Abstract
The paper presents RuBQ, the first Russian knowledge base question answering (KBQA) dataset. The high-quality dataset consists of 1,500 Russian questions of varying complexity, their English machine translations, SPARQL queries to Wikidata, reference answers, as well as a Wikidata sample of triples containing entities with Russian labels. The dataset creation started with a large collection of question-answer pairs from online quizzes. The data underwent automatic filtering, crowd-assisted entity linking, automatic generation of SPARQL queries, and their subsequent in-house verification.
The freely available dataset will be of interest for a wide community of researchers and practitioners in the areas of Semantic Web, NLP, and IR, especially for those working on multilingual question answering. The proposed dataset generation pipeline proved to be efficient and can be employed in other data annotation projects.
V. Korablinov—Work done as an intern at JetBrains Research.
Access provided by Autonomous University of Puebla. Download conference paper PDF
Similar content being viewed by others
Keywords
-
Resource location: http://doi.org/10.5281/zenodo.3835913
-
Project page: https://github.com/vladislavneon/RuBQ
1 Introduction
Question answering (QA) addresses the task of returning a precise and concise answer to a natural language question posed by the user. QA received a great deal of attention both in academia and industry. Two main directions within QA are Open-Domain Question Answering (ODQA) and Knowledge Base Question Answering (KBQA). ODQA searches for the answer in a large collection of text documents; the process is often divided into two stages: 1) retrieval of potentially relevant paragraphs and 2) spotting an answer span within the paragraph (referred to as machine reading comprehension, MRC). In contrast, KBQA uses a knowledge base as a source of answers. A knowledge base is a large collection of factual knowledge, commonly structured in subject–predicate–object (SPO) triples, for example (Vladimir_Nabokov, spouse, Véra_Nabokov).
A potential benefit of KBQA is that it uses knowledge in a distilled and structured form that enables reasoning over facts. In addition, knowledge base structure is inherently language-independent – entities and predicates are assigned unique identifiers that are tied to specific languages through labels and descriptions, – which makes KBs more suitable for multilingual QA. The task of KBQA can be formulated as a translation from natural language question into a formal KB query (expressed in SPARQL, SQL, or \(\lambda \)-calculus). In many real-life applications, like in Jeopardy! winning IBM Watson [15] and major search engines, hybrid QA systems are employed – they rely on both text document collections and structured knowledge bases.
High-quality annotated data is crucial for measurable progress in question answering. Since the advent of SQuAD [27], a wide variety of datasets for machine reading comprehension have emerged, see a recent survey [39]. We are witnessing a growing interest in multilingual question answering, which leads to the creation of multilingual MRC datasets [1, 8, 24]. Multilingual KBQA is also an important research problem and a promising application [9, 16]. Russian is among top-10 languages by its L1 and L2 speakersFootnote 1; it has a Cyrillic script and a number of grammar features that make it quite different from e.g. English and Chinese – the languages most frequently used in NLP and Semantic Web research.
In this paper we present RuBQ (pronounced [‘rubik]) – Russian Knowledge Base Questions, a KBQA dataset that consists of 1,500 Russian questions of varying complexity along with their English machine translations, corresponding SPARQL queries, answers, as well as a subset of Wikidata covering entities with Russian labels. To the best of our knowledge, this is the first Russian KBQA and semantic parsing dataset. To construct the dataset, we started with a large collection of trivia Q&A pairs harvested on the Web. We built a dedicated recall-oriented Wikidata entity linking tool and verified the obtained answers’ candidate entities via crowdsourcing. Then, we generated paths between possible question entities and answer entities and carefully verified them.
The freely available dataset is of interest for a wide community of Semantic Web, natural language processing (NLP), and information retrieval (IR) researchers and practitioners, who deal with multilingual question answering. The proposed dataset generation pipeline proved to be efficient and can be employed in other data annotation projects.
2 Related Work
Table 1 summarizes the characteristics of KBQA datasets that have been developed to date. These datasets vary in size, underlying knowledge base, presence of questions’ logical forms and their formalism, question types and sources, as well as the language of the questions.
The questions of the earliest Free917 dataset [7] were generated by two people without consulting a knowledge base, the only requirement was a diversity of questions’ topics; each question is provided with its logical form to query Freebase. Berant et al. [3] created WebQuestions dataset that is significantly larger but does not contain questions’ logical forms. Questions were collected through Google suggest API: authors fed parts of the initial question to the API and repeated the process with the returned questions until 1M questions were reached. After that, 100K randomly sampled questions were presented to MTurk workers, whose task was to find an answer entity in Freebase. Later studies have shown that only two-thirds of the questions in the dataset are completely correct; many questions are ungrammatical and ill-formed [37, 38]. Yih et al. [38] enriched 81.5% of WebQuestions with SPARQL queries and demonstrated that semantic parses substantially improve the quality of KBQA. They also showed that semantic parses can be obtained at an acceptable cost when the task is broken down into smaller steps and facilitated by a handy interface. Annotation was performed by five people familiar with Freebase design, which hints at the fact that the task is still too tough for crowdsourcing. WebQuestions were used in further studies aimed to generate complex questions [2, 31]. SimpleQuestions [5] is the largest manually created KBQA dataset to date. Instead of providing logical parses for existing questions, the approach explores the opposite direction: based on formal representation, a natural language question is generated by crowd workers. First, the authors sampled SPO triples from a Freebase subset, favoring non-frequent subject–predicate pairs. Then, the triples were presented to crowd workers, whose task was to generate a question about the subject, with the object being the answer. This approach doesn’t guarantee that the answer is unique – Wu et al. [37] estimate that SOTA results on the dataset (about 80% correct answers) reach its upper bound, since the rest of the questions are ambiguous and cannot be answered precisely. The dataset was used for the fully automatic generation of a large collection of natural language questions from Freebase triples with neural machine translation methods [29]. Dieffenbach et al. [11] succeeded in a semi-automatic matching of about one-fifth of the dataset to Wikidata.
The approach behind FreebaseQA dataset [19] is the closest to our study – it builds upon a large collection of trivia questions and answers (borrowed largely from TriviaQA dataset for reading comprehension [20]). Starting with about 130K Q&A pairs, the authors run NER over questions and answers, match extracted entities against Freebase, and generate paths between entities. Then, human annotators verify automatically generated paths, which resulted in about 28 K items marked relevant. Manual probing reveals that many questions’ formal representations in the dataset are not quite precise. For example, the question eval-25: Who captained the Nautilus in 20,000 Leagues Under The Sea? is matched with the relation book.book.characters that doesn’t represent its meaning and leads to multiple answers along with a correct one (Captain Nemo). Our approach differs from the above in several aspects. We implement a recall-oriented IR-based entity linking since many questions involve general concepts that cannot be recognized by off-the-shelf NER tools. After that, we verify answer entities via crowdsourcing. Finally, we perform careful in-house verification of automatically generated paths between question and answer entities in KB. We can conclude that our pipeline leads to a more accurate representation of questions’ semantics.
The questions in the KBQA datasets can be simple, i.e. corresponding to a single fact in the knowledge base, or complex. Complex questions require a combination of multiple facts to answer them. WebQuestions consists of 85% simple questions; SimpleQuestions and 30M factoid QA Corpus contain only simple questions. Many studies [2, 12, 13, 21, 28, 31] purposefully target complex questions.
The majority of datasets use Freebase [4] as target knowledge base. Freebase was discontinued and exported to Wikidata [25]; the latest available Freebase dump dates back to early 2016. QALD [33] and both versions of LC-QuAD [13, 32] use DBpedia [22]. LC QuAD 2.0 [13] and ComplexSequentialQuestions [28] use Wikidata [36], which is much larger, up-to-date, and has more multilingual labels and descriptions. The majority of datasets, where natural language questions are paired with logical forms, employ SPARQL as a more practical and immediate option compared to lambda calculus.
Existing KBQA datasets are almost exclusively English, with Chinese MSParS dataset being an exception [12]. QALD-9 [33], the latest edition of QALD shared task,Footnote 2 contains questions in 11 languages: English, German, Russian, Hindi, Portuguese, Persian, French, Romanian, Spanish, Dutch, and Italian. The dataset is rather small; at least Russian questions appear to be non-grammatical machine translations.Footnote 3
There are several studies on knowledge base question generation [14, 17, 21, 29]. These works vary in the amount and form of supervision, as well as the structure and the complexity of the generated questions. However, automatically generated questions are intended primarily for training; the need for high-quality, human-annotated data for testing still persists.
3 Dataset Creation
Following previous studies [19, 20], we opted for quiz questions that can be found in abundance online along with the answers. These questions are well-formed and diverse in terms of properties and entities, difficulty, and vocabulary, although we don’t control these properties directly during data processing and annotation.
The dataset generation pipeline consists of the following steps: 1) data gathering and cleaning; 2) entity linking in answers and questions; 3) verification of answer entities by crowd workers; 4) generation of paths between answer entities and question candidate entities; 5) in-house verification/editing of generated paths. In parallel, we created a Wikidata sample containing all entities with Russian labels. This snapshot ensures reproducibility – a reference answer may change with time as the knowledge base evolves. In addition, the smaller dataset requires less powerful hardware for experiments with RuBQ. In what follows we elaborate on these steps.
3.1 Raw Data
We mined about 150,000 Q&A pairs from several open Russian quiz collections on the Web.Footnote 4 We found out that many items in the collection aren’t actual factoid questions, for example, cloze quizzes (Leonid Zhabotinsky was a champion of Olympic games in …[Tokyo]Footnote 5), crossword, definition, and multi-choice questions, as well as puzzles (Q: There are a green one, a blue one, a red one and an east one in the white one. What is this sentence about? A: The White House). We compiled a list of Russian question words and phrases and automatically removed questions that don’t contain any of them. We also removed duplicates and crossword questions mentioning the number of letters in the expected answer. This resulted in 14,435 Q&A pairs.
3.2 Entity Linking in Answers and Questions
We implemented an IR-based approach for generating Wikidata entity candidates mentioned in answers and questions. First, we collected all Wikidata entities with Russian labels and aliases. We filtered out Wikimedia disambiguation pages, dictionary and encyclopedic entries, Wikimedia categories, Wikinews articles, and Wikimedia list articles. We also removed entities with less than four outgoing relations – we used this simple heuristic to remove less interconnected items that can hardly help solving KBQA tasks. These steps resulted in 4,114,595 unique entities with 5,430,657 different labels and aliases.
After removing punctuation, we indexed the collection with Elasticsearch using built-in tokenization and stemming. Each text string (question or answer) produces three types of queries to the Elasticsearch index: 1) all token trigrams; 2) capitalized bigrams (many named entities follow this pattern, e.g. Alexander Pushkin, Black Sea); and 3) free text query containing only nouns, adjectives, and numerals from the original string. N-gram queries (types 1 and 2) are run as phrase queries, whereas recall-oriented free text queries (type 3) are executed as Elasticsearch fuzzy search queries. Results of the latter search are re-ranked using a combination of BM25 scores from Elasticsearch and aggregated page view statistics of corresponding Wikipedia articles.Footnote 6 Finally, we combine search results preserving the type order and retain top-10 results for further processing. The proposed approach effectively combines precision- (types 1 and 2) and recall-oriented (type 3) processing.
3.3 Crowdsourcing Annotations
Entity candidates for answers obtained through the entity linking described above were verified on Yandex.Toloka crowdsourcing platform.Footnote 7 Crowd workers were presented with a Q&A pair and a ranked list of candidate entities. In addition, they could consult a Wikipedia page corresponding to the Wikidata item, see Fig. 1. The task was to select a single entity from the list or the None of the above option. The average number of candidates on the list is 5.43.
Crowd workers were provided with a detailed description of the interface and a variety of examples. To proceed to the main task, crowd workers had to first pass a qualification consisting of 20 tasks covering various cases described in the instruction. We also included 10% of honeypot tasks for live quality monitoring. These results are in turn used for calculating confidence of the annotations obtained so far as a weighted majority vote (see details of the approach in [18]). Confidence value governs overlap in annotations: if the confidence is below 0.85, the task is assigned to the next crowd worker. We hired Toloka workers from the best 30% cohort according to internal rating. As a result, the average confidence for the annotation is 98.58%; the average overlap is 2.34; average time to complete a task is 19 s.
In total, 9,655 out of 14,435 answers were linked to Wikidata entities. Among the matched entities, the average rank of the correct candidate appeared to be 1.5. The combination of automatic candidate generation and subsequent crowdsourced verification proved to be very efficient. A possible downside of the approach is a lower share of literals (dates and numerical values) in the annotated answers. We could match only a fraction of those answers with Wikidata: Wikidata’s standard formatted literals may look completely different even if representing the same value. Out of 1,255 date and numerical answers, 683 were linked to a Wikidata entity such as a particular year. For instance, the answer for In what year was Immanuel Kant born? matches Q6926 (year 1724), whereas the corresponding Wikidata value is . Although the linkage is deemed correct, this barely helps generate a correct path between question and answer entities.
3.4 Path Generation and In-House Annotation
We applied entity linking described above to the 9,655 questions with verified answers and obtained 8.56 candidate entities per question on average. Next, we generated candidate subgraphs spanning question and answer entities, restricting the length between them by two hops.Footnote 8
We investigated the option of filtering out erroneous question entities using crowdsourcing analogous to answer entity verification. A pilot experiment on a small sample of questions showed that this task is much harder – we got only 64% correct matches on a test set. Although the average number of generated paths decreased (from 1.9 to 0.9 and from 6.2 to 3.5 for paths of length one and two, respectively), it also led to losing correct paths for 14% of questions. Thus, we decided to perform an in-house verification of the generated paths. The work was performed by the authors of the paper.
After sending queries to the Wikidata endpoint, we were able to find chains of length one or two for 3,194 questions; the remaining 6,461 questions were left unmatched. We manually inspected 200 random unmatched questions and found out that only 10 of them could possibly be answered with Wikidata, but the required facts are missing in the KB.
Out of 2,809 1-hop candidates corresponding to 1,799 questions, 866 were annotated as correct. For the rest 2,328 questions, we verified 3,591 2-hop candidates, but only 55 of them were deemed correct. 279 questions were marked as answerable with Wikidata. To increase the share of complex questions in the dataset, we manually constructed SPARQL queries for them.
Finally, we added 300 questions marked as non-answerable over Wikidata, although their answers are present in the knowledge base. The majority of them are unanswerable because semantics of the question cannot be expressed using the existing Wikidata predicates, e.g. How many bells does the tower of Pisa have? (7). In some cases, predicates do exist and a semantically correct SPARQL query can be formulated, but the statement is missing in the KG thus the query will return an empty list, e.g. What circus was founded by Albert Salamonsky in 1880? (Moscow Circus on Tsvetnoy Boulevard). These adversarial examples are akin to unanswerable questions in the second edition of SQuAD dataset [26]; they make the task more challenging and realistic.
4 RuBQ Dataset
4.1 Dataset Statistics
Our dataset has 1,500 unique questions in total. It mentions 2,357 unique entities – 1,218 in questions and 1,250 in answers. There are 242 unique relations in the dataset. The average length of the original questions is 7.99 words (median 7); machine-translated English questions are 10.58 words on average (median 10). 131 questions have more than one correct answer. For 1,154 questions the answers are Wikidata entities, and for 46 questions the answers are literals. We consider empty answers to be correct for 300 unanswerable questions and do not provide answer entities for them.
Inspired by a taxonomy of query complexity in LC-QuAD 2.0 [13], we annotated obtained SPARQL queries in a similar way. The query type is defined by the constraints in the SPARQL query, see Table 2. Note that some queries have multiple type tags. For example, SPARQL query for the question How many moons does Mars have? is assigned 1-hop and count types and therefore isn’t simple in terms of SimpleQuestions dataset.
Taking into account RuBQ’s modest size, we propose to use the dataset primarily for testing rule-based systems, cross-lingual transfer learning models, and models trained on automatically generated examples, similarly to recent MRC datasets [1, 8, 24]. We split the dataset into development (300) and test (1,200) sets in such a way to keep a similar distribution of query types in both subsets.
4.2 Dataset Format
For each entry in the dataset, we provide: the original question in Russian, original answer text (may differ textually from the answer entity’s label retrieved from Wikidata), SPARQL query representing the meaning of the question, a list of entities in the query, a list of relations in the query, a list of answers (a result of querying the Wikidata subset, see below), and a list of query type tags, see Table 3 for examples. We also provide machine-translated English questions obtained through Yandex.Translate without any post-editing.Footnote 9 The reason to include them into the dataset is two-fold: 1) the translations, although not perfectly correct, help understand the questions’ meaning for non-Russian speakers and 2) they are ready-to-use for cross-lingual QA experiments (as we did with English QA system QAnswer). RuBQ is distributed under CC BY-SA license and is available in JSON format.
The dataset is accompanied by RuWikidata8M – a Wikidata sample containing all the entities with Russian labels.Footnote 10 It consists of about 212M triples with 8.1M unique entities. As mentioned before, the sample guarantees the correctness of the queries and answers and makes the experiments with the dataset much simpler. For each entity, we executed a series of CONSTRUCT SPARQL queries to retrieve all the truthy statements and all the full statements with their linked data.Footnote 11 We also added all the triples with subclass of (P279) predicate to the sample. This class hierarchy can be helpful for question answering task in the absence of an explicit ontology in Wikidata. The sample contains Russian and English labels and aliases for all its entities.
4.3 Baselines
We provide two RuBQ baselines from third-party systems – DeepPavlov and QAnswer – that illustrate two possible approaches to cross-lingual KBQA.
To the best of our knowledge, the KBQA libraryFootnote 12 from an open NLP framework DeepPavlov [6] is the only freely available KBQA implementation for Russian language. The library uses Wikidata as a knowledge base and implements the standard question processing steps: NER, entity linking, and relation detection. According to the developers of the library, they used machine-translated SimpleQuestions and a dataset for zero-shot relation extraction [23] to train the model. The library returns a single string or not found as an answer. We obtained an answer entity ID using reverse ID-label mapping embedded in the model. If no ID is found, we treated the answer as a literal.
QAnswer [10] is a rule-based KBQA system that answers questions in several languages using Wikidata. QAnswer returns a (possibly empty) ranked list of Wikidata item IDs along with a corresponding SPARQL query. We obtain QAnswer’s results by sending RuBQ questions machine-translated into English to its API.Footnote 13
QAnswer outperforms DeepPavlov in terms of precision@1 on the answerable subset (16% vs. 13%), but demonstrates a lower accuracy on unanswerable questions (43% vs. 73%). Table 4 presents detailed results. In contrast to DeepPavlov, QAnswer returns a ranked list of entities as a response to the query, and for 23 out of 131 questions with multiple correct answers, it managed to perfectly match the set of answers. For eight questions with multiple answers, QAnswer’s top-ranked answers were correct, but the lower-ranked ones contained errors. To facilitate different evaluation scenarios, we provide an evaluation script that calculates precision@1, exact match, and precision/recall/F1 measures, as well as the breakdown of results by query types.
5 Conclusion and Future Work
We presented RuBQ – the first Russian dataset for Question Answering over Wikidata. The dataset consists of 1,500 questions, their machine translations into English, and annotated SPARQL queries. 300 RuBQ questions are unanswerable, which poses a new challenge for KBQA systems and makes the task more realistic. The dataset is based on a collection of quiz questions. The data generation pipeline combines automatic processing, crowdsourced and in-house verification, and proved to be very efficient. The dataset is accompanied by a Wikidata sample of 212M triples that contain 8.1M entities with Russian and English labels, and an evaluation script. The provided baselines demonstrate the feasibility of the cross-lingual approach in KBQA, but at the same time indicate there is ample room for improvements. The dataset is of interest for a wide community of researchers in the fields of Semantic Web, Question Answering, and Semantic Parsing.
In the future, we plan to explore other data sources and approaches for RuBQ expansion: search query suggest APIs as for WebQuestions [3], a large question log [35], and Wikidata SPARQL query logs.Footnote 14 We will also address complex questions and questions with literals as answers, as well as the creation of a stronger baseline for RuBQ.
Notes
- 1.
- 2.
See overview of previous QALD datasets in [34].
- 3.
We manually verified all the 558 Russian questions in the QALD-9 dataset – only two of them happen to be grammatical.
- 4.
http://baza-otvetov.ru, http://viquiz.ru, and others.
- 5.
Hereafter English examples are translations from original Russian questions and answers.
- 6.
- 7.
- 8.
We examined the sample and found out that there are only 12 questions with distances between question and answer entities in the Wikidata graph longer than two.
- 9.
- 10.
https://zenodo.org/record/3751761, project’s page on github points here.
- 11.
Details about Wikidata statement types can be found here: https://www.mediawiki.org/wiki/Wikibase/Indexing/RDF_Dump_Format#Statement_types.
- 12.
http://docs.deeppavlov.ai/en/master/features/models/kbqa.html. The results reported below are as of April 2020; a newer model has been released in June 2020.
- 13.
- 14.
References
Artetxe, M., Ruder, S., Yogatama, D.: On the cross-lingual transferability of monolingual representations. arXiv preprint arXiv:1910.11856 (2019)
Bao, J., Duan, N., Yan, Z., Zhou, M., Zhao, T.: Constraint-based question answering with knowledge graph. In: Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pp. 2503–2514 (2016)
Berant, J., Chou, A., Frostig, R., Liang, P.: Semantic parsing on freebase from question-answer pairs. In: Proceedings of the 2013 conference on empirical methods in natural language processing, pp. 1533–1544 (2013)
Bollacker, K., Evans, C., Paritosh, P., Sturge, T., Taylor, J.: Freebase: a collaboratively created graph database for structuring human knowledge. In: Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pp. 1247–1250 (2008)
Bordes, A., Usunier, N., Chopra, S., Weston, J.: Large-scale simple question answering with memory networks. arXiv preprint arXiv:1506.02075 (2015)
Burtsev, M., et al.: Deeppavlov: Open-source library for dialogue systems. In: Proceedings of ACL 2018, System Demonstrations, pp. 122–127 (2018)
Cai, Q., Yates, A.: Large-scale semantic parsing via schema matching and lexicon extension. In: Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pp. 423–433 (2013)
Clark, J.H., et al.: TyDi QA: a benchmark for information-seeking question answering in typologically diverse languages. arXiv preprint arXiv:2003.05002 (2020)
Diefenbach, D., Both, A., Singh, K., Maret, P.: Towards a question answering system over the semantic web. arXiv preprint arXiv:1803.00832 (2018)
Diefenbach, D., Giménez-García, J., Both, A., Singh, K., Maret, P.: QAnswer KG: designing a portable question answering system over RDF data. In: Hart, A., et al. (eds.) ESWC 2020. LNCS, vol. 12123, pp. 429–445. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-49461-2_25
Diefenbach, D., Tanon, T.P., Singh, K.D., Maret, P.: Question answering benchmarks for wikidata. In: ISWC (Posters & Demonstrations) (2017)
Duan, N.: Overview of the NLPCC 2019 shared task: open domain semantic parsing. In: Tang, J., Kan, M.-Y., Zhao, D., Li, S., Zan, H. (eds.) NLPCC 2019. LNCS (LNAI), vol. 11839, pp. 811–817. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-32236-6_74
Dubey, M., Banerjee, D., Abdelkawi, A., Lehmann, J.: LC-QuAD 2.0: a large dataset for complex question answering over wikidata and DBpedia. In: Ghidini, C., et al. (eds.) ISWC 2019. LNCS, vol. 11779, pp. 69–78. Springer, Cham (2019). https://doi.org/10.1007/978-3-030-30796-7_5
Elsahar, H., Gravier, C., Laforest, F.: Zero-shot question generation from knowledge graphs for unseen predicates and entity types. In: NAACL, pp. 218–228 (2018)
Ferrucci, D., et al.: Building watson: an overview of the deepQA project. AI Mag. 31(3), 59–79 (2010)
Hakimov, S., Jebbara, S., Cimiano, P.: AMUSE: multilingual semantic parsing for question answering over linked data. In: d’Amato, C., et al. (eds.) ISWC 2017. LNCS, vol. 10587, pp. 329–346. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-68288-4_20
Indurthi, S.R., Raghu, D., Khapra, M.M., Joshi, S.: Generating natural language question-answer pairs from a knowledge graph using a RNN based question generation model. In: Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics, pp. 376–385 (2017)
Ipeirotis, P.G., Provost, F., Sheng, V.S., Wang, J.: Repeated labeling using multiple noisy labelers. Data Min. Knowl. Discov. 28(2), 402–441 (2014)
Jiang, K., Wu, D., Jiang, H.: FreebaseQA: a new factoid QA data set matching trivia-style question-answer pairs with Freebase. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pp. 318–323 (2019)
Joshi, M., Choi, E., Weld, D.S., Zettlemoyer, L.: TriviaQA: a large scale distantly supervised challenge dataset for reading comprehension. In: ACL, pp. 1601–1611 (2017)
Keysers, D., et al.: Measuring compositional generalization: a comprehensive method on realistic data. In: ICLR (2020)
Lehmann, J., et al.: DBpedia-a large-scale, multilingual knowledge base extracted from Wikipedia. Seman. Web 6(2), 167–195 (2015)
Levy, O., Seo, M., Choi, E., Zettlemoyer, L.: Zero-shot relation extraction via reading comprehension. In: CoNLL, pp. 333–342 (2017)
Lewis, P., Oğuz, B., Rinott, R., Riedel, S., Schwenk, H.: MLQA: evaluating cross-lingual extractive question answering. arXiv preprint arXiv:1910.07475 (2019)
Pellissier Tanon, T., Vrandečić, D., Schaffert, S., Steiner, T., Pintscher, L.: From freebase to wikidata: the great migration. In: Proceedings of the 25th international conference on world wide web, pp. 1419–1428 (2016)
Rajpurkar, P., Jia, R., Liang, P.: Know what you don’t know: unanswerable questions for SQuAD. In: ACL, pp. 784–789 (2018)
Rajpurkar, P., Zhang, J., Lopyrev, K., Liang, P.: SQuAD: 100,000+ questions for machine comprehension of text. In: EMNLP, pp. 2383–2392 (2016)
Saha, A., Pahuja, V., Khapra, M.M., Sankaranarayanan, K., Chandar, S.: Complex sequential question answering: towards learning to converse over linked question answer pairs with a knowledge graph. arXiv preprint (2018)
Serban, I.V., et al.: Generating factoid questions with recurrent neural networks: the 30M factoid question-answer corpus. In: ACL, pp. 588–598 (2016)
Su, Y., et al.: On generating characteristic-rich question sets for QA evaluation. In: Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pp. 562–572 (2016)
Talmor, A., Berant, J.: The Web as a knowledge base for answering complex questions. In: NAACL, pp. 641–651 (2018)
Trivedi, P., Maheshwari, G., Dubey, M., Lehmann, J.: LC-QuAD: a corpus for complex question answering over knowledge graphs. In: d’Amato, C., et al. (eds.) ISWC 2017. LNCS, vol. 10588, pp. 210–218. Springer, Cham (2017). https://doi.org/10.1007/978-3-319-68204-4_22
Usbeck, R., Gusmita, R.H., Axel-Cyrille Ngonga Ngomo, Saleem, M.: 9th challenge on question answering over linked data (QALD-9). In: SemDeep-4, NLIWoD4, and QALD-9 Joint Proceedings, pp. 58–64 (2018)
Usbeck, R., et al.: Benchmarking question answering systems. Semant. Web 10(2), 293–304 (2019)
Völske, M., et al.: What users ask a search engine: analyzing one billion Russian question queries. In: Proceedings of the 24th ACM International on Conference on Information and Knowledge Management, pp. 1571–1580 (2015)
Vrandečić, D., Krötzsch, M.: Wikidata: a free collaborative knowledgebase. Commun. ACM 57(10), 78–85 (2014)
Wu, Z., Kao, B., Wu, T.H., Yin, P., Liu, Q.: PERQ: Predicting, explaining, and rectifying failed questions in KB-QA systems. In: Proceedings of the 13th International Conference on Web Search and Data Mining, pp. 663–671 (2020)
Yih, W.T., Richardson, M., Meek, C., Chang, M.W., Suh, J.: The value of semantic parse labeling for knowledge base question answering. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pp. 201–206 (2016)
Zhang, X., Yang, A., Li, S., Wang, Y.: Machine reading comprehension: a literature review. arXiv preprint arXiv:1907.01686 (2019)
Acknowledgments
We thank Mikhail Galkin, Svitlana Vakulenko, Daniil Sorokin, Vladimir Kovalenko, Yaroslav Golubev, and Rishiraj Saha Roy for their valuable comments and fruitful discussion on the paper draft. We also thank Pavel Bakhvalov, who helped collect RuWikidata8M sample and contributed to the first version of the entity linking tool. We are grateful to Yandex.Toloka for their data annotation grant. PB acknowledges support by Ural Mathematical Center under agreement No. 075-02-2020-1537/1 with the Ministry of Science and Higher Education of the Russian Federation.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Copyright information
© 2020 Springer Nature Switzerland AG
About this paper
Cite this paper
Korablinov, V., Braslavski, P. (2020). RuBQ: A Russian Dataset for Question Answering over Wikidata. In: Pan, J.Z., et al. The Semantic Web – ISWC 2020. ISWC 2020. Lecture Notes in Computer Science(), vol 12507. Springer, Cham. https://doi.org/10.1007/978-3-030-62466-8_7
Download citation
DOI: https://doi.org/10.1007/978-3-030-62466-8_7
Published:
Publisher Name: Springer, Cham
Print ISBN: 978-3-030-62465-1
Online ISBN: 978-3-030-62466-8
eBook Packages: Computer ScienceComputer Science (R0)