=Paper=
{{Paper
|id=Vol-1963/paper505
|storemode=property
|title=Using Word Embeddings for Search in Linked Data with Ontodia
|pdfUrl=https://ceur-ws.org/Vol-1963/paper505.pdf
|volume=Vol-1963
|authors=Gerhard Wohlgenannt,Nikolay Klimov,Dmitry Mouromtsev,Daniil Razdyakonov,Dmitry Pavlov,Yury Emelyanov
|dblpUrl=https://dblp.org/rec/conf/semweb/WohlgenanntKMRP17a
}}
==Using Word Embeddings for Search in Linked Data with Ontodia==
https://ceur-ws.org/Vol-1963/paper505.pdf
Using Word Embeddings for Search in Linked
Data with Ontodia
Gerhard Wohlgenannt1 , Nikolay Klimov1 , Dmitry Mouromtsev1 , Daniil
Razdyakonov2 , Dmitry Pavlov2 , and Yury Emelyanov2
1
Intern. Lab. of Information Science and Semantic Technologies, ITMO University,
St. Petersburg, Russia http://en.ifmo.ru/en
2
Vismart Ltd., St. Petersburg, Russia https://vismart.biz
Abstract. Ontodia is an open-source diagramming and visual explo-
ration tool for linked data and ontologies. Here, we present an extension
of the Ontodia data query functionalities. We evaluate different types
and configurations of word embeddings for improving recall and flexibil-
ity of the Ontodia natural language interface. The demonstration will
focus especially on the new query functionalities, where Ontodia will be
applied to Wikidata as underlying dataset.
Keywords: Ontodia, word embeddings, Wikidata, Linked Data, visual
interface
1 Introduction
One of the key challenges of the Semantic Web and Linked Data is to make
the contents of datasets available to users which lack the skills to write semantic
queries and understand the underlying data schemata [1]. Ontodia [4] provides vi-
sual data exploration for Linked Data in a step-by-step diagrammatic approach.
In this demonstration, we combine Ontodia with Wikidata as a knowledge graph.
Wikidata3 is a free and open knowledge base and the central data storage for
projects like Wikipedia.
Ontodia allows users to understand and explore datasets, and to answer
information needs in a diagrammatic way. It combines its visual interface with
textual search for entities and properties in the dataset. Currently the retrieval of
entities and properties is limited to exact lexical matches with the labels defined
in the dataset. As a simple example, if a user searches for persons married with
a given person, they might not get any results if they did not use the correct
label spouse for their query.
In this work, we describe and evaluate improvements to the natural language
(NL) interface of Ontodia. The system finds and ranks properties related to a
user query using distributional semantics. We evaluate various types of word
embeddings against the pre-defined set of aliases for the Wikidata properties. In
the demonstration, we plan to present Ontodia and its improved search features,
and to discuss our experiences of using word embeddings in querying linked data.
3
https://www.wikidata.org
2 Authors Suppressed Due to Excessive Length
2 Related Work
We utilize distributional semantics in the form of word embeddings to enrich
the Ontodia NL interface. Word embeddings are language modeling techniques
that transform the vocabulary of a given corpus into a continuous and low-
dimensional vector space representation. Word embeddings have been applied
successfully to many NL processing problems, from word similarity estimations
to more complex tasks [3]. Shekarpour et al. [6] describe the challenges in Ques-
tion Answering on Linked Data. Word embeddings and deep learning techniques
are prominently listed as promising techniques for future investigation.
3 System Description
Ontodia4 is a free online RDF and OWL diagramming tool. One of its main use-
cases is the visual exploration of linked data sets, and the sharing of information
found. In the presented application, Ontodia is used inside the metaphacts5
platform to explore the Wikidata dataset.
The application described in this demonstration paper is found at:
http://ontodia-prop-suggest.apps.vismart.biz/wikidata.html. To use
the system, a user can search eg. for “Van Gogh” among the Instances, and
then pull the entity onto the canvas in the center of the window.
Fig. 1. Searching the properties related to “family” of entity Van Gogh
In this demo, we focus on improving the NL interface of Ontodia. After
searching instances and classes in the left-hand side of the interface, users can
filter the list of properties of the entities displayed on the canvas. Previously,
only properties with labels exactly matching the search term were found. For
this demo, we experimented with different word embedding models to find and
rank properties related to the user input. In a nutshell, using models trained
4
http://www.ontodia.org
5
http://www.metaphacts.com
Word Embeddings for Search in Linked Data 3
on a Wikipedia corpus, a representation of every Wikidata property is created
by the vectorial sum of the words in its label (and description text). Then we
compute the similarity of the user input (represented by the vectorial sum of its
words) with the entity properties, and rank the properties by similarity to the
user input. The goal is to make the NL interface more powerful and improve user
experience and ease-of-use. Figure 1 shows the result of searching for “family”
relations for entity Van Gogh.
4 Evaluation
In this section we give an overview of the evaluation results for various word
embedding models for the task of property suggestion for user input terms.
4.1 Evaluation Setup
Wikidata includes 3323 different properties and 4603 property aliases. Although
the data quality of the aliases is sometimes questionable, the aliases are a suffi-
ciently large gold standard dataset to evaluate our property suggestion modules.
Like for user input, for any of the aliases, first we extract the contained words,
and remove stopwords. Then the tool creates the vectorial sum of the alias words
using the pre-trained word embedding models (see below), and ranks all 3323
Wikidata properties by cosine similarity with the alias vector. We apply various
evaluation metrics. Due to space limitations, here we include only the ratio of
correct property suggestions for aliases in the top-N of the property ranking,
and the mean reciprocal rank (MRR).
Also, we conduct experiments with many types, settings and different training
corpora for the word embeddings. The best results are encountered with fast-
Text [2] embeddings trained on a Wikipedia corpus, and LexVec [5] embeddings
trained on Wikipedia and a news corpus, both with vectors of 300 dimensions.
4.2 Results
In Table 1 we present the evaluation results for the fastText and LexVec embed-
dings. We distinguish models which use only the words from the property label
to create the property vector (not using description) and property representa-
tions that build the vectors by making vectorial sums which include both the
words from the property labels and the property descriptions.
Using the property description text to create the property representations
improves service quality. Over all metrics, the fastText models perform best. This
may result from fastText using word morphology together with the word2vec
skipgram model, which is well suited for the task at hand.
In the real world application in visual data exploration in Ontodia the task
is much easier as compared to the evaluation in Table 1. When we randomly
pick entities from Wikidata – we evaluated with 1154 random entities, these
have on average 29.5 properties per entity. When evaluating randomly picked
4 Authors Suppressed Due to Excessive Length
Top 1 Top 3 Top 10 MRR
fastText using description 38.12 % 55.13% 70.22% 0.493
LexVec using description 37.21% 53.45% 67.98% 0.464
fastText not using description 36.10% 51.94% 65.25% 0.478
LexVec not using description 34.74% 50.03% 62.91% 0.447
Table 1. Ratio of test terms (aliases) with the correct property suggestion in the top-N
of the results, depending on word embedding method used and settings; and MRR.
entities using the same gold standard data, 70.5% of first ranked properties are
the correct ones, the Top-3 score is 85%, and the MRR is 0.80. The runtime of
queries is below 0.01 seconds, fast enough for interactive systems.
5 Conclusions
In this demonstration paper we present and evaluate the improved query func-
tionalities of Ontodia, which apply vector similarity with word embeddings be-
tween query terms and entity properties. Extensive evaluations show the method
is sufficiently accurate for integration in Linked Data visualization tools. In fu-
ture work we apply the method to other datasets, and the search for entities.
6 Acknowledgments
This work was supported by the Government of the Russian Federation (Grant
074-U01) through the ITMO Fellowship and Professorship Program.
References
1. Augenstein, I., Gentile, A.L., Norton, B., Zhang, Z., Ciravegna, F.: Mapping key-
words to linked data resources for automatic query expansion. In: Cimiano, P.e.a.
(ed.) The Semantic Web: ESWC 2013 Satellite Events, Montpellier, France. pp.
101–112. Springer LNCS, Berlin, Heidelberg (2013)
2. Bojanowski, P., Grave, E., Joulin, A., Mikolov, T.: Enriching word vectors with
subword information. arXiv preprint arXiv:1607.04606 (2016)
3. Ghannay, S., Favre, B., Estve, Y., Camelin, N.: Word embedding evaluation and
combination. In: et al., N.C. (ed.) Proc. of the Tenth International Conference on
Language Resources and Evaluation (LREC 2016). ELRA, Paris, France (May 2016)
4. Mouromtsev, D., Pavlov, D., Emelyanov, Y., Morozov, A., Razdyakonov, D., Galkin,
M.: The simple web-based tool for visualization and sharing of semantic data and
ontologies. In: International Semantic Web Conference (Posters & Demos) (2015)
5. Salle, A., Idiart, M., Villavicencio, A.: Enhancing the lexvec distributed word
representation model using positional contexts and external memory. CoRR
abs/1606.01283 (2016), http://arxiv.org/abs/1606.01283
6. Shekarpour, S., Lukovnikov, D., Kumar, A.J., Endris, K.M., Singh, K., Thakkar,
H., Lange, C.: Question answering on linked data: Challenges and future directions.
CoRR abs/1601.03541 (2016), http://arxiv.org/abs/1601.03541