SemMemDB: In-Database Knowledge Activation

Semantic memory techniques are currently facing significant scalability challenges due to the growth of problem complexity and data size. In this paper, we present SemMemDB, an in-database module for semantic memories. This module defines a novel relational model for semantic memories and an efficient SQL-based spreading activation algorithm. We provide an intuitive, easy-to-use API for the users to run retrieval tasks. SemMemDB is the first in-database query engine for semantic memories. The benefits of our in-database approach includes: 1) Databases have mature query engine and optimizer that generates efficient execution plans; 2) Databases provide massive storage that supports large amount of data; 3) SQL is one of the industry standards used widely in various systems and frameworks. We provide extensive evaluation on DBPedia, a web-scale knowledge base constructed from Wikipedia articles. Experiment results show our system runs over 500 times faster than previous works.

IMPORTANT: SemMemDB is still early release and under active research.

###Prerequisites

• Linux
• Python 2.7
• PostgreSQL 9.2

###Installation Guide Project directories:

- PROJECT_DIR
+- data
+- csv
+- example
+- paper
+- scripts
+- sql

Installation steps:

• Download the DBPedia data into data directory.
• Run PostgreSQL server, create a database, say semmemdb:

$ postgres -D ${DATA_PATH} &
$ createdb semmemdb

• In example/setup.sh, fill in the project path and database name:

PROJECT_DIR=/path/to/project
DB_NAME=semmemdb

• Run setup.sh:

$ ${PROJECT_DIR}/example/setup.sh

This completes the installation.

To run an example, login into PostgreSQL, populate the query table, and call activate() procedure:

$ psql semmemdb
psql (9.2.3)
Type "help" for help.

semmemdb=# SELECT * FROM semmemdb.activate() LIMIT 10;
 node |         s          
------+--------------------
    2 | -0.363317136968319
    3 | -0.594135148094781
    4 | -0.594135148094781
    5 | -0.594135148094781
    6 | -0.594135148094781
    7 | -0.594135148094781
    8 | -0.594135148094781
    9 | -0.594135148094781
   10 | -0.594135148094781
   11 | -0.594135148094781

The name of nodes are stored in the semmemdb.entities table.