In the era of daily information overload, personalized retrieval applications represent a crucial solution to provide suggestions to users. The research and industrial community have devoted an unprecedented efort to propose approaches and architectures to extract relevant and tailored information from every shred of knowledge. Inspired by the advances in knowledge graph, Graph Convolutional Networks, Link Prediction, and Recommender Systems research, this study aims to meet their cutting-edge research needs of holistic datasets by largely expanding the information available for two well-known recommendation datasets in the book and movie domains, i.e., LibraryThing and MovieLens 25M. We collect the associated knowledge graphs (KGs) for each of them and publish a mapping for each item in the two datasets with the corresponding entities in the original Wikidata, DBpedia, and Freebase KGs. Our work is available at https://github.com/sisinflab/Augmented-and-Linked-Open-Datasets-for-Recommendation.
and state-of-the-art. Knowledge graphs (KGs) have the advantage of covering a variety of heterogeneous Recommender Systems (RSs) constitute the backbone domains, thus making it easier to transpose the advances of online platforms, pursuing the purpose of suggest- in one domain to another. The adoption of knowledge ing to customers novel products or services that might graphs as a source of structured information has genermeet their interests and needs. The pivotal role of rec- ated several advancements in the tasks strictly related to ommender systems [1] has been advantaged by the no- recommendation [4]: knowledge completion [5], prefertable accuracy provided by collaborative filtering meth- ence elicitation [6], and user modeling [7, 8]. Given the ods [2]. Collaborative filtering (CF) is the most tradi- graph nature of data, several diferent recommendation tional and still prominent recommendation approach. CF approaches have been proposed to tackle: knowledge models calculate recommendations leveraging similari- transfer across domains [9], interpretable/explainableties in interaction/preference patterns of similar users. recommendation [10, 11, 12], user modeling [7, 13, 14], However, recommender systems generally have to face and graph-based recommendation [15, 16, 17, 18, 19]. Aldata sparsity issues due to the availability of little infor- though the KGs are all diferent, every method needs two mation about users’ preferences, which often causes the side information: a mapping that establishes a connecso-called cold-start problem. To compensate for this lack tion between items within the catalog and entities in the of information, an increasing number of approaches com- KGs, and a set of RDF triples that utilize this mapping to bine collaborative information with auxiliary content describe the relevant domain. attributes, such as tags, metadata, and geographical data. Unfortunately, dealing with these information sources Content-based Filtering (CBF) models assemble a pro- raises two important issues for researchers. First, it is ifle of user interests based on the description of the items usually a time and computationally expensive task due to that the user has previously consumed. Even though one the limits and constraints of KG services that make this could leverage this information using traditional CBF operation frustrating. For instance, the public DBpedia approaches like Vector Space Models and multimedia sig- Virtuoso SPARQL endpoint sets limits both on the rate nals [3], knowledge graphs gave birth to specific research of requests (whose violation could lead to a permanent directions that comprise their own literature, models, ban) and on the maximum SPARQL query solution size1. Second, the information available on the Web rapidly and constantly changes over time. This causes the diferent retrieval operations of item attributes to be inconsistent and lead to diferent outcomes in recommendation research, thus dangerously jeopardizing the reproducibility of experiments [20, 21].
This study aims to tackle these issues by releasing an 1https://www.dbpedia.org/blog/new-dbpedia-usage-report-2021/ extensive collection of side information exploited in tasks and exploit implicit entity preferences to infer the plauinvolving KGs. We focus on two well-known recommen- sibility of KG facts. Wang et al. [33] explore the intents dation datasets in the community and detail the steps to behind a user-item interaction by using auxiliary informagather the collection. In detail, the contribution of the tion in Amazon Books, LastFM and Alibaba-iFashion [34]. work at hand is manifold: Sun et al. [35] present an approach for KG embedding able to model and infer various relation patterns. They • a mapping between the items in the LibraryThing evaluate the proposed model on four widely used KGs, (LT) and MovieLens 25M (ML25M) catalogs and the i.e., FB15k [36], WN18 [36], FB15k-237, and WN18RR. entities available in three well-known knowl- Zhang et al. [37] use MovieLens 1M and IntentBooks [38] edge graphs: Wikidata [22], DBpedia [23], and Free- to evaluate CKE, a knowledge-aware model extracting base [24]; feature representations from a knowledge base and cap• the retrieval of DBpedia and Wikidata KG’s tures relationships between users and items. This large triples up to two hops to collect structured informa- amount of work using side information and enriched tion connected to these resources, thus providing per- datasets shows the potential of these tools. However, sistent and ready-to-use enriched datasets for perform- often these works use datasets whose linking strategies ing reproducible experiments; are highly diferent from each other and lack details for their reproducibility, putting reuse within other works • a methodology for datasets’ items linking as an at risk.
approach to promote the emergence of other state-of- For instance, concerning MovieLens and LibraryThing, the-art enriched datasets; which we treat in this paper, few significant attempts exist to link these datasets to KGs. Wang et al. [39] link • an evaluation of the linking methodology by com- MovieLens 1M with IMDb information, such as genres, paring it with a state-of-the-art entity linker showing actors, directors, and writers to use it in a knowledgethat traditional linking approaches are inefective for aware model. Ostuni et al. [40] publish a mapping of this task. MovieLens 1M to the DBpedia KG, while Zhao et al. [41] The remainder of this paper is organized as follows. propose KB4Rec, a public linking of MovieLens 20M, Section 2 surveys the existing dataset and the knowledge- LastFM 1B, and Amazon Book to the popular KB Freebase. aware recommendation works that use them. In Sec- Moreover, to the best of our knowledge, only Di Noia tions 3 and 4, we describe the outcome of our work, et al. [42] in 2016 succeded linking LibraryThing to a along with details about challenges we have faced and KG, specifically to DBpedia, likely due to some domain methodologies we have adopted. In Section 5, we eval- limitations detailed in Section 4. uate the adopted text similarity-based solution for our These eforts are often not complemented by details work. Finally, Section 6 concludes this paper, proposing about the adopted linking methodologies and do not proand discussing future directions. vide an evaluation of their outcome. In addition, previous literature provides linked datasets but does not publicly release the RDF triples retrievable by exploring the KGs. 2. Knowledge Graphs in This requires the researchers to scan the KGs each time, Recommendation Works thus causing high bandwidth, resource, and time consumption. It is worth highlighting that entities and triples According to Sun et al. [25], the ten most used datasets in of the KGs change over time, thus producing a mismatch of the recent literature (sorted in descending order of adop- the retrieved triples that causes reproducibility issues. tion) are ML-1M [26], Netflix, LastFM, Yelp, ML-20M, These drawbacks have animated us to release extensive Epinions [27], ML-10M, ML-100K, Book-Crossing [28], linkings of the large MovieLens 25M and LibraryThing Amazon Music. Unsurprisingly, they relate with stream- datasets to DBpedia, Wikidata, and Freebase and describe ing and e-commerce platforms and cover the typical do- our methodology in detail. We provide researchers with mains of movies, music, books, and shops. When linked a vast, up-to-date, and consistent-over-time collection data is available, it can be leveraged and adopted for ad- of RDF triples. The proposed methodology permits the dressing several tasks. For instance, Piao and Breslin [29] reproducibility of the results and helps researchers in link MovieLens 1M and DBook datasets to the DBpedia extracting novel datasets for other domains. Finally, we KG and investigate the idea of transfer learning between propose an evaluation of our linking strategy, thus enitem recommendations and KG completion tasks leverag- hancing the trust of the research community in using the ing a co-factorization model. Also Vagliano et al. [30] and brand-new enriched datasets.
Alshammari and Nasraoui [31] use knowledge-aware versions of MovieLens for powering their models. He et al. [5] adopt in their model the public linking KB4Rec [32]
One of the main contributions of this work is providing the community with two enriched versions of MovieLens 25M (ML25M) and LibraryThing (LT), thus guaranteeing the reproducibility of any experiment making use of additional knowledge. Table 1 reports the the overall statistics for the original datasets, while Table 2 summarizes the collected resources for both datasets. Each resource is extensively detailed in the GitHub repository.
The MovieLens datasets are the result of a collective efort of thousands of people expressing explicit preferences about movies [26]. These datasets constantly increase in size regarding ratings, users, and items, with the release that we hereby consider including about 25 million ratings. We have complemented the data, DBpedia, and Freebase knowledge graphs. With this information available, we have explored the Wikidata and DBpedia graphs up to the second hop [43, 44] by querying the public SPARQL endpoints to retrieve the subgraphs originating from the entities associated with the dataset items. ing the item ID and the URIs referring to the same entity in the three knowledge bases. On the other hand, the facts retrieved from the knowledge graphs have been ( −
→− work (RDF). RDF provides a simple data model to encode knowledge in a structured way through a set of triples in the form ⟨, ,
⟩. Each triple represents the connection between the subject and the object through a predicate . We define
exploration at 1-hop of as the operation of retrieving the triples having as the subject. Generalizing, the exploration at -hop is the operation to retrieve the triples having as subjects all the objects obtained at the exploration from 1-hop to 1)-hops. In the proposed resources, we collected the triples obtained exploring DBpedia and Wikidata at
LibraryThing. The LT dataset has been collected in 2011 from the namesake website, which allows users to items in MovieLens 25M with links to resources of Wiki- In this section, we present methodologies, challenges, create a virtual catalog of the books they own or have read. The dataset contains item reviews with ratings from
received. The linking of LibraryThing items to DBpedia and Wikidata URIs followed a similarity-based approach, due to some information mismatching that we discuss in Section 4 and that made dificult to obtain a definite mapping. As a consequence, the item-URI linking comes with a similarity, that accounts for a score that a book actually corresponds to the associated resource on Wikidata and DBpedia. Hence, one can set a confidence threshold, bearing in mind that the lower is the threshold, the higher is the number of linked entities, but the higher is the occurrence of false positives. Although a threshold in the range of [0.5, 0.6] may represent an acceptable tradeof between true positives and false positive, we have not fil tered the retrieved information, leaving each researcher the choice to select it. Regardless of the linking similarity scores, we have linked the items to Freebase entities. Additionally, we have provided all the triples retrievable at July 2023 from DBpedia and Wikidata by querying the SPARQL endpoints using the linked entities and exploring the graphs at the first and at the second hop.
and solutions adopted for performing the item linkings described in Section 3. Our goal is to sketch best practices and highlight issues that raise from item linking, with the aim of encouraging the discussion about the need of a common workflow to collect augumented datasets.
try to fill the gap between the items in the catalog and
item description linking, as suggested by Di Noia [45].
a map from internal item IDs to the identifiers of the movies in the massive Internet Movie Database (IMDb) and in The Movie Database (TMDb). We have exploited this connection and adopted a direct item linking strategy, querying the Wikidata endpoint with a SPARQL query.
the first query uses the IMDb identifier to find the corresponding entity on Wikidata; (ii) if the first attempt does not succeed, another query is sent to the endpoint using the TMDb identifier. Once the resource has been found in Wikidata, we reconstruct the corresponding
On the one hand, item linkings are collected compris- catalog with the corresponding resources in the Linked provided in the standard Resource Description Frame- solutions to address this goal, i.e., direct item linking and DBpedia URI from the English Wikipedia page URL. The instance ∈ with its title ∈ and its author linking approach chosen for MovieLens 25M does not ∈ , we aim to create a direct mapping → , generate false positives thanks to the accurate references where ∈ is characterized by a title ∈ from MovieLens 25M items to IMDBb and TMDb IDs, and an author ∈ , based upon the similarities and to the explicit reference to those IDs in the Wiki- between titles labels and and authors labels and data Knowledge Graph. However, for each item in the . dataset, at least one SPARQL query had to be performed. To this end, we consider and combine three popular Therefore, the researcher should take into account the costs similarity metrics, namely the Jaro [46], Ratio [47], and for establishing a huge number of HTTP requests and the -grams [48] similarities. We assume string similarity limitations that an endpoint may place. could be a reasonable solution to ensure that two entities are the same. Furthermore, by calculating and LibraryThing. LibraryThing collects books using in- combining the similarities between title names and auternal identifiers without sharing the international iden- thors, we may address critical situations where titles tifier or a website page. Moreover, books are usually sold are the same, but authors do not match, and vice versa. in diferent editions and LibraryThing manages this in- Then, by considering two strings 1 and 2, we define formation treating them as completely diferent books. _(1, 2) as the average of the these similarity Although book entities in Wikidata and DBpedia are gen- measures. Concerning the and sets, and takerally linked to a page of the LibraryThing website, a ing into account an instance ∈ and an instance relevant issue comes from the fact that the LibraryThing ∈ , we perform _(, ) (i.e., the similardataset and the Wikipedia/DBpedia KGs often link to ity between title labels) and _(, ) (i.e., the diferent editions of the same book — thus generating a similarity between author labels). Subsequently, we mulmismatching of the identifiers. To overcome this issue, tiply them to obtain the combined similarity measure we adopt the item description linking strategy. With this (, ) = _(, ) * _(, ). We set approach, we exploit a similarity function to compare a _(, ) = 1 if the labels of the author and/or resource’s title and author available in the dataset with are missing, so that the title labels are the only ones the corresponding information of the books from Wiki- to contribute to the similarity since 1 is the neutral factor data or DBpedia. Since the similarity is computed with for the multiplication. We perform this operation for a dedicated ofline operation, this approach requires to each instance ∈ , obtaining a number of triples query only once Wikidata and DBpedia to retrieve all the ≜ (, , (, )) equals to | | × | |. Finally, we books in these KGs along with their titles and authors. select at most one triple with max((, )) for each The linking task is then realized with a post-hoc anal- combination of and . In this respect, we obtain a direct ysis. To summarize, we created the collection of the and unambiguous mapping between the book ID of and books’ URI instances available on Wikidata and DBpedia the URI in . However, the linking may not be unique at with their titles () and authors (). Then, we created this stage because some diferent URIs may be associated the collection of books IDs from LibraryThing with with the same ID. In such cases, we make the linking their titles () and authors (). Finally, given an unique by only considering the one characterized by the
From DBpedia and Wikidata to Freebase. Both the DBpedia and Wikidata KG provide between their entities and Freebase KG. Therefore, we get from the both KGs all the object entities referenced by these predicates to accomplish this linking stage. As for MovieLens 25M, a Freebase reference was searched by exploring DBpedia, while for the entities not linked to Freebase yet, a reference was searched by querying Wikidata. Concerning LibraryThing, the solution adopted to perform the item linking led us to generate two diferent comprehensive resources, one for each KG. uation). Finally, we link a URI resource of DBpedia to the iflm having the highest similarity score and compare this linking with the ground truth. It is worth mentioning that side information could be diverse even though it refers to the same subject. For clarity, one can easily understand that "John Lasseter" and "J. Lasseter" are correct information about Toy Story’s director. However, even though 5. Evaluation this situation is common, an automated agent finds it more challenging to identify the same entity. Hence, we This section describes the limitations of existing named repeated the experiment to evaluate the robustness of entity linkers [49, 50, 51] for this linking task and the the linkers by introducing a random noise on the strings. need for an ad-hoc linker for the book domain. Then, an Specifically, each character is deleted e/o perturbed with experimental evaluation is conducted to assess if the pro- a probability of 0.15. posed solution can overcome the performance of existing The results in Table 3 show the eficacy of the prolinkers. In detail, diferently from a traditional linking posed solution. With the support of the original side task where a textual context is provided, only titles and information only, our linker successfully links 40 189 of authors are retrieved from LibraryThing. Although this the items, an acceptable performance compared to DBpecould already be a reasonable motivation, we compare dia Spotlight (15 483). For what concerns the robustness the solution with a state-of-the-art linker, DBpedia Spot- analysis, DBpedia Spotlight performance collapses up light, a framework for automatic annotation of DBpedia to 1755 of correct links, while the proposed linker loses entities from natural language text [50, 52]. only 4550 links. In conclusion, the results justify the
However, a ground truth is needed to evaluate the adoption of a tailored linker for this peculiar application methods. Unfortunately, LibraryThing does not provide scenario. this information, but MovieLens does. In fact, the MovieLens dataset contains titles, IMDb, and TMDb IDs. Therefore, we can exploit it to evaluate the performance of the 6. Conclusion and Future Work linkers by fixing the available side information, i.e., the label of the film and its director taken from IMDb. This work provides supplementary data of two well
On the one hand, to get DBpedia Spotlight working known and widely used recommendation datasets (i.e., at its best, we supply this framework with a context by MovieLens 25M and LibraryThing) that can be utilized generating sentences exploiting the available side infor- in tasks involving Knowledge Graphs. Inspired by the mation. For instance, given as metadata the film title " Toy new information needs of cutting-edge research, we proStory" and its director "John Lasseter", we ask DBpedia vide mappings to three well-known Knowledge Graphs: spotlight to annotate DBpedia entities in the sentence DBpedia, Wikidata, and Freebase. To ensure reproducibil"The film Toy Story is directed by John Lasseter ". Then, DB- ity of the future experiments, we also provide domain pedia Spotlight should be capable of taking out, among Knowledge Graphs at the first and second hop of explothe other recognized entities, the resource URI in the ration in the form of RDF triples, thus enabling reliable DBpedia Knowledge Graph (KG) of the film " Toy Story", Graph-based and Knowledge-aware recommendation rewhich should correspond to the one found with direct search. Furthermore, they are aligned and expose the item-linking. On the other hand, we follow the pipeline same additional information to ease the researcher’s efproposed for the book domain (see Section 4), retrieving fort in conducting evaluation and comparisons. For fuall the films in the DBpedia KG with their labels and direc- ture work, we encourage the enrichment of datasets from tors (corresponding to the book titles and their authors other domains such as music and shops and their intein the book domain) and computing string similarities gration with spectrograms and images, respectively, to with IMDb-derived data (the candidate items in this eval- design novel multi-modal recommendation methods. neering and Knowledge Management, IC3K 2019, and Data Mining, San Francisco, CA, USA, August Volume 1: KDIR, Vienna, Austria, September 17- 13-17, 2016, ACM, 2016, pp. 353–362. URL: https: 19, 2019, ScitePress, 2019, pp. 83–88. URL: https: //doi.org/10.1145/2939672.2939673. doi:10.1145/ //doi.org/10.5220/0008070900830088. doi:10.5220/ 2939672.2939673.
0008070900830088. [38] A. Uyar, F. M. Aliyu, Evaluating search fea[32] W. X. Zhao, G. He, K. Yang, H. Dou, J. Huang, tures of google knowledge graph and bing satori: S. Ouyang, J. Wen, Kb4rec: A data set for Entity types, list searches and query interfaces, linking knowledge bases with recommender Online Inf. Rev. 39 (2015) 197–213. URL: https: systems, Data Intell. 1 (2019) 121–136. //doi.org/10.1108/OIR-10-2014-0257. doi:10.1108/ URL: https://doi.org/10.1162/dint_a_00008. OIR-10-2014-0257.
doi:10.1162/dint\_a\_00008. [39] X. Wang, D. Wang, C. Xu, X. He, Y. Cao, T. Chua, [33] X. Wang, T. Huang, D. Wang, Y. Yuan, Z. Liu, Explainable reasoning over knowledge graphs for X. He, T. Chua, Learning intents behind inter- recommendation, in: The Thirty-Third AAAI actions with knowledge graph for recommenda- Conference on Artificial Intelligence, AAAI 2019, tion, in: J. Leskovec, M. Grobelnik, M. Najork, The Thirty-First Innovative Applications of ArJ. Tang, L. Zia (Eds.), WWW ’21: The Web Con- tificial Intelligence Conference, IAAI 2019, The ference 2021, Virtual Event / Ljubljana, Slovenia, Ninth AAAI Symposium on Educational Advances April 19-23, 2021, ACM / IW3C2, 2021, pp. 878– in Artificial Intelligence, EAAI 2019, Honolulu, 887. URL: https://doi.org/10.1145/3442381.3450133. Hawaii, USA, January 27 - February 1, 2019, AAAI doi:10.1145/3442381.3450133. Press, 2019, pp. 5329–5336. URL: https://doi.org/ [34] W. Chen, P. Huang, J. Xu, X. Guo, C. Guo, F. Sun, 10.1609/aaai.v33i01.33015329. doi:10.1609/aaai.
C. Li, A. Pfadler, H. Zhao, B. Zhao, POG: person- v33i01.33015329. alized outfit generation for fashion recommenda- [40] V. C. Ostuni, T. D. Noia, E. D. Sciascio, R. Mition at alibaba ifashion, in: A. Teredesai, V. Ku- rizzi, Top-n recommendations from implicit feedmar, Y. Li, R. Rosales, E. Terzi, G. Karypis (Eds.), back leveraging linked open data, in: Q. Yang, Proceedings of the 25th ACM SIGKDD Interna- I. King, Q. Li, P. Pu, G. Karypis (Eds.), Seventh ACM tional Conference on Knowledge Discovery & Data Conference on Recommender Systems, RecSys ’13, Mining, KDD 2019, Anchorage, AK, USA, August Hong Kong, China, October 12-16, 2013, ACM, 2013, 4-8, 2019, ACM, 2019, pp. 2662–2670. URL: https: pp. 85–92. URL: https://doi.org/10.1145/2507157. //doi.org/10.1145/3292500.3330652. doi:10.1145/ 2507172. doi:10.1145/2507157.2507172. 3292500.3330652. [41] W. X. Zhao, G. He, H. Dou, J. Huang, S. Ouyang, [35] Z. Sun, Z. Deng, J. Nie, J. Tang, Rotate: Knowledge J. Wen, Kb4rec: A dataset for linking knowledge graph embedding by relational rotation in complex bases with recommender systems, in: A. Cuzspace, in: 7th International Conference on Learning zocrea, F. Bonchi, D. Gunopulos (Eds.), ProceedRepresentations, ICLR 2019, New Orleans, LA, USA, ings of the CIKM 2018 Workshops co-located with May 6-9, 2019, OpenReview.net, 2019. URL: https: 27th ACM International Conference on Informa//openreview.net/forum?id=HkgEQnRqYQ. tion and Knowledge Management (CIKM 2018), [36] A. Bordes, N. Usunier, A. García-Durán, J. Weston, Torino, Italy, October 22, 2018, volume 2482 of O. Yakhnenko, Translating embeddings for CEUR Workshop Proceedings, CEUR-WS.org, 2018. modeling multi-relational data, in: C. J. C. Burges, URL: http://ceur-ws.org/Vol-2482/paper25.pdf. L. Bottou, Z. Ghahramani, K. Q. Weinberger [42] T. Di Noia, V. C. Ostuni, P. Tomeo, E. Di Sciascio, (Eds.), Advances in Neural Information Processing Sprank: Semantic path-based ranking for top-N Systems 26: 27th Annual Conference on Neural recommendations using linked open data, ACM Information Processing Systems 2013. Proceedings TIST 8 (2016) 9:1–9:34. URL: https://doi.org/10.1145/ of a meeting held December 5-8, 2013, Lake Tahoe, 2899005. doi:10.1145/2899005.
Nevada, United States, 2013, pp. 2787–2795. URL: [43] A. Ragone, P. Tomeo, C. Magarelli, T. D. Noia, https://proceedings.neurips.cc/paper/2013/hash/ M. Palmonari, A. Maurino, E. D. Sciascio, Schema1cecc7a77928ca8133fa24680a88d2f9-Abstract. summarization in linked-data-based feature selechtml. tion for recommender systems, in: A. Sefah, B. Pen[37] F. Zhang, N. J. Yuan, D. Lian, X. Xie, W. Ma, Col- zenstadler, C. Alves, X. Peng (Eds.), Proceedings of laborative knowledge base embedding for recom- the Symposium on Applied Computing, SAC 2017, mender systems, in: B. Krishnapuram, M. Shah, Marrakech, Morocco, April 3-7, 2017, ACM, 2017, A. J. Smola, C. C. Aggarwal, D. Shen, R. Rastogi pp. 330–335. URL: https://doi.org/10.1145/3019612. (Eds.), Proceedings of the 22nd ACM SIGKDD In- 3019837. doi:10.1145/3019612.3019837. ternational Conference on Knowledge Discovery [44] V. W. Anelli, T. D. Noia, E. D. Sciascio, A. Ferrara, A. C. M. Mancino, Sparse feature factorization for recommender systems with knowledge graphs, in: H. J. C. Pampín, M. A. Larson, M. C.
Willemsen, J. A. Konstan, J. J. McAuley, J. GarciaGathright, B. Huurnink, E. Oldridge (Eds.), RecSys ’21: Fifteenth ACM Conference on Recommender Systems, Amsterdam, The Netherlands, 27 September 2021 - 1 October 2021, ACM, 2021, pp. 154– 165. URL: https://doi.org/10.1145/3460231.3474243.
doi:10.1145/3460231.3474243. [45] T. Di Noia, Knowledge-enabled recommender systems: Models, challenges, solutions, in: KDWeb, volume 1959 of CEUR Workshop Proceedings, CEUR
WS.org, 2017. [46] M. A. Jaro, Advances in record-linkage methodology as applied to matching the 1985 census of tampa, florida, Journal of the American Statistical Association 84 (1989) 414–420. URL: https://www.tandfonline. com/doi/abs/10.1080/01621459.1989.10478785. doi:10.1080/01621459.1989.10478785.
arXiv:https://www.tandfonline.com/doi/pdf/10.1080/01621459.1989.10478785. [47] R. A. Wagner, M. J. Fischer, The string-to-string
correction problem, J. ACM 21 (1974) 168–173. [48] G. Kondrak, N -gram similarity and distance, in:
SPIRE, volume 3772 of Lecture Notes in Computer
Science, Springer, 2005, pp. 115–126. [49] A. Moro, A. Raganato, R. Navigli, Entity linking meets word sense disambiguation: a unified approach, Trans. Assoc. Comput. Linguistics 2 (2014) 231–244. URL: https://tacl2013.cs.columbia.edu/ojs/ index.php/tacl/article/view/291. [50] P. N. Mendes, M. Jakob, A. García-Silva, C. Bizer,
Dbpedia spotlight: shedding light on the web of documents, in: C. Ghidini, A. N. Ngomo, S. N.
Lindstaedt, T. Pellegrini (Eds.), Proceedings the 7th International Conference on Semantic Systems, ISEMANTICS 2011, Graz, Austria, September 7-9, 2011, ACM International Conference Proceeding Series, ACM, 2011, pp. 1–8. URL: https://doi.org/ 10.1145/2063518.2063519. doi:10.1145/2063518.
2063519. [51] G. Rizzo, R. Troncy, NERD: A framework for unifying named entity recognition and disambiguation extraction tools, in: W. Daelemans, M. Lapata, L. Màrquez (Eds.), EACL 2012, 13th Conference of the European Chapter of the Association for Computational Linguistics, Avignon, France, April 2327, 2012, The Association for Computer Linguistics, 2012, pp. 73–76. URL: https://aclanthology.org/
E12-2015/. [52] J. Daiber, M. Jakob, C. Hokamp, P. N. Mendes, Improving eficiency and accuracy in multilingual entity extraction, in: M. Sabou, E. Blomqvist, T. D.
Noia, H. Sack, T. Pellegrini (Eds.), I-SEMANTICS