Upload script to reproduce dataset

relation_templates/relation_templates_brands.csv

@@ -0,0 +1,11 @@
+uri,relation,template
+P17,country,In what country is [subj] located?
+P112,founded by,Who founded [subj]?
+P169,chief executive officer,Who is the CEO of [subj]?
+P127,owned by,Who is the owner of [subj]?
+P452,industry,In what industry is [subj]?
+P138,named after,Who was [subj] named after?
+P571,inception,In what year was [subj] founded?
+P159,headquarters location,In what country is the headquarters of [subj]?
+P822,mascot,What or who is the mascot of [subj]?
+P1056,product or material produced,What is the product or material produced by [subj]?

relation_templates/relation_templates_celebs.csv

@@ -0,0 +1,40 @@
+uri,relation,template
+P569,date of birth,In what year was [subj] born?
+P54,member of sports team,What sports team does [subj] play for?
+P106,occupation,What is the occupation of [subj]?
+P27,country of citizenship,What is the country of citizenship of [subj]?
+P19,place of birth,In what [obj_type] was [subj] born?
+P641,sport,What sport does [subj] play?
+P1412,"languages spoken, written or signed",What language does [subj] speak?
+P413,position played on team / speciality,What sports position does [subj] play?
+P50,author,Who is the author of [subj]?
+P20,place of death,In what [obj_type] did [subj] die?
+P495,country of origin,What is the country of origin of [subj]?
+P69,educated at,What is the alma mater of [subj]?
+P407,language of work or name,What is the language of [subj]?
+P1598,consecrator,Who is the consecrator of [subj]?
+P123,publisher,Who is the publisher of [subj]?
+P40,child,Who is the child of [subj]?
+P140,religion or worldview,What is the religion of [subj]?
+P57,director,Who was the director of [subj]?
+P264,record label,What is the record label of [subj]?
+P58,screenwriter,Who was the screenwriter for [subj]?
+P103,native language,What is the native language of [subj]?
+P1532,country for sport,What country does [subj] play for?
+P22,father,Who is the father of [subj]?
+P86,composer,Who was the composer of [subj]?
+P175,performer,Who is the performer of [subj]?
+P118,league,What sports league does [subj] play in?
+P26,spouse,Who is the spouse of [subj]?
+P162,producer,Who was the producer of [subj]?
+P3373,sibling,Who is the sibling of [subj]?
+P98,editor,Who is the editor of [subj]?
+P25,mother,Who is the mother of [subj]?
+P676,lyrics by,Who wrote the lyrics of [subj]?
+P166,award received,What award did [subj] receive?
+P451,unmarried partner,Who is the partner of [subj]?
+P800,notable work,What notable [obj_type] did [subj] create?
+P1303,instrument,What instrument does [subj] play?
+P39,position held,What position is held by [subj]?
+P101,field of work,What is the field of work of [subj]?
+P102,member of political party,Which political party is [subj] affiliated with?

relation_templates/relation_templates_landmarks.csv

@@ -0,0 +1,12 @@
+uri,relation,template
+P17,country,In what country is [subj] located?
+P112,founded by,Who founded [subj]?
+P127,owned by,Who is the owner of [subj]?
+P149,architectural style,What is the architectural style of [subj]?
+P571,inception,In what year was [subj] created?
+P84,architect,Who was the architect of [subj]?
+P580,start time,In what year did [subj] get its heritage designation?
+P2048,hight,How high is [subj]?
+P131,located in the administrative territorial entity,In what administrative territorial entity is [subj] located?
+P669,located on street,On what street is [subj] located?
+P1435,heritage designation,What is the heritage designation of [subj]?

relation_templates/relation_templates_paintings.csv

@@ -0,0 +1,13 @@
+uri,relation,template
+P127,owned by,Who is the owner of [subj]?
+P170,creator,Who painted [subj]?
+P186,made from material,From what material is [subj] made?
+P195,collection,A part of what collection is [subj]?
+P2348,time period,From what artistic time period is [subj]?
+P547,commemorates,What does [subj] commemorate?
+P138,named after,Who was [subj] named after?
+P571,inception,In what year was [subj] painted?
+P17,country,In what country is [subj] located?
+P495,country of origin,In what country was [subj] created?
+P88,commissioned by,Who commissioned [subj]?
+

scripts/build_dataset.py

@@ -0,0 +1,91 @@
+from dotenv import load_dotenv
+load_dotenv()
+from dataset_utils import *
+import argparse
+
+'''
+This is a script to build (or expand) the PopVQA dataset.
+Before running this script, make sure your directory contains the following:
+1. A CSV file with the base dataframe containing the columns 's_uri' (from wikidata) and 'type' (to match against the relation templates).
+2. a dir named 'relation_templates' containing CSV files with the relation templates for each type. The templates are triplets 'uri' - 'relation' - 'template'.
+See the existing files for reference.
+Run the script with the following command:
+python build_dataset.py --base-df <path_to_base_df> --start <start_step> --end <end_step>
+'''
+
+def main(args):
+    dir_name, file_name = os.path.split(args.base_df)
+    base_name, _ = os.path.splitext(file_name)
+    base_df = pd.read_csv(args.base_df).drop_duplicates('s_uri')
+    assert 'type' in base_df.columns, "The base dataframe must contain a 'type' column."
+
+    types = base_df['type'].unique()
+    
+    for entity_type in base_df['type'].unique():
+        template_path = os.path.join(dir_name, "relation_templates", f"relation_templates_{entity_type}.csv")
+        assert os.path.isfile(template_path), f"Missing relation template for type '{entity_type}' at: {template_path}"
+
+    all_question_dfs = []
+
+    for entity_type in types:
+        type_df = base_df[base_df['type'] == entity_type].copy()
+        type_dir = os.path.join(dir_name, entity_type)
+        os.makedirs(type_dir, exist_ok=True)
+
+        template_path = os.path.join(dir_name, "relation_templates", f"relation_templates_{entity_type}.csv")
+        templates = pd.read_csv(template_path)
+
+        print(f"Processing type: {entity_type}")
+        if args.start <= 0:
+            subject_to_relation = get_all_properties(type_df)
+            subject_to_relation = subject_to_relation[subject_to_relation['r_uri'].isin(templates['uri'])]
+            subject_to_relation.to_csv(os.path.join(type_dir, f"{base_name}_subject_to_relation.csv"), index=False)
+
+        if args.start <= 1 and args.end >= 1:
+            if args.start == 1:
+                subject_to_relation = pd.read_csv(os.path.join(type_dir, f"{base_name}_subject_to_relation.csv"))
+            aliases = get_aliases(subject_to_relation)
+            aliases.to_csv(os.path.join(type_dir, f"{base_name}_all_aliases.csv"), index=False)
+
+        if args.start <= 2 and args.end >= 2:
+            if args.start == 2:
+                subject_to_relation = pd.read_csv(os.path.join(type_dir, f"{base_name}_subject_to_relation.csv"))
+                aliases = pd.read_csv(os.path.join(type_dir, f"{base_name}_all_aliases.csv"))
+            a_types = attribute_type(subject_to_relation)
+            a_types.to_csv(os.path.join(type_dir, f"{base_name}_complete_attribute_types.csv"), index=False)
+
+        if args.start <= 3 and args.end >= 3:
+            if args.start == 3:
+                subject_to_relation = pd.read_csv(os.path.join(type_dir, f"{base_name}_subject_to_relation.csv"))
+                aliases = pd.read_csv(os.path.join(type_dir, f"{base_name}_all_aliases.csv"))
+                a_types = pd.read_csv(os.path.join(type_dir, f"{base_name}_complete_attribute_types.csv"))
+            triplets = aggregate_triplets(type_df, aliases, subject_to_relation, a_types, add_unesco=False)
+            triplets.to_csv(os.path.join(type_dir, f"{base_name}_question_triplets.csv"), index=False)
+
+        if args.start <= 4 and args.end >= 4:
+            if args.start == 4:
+                triplets = pd.read_csv(os.path.join(type_dir, f"{base_name}_question_triplets.csv"))
+            triplets = build_prompts(type_df, triplets, templates)
+            triplets['type'] = entity_type
+            triplets.to_csv(os.path.join(type_dir, f"{base_name}_questions.csv"), index=False)
+            all_question_dfs.append(triplets)
+
+    # Combine all question files and write to the top-level directory
+    if all_question_dfs:
+        combined_df = pd.concat(all_question_dfs, ignore_index=True)
+        combined_df.to_csv(os.path.join(dir_name, f"{base_name}_all_questions.csv"), index=False)
+        print(f"Combined questions file saved to {os.path.join(dir_name, f'{base_name}_all_questions.csv')}")
+
+
+def get_exp_parser():
+    parser = argparse.ArgumentParser(add_help=False)
+    parser.add_argument('--base-df', type=str)
+    parser.add_argument('--start', type=int, default=0, help="Start step for building the dataset.")
+    parser.add_argument('--end', type=int, default=4, help="End step for building the dataset.")
+    return parser
+
+
+if __name__ == "__main__":
+    parser = get_exp_parser()
+    args = parser.parse_args()
+    main(args)

scripts/dataset_utils.py

@@ -0,0 +1,343 @@
+import os
+import pandas as pd
+import numpy as np
+from ast import literal_eval
+from SPARQLWrapper import SPARQLWrapper, JSON
+from tqdm import tqdm
+from urllib.parse import urlparse
+import requests
+import re
+from ast import literal_eval
+from tqdm import tqdm
+tqdm.pandas()
+
+from dotenv import load_dotenv
+load_dotenv()
+
+DATA_DIR = os.environ['DATA_DIR']
+
+replacements = {"celebs":"the subject of this image",
+                "brands":"the brand in this image",
+                "landmarks":"the place in this image",
+                "paintings":"the painting in this image",
+                }
+
+def best_obj_type(obj_types):
+        if type(obj_types) == str:
+            obj_types = literal_eval(obj_types)
+        prioritized_obj_types = ["city", "capital city", 'metropolis', 'country', 'occupation', 'language', 'type of sport', 'music genre'] # 'cinematic technique', 'team sport'
+        for ot in prioritized_obj_types:
+            if ot in obj_types:
+                return ot
+            for ot_ in obj_types:
+                if "university" in ot_:
+                    return "university"
+                if "city" in ot_:
+                    return "city"
+        return obj_types[0]
+
+def replace_for_image(row):
+    replace_with = replacements[row['type']]
+    return row["template"].replace("[subj]", replace_with)
+
+class SPARQL:
+    def __init__(self):
+        self.agent = "'Mozilla/5.0 (Macintosh; Intel Mac OS X 10_11_5) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/50.0.2661.102 Safari/537.36'"
+        self.sparql = SPARQLWrapper("https://query.wikidata.org/sparql", agent=self.agent)
+        self.sparql.setReturnFormat(JSON)
+
+    def parse_value(self, value):
+        parsed_uri = urlparse(value)
+        if all([parsed_uri.scheme, parsed_uri.netloc]):
+            return parsed_uri.path.split('/')[-1]
+        return value 
+
+    def execute(self, query):
+        records = []     
+        try:
+            self.sparql.setQuery(query)
+            responses = self.sparql.query().convert()
+            for response in responses['results']['bindings']:
+                record = {}
+                for key in response:
+                    record[key] = self.parse_value(response[key]['value'])
+                records.append(record)
+            if records == 0:
+                print("request failed")
+        except Exception as e:
+            print(e)
+        return pd.DataFrame(records)
+
+
+def add_aliases(df):
+    def _query(uris):
+        return f'''
+SELECT ?s_uri ?alias
+WHERE {{
+    {{VALUES ?s_uri {{ {" ".join([f"wd:{uri}" for uri in uris])} }} }}
+    ?s_uri skos:altLabel ?alias.
+    FILTER(LANG(?alias) = "en")
+}}
+''' 
+    sparql = SPARQL()
+
+    uris = list(set(df["s_uri"].tolist()))
+    uri_chunks = [uris[i:i+100] for i in range(0, len(uris), 100)]
+
+    aliases = pd.concat([sparql.execute(_query(chunk)) for chunk in tqdm(uri_chunks)])
+    aliases = aliases.groupby("s_uri")["alias"].agg(list).reset_index(name="aliases")
+    res = pd.merge(df, aliases, how='left', on='s_uri')
+    res['aliases'] = res['aliases'].fillna('[]')
+    return res
+
+def get_aliases(df):
+    def _query(uris): 
+        return f'''
+SELECT ?uri ?alias
+WHERE {{
+    {{VALUES ?uri {{ {" ".join([f"wd:{uri}" for uri in uris])} }} }}
+    ?uri skos:altLabel ?alias.
+    FILTER(LANG(?alias) = "en")
+}}
+''' 
+    sparql = SPARQL()
+
+    uris = list(set(df["s_uri"].tolist()))# + df["a_uri"].tolist()))
+    uri_chunks = [uris[i:i+100] for i in range(0, len(uris), 100)]
+
+    aliases = pd.concat([sparql.execute(_query(chunk)) for chunk in tqdm(uri_chunks)])
+    aliases = aliases.groupby("uri")["alias"].agg(list).reset_index(name="aliases")
+    return aliases
+
+def add_images(df):
+    def _query(uris):
+        return f'''
+SELECT ?s_uri ?image
+WHERE {{
+    {{VALUES ?s_uri {{ {" ".join([f"wd:{uri}" for uri in uris])} }} }}
+    ?s_uri wdt:P18 ?image .
+}}
+''' 
+    sparql = SPARQL()
+
+    uris = list(set(df["s_uri"].tolist()))
+    uri_chunks = [uris[i:i+100] for i in range(0, len(uris), 100)]
+
+    images = pd.concat([sparql.execute(_query(chunk)) for chunk in tqdm(uri_chunks)])
+    images['image'] = 'http://commons.wikimedia.org/wiki/Special:FilePath/' + images['image']
+    res = pd.merge(df, images, how='inner', on='s_uri')
+    return res
+
+
+def get_attribute(df, attribute_name, attribute_id):
+    def _query(uris):
+        return f'''
+SELECT ?s_uri ?attribute_name
+WHERE {{
+    {{VALUES ?s_uri {{ {" ".join([f"wd:{uri}" for uri in uris])} }} }}
+    ?s_uri wdt:{attribute_id} ?{attribute_name} .
+}}
+''' 
+    sparql = SPARQL()
+
+    uris = list(set(df["s_uri"].tolist()))
+    uri_chunks = [uris[i:i+100] for i in range(0, len(uris), 100)]
+
+    attributes = pd.concat([sparql.execute(_query(chunk)) for chunk in tqdm(uri_chunks)])
+    attributes = attributes.groupby("s_uri")[attribute_name].agg(list).reset_index(name=attribute_name)
+
+    res = pd.merge(df, attributes, how='inner', on='s_uri')
+    return res
+
+def extract_year(timestamp):
+    parts = timestamp.split('-')
+    neg = False
+    if parts[0] == '':
+        year = parts[1]
+        neg = True
+    else:
+        year = parts[0]
+    if year.isdigit():
+        return str(-int(year)) if neg else str(int(year)) 
+    return np.nan
+
+def get_all_properties(df):
+    def _query(relation_ids): 
+        return f'''
+SELECT ?item ?itemLabel ?wd ?wdLabel ?ps_ ?ps_Label WHERE {{
+  VALUES ?item {{ 
+    {" ".join([f"wd:{id}" for id in relation_ids])}
+  }}
+  ?item ?p ?statement .
+  ?statement ?ps ?ps_ .
+  ?wd wikibase:claim ?p .
+  ?wd wikibase:statementProperty ?ps .
+  
+  SERVICE wikibase:label {{ bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en". }}
+}}
+'''
+    sparql = SPARQL()
+    
+    # df = pd.read_csv(origin)
+    subjects = df["s_uri"].to_list()
+    subject_chunks = [subjects[i:i+20] for i in range(0, len(subjects), 20)]
+    
+    df = pd.concat([sparql.execute(_query(chunk)) for chunk in tqdm(subject_chunks)])
+    df = df[~df["wdLabel"].str.contains(r"ID|category|template|username|instance of|gallery|article|handle|url|wiki|copyright|classification|website|described|tag|archive|reddit|profile|image|list|file", case=False, na=False)]
+    tmp = df[(df['wd'] == 'P569') | (df['wd'] == 'P571')].copy()
+    tmp['ps_Label'] = tmp['ps_Label'].apply(extract_year)
+    tmp.dropna(subset=['ps_Label'], inplace=True)
+    tmp['ps_'] = 'Q000'
+    df = df[~((df['wd'] == 'P569') | (df['wd'] == 'P571'))]
+    df = df[~df["ps_Label"].str.contains(r'\d', na=False)]
+    df = df[df["ps_"].apply(lambda s: bool(re.fullmatch(r"Q\d+", s)))]
+    df = pd.concat([df, tmp])
+    df = df[["item", "itemLabel", "wd", "wdLabel", "ps_", "ps_Label"]]
+    df = df.rename(
+        columns = {
+            "item": "s_uri",
+            "itemLabel": "subject",
+            "wd": "r_uri",
+            "wdLabel": "relation",
+            "ps_": "a_uri",
+            "ps_Label": "attribute",
+        }
+    )
+    return df
+
+
+def attribute_type(df):
+    def _query(uris):
+        return f'''
+SELECT ?uri ?typeLabel
+WHERE {{
+  {{VALUES ?uri {{ {" ".join([f"wd:{uri}" for uri in uris])} }} }}
+  ?uri wdt:P31 ?type.
+  SERVICE wikibase:label {{ bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en". }}
+}}   
+''' 
+    sparql = SPARQL()
+    
+    uris = df["a_uri"].drop_duplicates().to_list()
+    uri_chunks = [uris[i:i+100] for i in range(0, len(uris), 100)]
+    a_types = pd.concat([sparql.execute(_query(chunk)) for chunk in tqdm(uri_chunks)])
+    a_types = a_types.groupby("uri")["typeLabel"].agg(list).reset_index(name="a_type")
+    a_types['a_type'] = a_types['a_type'].apply(lambda x: x if type(x) == list else [])
+    a_types = pd.concat([a_types, pd.DataFrame([{'uri': 'Q000', 'a_type': str(['year'])}])])
+    return a_types
+
+
+def get_wikidata_id(name):
+    url = "https://www.wikidata.org/w/api.php"
+    params = {
+        "action": "wbsearchentities",
+        "format": "json",
+        "language": "en",
+        "search": name
+    }
+    response = requests.get(url, params=params).json()
+    if 'search' in response and response['search']:
+        return response['search'][0]['id']
+    return None
+
+def add_wikidata_ids(df, name_col="subject"):
+    df["wikidata_id"] = df[name_col].apply(get_wikidata_id)
+    return df
+
+
+def add_unesco_question(base_df):
+    def _query(qids): 
+        return f"""
+    SELECT ?item ?itemLabel ?startTime WHERE {{
+      VALUES ?item {{{' '.join(f'wd:{qid}' for qid in qids)}}}
+      ?item p:P1435 ?heritageStatement.
+      ?heritageStatement ps:P1435 wd:Q9259.
+      OPTIONAL {{
+        ?heritageStatement pq:P580 ?startTime.
+      }}
+      SERVICE wikibase:label {{ bd:serviceParam wikibase:language "[AUTO_LANGUAGE],en". }}
+    }}
+    """
+    sparql = SPARQL()
+    
+    df = base_df[base_df['type'] == 'landmarks']
+    subjects = df["s_uri"].to_list()
+    subject_chunks = [subjects[i:i+20] for i in range(0, len(subjects), 20)]
+    
+    df = pd.concat([sparql.execute(_query(chunk)) for chunk in tqdm(subject_chunks)])
+    df.dropna(subset=['startTime'], inplace=True)
+    df['startTime'] = df['startTime'].apply(extract_year)
+    df = df.rename(
+        columns = {
+            "item": "s_uri",
+            "startTime": "attribute",
+            "itemLabel": "subject",
+        }
+    )
+    df['possible_answers'] = df['attribute'].apply(lambda x: str([x]))
+    df['r_uri'] = 'P580'
+    df['relation'] = 'start time'
+    df['a_uri'] = 'P580'
+    df['a_type'] = str(['year'])
+    return df    
+
+
+def aggregate_triplets(base, aliases, relations, attributes, add_unesco=False):
+    subjects = base[['s_uri']]
+    relations = relations.merge(subjects, on="s_uri")
+    aliases = pd.read_csv("data/all_aliases.csv", index_col=0)
+    if type(aliases.iloc[0]['aliases']) == str:
+        aliases["aliases"] = aliases["aliases"].apply(lambda x: literal_eval(x))
+    if type(attributes.iloc[0]['a_type']) == str:
+        attributes["a_type"] = attributes["a_type"].apply(lambda x: literal_eval(x))
+
+    relations = relations.merge(aliases, left_on="a_uri", right_on="uri", how="left")
+    relations = relations.drop(columns=["uri"])
+    relations["possible_answers"] = relations['aliases'].apply(lambda x: x if type(x) == list else [])
+    relations["possible_answers"] = relations.progress_apply(lambda x: x["possible_answers"] + [x["attribute"]], axis=1)
+    
+    agg_funcs = {col: 'first' for col in relations.columns if col not in ['s_uri', 'r_uri', 'possible_answers']}
+    agg_funcs['possible_answers'] = sum
+    relations = relations.groupby(['s_uri', 'r_uri'], as_index=False).agg(agg_funcs)
+    
+    relations = relations.drop(columns=["aliases"])
+    relations = relations.merge(attributes, left_on="a_uri", right_on="uri", how="left")
+    relations = relations.drop(columns=["uri"])
+    
+    if add_unesco:
+        unesco = add_unesco_question(base)
+        relations = pd.concat([relations, unesco])
+    
+    return relations
+
+
+def subj_substitute(row):
+    if row['type'] == 'brands':
+        return f"the brand {row['subject']}"
+    if row['type'] == 'paintings':
+        return f"the painting {row['subject']}"
+    return row['subject']
+
+
+def build_prompts(base_df, triplets, templates):
+    subjects = base_df[["s_uri", "subject"]]
+    base_df = base_df[["s_uri", "type"]]
+    triplets = triplets.drop("subject", axis=1)
+    triplets = triplets.merge(subjects, on=["s_uri"])
+    triplets = triplets.merge(base_df, on=["s_uri"], how='left')
+    triplets = triplets.merge(templates[["uri", "template"]], left_on="r_uri", right_on="uri")
+    triplets = triplets.drop(columns=["uri"])
+    triplets = triplets.dropna()
+    
+    query_counts = triplets.drop_duplicates(["s_uri", "r_uri"]).groupby(["s_uri"])["r_uri"].count().reset_index(name="count")
+    triplets = triplets.merge(query_counts[query_counts["count"] > 1][["s_uri"]], on="s_uri")
+
+    triplets["question_for_image"] = triplets.progress_apply(replace_for_image, axis=1)
+    triplets["question_for_image"] = triplets.progress_apply(lambda row: row["question_for_image"].replace("[obj_type]", best_obj_type(row["a_type"])) if len(row["a_type"]) > 0 else row["question"], axis=1)
+    triplets["question"] = triplets.progress_apply(lambda row: row["template"].replace("[subj]", subj_substitute(row)), axis=1)
+    triplets["question"] = triplets.progress_apply(lambda row: row["question"].replace("[obj_type]", best_obj_type(row["a_type"])) if len(row["a_type"]) > 0 else row["question"], axis=1)
+    triplets = triplets.drop(columns=["template"])
+    triplets = triplets[['type','subject','question_for_image','question','possible_answers', 'relation', 's_uri', 'r_uri','a_uri','attribute','a_type']]
+    return triplets
+
+