src/substitutions_visualizer.py

import pandas as pd
import streamlit as st
from jiwer import process_words

from visual_eval.evaluator import extract_substitution_samples, HebrewTextNormalizer

subs_table_styles = """
<style>
    .sub-table {
        background: white;
        width: 100%;
        border-collapse: collapse;
        color: black;
    }
    .sub-row {
        cursor: pointer;
        transition: all 0.2s;
    }
    .sub-row .txt {
        text-align: center;
    }
    .sub-row:nth-child(even):hover {
        background: #eee;
    }
    .sub-row:nth-child(odd):hover {
        background: #eee;
    }
    .sub-row:nth-child(even):hover + .sub-row {
        background: #eee;
    }
    .sub-row:nth-child(even):has(+ .sub-row:hover) {
        background: #eee;
    }
    .sub-row.ref {
        color: green;
    }
    .sub-row.ref .ctx {
        text-align: end;
    }
    .sub-row.hyp {
        color: red;
        border-bottom: 1px solid black;
    }
    .sub-row.hyp .ctx {
        text-align: start;
    }
</style>
"""


@st.cache_data
def visualize_substitutions(ref, hyp):
    norm = HebrewTextNormalizer()
    wer_word_output = process_words(norm(ref), norm(hyp))
    subs_rows = []
    for sample in extract_substitution_samples(wer_word_output):
        subs_rows.append(
            {
                "ref": " ".join(sample.ref),
                "hyp": " ".join(sample.hyp),
                "hyp_ctx": " ".join(
                    wer_word_output.hypotheses[0][slice(*sample.hyp_context_span)]
                ),
                "ref_ctx": " ".join(
                    wer_word_output.references[0][slice(*sample.ref_context_span)]
                ),
            }
        )

    sub_rows_html = []
    for row in subs_rows:
        sub_rows_html.append(
            f"""
            <tr class="sub-row ref">
                <td class="ctx">{row['ref_ctx']}</td>
                <td class="txt">{row['ref']}</td>
                <td></td>
            </tr>
            <tr class="sub-row hyp">
                <td></td>
                <td class="txt">{row['hyp']}</td>
                <td class="ctx">{row['hyp_ctx']}</td>
            </tr>
            """
        )

    st.subheader("Substitutions List")

    table_html = f"""
{subs_table_styles}
<table class="sub-table" dir="rtl" lang="he">
    <tr>
        <th style="text-align: end;">Ref Context</th>
        <th style="text-align: center;">Ref/Hyp</th>
        <th style="text-align: start;">Hyp Context</th>
    </tr>
    {"".join(sub_rows_html)}
</table>
"""

    st.html(table_html)