Synthesizing information from collections of tables embedded within scientific and technical documents is increasingly critical to emerging knowledge-driven applications. Given their structural heterogeneity, highly domain-specific content, and difuse context, inferring a precise semantic understanding of such tables is traditionally better accomplished through linking tabular content to concepts and entities in reference knowledge graphs. However, existing tabular data discovery systems are not designed to adequately exploit these explicit, human-interpretable semantic linkages. Moreover, given the prevalence of misinformation, the level of confidence in the reliability of tabular information has become an important, often overlooked, factor in discovery over open datasets. We describe a preliminary implementation of a discovery engine that enables table-based semantic search and retrieval of tabular information from a linked knowledge graph of scientific tables. We discuss the viability of semantics-guided tabular data analysis operations, including on-the-fly table generation under reliability constraints, within discovery scenarios motivated by intelligence production from documents.
Scientific tables, semantic tabular data discovery, on-the-fly table generation, data fusion © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License proach and preliminary implementation of a tabular data discovery system driven by knowledge graph technology.
Tables are ubiquitous across domains in organizing and concisely communicating information in a structured form. Increasing democratization of generative AI techniques and the popularity of their application in the comprehension, creation, and refinement of complex multimodal digital content such as technical documents is triggering a fresh revisit of tables—often, a key structured data component embedded within various kinds of published documents ranging from scientific papers and preprint articles to patents and contractual agreements to intelligence reports and impact assessment statements. We use the term “scientific tables” to denote this sub-category of tabular data objects whose primary role (alongside other structured artifacts like charts) is one of supplementing a document’s textual content with vital visual cues, and whose access and interpretation is nEvelop-O (A. Joshi) (A. Joshi)
Tables in scientific documents often capture a point-intime summary view or meta-analysis over a more comprehensive set of information—–including over large structured datasets hosted across diverse locations on the web (e.g., open data portals, data markets, research data repositories, etc.). While these latter datasets typically feed data preparation tasks aimed at automating data science and analysis pipelines, we focus more on the former kind of tables to automate technical content generation pipelines. Specifically, we seek to enable technical experts and analysts to (i) eficiently discover useful information in existing scientific tables, (ii) analyze knowledge extracted from across these tables in the wider context of their visual appearance and reliability, and (iii) present their learnings and valuable information (again, in the form of scientific tables) for possible inclusion in new documents or reports. Our work is motivated by how intelligence community standards instruct analysts to “incorporate efective visual presentations” of information (including via tables) to enhance the overall usefulness of intelligence reports [10]. Additionally, we derive inspiration from a recent trend of scientists resorting to AI and conversational search engines as an evolving modern-day lazyweb personification for generating tabular content in lieu of conducting the research themselves.
The very practices seeking to ease human comprehension of scientific tables also introduce challenges for machinedriven table understanding. Scientific tables exhibit certain distinctive characteristics borne out of the general circumstances of their creation: 1. High structural heterogeneity: Constrained by ‘publication real estate’ and desire to place tables in close proximity to any accompanying text, collections of scientific tables display high structural variability, even more so than web tables. Data discovery and integration systems with data and schema matching techniques that assume ‘well-structured’ or relational tables do not adequately address this complexity. semantics of its content. Based on how tables are visually formatted to optimize informational content for human consumption, this context may include inferred semantics of other cells in a row or column, table captions or other descriptive text (from within the body of the containing documents) that refers to these tables.
reviewed preprints, there is a dearth of approaches to tackle the lack of information reliability of scientific tables , as is the case with web tables and open datasets.
In [8], we describe in detail our solution to address some of these challenges via extensive structural characterization of over 120,000 tables drawn from scientific publications, and by matching tabular content to reference knowledge graphs with high precision. We also developed a practical entity linker [9], adaptable to different domains, to eficiently match COVID-19-related scientific tables to Wikidata [ 12].
meta-analyses information about phase I clinical trials for COVID-19 vaccines developed around the world. Unless such information is centrally curated, it will be more expeditiously available directly within scientific tables in published documents (e.g., figure 1).
While some publisher sea.rch services [13] can specif2. Domain-specific entities: Like open datasets, sci- ically return tables from relevant papers, such matches entific tables typically contain more numerical cell are not based on tabular data. In reality, most tabular content than text. Where they do contain text, it is content in scientific documents is not directly accessible usually in the form of literals or idiomatic strings and even in instances where they are internally maintained entities specific to a scientific domain. Data semantics as machine-readable (e.g., HTML-formatted) objects. As[11] play a key role in disambiguating such content. suming that analyst queries are best served by synthesizing tabular responses and that the primary source of 3. As with web tables, scientific tables exhibit difuse information for doing so are existing scientific tables, one context wherein one must draw upon additional con- could potentially adapt current semantic schema matchtextual information that lies outside an individual ing techniques to help discover ranked lists of tables with table cell (or, even an entire table body) to infer the matching content. However, as with open datasets, it is unlikely that individual scientific tables will contain all information requested by queries. Instead, relevant “scientific views” must be composed on the fly by suitably merging content from multiple scientific tables. We introduce technical challenges with on-the-fly generation of relational scientific tables in response to search requests under contextual constraints and briefly describe our heuristic approach to address them in section 3.3. 3. Technical Approach parse an input table-based search request into an intermediate query plan comprising a set of abstract foundational primitives: SELECT, FILTER, RANK (and FUSE). Akin to relational algebra operators, SELECT logically returns a list of identifiers for all tables that semantically match the query table. FILTER prunes this list by applying one or more temporal, cell coverage-based, or reliability-based constraints on matching tables. RANK orthogonally reorders the list of tables based on some ranking criteria.
Any table-based search request can be expressed as a query plan comprising these primitives.
of knowledge-based analysis to generate new tables on the fly using information integrated from multiple scientific tables. Our goal is to populate a result table with new knowledge, potentially even one cell at a time.
Intermediate query plans are automatically translated into subgraph triple pattern-matching queries for execution against our KG of scientific tables. We built a • We first constructed a KG from scientific tables tabular data discovery engine to incrementally construct [8]—Specifically, we perform column type annota- SPARQL queries by adding or modifying ad hoc graph tion (CTA) and cell entity annotation (CEA) [14]—i.e., patterns corresponding to each primitive instance in a header and body cells are automatically annotated with query plan. As depicted in figure 2, the discovery of relWikidata concepts and entities respectively (e.g., the evant existing scientific tables without on-the-fly table cell with content ‘Platform’ from the table in figure 1 is generation requires a single pass over the engine’s comlinked to entity with QID: Q108028785: “Vaccine Plat- ponents. This engine is also responsible for packaging form”). Each table’s structural assessments, inferred the results of SPARQL query execution in the form of semantics, and provenance-based estimates of reliabil- relational result tables for easy consumption. ity are all encoded as RDF triples in accordance with By translating query plans into SPARQL as shown, linked data principles. our engine is efectively emulating “database-like” analysis against tables in our KG—where each primitive’s • We then designed and implemented a prototype system implementation is driven by explicit semantic linkages for ultimately discovering tabular data from this KG automatically inferred for each table. SELECT compares of scientific tables. As an initial step towards discov- the set of header-cell QIDs for the query table against ery and on-the-fly generation [ 15], we first developed those for each KG table. If the two sets of QIDs overfoundational capabilities—including a query model and lap completely, or if set overlap exceeds some threshdiscovery engine—to enable table-based search over old specified in the request, then the candidate table is our KG under rich contextual constraints. Finally, we deemed semantically similar and included in the results. extended these foundational capabilities with a pre- By default, RANK sorts a list of result tables based on this liminary, heuristic approach to identify and fuse the header-cell coverage metric (i.e., tables with maximum content of semantically compatible scientific tables on QID set overlap are ranked higher). Besides exact comthe fly via union operations. Our overall approach parison of header-cell QIDs, the engine also optionally is analogous to the ‘reference architecture’ approach supports QID similarity based on pre-trained knowledge described in [16] to discover project-join views. graph embeddings from Wembedder [17]. While a detailed algorithmic description of query trans3.1. Table-based Semantic Search lation is beyond the scope of this paper, in general, given a search request, a Query Parser identifies and assembles Search requests against our KG can take the form of a key- (in a specific order) all information needed to formulate word list or a (potentially partially-specified query-by- a SPARQL SELECT query—including: return variables example) [16]) input table—–which can then be semanti- and limits, (subject, predicate and object) for base triple cally resolved to reduce ambiguity—along with any asso- patterns, entire subgraph patterns (if applicable), variciated contextual constraints. In response to a request, able bindings and clauses for the query constraints, and we match the semantics of the query table with inferred ranking preferences. A SPARQL Formulator then acts on semantics of tables in our KG. Since KG queries operate at these inputs one by one in the prescribed order, expandthe granularity of low-level subgraph triple patterns, we ing them into actual triple patterns, nested sub-queries, elevate scientific tables and relational-style analysis oper- and clauses (FILTER, HAVING, OPTIONAL, etc.) as reations on these tables to first-class citizens in our KG. We quired, to produce a functional SPARQL query.
On-the-fly generation of tables or views by fusing the relevant portions of a list of existing tables brings additional complex challenges to knowledge-based analysis, e.g., de- 4. Conclusions termining if a pair of candidate tables are compatible for merge operations based on their inferred semantics, de- Supporting tabular data discovery over collections of tatermining the order in which a list of candidate tables bles published in scientific documents presents unique must be merged to produce an optimal tabular response, challenges that require use of the explicit meaning of assessing the order in which diferent kinds of merge scientific tables as inferred by linking them to reference operations (e.g., union, join, cell-expand) need to be ap- knowledge graphs. We described our approach and implied on the tables, etc. Moreover, the context needed to plementation of a novel engine that can discover tabular efectively merge content across tables may come from data from knowledge graphs of scientific tables with early other sources like reliability assessments captured in our support for automatically merging content from multiple KG. Knowledge derived from other modalities such as tables on the fly via union operations. While preliminary, text may act as a ‘bridge’ between a pair of tables where we believe our approach is foundational and can be highly compatibility may not be established based on semantics efective when expanded to cover on-the-fly cell-based of table content alone. table expansion techniques. We believe this work will
To demonstrate on-the-fly table generation capability, motivate new research directions in knowledge-guided we implemented a preliminary dynamic approach to fuse scientific table generation and analysis at large scales. content of relevant tables via union operations. Ours is a heuristic-based approach that greedily seeks to maximize the amount of populated cells in the result table via new Acknowledgments rows while satisfying provenance-based reliability constraints. Our engine breaks up a list of initial candidate tables into groups where tables in each group have identical header-cell set overlap with the query table. It then performs a union of tables within each group to create super-tables. If there is partial set overlap across supertables, it performs a union across groups in decreasing order of their set overlap sizes. Finally, these merged
Ofice of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via [2021-21022600004]. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the oficial policies, either expressed or implied, of ODNI, IARPA, or the U.S. Government. [14] N. Abdelmageed, J. Chen, V. Cutrona, V. Efthymiou,
O. Hassanzadeh, M. Hulsebos, E. Jiménez-Ruiz, J. Se[1] M. Cafarella, A. Halevy, H. Lee, J. Madhavan, C. Yu, queda, K. Srinivas, Results of SemTab 2022, volD. Z. Wang, E. Wu, Ten Years of Webtables, Proc. ume 3320 of CEUR Workshop Proceedings, 2022. URL: VLDB Endow. 11 (2018) 2140–2149. doi:10.14778/ https://ceur-ws.org/Vol-3320/paper0.pdf. 3229863.3240492. [15] S. Zhang, K. Balog, On-the-fly Table Generation, in: [2] S. Zhang, K. Balog, Web Table Extraction, Retrieval, The 41st International ACM SIGIR Conference on and Augmentation: A Survey, ACM Trans. Intell. Research & Development in Information Retrieval, Syst. Technol. 11 (2020). doi:10.1145/3372117. SIGIR ’18, Association for Computing Machinery, [3] A. Chapman, E. Simperl, L. Koesten, G. Konstan- New York, NY, USA, 2018, p. 595–604. doi:10.1145/ tinidis, L.-D. Ibáñez, E. Kacprzak, P. Groth, Dataset 3209978.3209988.
Search: A Survey, The VLDB Journal 29 (2019) [16] Y. Gong, Z. Zhu, S. Galhotra, R. C. Fernan251–272. doi:10.1007/s00778-019-00564-x. dez, Ver: View Discovery in the Wild, 2022. [4] D. Brickley, M. Burgess, N. Noy, Google Dataset arXiv:2106.01543.
Search: Building a Search Engine for Datasets in [17] F. Årup Nielsen, Wembedder: Wikidata entity eman Open Web Ecosystem, in: The World Wide Web bedding web service, 2017. arXiv:1710.04099. Conference, WWW ’19, Association for Computing Machinery, New York, NY, USA, 2019, p. 1365–1375.
doi:10.1145/3308558.3313685. [5] Elicit: The AI Research Assistant, Ought, 2022.
https:elicit.org. [6] REASON - Rapid Explanation, Analysis and Sourcing ONline, IARPA, 2023. https://www.iarpa.gov/ research-programs/reason. [7] PMC Open Access Subset, National Library of
Medicine, Bethesda, MD, 2003. https://www.ncbi.
nlm.nih.gov/pmc/tools/openftlist/. [8] V. Mulwad, V. S. Kumar, J. Weisenberg Williams,
T. Finin, S. Dixit, A. Joshi, Towards Semantic Exploration of Tables in Scientific Documents, in: ESWC 2023 Workshops and Tutorials Joint Proceedings, volume 3443 of CEUR Workshop Proceedings, 2023. URL: https://ceur-ws.org/Vol-3443/ESWC_ 2023_SemTech4STLD_paper_2.pdf. [9] V. Mulwad, T. Finin, V. S. Kumar, J. Weisenberg Williams, S. Dixit, A. Joshi, A Practical Entity Linking System for Tables in Scientific Literature, in: Proceedings of the Workshop on Scientific Document Understanding (SDU 2023), colocated with 37th AAAI Conference on Artificial
Inteligence (AAAI), 2023. [10] Army Techniques Publication (ATP) 2-33.4. Intelligence Analysis, 2020. URL: https://irp.fas.org/ doddir/army/atp2-33-4.pdf. [11] U. Khurana, K. Srinivas, S. Galhotra, H. Samulowitz,
A Vision for Semantically Enriched Data Science, 2023. arXiv:2303.01378. [12] D. Vrandečić, M. Krötzsch, Wikidata: A Free Collaborative Knowledgebase, Commun. ACM 57 (2014) 78–85. doi:10.1145/2629489. [13] The New England Journal of Medicine - Advanced
Search, NEJM, 2023. https://www.nejm.org/ search?searchType=advancedSearch&allWords= covid19&searchWithin=fullText&objectType= nejm-media&mediaType=Table.