We present a dataset for graph-based question answering. The dataset consists of <question; candidate answer> pairs. For each candidate, we present a graph that is obtained by finding the shortest path between named entities mentioned in a question and a candidate answer. As a knowledge graph, we adopted Wikidata. Our dataset has the following fields:

• sample_id - an identifier for <question, candidate answer>;

• question - question text;

• questionEntity - comma-separated list of names (textual strings) for Wikidata concepts mentioned in a given question;

• answerEntity - a textual name of candidate answer (candidate is a concept from Wikidata) for the given question;

• groundTruthAnswerEntity - a textual name of ground truth answer (answer is a concept from Wikidata) for the given question;

• answerEntityId - a Wikidata id of candidate answer (see "answerEntity" column). Example: "Q2599";

• questionEntityId - a comma-separated list of Wikidata ids for concepts mentioned in a given question (list of ids for mentions from "questionEntity" column);

• groundTruthAnswerEntityId - a Wikidata id of ground truth answer (see "answerEntity" column). Example: "Q148234";

• correct - either "True" or "False". The field indicates whether a <question, answer candidate> is correct, i.e., candidate answer is a true answer to the given question;

• graph - a shortest-path graph for a given <question, candidate answer> pair. The graph is obtained by taking the shortest paths from all mentioned concepts ("questionEntityId" column) to a candidate answer("answerEntityId" column) in the knowledge graph of Wikidata. The graph is stored in "node-link" JSON format from NetworkX. You can import the graph using the node_link_graph.

For an example on how to draw shortest path graphs using NetworkX, please see the "Data vizualization" section. You can use a Wikidata id to access Wikidata concepts. E.g., Q148234 or Q51056.

Train dataset is available here. It can be used for initial experiments. Our dataset is also available at HuggingFace.

For your convenience, we provide a visualization script question-candidate graphs to help you better understand the graph structures. You can run the visualization script as follows:

python visualization/draw_random_question_graphs.py --input_tsv "data/tsv/train.tsv" \
--num_questions 10 \
--output_dir "question_graph_examples/"

Pre-vizualized graphs for all candidate answers of 10 random questions are available here.

We propose two BERT-based baselines:

• A BERT-based binary classifier which does not take graphs into account.

• A BERT-based binary classifier which trains on concatenated samples consisting of: (i) question, (ii) linearized graph of a candidate answer.

We see the task as a binary classification. Given a question, answer candidate, and a subgraph from knowledge grpah, your goal is assign a binary label to each pair.

For evaluation, we will adopt precision, recall, F1-measure, and accuracy. For your convenience, we publish an evaluation script. The evaluation script can be executed as follows:

python visualization/draw_random_question_graphs.py \
--predictions_path <path to tsv-file with predicted labels> \
--gold_labels_path "data/tsv/train.tsv"