Interacting with Large Language Models (LLMs) via declarative queries is increasingly popular for tasks like question answering and data extraction, thanks to their ability to process vast unstructured data. However, LLMs often struggle with answering complex factual questions, exhibiting low precision and recall in the returned data. This challenge highlights that executing queries on LLMs remains a largely unexplored domain, where traditional data processing assumptions often fall short. Conventional query optimization, typically cost-driven, overlooks LLM-specific quality challenges such as contextual understanding. Just as new physical operators are designed to address the unique characteristics of LLMs, optimization must consider these quality challenges. Our results highlight that adhering strictly to conventional query optimization principles fails to generate the best plans in terms of result quality. To tackle this challenge, we present a novel approach to enhance SQL results by applying query optimization techniques specifically adapted for LLMs. We introduce a database system, GALOIS, that sits between the query and the LLM, effectively using the latter as a storage layer. We design alternative physical operators tailored for LLM-based query execution and adapt traditional optimization strategies to this novel context. For example, while pushing down operators in the query plan reduces execution cost (fewer calls to the model), it might complicate the call to the LLM and deteriorate result quality. Additionally, these models lack a traditional catalog for optimization, leading us to develop methods to dynamically gather such metadata during query execution. Our solution is compatible with any LLM and balances the trade-off between query result quality and execution cost. Experiments show up to 144% quality improvement over questions in Natural Language and 29% over direct SQL execution, highlighting the advantages of integrating database solutions with LLMs.

Tabular Natural Language Inference (TNLI) involves machine learning models that assess whether structured tabular data supports or contradicts a hypothesis formulated in natural language. TNLI models typically require large sets of training examples, which are costly to produce manually. In this demonstration, we present Tenet, a system for the automatic generation of training examples for TNLI applications. Existing TNLI training approaches either depend on costly human annotation or generate simplistic examples that lack data diversity and complex reasoning. In contrast, Tenet can start from a small set of manually annotated examples to automatically generate a large and diverse training dataset. Tenet is based on the idea that SQL queries are the right tool for obtaining rich and complex generated examples. To ensure data variety, evidence-queries extract cell values from tables based on diverse data patterns. Once the relevant data are identified, semantic queries define different ways to interpret it using SQL clauses. These interpretations are then verbalized as text to create annotated examples for TNLI. This demonstration offers an interactive experience where users will be able to select evidence from tabular data, inspect and refine generated queries, and observe how Tenet transforms structured data into natural language hypotheses. By engaging with different scenarios, users will see how Tenet enables the rapid creation of high-quality TNLI datasets, leading to inference models with performance comparable to those trained on manually crafted examples.

International audience

Several applications, such as text-to-SQL and computational fact checking, exploit the relationship between relational data and natural language text. However, state of the art solutions simply fail in managing "data-ambiguity", i.e., the case when there are multiple interpretations of the relationship between text and data. Given the ambiguity in language, text can be mapped to different subsets of data, but existing training corpora only have examples in which every sentence/question is annotated precisely w.r.t. the relation. This unrealistic assumption leaves the target applications unable to handle ambiguous cases. To tackle this problem, we present an end-to-end solution that, given a table D, generates examples that consist of text, annotated with its data evidence, with factual ambiguities w.r.t. D. We formulate the problem of profiling relational tables to identify row and attribute data ambiguity. For the latter, we propose a deep learning method that identifies every pair of data ambiguous attributes and a label that describes both columns. Such metadata is then used to generate examples with data ambiguities for any input table. To enable scalability, we finally introduce a SQL approach that can generate millions of examples in seconds. We show the high accuracy of our solution in profiling relational tables and report on how our automatically generated examples lead to drastic quality improvements in two fact-checking applications, including a website with thousands of users, and in a text-to-SQL system.

Tabular data is becoming increasingly important in Natural Language Processing (NLP) tasks, such as Tabular Natural Language Inference (TNLI). Given a table and a hypothesis expressed in NL text, the goal is to assess if the former structured data supports or refutes the latter. In this work, we focus on the role played by the annotated data in training the inference model. We introduce a system, Tenet, for the automatic augmentation and generation of training examples for TNLI. Given the tables, existing approaches are either based on human annotators, and thus expensive, or on methods that produce simple examples that lack data variety and complex reasoning. Instead, our approach is built around the intuition that SQL queries are the right tool to achieve variety in the generated examples, both in terms of data variety and reasoning complexity. The first is achieved by evidence-queries that identify cell values over tables according to different data patterns. Once the data for the example is identified, semantic-queries describe the different ways such data can be identified with standard SQL clauses. These rich descriptions are then verbalized as text to create the annotated examples for the TNLI task. The same approach is also extended to create counterfactual examples, i.e., examples where the hypothesis is false, with a method based on injecting errors in the original (clean) table. For all steps, we introduce generic generation algorithms that take as input only the tables. For our experimental study, we use three datasets from the TNLI literature and two crafted by us on more complex tables. Tenet generates human-like examples, which lead to the effective training of several inference models with results comparable to those obtained by training the same models with manually-written examples.

Applications such as computational fact checking and data-to-text generation exploit the relationship between relational data and natural language text. Despite promising results in these areas, state of the art solutions simply fail in managing "data-ambiguity", i.e., the case when there are multiple interpretations of the relationship between the textual sentence and the relational data. To tackle this problem, we introduce Pythia, a system that, given a relational table D, generates textual sentences that contain factual ambiguities w.r.t. the data in D. Such sentences can then be used to train target applications in handling data-ambiguity. In this demonstration, we first show how our system generates data ambiguous sentences for a given table in an unsupervised fashion by data profiling and query generation. We then demonstrate how two existing applications benefit from Pythia's generated sentences, improving the state-of-the-art results. The audience will interact with Pythia by changing input parameters in an interactive fashion, including the upload of their own dataset to see what data ambiguous sentences are generated for it.

We present Falcon, an interactive, deterministic, and declarative data cleaning system, which uses SQL update queries as the language to repair data. Falcon does not rely on the existence of a set of pre-defined data quality rules. On the contrary, it encourages users to explore the data, identify possible problems, and make updates to fix them. Bootstrapped by one user update, Falcon guesses a set of possible sql update queries that can be used to repair the data. The main technical challenge addressed in this paper consists in finding a set of sql update queries that is minimal in size and at the same time fixes the largest number of errors in the data. We formalize this problem as a search in a lattice-shaped space. To guarantee that the chosen updates are semantically correct, Falcon navigates the lattice by interacting with users to gradually validate the set of sql update queries. Besides using traditional one-hop based traverse algorithms (e.g., BFS or DFS), we describe novel multi-hop search algorithms such that Falcon can dive over the lattice and conduct the search efficiently. Our novel search strategy is coupled with a number of optimization techniques to further prune the search space and efficiently maintain the lattice. We have conducted extensive experiments using both real-world and synthetic datasets to show that Falcon can effectively communicate with users in data repairing.