Query-Based Keyphrase Extraction from Long Documents
Martin Docekal, Pavel Smrz
Brno University of Technology
idocekal@ﬁt.vutbr.cz, smrz@ﬁt.vutbr.cz
Abstract
Transformer-based architectures in natural language process-
ing force input size limits that can be problematic when long
documents need to be processed. This paper overcomes this
issue for keyphrase extraction by chunking the long docu-
ments while keeping a global context as a query deﬁning the
topic for which relevant keyphrases should be extracted. The
developed system employs a pre-trained BERT model and
adapts it to estimate the probability that a given text span
forms a keyphrase. We experimented using various context
sizes on two popular datasets, Inspec and SemEval, and a
large novel dataset. The presented results show that a shorter
context with a query overcomes a longer one without the
query on long documents.1
Introduction
Keyphrase refers to a short language expression describ-
ing the content of a longer text. Due to their concise form,
keyphrases can be used for a quick familiarization with
a document. They also improve the ﬁndability of docu-
ments or passages within them. In the bibliographic records,
keyphrase descriptors enable ﬂexible indexing.
Whether a text span is a keyphrase depends on the context
of that span because a keyphrase for a speciﬁc topic may
not be a keyphrase for another topic. The presented work
builds on the idea that the topic can be explicitly given as
an input to the keyphrase extraction algorithm in the form
of a query. We approximate such a query with a document’s
title in our experiments. We also investigate the inﬂuence of
context size and document structure on the results.
Related Work
Traditional approaches to keyphrase extraction involve
graph-based methods, such as TextRank (Mihalcea and Ta-
rau 2004) and RAKE (Rose et al. 2010). Recently, many
types of neural networks have been used for the task (Lin
and Wang 2019; Sahrawat et al. 2020). Most of the deep
learning work assumes the existence of a title and an abstract
of the document and extracts keyphrases from them because
Copyright © 2022 by the authors. All rights reserved.
1The code is available at https://github.com/KNOT-FIT-
BUT/QBEK.they struggle with longer inputs such as whole scientiﬁc ar-
ticles (Kontoulis, Papagiannopoulou, and Tsoumakas 2021).
Some works try to overcome this limitation by ﬁrst creating
a document summary and then extracting keyphrases from
it (Kontoulis, Papagiannopoulou, and Tsoumakas 2021; Liu
and Iwaihara 2021). Our research follows an alternative
path, compensating for the limited context by a query spec-
ifying a topic.
Model
First, a document is split into parts (contexts), which are fur-
ther processed independently. Then, the devised model esti-
mates the probability that a given continuous text span forms
a keyphrase. It looks for boundaries tsandte, corresponding
to the text span’s start and end, respectively. The inspiration
for this approach comes from the task of reading compre-
hension, where a similar technique is used to search for po-
tential answers to a question in an input text (Devlin et al.
2019). Formally, the model estimates probability:
P(ts; tejx) =P(tsjx)P(tejx); (1)
where xis the input sequence. It assumes that the start and
the end of a text span < ts; te>are independent. The prob-
abilities P(tsjx)andP(tejx)are obtained in the following
way:
P(tjx) =sigmoid (wT
BERT (x)[] +b); (2)
wherestands for the start and end positions. Weights
wsandweare learnable vectors, bis the scalar bias and
BERT (x)[i]is BERT (Devlin et al. 2019) vector representa-
tion of a token from sequence xon position i. See the model
illustration in Figure 1.
The task of predicting whether a given token is the start
token of a span or the end token could be seen as binary clas-
siﬁcation with two classes start/end andnot start/not end ,
respectively. The binary cross-entropy ( BCE ) loss function
is used for training in the following way:
BCE( vs; gs) + BCE( ve; ge); (3)
where vsis a vector of probabilities that a token is the start
token of a span, for each token in the input, and veis vec-
tor computed analogously, but for the ends. The gsandge
are vectors of ground truth probabilities of starts or ends,
respectively.x = [CLS] title [SEP] context [SEP]
P(t |x)s
P(t |x)e
BERT
BERT
Linear Layer
SigmoidBERT(x)[i]Figure 1: Illustration of our model with query at the input.
We work with two types of inputs. One consists of a text
fragment such as a sentence, and the other uses a query (doc-
ument title) and a text segment, separated by a special token.
Various context sizes are explored in our experiments. The
context size determines how big is the document part the
model sees at once. Every context part of a document is pro-
cessed independently. The ﬁnal list of keyphrases is created
by collecting keyphrase spans with their scores and selecting
the top ones.
Datasets
Besides two standard datasets for keyphrase extraction, we
created and used a novel dataset of long documents, referred
to a Library, and we also prepared an unstructured version
of the SemEval-2010 dataset. A comparison of the datasets
is given in Table 1.
dataset train val. testsentences
(train)
SemEval-2010 130 14 100 66 428 (72%)
Unstructured-
SemEval-2010130 14 100 45 346 (67%)
Inspec 1 000 500 500 5 894 (25%)
Library 48 879 499 499 298 217 589 (94%)
Table 1: The number of documents in each split along with
the total number of sentences in a train set. The percentage in
the sentences column is the proportion of sentences without
keyphrases.
We had to annotate the spans that represent given
keyphrases in the text as the discussed datasets provide just a
list of associated keyphrases with no information about their
actual positions. The search was case insensitive and the
Porter stemmer was utilized for the SemEval and Hulth2003
(Inspec) datasets. For the Library dataset, as it is in Czech,
theMorphoDiTa lemmatizer2was used.
SemEval-2010 (Kim et al. 2010) consists of whole
plain texts from scientiﬁc articles. The dataset provides
keyphrases provided by authors and readers. As it is
common practice (Kim et al. 2010; Kontoulis, Papa-
giannopoulou, and Tsoumakas 2021), we use a combina-
tion of both in our experiments. Our validation dataset was
2http://hdl.handle.net/11858/00-097C-0000-0023-43CD-0created by randomly choosing a subset of the train set. As
the original data source does not explicitly provide the ti-
tles, which we need to formulate a query, we have manually
extracted the title of each document from the plain text.
Documents in this dataset have a well-formed structure.
They contain a title and abstract and are divided into sections
introduced with a headline. As we want to investigate the
inﬂuence of such structure on results, we have made an un-
structured version of this dataset. We downloaded the orig-
inal PDFs and used the GROBID3to get a structured XML
version of them. We kept only the text from the document’s
main body while the parts such as title, abstract, authors,
or section headlines were removed. Nevertheless, document
keyphrase annotations remain the same. We call this dataset
Unstructured-SemEval-2010. The name SemEval is used to
name these two collectively.
Inspec (Hulth 2003) contains a set of title-abstract pairs
collected from scientiﬁc articles. For each abstract, there are
two sets of keyphrases — controlled , which are restricted to
the Inspec thesaurus, and uncontrolled that can contain any
suitable keyphrase. To be comparable with previous works
(Hulth 2003; Liu and Iwaihara 2021), we used only the un-
controlled set in our experiments.
Library is a newly created dataset that takes advantage
of a large body of scanned documents, provided by Czech
libraries, that were converted to text by OCR software. This
way of getting the text is unreliable, so the dataset contains
many errors on the word and character level. The dataset
builds on the documents where the language was recognized
as ‘Czech’ by the OCR software.
All used documents in the original data source are rep-
resented by their signiﬁcant content (the average number
of characters per document is 529 276) and metadata. The
metadata contains (not for all) keyphrases and document lan-
guage annotations. We did not ask annotators to annotate
each document. Instead, we selected metadata ﬁelds used by
librarians as keyphrase annotations. So, our data and meta-
data come from the real-world environment of Czech li-
braries. We have ﬁltered out all documents with less than
ﬁve keyphrases.
Documents come from more than 290 categories. Vari-
ous topics such as mathematics, psychology, belles lettres,
music, and computer science are covered. Not all anno-
tated keyphrases can be extracted from the text. Consid-
ering the lemmatization, the test set annotations contain
about 53% of extractive keyphrases. Bibliographic ﬁeld Title
(MARC 2454) is used as the query. Note that the ﬁeld may
contain additional information to the title, such as authors.
Experimental Setup
The implemented system builds on PyTorch5and PyTorch
Lightning6tools. The BERT part of the model uses the
3https://github.com/kermitt2/grobid
4https://www.loc.gov/marc/bibliographic/bd245.html
5https://pytorch.org/
6https://www.pytorchlightning.ai/1359 19 370:120:140:160:180:2
input size [sentences]F1@10sentences only sentences + query
sentences only (unst) sentences + query (unst)
Figure 2: Results for SemEval-2010 and Unstructured-
SemEval-2010 test set. The light red and blue areas are con-
ﬁdence intervals with a conﬁdence level of 0.95. Each point
corresponds to an average of ﬁve runs.
implementation of BERT by Hugging Face7and it is ini-
tialized with pre-trained weights of bert-base-multilingual-
cased . These weights are also optimized during ﬁne-tuning.
The Adam optimiser with weight decay (Loshchilov and
Hutter 2017) is used in all the experiments. The evaluation
during training on the validation set is done every 4 000 op-
timization steps for the Library dataset and every 50 steps
for Inspec (25 for whole abstracts with titles). For SemEval
datasets, the number of steps differs among experiments.
Early stopping with patience 3 is applied, so the training
ends when the model stops improving. Batch size 128 is
used for experiments with the Library dataset, and batch
size 32 is used for Inspec and SemEval datasets. The learn-
ing rate 1E-06 is used for the experiments with SemEval
datasets, while it is set to the value of 1E-05 for all other
datasets. Inputs longer than a maximum input size are split
into sequences of roughly the same size in a way that for-
bids splitting of keyphrase spans. In edge cases (split is not
possible), the input is truncated. No predicted span is longer
than 6 tokens.
The ofﬁcial script for SemEval-2010 is used for evalua-
tion. However, the normalization of keyphrases is different
for the Library dataset as we have used the mentioned Czech
lemmatizer instead of the original stemmer. We use the F1
over the top ﬁve (F1@5) candidates for the Library dataset
and over the top ten (F1@10) for the rest.
Experiments
The performed experiments investigate the inﬂuence of
queries on four different datasets, the output quality with
various context sizes, and the impact of the document struc-
ture.
The ﬁrst set of experiments is performed on long docu-
ments with a well-formed structure from the SemEval-2010
7https://huggingface.co/1359 19 370255075
input size [sentences]too long inputs [%]sentences only sentences + query
Figure 3: The proportion of inputs longer than the maximum
input size for SemEval-2010 train set.
1359 19 370255075
input size [sentences]proportion [%]
Figure 4: The proportion of contexts containing at least one
section headline as a substring for SemEval-2010 test set.
dataset and compares them with SemEval’s unstructured
version. Figure 2 shows that inputs with a query are bet-
ter than those without a query, but the last point. For the
structured input, it can be seen that from the point with 19
sentences, the performance of input with query stops with
the fast growth. It correlates with Figure 3 showing the sat-
uration of the model input. Notice that from 19 sentences,
the input becomes more saturated, and the splitting strategy
starts shrinking contexts.
It is not surprising that the nominal values are lower for
unstructured inputs. On the other hand, it is clear that the
query has a bigger inﬂuence on the unstructured version, es-
pecially for short context sizes, because the average abso-
lute difference among results (with- and without a query) for
each context size is 2.2% compared to 1.29% for the struc-
tured one.
Looking at the curve of results with a query on an un-
structured version, we assume that the model can exploit a
document structure without explicitly tagging it with special
tokens because additional context size above the three sen-
tences is not much beneﬁcial compared to the case with the
document structure. This hypothesis is supported by the fact
that the proportion of the context containing structured infor-
mation grows with context size, as is demonstrated in Figure
4 showing the proportion of contexts containing a section
headline.
The second set of experiments was performed on our Li-
brary dataset. The results can be seen in Figure 5. We have
chosen F1@5 because only approximately half of the doc-
uments have ten and more keyphrases. Again, the results
show that queries are beneﬁcial. Also, it can be seen that
the shape of the query curve is similar to Unstructured-
SemEval-2010. The average absolute difference between the
version with and without query is now 3.1%. For F1@10, it1359 19 370:10:15
input size [sentences]F1@5sentences only sentences + query
Figure 5: Results for Library test set for various context
sizes. The light red area symbolizes a conﬁdence interval
with a conﬁdence level of 0.95. Each point is average from
three runs.
Inspec SemEval 2010
F1@10 F1@10
TextRank 15.28 6.55
KFBS + BK-Rank 46.62 15.59
DistilRoBERTa + TF-IDF - 16.2
context
[sentences]query
whole document 7 39.67 -
17 40.26 14.96
3 39.95 16.06
197 - 17.18
3 - 18.56
Table 2: Comparison of achieved results with other work.
KFBS + BK-Rank and TextRank is from (Liu and Iwaihara
2021). The DistilRoBERTa + TF-IDF is from (Kontoulis,
Papagiannopoulou, and Tsoumakas 2021). Our results are
averages from ﬁve runs.
is 2.3, which is close to the value for the unstructured version
of SemEval.
The last set of experiments is done on the Inspec dataset,
which has only titles and abstracts. The purpose is to in-
vestigate the inﬂuence of a query on short inputs containing
mainly salient sentences. Results are summarized in Table 2,
which also compares our results with other works. It shows
that the results for a single sentence, a single sentence with
a title, and whole abstract with a title are similar. The ex-
planation can be that the abstract contains mainly salient
sentences containing keyphrases, and also, the abstract it-
self deﬁnes the topic of the article. A similar observation
is presented in (Liu and Iwaihara 2021), where the version
without summarization gives similar results as the extraction
performed on a summary.
Conclusions
We have conducted experiments that show that query-based
keyphrase extraction is promising for processing long doc-
uments. Our experiments show the relationship between
the context size and the performance of the BERT-basedkeyphrase extractor. The developed model was evaluated on
four datasets; one of them is non-English. The datasets al-
lowed us to ﬁnd when the query-based approach is beneﬁ-
cial and when not. It was shown that a query gives no bene-
ﬁt when extracting keyphrases from abstracts. On the other
hand, it is beneﬁcial for long documents, particularly those
without a well-formed document structure on short context
sizes.
Acknowledgment
This work was supported by the Technology Agency of the
Czech Republic, Grant FW03010656 – MASAPI: Multilin-
gual assistant for searching, analysing and processing infor-
mation and decision support. The computation used the in-
frastructure supported by the Ministry of Education, Youth
and Sports of the Czech Republic through the e-INFRA CZ
(ID:90140).
References
Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019.
BERT: Pre-training of deep bidirectional transformers for lan-
guage understanding. In Proceedings of the 2019 NAACL Con-
ference: Human Language Technologies, Volume 1 (Long and
Short Papers) , 4171–4186. Minneapolis, Minnesota: Associa-
tion for Computational Linguistics.
Hulth, A. 2003. Improved automatic keyword extraction given
more linguistic knowledge. In Proceedings of the 2003 EMNLP
Conference , 216–223.
Kim, S. N.; Medelyan, O.; Kan, M.-Y .; and Baldwin, T. 2010.
SemEval-2010 task 5 : Automatic keyphrase extraction from
scientiﬁc articles. In Proceedings of the 5th SemEval , 21–26.
Uppsala, Sweden: Association for Computational Linguistics.
Kontoulis, C. G.; Papagiannopoulou, E.; and Tsoumakas, G.
2021. Keyphrase extraction from scientiﬁc articles via extrac-
tive summarization. In Proceedings of the Second SDP Work-
shop , 49–55. Online: Association for Computational Linguis-
tics.
Lin, Z.-L., and Wang, C.-J. 2019. Keyword extraction
with character-level convolutional neural tensor networks. In
Paciﬁc-Asia Conference on Knowledge Discovery and Data
Mining , 400–413. Springer.
Liu, T., and Iwaihara, M. 2021. Supervised learning of
keyphrase extraction utilizing prior summarization. In Ke, H.-
R.; Lee, C. S.; and Sugiyama, K., eds., Towards Open and
Trustworthy Digital Societies , 157–166. Cham: Springer In-
ternational Publishing.
Loshchilov, I., and Hutter, F. 2017. Decoupled weight decay
regularization. arXiv preprint arXiv:1711.05101 .
Mihalcea, R., and Tarau, P. 2004. Textrank: Bringing order
into text. In Proceedings of the 2004 conference on EMNLP ,
404–411.
Rose, S.; Engel, D.; Cramer, N.; and Cowley, W. 2010. Au-
tomatic keyword extraction from individual documents. Text
mining: applications and theory 1:1–20.
Sahrawat, D.; Mahata, D.; Zhang, H.; Kulkarni, M.; Sharma,
A.; Gosangi, R.; Stent, A.; Kumar, Y .; Shah, R. R.; and Zim-
mermann, R. 2020. Keyphrase extraction as sequence labeling
using contextualized embeddings. In ECIR , 328–335. Springer.