The Web stores a large volume of web-tables with semi-structured data. The Semantic Web community considers them as a valuable source for the knowledge graph population. Interrelated named entities can be extracted from web-tables and mapped to a knowledge graph. It generally requires reconstructing the semantics missing in web-tables to interpret them according to their meaning. This paper discusses prospects of an end-to-end solution for the knowledge graph population by entities extracted from web-tables of predefined types. The discussion covers theoretical foundations both for transforming data from web-tables to entity sets (table analysis) and for mapping entities, attributes, and relations to a knowledge graph (semantic table annotation). Unlike general-purpose text mining and web-scraping tools, we aim at developing a solution that takes into account the relational nature of the information represented in web-tables. In contrast to the table-specific proposals, our approach implies both the table analysis and the semantic table annotation.
The Web stores a large volume of tables. The exploration of the Web crawl discovered hundreds of millions of web-tables containing relational data [1,2]. There are at least billiards of valuable facts that can be extracted from web-tables. All of these make web-tables an attractive data source in various applications, such as knowledge base construction [3,4], question-answering systems, and table augmentation [5,6,7]. However, in general, web-tables are not interpretable by computer programs. Their original representation does not provide all explicit semantics required to interpret them according to their meaning. This hinders the wide usage of such tabular data in practice.
Reconstruction of the semantics missing in tables is commonly referred to as the "table understanding". This problem was first formulated by M. Hurst in 2000 [8,9]. Over the two last decades, hundreds of papers devoted to its issues were published [6,7,10,11]. The literature survey shows that this topic continues to rapidly develop in several communities such as document understanding, semantic web, and end-user programming. The last 3 years were marked by an extraordinary growth of proposals based on a novel apparatus, namely, deep learning, word and entity embedding, and knowledge graphs. There several are challenges for the expert community. One of them is to develop of a common theoretical and technological basis applicable to various digital environments and formats for representing tabular data in the Web (such as print-oriented documents, spreadsheets, and web-pages). Our approach is addressed to this challenge, namely the extraction and semantic interpretation of data from web-tables represented in HTML-format (Fig. 1).
The recent surveys of the thematic literature [6,7,10,11] note that the problem of table understanding remains open. The review [6] revealed that the majority of the works focuses mainly on the tasks of discrimination and semantic interpretation of web-tables. Roldán et al. [7] indicated that none of the known solutions is complete. They do not provide all steps of the table understanding. This is also confirmed by Burdick et al. [10]. As reported in [7], many table design properties are not taken into account by the state-of-the-art solutions. This often hinders their practical application.
The existing models of table representation do not completely reflect the complexity of the structure of real tables. One of the commonly-used assumptions is "all cell values are atomic". They assume that any non-blank cell contains only one functional data item. To the best of our knowledge, all competitive solutions follow this simplification. However, a real cell can have several data items with the same or different functions. The latter should be taken to account in order to extend the range of cases for table processing.
The novelty of our proposal is established by the following. First, we propose an end-to-end solution covering the stages from extracting data from syntactically tagged tables to their semantic interpretation (i.e. mapping extracted data and metadata to a cross-domain knowledge graph). Second, we take into account structured cells which content should be decomposed into several atomic data items with different functional roles. Third, our proposal can show the applicability of some promising techniques (cell embeddings, contextualized word embeddings, entity embedding) to the tasks of table understanding.
Unlike the general-purpose text mining and web-scraping tools, our solution takes into account the relational nature of the information represented in web-tables. In contrast to similar proposals that target data extraction from web-tables, we cover a wider range of cases by involving the structured content of a cell. Moreover, the competitive techniques are limited either by data extraction stage or the stage of semantic table interpretation, whereas, our approach implies both of them. Therefore, the expected results could be applied in the knowledge base population.
The approach to the data extraction from web-tables includes two stages: (i) classifying webtables by predefined types; (ii) extracting entity sets from web-tables, using algorithms appropriated to the corresponding types.
In the last decade, several taxonomies of web-table types were published in the last decade [12,13,14]. All taxonomies describe three common types of web-tables with relational data (Fig. 2). Eberius et al. refer to them as "vertical listing", "horizontal listing", and "matrix". This taxonomy is used by the latest proposals for the table type classification [15,16,17,18]. We also rely on this taxonomy. This will allow us to quantitatively compare our results with others.
We plan to develop a deep neural network model for classifying web-tables based on DeepTable1 [17], the ad-hoc architecture that provides four main blocks: (i) Embedding layer for extracting vector representation of cell tokens; (ii) LSTM (recurrent neural network) for identifying semantic dependencies between tokens in a cell; (iii) MLP (multilayer perceptron) for identifying non-linear dependencies between all cells in a table; (iv) Softmax as a classification
layer.
An open collection of tagged tables extracted from biomedical research papers (PubMed Central 2 ) can be used as training data. To select a basic tool for contextualized vector representation of words, we propose to try several variations (ELMo 3 [19], fastText 4 [20], etc.). Some classifiers can be trained for each variation. This allows us to compare their accuracy and choose the best for this task.
We propose to develop algorithms that target three table types of [14]. The algorithms should analyze the logical structure of web-tables by using built-in rules and trained classifiers dealing with these types. It is important to note that web-tables mix data and metadata. Moreover, one cell may contain several values of both data and metadata. The extraction of the logical structure requires: (i) to associate each cell value (data item) with its functional role (data and metadata); (ii) to associate data values with metadata ones; (iii) to group data belonging to one record (entity).
After the table type classification, the extraction of data and metadata values becomes a "cornerstone" step. We plan to implement a classifier based on machine learning algorithms to assign functional roles to data items. A promising approach to encoding cell context in the vector representation named "cell embedding" was recently proposed in [21,22]. In our case, the cell context can be deduced automatically by using the built-in type-specific table structure analysis. This approach should allow cells classification taking into account the properties of their layout and formatting, as well as the semantic similarity of their text content. We plan to reduce the number of possible false-positive and false-negative errors by using some table-specific constraints. To associate data with metadata and group data values with records, we also suggest using rule-based analysis of the table structure. The extracted entity sets can be represented in JSON notation compatible with TOMATE 5 [18], a recently published framework for the performance evaluation of web-table data extraction tools.
There are 3 main approaches to the semantic table interpretation, namely: (i) ontology matching;
(ii) entity lookup and wikification; (iii) vector representations of knowledge graphs (entity embedding). The recent study [23] showed experimentally that a hybrid approach combining lookup services and entity the efficient We plan to exploit such a hybrid using the available toolset (DBpedia Lookup6 , DBpedia SPARQL Endpoint7 , DBpedia Spotlight8 , RDF2Vec9 , KGloVe10 , and Wikipedia2Vec11 ).
The end-to-end semantic table interpretation includes 3 stages:
• Entity linking -CEA (Cell-Entity Annotation).
• Attribute-concept matching -CTA (Column-Type Annotation).
• Relation extraction -CPA (Column-Property Annotation).
As a result, this enables knowledge graph augmentation. Fig. 3 shows two examples where webtables from Wikipedia 1213 (Fig. 3, a, b) are normalized and enriched by the semantic annotation (Fig. 3, c, d), i.e. links to a knowledge graph.
The proposed solution should provide for the following: (i) identifying a subject column containing names of entities listed in a table ; (ii) lookup a set of candidate 𝐾𝐺-instances for each entity; (iii) entity disambiguation in cases when several candidate KG-instances are associated with an entity. The subject column is selected among potential keys that contain entity mentions. We are limited by the trivial case when there is only one candidate subject column. (Note that the general case requires the end-to-end semantic table interpretation).
As the main tool for linking entities, we propose to use the vector representations of subsets from a knowledge graph. The initial lookup of candidate KG-instances can be performed by using SPARQL-queries to the knowledge graph. Such queries are composed of surface forms contained in the text of cells. Each KG-instance can be encoded as a vector representation of the entity by the existing algorithms, such as RDF2Vec [24], KGloVe [25], or Wikipedia2Vec [26]. The formed vector model should allow us to use some semantic similarity metrics [27] to rank candidate KG-instances by relevance to the entity.
The approach to the entity disambiguation relies on the assumption proposed by [28] which implies that that the most relevant KG-instances from the candidate sets have the highest semantic similarity values in pairwise matching. This can be explained by the following example from [28]. Let a column contain 3 mentions: "USA", "China", and "India". They should be matched to 3 sets of candidate KG-instances respectively: "USA" → ["University of South Alabama (University)", "United States of America (Country)"], "China" → ["People's Rep. of China (Country)", "China (Band)", "China, Kagoshima (City)"], "India" → ["India (Country)", "India (George W. Bush's cat)", "India (Xandria album)"]. Among all pairs of KG-instances, "United States of America (Country)", "People's Rep. of China (Country)" and "India (Country)" would be the most semantically (they the knowledge graph).
Thus, this approach should allow us to rank candidate KG-instances and select from them the reference KG-instances for specific mentions. For example, the table showed in Fig. 3, a contains the surface form "London" that can mean "Location" or "Degree powers". Obviously, in the context of the column [Oxford, Cambridge, London, Durham, Stockton] it should be assigned to the instance of "Location" while in the context of the column [Full, London, Taught] it corresponds to the instance of "Degree powers".
In practice, many tables are not accompanied by metadata (named attributes). Generally, to map a column to a KG-class, first it is needed to associate the entities listed in the column with the reference KG-instances. After that, it is possible to form an index of all candidate KG-classes to which the reference KG-instances belong. Among them, the KG-class which is most relevant to all column values is selected. For example, in Fig. 3, a three columns should be matched to KG-classes (Fig. 3, b) as follows:
"University" -> o:EducationalInstitution "Location" -> o:Location "Degree powers" -> o:DegreePowers While the rest of columns are corresponded to KG-properties of o:EducationalInstitution (KG-class) as follows:
"Established" -> o:Established "Num. of students" -> o:NumOfStudents "Tuition fee" -> o:Tuition(£)
In the cases when entity linking (CEA-stage) fails, we propose to use ANN-models to predict the KG-class of a column based on ColNet algorithms [29,30]. To map a column of literal values (NUMERIC, DATE, CURRENCY, etc.) to a KG-datatype, it is enough to recognize standard named entities. This is reached by using regular expressions and NER-models available in popular NLP-libraries (e.g. Stanford CoreNLP 14 , AllenNLP 15 ).
To map pairs of columns (< 𝐸, 𝑃 >, where 𝐸 is a subject, 𝑃 is not a subject) with KG-properties, we plan to use entity relatedness metrics [31]. It is assumed that these metrics will allow ranking the index of candidate KG-properties and choosing the most relevant ones. For example, two web-tables showed in Fig. 3
An entity set represented as linked data (RDF-triples with URI-references to concepts in a knowledge graph) should be suitable for further interpretation. In particular, other facts (RDFtriples) can be inferred from them and asserted to the knowledge graph. Such restored semantics would provide populating the existing knowledge graphs with new entities extracted from web-tables. For example, Fig. 4 shows the terminological level (TBox) of a knowledge graph constructed by using 49 tables scrapped from Wikipedia pages in "Category: Universities in the United Kingdom". The facts extracted from these tables can be asserted into the ABox component of the knowledge graph.
We plan to demonstrate the applicability of the proposed solution by an illustrative example of populating a domain-specific knowledge graph (ABox component) in the area of industrial safety expertise. This should cover aligning entity set records with the structure of the knowledge graph (row-to-instance matching) and synthesizing new KG-instances and KG-properties. Tables extracted from real reports on industrial safety expertise may be used as a source of domain data.
Our previous work [32,33,34] was aimed at data extraction from spreadsheets driven by user-defined rules. We proposed end-user programming as the main approach. This allowed us to support specific tricks of table layout, formatting, and content. However, scaling such solutions may be challenging when there are ambiguous tricks applied within source tables. Nonetheless, a solution for the Web should be easily This is possible there types of web-tables. The latter is needed to classify them and select type-specific algorithms of analysis and interpretation. Thus, our previous approach is suitable for spreadsheet sources, but not for the Web.
The current proposal aims to fill this gap by the development of a scalable solution for web-tables. The expected results contribute to the following: (i) data extraction, including algorithms for classifying web-tables by types of taxonomy and extracting entity sets from tables of predefined types, (ii) semantic table annotation, including algorithms for mapping entities, attributes, and relations to concepts of an external knowledge graph, (iii) open software for implementing the functionality of the extraction and semantic annotation of tabular data in applications of the knowledge graph population.
To the best of our knowledge, all existing proposals for data extraction from web-tables exploit a specific constraint: "any cell contains only one atomic data item". This constraint can be eliminated in the proposed solution. We argue that the structured content of a cell can be decomposed into several data items. Moreover, all proposals implement the semantic table interpretation only for entity sets, not pivots. We plan to study both kinds of tabular data. We think this can expand the range of cases to be processed.
We propose to apply the state-of-the-art methods and tools, including contextualized word embeddings, vector representations of knowledge graphs, entity lookup services, as well as metrics of semantic similarity and entity relatedness. The applicability of some of these tools for the considered issues remains poorly studied. The expected results could demonstrate the promise of the use of these techniques.
The expected results could be useful to intellectualize software for tabular data extraction and integration in scientific and industrial applications. It can be of particular interest in areas with the intensive use of tabular data (e.g., finance, government statistics, and business management) to form linked open data.
This work was supported by the Russian Science Foundation (Grant No. 18-71-10001).