ABNIRML: Analyzing the Behavior of Neural IR Models
Sean MacAvaney†∗
Sergey Feldman‡
Nazli Goharian†
Doug Downey‡
Arman Cohan‡§
†IR Lab, Georgetown University, Washington, DC, USA
‡Allen Institute for AI, Seattle, WA, USA
§Paul G. Allen School of Computer Science, University of Washington, WA, USA
{sean,nazli}@ir.cs.georgetown.edu
{sergey,dougd,armanc}@allenai.org
Astratto
Pretrained contextualized language models
such as BERT and T5 have established a
new state-of-the-art for ad-hoc search. How-
ever, it is not yet well understood why these
methods are so effective, what makes some
variants more effective than others, and what
pitfalls they may have. We present a new
comprehensive framework for Analyzing the
Behavior of Neural IR ModeLs (ABNIRML),
which includes new types of diagnostic probes
that allow us to test several characteristics—
such as writing styles, factuality, sensitivity
to paraphrasing and word order—that are not
addressed by previous techniques. To demon-
strate the value of the framework, we conduct
an extensive empirical study that yields in-
sights into the factors that contribute to the
neural model’s gains, and identify potential
unintended biases the models exhibit. Some
of our results confirm conventional wisdom,
Per esempio, that recent neural ranking mod-
els rely less on exact term overlap with the
query, and instead leverage richer linguistic
informazione, evidenced by their higher sen-
sitivity to word and sentence order. Other
results are more surprising, such as that some
models (per esempio., T5 and ColBERT) are biased
towards factually correct (rather than simply
relevant) texts. Further, some characteristics
vary even for the same base language model,
and other characteristics can appear due to
random variations during model training.1
1 introduzione
Pre-trained contextualized language models such
as BERT (Devlin et al., 2019) are state-of-the-art
for a wide variety of natural language processing
compiti (Xia et al., 2020). In Information Retrieval
∗Currently at the University of Glasgow. Work done in
part during an internship at the Allen Institute for AI.
1Code: https://github.com/allenai/abnriml.
224
large
these models have brought about
(IR),
improvements in the task of ad-hoc retrieval—
ranking documents by their relevance to a tex-
tual query (Lin et al., 2020; Nogueira and Cho,
2019; MacAvaney et al., 2019UN; Dai and Callan,
2019B)—where the models increasingly dominate
competition leaderboards (Craswell et al., 2019;
Dalton et al., 2019).
Despite this success, little is understood about
why pretrained language models are effective for
ad-hoc ranking. Previous work has shown that
traditional IR axioms, Per esempio, that increased
term frequency should correspond to higher rel-
evance, do not explain the behavior of recent
neural models (Cˆamara and Hauff, 2020). Outside
of IR, others have examined what characteristics
contextualized language models learn in general
(Liu et al., 2019UN; Rogers et al., 2020; Loureiro
it remains unclear if these
et al., 2020), Ma
qualities are valuable for ad-hoc ranking specif-
ically. Così, new approaches are necessary to
characterize models.
We propose a new framework aimed at An-
alyzing the Behavior of Neural
IR ModeLs
(ABNIRML2), which aims to probe the sensitivity
of ranking models on specific textual properties.
Probes consist of samples comprising a query
and two contrastive documents. We propose three
strategies for building probes. The ‘‘measure and
match’’ strategy (akin to the diagnostic datasets
proposed by Rennings et al. [2019]) constructs
probing samples by controlling one measurement
(per esempio., term frequency) and varying another (per esempio.,
document length) using samples from an existing
IR collection. Unlike Rennings et al. (2019), our
framework generalizes the idea to any measurable
characteristic, rather than relying chiefly on prior
proposed IR axioms. A second strategy, ‘‘textual
2Pronounced /ab’n@rm@l/ , similar to ‘‘abnormal’’.
Operazioni dell'Associazione per la Linguistica Computazionale, vol. 10, pag. 224–239, 2022. https://doi.org/10.1162/tacl a 00457
Redattore di azioni: Tim Baldwin. Lotto di invio: 7/2021; Lotto di revisione: 10/2021; Pubblicato 3/2022.
C(cid:3) 2022 Associazione per la Linguistica Computazionale. Distribuito sotto CC-BY 4.0 licenza.
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
4
5
7
2
0
0
2
6
9
8
/
/
T
l
UN
C
_
UN
_
0
0
4
5
7
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3
manipulation,’’ probes the effect that altering the
text of a document text has on its ranking. Finalmente,
a ‘‘dataset transfer’’ strategy constructs probes
from non-IR datasets. The new probes allow us
to isolate model characteristics—such as sensi-
tivity to word order, degree of lexical simplicity,
or even factuality—that cannot be analyzed using
other approaches.
Using our new framework, we perform the first
large-scale analysis of neural IR models. Noi veniamo-
pare today’s leading ranking techniques, including
those using BERT (Devlin et al., 2019) and T5
(Raffel et al., 2020), methods focused on effi-
ciency like DocT5Query (Nogueira et al., 2020)
and EPIC (MacAvaney et al., 2020), and dense re-
trieval models like ANCE (Xiong et al., 2021)
and ColBERT (Khattab and Zaharia, 2020).3
Some of our results establish widely believed, Ma
not-yet-verified, conjectures about neural mod-
els. Per esempio, we show that neural models
can exploit richer linguistic signals than classical
term-matching metrics like BM25: When control-
ling for term frequency match, the neural models
detect document relevance much more accurately
than the BM25 baseline. Allo stesso modo, unlike prior
approcci, rankers based on BERT and T5 are
heavily influenced by word order: Shuffling the
words in a document consistently lowers the doc-
ument’s score relative to the unmodified version,
and neural rankers show a sensitivity to sentence
order that is completely absent in classical mod-
els. Other findings from ABNIRML are more
surprising. Per esempio, we find that the T5 and
ColBERT models we examine prefer answers that
are factually correct, implying that they encode
and utilize some real-world knowledge. Further,
although this knowledge may be a result of the
model’s pre-training process, it is not necessarily
utilized as a ranking signal, given that other mod-
els that use the same base language model do not
have the same preference. Our battery of probes
also uncover a variety of other findings, including
that adding additional text to documents can of-
ten exhibit adverse behavior in neural models—
decreasing the document’s score when the added
text is relevant, and increasing the score when the
added text is irrelevant.
3Although a multitude of other models exist, it is im-
practical to investigate them all. We instead focus on a
representative sample of the recent and successful models
and well-known baselines to provide context.
In summary, we present a new framework
(ABNIRML) for performing analysis of ad-hoc
ranking models. We then demonstrate how the
framework can provide insights into ranking model
characteristics by providing the most comprehen-
sive analysis of neural ranking models to date.
Our software implementation of the framework is
easily extensible, facilitating the replication of our
results and further analyses in future work.
2 ABNIRML
In order to characterize the behavior of ranking
models we construct several diagnostic probes.
Each probe aims to evaluate specific properties
of ranking models and probe their behavior (if
they are heavily influenced by term matching,
discourse and coherence, conciseness/verbosity,
writing styles, eccetera.). We formulate three differ-
ent approaches to construct probes (Measure and
Match, Textual Manipulation, and Dataset Transfer).
In ad-hoc ranking, a query (expressed in natural
lingua) is submitted by a user to a search engine,
and a ranking function provides the user with a list
of natural language documents sorted by relevance
to the query. More formally, let R(q, D) ∈ R be
a ranking function, which maps a given query q
and document d (each being a natural-language
sequence of terms) to a real-valued ranking score.
At query time, documents in a collection D are
scored using R(·) for a given query q, and ranked
by the scores (conventionally, sorted descending
by score). Learning-to-rank models optimize a set
of parameters for the task of relevance ranking
based on training data.
2.1 Document Pair Probing
We utilize a document pair probing strategy, In
which probes are composed of samples, each of
which consists of a query and two documents that
differ primarily in some characteristic of interest
(per esempio., succinctness). The ranking scores of the two
documents are then compared (with respect to
the query). This allows the isolation of particular
model preferences. For instance, a probe could
consist of summarized and full texts of news
articles; models that consistently rank summaries
over full texts prefer succinct text.
More formally, each document pair probe con-
sists of a collection of samples S, where each
(cid:5)q, d1, d2(cid:6) ∈ S is a 3-tuple consisting of a query
(or query-like text, q), and two documents (O
225
l
D
o
w
N
o
UN
D
e
D
F
R
o
M
H
T
T
P
:
/
/
D
io
R
e
C
T
.
M
io
T
.
e
D
tu
/
T
UN
C
l
/
l
UN
R
T
io
C
e
–
P
D
F
/
D
o
io
/
.
1
0
1
1
6
2
/
T
l
UN
C
_
UN
_
0
0
4
5
7
2
0
0
2
6
9
8
/
/
T
l
UN
C
_
UN
_
0
0
4
5
7
P
D
.
F
B
sì
G
tu
e
S
T
T
o
N
0
7
S
e
P
e
M
B
e
R
2
0
2
3
Figura 1: Overview of strategies for constructing probes. Each probe in ABNIRML is composed of samples, each
of which consists of a query (q) and two documents (d1 and d2).
document-like texts, d1 and d2). The relation-
ship between d1 and d2 (with respect to q) for
each sample defines the probe. Per esempio, UN
probe testing summarization could be defined as:
(1) d2 is a summary of d1, E (2) d1 is relevant to
query q.
Almost all of our probes are directional, Dove
d2 has some attribute that d1 lacks, and we measure
the effect of this attribute on ranking. Specifically,
each sample in the probe is scored as: (+1) scoring
d1 above d2 (a positive effect), (−1) scoring d2
above d1 (a negative effect), O (0) a neutral effect.
Formalmente, the effect eff (·) of a given sample is
defined as:
eff (q, d1, d2) =
(cid:2)
1 R(q, d1) − R(q, d2) > δ
−1 R(q, d1) − R(q, d2) < −δ
0 −δ ≤ R(q, d1) − R(q, d2) ≤ δ
(1)
The parameter δ adjusts how large the score dif-
ference between the scores of d1 and d2 must be in
order to count as positive or negative effect. This
allows us to disregard small changes to the score
that are unlikely to affect the final ranking. In
practice, δ depends on the ranking model because
each model scores on different scales. Therefore
we tune δ for each model (see Section 3.3).
Symmetric probes are different from directional
ones in that d1 and d2 are exchangeable; for ex-
ample, we experiment with one symmetric probe
in which d1 and d2 are paraphrases of each other.
For symmetric probes only the magnitude of score
difference is meaningful, and thus eff outputs 1 if
the absolute value of the difference is larger than
δ, and 0 otherwise.
A model’s performance on a particular probe is
summarized by a single score s that averages the
effect of all samples in the probe:
s =
1
|S|
(cid:3)
(cid:5)q,d1,d2(cid:6)∈S
eff (q, d1, d2)
(2)
Note that this score is in the interval [−1, 1]
for directional probes and [0, 1] for symmetric
probes. For directional probes, positive scores
indicate a stronger preference towards documents
from group 1 (d1 documents), and negative scores
indicate a preference towards documents from
group 2 (d2 documents). Scores near 0 indicate
no strong preference or preferences that are split
roughly evenly; disentangling these two cases
requires analyzing individual effect scores.
There are several important differences between
our setup and the ‘‘diagnostic dataset’’ approach
proposed by Rennings et al. (2019). First, by
including the δ threshold, we ensure that our
probes measure differences that can affect the fi-
nal order in ranked lists. Second, by including
the ‘‘neutral effect’’ case in our scoring function,
we distinguish between cases in which d1 or d2
are preferred and cases where neither document
is strongly preferred. And finally, our probes are
aimed at describing model behavior, rather than
evaluating models. For instance, one of our tests
measures whether the model prefers succinct or
elaborative text—whether this preference is de-
sirable depends on the application or even the
particular user.
2.2 Document Pair Probing Strategies
In this work, we explore three strategies for design-
ing document pair probes. As discussed below,
the strategies have different strengths and weak-
nesses. When used in concert, they allow us to
characterize a wide variety of model behaviors.
Figure 1 provides an overview of the strategies.
2.2.1 Measure and Match Probes (MMPs)
Some surface-level characteristics of documents,
such as its Term Frequency (TF) for a given query,
are both easy to measure and valuable for char-
acterizing models. Comparing the ranking scores
226
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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 two documents that differ in one characteris-
tic but are otherwise similar yields evidence of
how the characteristic influences model behavior.
Measure and Match Probes (MMPs) follow such
an approach. MMPs involve first measuring the
characteristics of judged query-document pairs in
an IR dataset. Then, the pairs are matched to form
probe samples consisting of a control (a charac-
teristic that approximately matches between the
documents, such as document length), and a vari-
able (which differs between documents, such as
TF). Probes used in previous work to verify ex-
isting ranking axioms (Cˆamara and Hauff, 2020;
Rennings et al., 2019)4 are instances of MMPs.
For our experiments, we design MMPs to ex-
plore the relationship between the primary IR
objective (document relevance) and the classical
IR ranking signal (TF, potentially controlling for
document length). We are motivated to explore
this relationship because TF has long been used
as a core signal for ranking algorithms; a depar-
ture from monotonically increasing the score of a
document as TF increases would represent a fun-
damental shift in the notion of relevance scoring
(Fang et al., 2004). Specifically, we explore the
following characteristics in MMPs:
• Relevance: the human-assessed graded rel-
evance score of a document to the given
query.
• Length: the document length, in total number
of non-stopword tokens.
• TF: the individual Porter-stemmed Term Fre-
quencies of non-stopword terms from the
query. To determine when the TF of two
documents are different, we use the condi-
tion that the TF of at least one query term in
d1 must be greater than the same term in d2,
and that no term in d1 can have a lower TF
than the corresponding term in d2.
• Overlap:
the proportion of non-stopword
terms in the document that appear in the
query. Put another way, the total TF divided
by the document length.
Each of these characteristics is used as both a
variable (matching based on differing values) and
a control (matching based on identical values). In
our experiments, we examine all pairs of these
4An example is TFC1 from Fang et al. (2004), which
suggests that higher TFs should be mapped to higher scores.
characteristics, greatly expanding upon IR axioms
investigated in prior work.
We note that the MMPs that we explore in this
work do not cover all prior IR axioms. For instance,
axioms SMTC1–3, proposed by Fang and Zhai
(2006), suggest behaviors related to the occurrence
of semantically similar terms. Although MMPs
can be constructed to test these, we assert that
other types of probes are more suitable to testing
these behaviors. We test textual fluency, formality,
and simplicity (all of which are specific types
of semantic similarity) while controlling for the
meaning of the text using dataset transfer probes
(Section 2.2.3).
2.2.2 Textual Manipulation Probes (TMPs)
Not all characteristics are easily captured with
MMPs. For instance, it would be difficult to probe
the sensitivity to word order with MMPs; it is un-
likely to find naturally occurring document pairs
that use the same words but in a different or-
der. Nevertheless, it is valuable to understand the
extent to which models are affected by character-
istics like this, given that traditional bag-of-words
models are unaffected by word order and that there
is evidence that word order is unimportant when
fine-tuning recent neural models (Sinha et al., 2021;
Alleman et al., 2021). To overcome these limi-
tations, we propose Textual Manipulation Probes
(TMPs). TMPs apply a manipulation function to
scored documents from an existing IR dataset.
For example, for probing word order, we can
use a simple manipulation function that, given a
document d1, creates a corresponding synthetic
document d2 by shuffling the order of the words
in each sentence. The degree to which a model
prefers d1 is then a measure of its preference for
proper word order. Prior work that uses a similar
approach for probing ranking methods include the
collection perturbation tests of Fang et al. (2011)
(which perform operations like removing docu-
ments from the collection and deleting individual
terms from documents) and a diagnostic dataset
proposed by Rennings et al. (2019) (which tests
the effect of duplicating the document: an adap-
tation of a traditional ranking axiom). Although
TMPs allow probing a wider variety of charac-
teristics than MMPs, we note that they involve
constructing artificial data; d2 may not resemble
documents seen in practice. Despite this, their
versatility make TMPs an attractive choice for a
variety of characteristics.
227
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
We now detail the specific TMPs we explore
in our experiments. We use TMPs to verify a key
difference we expect to hold between neural mod-
els and previous rankers: Because neural models
are pretrained on large bodies of running text,
they should make better use of richer linguistic
features like word order. We investigate this with
TMPs that shuffle words in the document. We
also probe which aspects of word order are im-
portant, through TMPs that only shuffle a small
number of non-content words (prepositions) and
TMPs that only shuffle the sentence order, but not
the individual words within each sentence. Fur-
ther, another important distinction of pretrained
neural models is that they process unaltered text,
without classical normalization like stopword re-
moval or lemmatization; we introduce TMPs that
study these manipulations.5 Recognizing changes
such as lemmatization and word shuffling can
drastically alter the text, we also include a more
subtle TMP that applies typical typograhpical er-
rors (typos) by replacing words with common
misspellings.6 We also evaluate the recent, ef-
fective technique of using neural models to add
content (DocT5Query terms [Nogueira et al.,
2020]) to each document to aid IR, and con-
trast this with a complementary TMP that adds a
non-relevant sentence to the document.
2.2.3 Dataset Transfer Probes (DTPs)
Even with MMPs and TMPs, some characteristics
may still be difficult to measure. For instance, for
attributes like textual fluency (the degree to which
language sounds like a native speaker wrote it), we
would need to find pairs of otherwise-similar doc-
uments with measurable differences in fluency (for
an MMP) or identify ways to automatically ma-
nipulate fluency (for a TMP), both of which would
be difficult. To probe characteristics like these, we
propose Dataset Transfer Probes (DTPs). In this
setting, a dataset built for a purpose other than
ranking is repurposed to probe a ranking model’s
behavior. For example, one could create a DTP
from a dataset of human-written textual fluency
pairs (e.g., from the JFLEG dataset [Napoles et al.,
2017]) to sidestep challenges in both measurement
5We use SpaCy’s (Honnibal and Montani, 2017) lemma-
tizer, rather than, e.g., a stemmer, because the outputs from
a stemming function like Porter are often not found in the
lexicon of models like BERT.
6We use this
in English
text: https://en.wikipedia.org/wiki/Commonly
misspelled English words.
list of common errors
and manipulation. Text pair datasets are abundant,
allowing us to probe a wide variety of character-
istics, like fluency, formality, and succinctness.
With these probes, d1 and d2 can be easily defined
by the source dataset. In some cases, external in-
formation can be used to infer a corresponding q,
such as using the title of the article as a query for
news article summarization tasks, a technique that
has been studied before to train ranking models
(MacAvaney et al., 2019b). In other cases, queries
can be artificially generated, as long as the text
resembles a likely query.
We first use DTPs to investigate the important
question of whether models exhibit confounding
preferences for stylistic features of text are at least
partially independent of relevance. Specifically,
we first investigate paraphrases in general, and
then move on to check the specific qualities of
fluency, formality, simplicity, lexical bias, and
succinctness. We then use DTPs to test the ca-
pacity of models to encode and utilize real-world
knowledge through probes that measure a model’s
tendency to select factual answers.
The TMPs described in the previous section
probe the sensitivity of models to word order. In
this case, the words remain the same, but meaning
is altered. It is natural to wonder whether model
behaviors would be similar if the meaning is pre-
served when using different words. This motivates
a paraphrase DTP. We construct this probe from
the Microsoft Paraphrase Corpus (MSPC).7 We
select d1 and d2 from all text pairs labeled as
paraphrases. Note that this is the first example of
a symmetric probe, as there is no directionality
in the paraphrase relation; the assignment of d1
and d2 is arbitrary. We generate q by randomly
selecting a noun chunk that appears in both ver-
sions of the text, ensuring a query that is relevant
to both texts. (If no such chunk exists, we discard
the sample.) By selecting a noun chunk, the query
remains reasonably similar to a real query.
Although the paraphrase probe can tell us
whether models distinguish between text with
similar meaning, it cannot tell us what character-
istics it favors when making such a distinction.
To gain insights here, we propose several direc-
tional probes based on stylistic differences that
result in similar meanings. One such characteris-
tic is textual fluency. We propose a DTP using the
7https://www.microsoft.com/en-us/download
/details.aspx?id=52398.
228
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
JFLEG dataset (Napoles et al., 2017). This dataset
contains sentences from English-language fluency
tests. Each non-fluent sentence is corrected for flu-
ency by four fluent English speakers to make the
text sound ‘‘natural’’ (changes include grammar
and word usage changes). We treat each fluent
text as a d1 paired with the non-fluent d2, and use
the strategy used for paraphrases to generate q.
We probe formality by building a DTP from the
GYAFC dataset (Rao and Tetreault, 2018). This
dataset selects sentences from Yahoo Answers
and has four annotators make edits to the text
that either improve the formality (for text that is
informal), or reduce the formality (for text that is
already formal). We treat formal text as d1 and
informal text as d2. Because the text came from
Yahoo Answers, we can link the text back to the
original questions using the Yahoo L6 dataset.8
We treat the question (title) as q. In cases where
we cannot find the original text or there are no
overlapping non-stopword lemmas from q in both
d1 and d2, we discard the sample.
The simplicity of text indicates the ease of
reading a particular text. We test the effect of
lexical text simplicity using the WikiTurk dataset
provided by Xu et al. (2016). In this dataset, sen-
tences from Wikipedia were edited to make them
simpler by Amazon Mechanical Turk workers.
We treat the simplified text as d1, the original text
as d2, and we use the query construction technique
from the paraphrase probe for q.
Text can also express similar ideas but with dif-
fering degrees of subjectivity or bias. We construct
a neutrality DTP using the Wikipedia Neutrality
Corpus (WNC) dataset (Pryzant et al., 2020). This
corpus consists of sentences that were corrected
by Wikipedia editors to enforce the platform’s
neutral point of view. We use the neutral text as
d1, the biased text as d2, and we use the query
construction technique from the paraphrase probe
for q.
An idea can also be expressed in greater or
lesser detail. To probe whether models have a
preference for succinctness, we construct DTPs
from summarization datasets, using the assump-
tion that a document’s summary will be more
succinct than its full text. We utilize two datasets
to conduct this probe: XSum (Narayan et al., 2018)
8https://webscope.sandbox.yahoo.com
/catalog.php?datatype=l&did=11.
and CNN/DailyMail (See et al., 2017). The for-
mer uses extremely concise summaries from BBC
articles, usually consisting of a single sentence.
The CNN/DailyMail dataset uses slightly longer
bullet point list summaries, usually consisting of
around 3 sentences. For these probes, we use the
title of the article as q, the summarized text as
d1, and the article body as d2. When there is no
overlap between the non-stopword lemmas of q in
both d1 and d2, we discard the samples. We further
sub-sample the dataset at 10% because the datasets
are already rather large. To handle the long full
text in BERT and EPIC, we use the passage ag-
gregation strategy proposed by MacAvaney et al.
(2019a).
Moving beyond probes that express similar
ideas, we explore the extent to which models are
aware of real-world knowledge using a factual-
ity probe. This probe is motivated by the intuition
that contextualized language models may be mem-
orizing facts from the pre-training corpus when
determining relevance. We construct this probe
from the Natural Questions (Kwiatkowski et al.,
2019) dataset. We make use of the known answer
text from NQ by replacing it with a similar answer.
Similar answers must be of the same entity type9
and have the same number of non-stopword to-
kens. We discard samples where the question text
contains the answer text (e.g., this-or-that ques-
tions). We use the factual text as d1, the non-factual
version as d2, and the question text as q. Note that
this probe can be considered both a DTP and a
TMP. We decide to consider it to primarily be a
DTP because it makes use of data specific to this
external dataset (i.e., answer strings).
3 Experimental Setup
3.1 Datasets
We use the MS-MARCO passage dataset (Campos
et al., 2016) to train the neural ranking mod-
els. The training subset contains approximately
809k natural-language questions from a query
log (with an average length of 7.5 terms) and
8.8 million candidate answer passages (with an
average length of 73.1 terms). Due to its scale
in number of queries, it is shallowly annotated,
almost always containing fewer than 3 positive
judgments per query. This dataset is frequently
9Entities extracted using SpaCy: Person (PER), Location
(LOC), Geo-Political Entity (GPE), Nationality/Religion/etc.
(NORP), or Organization (ORG).
229
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
used for training neural ranking models. Impor-
tantly, it also has been shown to effectively transfer
relevance signals to other collections (Nogueira
and Lin, 2019), making it suitable for use with
DTPs, which may include text from other domains.
We build MMPs and TMPs using the TREC
Deep Learning 2019 passage dataset (Craswell
et al., 2019) and the ANTIQUE passage rank-
ing dataset (Hashemi et al., 2020). TREC DL
uses the MS-MARCO passage collection and has
43 queries with deep relevance judgments (on av-
erage, 215 per query). The judgments are graded
as highly relevant (7%), relevant (19%), topical
(17%), and non-relevant (56%), allowing us to
make more fine-grained comparisons. We use
the test subset of ANTIQUE, which contains
200 queries with 33 judgments per query. These
judgments are graded as convincing (20%), pos-
sibly correct (18%), on-topic (37%), and off-topic
(25%). We opt to perform our analysis in a pas-
sage ranking setting to eliminate effects of long
document aggregation—which is challenging for
some neural models given a maximum sequence
length in the underlying model—given that this
is an area with many model varieties that is still
under active investigation (Li et al., 2020).
3.2 Models
We compare a sample of several models covering
a traditional lexical model (BM25), a conventional
learning-to-rank approach (LightGBM), and neu-
ral models based on contextualized language
models. We include two models that focus on
query-time computational efficiency, and two rep-
resentative models that use dense retrieval. The
neural models represent a sample of the recent
state-of-the-art ranking models. For each model,
we provide the MRR (minimum relevance of
2) performance on the TREC DL 2019 pas-
sage benchmark when re-ranking the provided
candidate passages.
BM25. We use the Terrier (Ounis et al., 2006)
implementation of BM25 with default parameters.
BM25 is an unsupervised model that incorporates
the lexical features of term frequency (TF), inverse
document frequency (IDF), and document length.
(TREC DL 2019 MRR: 0.627.)
WMD. As a second unsupervised model, we
use the Word Mover’s Distance (Kusner et al., 2015)
(non-contextualized) GloVe (Pennington
over
(glove-wiki-
et
embeddings
al., 2014)
gigaword-100). We use the implementation
from the Gensim (Rehurek and Sojka, 2011)
Python package. (TREC DL 2019 MRR: 0.364.)
SBERT. As an unsupervised model based on a
contextualized language model, we use SBERT’s
and Gurevych, 2019) pre-trained
(Reimers
Bi-encoder model, trained on Semantic Textual
Similarity, Natural Language Inference, and
Quora Duplicate Question Detection data in
multiple languages.10 This approach has been
shown by Litschko et al. (2021) to effectively
perform cross-lingual retrieval. (TREC DL 2019
MRR: 0.465.)
search engine
LGBM (Ke et al., 2017). As a non-neural
learning-to-rank baseline, we use the Light Gradi-
ent Boosting Machine model currently used by the
(Feldman,
Semantic Scholar
2020).11 This public model was trained on click-
through data from this search engine, meaning
that it services various information needs (e.g.,
navigational and topical queries). Not all of the
model’s features are available in our setting (re-
cency, in-links, etc.), so we only supply the text-
based features like lexical overlap and scores from
a light-weight language model (Heafield et al.,
2013). (TREC DL 2019 MRR: 0.580.)
VBERT (Devlin et al., 2019). We use a
BERT model that uses a linear ranking layer atop
a BERT pretrained transformer language model
(Nogueira and Cho, 2019; MacAvaney et al.,
2019a; Dai and Callan, 2019b). (This setup goes by
several names in the literature, including Vanilla
BERT (VBERT), monoBERT, BERT-CAT, etc.)
We fine-tune the bert-base-uncased model
for this task using the official training sequence of
the MS-MARCO passage ranking dataset. (TREC
DL 2019 MRR: 0.809.)
T5 (Raffel et al., 2020). The Text-To-Text
Transformer ranking model (Nogueira and Lin,
2019) scores documents by predicting whether
the concatenated query, document, and control to-
kens is likely to generate the term ‘true’ or ‘false’.
We use the models released by the authors, which
were tuned on the MS-MARCO passage ranking
dataset. We test both the t5-base (T5-B) and
t5-large (T5-L) models to gain insights into
the effect of model size. (TREC DL 2019 MRR:
0.868 (T5-B), 0.857 (T5-L).)
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
10distilbert-multilingual-nli-stsb-quora-ranking.
11https://github.com/allenai/s2search.
230
EPIC (MacAvaney et al., 2020). This is an
efficiency-focused BERT-based model, which
separately encodes query and document content
into vectors that are the size of the source lexicon
(where each element represents the importance of
the corresponding term in the query/document).
We use the bert-base-uncased model, and
tune the model for ranking using the train split of
the MS-MARCO passage ranking dataset with the
code released by the EPIC authors with default
settings. (TREC DL 2019 MRR: 0.809.)
DT5Q (Nogueira and Lin, 2019). The T5 vari-
ant of the Doc2Query model (DT5Q) generates
additional terms to add to a document using a T5
model. The expanded document can be efficiently
indexed, boosting the weight of terms likely to
match queries. We use the model released by the
authors, which was trained using the MS-MARCO
passage training dataset. For our probes, we gen-
erate four queries to add to each document. As
was done in the original paper, we use BM25 as
a scoring function over the expanded documents.
(TREC DL 2019 MRR: 0.692.)
ANCE (Xiong et al., 2021). This
is a
representation-based dense retrieval model that
is trained using a contrastive learning technique.
It is designed for single-stage dense retrieval.
We use the model weights released by the orig-
inal authors, which is based on the RoBERTa
(Liu et al., 2019b) base model. (TREC DL 2019
MRR: 0.852.)
ColBERT (Khattab and Zaharia, 2020). This
is a two-stage dense retrieval approach that uses
multiple representations for each document (one
per WordPiece token). It makes use of both a
first-stage approximate nearest neighbor search to
find candidate documents and a re-ranking stage
to calculate the precise ranking scores. It is based
on the bert-base-uncased model. We use
the model weights released by the original authors.
(TREC DL 2019 MRR: 0.873.)
3.3 Choosing δ
Recall that δ indicates the minimum absolute dif-
ference of scores in a document pair probe to
have a positive or negative effect. Because each
model scores documents on a different scale,
we empirically choose a δ per model. We do
this by re-ranking the official set from TREC
DL 2019. Among the top 10 results, we calcu-
late the differences between each adjacent pair
of scores (i.e., {R(q, d1) − R(q, d2), R(q, d2) −
R(q, d3), . . . , R(q, d9) − R(q, d10)}, where di is
the ith highest scored document for q). We set δ
to the median difference. By setting the threshold
this way, we can expect the differences captured
by the probes to have an effect on the final rank-
ing score at least half the time. We explore this
further in Section 4.1. Note that choosing a con-
stant δ over one that is assigned per-query allows
for testing probes where a complete corpus is not
available, as is the case for some DTPs.
3.4 Significance Testing
We use a two-sided paired t-test to determine the
significance (pairs of R(q, d1) and R(q, d2)). We
use a Bonferroni correction over each table to
correct for multiple tests, and test for p < 0.01.
3.5 Software and Libraries
We use the following software to conduct our
experiments: PyTerrier (Macdonald et al., 2021),
OpenNIR (MacAvaney, 2020), ir datasets
(MacAvaney et al., 2021), Transformers (Wolf
et al., 2019), sentence-transformers (Reimers and
Gurevych, 2019), Anserini (Yang et al., 2018),
and Gensim (Rehurek and Sojka, 2011).
4 Results and Analysis
We present results for MMPs in Table 1, TMPs
in Table 2, and DTPs in Table 3 and highlight our
key findings in the order they appear in the tables.
Contextualized language models can dis-
tinguish relevance grades when TF is held
constant. From Table 1, we see that SBERT,
VBERT, EPIC, T5, ColBERT, and ANCE all are
able to distinguish relevance when term frequency
is constant with at least a score of +0.18 across
both datasets. Perhaps surprisingly, this is even
true for our transfer SBERT model, which is not
trained on relevance ranking data. These results
are in contrast with models that score lexically
(BM25, LGBM, and DT5Q), which score at most
+0.10. The contextualized language models also
perform better at distinguishing relevance grades
than the other models when length and overlap are
held constant, though by a lesser margin. When
controlling for model type, it appears that the
model’s size is related to its effectiveness in this
setting: The large version of T5 (T5-L, +0.53)
performs better the base model (T5-B, +0.43).
231
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
Variable
Control
BM25 WMD SBERT LGBM DT5Q VBERT
EPIC T5-B
T5-L ColBERT ANCE Samples
TREC DL 2019
Relevance Length
Length
TF
Overlap
Length
TF
Overlap
+0.40 +0.27 +0.43 +0.40 +0.48 +0.58 +0.54 +0.61 +0.66
−0.03 +0.11 +0.25 +0.04 +0.10 +0.34 +0.27 +0.43 +0.53
TF
+0.41 +0.15 +0.39 +0.34 +0.47 +0.55 +0.50 +0.61 +0.65
Overlap
∗−0.01 −0.08 +0.01 +0.00
Relevance −0.05 −0.10 −0.01 +0.04 −0.07
−0.14 −0.08 ∗+0.02 +0.02 −0.09 −0.09 −0.15 +0.01 ∗−0.00
TF
+0.51 ∗+0.02 +0.15 +0.26 +0.24 +0.20 +0.11 +0.19 +0.18
Overlap
Relevance +0.88 +0.49 +0.34 +0.50 +0.73 +0.41 +0.48 +0.38 +0.42
+1.00 +0.65 +0.46 +0.59 +0.84 +0.54 +0.61 +0.51 +0.53
Length
+0.79 ∗+0.02 +0.18 +0.37 +0.36 +0.26 +0.17 +0.26 +0.24
Overlap
Relevance +0.70 +0.47 +0.22 +0.20 +0.52 +0.19 +0.25 +0.17 +0.18
+0.75 +0.59 +0.32 +0.35 +0.59 +0.31 +0.35 +0.28 +0.29
Length
+0.88 +0.25 ∗−0.00 −0.03 +0.47 +0.11 +0.17 +0.04 +0.06
TF
ANTIQUE
−0.17 ∗−0.09 +0.12 −0.15 −0.09 +0.23 ∗−0.01 +0.26 +0.35
−0.07 ∗+0.01 +0.18 ∗+0.02 +0.04 +0.23 +0.23 +0.34 +0.46
TF
∗−0.01 ∗+0.00 +0.26 ∗+0.03 +0.12 +0.39 +0.16 +0.42 +0.47
Overlap
Relevance +0.04 ∗−0.07 +0.13 +0.23 +0.02 −0.07 +0.22 +0.12 +0.17
−0.47 ∗−0.09 +0.12 +0.04 −0.23 −0.13 +0.25 +0.03 +0.19
TF
∗+0.04 +0.35 +0.12 +0.17
+0.67 ∗+0.07 +0.17 +0.33 +0.34
Overlap
Relevance +0.69 ∗+0.23 +0.37 +0.57 +0.56 +0.24 +0.53 +0.38 +0.42
+1.00 ∗+0.48 +0.50 +0.68 +0.84 +0.39 +0.59 +0.36 +0.31
Length
∗+0.08 +0.35 +0.15 +0.22
+0.92 ∗+0.06 +0.14 +0.36 +0.45
Overlap
Relevance +0.42 ∗+0.29 +0.09 +0.01 +0.35 +0.21 ∗+0.01 +0.07 ∗+0.03
+0.67 ∗+0.33 +0.22 +0.35 +0.48 +0.10 +0.25 +0.13 ∗+0.08
Length
+0.87 ∗+0.21 +0.07 −0.05 +0.44 +0.14 −0.13 −0.01 −0.07
TF
+0.61 +0.53
+0.47 +0.45
+0.60 +0.49
∗+0.00 +0.00
∗+0.03 +0.06
+0.18 +0.15
+0.39 +0.35
+0.53 +0.47
+0.25 +0.19
+0.14 +0.18
+0.27 +0.30
∗+0.03 +0.04
+0.13 +0.24
+0.28 +0.33
+0.31 +0.36
+0.17 +0.23
+0.15 +0.24
+0.21 +0.28
+0.46 +0.45
+0.55 +0.40
+0.25 +0.28
+0.04 ∗−0.01
+0.20 +0.18
∗−0.00 −0.08
19676
31619
4762
515401
88582
3963
303058
19770
2294
357470
20819
13980
2257
5586
1211
36164
8296
902
19900
1397
553
27539
1224
4498
Relevance Length
Table 1: Results of Measure and Match Probes (MMPs) on the TREC DL 2019 and ANTIQUE
datasets. Positive scores indicate a preference towards a higher value of the variable. Scores marked
with * are not statistically significant (see Section 3.4).
Probe
Dataset BM25 WMD SBERT LGBM DT5Q VBERT EPIC T5-B T5-L ColBERT ANCE Samples
Lemmatize
Shuf. Sents.
Shuf. Words
Rem. Stops/Punct DL19
ANT
DL19
ANT
DL19
ANT
DL19
ANT
Shuf. Prepositions DL19
ANT
DL19
ANT
DL19
ANT
DL19
ANT
+ Non-Rel Sent.
+ DocT5Query
Typos
∗+0.00 ∗−0.09 −0.23 −0.20 −0.04 +0.18 −0.78 −0.74 −0.80
∗+0.04 ∗−0.19 −0.38 −0.24 −0.07 −0.25 −0.78 −0.64 −0.81
+0.00 −0.18 +0.05 −0.02 ∗+0.01 −0.04 −0.25 −0.42 −0.44
+0.04 ∗−0.01 −0.04 −0.09 +0.00 −0.22 −0.25 −0.30 −0.47
∗+0.00 −0.21 −0.06 −0.25 −0.11 −0.38 −0.76 −0.65 −0.76
∗+0.04 ∗−0.11 −0.10 −0.25 −0.13 −0.61 −0.67 −0.65 −0.75
∗−0.00 ∗−0.01 −0.06 ∗−0.00 ∗−0.02 −0.13 −0.19 −0.20 −0.14
∗−0.00 ∗−0.02 −0.04 ∗−0.00 ∗−0.02 −0.17 −0.20 −0.22 −0.22
+0.01 −0.21 −0.02 −0.02 ∗+0.02 −0.01 −0.11 −0.28 −0.31
+0.05 ∗−0.11 −0.04 −0.03 +0.01 −0.12 −0.16 −0.30 −0.36
−0.23 −0.17 ∗+0.07 −0.15 −0.18 −0.09 −0.50 −0.37 −0.27
−0.32 ∗−0.41 −0.09 −0.27 −0.27 −0.40 −0.45 −0.38 −0.40
+0.34 +0.45 +0.33 +0.41 +0.15 −0.22 −0.63 −0.54 −0.60
+0.34 ∗+0.14 +0.17 +0.32 ∗+0.03 −0.42 −0.13 −0.67 −0.68
−0.03 −0.10 +0.34 +0.20 +0.04 +0.26 +0.11 +0.33 +0.27
+0.25 ∗+0.04 +0.38 +0.31 +0.25 +0.08 +0.47 +0.28 +0.30
−0.68 −0.59
−0.74 −0.70
−0.38 −0.31
−0.25 −0.31
−0.76 −0.40
−0.67 −0.58
−0.14 −0.10
−0.13 −0.14
−0.18 −0.24
−0.18 −0.29
−0.42 −0.20
−0.56 −0.36
−0.50 −0.47
∗−0.10 −0.37
+0.33 +0.39
+0.34 +0.41
9259
6540
9259
6392
9260
6545
7290
4211
9239
6186
8982
5551
9260
6589
9260
6346
Table 2: Results of Text Manipulation Probes (TMPs) on the TREC DL 2019 and ANTIQUE datasets.
Positive scores indicate a preference for the manipulated document text; negative scores prefer the
original text. Scores marked with * are not statistically significant (see Section 3.4).
Models generally have similar sensitivity
length, TF, and overlap on
to document
TREC DL 2019. With the exception of mod-
els that use BM25 for scoring (BM25 and
DT5Q), all the models we explore have sim-
ilar behaviors when varying length, TF, and
overlap. This suggests that although signals like
TF are not required for EPIC, BERT, and T5
to rank effectively,
they still remain an im-
portant signal when available. There are bigger
differences between models when exploring the
ANTIQUE dataset, suggesting differences in the
models’ capacity to generalize. We note that some
of the largest differences relate to the relevance
measurement, highlighting the differences in label
definitions between the two datasets.
232
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
0.94
0.60
0.85
∗0.89
∗0.00
Formality
T5-B
0.91
EPIC
0.65
BM25 WMD SBERT LGBM DT5Q VBERT
0.82
Probe
Dataset
T5-L ColBERT ANCE Samples
Paraphrase MSPC
3421
0.88
0.76
0.90
+0.03 ∗−0.07 ∗+0.00 ∗−0.00 ∗+0.02 +0.10 +0.22 +0.14 +0.07
+0.24 +0.17
5073
JFLEG
Fluency
(spellchecked) ∗+0.01 ∗+0.05 ∗−0.03 ∗−0.00 ∗−0.01 +0.07 +0.20 +0.14 +0.13
+0.18 +0.09
5187
−0.03 ∗−0.03 −0.15 −0.09 −0.07 −0.05 +0.16 ∗+0.01 +0.15
−0.07 ∗+0.05
6721
GYAFC
+0.04 ∗+0.01 ∗−0.11 −0.04 −0.01 ∗+0.04 +0.19 ∗+0.11 +0.23
∗+0.01 +0.08
- entertain.
2960
−0.08 ∗−0.05 −0.18 −0.13 −0.11 −0.12 ∗+0.13 −0.08 ∗+0.08
−0.14 ∗+0.03
3761
- family
+0.13 ∗+0.21 +0.07 ∗−0.00 +0.05 ∗−0.03 ∗−0.01 −0.08 −0.13
+0.01 ∗−0.03
17849
+0.31 ∗+0.34 +0.11 ∗+0.00 +0.13 +0.11 +0.07 −0.00 +0.03
+0.13 −0.00 178252
+0.66 ∗+0.91 +0.58 +0.18 +0.66 +0.49 +0.18 −0.09 +0.07
+0.33 +0.47
17938
+0.37 ∗+0.74 ∗+0.02 −0.43 +0.41 +0.16 −0.72 −0.58 −0.54
−0.33 −0.28
7154
∗−0.01 +0.54 −0.37 −0.80 +0.06 −0.26 −0.93 −0.63 −0.58
−0.71 −0.56
18930
∗−0.00 +0.16 −0.02 −0.00 −0.02 ∗−0.02 −0.07 +0.10 +0.14
+0.04 +0.04
72983
∗−0.00 +0.22 +0.02 +0.00 ∗+0.00 +0.09 +0.00 +0.27 +0.30
+0.22 +0.12
33528
+0.14 ∗+0.11
∗−0.03 +0.21 ∗−0.12 ∗−0.02 ∗−0.02 ∗+0.01 ∗+0.02 +0.28 +0.29
962
+0.26 +0.10
+0.02 +0.30 +0.01 +0.01 +0.03 +0.07 +0.07 +0.25 +0.33
4250
+0.01 +0.34 +0.01 ∗+0.00 ∗+0.01 +0.07 −0.01 +0.33 +0.38
+0.19 +0.13
13831
Simplicity WikiTurk
Neutrality WNC
Succinctness XSum
CNN
Daily Mail
NQ: PER
NQ: GPE
NQ: LOC
NQ: NORP
NQ: ORG
Factuality
Table 3: Results of Dataset Transfer Probes (DTPs). The paraphrase probe is unsigned, as it is
symmetric. Positive scores indicate a preference for fluent, formal, simplified, neutral (non-biased),
succinct, and factual text. Scores marked with * are not statistically significant (see Section 3.4).
Trained contextualized language models
are adversely affected by heavily destructive
pre-processing steps. From Table 2, we find that
removing stopwords and punctuation, performing
lemmatization, and shuffling words negatively im-
pacts most models across both datasets. Perhaps
this is expected, given that this text is dissimilar to
the text the models were pre-trained on. However,
we note that the transfer SBERT model is far less
affected by these operations, suggesting that these
characteristics are not intrinsic to the contextual-
ized language models, but rather a consequence
of training them for relevance ranking. To gain
further insights into the importance of word order,
we control for local word order by only shuffling
sentence order. We see that an effect remains for
the contextualized models, though it is substan-
tially reduced. This suggests that discourse-level
signals (e.g., what topics are discussed earlier in
a document) have some effect on the models,
or the models encode some positional bias (e.g.,
preferring answers at the start of documents). To
understand if the word usage of particular terms is
important (rather than overall coherence), we also
try shuffling only the prepositions in the sentence.
We find that this has an effect on some models
(most notably, both T5 models and ANCE), but
not other models, suggesting that some end up
learning that although these terms have meaning
in the text, they are often unimportant when it
comes to ranking.
Lexical models handle typographical errors
better than trained contextualized language
models. In all but one case (ANCE DL19), BM25,
LGBM, and DT5Q are negatively affected by
typographical errors less than the trained con-
textualized language models. This is a surprising
result, given that contextualized language mod-
els should be able to learn common misspellings
and treat them similarly to the original words
(the transfer SBERT model largely ignores ty-
pos). This problem is particularly apparent for
EPIC and ColBERT, which perform matching on
the WordPiece level.
Trained contextualized models behave unex-
pectedly when additional content is introduced
in documents. We find that models that rely
heavily on unigram matching (e.g., BM25) and
the transfer SBERT model respond positively
to the addition of DocT5Query terms. Even the
DocT5Query model itself sees an additional boost,
suggesting that weighting the expansion terms
higher in the document may further improve the
effectiveness of this model. However, the contex-
tualized models often respond negatively to these
additions. We also find that adding non-relevant
sentences to the end of relevant documents often
increases the ranking score of contextualized mod-
els. This is in contrast with models like BM25, in
which the scores of relevant documents decrease
with the addition of non-relevant information.
From the variable length MMPs, we know that
this increase in score is likely not due to increas-
ing the length alone. Such characteristics may
pose a risk to ranking systems based on contex-
tualized models, in which content sources could
aim to increase their ranking simply by adding
non-relevant content to their documents.
233
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
Paraphrasing text can drastically change
ranking scores. In Table 3, we observe high
scores across most models for the paraphrase
probe. For BM25, this is because the document
lengths differ to a substantial degree. Contextual-
ized models—which one may expect to handle
semantic equivalences like these well—assign
substantially different scores for paraphrases up
to 94% of the time. To dig into specific stylis-
tic differences that could explain the paraphrase
discrepancies, we explore fluency, formality, sim-
plicity, and neutrality. We find that fluency and
formality have a greater effect than simplicity and
neutrality. Most notably, EPIC and ColBERT pre-
fer fluent text with scores of +0.18 to +0.24,
while lexical models have low or insignificant
differences. Meanwhile, EPIC and T5-L prefer
formal text, while ColBERT and T5-B either pre-
fer informal text or have no significant differ-
ences. Finally, the largest preferences observed
for simple and neutral text are from BM25—
which are likely a consequence of reduced docu-
ment lengths.
Model behaviors vary considerably with suc-
cinctness. First, BM25 has a strong (+0.66)
preference for the summaries in XSum, a mod-
erate preference for summaries in CNN (+0.37),
and no significant preference for Daily Mail. This
suggests different standards among the various
datasets—for example, XSum (BBC) must use
many of the same terms from the titles in the sum-
maries, and provide long documents (reducing the
score) that may not repeat terms from the title
much. WMD also appears to be heavily affected
by summaries, though in two of the three probes,
there is insufficient evidence to claim significance.
The preference for summaries in XSum can be
seen across all models except T5-B, which very
slightly favors the full text. Although most contex-
tualized models prefer the full text for CNN and
Daily Mail, VBERT prefers summaries for CNN
(+0.16) while it prefers full text for Daily Mail
(−0.26). Such discrepancies warrant exploration
in future work.
WMD, T5, and ColBERT are biased towards
factual answers. From our factuality probes, we
see that most models have little preference for
factual passages. However, WMD, both T5 vari-
ants, and ColBERT are biased towards answers
that contain factually correct information. For T5
and ColBERT, this suggests that these models
both learn some real-world information (likely in
pre-training), and use this information as a signal
when ranking. The larger size of T5-L appears
to equip it with more knowledge, particularly
about people, nationalities, and organizations.
Curiously, although ColBERT exploits this infor-
mation, the VBERT model (which uses the same
base language model) does not appear to learn to
use this information. For WMD, which doesn’t
have nearly the modeling capacity of T5 and Col-
BERT, the preference for factual information must
be due to the fact that the word embeddings of
the question are more similar to the word em-
beddings from the factual phrase than to those of
the non-factual phrase. Although the contextual-
ized language models should have the capacity to
learn these trends and make similar decisions, this
would be subject to such trends being present and
distinguishable during fine-tuning. This suggests
that using WMD over contextualized word em-
beddings may also improve the capacity of models
to select factual answers.
4.1 Effect of δ
Recall that δ defines the model-specific threshold
at which a difference in ranking score is con-
sidered important. To test the importance of the
selection of δ, we test all probes while varying
this parameter. Because the suitable values de-
pend upon the range of scores for the particular
ranker, we select δ by percentile among differ-
ences in the top 10 scoring passages of TREC DL
2019. Figure 2 provides a representative sample
of these plots. We find that for low percentiles
(corresponding to settings where minute changes
in score are considered important), the scores and
rankings of systems can sometimes be unstable
(e.g., see BM25 and DT5Q in (c)). This suggests
that there are variations of the score distributions
close to 0. However, we remind the reader that
such differences are unlikely to have impactful
changes in a real ranked list. We find that by
the 50th percentile of δ (i.e., the value we use
for our experiments), the rankings of the systems
produced by ABNIRML are generally stable. In
most cases, the scores are stable as well, though in
some cases drifting occurs (e.g., (c)). With a large
δ, nearly no differences are considered important.
In (c), we observe that L-GBM has no sensitivity
to the paraphrases present in the probe, regardless
of δ. These observations validate our technique
for choosing δ.
234
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
Figure 2: Plots of scores for three representative probes when varying δ to the specified percentile in TREC DL
2019. The vertical dashed line indicates the operational point of our experiments (the median value).
Probes
VBERT Stdev.
EPIC Stdev.
5 Related Work
MMP
TMP
DTP
3.5
11.2
9.5
3.6
17.1
8.9
Table 4: Average standard deviations (square
root of average variance) of 5 VBERT and
EPIC models, by probe type.
4.2 Effect of Model Training
We observed that identical and similar base lan-
guage models can differ in the behaviors they
exhibit. To gain a better understanding of the ori-
gin of these differences, we probe 5 versions of
the VBERT and EPIC models, each trained with
different random seeds. We calculate the standard
deviations of the performance over all the probes
and report the average standard deviation for each
probe type in Table 4. We find that among all
probe types, MMPs are the most stable across ran-
dom initializations and TMPs are the least stable.
Curiously, the Stopword / punctuation removal
TMP is the least stable probe across both models,
with a stdev of 0.24 for VBERT and 0.27 for
EPIC. In the case of VBERT, the probe score
ranged from −0.33 to +0.31, highlighting that
unexpected qualities can appear in models simply
due to random variations in the training process.
This is despite the fact that this probe is highly
robust to the cutoff threshold on individual models
(as seen in Figure 2(b)). Another probe with par-
ticularly high variance are the succinctness probe
for VBERT using the CNN dataset, with a stdev
of 0.23, and can either learn to prefer succinct
(+0.15) or elaborative (−0.42) text, again due to
the random initialization. These findings highlight
that some biases can be introduced in the model
training process randomly, rather than as a result
of the pre-training process or model architecture.
Pretrained contextualized language models are
neural networks that are initially trained on lan-
guage modeling objectives and are later fine-tuned
on task-specific objectives (Peters et al., 2018).
Well-known models include ELMo (Peters et al.,
2018), BERT (Devlin et al., 2019), and T5
(Raffel et al., 2020). These models can effectively
transfer signals to the task of ad-hoc retrieval,
either by using the model directly (i.e., vanilla or
mono models) (Nogueira and Cho, 2019) or by
using the outputs as features into a larger model
(MacAvaney et al., 2019a). There has been con-
siderable work in this area; we refer readers to
Lin et al.
for a comprehensive sur-
vey on these techniques. We shed light on the
mechanisms, strengths, and weaknesses of this
burgeoning body of work.
(2020)
Diagnostic datasets, proposed by Rennings et al.
(2019), reformulate traditional ranking axioms—
for example, that documents with a higher term
frequency should receive a higher ranking score
(Fang et al., 2004)—as empirical tests for ana-
lyzing ranking models. Rennings et al. studied
neural ranking architectures that predate the rise
of contextualized language models for ranking,
and focused on just four axioms. Cˆamara and
Hauff (2020) extended this work by adding five
more previously proposed ranking axioms (e.g.,
term proximity [Tao and Zhai, 2007], and word
semantics [Fang and Zhai, 2006]) and evaluating
on a distilled BERT model. They found that the
axioms are inadequate to explain the ranking ef-
fectiveness of their model. V¨olske et al. (2021)
examine the extent to which these axioms, when
acting in concert, explain ranking model decisions.
Unlike these prior lines of work, we propose new
probes that shed light onto possible sources of ef-
fectiveness, and test against current leading neural
ranking architectures.
235
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
Although some insights about the effectiveness
rank-
of contextualized language models for
ing have been gained using existing datasets
(Dai and Callan, 2019b) and indirectly through
various model architectures (Nogueira et al.,
2019; Dai and Callan, 2019a; MacAvaney et al.,
2020, 2019a; Hofst¨atter et al., 2020; Khattab and
Zaharia, 2020), they only provide circumstan-
tial evidence. For instance, several works show
how contextualized embedding similarity can be
effective, but this does not imply that vanilla mod-
els utilize these signals for ranking. Rather than
proposing new ranking models, in this work we
analyze the effectiveness of existing models us-
ing controlled diagnostic probes, which allows us
to gain insights into the particular behaviors and
preferences of the ranking models.
Outside of the work in IR, others have devel-
oped techniques for investigating the behavior of
contextualized language models in general. Al-
though probing techniques (Tenney et al., 2019)
and attention analysis (Serrano and Smith, 2019)
can be beneficial for understanding model capabil-
ities, these techniques cannot help us characterize
and quantify the behaviors of neural ranking mod-
els. CheckList (Ribeiro et al., 2020) and other
challenge set techniques (McCoy et al., 2019)
differ conceptually from our goals; we aim to char-
acterize the behaviors to understand the qualities
of ranking models, rather than provide additional
measures of model quality.
6 Conclusion
We presented a new framework (ABNIRML) for
analyzing ranking models based on three probing
strategies. By using probes from each strategy,
we demonstrated that a variety of insights can be
gained about the behaviors of recently-proposed
ranking models, such as those based on BERT and
T5. Our analysis is, to date, the most extensive
analysis of the behaviors of neural ranking mod-
els, and sheds light on several unexpected model
behaviors. For instance, adding non-relevant text
can increase a document’s ranking score, even
though the models are largely not biased towards
longer documents. We also see that the same
base language model used with a different ranking
architecture can yield different behaviors, such
as higher sensitivity to shuffling a document’s
text. We also find that some models learn to uti-
lize real-world knowledge in the ranking process.
Finally, we observe that some strong biases can
appear simply by chance during the training
process. This motivates future investigations on
approaches for stabilizing training processes and
avoiding the introduction of unwanted biases.
References
Matteo Alleman, J. Mamou, M. D. Rio, Hanlin
Tang, Yoon Kim, and SueYeon Chung.
2021. Syntactic perturbations reveal
repre-
sentational correlates of hierarchical phrase
structure
in pretrained language models.
arXiv, abs/2104.07578. https://doi.org
/10.18653/v1/2021.repl4nlp-1.27
Arthur Cˆamara and Claudia Hauff. 2020. Di-
agnosing BERT with retrieval heuristics. In
ECIR. https://doi.org/10.1007/978
-3-030-45439-5_40
Daniel Fernando Campos, T. Nguyen, M.
Rosenberg, Xia Song, Jianfeng Gao, Saurabh
Tiwary, Rangan Majumder, L. Deng, and
Bhaskar Mitra. 2016. MS MARCO: A hu-
man generated machine reading comprehension
dataset. arXiv, abs/1611.09268.
Nick Craswell, Bhaskar Mitra, Emine Yilmaz,
Daniel Campos, and Ellen M. Voorhees. 2019.
Overview of the TREC 2019 deep learning
track. In TREC.
Zhuyun Dai and J. Callan. 2019a. Context-aware
sentence/passage term importance estimation
for first stage retrieval. arXiv, abs/1910.10687.
Zhuyun Dai and J. Callan. 2019b. Deeper text
understanding for ir with contextual neural lan-
guage modeling. SIGIR. https://doi.org
/10.1145/3331184.3331303
Jeffrey Dalton, Chenyan Xiong, and Jamie
Callan. 2019. CAsT 2019: The conversational
assistance track overview. In TREC.
J. Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. BERT: Pre-training
of deep bidirectional transformers for language
understanding. In NAACL-HLT.
Hui Fang, T. Tao, and ChengXiang Zhai. 2004. A
formal study of information retrieval heuristics.
In SIGIR. https://doi.org/10.1145
/1008992.1009004
236
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
Hui Fang, T. Tao, and ChengXiang Zhai.
2011. Diagnostic evaluation of information
retrieval models. ACM Transactions on Man-
agement
Information Systems, 29:7:1–7:42.
https://doi.org/10.1145/1961209
.1961210
Hui Fang and ChengXiang Zhai. 2006. Semantic
term matching in axiomatic approaches to infor-
mation retrieval. In SIGIR ’06. https://doi
.org/10.1145/1148170.1148193
Sergey Feldman. 2020. Building a better search
engine for semantic scholar. Blog post.
Helia Hashemi, Mohammad Aliannejadi, Hamed
Zamani, and Bruce Croft. 2020. ANTIQUE: A
non-factoid question answering benchmark. In
ECIR. https://doi.org/10.1007/978
-3-030-45442-5_21
Kenneth Heafield, Ivan Pouzyrevsky, Jonathan H.
Clark, and Philipp Koehn. 2013. Scalable mod-
ified Kneser-Ney language model estimation.
In ACL.
Sebastian Hofst¨atter, Markus Zlabinger, and A.
Hanbury. 2020. Interpretable and time-budget-
constrained contextualization for re-ranking.
In ECAI.
Matthew Honnibal and Ines Montani. 2017. spaCy
2: Natural language understanding with Bloom
embeddings, convolutional neural networks and
incremental parsing. To appear.
Guolin Ke, Q. Meng, Thomas Finley, Taifeng
Wang, Wei Chen, Weidong Ma, Qiwei Ye, and
T. Liu. 2017. LightGBM: A highly efficient
gradient boosting decision tree. In NIPS.
O. Khattab and M. Zaharia. 2020. ColBERT:
Efficient and effective passage search via
contextualized late interaction over BERT.
In SIGIR. https://doi.org/10.1145
/3397271.3401075
Matt J. Kusner, Yu Sun, Nicholas I. Kolkin,
and Kilian Q. Weinberger. 2015. From word
embeddings to document distances. In ICML.
Tom Kwiatkowski, Jennimaria Palomaki, Olivia
Redfield, Michael Collins, Ankur Parikh, Chris
Alberti, Danielle Epstein,
Illia Polosukhin,
Matthew Kelcey, Jacob Devlin, Kenton Lee,
Kristina N. Toutanova, Llion Jones, Ming-Wei
Chang, Andrew Dai, Jakob Uszkoreit, Quoc
Le, and Slav Petrov. 2019. Natural Questions:
a benchmark for question answering research.
In TACL. https://doi.org/10.1162
/tacl_a_00276
Canjia Li, Andrew Yates, Sean MacAvaney,
Ben He, and Yingfei Sun. 2020. Parade: Pas-
sage representation aggregation for document
reranking. arXiv, abs/2008.09093.
Jimmy Lin, Rodrigo Nogueira, and A. Yates.
2020. Pretrained transformers for text ranking:
BERT and beyond. arXiv, abs/2010.06467.
Robert Litschko,
Ivan Vuli’c, Simone Paolo
Ponzetto, and Goran Glavavs. 2021. Evaluating
multilingual
text encoders for unsupervised
cross-lingual retrieval. In ECIR. https://doi
.org/10.1007/978-3-030-72113-8 23
Nelson F. Liu, Matt Gardner, Yonatan Belinkov,
Matthew E. Peters, and Noah A. Smith. 2019a.
Linguistic knowledge and transferability of
contextual representations. In NAACL-HLT.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei
Du, Mandar Joshi, Danqi Chen, Omer Levy,
M. Lewis, Luke Zettlemoyer, and Veselin
Stoyanov. 2019b. RoBERTa: A robustly op-
timized BERT pretraining approach. arXiv,
abs/1907.11692.
D. Loureiro, Kiamehr Rezaee, Mohammad Taher
Pilehvar, and Jos´e Camacho-Collados. 2020.
Language models and word sense disam-
biguation: An overview and analysis. arXiv,
abs/2008.11608.
Sean MacAvaney. 2020. OpenNIR: A complete
neural ad-hoc ranking pipeline. In WSDM.
Sean MacAvaney, Franco Maria Nardini, Raffaele
Perego, Nicola Tonellotto, Nazli Goharian,
and Ophir Frieder. 2020. Expansion via pre-
diction of importance with contextualization.
In SIGIR. https://doi.org/10.1145
/3397271.3401262
Sean MacAvaney, Andrew Yates, Arman Cohan,
and Nazli Goharian. 2019a. CEDR: Contextu-
alized embeddings for document ranking. In
SIGIR.
Sean MacAvaney, Andrew Yates, Sergey Feldman,
Doug Downey, Arman Cohan, and Nazli
Goharian. 2021. Simplified data wrangling with
ir datasets. In SIGIR. https://doi.org
/10.1145/3404835.3463254
237
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
Sean MacAvaney, Andrew Yates, Kai Hui, and
Ophir Frieder. 2019b. Content-based weak
supervision for ad-hoc re-ranking. In SIGIR.
https://doi.org/10.1145/3331184
.3331316
Craig Macdonald, Nicola Tonellotto, Sean
MacAvaney, and Iadh Ounis. 2021. PyTerrier:
Declarative experimentation in python from
BM25 to dense retrieval. In CIKM.
vectors for word representation. In EMNLP.
https://doi.org/10.3115/v1/D14
-1162
Matthew Peters, Mark Neumann, Mohit Iyyer,
Matt Gardner, Christopher Clark, Kenton Lee,
and Luke Zettlemoyer. 2018. Deep contex-
In NAACL-
tualized word representations.
HLT. https://doi.org/10.18653/v1
/N18-1202
R. Thomas McCoy, Ellie Pavlick, and Tal Linzen.
2019. Right for the wrong reasons: Diagnos-
language
ing syntactic heuristics in natural
inference. In ACL. https://doi.org/10
.1145/3459637.3482013
Reid Pryzant, Richard Diehl Martinez, Nathan
Dass, S. Kurohashi, Dan Jurafsky, and Diyi
Yang. 2020. Automatically neutralizing sub-
jective bias in text. In AAAI. https://doi
.org/10.1609/aaai.v34i01.5385
Courtney Napoles, Keisuke Sakaguchi, and Joel
Tetreault. 2017. JFLEG: A fluency corpus and
benchmark for grammatical error correction.
In EACL. https://doi.org/10.18653
/v1/E17-2037
Shashi Narayan, Shay B. Cohen, and Mirella
Lapata. 2018. Don’t give me the details,
just the summary! Topic-aware convolutional
neural networks
extreme summariza-
tion. In EMNLP. https://doi.org/10
.18653/v1/D18-1206
for
Rodrigo Nogueira and Kyunghyun Cho. 2019.
arXiv,
re-ranking with BERT.
Passage
abs/1901.04085.
Rodrigo Nogueira, Zhiying Jiang, and Jimmy Lin.
2020. Document ranking with a pretrained
sequence-to-sequence model. arXiv, abs/2003
.06713. https://doi.org/10.18653
/v1/2020.findings-emnlp.63
Rodrigo Nogueira and Jimmy Lin. 2019. From
doc2query to docTTTTTquery. Self-published.
Rodrigo Nogueira, Wei Yang, Jimmy Lin, and
Kyunghyun Cho. 2019. Document expansion
by query prediction. arXiv, abs/1904.08375.
I. Ounis, G. Amati, V. Plachouras, B. He, C.
Macdonald, and C. Lioma. 2006. Terrier: A
high performance and scalable information
retrieval platform. In Proceedings of ACM
SIGIR’06 Workshop on Open Source Information
Retrieval (OSIR 2006). https://doi.org
/10.1007/978-3-540-31865-1_37
Jeffrey
Pennington, Richard
and
Christopher D. Manning. 2014. GloVe: Global
Socher,
Colin Raffel, Noam Shazeer, Adam Roberts,
Katherine Lee, Sharan Narang, Michael
Matena, Yanqi Zhou, Wei Li, and Peter J. Liu.
2020. Exploring the limits of transfer learning
with a unified text-to-text transformer. Journal
of Machine Learning Research, 21(140).
Sudha Rao and J. Tetreault. 2018. Dear sir or
madam, may I introduce the YAFC corpus:
Corpus, benchmarks and metrics for formality
style transfer. In NAACL-HLT. https://doi
.org/10.18653/v1/N18-1012
Radim Rehurek and Petr Sojka. 2011. Gensim–
Python framework for vector space modelling.
NLP Centre, Faculty of Informatics, Masaryk
University, Brno, Czech Republic, 3(2).
Nils Reimers
and Iryna Gurevych. 2019.
Sentence-BERT: Sentence embeddings using
Siamese BERT-networks. In EMNLP. Associa-
tion for Computational Linguistics. https://
doi.org/10.18653/v1/D19-1410
D. Rennings, Felipe Moraes, and C. Hauff. 2019.
An axiomatic approach to diagnosing neural IR
models. In ECIR. https://doi.org/10
.1007/978-3-030-15712-8_32
Marco T´ulio Ribeiro, Tongshuang Wu, Carlos
Guestrin, and Sameer Singh. 2020. Beyond ac-
curacy: Behavioral testing of NLP models with
checklist. In ACL. https://doi.org/10
.24963/ijcai.2021/659
Anna Rogers, O. Kovaleva, and Anna Rumshisky.
2020. A primer in BERTology: What we know
about how BERT works. TACL. https://
doi.org/10.1162/tacl a 00349
238
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
A. See, Peter J. Liu, and Christopher D. Manning.
2017. Get to the point: Summarization with
pointer-generator networks. In ACL. https://
doi.org/10.18653/v1/P17-1099
Sofia Serrano and Noah A. Smith. 2019. Is at-
tention interpretable? In ACL. https://doi
.org/10.18653/v1/P19-1282
Koustuv Sinha, Robin Jia, Dieuwke Hupkes, J.
Pineau, Adina Williams, and Douwe Kiela.
2021. Masked language modeling and the
distributional hypothesis: Order word matters
pre-training for little. arXiv, abs/2104.06644.
https://doi.org/10.18653/v1/2021
.emnlp-main.230
T. Tao and ChengXiang Zhai. 2007. An explo-
ration of proximity measures in information
retrieval. In SIGIR. https://doi.org/10
.1145/1277741.1277794
Ian Tenney, Dipanjan Das, and Ellie Pavlick.
2019. BERT rediscovers the classical NLP
pipeline. In ACL. https://doi.org/10
.18653/v1/P19-1452
arXiv, abs/2106.08019. https://doi.org
/10.1145/3471158.3472256
Thomas Wolf, Lysandre Debut, Victor Sanh,
Julien Chaumond, Clement Delangue, Anthony
Moi, Pierric Cistac, Tim Rault, R’emi Louf,
Morgan
Jamie Brew.
2019. HuggingFace’s Transformers: State-of-
the-art natural
language processing. arXiv,
abs/1910.03771.
Funtowicz,
and
Patrick Xia, Shijie Wu, and B. Van Durme.
2020. Which *BERT? a survey organizing
contextualized encoders. In EMNLP.
Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung
Tang, Jialin Liu, Paul Bennett, Junaid Ahmed,
and Arnold Overwijk. 2021. Approximate near-
est neighbor negative contrastive learning for
dense text retrieval. arXiv, abs/2007.00808.
Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze
Chen, and Chris Callison-Burch. 2016. Opti-
mizing statistical machine translation for text
simplification. TACL, 4. https://doi.org
/10.1162/tacl_a_00107
Michael V¨olske, A. Bondarenko, Maik Fr¨obe,
Matthias Hagen, Benno Stein, Jaspreet Singh,
and Avishek Anand. 2021. Towards ax-
iomatic explanations for neural ranking models.
Peilin Yang, Hui Fang, and Jimmy Lin. 2018.
Anserini: Reproducible ranking baselines using
Lucene. J. Data and Information Quality, 10.
https://doi.org/10.1145/3239571
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
4
5
7
2
0
0
2
6
9
8
/
/
t
l
a
c
_
a
_
0
0
4
5
7
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
239
