Skip to content

Latest commit

 

History

History
144 lines (119 loc) · 8.48 KB

README.md

File metadata and controls

144 lines (119 loc) · 8.48 KB
Raw

Experiments for ISWC 2024

This README describes and details the experiments that were run for the ISWC 2024 submission.

5.1 Relevant Event Retrieval

For the search experiments, we created a main folder with the data in the root directory: data-test, with two sub-folders: dbpedia and wikidata. Each sub-folder had three folders: config, gs_events and referents.

All scripts are in this folder. All example commands are run from root directory of the repo.

  • Parameter selection for the search (subset of 12 events)
    • Python script: run_all_grid_search.py
    • Example command: 
      python experiments_run/run_all_grid_search.py -t <type-system> -e experiments_run/grid-search-events.csv
  • Main results for the search (all events)
    • Python script: run_all_search.py
    • Example command: 
      python experiments_run/run_all_search.py -t <type-system> -e experiments_run/all-search-events.csv

The following table shows information on events and their sub-events across datasets. All types of events are taken into account, which include mainly historical events, sports events and political events such as elections.

Dataset Nb. of sub-events = 1 Nb. of sub-events > 1 Nb. of sub-events > 10 Final
Wikidata 203,988 238,094 2,408 341
DBpedia 84,599 95,504 1,333 250
YAGO4 70,738 76,682 993 306

The following 12 events were used for parameter selection for the search:

  • World War I
  • American Indian Wars
  • Mediterranean and Middle East Theatre of World War II
  • French RevolutionCold War
  • European Theatre of World War II
  • Napoleonic Wars
  • Coalition Wars
  • European Theatre of World War I
  • Pacific War
  • Russian Civil War
  • Yugoslav Wars

5.2 Event-centric KG Population

The following table shows the labels that were used to retrieve information from triples. If a predicate contained any of the labels, it would add information in the graph based on its narrative dimension and SEM predicate. As an example, let us take an input triple (s, p, o). If the string of p contains the substring "person", then a triple (s, sem:hasActor, o) is added to the output graph.

Narrative dimension SEM predicate Labels
Who sem:hasActor person, combatant, commander, participant
When (begin) sem:hasBeginTimeStamp start time, date, point in time
When (end) sem:hasEndTimeStamp end time
Where sem:hasPlace place, location, country
Part of sem:subEventOf partof, part of
Part of (inverse) sem:hasSubEvent has part, significant event

For the event-centric KG generation experiments, we extracted the data that we needed from the search experiments, and created a new folder, data_ng_building.

  • Extracting data

    • Python script: get_data_ng_building.py. There are some parameters to change in PARAMS (start and end date of experiments, folder_gs if different)
    • Example command: 
      python experiments_run/get_data_ng_building.py
    • The data was saved in a new folder: data_ng_building

  • Event-centric KG population (EC-KG-P) from KG

    • This includes EC-KG-P from the output of graph search, EC-KG-P with our system from ground truth events, EC-KG-P from EventKG
    • Python script: build_ng_from_search.py
    • Example command: 
      python experiments_run/build_ng_from_search.py --folder data_ng_building/

  • Metrics for EC-KG-P from KG

    • Python script: get_metrics.py
    • Example command (comparing with all ground truth events): 
      python experiments_run/get_metrics.py --folder data_ng_building/ --output_name eventkg_vs_generation.json --graph_c_path generation_ng.ttl --graph_gs_path eventkg_ng.ttl
    • Example command (comparing with output of graph search): 
      python experiments_run/get_metrics.py --folder data_ng_building/ --output_name eventkg_vs_search.json --graph_c_path search_ng.ttl --graph_gs_path eventkg_ng.ttl

  • Aggregating results (as in the paper):

    • Python script: get_table_results.py
    • Example command (comparing with all ground truth events): 
      python experiments_run/get_table_results.py --folder data_ng_building/ --metric eventkg_vs_generation.json --label <label>
    • Example command (comparing with output of graph search): 
      python experiments_run/get_table_results.py --folder data_ng_building/ --metric eventkg_vs_search.json --label <label>

  • Event-centric KG Population from text

    • These are the experiments to generate KGs from the DBpedia abstracts
    • First you need to set up a local DBpedia, you can follow the steps on this link: https://github.com/MartinoMensio/spacy-dbpedia-spotlight
    • Python script: build_kg_with_frames.py
    • Example command: 
      python experiments_run/build_kg_with_frames.py --folder data_ng_building/dbpedia/

  • Annotating and analysing (causation) frames

    • Analysing
      • Python script: get_csv_analyse_frame.py
      • Example command 
        python experiments_run/get_csv_analyse_frame.py --folder_input data_ng_building/dbpedia/ --folder_output experiments_run/ng_analysis
    • Extracting causation frames for manual annotation
      • Python script: extract_causation_for_annot.py
      • Example command 
        python experiments_run/extract_causation_for_annot.py --csv experiments_run/ng_analysis/df_causation.csv --folder experiments_run/ng_analysis/

Furthermore, the manually annotated causation frames can be found in the experiments_run/annotated folder.

The following table shows the F1, Precision and Recall scores of the narrative graphs generated from the output of the search algorithm. Our end-to-end system achieves an overall (for all predicates) F1 score of 51.7% and 49.2% respectively. APrecision is higher for DBpedia, while recall tends to be higher for Wikidata. The results are furthermore lower than those in the paper, which is expected since the output of the search also contains events that are not in the ground truth events. Consequently, for each of such event, all generated triples will not be in the ground truth from EventKG.

Pred F1 (DB) F1 (WD) Precision (DB) Precision (WD) Recall (DB) Recall (WD)
all 51.7 49.2 79.3 41.1 38.4 61.4
sem:hasActor 48.5 27.6 76.5 35.1 35.5 22.8
sem:hasBeginTimeStamp 62.6 50.3 79.5 38.0 51.6 74.4
sem:hasEndTimeStamp 62.7 48.3 79.5 38.1 51.7 65.9
sem:hasPlace 48.2 59.2 81.6 50.7 34.2 71.2

5.3 User study

All the content related to the user studies are in the experiments_run/usage_ng folder:

  • The scripts permit to get the prompts, the answers, and the grounding triples, for all types of prompting.
  • experiments_run/usage_ng/human_evaluation contains the final results from the forms, and a notebook to analyse the results.
  • experiments_run/usage_ng/qa_prompt_answer contains the saved prompts and answers for the three prompting techniques.
  • The triples used from the event-centric KG are in the french_rev_frame_generation_all.ttl file. To query it, we set up a local GraphDB repository.

Link to the forms: