Codes for ISWC 2021 research track paper "EDG-based Question Decomposition for Complex Question Answering over Knowledge Bases".

EDGQA is a QA system over knowledge bases based on Entity-Description Graphs (EDGs). Currently EDGQA has been implemented for DBpedia, tested on LC-QuAD 1.0 and QALD-9.

The above figure shows SPARQL (a. and c.) and EDG (b. and d.) on two exemplar natural language questions. The dashed line connects a description to an intermediate entity. The types of nodes and edges of EDG is defined as follows.

By generating such EDGs, questions are represented as a combination of entities and their description, providing a structure for understanding and answering complex questions. More information are detailed in our paper.

• JDK 1.8.0
• Maven
• python 3.6

Knowledge base dumps and linking systems are also needed.

In EDGQA, DBpedia 1604 (for lcquad) and DBpedia 1610 (for qald-9) is stored in Virtuoso.

You can deploy the dbpedia locally or use the online endpoint (can be incompatible to the datasets). Then fill the server address and port in src/main/java/cn/edu/nju/ws/edgqa/utils/kbutil/KBUtil.java.

• DBpedia 1604
• Virtuoso

Earl, Falcon, Dexter are used in EDGQA.

See directory linking_tools, and follow the instructions to set up the three linking systems. Then fill in the server address and port in src/main/java/cn/edu/nju/ws/edgqa/utils/linking/LinkingTool.java.

For more information:

• Dexter
• EARL
• Falcon

EDGQA employs bert-based classifier as semantic matching models for relation detection and query reranking.

See directory models to deploy the models correctly.

Program arguments are defined in src/main/java/cn/edu/nju/ws/edgqa/main/QAArgs.java.

Running settings for Intellij IDEA 2019.3 above versions are stored in EDGQA/.run.

Run src/main/java/cn/edu/nju/ws/edgqa/main/EDGQA.java by following CLI arguments:

-d --dataset: 'lc-quad', 'qald-9'
-tr --train: 'true' for training set, 'false' for test set
-r --run: 'autotest', 'single', or 'serial_number'
-uc --use_cache: 'true' for using linking cache, 'false' otherwise
-cc --create_cache: 'true' for creating linking cache, 'false' otherwise
-gll --global_linking: 'true' for using global linking, 'false' otherwise
-lll --local_linking: 'true' for using local linking, 'false' otherwise
-qd --question_decomposition: 'true' for using EDG to decompose the question
-rr --reranking: 'true' for re-ranking by EDG block, 'false' otherwise

Because the linking tools consume a lot of time, caching the linking results of the test queries helps improve the speed of the test. The cache needs to be built the first time the QA is run and is available when it is run again. Use the arguments use_cache and create_cache above to set the cache tool.

[Optional] The cache of linking results is available on this Google Drive dir. Please set it to the current directory, i.e., EDGQA/cache. It is optional for running EDGQA.

Run cn/edu/nju/ws/edgqa/handler/PointerNetworkQA.java by following CLI arguments:

--dataset: 'lc-quad', 'qald-9'

• PKUmod paraphrase dict
• QALD-9
• BERT by Google Research
• java-string-similarity

Feel free to create a GitHub Issue or send an e-mail. We look forward to receiving your feedback.