Skip to content

Latest commit

 

History

History
81 lines (56 loc) · 4.35 KB

README.md

File metadata and controls

81 lines (56 loc) · 4.35 KB
Raw

NeoDUDES - A Compositional Question Answering System Using DUDES

Results

The results and other files generated during the experiments from the paper can be found in results/. For further details please see the README.md there.

Setup

The easiest way to run the NeoDUDES question answering system is using Docker. Either run

docker pull neodudes:latest

on your machine or build the image yourself:

docker build . -t neodudes

After that, you can run the container as follows:

docker run -e DBPEDIA_SPOTLIGHT_ENDPOINT='http://172.17.0.1:2222/rest' -e DBPEDIA_ENDPOINT='http://172.17.0.1:8890/sparql' -it neodudes

The container expects two environment variables to be set:

  • DBPEDIA_SPOTLIGHT_ENDPOINT: URL of a running DBpedia Spotlight https://www.dbpedia-spotlight.org/ instance accessible for the Docker container
  • DBPEDIA_ENDPOINT: URL of a SPARQL endpoint serving the necessary triples for the benchmark.

You can start a DBpedia Spotlight instance for English locally in the background using the following command:

docker run -tid --restart unless-stopped --name dbpedia-spotlight.en --mount source=spotlight-model,target=/opt/spotlight -p 2222:80 dbpedia/dbpedia-spotlight spotlight.sh en

For the SPARQL endpoint, in case of QALD-9, the triples used by our approach can be found here: https://ag-sc.techfak.uni-bielefeld.de/download/dudes/2016.ttl.zst. To serve them, you can use, e.g., Virtuoso for which you can find a prepared data directory here: https://ag-sc.techfak.uni-bielefeld.de/download/dudes/virtuoso2016.tar.zst

Please note that in case of local instances you need to take care that they are accessible from the container, e.g. by finding out the corresponding IP addresses or using some kind of docker-compose setup. URLs like 'http://localhost:8890/sparql' will likely not work (see https://stackoverflow.com/questions/24319662/from-inside-of-a-docker-container-how-do-i-connect-to-the-localhost-of-the-mach).

In case you want to setup the project locally, the Dockerfile and requirements.txt might give you a few hints which packages you need to install.

Benchmark

Inside the container, there are a few prepared scripts which you can run:

  • make_docs.sh to generate the Sphinx documentation of the project
  • qald-eval-test.sh and 'qald-eval-train.sh' for running the QALD-9 train or test benchmark, respectively
  • qald-rpc.sh for starting the DUDES RPC server dealing with tagging input questions and scoring SPARQL queries with the provided models
  • src/llm/prompting.py for experiments with GPT - for this make sure to provide valid OpenAI organization and API key in the .env file. An example evnironment file can be found in sample.env
  • src/llm/query_score_training.py for training query scoring models - a Slurm job array is given with query_score_train.sarray and query_score_train_best.sarray, also illustrating relevant command line options
  • qald-eval-newpipeline.py for running the QALD-9 benchmark, using the files src/lemon/resources/qald/QALD9_train-dataset-raw.csv and src/lemon/resources/qald/QALD9_test-dataset-raw.csv

Running the benchmark can be done using qald-eval-newpipeline.py, which by default launches four processes running in parralel. To change the number of spawned processes, you need to change the cpu_count variable.

Citation

Please consider citing our work if you find the provided resources useful:

@InProceedings{schmidt-etal-2025-neodudes,
    author="Schmidt, David Maria
    and Elahi, Mohammad Fazleh
    and Cimiano, Philipp",
    editor="Alam, Mehwish
    and Rospocher, Marco
    and van Erp, Marieke
    and Hollink, Laura
    and Gesese, Genet Asefa",
    title="Lexicalization Is All You Need: Examining the Impact of Lexical Knowledge in a Compositional QALD System",
    booktitle="Knowledge Engineering and Knowledge Management",
    year="2025",
    publisher="Springer Nature Switzerland",
    address="Cham",
    pages="102--122",
    isbn="978-3-031-77792-9"
}