and an expected answer
(yes or no), the task is to write a strategy question
about the term with the expected answer, 和
facts required to answer the question.
Priming with Wikipedia Terms Writing
strategy questions from scratch is difficult. 到
inspire worker creativity, we ask to write questions
about terms they are familiar with or can easily
understand. The terms are titles of ‘‘popular’’2
Wikipedia pages. We provide workers only with a
short description of the given term. 然后, 工人
use their background knowledge and Web search
skills to form a strategy question.
Controlling the Answer Distribution We ask
workers to write questions where the answer is
set to be ‘yes’ or ‘no’. To balance the answer
分配, the expected answer is dynamically
sampled inversely proportional to the ratio of ‘yes’
and ‘no’ questions collected until that point.
Model-in-the-Loop Filtering To ensure ques-
tions are challenging and reduce recurring
language and reasoning patterns, questions are
only accepted when verified by two sets of online
solvers. We deploy a set of 5 pre-trained models
(termed PTD) that check if the question is too easy.
If at least 4 在......之外 5 answer the question correctly,
it is rejected. 第二, we use a set of 3 型号
(called FNTD) that are continuously fine-tuned on
our collected data and are meant to detect biases
in the current question set. A question is rejected
if all 3 solvers answer it correctly. The solvers are
ROBERTA (刘等人。, 2019) models fine-tuned on
different auxiliary datasets; details in §5.1.
Auxiliary Sub-Task We ask workers to provide
the facts required to answer the question they have
written, for several reasons: 1) it helps workers
frame the question writing task and describe the
reasoning process they have in mind, 2) it helps
reviewing their work, 和 3) it provides useful
information for the decomposition step (§3.2).
2We filter pages based on the number of contributors and
3.2 Strategy Question Decomposition (SQD)
A
这
问题
相应的
和
一次
facts are written, we generate the strategy
question decomposition (数字 2, middle row).
We annotate decompositions before matching
evidence in order to avoid biases stemming from
seeing the context.
The decomposition strategy for a question is
not always obvious, which can lead to undesirable
explicit decompositions. 例如, a possible
explicit decomposition for Q1 (数字 1) 可能
是 (1) What items did Aristotle use? (2) Is laptop
在 #1?; but the first step is not feasible. To guide
the decomposition, we provide workers with the
facts written in the CQW task to show the strategy
of the question author. 显然, 可以有
many valid strategies and the same strategy can
be phrased in multiple ways—the facts only serve
as a soft guidance.
Task Description Given a strategy question, A
yes/no answer, and a set of facts, the task is to
write the steps needed to answer the question.
Auxiliary Sub-task We observe that in some
案例, annotators write explicit decompositions,
which often lead to infeasible steps that cannot be
answered from the corpus. To help workers avoid
explicit decompositions, we ask them to specify,
for each decomposition step, a Wikipedia page
they expect to find the answer in. This encourages
workers to write decomposition steps for which it
is possible to find answers in Wikipedia, and leads
to feasible strategy decompositions, 只有一个
small overhead (the workers are not required to
read the proposed Wikipedia page).
3.3 Evidence Matching (EVM)
We now have a question and its decomposition.
To ground them in context, we add a third task of
evidence matching (数字 2, bottom row).
Task Description Given a question and its
分解 (a list of single-step questions), 这
task is to find evidence paragraphs on Wikipedia
for each retrieval step. Operation steps that do not
require retrieval (§2.2) are marked as operation.
Controlling the Matched Context Workers
search for evidence on Wikipedia. We index
Wikipedia3 and provide a search interface where
workers can drag-and-drop paragraphs from the
the number of backward links from other pages.
3We use the Wikipedia Cirrus dump from 11/05/2020.
351
我
D
哦
w
n
哦
A
d
e
d
F
r
哦
米
H
t
t
p
:
/
/
d
我
r
e
C
t
.
米
我
t
.
e
d
你
/
t
A
C
我
/
我
A
r
t
我
C
e
–
p
d
F
/
d
哦
我
/
.
1
0
1
1
6
2
/
t
我
A
C
_
A
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
我
A
C
_
A
_
0
0
3
7
0
p
d
.
F
乙
y
G
你
e
s
t
t
哦
n
0
7
S
e
p
e
米
乙
e
r
2
0
2
3
results shown on the search interface. 这
guarantees that annotators choose paragraphs
we included in our index, at a pre-determined
paragraph-level granularity.
3.4 Data Verification Mechanisms
Task Qualifications For each task, we hold
qualifications that test understanding of the task,
and manually review several examples. Workers
who follow the requirements are granted access
to our tasks. Our qualifications are open to
workers from English-speaking countries who
这
have high reputation scores. 此外,
authors regularly review annotations to give
feedback and prevent noisy annotations.
Real-time Automatic Checks For CQW, 我们
use heuristics to check question validity, 为了
例子, whether it ends with a question mark,
and that it doesn’t use language that characterizes
explicit multi-hop questions (例如, 拥有
multiple verbs). For SQD, we check that
这
decomposition structure forms a directed acyclic
图形, 那
是: (我) each decomposition step is
referenced by (至少) one of the following steps,
such that all steps are reachable from the last step;
和 (二) steps don’t form a cycle. In the EVM task,
a warning message is shown when the worker
marks an intermediate step as an operation (一个
unlikely scenario).
Inter-task Feedback At each step of
这
pipeline, we collect feedback about previous steps.
To verify results from the CQW task, we ask
workers to indicate whether the given answer is
不正确 (in the SQD, EVM tasks), or if the
question is not definitive (in the SQD task) (§2.1).
相似地, to identify non-feasible questions or
decompositions, we ask workers to indicate if
there is no evidence for a decomposition step (在
the EVM task).
Evidence Verification Task After the EVM
它是
step, each example comprises a question,
回答, 分解, and supporting evidence.
To verify that a question can be answered
by executing the decomposition steps against
the matched evidence paragraphs, we construct
an additional evidence verification task (EVV).
In this task, workers are given a question,
its decomposition and matched paragraphs, 和
are asked to answer
the question in each
decomposition step purely based on the provided
段落. Running EVV on a subset of examples
during data collection helps identify issues in the
pipeline and in worker performance.
4 The STRATEGYQA Dataset
We run our pipeline on 1,799 Wikipedia terms,
allowing a maximum of 5 questions per term. 我们
update our online fine-tuned solvers (FNTD) every
1K questions. Every question is decomposed once,
and evidence is matched for each decomposition
经过 3 different workers. The cost of annotating a
full example is $4.
To encourage diversity in strategies used
in the questions, we recruited new workers
throughout data collection. 而且, periodic
updates of the online solvers prevent workers
from exploiting shortcuts, since the solvers adapt
to the training distribution. 全面的, there were 29
question writers, 19 decomposers, 和 54 证据
matchers participating in the data collection.
We collected 2,835 问题, out of which 55
were marked as having an incorrect answer during
SQD (§3.2). This results in a collection of 2,780
verified strategy questions, for which we create an
annotator-based data split (Geva et al., 2019). 我们
now describe the dataset statistics (§4.1), analyze
the quality of the examples (§4.2), and explore the
reasoning skills in STRATEGYQA (§4.3).
4.1 Dataset Statistics
We observe (桌子 4) that the answer distribution
is roughly balanced (yes/no). 而且, 问题
are short (< 10 words), and the most common
trigram occurs in roughly 1% of the examples.
This indicates that the language of the questions
is both simple and diverse. For comparison,
the average question length in the multi-
hop datasets HOTPOTQA (Yang et al., 2018)
and COMPLEXWEBQUESTIONS (Talmor and Berant,
2018) is 13.7 words and 15.8 words, respectively.
Likewise, the top trigram in these datasets occurs
in 9.2% and 4.8% of their examples, respectively.
More than half of the generated questions are
filtered by our solvers, pointing to the difficulty
of generating good strategy questions. We release
all 3,305 filtered questions as well.
To characterize the reasoning complexity
required to answer questions in STRATEGYQA, we
examine the decomposition length and the number
of evidence paragraphs. Figure 3 and Table 4
(bottom) show the distributions of these properties
352
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
# of questions
% ‘‘yes’’ questions
# of unique terms
# of unique decomposition
steps
# of unique evidence
paragraphs
# of occurrences of the top
trigram
# of question writers
# of filtered questions
Avg. question length (words)
Avg. decomposition length
(steps)
Avg. # of paragraphs per
question
Test
490
Train
2290
46.8% 46.1%
1333
6050
442
1347
9251
2136
31
5
23
2821
9.6
2.93
6
484
9.8
2.92
2.33
2.29
Table 4: STRATEGYQA statistics. Filtered questions
were rejected by the solvers (§3.1). The train and
test sets of question writers are disjoint. The ‘‘top
trigram’’ is the most common trigram.
Figure 3: The distributions of decomposition length
(left) and the number of evidence paragraphs (right).
The majority of the questions in STRATEGYQA require
a reasoning process comprised of ≥ 3 steps, of which
about 2 steps involve retrieving external knowledge.
are centered around 3-step decompositions and 2
evidence paragraphs, but a considerable portion
of the dataset requires more steps and paragraphs.
4.2 Data Quality
Do questions in STRATEGYQA require multi-
step implicit reasoning? To assess the quality
of questions, we sampled 100 random examples
from the training set, and had two experts (authors)
independently annotate whether the questions
satisfy the desired properties of strategy questions
(§2.1). We find that most of the examples (81%)
are valid multi-step implicit questions, 82% of
353
implicit
explicit
multi-step
81
14.5
95.5
single-step
1
3.5
4.5
82
18
100
Table 5: Distribution over the implicit and
multi-step properties (§2) in a sample of 100
STRATEGYQA questions, annotated by two
experts (we average the expert decisions).
Most questions are multi-step and implicit.
Annotator agreement is substantial for both
the implicit (κ = 0.73) and multi-step
(κ = 0.65) properties.
questions are implicit, and 95.5% are multi-step
(Table 5).
Do questions in STRATEGYQA have a definitive
answer? We let experts review the answers to
100 random questions, allowing access to the
Web. We then ask them to state for every question
whether they agree or disagree with the provided
answer. We find that the experts agree with the
answer in 94% of the cases, and disagree only
in 2%. For the remaining 4%, either the question
was ambiguous, or the annotators could not find a
definite answer on the Web. Overall, this suggests
that questions in STRATEGYQA have clear answers.
What is the quality of the decompositions?
We randomly sampled 100 decompositions and
asked experts to judge their quality. Experts
judged if the decomposition is explicit or utilizes a
strategy. We find that 83% of the decompositions
validly use a strategy to break down the question.
The remaining 17% decompositions are explicit,
however, in 14% of the cases the original question
is already explicit. Second, experts checked if
the phrasing of the decomposition is ‘‘natural’’,
that is, it reflects the decomposition of a person
that does not already know the answer. We
find that 89% of the decompositions express a
‘‘natural’’ reasoning process, while 11% may
depend on the answer. Last, we asked experts
to indicate any potential
logical flaws in the
decomposition, but no such cases occurred in
the sample.
Would different annotators use the same
decomposition
50
sample
different workers
and
examples,
decompose
the
the
decomposition pairs, we find that a) for all pairs,
strategy? We
two
let
questions. Comparing
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
the last step returns the same answer, b) in 44
out of 50 pairs, the decomposition pairs follow the
same reasoning path, and c) in the other 6 pairs, the
decompositions either follow a different reasoning
process (5 pairs) or one of the decompositions is
explicit (1 pair). This shows that different workers
usually use the same strategy when decomposing
questions.
Is the evidence for strategy questions in
Wikipedia? Another
important property is
whether questions
in STRATEGYQA can be
answered based on context from our corpus,
Wikipedia, given that questions are written
independently of the context. To measure evidence
coverage, in the EVM task (§3.3), we provide
workers with a checkbox for every decomposition
step,
indicating whether only partial or no
evidence could be found for that step. Recall
that three different workers match evidence for
each decomposition step. We find that 88.3% of
the questions are fully covered: Evidence was
matched for each step by some worker. Moreover,
in 86.9% of the questions, at least one worker
found evidence for all steps. Last, in only 0.5%
of the examples were all three annotators unable
to match evidence for any of the steps. This
suggests that overall, Wikipedia is a good corpus
for questions in STRATEGYQA that were written
independently of the context.
Do matched paragraphs provide evidence?
We assess the quality of matched paragraphs
by analyzing both example-level and step-level
annotations. First, we sample 217 decomposition
paragraphs
steps with
matched by one of the three workers. We let
3 different crowdworkers decide whether the
paragraphs provide evidence for the answer to
that step. We find that in 93% of the cases, the
majority vote is that the evidence is valid.4
corresponding
their
succeeds. In the last 12 cases indeed evidence is
missing, and is possibly absent from Wikipedia.
Lastly, we let experts review the paragraphs
matched by one of the three workers to all the
decomposition steps of a question, for 100 random
questions. We find that for 79 of the questions the
matched paragraphs provide sufficient evidence
for answering the question. For 12 of the 21
questions without sufficient evidence, the experts
indicated they would expect to find evidence in
Wikipedia, and the worker probably could not find
it. For the remaining 9 questions, they estimated
that evidence is probably absent from Wikipedia.
In conclusion, 93% of the paragraphs matched
at the step-level were found to be valid. Moreover,
when considering single-worker annotations,
the questions are matched with
∼80% of
paragraphs that provide sufficient evidence for
all retrieval steps. This number increases to
88% when aggregating the annotations of three
workers.
by
for
compare
Do different annotators match the same
paragraphs? To
evidence
the
different
paragraphs matched
evidence
workers, we
a given
check whether
decomposition step, the same paragraph IDs are
retrieved by different annotators. Given two non-
empty sets of paragraph IDs P1, P2, annotated by
two workers, we compute the Jaccard coefficient
J(P1, P2) = |P1∩P2|
|P1∪P2| . In addition, we take the sets
of corresponding Wikipedia page IDs T1, T2 for
the matched paragraphs, and compute J(T1, T2).
Note that a score of 1 is given to two identical
sets, while a score of 0 corresponds to sets that are
disjoint. The average similarity score is 0.43 for
paragraphs and 0.69 for pages. This suggests that
evidence for a decomposition step can be found in
more than one paragraph in the same page, or in
different pages.
Next, we analyze annotations of the verification
task (§3.4), where workers are asked to answer all
decomposition steps based only on the matched
paragraphs. We find that the workers could answer
sub-questions and derive the correct answer in 82
out of 100 annotations. Moreover, in 6 questions
indeed there was an error in evidence matching,
but another worker who annotated the example
was able to compensate for the error, leading to
88% of the questions where evidence matching
4With moderate annotator agreement of κ = 0.42.
4.3 Data Diversity
We aim to generate creative and diverse questions.
We now analyze diversity in terms of the required
reasoning skills and question topic.
Reasoning Skills To explore the required
reasoning skills in STRATEGYQA, we sampled
100 examples and let
two experts (authors)
discuss and annotate each example with a) the
type of strategy for decomposing the question,
and b) the required reasoning and knowledge
354
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
Strategy
Physical
Biological
Historical
Temporal
Definition
Cultural
Religious
Example
Can human nails carve
a statue out of quartz?
Is a platypus immune
from cholera?
Were mollusks an
ingredient in the color
purple?
Did the 40th president
of the United States
forward lolcats to
his friends?
Are quadrupeds
represented on Chinese
calendar?
Would a compass
attuned to Earth’s
magnetic field be a bad
gift for a Christmas elf?
Was Hillary Clinton’s
deputy chief of staff in
2009 baptised?
Entertainment Would Garfield enjoy a
Sports
trip to Italy?
Can Larry King’s
ex-wives form a water
polo team?
%
13
11
10
10
8
5
5
4
4
Table 6: Top strategies in STRATEGYQA and their
frequency in a 100 example subset (accounting
for 70% of the analyzed examples).
Figure 4: Reasoning skills in STRATEGYQA; each skill is
associated with the proportion of examples it is required
for. Domain-related and logical reasoning skills are
marked in blue and orange (italic), respectively.
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
skills per decomposition step. We then aggregate
labels (e.g., botanical → biological)
similar
and compute the proportion of examples each
strategy/reasoning skill is required for (an example
can have multiple strategy labels).
Table 6 demonstrates
the top strategies,
showing that STRATEGYQA contains a broad set
of strategies. Moreover, diversity is apparent
in terms of both domain-related
(Figure 4)
reasoning (e.g., biological and technological) and
logical functions (e.g., set
inclusion and ‘‘is
member of’’). While the reasoning skills sampled
from questions in STRATEGYQA do not necessarily
reflect their prevalence in a ‘‘natural’’ distribution,
we argue that promoting research on methods
Figure 5: The top 15 categories of terms used to prime
workers for question writing and their proportion.
for inferring strategies is an important research
direction.
Question Topics As questions in STRATEGYQA
were triggered by Wikipedia terms, we use the
‘‘instance of’’ Wikipedia property to characterize
the topics of questions.5 Figure 5 shows the
distribution of topic categories in STRATEGYQA.
The distribution shows STRATEGYQA is very
diverse, with the top two categories (‘‘human’’
5It is usually a 1-to-1 mapping from a term to a Wikipedia
category. In cases of 1-to-many, we take the first category.
355
the
further
compare
diversity
and ‘‘taxon’’; i.e., a group of organisms) covering
only a quarter of the data, and a total of 609 topic
categories.
We
of
STRATEGYQA to HOTPOTQA, a multi-hop QA
dataset over Wikipedia paragraphs. To this end,
we sample 739 pairs of evidence paragraphs
associated with a single question in both datasets,
and map the pair of paragraphs to a pair of
Wikipedia categories using the ‘‘instance of’’
property. We find that
there are 571 unique
category pairs in STRATEGYQA, but only 356
unique category pairs in HOTPOTQA. Moreover,
the top two category pairs in both of the datasets
(‘‘human-human’’, ‘‘taxon-taxon’’) constitute 8%
and 27% of
the cases in STRATEGYQA and
HOTPOTQA, respectively. This demonstrates the
creativity and breadth of category combinations
in STRATEGYQA.
4.4 Human Performance
To see how well humans answer
strategy
questions, we sample a subset of 100 questions
from STRATEGYQA and have experts (authors)
answer questions, given access to Wikipedia
articles and an option to reveal the decomposition
for every question. In addition, we ask them to
provide a short explanation for the answer, the
number of searches they conducted to derive the
answer, and to indicate whether they have used
the decomposition. We expect humans to excel at
coming up with strategies for answering questions.
Yet, humans are not necessarily an upper bound
because finding the relevant paragraphs is difficult
and could potentially be performed better by
machines.
Table 7 summarizes
the results. Overall,
humans infer the required strategy and answer the
questions with high accuracy. Moreover, the low
number of searches shows that humans leverage
background knowledge, as they can answer some
of the intermediate steps without search. An error
analysis shows that the main reason for failure
(10%) is difficulty to find evidence, and the rest
of the cases (3%) are due to ambiguity in the
question that could lead to the opposite answer.
5 Experimental Evaluation
In this section, we conduct experiments to
answer the following questions: a) How well
do pre-trained language models (LMs) answer
356
Answer accuracy
Strategy match
Decomposition usage
Average # searches
87%
86%
14%
1.25
Table 7: Human performance in answering
questions. Strategy match is computed by
comparing the explanation provided by the expert
with the decomposition. Decomposition usage
and the number of searches are computed based
on information provided by the expert.
strategy questions? b) Is retrieval of relevant
context helpful? and c) Are decompositions useful
for answering questions that require implicit
knowledge?
5.1 Baseline Models
Answering strategy questions requires external
knowledge that cannot be obtained by training
on STRATEGYQA alone. Therefore, our models and
online solvers (§3.1) are based on pre-trained
LMs, fine-tuned on auxiliary datasets that require
reasoning. Specifically, in all models we fine-tune
ROBERTA (Liu et al., 2019) on a subset of:
• BOOLQ (Clark et al., 2019): A dataset for
Boolean question answering.
• MNLI (Williams et al., 2018): A large natural
language inference (NLI) dataset. The task
is to predict if a textual premise entails,
contradicts, or is neutral with respect to the
hypothesis.
• TWENTY QUESTIONS (20Q): A collection of
50K short commonsense Boolean questions.6
• DROP (Dua et al., 2019): A large dataset for
numerical reasoning over paragraphs.
Models are trained in two configurations:
• No context : The model is fed with the
question only, and outputs a binary prediction
using the special CLS token.
• With context : We use BM25 (Robertson
et al., 1995) to retrieve context from our
corpus, while removing stop words from all
queries. We examine two retrieval methods:
a) question-based retrieval: by using the
question as a query and taking the top
6https://github.com/allenai/twentyquestions.
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
k = 10 results, and b) decomposition-
based retrieval: by initiating a separate query
for each (gold or predicted) decomposition
step and concatenating the top k = 10
results of all steps (sorted by retrieval
is fed
score). In both cases,
with the question concatenated to the
retrieved context, truncated to 512 tokens
(the maximum input length of ROBERTA),
and outputs a binary prediction.
the model
Model
Solver group(s)
ROBERTA∅(20Q)
ROBERTA∅(20Q+BOOLQ)
ROBERTA∅(BOOLQ)
ROBERTAIR-Q(BOOLQ)
ROBERTAIR-Q(MNLI+BOOLQ)
PTD, FNTD
PTD, FNTD
PTD, FNTD
PTD
PTD
Table 8: QA models used as online solvers during
data collection (§3.1). Each model was fine-tuned
on the datasets mentioned in its name.
Predicting Decompositions We train a seq-
to-seq model, termed BARTDECOMP, that, given a
question, generates its decomposition token-by-
token. Specifically, we fine-tune BART (Lewis
et al., 2020) on STRATEGYQA decompositions.
Baseline Models As our base model, we train
a model as follows: We take a ROBERTA (Liu
et al., 2019) model and fine-tune it on DROP, 20Q
and BOOLQ (in this order). The model is trained
on DROP with multiple output heads, as in Segal
et al. (2020), which are then replaced with a single
Boolean output.7 We call this model ROBERTA*.
We use ROBERTA* and ROBERTA to train
the following models on STRATEGYQA: without
(ROBERTA*∅), with question-based
context
retrieval
(ROBERTA*IR-Q, ROBERTAIR-Q), and
with predicted decomposition-based retrieval
(ROBERTA*IR-D).
We also present four oracle models:
• ROBERTA*ORA-P: Uses the gold paragraphs
(no retrieval).
• ROBERTA*IR-ORA-D: Performs retrieval with
the gold decomposition.
• ROBERTA*last-step
ORA-P-D: Exploits both the gold
decomposition and the gold paragraphs. We
fine-tune ROBERTA on BOOLQ and SQUAD
(Rajpurkar et al., 2016) to obtain a model
that can answer single-step questions. We
then run this model on STRATEGYQA to obtain
answers for all decomposition sub-questions,
and replace all placeholder references with
7For brevity, exact details on model
training and
hyper-parameters will be released as part of our codebase.
Model
MAJORITY
ROBERTA*∅
ROBERTAIR-Q
ROBERTA*IR-Q
ROBERTA*IR-D
ROBERTA*IR-ORA-D
ROBERTA*ORA-P
ROBERTA*last-step-raw
ORA-P-D
ROBERTA*last-step
ORA-P-D
Accuracy Recall@10
53.9
63.6 ± 1.3
53.6 ± 1.0
63.6 ± 1.0
61.7 ± 2.2
62.0 ± 1.3
70.7 ± 0.6
65.2 ± 1.4
72.0 ± 1.0
-
-
0.174
0.174
0.195
0.282
-
-
-
Table 9: QA accuracy (with standard deviation
across 7 experiments), and retrieval performance,
measured by Recall@10, of baseline models on
the test set.
the predicted answers. Last, we fine-tune
ROBERTA* to answer the last decomposition
step of STRATEGYQA, for which we have
supervision.
• ROBERTA*last-step-raw
ORA-P-D : ROBERTA* that
is
fine-tuned to predict
from
the gold paragraphs and the last step of
the gold decomposition, without replacing
placeholder references.
the answer
Online Solvers For the solvers integrated in the
data collection process (§3.1), we use three no-
context models and two question-based retrieval
models. The solvers are listed in Table 8.
5.2 Results
Strategy QA performance Table 9 summarizes
the results of all models (§5.1). ROBERTA*IR-Q
substantially outperforms ROBERTAIR-Q, indicat-
ing that fine-tuning on related auxiliary datasets
before STRATEGYQA is crucial. Hence, we focus
on ROBERTA* for all other results and analysis.
Strategy questions pose a combined challenge
of retrieving the relevant context, and deriving
the answer based on that context. Training
357
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
without context shows a large accuracy gain of
53.9 → 63.6 over the majority baseline. This
is far from human performance, but shows that
some questions can be answered by a large LM
fine-tuned on related datasets without retrieval.
On the other end, training with gold paragraphs
raises performance to 70.7. This shows that
high-quality retrieval lets the model effectively
reason over the given paragraphs. Last, using
both gold decompositions and retrieval further
increases performance to 72.0, showing the utility
of decompositions.
Focusing on retrieval-based methods, we
observe that question-based retrieval
reaches
an accuracy of 63.6 and retrieval with gold
decompositions results in an accuracy of 62.0. This
shows that the quality of retrieval even with gold
decompositions is not high enough to improve the
63.6 accuracy obtained by ROBERTA*∅, a model
that uses no context. Retrieval with predicted
decompositions results in an even lower accuracy
of 61.7. We also analyze predicted decompositions
below.
Retrieval Evaluation A question decomposi-
tion describes the reasoning steps for answering
the question. Therefore, using the decomposi-
tion for retrieval may help obtain the relevant
context and improve performance. To test this,
we directly compare performance of question-
and decomposition-based retrieval with respect
to the annotated gold paragraphs. We compute
Recall@10, that is, the fraction of the gold para-
graphs retrieved in the top-10 results of each
method. Since there are 3 annotations per ques-
tion, we compute Recall@10 for each annotation
and take the maximum as the final score. For a fair
comparison, in decomposition-based retrieval, we
use the top-10 results across all steps.
Results (Table 9) show that retrieval per-
low, partially explaining why
formance is
retrieval models do not
improve performance
compared to ROBERTA*∅, and demonstrating
the retrieval challenge in our
setup. Gold
decomposition-based retrieval substantially out-
performs question-based retrieval, showing that
using the decomposition for retrieval is a promis-
ing direction for answering multi-step questions.
Still, predicted decomposition-based retrieval
does not improve retrieval compared to question-
based retrieval, showing better decomposition
models are needed.
To understand the low retrieval scores, we
analyzed the query results of 50 random
decomposition steps. Most
failure cases are
due to the shallow pattern matching done by
BM25—for example, failure to match synonyms.
indeed there is little word
This shows that
overlap between decomposition steps and the
evidence, as intended by our pipeline design. In
other examples, either a key question entity was
missing because it was represented by a reference
token, or the decomposition step had complex
language, leading to failed retrieval. This analysis
suggests that advances in neural retrieval might
be beneficial for STRATEGYQA.
Human Retrieval Performance To quantify
human performance in finding gold paragraphs,
we ask experts to find evidence paragraphs for
100 random questions. For half of the questions
we also provide decomposition. We observe
average Recall@10 of 0.586 and 0.513 with
and without the decomposition, respectively. This
shows that humans significantly outperform our
IR baselines. However, humans are still far from
covering the gold paragraphs, since there are
multiple valid evidence paragraphs (§4.2), and
retrieval can be difficult even for humans. Lastly,
using decompositions improves human retrieval,
showing decompositions indeed are useful for
finding evidence.
Interestingly,
Predicted Decompositions Analysis shows that
BARTDECOMP’s decompositions are grammatical
and well-structured.
the model
generates strategies, but often applies them to
questions incorrectly. For example, the question
Can a lifeboat rescue people in the Hooke Sea? is
decomposed to 1) What is the maximum depth of
the Hooke Sea? 2) How deep can a lifeboat dive?
3) Is #2 greater than or equal to #1?. While the
decomposition is well-structured, it uses a wrong
strategy (lifeboats do not dive).
6 Related Work
Prior work has typically let annotators write
questions based on an entire context (Khot et al.,
2020a; Yang et al., 2018; Dua et al., 2019;
Mihaylov et al., 2018; Khashabi et al., 2018).
In this work, we prime annotators with minimal
information (few tokens) and let them use their
358
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
imagination and own wording to create questions.
A related priming method was recently proposed
by Clark et al. (2020), who used the first 100
characters of a Wikipedia page.
Among multi-hop reasoning datasets, our
it requires implicit
in that
dataset stands out
decompositions. Two recent datasets (Khot et al.,
2020a; Mihaylov et al., 2018) have considered
questions requiring implicit facts. However, they
are limited to specific domain strategies, while in
our work we seek diversity in this aspect.
Most multi-hop reasoning datasets do not fully
annotate question decomposition (Yang et al.,
2018; Khot et al., 2020a; Mihaylov et al., 2018).
This issue has prompted recent work to create
question decompositions for existing datasets
(Wolfson et al., 2020), and to train models that
generate question decompositions (Perez et al.,
2020; Khot et al., 2020b; Min et al., 2019). In
this work, we annotate question decompositions
as part of the data collection.
7 Conclusion
We present STRATEGYQA,
the first dataset of
implicit multi-step questions requiring a wide-
range of reasoning skills. To build STRATEGYQA,
we introduced a novel annotation pipeline for
eliciting creative questions
that use simple
language, but cover a challenging range of
diverse strategies. Questions in STRATEGYQA are
annotated with decomposition into reasoning steps
and evidence paragraphs, to guide the ongoing
research towards addressing implicit multi-hop
reasoning.
Acknowledgments
We thank Tomer Wolfson for helpful feedback
and the REVIZ team at Allen Institute for
AI, particularly Michal Guerquin and Sam
Skjonsberg. This research was supported in part
by the Yandex Initiative for Machine Learning,
and the European Research Council (ERC) under
the European Union Horizons 2020 research
and innovation programme (grant ERC DELPHI
802800). Dan Roth is partly supported by ONR
contract N00014-19-1-2620 and DARPA contract
FA8750-19-2-1004, under the Kairos program.
This work was completed in partial fulfillment for
the PhD degree of Mor Geva.
References
Max Bartolo, Alastair Roberts, Johannes Welbl,
Sebastian Riedel,
and Pontus Stenetorp.
2020. Beat the AI: Investigating adversarial
reading comprehen-
human annotation for
sion. Transactions of
the Association for
Computational Linguistics, 8:662–678. DOI:
https://doi.org/10.1162/tacl a
00338
Christopher Clark, Kenton Lee, Ming-Wei Chang,
Tom Kwiatkowski, Michael Collins, and
Kristina Toutanova. 2019. BoolQ: Exploring
the surprising difficulty of natural yes/no
the 2019
In Proceedings of
questions.
Conference of the North American Chapter of
the Association for Computational Linguistics:
Human Language Technologies, Volume 1
(Long and Short Papers), pages 2924–2936,
for
Minneapolis, Minnesota. Association
Computational Linguistics.
Jonathan H. Clark, Eunsol Choi, Michael
Collins, Dan Garrette, Tom Kwiatkowski,
Vitaly Nikolaev, and Jennimaria Palomaki.
2020. TyDi QA: A benchmark for information-
seeking question answering in typologically
diverse languages. Transactions of the Asso-
ciation for Computational Linguistics (TACL),
8:454–470. DOI: https://doi.org/10
.1162/tacl a 00317
Jay DeYoung, Sarthak Jain, Nazneen Fatema
Rajani, Eric Lehman, Caiming Xiong, Richard
Socher, and Byron C. Wallace. 2020. ERASER:
A benchmark to evaluate rationalized NLP
models. In Proceedings of the 58th Annual
Meeting of the Association for Computational
Linguistics, pages 4443–4458. Association for
Computational Linguistics. DOI: https://
doi.org/10.18653/v1/2020.acl-main
.408
Dheeru Dua, Yizhong Wang, Pradeep Dasigi,
Gabriel Stanovsky, Sameer Singh,
and
Matt Gardner. 2019. DROP: A reading
comprehension benchmark requiring discrete
reasoning over paragraphs. In North American
Chapter of the Association for Computational
Linguistics (NAACL).
Mor Geva, Yoav Goldberg, and Jonathan Berant.
the
2019. Are we modeling the task or
359
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
annotator? An investigation of annotator bias
language understanding datasets.
in natural
the 2019 Conference on
In Proceedings of
Empirical Methods
in Natural Language
Processing and the 9th International Joint
Conference on Natural Language Processing
(EMNLP-IJCNLP), pages 1161–1166, Hong
Kong, China. Association for Computational
Linguistics.
the 2018 Conference of
Suchin Gururangan, Swabha Swayamdipta, Omer
Levy, Roy Schwartz, Samuel Bowman, and
Noah A. Smith. 2018. Annotation artifacts
in natural language inference data. In Pro-
ceedings of
the
North American Chapter of the Association
for Computational Linguistics: Human Lan-
guage Technologies, Volume 2 (Short Papers),
pages 107–112, New Orleans, Louisiana.
Association for Computational Linguistics.
DOI: https://doi.org/10.18653/v1
/N18-2017
Yichen
Jiang
and Mohit Bansal.
2019.
Avoiding reasoning shortcuts: Adversarial
evaluation, training, and model development
for multi-hop QA. In Association for Compu-
tational Linguistics (ACL). DOI: https://
doi.org/10.18653/v1/P19-1262,
PMID: 31353678
Daniel Khashabi, Snigdha Chaturvedi, Michael
Roth, Shyam Upadhyay, and Dan Roth. 2018.
Looking beyond the surface: A challenge
set for reading comprehension over multi-
the 2018
ple sentences. In Proceedings of
Conference of
the North American Chap-
ter of
the Association for Computational
Linguistics: Human Language Technologies,
Volume 1 (Long Papers), pages 252–262,
New Orleans, Louisiana. Association for Com-
putational Linguistics, DOI: https://doi
.org/10.18653/v1/N18-1023
Tushar Khot, Peter Clark, Michal Guerquin,
Peter Jansen, and Ashish Sabharwal. 2020a.
for question answering
QASC: A dataset
In AAAI. DOI:
via sentence composition.
https://doi.org/10.1609/aaai.v34i05
.6319
Tushar Khot, Daniel Khashabi, Kyle Richardson,
Peter Clark, and Ashish Sabharwal. 2020b.
360
Text modular networks: Learning to decompose
tasks in the language of existing models. arXiv
preprint arXiv:2009.00751.
Tom Kwiatkowski, Jennimaria Palomaki, Olivia
Redfield, Michael Collins, Ankur Parikh, Chris
Alberti, Danielle Epstein,
Illia Polosukhin,
Jacob Devlin, Kenton Lee, Kristina Toutanova,
Llion Jones, Matthew Kelcey, Ming-Wei
Chang, Andrew M. Dai, Jakob Uszkoreit,
Quoc Le, and Slav Petrov. 2019. Natural
questions: A benchmark for question answering
research. Transactions of the Association for
Computational Linguistics (TACL), 7:453–466.
DOI: https://doi.org/10.1162/tacl
a 00276
Mike Lewis, Yinhan Liu, Naman Goyal, Marjan
Ghazvininejad, Abdelrahman Mohamed, Omer
Levy, Veselin Stoyanov, and Luke Zettlemoyer.
2020. BART: Denoising sequence-to-sequence
pre-training for natural language generation,
translation, and comprehension. In Proceedings
of the 58th Annual Meeting of the Association
for Computational Linguistics, pages 7871–7880.
Association for Computational Linguistics.
DOI: https://doi.org/10.18653/v1
/2020.acl-main.703
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei
Du, Mandar
Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and
Veselin Stoyanov. 2019. RoBERTa: A robustly
optimized BERT pretraining approach. arXiv
preprint arXiv:1907.11692.
for
armor
Todor Mihaylov, Peter Clark, Tushar Khot,
2018. Can
Sabharwal.
a
and Ashish
electricity? A
conduct
suit
of
question
book
open
new dataset
the 2018
In Proceedings of
answering.
Conference on Empirical Methods in Natural
Language Processing,
2381–2391,
Brussels, Belgium. Association for Compu-
tational Linguistics. DOI: https://doi
.org/10.18653/v1/D18-1260
pages
Sewon Min, Victor Zhong, Luke Zettlemoyer,
and Hannaneh Hajishirzi. 2019. Multi-hop
reading
question
decomposition and rescoring. In Proceedings of
the 57th Annual Meeting of the Association for
Computational Linguistics, pages 6097–6109,
comprehension
through
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
pages 6418–6428, Florence,
Italy. Associ-
ation for Computational Linguistics. DOI:
https://doi.org/10.18653/v1/P19
-1644
Alon Talmor and Jonathan Berant. 2018. The web
as a knowledge-base for answering complex
In North American Chapter of
questions.
the Association for Computational Linguistics
(NAACL). DOI: https://doi.org/10
.18653/v1/N18-1059
Johannes Welbl, Pontus Stenetorp, and Sebastian
Riedel. 2018. Constructing datasets for multi-
hop reading comprehension across documents.
Transactions of the Association for Computa-
tional Linguistics (TACL), 6:287–302. DOI:
https://doi.org/10.1162/tacl a
00021
Adina Williams, Nikita Nangia, and Samuel
Bowman. 2018. A broad-coverage challenge
corpus for sentence understanding through
2018
In Proceedings of
inference.
Conference of the North American Chapter of
the Association for Computational Linguistics:
Human Language Technologies, Volume 1
(Long Papers), pages 1112–1122, New Orleans,
Louisiana, Association for Computational
Linguistics. DOI: https://doi.org/10
.18653/v1/N18-1101
the
Tomer Wolfson, Mor Geva, Ankit Gupta, Matt
Gardner, Yoav Goldberg, Daniel Deutch,
and Jonathan Berant. 2020. Break it down:
A question understanding benchmark. Trans-
actions of the Association for Computational
(TACL), DOI: https://doi
Linguistics
.org/10.1162/tacl a 00309
Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua
Bengio, William Cohen, Ruslan Salakhutdinov,
and Christopher D. Manning. 2018. HotpotQA:
A dataset for diverse, explainable multi-hop
In Empirical Methods
question answering.
in Natural Language Processing (EMNLP),
DOI: https://doi.org/10.18653/v1
/D18-1259, PMCID: PMC6156886
Florence,
for Com-
Italy. Association
putational Linguistics. DOI: https://
doi.org/10.18653/v1/P19-1613
Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng
Gao, Saurabh Tiwary, Rangan Majumder, and
Li Deng. 2016. MS MARCO: A human
generated machine reading comprehension
dataset. In Workshop on Cognitive Computing
at NIPS.
Ethan Perez, Patrick Lewis, Wen-tau Yih,
Kyunghyun Cho, and Douwe Kiela. 2020.
Unsupervised question decomposition for
question answering. In Proceedings of
the
2020 Conference on Empirical Methods
in Natural Language Processing (EMNLP),
pages 8864–8880.
Pranav Rajpurkar,
Jian Zhang, Konstantin
Lopyrev, and Percy Liang. 2016. SQuAD:
100,000+ questions for machine comprehen-
sion of text. In Empirical Methods in Natural
(EMNLP). DOI:
Language
https://doi.org/10.18653/v1/D16
-1264
Processing
Stephen Robertson, S. Walker, S. Jones, M. M.
Hancock-Beaulieu, and M. Gatford. 1995.
the
Okapi
Third Text REtrieval Conference (TREC-3),
pages 109–126. Gaithersburg, MD: NIST.
In Overview of
at TREC-3.
Elad Segal, Avia Efrat, Mor Shoham, Amir
Globerson, and Jonathan Berant. 2020. A
simple and effective model
for answering
multi-span questions. In Proceedings of the
2020 Conference on Empirical Methods in
Natural Language Processing
(EMNLP),
pages 3074–3080. DOI: https://doi.org
/10.18653/v1/2020.emnlp-main.248
Alane Suhr, Stephanie Zhou, Ally Zhang,
Iris Zhang, Huajun Bai, and Yoav Artzi.
2019. A corpus for reasoning about natural
language grounded in photographs. In Pro-
ceedings of the 57th Annual Meeting of the
Association for Computational Linguistics,
361
l
D
o
w
n
o
a
d
e
d
f
r
o
m
h
t
t
p
:
/
/
d
i
r
e
c
t
.
m
i
t
.
e
d
u
/
t
a
c
l
/
l
a
r
t
i
c
e
-
p
d
f
/
d
o
i
/
.
1
0
1
1
6
2
/
t
l
a
c
_
a
_
0
0
3
7
0
1
9
2
4
1
0
4
/
/
t
l
a
c
_
a
_
0
0
3
7
0
p
d
.
f
b
y
g
u
e
s
t
t
o
n
0
7
S
e
p
e
m
b
e
r
2
0
2
3
