app.py

import gradio as gr
import pandas as pd
from datasets import load_dataset
import numpy as np
from functools import lru_cache
import re
from collections import Counter
import editdistance

# Cache the dataset loading to avoid reloading on refresh
@lru_cache(maxsize=1)
def load_data():
    try:
        dataset = load_dataset("GenSEC-LLM/SLT-Task1-Post-ASR-Text-Correction", split="test")
        return dataset
    except Exception:
        # Fallback to explicit file path if default loading fails
        return load_dataset("parquet", 
                          data_files="https://huggingface.co/datasets/GenSEC-LLM/SLT-Task1-Post-ASR-Text-Correction/resolve/main/data/test-00000-of-00001.parquet")

# Preprocess text for better WER calculation
def preprocess_text(text):
    if not text or not isinstance(text, str):
        return ""
    text = text.lower()
    text = re.sub(r'[^\w\s]', '', text)
    text = re.sub(r'\s+', ' ', text).strip()
    return text

# N-gram scoring for hypothesis ranking
def score_hypothesis(hypothesis, n=4):
    if not hypothesis:
        return 0
    
    words = hypothesis.split()
    if len(words) < n:
        return len(words)
    
    ngrams = []
    for i in range(len(words) - n + 1):
        ngram = ' '.join(words[i:i+n])
        ngrams.append(ngram)
    
    unique_ngrams = len(set(ngrams))
    total_ngrams = len(ngrams)
    score = len(words) + unique_ngrams/max(1, total_ngrams) * 5
    return score

# N-gram ranking approach
def get_best_hypothesis_lm(hypotheses):
    if not hypotheses:
        return ""
    
    if isinstance(hypotheses, str):
        return hypotheses
    
    hypothesis_list = [preprocess_text(h) for h in hypotheses if isinstance(h, str)]
    
    if not hypothesis_list:
        return ""
    
    scores = [(score_hypothesis(h), h) for h in hypothesis_list]
    best_hypothesis = max(scores, key=lambda x: x[0])[1]
    return best_hypothesis

# Subwords voting correction approach
def correct_hypotheses(hypotheses):
    if not hypotheses:
        return ""
    
    if isinstance(hypotheses, str):
        return hypotheses
    
    hypothesis_list = [preprocess_text(h) for h in hypotheses if isinstance(h, str)]
    
    if not hypothesis_list:
        return ""
    
    word_lists = [h.split() for h in hypothesis_list]
    lengths = [len(words) for words in word_lists]
    
    if not lengths:
        return ""
    
    most_common_length = Counter(lengths).most_common(1)[0][0]
    filtered_word_lists = [words for words in word_lists if len(words) == most_common_length]
    
    if not filtered_word_lists:
        return max(hypothesis_list, key=len)
    
    corrected_words = []
    for i in range(most_common_length):
        position_words = [words[i] for words in filtered_word_lists]
        most_common_word = Counter(position_words).most_common(1)[0][0]
        corrected_words.append(most_common_word)
    
    return ' '.join(corrected_words)

# Calculate WER
def calculate_simple_wer(reference, hypothesis):
    if not reference or not hypothesis:
        return 1.0
    
    ref_words = reference.split()
    hyp_words = hypothesis.split()
    
    distance = editdistance.eval(ref_words, hyp_words)
    
    if len(ref_words) == 0:
        return 1.0
    return float(distance) / float(len(ref_words))

# Calculate WER for a group of examples with multiple methods
def calculate_wer_methods(examples, max_samples=200):
    if not examples or len(examples) == 0:
        return np.nan, np.nan, np.nan
    
    # Limit sample size for efficiency
    if hasattr(examples, 'select'):
        items_to_process = examples.select(range(min(max_samples, len(examples))))
    else:
        items_to_process = examples[:max_samples]
    
    wer_values_no_lm = []
    wer_values_lm_ranking = []
    wer_values_n_best_correction = []
    
    for ex in items_to_process:
        # Get reference transcription
        transcription = ex.get("transcription")
        if not transcription or not isinstance(transcription, str):
            continue
        
        reference = preprocess_text(transcription)
        if not reference:
            continue
        
        # Get 1-best hypothesis for baseline
        input1 = ex.get("input1")
        if input1 is None and "hypothesis" in ex and ex["hypothesis"]:
            if isinstance(ex["hypothesis"], list) and len(ex["hypothesis"]) > 0:
                input1 = ex["hypothesis"][0]
            elif isinstance(ex["hypothesis"], str):
                input1 = ex["hypothesis"]
        
        # Get n-best hypotheses for other methods
        n_best_hypotheses = ex.get("hypothesis", [])
        
        # Method 1: No LM (1-best ASR output)
        if input1 and isinstance(input1, str):
            no_lm_hyp = preprocess_text(input1)
            if no_lm_hyp:
                wer_no_lm = calculate_simple_wer(reference, no_lm_hyp)
                wer_values_no_lm.append(wer_no_lm)
        
        # Method 2: N-gram ranking
        if n_best_hypotheses:
            lm_best_hyp = get_best_hypothesis_lm(n_best_hypotheses)
            if lm_best_hyp:
                wer_lm = calculate_simple_wer(reference, lm_best_hyp)
                wer_values_lm_ranking.append(wer_lm)
        
        # Method 3: Subwords voting correction
        if n_best_hypotheses:
            corrected_hyp = correct_hypotheses(n_best_hypotheses)
            if corrected_hyp:
                wer_corrected = calculate_simple_wer(reference, corrected_hyp)
                wer_values_n_best_correction.append(wer_corrected)
    
    # Calculate average WER for each method
    no_lm_wer = np.mean(wer_values_no_lm) if wer_values_no_lm else np.nan
    lm_ranking_wer = np.mean(wer_values_lm_ranking) if wer_values_lm_ranking else np.nan
    n_best_correction_wer = np.mean(wer_values_n_best_correction) if wer_values_n_best_correction else np.nan
    
    return no_lm_wer, lm_ranking_wer, n_best_correction_wer

# Get WER metrics by source
def get_wer_metrics(dataset):
    # Group examples by source
    examples_by_source = {}
    
    for ex in dataset:
        source = ex.get("source", "unknown")
        # Skip all_et05_real as requested
        if source == "all_et05_real":
            continue
            
        if source not in examples_by_source:
            examples_by_source[source] = []
        examples_by_source[source].append(ex)
    
    # Get all unique sources
    all_sources = sorted(examples_by_source.keys())
    
    # Calculate metrics for each source
    source_results = {}
    for source in all_sources:
        examples = examples_by_source.get(source, [])
        count = len(examples)
        
        if count > 0:
            no_lm_wer, lm_ranking_wer, n_best_wer = calculate_wer_methods(examples)
        else:
            no_lm_wer, lm_ranking_wer, n_best_wer = np.nan, np.nan, np.nan
        
        source_results[source] = {
            "Count": count,
            "No LM Baseline": no_lm_wer,
            "N-best LM Ranking": lm_ranking_wer,
            "N-best Correction": n_best_wer
        }
    
    # Calculate overall metrics
    filtered_dataset = [ex for ex in dataset if ex.get("source") != "all_et05_real"]
    total_count = len(filtered_dataset)
    
    sample_size = min(500, total_count)
    sample_dataset = filtered_dataset[:sample_size]
    no_lm_wer, lm_ranking_wer, n_best_wer = calculate_wer_methods(sample_dataset)
    
    source_results["OVERALL"] = {
        "Count": total_count,
        "No LM Baseline": no_lm_wer,
        "N-best LM Ranking": lm_ranking_wer,
        "N-best Correction": n_best_wer
    }
    
    # Create flat DataFrame with labels in the first column
    rows = []
    
    # First add row for number of examples
    example_row = {"Methods": "Number of Examples"}
    for source in all_sources + ["OVERALL"]:
        example_row[source] = source_results[source]["Count"]
    rows.append(example_row)
    
    # Then add rows for each WER method with simplified names
    no_lm_row = {"Methods": "No LM"}
    lm_ranking_row = {"Methods": "N-gram Ranking"}
    n_best_row = {"Methods": "Subwords Voting"}
    
    # Add the additional methods from the figure
    llama_lora_row = {"Methods": "LLaMA-7B-LoRA"}
    nb_oracle_row = {"Methods": "N-best Oracle (o_nb)"}
    cp_oracle_row = {"Methods": "Compositional Oracle (o_cp)"}
    
    # Populate the existing methods
    for source in all_sources + ["OVERALL"]:
        no_lm_row[source] = source_results[source]["No LM Baseline"]
        lm_ranking_row[source] = source_results[source]["N-best LM Ranking"]
        n_best_row[source] = source_results[source]["N-best Correction"]
        
        # Add hardcoded values for the additional methods based on the figure
        # Default to NaN for sources not in the figure
        llama_lora_row[source] = np.nan
        nb_oracle_row[source] = np.nan
        cp_oracle_row[source] = np.nan
    
    # Add hardcoded values from the figure for each source
    # CHiME-4
    if "test_chime4" in all_sources:
        llama_lora_row["test_chime4"] = 6.6 / 100  # Convert from percentage
        nb_oracle_row["test_chime4"] = 9.1 / 100
        cp_oracle_row["test_chime4"] = 2.8 / 100
    
    # Tedlium-3
    if "test_td3" in all_sources:
        llama_lora_row["test_td3"] = 4.6 / 100
        nb_oracle_row["test_td3"] = 3.0 / 100
        cp_oracle_row["test_td3"] = 0.7 / 100
    
    # CommonVoice (CV-accent)
    if "test_cv" in all_sources:
        llama_lora_row["test_cv"] = 11.0 / 100
        nb_oracle_row["test_cv"] = 11.4 / 100
        cp_oracle_row["test_cv"] = 7.9 / 100
    
    # SwitchBoard
    if "test_swbd" in all_sources:
        llama_lora_row["test_swbd"] = 14.1 / 100
        nb_oracle_row["test_swbd"] = 12.6 / 100
        cp_oracle_row["test_swbd"] = 4.2 / 100
    
    # LRS2
    if "test_lrs2" in all_sources:
        llama_lora_row["test_lrs2"] = 8.8 / 100
        nb_oracle_row["test_lrs2"] = 6.9 / 100
        cp_oracle_row["test_lrs2"] = 2.6 / 100
    
    # CORAAL
    if "test_coraal" in all_sources:
        llama_lora_row["test_coraal"] = 19.2 / 100
        nb_oracle_row["test_coraal"] = 21.8 / 100
        cp_oracle_row["test_coraal"] = 10.7 / 100
    
    # LibriSpeech Clean
    if "test_ls_clean" in all_sources:
        llama_lora_row["test_ls_clean"] = 1.7 / 100
        nb_oracle_row["test_ls_clean"] = 1.0 / 100
        cp_oracle_row["test_ls_clean"] = 0.6 / 100
    
    # LibriSpeech Other
    if "test_ls_other" in all_sources:
        llama_lora_row["test_ls_other"] = 3.8 / 100
        nb_oracle_row["test_ls_other"] = 2.7 / 100
        cp_oracle_row["test_ls_other"] = 1.6 / 100
    
    # Calculate overall averages for the three additional methods
    llama_values = []
    nb_oracle_values = []
    cp_oracle_values = []

    for source in all_sources:
        if pd.notna(llama_lora_row[source]):
            llama_values.append(llama_lora_row[source])
        if pd.notna(nb_oracle_row[source]):
            nb_oracle_values.append(nb_oracle_row[source])
        if pd.notna(cp_oracle_row[source]):
            cp_oracle_values.append(cp_oracle_row[source])

    # Print collected values for debugging
    print(f"LLaMA values: {llama_values}")
    print(f"N-best Oracle values: {nb_oracle_values}")
    print(f"Compositional Oracle values: {cp_oracle_values}")

    # Calculate overall values - with hardcoded fallbacks 
    if llama_values:
        llama_overall = np.mean(llama_values)
    else:
        # Calculate from the table data: average of (6.6, 19.2, 11.0, 8.8, 1.7, 3.8, 14.1, 4.6) / 100
        llama_overall = 0.0873  # 8.73%
    llama_lora_row["OVERALL"] = llama_overall

    if nb_oracle_values:
        nb_oracle_overall = np.mean(nb_oracle_values)
    else:
        # Calculate from the table data: average of (9.1, 21.8, 11.4, 6.9, 1.0, 2.7, 12.6, 3.0) / 100
        nb_oracle_overall = 0.0856  # 8.56%
    nb_oracle_row["OVERALL"] = nb_oracle_overall

    if cp_oracle_values:
        cp_oracle_overall = np.mean(cp_oracle_values)
    else:
        # Calculate from the table data: average of (2.8, 10.7, 7.9, 2.6, 0.6, 1.6, 4.2, 0.7) / 100
        cp_oracle_overall = 0.0389  # 3.89%
    cp_oracle_row["OVERALL"] = cp_oracle_overall
    
    # Add rows in the desired order
    rows.append(no_lm_row)
    rows.append(lm_ranking_row)
    rows.append(n_best_row)
    rows.append(llama_lora_row)
    rows.append(nb_oracle_row)
    rows.append(cp_oracle_row)
    
    # Create DataFrame from rows
    result_df = pd.DataFrame(rows)
    
    return result_df

# Format the dataframe for display
def format_dataframe(df):
    df = df.copy()
    
    # Find the rows containing WER values
    wer_row_indices = []
    for i, method in enumerate(df["Methods"]):
        if method not in ["Number of Examples"]:
            wer_row_indices.append(i)
    
    # Format WER values
    for idx in wer_row_indices:
        for col in df.columns:
            if col != "Methods":
                value = df.loc[idx, col]
                if pd.notna(value):
                    df.loc[idx, col] = f"{value:.4f}"
                else:
                    df.loc[idx, col] = "N/A"
    
    return df

# Main function to create the leaderboard
def create_leaderboard():
    dataset = load_data()
    metrics_df = get_wer_metrics(dataset)
    return format_dataframe(metrics_df)

# Create the Gradio interface
with gr.Blocks(title="ASR Text Correction Leaderboard") as demo:
    gr.Markdown("# HyPoraside N-Best WER Leaderboard (Test Data)")
    gr.Markdown("Word Error Rate (WER) metrics for different speech sources with multiple correction approaches")
    
    with gr.Row():
        refresh_btn = gr.Button("Refresh Leaderboard")
    
    with gr.Row():
        gr.Markdown("### Word Error Rate (WER)")
    
    with gr.Row():
        try:
            initial_df = create_leaderboard()
            leaderboard = gr.DataFrame(initial_df)
        except Exception:
            leaderboard = gr.DataFrame(pd.DataFrame([{"Error": "Error initializing leaderboard"}]))
    
    def refresh_and_report():
        return create_leaderboard()
    
    refresh_btn.click(refresh_and_report, outputs=[leaderboard])

if __name__ == "__main__":
    demo.launch()