14dtj / cocoqa

An auto QA system about code convention and styles based on knowledge base

CocoQa

An auto QA system about code convention and styles based on knowledge base

Ontology & RDF

We construct a knowledge base called CCBase in the field of coding convention. Ontology folder contains the ontology of CCBase, including subject types, relations and so on, which is in the form of owl. RDF folder contains some entities of our CCBase.

Plugin

We developed an Intellij Idea plugin. You can download plugin.zip and install it by the instructions of Idea.

Introduction

CocoQa is a QA system for coding conventions. CocoQa combines the advantages of the KB- and the corpus-based QA systems: it stores in a knowledge base coding conventions collected from online resources; it builds CCBase, a knowledge graph introduced in [3], for modeling coding conventions; and it answers questions about coding conventions by querying the graph.

As the above figure shows, the CocoQa tool consists of four components: (1) a coding convention knowledge graph constructed from online resources, (2) a subgraph matcher that understands natural language questions and performs SPARQL queries over the knowledge graph, (3) a machine comprehenser that employs a deep neural network model to answer questions. The comprehenser searches answers from all textual paragraphs attached to the knowledge graph, and (4) a ranker that ranks answers retrieved by the subgraph matcher and the machine comprehenser via a logistic regression classifier.

The subgraph matcher generates answers to a question by querying entities and their relations in CCBase. Given a question, it casts the question into a SPARQL query followed by querying CCBase. The procedure and a running example is available here.

We also integrate an end-to-end machine comprehension approach that applies a deep neural network to answer the query on textual paragraphs attached in CCBase. The introduction and procedure to run machine comprehension QA could be found here . A logistic regression classifier is trained to merge and rank the answers, and retrieve the top ones.

Install

It is a web application and we developed a graphical interface. After installing the following python packages, you can run FLASK server and experience CocoQa.
Dependencies
First of all, you should download Stanford CoreNLP toolkit and run the server.

cd stanford-corenlp-full-${version}
java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 15000

Then, Stanford CoreNLP server will listen on port 9000 for requests.

After that, you should run MRCQA and Ranker. At last, you can start to run CCQA.

Now, you can visit localhost:$port/ordinary to experience our system!