app.py

import json
import os
import random
from urllib.request import urlopen

import pandas as pd
import plotly.graph_objects as go
import gradio as gr

# -------------------------------------------------------------------
# Load data
# -------------------------------------------------------------------

DATA_SOURCE = "https://os.zhdk.cloud.switch.ch/115-canonical-processed-final/langident/langident-lid-ensemble_multilingual_v2-0-2/langid-ocrqa_v2-0-0.json"

with urlopen(DATA_SOURCE) as response:
    data = json.load(response)

# -------------------------------------------------------------------
# Flatten yearly OCRQA data
# -------------------------------------------------------------------

rows = []

for media in data.get("media_list", []):
    provider = media.get("data_provider")
    newspaper = media.get("media_title")

    for stats in media.get("media_statistics", []):
        if stats.get("granularity") != "year":
            continue

        try:
            year = int(stats["element"].rsplit("-", 1)[-1])
        except Exception:
            continue

        media_stats = stats.get("media_stats", {})
        avg_ocrqa = media_stats.get("avg_ocrqa")
        if avg_ocrqa is None:
            continue

        rows.append(
            {
                "provider": provider,
                "newspaper": newspaper,
                "year": year,
                "avg_ocrqa": avg_ocrqa,
                "issues": media_stats.get("issues"),
                "content_items_out": media_stats.get("content_items_out"),
            }
        )

df = pd.DataFrame(rows).sort_values(["provider", "newspaper", "year"])

if df.empty:
    raise ValueError("No yearly OCRQA data found.")

# -------------------------------------------------------------------
# Alias lookups (ALL-ALIAS.jsonl)
# -------------------------------------------------------------------

media_title_map: dict[str, str] = {}  # media_alias  → full title
provider_name_map: dict[str, str] = {}  # provider_alias → full name

_alias_path = os.path.join(
    os.path.dirname(os.path.abspath(__file__)), "ALL-ALIAS.jsonl"
)
with open(_alias_path, encoding="utf-8") as _f:
    for _line in _f:
        _line = _line.strip()
        if not _line:
            continue
        _entry = json.loads(_line)
        media_title_map[_entry["media_alias"].strip()] = _entry["media_title"]
        _pa = _entry["provider_alias"].strip()
        if _pa not in provider_name_map:
            provider_name_map[_pa] = _entry["provider_name"]


def newspaper_label(alias: str) -> str:
    alias = alias.strip()
    title = media_title_map.get(alias, alias)
    return f"{title} [{alias}]" if title != alias else alias


def provider_label(alias: str) -> str:
    alias = alias.strip()
    name = provider_name_map.get(alias, alias)
    # Strip a trailing "(ALIAS)" already embedded in provider_name
    suffix = f"({alias})"
    if name.endswith(suffix):
        name = name[: -len(suffix)].strip()
    return f"{name} [{alias}]"


provider_options = [("All", "All")] + sorted(
    [(provider_label(p), p) for p in df["provider"].dropna().unique()],
    key=lambda x: x[0],
)

# -------------------------------------------------------------------
# Rankings
# -------------------------------------------------------------------

ranking_by_provider = (
    df.groupby(["provider", "newspaper"], as_index=False)["avg_ocrqa"]
    .mean()
    .rename(columns={"avg_ocrqa": "mean_ocrqa"})
)

ranking_global = (
    df.groupby("newspaper", as_index=False)["avg_ocrqa"]
    .mean()
    .rename(columns={"avg_ocrqa": "mean_ocrqa"})
)


def get_ranked_df(provider="All", query=""):
    if provider == "All":
        ranked = ranking_global.copy()
    else:
        ranked = ranking_by_provider.loc[
            ranking_by_provider["provider"] == provider, ["newspaper", "mean_ocrqa"]
        ].copy()

    ranked = ranked.sort_values(
        ["mean_ocrqa", "newspaper"], ascending=[False, True]
    ).reset_index(drop=True)

    if query:
        q = query.strip()

        def _matches(alias: str) -> bool:
            if q in alias:
                return True
            return q in media_title_map.get(alias.strip(), "")

        ranked = ranked[ranked["newspaper"].apply(_matches)].reset_index(drop=True)

    return ranked


def choose_newspapers(ranked, n_best, n_worst, n_random, seed=13):
    ranked_names = ranked["newspaper"].tolist()

    best = ranked_names[: int(n_best)] if n_best > 0 else []
    worst = ranked_names[-int(n_worst) :] if n_worst > 0 else []

    remaining_for_random = [
        n for n in ranked_names if n not in set(best) and n not in set(worst)
    ]

    rng = random.Random(seed)
    n_random = min(int(n_random), len(remaining_for_random))
    random_pick = rng.sample(remaining_for_random, n_random) if n_random > 0 else []

    selected = best + worst + random_pick

    # Deduplicate while preserving order
    selected = list(dict.fromkeys(selected))

    # Choices should remain OCRQA-ranked, not in selection order
    choices = ranked_names

    return choices, selected


def update_newspapers(provider, query, n_best, n_worst, n_random):
    ranked = get_ranked_df(provider, query)
    choices, selected = choose_newspapers(ranked, n_best, n_worst, n_random)
    labeled_choices = [(newspaper_label(n), n) for n in choices]
    return gr.update(choices=labeled_choices, value=selected)


def make_plot(provider, selected_newspapers):
    if not selected_newspapers:
        fig = go.Figure()
        fig.update_layout(
            title="Select one or more newspapers",
            xaxis_title="Year",
            yaxis_title="Average OCRQA",
            yaxis=dict(range=[0, 1.05]),
            template="plotly_white",
            height=650,
        )
        return fig

    subset = df.copy() if provider == "All" else df[df["provider"] == provider].copy()
    subset = subset[subset["newspaper"].isin(selected_newspapers)]

    if subset.empty:
        fig = go.Figure()
        fig.update_layout(
            title="No data for the current selection",
            xaxis_title="Year",
            yaxis_title="Average OCRQA",
            yaxis=dict(range=[0, 1.05]),
            template="plotly_white",
            height=650,
        )
        return fig

    # Preserve ranking order in legend/traces
    ranked = get_ranked_df(provider, "")
    ranked_order = [
        n for n in ranked["newspaper"].tolist() if n in set(selected_newspapers)
    ]

    fig = go.Figure()

    for newspaper in ranked_order:
        dfn = subset[subset["newspaper"] == newspaper].sort_values("year")
        if dfn.empty:
            continue

        fig.add_trace(
            go.Scatter(
                x=dfn["year"],
                y=dfn["avg_ocrqa"],
                mode="markers",
                name=newspaper_label(newspaper),
                customdata=dfn[["issues", "content_items_out"]].values,
                hovertemplate=(
                    "<b>%{fullData.name}</b><br>"
                    "Year: %{x}<br>"
                    "Average OCRQA: %{y:.3f}<br>"
                    "Issues: %{customdata[0]}<br>"
                    "Content items: %{customdata[1]}"
                    "<extra></extra>"
                ),
            )
        )

    year_min = subset["year"].min()
    year_max = subset["year"].max()
    if year_max - year_min < 10:
        mid = (year_min + year_max) / 2
        year_min = int(mid - 5)
        year_max = int(mid + 5)

    provider_display = provider if provider == "All" else provider_label(provider)
    fig.update_layout(
        title=f"OCRQA by newspaper — provider: {provider_display}",
        xaxis_title="Year",
        xaxis=dict(range=[year_min - 1, year_max + 1]),
        yaxis_title="Average OCRQA",
        yaxis=dict(range=[0, 1.05]),
        template="plotly_white",
        height=650,
    )
    return fig


# -------------------------------------------------------------------
# Initial state
# -------------------------------------------------------------------

initial_provider = "All"
initial_query = ""
initial_best = 10
initial_worst = 0
initial_random = 0

initial_ranked = get_ranked_df(initial_provider, initial_query)
initial_choices, initial_selected = choose_newspapers(
    initial_ranked, initial_best, initial_worst, initial_random
)

# -------------------------------------------------------------------
# UI
# -------------------------------------------------------------------

with gr.Blocks() as demo:
    gr.Markdown("## OCR Quality Assessment exploration")
    gr.Markdown(
        "For details on how OCRQA scores are computed, see the <a"
        ' href="https://huggingface.co/spaces/impresso-project/ocrqa-demo"'
        ' target="_blank">OCRQA demo</a>.'
    )

    with gr.Row():
        provider = gr.Dropdown(
            choices=provider_options,
            value=initial_provider,
            label="Provider",
        )
        query = gr.Textbox(
            value=initial_query,
            label="Filter newspapers (case-sensitive)",
            placeholder="Type a newspaper title",
        )

    with gr.Row():
        n_best = gr.Slider(
            minimum=0,
            maximum=400,
            value=initial_best,
            step=1,
            label="Best OCRQA",
        )
        n_worst = gr.Slider(
            minimum=0,
            maximum=400,
            value=initial_worst,
            step=1,
            label="Worst OCRQA",
        )
        n_random = gr.Slider(
            minimum=0,
            maximum=400,
            value=initial_random,
            step=1,
            label="Random OCRQA",
        )

    newspaper = gr.Dropdown(
        choices=[(newspaper_label(n), n) for n in initial_choices],
        value=initial_selected,
        multiselect=True,
        label="Newspapers (filtered and ranked)",
    )

    plot = gr.Plot()

    selector_inputs = [provider, query, n_best, n_worst, n_random]

    for trigger in selector_inputs:
        trigger.change(
            fn=update_newspapers,
            inputs=selector_inputs,
            outputs=newspaper,
        )
        trigger.change(
            fn=lambda provider, newspaper: make_plot(provider, newspaper),
            inputs=[provider, newspaper],
            outputs=plot,
        )

    newspaper.change(
        fn=make_plot,
        inputs=[provider, newspaper],
        outputs=plot,
    )

    demo.load(
        fn=make_plot,
        inputs=[provider, newspaper],
        outputs=plot,
    )

demo.launch(ssr_mode=False)