song_matching.py

import pandas as pd
from sentence_transformers import SentenceTransformer, util

class SongMatcher:
    def __init__(self, songs_data_file, model_name="sentence-transformers/all-mpnet-base-v2"):
        """
        Initializes the SongMatcher with the songs data file and the SentenceTransformer model.
        :param songs_data_file: Path to the CSV file containing songs data
        :param model_name: Name of the SentenceTransformer model
        """
        self.songs_df = pd.read_csv(songs_data_file)
        self.sim_model = SentenceTransformer(model_name)

    def match_songs_with_sentiment(self, user_sentiment_label, user_sentiment_score, user_input, score_range=0.00625):
        """
        Matches songs from the dataset with the user's sentiment.
        :param user_sentiment_label: The sentiment label of the user input
        :param user_sentiment_score: The sentiment score of the user input
        :param user_input: Text input from the user
        :param score_range: Range for filtering songs based on sentiment score
        :return: DataFrame of top 5 matched songs
        """
        # Filter songs with the same sentiment label
        matched_songs = self.songs_df[self.songs_df['sentiment'] == user_sentiment_label]

        # Calculate the score range
        score_min = max(0, user_sentiment_score - score_range)
        score_max = min(1, user_sentiment_score + score_range)

        # Further filter songs whose scores fall within the specified range
        matched_songs = matched_songs[(matched_songs['score'] >= score_min) & (matched_songs['score'] <= score_max)]

        # Compute similarity between user input and song lyrics
        input_vector = self.sim_model.encode(user_input)
        matched_songs['similarity'] = matched_songs['seq'].apply(lambda x: util.pytorch_cos_sim(self.sim_model.encode(x), input_vector))

        # Select the top five songs based on similarity and return
        top_5 = matched_songs.nlargest(5, 'similarity')
        return top_5[['song', 'artist', 'seq', 'similarity', 'sentiment', 'score']]