File size: 6,699 Bytes
5270fb9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 |
import os
from typing import Dict, List, Optional, Protocol
import pandas as pd
import tqdm
import ujson
from nlp4web_codebase.ir.data_loaders import IRDataset
def round_dict(obj: Dict[str, float], ndigits: int = 4) -> Dict[str, float]:
return {k: round(v, ndigits=ndigits) for k, v in obj.items()}
def sort_dict(obj: Dict[str, float], reverse: bool = True) -> Dict[str, float]:
return dict(sorted(obj.items(), key=lambda pair: pair[1], reverse=reverse))
def save_ranking_results(
output_dir: str,
query_ids: List[str],
rankings: List[Dict[str, float]],
query_performances_lists: List[Dict[str, float]],
cid2tweights_lists: Optional[List[Dict[str, Dict[str, float]]]] = None,
):
os.makedirs(output_dir, exist_ok=True)
output_path = os.path.join(output_dir, "ranking_results.jsonl")
rows = []
for i, (query_id, ranking, query_performances) in enumerate(
zip(query_ids, rankings, query_performances_lists)
):
row = {
"query_id": query_id,
"ranking": round_dict(ranking),
"query_performances": round_dict(query_performances),
"cid2tweights": {},
}
if cid2tweights_lists is not None:
row["cid2tweights"] = {
cid: round_dict(tws) for cid, tws in cid2tweights_lists[i].items()
}
rows.append(row)
pd.DataFrame(rows).to_json(
output_path,
orient="records",
lines=True,
)
class TermWeightingFunction(Protocol):
def __call__(self, query: str, cid: str) -> Dict[str, float]: ...
def compare(
dataset: IRDataset,
results_path1: str,
results_path2: str,
output_dir: str,
main_metric: str = "recip_rank",
system1: Optional[str] = None,
system2: Optional[str] = None,
term_weighting_fn1: Optional[TermWeightingFunction] = None,
term_weighting_fn2: Optional[TermWeightingFunction] = None,
) -> None:
os.makedirs(output_dir, exist_ok=True)
df1 = pd.read_json(results_path1, orient="records", lines=True)
df2 = pd.read_json(results_path2, orient="records", lines=True)
assert len(df1) == len(df2)
all_qrels = {}
for split in dataset.split2qrels:
all_qrels.update(dataset.get_qrels_dict(split))
qid2query = {query.query_id: query for query in dataset.queries}
cid2doc = {doc.collection_id: doc for doc in dataset.corpus}
diff_col = f"{main_metric}:qp1-qp2"
merged = pd.merge(df1, df2, on="query_id", how="outer")
rows = []
for _, example in tqdm.tqdm(merged.iterrows(), desc="Comparing", total=len(merged)):
docs = {cid: cid2doc[cid].text for cid in dict(example["ranking_x"])}
docs.update({cid: cid2doc[cid].text for cid in dict(example["ranking_y"])})
query_id = example["query_id"]
row = {
"query_id": query_id,
"query": qid2query[query_id].text,
diff_col: example["query_performances_x"][main_metric]
- example["query_performances_y"][main_metric],
"ranking1": ujson.dumps(example["ranking_x"], indent=4),
"ranking2": ujson.dumps(example["ranking_y"], indent=4),
"docs": ujson.dumps(docs, indent=4),
"query_performances1": ujson.dumps(
example["query_performances_x"], indent=4
),
"query_performances2": ujson.dumps(
example["query_performances_y"], indent=4
),
"qrels": ujson.dumps(all_qrels[query_id], indent=4),
}
if term_weighting_fn1 is not None and term_weighting_fn2 is not None:
all_cids = set(example["ranking_x"]) | set(example["ranking_y"])
cid2tweights1 = {}
cid2tweights2 = {}
ranking1 = {}
ranking2 = {}
for cid in all_cids:
tweights1 = term_weighting_fn1(query=qid2query[query_id].text, cid=cid)
tweights2 = term_weighting_fn2(query=qid2query[query_id].text, cid=cid)
ranking1[cid] = sum(tweights1.values())
ranking2[cid] = sum(tweights2.values())
cid2tweights1[cid] = tweights1
cid2tweights2[cid] = tweights2
ranking1 = sort_dict(ranking1)
ranking2 = sort_dict(ranking2)
row["ranking1"] = ujson.dumps(ranking1, indent=4)
row["ranking2"] = ujson.dumps(ranking2, indent=4)
cid2tweights1 = {cid: cid2tweights1[cid] for cid in ranking1}
cid2tweights2 = {cid: cid2tweights2[cid] for cid in ranking2}
row["cid2tweights1"] = ujson.dumps(cid2tweights1, indent=4)
row["cid2tweights2"] = ujson.dumps(cid2tweights2, indent=4)
rows.append(row)
table = pd.DataFrame(rows).sort_values(by=diff_col, ascending=False)
output_path = os.path.join(output_dir, f"compare-{system1}_vs_{system2}.tsv")
table.to_csv(output_path, sep="\t", index=False)
# if __name__ == "__main__":
# # python -m lecture2.bm25.analysis
# from nlp4web_codebase.ir.data_loaders.sciq import load_sciq
# from lecture2.bm25.bm25_retriever import BM25Retriever
# from lecture2.bm25.tfidf_retriever import TFIDFRetriever
# import numpy as np
# sciq = load_sciq()
# system1 = "bm25"
# system2 = "tfidf"
# results_path1 = f"output/sciq-{system1}/results/ranking_results.jsonl"
# results_path2 = f"output/sciq-{system2}/results/ranking_results.jsonl"
# index_dir1 = f"output/sciq-{system1}"
# index_dir2 = f"output/sciq-{system2}"
# compare(
# dataset=sciq,
# results_path1=results_path1,
# results_path2=results_path2,
# output_dir=f"output/sciq-{system1}_vs_{system2}",
# system1=system1,
# system2=system2,
# term_weighting_fn1=BM25Retriever(index_dir1).get_term_weights,
# term_weighting_fn2=TFIDFRetriever(index_dir2).get_term_weights,
# )
# # bias on #shared_terms of TFIDF:
# df1 = pd.read_json(results_path1, orient="records", lines=True)
# df2 = pd.read_json(results_path2, orient="records", lines=True)
# merged = pd.merge(df1, df2, on="query_id", how="outer")
# nterms1 = []
# nterms2 = []
# for _, row in merged.iterrows():
# nterms1.append(len(list(dict(row["cid2tweights_x"]).values())[0]))
# nterms2.append(len(list(dict(row["cid2tweights_y"]).values())[0]))
# percentiles = (5, 25, 50, 75, 95)
# print(system1, np.percentile(nterms1, percentiles), np.mean(nterms1).round(2))
# print(system2, np.percentile(nterms2, percentiles), np.mean(nterms2).round(2))
# # bm25 [ 3. 4. 5. 7. 11.] 5.64
# # tfidf [1. 2. 3. 5. 9.] 3.58
|