=Paper=
{{Paper
|id=Vol-2699/paper14
|storemode=property
|title=ESSTER at the EYRE 2020 Entity Summarization Task
|pdfUrl=https://ceur-ws.org/Vol-2699/paper14.pdf
|volume=Vol-2699
|authors=Qingxia Liu,Gong Cheng,Yuzhong Qu
|dblpUrl=https://dblp.org/rec/conf/cikm/LiuCQ20
}}
==ESSTER at the EYRE 2020 Entity Summarization Task==
https://ceur-ws.org/Vol-2699/paper14.pdf
ESSTER at the EYRE 2020 Entity Summarization Task
Qingxia Liua , Gong Chenga and Yuzhong Qua
a State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing 210023, China
Abstract
Entity summaries provide human users with the key information about an entity. In this system paper, we present the im-
plementation of our entity summarizer ESSTER. It aims at generating entity summaries that contain structurally important
triples and exhibit high readability and low redundancy. For structural importance, we exploit the global and local character-
istics of properties and values in RDF data. For readability, we learn the familarity of properties from a text corpus. To reduce
redundancy, we perform logical reasoning and compute textual and numerical similarity between triples. ESSTER solves a
combinatorial optimization problem to integrate these features. It achieves state-of-the-art results on the ESBM v1.2 dataset.
Keywords
Entity summarization, readability, redundancy
1. Introduction predicate and the value π£ is π‘βs object. When π is the
object of π‘, the property π is the inverse of π‘βs predicate
In RDF data, an entity is described by a possibly large and the value π£ is π‘βs subject. For convenience, we
set (e.g., hundreds) of RDF triples. The entity summa- define prop(π‘) = π and val(π‘) = π£. Given an integer
rization task is to automatically generate a compact size constraint π, an entity summary π for π is a subset
summary to provide human users with the key infor- of desc(π) satisfying |π| β€ π.
mation about an entity. Specifically, an entity sum-
mary is a size-constrained subset of triples selected
from an entity description. Current methods [1, 2, 3, 4, 3. Implementation of ESSTER
5, 6] are mainly focused on selecting important triples,
but ignore the reading experience of human users. In ESSTER considers structural importance, readability,
this system paper, we present the implementation of and redundancy. Below we present their computation
our entity summarizer named ESSTER [7].1 It aims at and finally integrate them by solving a combinatorial
generating entity summaries of structural importance, optimization problem.
high readability, and low redundancy. Improving tex-
tual readability and reducing information redundancy 3.1. Structural Importance
help to enhance the reading experience of users. Ex-
periments on the ESBM v1.2 dataset [8] show that ES- We measure the structural importance of a triple π‘ from
STER achieves state-of-the-art results. two perspectives.
First, globally popular properties often reflect im-
portant aspects of entities, while globally unpopular
2. Task Definition values are informative. Therefore, we compute the
global importance of a triple as follows:
RDF data is a set of subject-predicate-object triples π .
For an entity π, its description desc(π) is the subset of glb(π‘) = ppopglobal (π‘) β
(1 β vpop(π‘)) ,
triples in π such that π is the subject or the object. Each log(pfreqglobal (π‘) + 1)
triple π‘ β desc(π) provides a property-value pair β¨π, π£β© ppopglobal (π‘) = , (1)
log(|πΈ| + 1)
for π. When π is the subject of π‘, the property π is π‘βs
log(vfreq(π‘) + 1)
vpop(π‘) = ,
Proceedings of the CIKM 2020 Workshops, October 19-20, 2020, log(|π | + 1)
Galway, Ireland
email: qxliu2013@smail.nju.edu.cn (Q. Liu); gcheng@nju.edu.cn where πΈ is the set of all entities described in RDF data π ,
(G. Cheng); yzqu@nju.edu.cn (Y. Qu) pfreqglobal (π‘) is the number of entity descriptions in π
url: http://ws.nju.edu.cn/~gcheng (G. Cheng);
http://ws.nju.edu.cn/~yzqu (Y. Qu)
where prop(π‘) appears, and vfreq(π‘) is the number of
orcid: 0000-0001-6706-3776 (Q. Liu); 0000-0003-3539-7776 (G. triples in π where val(π‘) is the value.
Cheng); 0000-0003-2777-8149 (Y. Qu) Second, multi-valued properties are intrinsically pop-
Β© 2020 Copyright for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0). ular compared with single-valued properties. To com-
CEUR
CEUR Workshop Proceedings (CEUR-WS.org)
pensate for this, we penalize multi-valued properties
http://ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
1 https://github.com/nju-websoft/ESSTER
by using local popularity. We compute the local im- and val(π‘π ) are equal, and rdfs:subPropertyOf is a
portance of a triple as follows: relation between prop(π‘π ) and prop(π‘π ).
Otherwise, we rely on the similarity between prop-
loc(π‘) = (1 β ppoplocal (π‘)) β
vpop(π‘) , erties and the similarity between values:
log(pfreqlocal (π‘) + 1) (2)
ppoplocal (π‘) =
log(|desc(π)| + 1)
, sim(π‘π , π‘π ) = max{simp (π‘π , π‘π ), simv (π‘π , π‘π ), 0} , (6)
where pfreqlocal (π‘) is the number of triples in desc(π) where for simp we use the ISub string similarity [9].
where prop(π‘) is the property. For simv , we differentiate between two cases.
Finally, we compute structural importance: In the first case, val(π‘π ) and val(π‘π ) are both numer-
ical values. We compute
πstruct (π‘) = πΌ β
glb(π‘) + (1 β πΌ) β
loc(π‘) , (3) {
β1 val(π‘π ) β
val(π‘π ) β€ 0 ,
where πΌ β [0, 1] is a parameter to tune. simv (π‘π , π‘π ) = min{val(π‘π ),val(π‘π )}
otherwise .
max{val(π‘π ),val(π‘π )}
(7)
3.2. Textual Readability In all other cases, we simply use ISub for simv .
To generate readable summaries, we measure the fa-
miliarity of a triple π‘ based on its property prop(π‘). A 3.4. Combinatorial Optimization
property is familiar to users if it is often used in an
We formulate entity summarization as a 0-1 quadratic
open-domain corpus. Specifically, given a text corpus
knapsack problem (QKP), and we solve it using a heuris-
of π΅ documents where π documents have been read
tic algorithm [10].
by the user, let π(π‘) be the number of documents where
Specifically, we define the profit of choosing two
the name of prop(π‘) appears. We compute
triples π‘π , π‘π for a summary:
min(π(π‘),π) π(π‘) π΅βπ(π‘) {
( π ) β
( πβπ )
π(π‘) = β π΅
β
familarity(π) , (1 β πΏ) β
(πstruct (π‘π ) + πtext (π‘π )) π = π ,
π=0 (π ) profitπ,π =
πΏ β
(βsim(π‘π , π‘π )) π β π,
familarity(π) =
log(π + 1)
. (8)
log(π΅ + 1) where πΏ β [0, 1] is a parameter to tune.
(4) Finally, our goal is to
Here, π represents the number of documents the user |desc(π)| |desc(π)|
has read where the name of prop(π‘) appears, based maximize β β profitπ,π β
π₯π β
π₯π ,
on which familarity(π) gives the degree of fami- π=1 π=π
larity of prop(π‘) to the user. However, it is difficult |desc(π)| (9)
to know π in practice, so π(π‘) computes the expected subject to β π₯π β€ π ,
value of familarity(π). For simplicity, we assume π=1
π is a constant. In the experiments we set π = 40 and π₯π β {0, 1} for all π = 1 β¦ |desc(π)| .
we use the Google Books Ngram2 as our corpus.
Finally, we compute textual readability:
4. Experiments
πtext (π‘) = log(π(π‘) + 1). (5)
4.1. Settings
3.3. Information Redundancy We use the ESBM v1.2 dataset [8]. It provides ground-
truth summaries under π = 5 and π = 10 for entities
To reduce redundancy in summaries, we measure the in DBpedia and LinkedMDB. We follow the provided
similarity between two triples π‘π , π‘π in various ways. training-development-test splits for 5-fold cross vali-
First, we perform logical reasoning to measure on- dation, and we use the training and development sets
tological similarity. We define sim(π‘π , π‘π ) = 1 if prop(π‘π ) for tuning our parameters πΌ and πΏ by grid search in the
and prop(π‘π ) are rdf:type, and rdfs:subClassOf range of 0β1 with 0.01 increments. We use F1 score as
is a relation between val(π‘π ) and val(π‘π ); or if val(π‘π ) the evaluation metric.
2 http://books.google.com/ngrams
Table 1 [2] G. Cheng, T. Tran, Y. Qu, RELIN: relatedness and
F1 Scores informativeness-based centrality for entity sum-
DBpedia LinkedMDB marization, in: ISWCβ11, Part I, 2011, pp. 114β
π = 5 π = 10 π = 5 π = 10 129. doi:10.1007/978-3-642-25073-6_8.
RELIN 0.242 0.455 0.203 0.258 [3] K. Gunaratna, K. Thirunarayan, A. P. Sheth,
DIVERSUM 0.249 0.507 0.207 0.358 FACES: diversity-aware entity summarization
FACES 0.270 0.428 0.169 0.263 using incremental hierarchical conceptual clus-
FACES-E 0.280 0.488 0.313 0.393 tering, in: AAAIβ15, 2015, pp. 116β122.
CD 0.283 0.513 0.217 0.331 [4] K. Gunaratna, K. Thirunarayan, A. P. Sheth,
LinkSUM 0.287 0.486 0.140 0.279 G. Cheng, Gleaning types for literals in RDF
BAFREC 0.335 0.503 0.360 0.402
triples with application to entity summarization,
KAFCA 0.314 0.509 0.244 0.397
MPSUM 0.314 0.512 0.272 0.423
in: ESWCβ16, 2016, pp. 85β100. doi:10.1007/
ESSTER 0.324 0.521 0.365 0.452 978-3-319-34129-3_6.
[5] A. Thalhammer, N. Lasierra, A. Rettinger,
LinkSUM: Using link analysis to summarize en-
4.2. Results tity data, in: ICWEβ16, 2016, pp. 244β261. doi:10.
1007/978-3-319-38791-8_14.
Table 1 presents the evaluation results. We compare [6] H. Kroll, D. Nagel, W.-T. Balke, BAFREC: Balanc-
with known results of existing unsupervised entity sum- ing frequency and rarity for entity characteriza-
marizers [8]. On DBpedia under π = 5, BAFREC [6] tion in linked open data, in: EYREβ18, 2018.
achieves the highest F1 score, and is closely followed [7] Q. Liu, G. Cheng, Y. Qu, Entity summarization
by ESSTER. In all the other three settings, ESSTER out- with high readability and low redundancy, Sci.
performs all the baselines. Overall, ESSTER achieves Sin. Inform. 50 (2020) 845β861. doi:10.1360/
state-of-the-art results on ESBM v1.2. SSI-2019-0291.
[8] Q. Liu, G. Cheng, K. Gunaratna, Y. Qu, ESBM:
an entity summarization benchmark, in:
5. Conclusion ESWCβ20, 2020, pp. 548β564. doi:10.1007/
In this system paper, we presented the implementa- 978-3-030-49461-2_32.
tion of our entity summarizer ESSTER. By integrat- [9] G. Stoilos, G. B. Stamou, S. D. Kollias, A string
ing structural importance, textual readability, and in- metric for ontology alignment, in: ISWCβ05,
formation redundancy via combinatorial optimization, 2005, pp. 624β637. doi:10.1007/11574620_45.
ESSTER achieves state-of-the-art results among unsu- [10] Z. Yang, G. Wang, F. Chu, An effective GRASP
pervised entity summarizers on the ESBM v1.2 dataset. and tabu search for the 0-1 quadratic knapsack
However, the results are not comparable with super- problem, Comput. Oper. Res. 40 (2013) 1176β
vised neural entity summarizers [11, 12]. 1185. doi: 10.1016/j.cor.2012.11.023 .
For the future work, we will consider more powerful [11] Q. Liu, G. Cheng, Y. Qu, Deeplens: Deep learning
measures of readability and redundancy, and will in- for entity summarization, in: DL4KGβ20, 2020.
corporate these features into a neural network model. [12] J. Li, G. Cheng, Q. Liu, W. Zhang, E. Kharlamov,
K. Gunaratna, H. Chen, Neural entity summa-
rization with joint encoding and weak supervi-
Acknowledgments sion, in: IJCAIβ20, 2020, pp. 1644β1650. doi:10.
24963/ijcai.2020/228.
This work was supported by the National Key R&D
Program of China (2018YFB1004300) and by the NSFC
(61772264).
References
[1] Q. Liu, G. Cheng, K. Gunaratna, Y. Qu, Entity
summarization: State of the art and future chal-
lenges, CoRR abs/1910.08252 (2019).