Search for: All records

With the emerging advancements of AI, validating data generated by AI models becomes a key challenge. In this work, we tackle the problem of validating tabular data generated by large language models (LLMs). By leveraging a recently proposed technique called Gen-T, we present a technique to verify if the data in the LLM table can be reclaimed (reproduced) using tables available in a given data lake (for example, tables used to train the LLM). Specically, we measure the number of data lake tables that support tuples (or partial tuples) in a generated table. We further provide suggestions for value replacements if a generated value cannot be reclaimed. Using this approach, users can evaluate their LLM-generated tables, consider potential modications for table values, and gauge how much support the modied table has from the data lake. We discuss two case studies showing that table values annotated with reclama- tion support scores, along with possible value replacements, can help users assess the trustworthiness of LLM-generated tables. 

We introduce the problem of Table Reclamation. Given a Source Table and a large table repository, reclamation finds a set of tables that, when integrated, reproduce the source table as closely as possible. Unlike query discovery problems like Query-by-Example or by-Target, Table Reclamation focuses on reclaiming the data in the Source Table as fully as possible using real tables that may be incomplete or inconsistent. To do this, we define a new measure of table similarity, called error-aware instance similarity, to measure how close a reclaimed table is to a Source Table, a measure grounded in instance similarity used in data exchange. Our search covers not only SELECT-PROJECT- JOIN queries, but integration queries with unions, outerjoins, and the unary operators subsumption and complementation that have been shown to be important in data integration and fusion. Using reclamation, a data scientist can understand if any tables in a repository can be used to exactly reclaim a tuple in the Source. If not, one can understand if this is due to differences in values or to incompleteness in the data. Our solution, Gen- T, performs table discovery to retrieve a set of candidate tables from the table repository, filters these down to a set of originating tables, then integrates these tables to reclaim the Source as closely as possible. We show that our solution, while approximate, is accurate, efficient and scalable in the size of the table repository with experiments on real data lakes containing up to 15K tables, where the average number of tuples varies from small (web tables) to extremely large (open data tables) up to 1M tuples. 

Existing techniques for unionable table search define unionability using metadata (tables must have the same or similar schemas) or column-based metrics (for example, the values in a table should be drawn from the same domain). In this work, we introduce the use of semantic relationships between pairs of columns in a table to improve the accuracy of the union search. Consequently, we introduce a new notion of unionability that considers relationships between columns, together with the semantics of columns, in a principled way. To do so, we present two new methods to discover the semantic relationships between pairs of columns. The first uses an existing knowledge base (KB), and the second (which we call a "synthesized KB") uses knowledge from the data lake itself. We adopt an existing Table Union Search benchmark and present new (open) benchmarks that represent small and large real data lakes. We show that our new unionability search algorithm, called SANTOS, outperforms a state-of-the-art union search that uses a wide variety of column-based semantics, including word embeddings and regular expressions. We show empirically that our synthesized KB improves the accuracy of union search by representing relationship semantics that may not be contained in an available KB. This result hints at a promising future of creating synthesized KBs from data lakes with limited KB coverage and using them for union search. 

Dataset discovery from data lakes is essential in many real application scenarios. In this paper, we propose Starmie, an end-to-end framework for dataset discovery from data lakes (with table union search as the main use case). Our proposed framework features a contrastive learning method to train column encoders from pre-trained language models in a fully unsupervised manner. The column encoder of Starmie captures the rich contextual semantic information within tables by leveraging a contrastive multi-column pre-training strategy. We utilize the cosine similarity between column embedding vectors as the column unionability score and propose a filter-and-verification framework that allows exploring a variety of design choices to compute the unionability score between two tables accordingly. Empirical results on real table benchmarks show that Starmie outperforms the best-known solutions in the effectiveness of table union search by 6.8 in MAP and recall. Moreover, Starmie is the first to employ the HNSW (Hierarchical Navigable Small World) index to accelerate query processing of table union search which provides a 3,000X performance gain over the linear scan baseline and a 400X performance gain over an LSH index (the state-of-the-art solution for data lake indexing).