SkeletonKBQA: Skeleton Parsing for Complex Question Answering over Knowledge Bases

Codes for a journal paper: "Skeleton Parsing for Complex Question Answering over Knowledge Bases" . If you meet any questions, please email to him (ywsun at smail.nju.edu.cn).

Project Structure:

File	Description
kbcqa	Codes of skeleton-based SP and IR approaches
skeletons	Skeleton Bank from three complex KBQA datasets

Skeleton Bank

We annotate and publish a skeleton bank of 15,166 questions from three KBQA datasets.

The skeleton bank is json format. An example:

{
"question": "People from the country with the capital Brussels speak what languages ?",
"skeleton": [
	{
		"question": "People from the country with the capital Brussels speak what languages ?",
		"text_span": "with the capital Brussels",
		"headword_index": 3,
		"attachment_relation": "nmod"
	},
	{
		"question": "People from the country speak what languages ?",
		"text_span": "from the country",
		"headword_index": 0,
		"attachment_relation": "nmod"
	}
]
}

Note that we will explain how to run the codes of kbcqa file below.

Requirements

• requirements.txt

Configuration

The cofiguration of SkeletonKBQA is in kbcqa/common/globals_args.py.

• root: root of all resources and datasets, default ../dataset.
• q_mode: a specific KBQA dataset: lcquad, graphq, and cwq.
• sutime: jar files path of SUTime Java library tool.
• corenlp_ip_port: ip port of Stanford CoreNLP server.
• dbpedia_pyodbc: odbc of DBpedia virtuoso server.
• dbpedia_sparql_html: web site of DBpedia virtuoso server.
• freebase_pyodbc: odbc of Freebase virtuoso server.
• freebase_sparql_html: web site of Freebase virtuoso server.

Common Resources

The zip file from google drive contains three parts:

• Stanford CoreNLP server
• SUTime Java library
• BERT pre-trained Models

Note that download, unzip the zip file, and then copy it to the root folder.

Knowledge Bases

• DBpedia 201604 version for LC-QuAD 1.0
• Freebase 2013 version for GraphQuestions
• Freebase latest version for ComplexWebQuestions 1.1

Note that download a virtuoso server and load the above KBs.

You only need to load a specific KB which is correspond to your KBQA dataset.

LC-QuAD 1.0 Resources

The zip file from google drive contains three parts:

• LC-QuAD 1.0 datasets
• Its skeleton parsing models
• Its corresponding KB entity-related Lexicons

Note that download, unzip the zip file, and then copy it to the root.

GraphQuestions Resources

The zip file from google drive contains three parts:

• GraphQuestions datasets
• Its skeleton parsing models
• Its corresponding KB entity-related Lexicons

Note that download, unzip the zip file, and then copy it to the root.

ComplexWebQuestions 1.1

The zip file from google drive contains three parts:

• ComplexWebQuestions 1.1 datasets
• Its skeleton parsing models
• Its corresponding KB entity-related Lexicons

Note that download, unzip the zip file, and then copy it to the root.

Run SkeletonKBQA

SkeletonKBQA contains two KBQA approaches: SSP and SIR.

• Skeleton-based semantic parsing approach (SSP) has four modules:
  • Ungrounded query generation
  • Entity linking
  • Candidate grounded query generation
  • Semantic matching

Note that the above four modules are correspond to the arguement module in kbcqa/method_sp/sp_pipeline.py.

Run the provided SSP scripts as:

bash run_ssp_LCQ.sh
bash run_ssp_GraphQ.sh
bash run_ssp_CWQ.sh

• Skeleton-based Information Retrieval approach (SIR) has three modules:
  • Node recogniztion and linking
  • Candidate grounded path generation
  • Semantic matching

Note that the above three modules are correspond to the arguement module in kbcqa/method_ir/ir_pipeline.py.

Run the provided SIR scripts as:

bash run_sir_LCQ.sh
bash run_sir_GraphQ.sh
bash run_sir_CWQ.sh

Contacts

If you have any difficulty or questions in running codes, reproducing experimental results, and skeleton parsing, please email to him (ywsun at smail.nju.edu.cn).