Swanton Pacific Ranch Chatbot - CSC580

This repository contains minimal Knowledge Graph Question Answering system to handle Swanton Pacific Ranch related questions and ParAI model to handle conversational questions.

Knowledge Graph Database: Ontotext GraphDB

Currently, we are using a free version of the GraphDB. Therefore, it's not hosted on the server.

Dependencies

1. rdf4j library (3.7.3) to add or delete data from the database.
2. Maven as a build and dependency management tool.
   • Maven works really well with Intellij. So, I recommend to use Intellij for your IDE.
3. ParlAI should be installed via pip.

Usage

To update the data in the knowledge graph with our wrapper classes,

1. Export graph as a Turtle file.
2. Pass its filename to KGModelWrapper as an argument, make changes to the wrapper, and write the model to a Turtle file. For example,

   KGModelWrapper kgModelWrapper = KGModelWrapper.createInstance(nameSpace, prefix, filename);
   kgModelWrapper.addStatementToModel("Location", RDF.TYPE, OWL.CLASS);
   kgModelWrapper.writeModel("testStatements.ttl");

3. Import the output Turtle file to the knowledge graph using its workbench.

To execute a query in the knowledge graph with our wrapper classes,

1. Create a Swanton knowledge graph instance with a model from KgModelWrapper.

Model model = kgModelWrapper.getKgModel();
SwantonKnowledgeGraph swantonKg = SwantonKnowledgeGraph.createKnowledgeGraph(model);

2. Write a query and execute the query.

queryString = "PREFIX spr:<http://swantonpacificranch.org/> \n";
queryString += "SELECT ?name WHERE { \n";
queryString += "\tspr:Swanton_Pacific_Ranch spr:owned_by ?name";
queryString += "}";

swantonKg.queryStatements(queryString);

3. queryStatements() return list of BindingSet. Each BindingSet should have a value with the name used in the executed query. In the above case, name.

List<BindingSet> result = swantonKg.queryStatements(queryString);

for (BindingSet solution: result) {
   System.out.println(solution.getValue("name"));
}

TODO

Most of our future work belongs to the Knowledge Graph Question Answering system.

• Better design the knowledge graph schema with a class hierarchy such as “Swanton Pacific Ranch is a subclass of location” so that we can see the relationship between nodes with graph visualization.
• Add more rules to the KG extraction system to turn the massive amount of data from documents into KG triplets more accurately.
• Come up with a logical way to handle complex questions that require multiple hops in the knowledge graph. Since it’s hard to automatically generate queries with multi-hops, we might need to use QA pairs as an alternative to handle those questions.
• Gather more question - query pairs to either use them as a template or train a neural semantic parser.
• make custom ParlAI model recognized globally