Wiki-Infobox-Completion

Overview

The graph below shows the coherence of the files in this repository. The description below the graph contains more information and links to the respective files.

flowchart TD


    subgraph pre-processing
        A(get_cities.py) --- R[[data/cities_125000.json]] 
        A --- S[[data/cities_250000.json]]
        R --- D(get_properties.py) --- T[[data/all_properties.json]] 
        D --- U[[data/properties_per_city.json]] 
        T --- E(extract_embeddings.py) --- V[[all_properties_with_emb.json]]
        V --- F(compute_threshold.py) 

        
    end
    
    F --- Config("config.py\n  - threshold\n  - model_name\n  - seed")
    Config -.- G

    F -.-> C(property_similarity.py) -.-> G & Z

    subgraph testing system
        U & V --- Z(test_system.py) --> '[[stdout]]
    end

    subgraph infobox completion
        S & R --- B(parse_infoboxes.py)
        B --- W[["data/infoboxes_125000.json"]]--> I
        B --- Y[["data/infoboxes_250000.json"]] --> I
        I(main.py)
        G(embeddings_alignment.py) -.-> I
        H(value_alignment.py) -.-> I
        I--> X[[data/completed_infoboxes.json]]
    end

Loading

get_cities.py retrieves cities from Wikidata that are an 'instance of' city and have a population of 125,000 or more. It also creates a file with cities with a population of 250,000 or more. These files get used by different parts of the program. This population threshold is to ensure that the cities are more likely have both an English and Dutch wikipedia page, and to keep computing power to a reasonable amount.

cities.json contains the cities found by get_cities.py with 125,000 inhabitants. For each city the file contains the name of that city in English and the link to the Wikidata page. cities_250000.json contains the > 250,000 inhabitants version.

parse_infoboxes.py retrieves every Wikipedia page in cities.json and retrieves the infoboxes. It processes the infoboxes into data/infoboxes.json, which contains all the property-value pairs for every city in both languages.

property_similarity.py contains the code to compute similarity between two properties in different languages using a pretrained multilingual large language model.

To test how well the property_similarity.py script works, there are two scripts for testing the system. get_properties.py uses the Wikidata links in cities.json to extract all the unique properties that are found in all the cities pages. For these properties it retrieves both the Dutch and English names for the properties, and their frequencies. This information is saved in all_properties.json. test_system.py will use these language pairs to evaluate the property_similarity.py method.

The final infobox completion happens in main.py, which uses the value alignment script and embedding alignment script to complete Dutch infoboxes with fields that only the English infoboxes contain. the embedding alignment script uses a threshold set by finding a Euclidean distance threshold that still achieves reasonable precision on the Wikidata test set. The final results are saved in cities-completed-250000.

The final test will be a qualitative evaluation, where we select a random sample of infoboxes from the system output to evaluate. The random sample of 40 completed infoboxes are saved in test-cities_250000.json. The test-cities files followed by a name are for annotating whether a mapped property was correct.

Running the code

To run the code, use a Python version of 3.10 or above.

Create a virtual environnment, activate it, and install the requirements:

python3 -m venv env
source env/bin/activaten
pip3 install -r requirements.txt

Then pick a file to run. The correct filenames are provided.