=Paper= {{Paper |id=Vol-2106/paper4 |storemode=property |title=Time-Aware and Corpus-Specific Entity Relatedness |pdfUrl=https://ceur-ws.org/Vol-2106/paper4.pdf |volume=Vol-2106 |authors=Nilamadhaba Mohapatra,Vasileios Iosifidis,Asif Ekbal,Stefan Dietze,Pavlos Fafalios |dblpUrl=https://dblp.org/rec/conf/esws/MohapatraIEDF18 }} ==Time-Aware and Corpus-Specific Entity Relatedness== https://ceur-ws.org/Vol-2106/paper4.pdf

Time-Aware and Corpus-Specific
Entity Relatedness

Nilamadhaba Mohapatra1,2 , Vasileios Iosifidis1 , Asif Ekbal2 ,
Stefan Dietze1 , and Pavlos Fafalios1
1
L3S Research Center, University of Hannover, Germany
{iosifidis, dietze, fafalios}@L3S.de
2
Indian Institute of Technology, Patna, India
{nilamadhaba.mtmc15, asif}@iitp.ac.in

Abstract. Entity relatedness has emerged as an important feature in a
plethora of applications such as information retrieval, entity recommen-
dation and entity linking. Given an entity, for instance a person or an
organization, entity relatedness measures can be exploited for generat-
ing a list of highly-related entities. However, the relation of an entity
to some other entity depends on several factors, with time and context
being two of the most important ones (where, in our case, context is de-
termined by a particular corpus). For example, the entities related to the
International Monetary Fund are different now compared to some years
ago, while these entities also may highly differ in the context of a USA
news portal compared to a Greek news portal. In this paper, we propose
a simple but flexible model for entity relatedness which considers time
and entity aware word embeddings by exploiting the underlying corpus.
The proposed model does not require external knowledge and is language
independent, which makes it widely useful in a variety of applications.

Keywords: Entity Relatedness · Word2Vec · Entity Embeddings

1 Introduction

Entity relatedness is the task of determining the degree of relatedness between
two entities. Measures for entity relatedness facilitate a wide variety of applica-
tions such as information retrieval, entity recommendation and entity linking.
Traditional approaches consider the structural similarity in a given graph and
lexical features [5, 6, 10–12, 15], or make use of Wikipedia-based entity distribu-
tions and embeddings [1, 2].
Recent works on related topics have supported the hypothesis that the con-
text of entities is temporal in nature. Specifically, [4] has shown that prior proba-
bilities often change across time, while [17] showed that the effectiveness of entity
recommendations for keyword queries is affected by the time dimension. In addi-
tion, [16] introduced the notion of contextual entity relatedness and showed that
entity relatedness is both time and context dependent. In that work, context
refers to topic (aspect), i.e., the goal is to find the most related entities given
2 N. Mohapatra et al.

an entity and an aspect (e.g., relationship or career are two different aspects for
a person). Finally, [13] showed that exploiting different temporal versions of a
knowledge base, each reflecting a different time period, can improve the accuracy
of entity relatedness.
On the other hand, entity relatedness is also strongly dependent on the corpus
context at hand. For instance, given a search application operating over a collec-
tion of German news articles from summer 2014 and the query entity 2014 FIFA
World Cup (https://en.wikipedia.org/wiki/2014_FIFA_World_Cup), a list
of related entities should include Germany national football team, Argentina na-
tional football team, Mario Götze, and Brazil. However, considering a collection
of Greek articles of the same time period, the top related entities might include
Greece national football team, Costa Rica national football team, and Sokratis
Papastathopoulos (entities that are not important or popular in German news).
To this end, our work considers entity relatedness as a measure which is
strongly dependent on both the time aspect and the corpus of the underlying
application. Contrary to existing approaches that train embeddings on general-
purpose corpora, in particular Wikipedia [1, 2], we compute entity relatedness
given a specific collection of documents that spans a particular time period. Our
approach exploits entities extracted from the underlying corpus for building
time and entity aware word embeddings. The evaluation results show that the
proposed model outperforms similar time and entity agnostic models.
The remainder of this paper is organized as follows: Section 2 defines the
problem of time-aware and corpus specific entity relatedness, Section 3 details
the proposed method, Section 4 presents evaluation results, and finally Section
5 concludes the paper and discusses interesting directions for future research.

2 Problem Definition

Let D be a corpus of documents, e.g., a collection of news articles, covering the
time period TD = [τs , τe ] (where τs , τe are two different time points with τs < τe ).
For a document d ∈ D, let Ed denote all entities mentioned in d extracted using
an entity linking method [14], where each entity is associated with a unique URI
in a knowledge base like Wikipedia. This list of extracted entities may include
persons, locations, etc., but also events (e.g., US 2016 presidential election) and
more abstract concepts such as democracy or abortion. Finally, let ED denote
all entities mentioned in documents of D.
Consider now a collection of documents D covering a time period TD and
a set of entities ED extracted from and prevalent in D. Given i) one or more
query entities Eq and ii) a time period of interest Tq ⊆ TD , the task “time-
aware and corpus-specific entity relatedness” focuses on finding a top-k list of
entities Ek ⊂ ED related to the query entities Eq in terms of both Tq and D. We
model this task as a ranking problem, where we first generate a list of candidate
entities Ec ⊂ ED and then rank this list of entities based on their relevance to
the query entities. For generating the list of candidate entities, one can follow
an approach similar to [17] which exploits Wikipedia links, DBpedia and entity
Time-Aware and Corpus-Specific Entity Relatedness 3

co-occurrences in the annotated corpus, or consider the connectedness of the
query entity with other entities in Social Media [3].

3 Approach
Word embeddings provide a distributed representation of words in the semantic
space. We generate word vectors following the distributional models proposed
in [8, 9] and the well-known Word2Vec tool.

3.1 Time-Aware Word Vector Similarity
We group all documents in D into time-specific subsets C = (C1 , . . . , Cn ) based
on a fixed time granularity ∆ (e.g., week, month, or year). We then preprocess
the documents of each subset Ci ∈ C and train a Word2Vec Continuous Bag
of Words (CBOW) model. Each model generates a matrix keys × dimension(d)
where keys represents a word from the corpus Ci for which a d-dimensional
vector exists in the trained model.
To find the relatedness score between a query entity eq ∈ Eq and a candidate
entity ec ∈ Ec , we compute the cosine similarity of their word vectors in the
corresponding trained model for the input time period Tq . In our experiments,
for finding the words that represent a query or candidate entity, we use the last
part of the entity’s Wikipedia URI by first replacing the underscore character
with the space character and removing any text in parentheses. Based on the
underlying knowledge base, one could use here other approaches, e.g., exploit
the entity’s label in DBpedia. For multi-word entities, we compute the average
vector of the constituent words, while for inputs consisting of more than one
entity, we compute the average of all entity vectors.

3.2 Considering the Entity Annotations
There are two limitations in dealing with the embedding vectors obtained in
the previous model: (i) for multi-word entities the average of the word vectors
of the constituent tokens is computed, i.e., n-grams are ignored (consider for
example the entity United Nations), and (ii) the same entity mention (surface
form) may refer to different entities (e.g., Kobe may refer to the basketball player
Kobe Bryant or the Japanese city). To cope with these problems, we exploit the
entity annotations ED . Recall that each extracted entity is associated with a
unique URI in a knowledge base (e.g., Wikipedia or DBpedia). As also suggested
in [9] for the case of phrases, we preprocess the documents and replace the entity
mentions with unique IDs. The modified text corpora are now used for training
the Word2Vec model, where the word vector for an entity is now calculated using
its unique ID.
Now, the similarity score between a query entity eq ∈ Eq and a candidate
entity ec ∈ Ec is the cosine similarity of their word vectors using the correspond-
ing modified (with entity IDs) Word2Vec model for the input time period Tq .
4 N. Mohapatra et al.

Formally:
Sim(eq , ec , Tq ) = cosine(wordvec(id(eq ), Tq ), wordvec(id(ec ), Tq )) (1)

where id(e) is the ID of entity e, and wordvec(w, Tq ) returns the word vector of
token w using the corresponding model for the input time period Tq .

3.3 Relaxing the Time Boundaries
An important event related to the query entities may happened very close to
the boundaries of the time period of interest Tq . This means that two entities
might be highly related some time before or after Tq . To cope with this problem,
the similarity score can also consider the Word2Vec models for the time periods
before and after Tq . Let Tq−1 and Tq+1 be the time periods of granularity ∆ before
and after Tq , respectively. The similarity score between a query entity eq ∈ Eq
and a candidate entity ec ∈ Ec can now be defined as follows:

Sim0 (eq , ec , Tq ) = Sim(eq , ec , Tq )·w1 +Sim(eq , ec , Tq−1 )·w2 +Sim(eq , ec , Tq+1 )·w3 (2)

where w1 , w2 and w3 are the weights of the models of Tq , Tq−1 and Tq+1 ,
respectively, with w1 + w2 + w3 = 1.0.
This modification can increase the ranking of an important candidate entity
that co-occurs frequently with a query entity some time before or after Tq , but
at the same time can decrease the ranking of an entity co-occurring with a query
entity during Tq . Thus, we should avoid using a very small w1 value.

4 Evaluation Results
Our objective is to evaluate the effectiveness of the proposed approach and com-
pare it with similar but time and entity agnostic models. We use the dataset
and ground truth provided by [17] for the problem of time-aware entity recom-
mendation3 . The dataset provides candidate entities and relevance judgments
for 22 keyword queries, where each query corresponds to a particular date range
(month). The dataset also provides a set of more than 8M news articles spanning
a period of 7 months (Jul’14-Jan’15), annotated with Wikipedia entities. For each
of the keyword queries, we manually specified the corresponding entities (since,
in our problem, the input is an unambiguous entity URI). For instance, for the
query Tour de France Nibali (with date range 07/2014), the query entities are:
Tour de France (https://en.wikipedia.org/wiki/Tour_de_France) and Vin-
cenzo Nibali (https://en.wikipedia.org/wiki/Vincenzo_Nibali). We also
remove the query entities from the set of candidate entities since these are the
input in our entity relatedness problem.
We build 7 CBOW models using the proposed approach for each month from
Jul’14 to Jan’15 by considering only the articles in English and using the default
3
http://km.aifb.kit.edu/sites/ter/
Time-Aware and Corpus-Specific Entity Relatedness 5

Word2Vec setting: 300 dimensions, 5 words window size, 5 minimum word count
(as also used in [9]). We also experimented with varied dimension (300, 400,
500, 600, 700) and context size (5, 8, 10) but the results were not significantly
improved. Then, we compare the effectiveness of our approach on ranking the
candidate entities with two entity-agnostic baselines (using the same setting):
i) one that considers word embeddings computed from the entire collection of
documents (i.e., time-agnostic), and ii) one that considers month-wise word em-
beddings (i.e., time-aware).4
Table 1 shows the results of normalized Discounted Cumulative Gain (nDCG)
[7] for different top-k lists, without considering time boundary relaxation (i.e.,
for w1 = 1.0). Regarding the two entity-agnostic baselines, we notice that the
time-aware modeling outperforms the time-agnostic one in all the cases. This
improvement is statistically significant for all values of k apart from k = 5 (paired
t-tests, a-level 5%). Regarding the proposed model (time and entity aware), the
evaluation results show that it outperforms the baselines in all the cases, while
the improvement is statistically significant for all values of k. This justifies that
this model solves the problem of ambiguity that occurs due to the variations of
different entity mentions.

Table 1. Evaluation results (‡ indicates statistically significant improvement).
nDCG@k Time+Entity Agnostic Entity Agnostic Time+Entity Aware
k=5 0.3210 0.3653 0.4999 ‡
k=10 0.3748 0.4113 ‡ 0.5402 ‡
k=20 0.4546 0.4971 ‡ 0.6115 ‡
k=30 0.5092 0.5704 ‡ 0.6562 ‡

Table 2 shows the effect of relaxing the time boundaries for different w1
values, where w2 = w3 . In general we see that the effect of relaxing the time
boundaries is very small in almost all cases. We notice that considering only the
query time period Tq provides the best results in all cases apart from k = 5
where w1 = 0.9 slightly performs better. Moreover, we notice that as the value
of w1 decreases (which means increased consideration of Tq−1 and Tq+1 ), the
effectiveness of the model gets worse. This is an expected result since, for the
majority of the query entities in the dataset, the important event related to these
entities did not happen very close to the time boundaries.
An example for which the relaxation of the time boundaries has a positive
impact on the ranking is for the entity 2014 FIFA World Cup. For w1 = 0.8,
nDCG@5 increases from 0.45 to 0.51. Notice that the time period of interest
for this entity is July 2014, however the tournament started on June 12, 2014.
Another example is the query entity Tim Cook (the CEO of Apple). For w1 = 0.8,
nDCG@5 increases from 0.58 to 0.62. Here the query time period is October 2014,
4
Note that our approach cannot be compared with [17] because this work addresses
the different problem of entity recommendation, where the input is a free-text query,
not an unambiguous entity URI like in our case.
6 N. Mohapatra et al.

however an important event related to Tim Cook happened at the end of the
month (he publicly announced that he is gay on October 29).
We see that the relaxation of time boundaries can positively affect entity
relatedness when: i) an important event related to the query entity happened
very close to the boundaries of the query time period, and ii) the query entity
actually corresponds to an event which spans a long time period. Detecting such
cases where time boundaries relaxation should be applied is beyond the scope
of this paper but an important direction for future work.

Table 2. Effect of time boundary relaxation.
nDCG@k w1 = 1.0 w1 = 0.9 w1 = 0.8 w1 = 0.7 w1 = 0.6
k=5 0.4999 0.5017 0.4990 0.4933 0.4890
k=10 0.5402 0.5332 0.5358 0.5296 0.5291
k=20 0.6115 0.6039 0.5971 0.5932 0.5893
k=30 0.6562 0.6517 0.6451 0.6403 0.6371

5 Conclusion
We have proposed a flexible model for entity relatedness that considers time-
dependent and entity-aware word embeddings by exploiting the corpus of the
underlying application. The results of a preliminary evaluation have shown that
the proposed approach significantly outperforms similar but time and entity-
agnostic models.
As regards future work, an interesting direction is to extend the proposed
method for supporting arbitrary time intervals, which may require joining the
results from many models of smaller granularity. However, for supporting very
short time periods (e.g., day), this may also require the creation of a large number
of Word2Vec models. Regarding the relaxation of time boundaries, there is a need
of methods that can identify the most reasonable time window to consider given
the query entity and time period (by detecting, for example, periods of increased
entity popularity [3]). Finally, we plan to extensively evaluate the effectiveness
of our approach and compare it with state-of-the-art entity relatedness methods
using a variety of corpora of different contexts and time periods.

Acknowledgements
The work was partially funded by the European Commission for the ERC Ad-
vanced Grant ALEXANDRIA under grant No. 339233.

References
1. Aggarwal, N., Buitelaar, P.: Wikipedia-based distributional semantics for entity
relatedness. In: 2014 AAAI Fall Symposium Series. vol. 192 (2014)
Time-Aware and Corpus-Specific Entity Relatedness 7

2. Basile, P., Caputo, A., Rossiello, G., Semeraro, G.: Learning to rank entity relat-
edness through embedding-based features. In: International Conference on Appli-
cations of Natural Language to Information Systems. pp. 471–477. Springer (2016)
3. Fafalios, P., Iosifidis, V., Stefanidis, K., Ntoutsi, E.: Multi-aspect Entity-centric
Analysis of Big Social Media Archives. In: 21st International Conference on Theory
and Practice of Digital Libraries (TPDL’17). Thessaloniki, Greece (2017)
4. Fang, Y., Chang, M.W.: Entity linking on microblogs with spatial and temporal
signals. Transactions of the Association for Computational Linguistics 2 (2014)
5. Hoffart, J., Seufert, S., Nguyen, D.B., Theobald, M., Weikum, G.: Kore: keyphrase
overlap relatedness for entity disambiguation. In: Proceedings of the 21st ACM
international conference on Information and knowledge management. pp. 545–554.
ACM (2012)
6. Hulpuş, I., Prangnawarat, N., Hayes, C.: Path-based semantic relatedness on linked
data and its use to word and entity disambiguation. In: International Semantic Web
Conference. pp. 442–457. Springer (2015)
7. Järvelin, K., Kekäläinen, J.: Cumulated gain-based evaluation of ir techniques.
ACM Transactions on Information Systems (TOIS) 20(4), 422–446 (2002)
8. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word repre-
sentations in vector space. arXiv preprint arXiv:1301.3781 (2013)
9. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed repre-
sentations of words and phrases and their compositionality. In: Advances in neural
information processing systems. pp. 3111–3119 (2013)
10. Milne, D., Witten, I.H.: An effective, low-cost measure of semantic relatedness
obtained from wikipedia links. In: In Proceedings of AAAI 2008 (2008)
11. Nunes, B.P., Dietze, S., Casanova, M.A., Kawase, R., Fetahu, B., Nejdl, W.: Com-
bining a co-occurrence-based and a semantic measure for entity linking. In: Ex-
tended Semantic Web Conference. pp. 548–562. Springer (2013)
12. Ponza, M., Ferragina, P., Chakrabarti, S.: A two-stage framework for computing
entity relatedness in wikipedia. In: Proceedings of the 2017 ACM on Conference
on Information and Knowledge Management. pp. 1867–1876. ACM (2017)
13. Prangnawarat, N., Hayes, C.: Temporal evolution of entity relatedness using
wikipedia and dbpedia. 3rd Workshop on Managing the Evolution and Preser-
vation of the Data Web (2017)
14. Shen, W., Wang, J., Han, J.: Entity linking with a knowledge base: Issues, tech-
niques, and solutions. IEEE Transactions on Knowledge and Data Engineering
27(2), 443–460 (2015)
15. Strube, M., Ponzetto, S.P.: Wikirelate! computing semantic relatedness using
wikipedia. In: AAAI. vol. 6, pp. 1419–1424 (2006)
16. Tran, N.K., Tran, T., Niederée, C.: Beyond time: Dynamic context-aware entity
recommendation. In: European Semantic Web Conference. Springer (2017)
17. Zhang, L., Rettinger, A., Zhang, J.: A probabilistic model for time-aware entity
recommendation. In: International Semantic Web Conference. Springer (2016)