bordirlines.py

import json
from copy import copy
from functools import lru_cache

import datasets
import pandas as pd

SUPPORTED_LANGUAGES = [
    "sl",
    "ur",
    "sw",
    "uz",
    "vi",
    "sq",
    "ms",
    "km",
    "hy",
    "da",
    "ky",
    "mg",
    "mn",
    "ja",
    "el",
    "it",
    "is",
    "ru",
    "tl",
    "so",
    "pt",
    "uk",
    "sr",
    "sn",
    "ht",
    "bs",
    "my",
    "ar",
    "hr",
    "nl",
    "bn",
    "ne",
    "hi",
    "ka",
    "az",
    "ko",
    "id",
    "fr",
    "es",
    "en",
    "fa",
    "lo",
    "iw",
    "th",
    "tr",
    "zht",
    "zhs",
    "ti",
    "tg",
    "control",
]
SYSTEMS = ["openai", "m3"]
MODES = ["qlang", "qlang_en", "en", "rel_langs"]
RELEVANCE_FILTERS = ["all", "relevant", "non-relevant"]
LLM_MODES = ["zeroshot", "fewshot"]

ROOT_DIR = "data"


class BordIRlinesConfig(datasets.BuilderConfig):
    def __init__(self, language, n_hits=10, **kwargs):
        super(BordIRlinesConfig, self).__init__(**kwargs)
        self.language = language
        self.n_hits = n_hits
        self.data_root_dir = ROOT_DIR


def load_json(path):
    with open(path, "r", encoding="utf-8") as f:
        return json.load(f)


@lru_cache
def replace_lang_str(path, lang):
    parent = path.rsplit("/", 2)[0]
    return f"{parent}/{lang}/{lang}_docs.json"


def get_label(human_bool, llm_bool, annotation_type):
    if annotation_type == "human":
        return human_bool
    elif annotation_type == "llm":
        return llm_bool
    else:
        return human_bool if human_bool is not None else llm_bool


class BordIRLinesDataset(datasets.GeneratorBasedBuilder):
    VERSION = datasets.Version("1.0.0")

    BUILDER_CONFIGS = [
        BordIRlinesConfig(
            name=lang,
            language=lang,
            description=f"{lang.upper()} dataset",
        )
        for lang in SUPPORTED_LANGUAGES
    ]

    def __init__(
        self,
        *args,
        relevance_filter="all",
        annotation_type=None,
        llm_mode="fewshot",
        viewpoint_filter=None,
        **kwargs,
    ):
        super().__init__(*args, **kwargs)
        self.relevance_filter = relevance_filter
        assert self.relevance_filter in RELEVANCE_FILTERS
        self.annotation_type = annotation_type
        self.llm_mode = llm_mode
        assert self.llm_mode in LLM_MODES
        self.viewpoint_filter = viewpoint_filter  # Filter for a specific viewpoint

    def _info(self):
        return datasets.DatasetInfo(
            description="IR Dataset for BordIRLines paper.",
            features=datasets.Features(
                {
                    "query_id": datasets.Value("string"),
                    "query": datasets.Value("string"),
                    "query_lang": datasets.Value("string"),
                    "territory": datasets.Value("string"),
                    "rank": datasets.Value("int32"),
                    "score": datasets.Value("float32"),
                    "doc_id": datasets.Value("string"),
                    "doc_text": datasets.Value("string"),
                    "doc_lang": datasets.Value("string"),
                    "viewpoint_human": datasets.Value("string"),
                    "viewpoint_llm": datasets.Value("string"),
                    "relevant_human": datasets.Value("bool"),
                    "relevant_llm": datasets.Value("bool"),
                }
            ),
        )

    def _split_generators(self, dl_manager):
        base_url = self.config.data_root_dir
        queries_path = f"{base_url}/queries.tsv"
        docs_path = dl_manager.download_and_extract(f"{base_url}/all_docs.json")
        human_annotations_path = dl_manager.download_and_extract(
            f"{base_url}/human_annotations.tsv"
        )
        llm_annotations_path = dl_manager.download_and_extract(f"{base_url}/llm_annotations.tsv")

        lang = self.config.language

        splits = []
        downloaded_data = {}

        for system in SYSTEMS:
            for mode in MODES:
                source = f"{system}.{mode}"
                downloaded_data[source] = dl_manager.download_and_extract(
                    {
                        "hits": f"{base_url}/{lang}/{system}/{mode}/{lang}_query_hits.tsv",
                        "docs": docs_path,
                        "queries": queries_path,
                        "human_annotations": human_annotations_path,
                        "llm_annotations": llm_annotations_path,
                    }
                )

                split = datasets.SplitGenerator(
                    name=f"{system}.{mode}",
                    gen_kwargs={
                        "hits_path": downloaded_data[source]["hits"],
                        "docs_path": downloaded_data[source]["docs"],
                        "queries_path": downloaded_data[source]["queries"],
                        "human_annotations_path": downloaded_data[source]["human_annotations"],
                        "llm_annotations_path": downloaded_data[source]["llm_annotations"],
                    },
                )
                splits.append(split)

        return splits

    def _skip_viewpoint(self, viewpoint_human, viewpoint_llm, query_entry):
        viewpoint = get_label(viewpoint_human, viewpoint_llm, self.annotation_type)
        if viewpoint is None:
            return True

        if self.viewpoint_filter == "Non-controllers":
            controller = query_entry["Controller"]
            if controller == "Unknown":
                return True

            claimants = copy(query_entry["Claimants"])
            claimants.remove(controller)
            return (
                not claimants or viewpoint not in claimants
            )  # skip if not a non-controller viewpoint

        # otherwise, handle the case where we want to filter for a specific viewpoint
        target_viewpoint = (
            query_entry["Controller"]
            if self.viewpoint_filter == "Controller"
            else self.viewpoint_filter
        )

        return target_viewpoint and viewpoint != target_viewpoint

    def _skip_relevance(self, relevant_human, relevant_llm):
        # Filtering logic based on relevance preference
        relevant = get_label(relevant_human, relevant_llm, self.annotation_type)
        target_relevant = {"relevant": True, "non-relevant": False}.get(self.relevance_filter, None)
        return target_relevant is not None and relevant != target_relevant
        # If "all", do not filter anything

    def _generate_examples(
        self, hits_path, docs_path, queries_path, human_annotations_path, llm_annotations_path
    ):
        n_hits = self.config.n_hits
        queries_df = pd.read_csv(queries_path, sep="\t").set_index("query_id")
        queries_df["Claimants"] = queries_df["Claimants"].str.split(";").map(set)
        counter = 0

        docs = load_json(docs_path)

        hits = pd.read_csv(hits_path, sep="\t")
        human_annotations = pd.read_csv(human_annotations_path, sep="\t")
        llm_annotations = pd.read_csv(llm_annotations_path, sep="\t")

        if n_hits:
            hits = hits.groupby("query_id").head(n_hits)

        # sort hits by query_id and rank
        hits["query_id_int"] = hits["query_id"].str[1:].astype(int)
        hits = hits.sort_values(by=["query_id_int", "rank"])
        hits = hits.drop(columns=["query_id_int"])

        human_map = human_annotations.set_index(["query_id", "doc_id"]).to_dict(orient="index")
        llm_map = llm_annotations.set_index(["query_id", "doc_id"]).to_dict(orient="index")

        for _, row in hits.iterrows():
            doc_id = row["doc_id"]
            doc_lang = row["doc_lang"]
            query_id = row["query_id"]
            query_entry = queries_df.loc[query_id]
            query_text = query_entry["query_text"]
            query_lang = query_entry["language"]

            # Get Human Data
            human_data = human_map.get((query_id, doc_id), {})

            relevant_human = human_data.get("relevant", None)
            viewpoint_human = human_data.get("territory", None)

            # Get LLM Data
            llm_data = llm_map.get((query_id, doc_id), {})
            relevant_llm = llm_data[f"relevant_{self.llm_mode}"]
            viewpoint_llm = llm_data[f"territory_{self.llm_mode}"]
            # Filtering logic based on viewpoint preference
            viewpoint_llm = viewpoint_llm.split(") ", 1)[-1] if not pd.isna(viewpoint_llm) else None

            if self.viewpoint_filter:
                do_skip = self._skip_viewpoint(viewpoint_human, viewpoint_llm, query_entry)
                if do_skip:
                    continue

            if self.relevance_filter != "all":
                do_skip = self._skip_relevance(relevant_human, relevant_llm)
                if do_skip:
                    continue

            yield (
                counter,
                {
                    "query_id": query_id,
                    "query": query_text,
                    "query_lang": query_lang,
                    "territory": row["territory"],
                    "rank": row["rank"],
                    "score": row["score"],
                    "doc_id": doc_id,
                    "doc_text": docs[doc_lang][doc_id],
                    "doc_lang": doc_lang,
                    "viewpoint_human": viewpoint_human,
                    "viewpoint_llm": viewpoint_llm,
                    "relevant_human": relevant_human,
                    "relevant_llm": relevant_llm,
                },
            )
            counter += 1