[Paper Review] Cognitive Database: A Step towards Endowing Relational Databases with Artificial Intelligence Capabilities

Rajesh Bordawekar, Bortik Bandyopadhyay+1 more|arXiv (Cornell University)|Dec 19, 2017

Semantic Web and Ontologies52 references19 citations

TL;DR

This paper introduces Cognitive Databases, a novel approach that leverages unsupervised word embedding from database-derived text to endow relational databases with AI capabilities. By transforming structured data into semantic vectors, it enables a new class of SQL-based cognitive intelligence (CI) queries that support semantic matching, inductive reasoning, and external knowledge integration—demonstrated via an Apache Spark prototype over multi-modal data.

ABSTRACT

We propose Cognitive Databases, an approach for transparently enabling Artificial Intelligence (AI) capabilities in relational databases. A novel aspect of our design is to first view the structured data source as meaningful unstructured text, and then use the text to build an unsupervised neural network model using a Natural Language Processing (NLP) technique called word embedding. This model captures the hidden inter-/intra-column relationships between database tokens of different types. For each database token, the model includes a vector that encodes contextual semantic relationships. We seamlessly integrate the word embedding model into existing SQL query infrastructure and use it to enable a new class of SQL-based analytics queries called cognitive intelligence (CI) queries. CI queries use the model vectors to enable complex queries such as semantic matching, inductive reasoning queries such as analogies, predictive queries using entities not present in a database, and, more generally, using knowledge from external sources. We demonstrate unique capabilities of Cognitive Databases using an Apache Spark based prototype to execute inductive reasoning CI queries over a multi-modal database containing text and images. We believe our first-of-a-kind system exemplifies using AI functionality to endow relational databases with capabilities that were previously very hard to realize in practice.

Motivation & Objective

• To address the limitation of traditional SQL in capturing latent semantic relationships across diverse data types in relational databases.
• To enable AI-driven querying capabilities without relying on external knowledge sources like WordNet or thesauri.
• To provide a holistic, context-aware view of relational data by encoding semantic relationships through unsupervised NLP techniques.
• To extend relational databases with a new class of analytics queries—cognitive intelligence (CI) queries—capable of inductive reasoning and semantic similarity.
• To demonstrate the feasibility and expressiveness of CI queries using a multi-modal database (text and images) via an Apache Spark prototype.

Proposed method

• Treat database schema and instance data (including text, numbers, dates, and images) as meaningful unstructured text for NLP processing.
• Apply word embedding techniques (e.g., Skip-gram or CBOW) to generate dense vector representations that encode inter- and intra-column semantic relationships.
• Integrate the trained word embedding model directly into the SQL query execution engine to enable vector-based semantic operations.
• Support CI queries that use vector similarity for semantic matching, analogical reasoning, and predictive inference over unseen entities.
• Enable use of external knowledge bases (e.g., Wikipedia) by aligning their embeddings with the database’s learned vectors.
• Implement the system on Apache Spark to support scalable training and inference over large, multi-modal databases.

Experimental results

Research questions

• RQ1 Can latent semantic relationships in relational databases be effectively captured using unsupervised NLP techniques like word embeddings?
• RQ2 To what extent can word embeddings derived from database content enable new forms of SQL-based analytics beyond traditional value-based predicates?
• RQ3 How can semantic vectors be used to support inductive reasoning tasks such as analogical queries and predictions for entities not present in the database?
• RQ4 Can external knowledge sources be seamlessly integrated with database-derived embeddings to enhance query expressiveness?
• RQ5 What are the performance and scalability characteristics of integrating learned semantic vectors into a standard SQL query engine?

Key findings

• The proposed Cognitive Database system successfully enables semantic matching and inductive reasoning via CI queries, allowing users to find employees 'most related' to a given individual like John Dolittle without explicit schema or join logic.
• The system supports predictive queries over entities not present in the database, such as identifying dangerous vacation packages using external knowledge about terms like 'accident' or 'death'.
• The prototype demonstrates the feasibility of executing complex CI queries on multi-modal data (text and images) using Apache Spark, showing that semantic reasoning can be natively integrated into SQL-based systems.
• The approach does not rely on external lexical resources such as WordNet or thesauri, making it self-contained and adaptable to domain-specific data.
• The system enables a dual view of data—both relational and semantic—by representing database tokens as vectors that encode contextual meaning.
• The authors report that the model captures both syntactic and semantic relationships across diverse data types, including numerical values, strings, dates, and images, through a unified embedding space.