The cross-lingual matching of authors and publications is a special case of the task of assigning a unique identifier to the same real-world entity in multilingual data sources. This paper presents the results of experiments with the several versions of a cross-lingual system designed to match, basing on a Russian-language data source, the authors and English-language publications. Since different heuristics have been tested in these versions of the system, we consider here only those that have given the best results. An important element of the system is its interactive visualization tool, which gives information on the distribution of publications by authors, as well as providing the ability to edit the results of the analysis. The visualization system is supplemented with methods for similarity matrices ordering. Experiments have shown that the main source of improving the quality of the matching and clustering algorithm is extending the set of confirmed publications. The approaches used in this system are applicable to solving the problem of linking named entities in various multilingual data sources.
Nowadays, entity resolution is being intensively investigated in the context of the integration of heterogeneous data sets. Collecting data from heterogeneous data sets and integrating them in a query able environment increases completeness and correctness as well as ensures a more effective analysis. Of special interest is the problem of cross-lingual entity resolution for multilingual data integration since local language data sources are often more complete and accurate than global data sources.
Although English is the main language for research and the Internet, a great
number of research publications belong to non-English authors and are translated from
various foreign languages, which make the task of integrating multiple data sources
even more difficult. Naturally, this poses the problem of the cross-language
disambiguation of named entities and, in particular, the cross-language matching of authors
and publications [
Our previous research has demonstrated that Russian names allowing several
transliterations represent a challenge for international data bases and knowledge graphs.
Experiments with several multilingual datasets have shown that Russian names
admitting several transliterations are often treated as homonyms, and several different
persons with identical name variations are treated as synonyms [
For example, the papers by Виктор Карлович Сабельфельд are assigned in the
Scopus data base [
Thanks to the projects such as ORCID [
Since similar situations are quite common, our main objective is to investigate various methods to enrich and improve an English-language data source(s) by comparing its (their) content with national data sources in local languages, such as the Russianlanguage data sources.
An algorithm for the cross-language identity resolution using the SBRAS Open
Archive is presented in [
To this end, a way of establishing correspondence between the Russian-named and
English-named entities has to be developed. The transliteration-based matching of
personal names was already described in [
The paper is organized as follows: first, we outline the related works and then present the essential datasets and metadata. After that, the matching and clustering algorithm and implementation details are described. Finally, we demonstrate an interactive visualization, which facilitates the comprehension of the matching and clustering results and allows users to improve them. 2
Entity resolution – identity resolution, deduplication, record linkage, object matching – refers to the task of identifying different representations of the same real-world object. In the context of relational data bases, the term entity reconciliation has been used for a long time. Another related term is link discovery, which describes the process of automatic search for the overlapping parts of heterogeneous data sets and linking individual records of these data sets by exploiting their specific properties. When talking about entities from multiple sources, it is common to use the term entity clustering. There are several groups of works closely related to the topic of our paper.
Entity resolution is an important step in any data integration pipeline as well as in
the area of information retrieval and question answering. Nowadays, data from any
domain are available on the Web and a lot of research has been dedicated to this
problem. A recent overview is presented in [
To avoid comparing new entities with all members of existing clusters, each cluster
creates a cluster representative, which is fused from all the properties of the cluster
members. The same idea is implemented as the DBpedia Global Id Management
module [
Therefore, much more numerous are the studies of cross-lingual KG alignment
aiming to match automatically entities in different languages from different data
sources. Recently, several embedding-based approaches have been proposed for
cross-lingual KG alignment, including MTransE [
More recent papers model the equivalent relations between entities by using the
graph convolutional networks (GCN), since they are able to generate the
neighborhood-aware embeddings of entities used to discover entity alignments [
Unlike the previous methods using entity embeddings to match entities, the paper
[
The SpringerLink digital library [
Other important reasons for choosing this data source are open access to its data and the emergence of the related SN SciGraph Linked Open Data platform, which also demonstrates the problems mentioned above.
The eLIBRARY.ru data base is the largest bibliographic data set used for computing the scientific rating of Russian researchers. eLIBRARY.ru stores data in the fields of science, technology, medicine and education on more than thirty million publications, more than 900,000 researchers and over 13,000 organizations (including over 3,000 officially registered). The A.P. Ershov Institute of Informatics Systems of the Siberian Branch of the Russian Academy of Sciences is an organization registered at eLIBRARY.ru; it regularly inputs and updates information concerning its employees’ publications. Moreover, you can find here a complete list of publications by Academician Andrei Petrovich Ershov created by Anna Andreyevna Bulyonkova.
The sets of metadata provided by eLIBRARY.ru are similar to those of SpringerLink, though access to these metadata is restricted. To be more specific, the list of publications of an author is freely available, but detailed metadata on his/her papers are not free. Therefore, our disambiguation algorithm is based on the data freely available at eLIBRARY.ru. Another essential difference between these two data sources is that the language of SpringerLink is English, and that of eLIBRARY.ru is Russian, even when it stores data on the English publications of Russian researchers. The main problem, hence, is how to match entities described in different languages. 4
The problem to solve is formulated as follows. Given an English language bibliographic data source(s), extract all publications potentially belonging to the person(s) specified by a Russian-language keyword and divide the extracted articles into subsets S1, S2, ..., Sn so that each subset of articles belongs to one real person.
The steps of the matching and clustering algorithm are: 1. Given a full Russian name, a set of extended transliterations is generated. 2. Elements of this set are used for the keyword search of publications in the Englishlanguage data source (for example, SpringerLink digital library). 3. An extended set of potential homonyms of the person, specified by the full Russian name, is used to extract groups of publications from a Russia-language data source (for example, eLIBRARY.ru). 4. All the publications extracted from SpringerLink are matched against the eLI
BRARY.ru groups of publications. 5. The papers unmatched at the previous step are further analyzed and clustered. 6. Interactive visualization of the clustering results makes it possible to analyze and to further refine it.
The extended transliteration and data extraction from the English-language source
SpringerLink are described in detail in [
The authors of the publications extracted from SpringerLink can be either homonyms or synonyms. The entity identification algorithm should process the list of publications and determine which of their authors are synonyms and which are homonyms. In other words, the list of publications should be clustered into the subsets S1, S2,…, Sn such that each subset of articles is authored by a single person and all his or her name variations are synonyms. The subset S1 should contain the articles authored by the person under consideration.
To this end, the list of publications S extracted from SpringerLink is matched against the lists of publications E extracted from eLIBRARY.ru. Note that the papers of eLIBRARY.ru are already clustered into the groups E1, E2,…, Em corresponding to individual authors. Therefore, if a paper si S is recognized as identical to a paper ej belonging to a group Em from eLIBRARY.ru, it is assigned to a group Sm.
eLIBRARY.ru specifies persons by their full Russian name in <LastName First Name Middle Name> format, affiliation and location of the employing organization.
Note that each person can have several homonyms and “partial” homonyms, when a short form of his/her name coincides with the short form of another person’s name. For example, five full homonyms, having the same name Петров Евгений Сергеевич, are described in eLIBRARY.ru along with two partial homonyms having two distinct middle names. The persons having identical full or short forms of their names can be erroneously identified as synonyms.
To prevent this kind of errors, our algorithm creates groups of confirmed eLIBRARY.ru papers for each potential homonym of a given author. Since we consider eLIBRARY.ru to be a reliable source of information about publications written by Russian authors, we use it to create confirmed groups of publications for an English data source. That is why the confirmed groups of papers are created by comparing the papers from SpringerLink and eLIBRARY.ru.
When comparing publications from the two data sources, two main possibilities are considered.
1. Publications in both data sources are described in English. The situation when a publication in eLIBRARY.ru contains a description in Russian and an Englishlanguage version of the publication is described in the "Versions" field of the publication also falls into this category. 2. The English data source contains an English description of the publication, and the Russian data source only contains a description of the Russian version.
In the first case, when the descriptions of both articles are given in English, the names of the publications and the lists of authors are compared. A paper si S is considered to be identical to a paper ej E if Title(si) = Title (ei) AND Authors(si) = Authors(ei). The title cannot identify a paper uniquely as some authors can have several publications with the same title. Nevertheless, the exact match of titles and author names can be considered as evidence that the papers were authored by the same person. However, some paper titles differ in SpringerLink and eLIBRARY.ru due to scanning errors. For example, the paper titled as SCHEMATOLOGY IN A MULTILANGUAGE OPTIMIZER in eLIBRARY.ru appears as Schematology in a MJ I/T Ilanguage OPT imizer in SpringerLink. In the absence of an exact match of the paper titles, both titles are stemmed by the Porter stemmer and their overlap score is calculated. If this score exceeds a threshold value, the titles are considered coinciding. The discovered matching is written in a special file for further user control.
In the second case, cross-language identification of paper and journal titles should be applied. Many Russian journals are first published in Russian and then translated in English. A typical example is the Программирование journal, which is published in English as Programming and Computer Software. About 40% of eLIBRARY.ru older entries have only Russian description and do not have an English counterpart. These publications, however, are very important for making confirmed paper groups as large as possible. This situation raises several problems. First, it is impossible to compare papers by the title when the title of an original paper is in Russian and the title of a translated paper is in English. Besides, the original and translated papers have disjoint sets of attributes, such as venue, ISSN, publication data, page numbers, etc.
Although SpringerLink provides information about the journal titles in the Latin alphabet only, every translated paper in the database mentions its Russian original. For example, the paper by V.E. Kotov Parallel programming with types of control has the label “Translated from Kibernetika, No. 3, pp. 1–13, May-June, 1979” in SpringerLink. Moreover, the SpringerLink database provides the ISSN of the translated version. This information suffices to find the Russian version of the paper if it is available in eLIBRARY.ru. The corresponding English-language article is marked as matched and the pair of papers is saved for further processing.
The average number of papers assigned to the confirmed groups during the matching step was about 69%, while the number of erroneously attributed publications was close to zero. The main reason why the system cannot assign some papers to their authors is data sparsity. To extend the set of the identified authors of papers, a clustering algorithm was applied to the unmatched papers. 5
The publications unmatched at the previous step are considered to be unconfirmed and should be further analyzed and clustered. The clustering algorithm for unconfirmed publications is based on comparing each unconfirmed publication with groups of confirmed publications based on various similarity metrics.
The unconfirmed publications are compared with groups of confirmed publications using the following attributes: titles of publications, lists of authors of publications, topics and keywords, dates of publication, venue of publication (journal title or conference title), and similarity of texts of publications (TF-IDF).
All the attributes are compared pair wise, which results in a number of scores that are summarized in the final step. When calculating the similarity scores the program adheres to the following rules.
Titles of papers similarity If an exact match of the paper titles A and B is found, the title_similarity_score is set to 1.0. Otherwise, the titles of the papers A and B are stemmed, and the title_similarity_score is set to the overlap ratio of their word lists.
Co-authors similarity For each confirmed group of publications, a list of all its co-authors is created, and the more often an author appears in the list of co-authors for a certain group of publications, the greater is his weight. Also, for each confirmed group of publications, a list of the author’s affiliations is created, which characterizes each group, and when comparing the authors of publications, not only the names of the authors, but also their affiliations are compared.
Subjects and keywords similarity The subject_similarity_score and keyword_similarity_score use Jaccard Index to evaluate the overlap ratio of the respective lists.
Date similarity The date_similarity_score is set to 0.1 if the timestamp difference of the papers A and B is less than five years. If the timestamps difference of the papers A and B is more than twenty five years, it is set to - 0.1.
Venue similarity The publication_venue_score (i.e., conference/journal title) is set to 0.1 if there is an exact match between their titles.
Text similarity Text_similarity_score is evaluated by TF_IDF and cosin similarity measure.
The final assignment likelihood is calculated as the sum of all the above scores.
All unconfirmed publications initially obtain a group number of -1. An unconfirmed publication joins a group of confirmed publications if its similarity with a given group exceeds a threshold value. Two groups of unconfirmed publications can be merged if the similarity value of publications inside two groups exceeds the threshold value. When merging two groups, the algorithm monitors that both groups do not belong to the set of the confirmed groups. If this happens, the merging does not occur, since the confirmed groups correspond to the publications by distinct authors.The weights of all the attributes involved in the comparison, as well as the threshold value, can be adjusted at the stage of interactive visualization. 6
To simplify the understanding and modification of the matching and clustering
results, several interlinked visualizations have been developed; they are described in
detail in [
Experiments have shown that to verify the matching and clustering results users
need to compare papers clustered in distinct groups. To this end they need to see all
the attributes of the entire set of publications in a single view. Therefore, a set of
visualizations was extended by a similarity matrix reordering module based on the
reorder.js library [
Two groups of experiments have been carried out. First, the available matrix ordering algorithms were applied to the similarity matrices, which contained information about the similarity of various publications in the English source, but did not contain information about the results of the comparison of the English-language and Russianlanguage data sources. In this case, all ordering algorithms showed different results.
Then the similarity matrices have been modified by adding a large similarity score to the publications that were classified as identical at the stage of the comparison of the English-language and Russian-language data sources. After that, all the matrix ordering algorithms have started to output the equivalent results. 7
The newly developed matching procedure provides the algorithm presented in this paper with the ability not only to cluster the papers correctly, but also to determine the exact identity of authors, including the name and location of the affiliating organization.
The program implementing the algorithm has been tested on a dataset of 100 persons employed by the IIS SB RAS at various time periods. Also, this dataset contains Academician A.P. Ershov, whose papers have been input into eLIBRARY.ru by the IIS SB RAS. The total number of papers found in SpringerLink for all Russian names in this dataset was 3,175. All the results obtained by the program were verified manually. For each person listed in the test dataset the following values were calculated: total number of papers found in SpringerLink for each Russian full name listed in the test dataset; number of articles actually authored by a researcher specified in the test dataset; number of papers that have been correctly recognized by the matching algorithm; number of papers that have been correctly recognized by the matching + clustering algorithm;
Experiments have shown that the main source of improving the quality of crosslingual entity resolution algorithm is extending the set of confirmed publications by matching publications and authors in Russian and English data sources. The combination of the matching algorithm and the clustering algorithm allows us to correctly recognize from 92 to 93 percent of publications.
This algorithm can be used to match any multilingual knowledge bases. The further development of these studies is supposed to focus on implementing new algorithms for the cross-language identification of entities and on implementing a fullfledged framework including tools for accessing various data sources, methods for establishing correspondence between different schemes of data sources, etc.
The author thanks Avramenko M.Yu., Paramoshin A.A., Isachenko V.V., and Eliseev E.S. for participating in the implementation of the various versions of the matching and clustering algorithm.