title = 'EXPLOITING SYNCHRONIZED LYRICS AND VOCAL FEATURES FOR MUSIC EMOTION DETECTION'
abstract = "One of the key points in music recommendation is authoring engaging playlists according to sentiment and emotions. While previous works were mostly based on audio for music discovery and playlists generation, we take advantage of our synchronized lyrics dataset to combine text representations and music features in a novel way; we therefore introduce the Synchronized Lyrics Emotion Dataset. Unlike other approaches that randomly exploited the audio samples and the whole text, our data is split according to the temporal information provided by the synchronization between lyrics and audio. This work shows a comparison between text-based and audio-based deep learning classification models using different techniques from Natural Language Processing and Music Information Retrieval domains. From the experiments on audio we conclude that using vocals only, instead of the whole audio data improves the overall performances of the audio classifier. In the lyrics experiments we exploit the state-ofthe-art word representations applied to the main Deep Learning architectures available in literature. In our benchmarks the results show how the Bilinear LSTM classifier with Attention based on fastText word embedding performs better than the CNN applied on audio. "
text = "EXPLOITING SYNCHRONIZED LYRICS AND VOCAL FEATURES FOR \nMUSIC EMOTION DETECTION \n\nABSTRACT \nOne of the key points in music recommendation is authoring engaging playlists according to sentiment and emotions. While previous works were mostly based on audio for music discovery and playlists generation, we take advantage of our synchronized lyrics dataset to combine text representations and music features in a novel way; we therefore introduce the Synchronized Lyrics Emotion Dataset. Unlike other approaches that randomly exploited the audio samples and the whole text, our data is split according to the temporal information provided by the synchronization between lyrics and audio. This work shows a comparison between text-based and audio-based deep learning classification models using different techniques from Natural Language Processing and Music Information Retrieval domains. From the experiments on audio we conclude that using vocals only, instead of the whole audio data improves the overall performances of the audio classifier. In the lyrics experiments we exploit the state-ofthe-art word representations applied to the main Deep Learning architectures available in literature. In our benchmarks the results show how the Bilinear LSTM classifier with Attention based on fastText word embedding performs better than the CNN applied on audio. \n\n1. INTRODUCTION \nMusic Emotion Recognition (MER) refers to the task of finding a relationship between music and human emotions [24,43]. Nowadays, this type of analysis is becoming more and more popular, music streaming providers are finding very helpful to present users with musical collections organized according to their feelings. The problem of Music Emotion Recognition was proposed for the first time in the Music Information Retrieval (MIR) community in 2007, during the annual Music Information Research Evaluation eXchange (MIREX) [14]"

ranker = KeyphrasesRanker()
set(ranker.extract_candidate_chunks(text))
{'100-dimensional word2vec',
 '] achieves relevant results',
 '] exploit contextual information',
 '] speech',
 '] techniques',
 'account',
 'acculturation',
 'acoustical properties',
 'adjectives',
 'advantage',
...

ranker = KeyphrasesRanker()
set(ranker.extract_candidate_words(text))
{'100-dimensional',
 'abstract',
 'account',
 'acculturation',
 'achieves',
 'acoustic',
 'acoustical',
 'adjective',
 'adjectives',
 'adjectives/tags/labels',
 'advantage',
 'aforementioned',
 'agents',
 'aim',
 'allocation',
 ...

ranker = KeyphrasesRanker()
texts = [title, abstract, text]
corpus, corpus_tfidf, dictionary = ranker.score_keyphrases_by_tfidf(texts, candidates='chunks')
d = {dictionary.get(id): value for doc in corpus_tfidf for id, value in doc}
print(json.dumps(d,indent=4))
{
   "exploiting synchronized lyrics and vocal features for music emotion detection": 1.0,
   "advantage": 0.02818991144288086,
   "audio": 0.12685460149296385,
   "audio classi\ufb01er": 0.01409495572144043,
   "audio for music discovery": 0.01409495572144043,
   "audio samples": 0.01409495572144043,
   "audio-based deep learning classi\ufb01cation models": 0.01409495572144043,
   "benchmarks": 0.01409495572144043,
   "bilinear lstm classi\ufb01er with attention": 0.01409495572144043,
   "cnn": 0.02818991144288086,
   "comparison": 0.01409495572144043,
   "data": 0.01409495572144043,
   "different techniques from natural language processing": 0.01409495572144043,
   "emotions": 0.098664690050083,
   "experiments on audio": 0.01409495572144043,
   "fasttext word": 0.01409495572144043,
...

ranker = KeyphrasesRanker()
ranker.score_keyphrases_by_textrank(text)
 [('audio', 0.020772356010959545),
 ('audio features', 0.015940980040515616),
 ('music', 0.015056242947485416),
 ('lyrics', 0.014441885556177585),
 ('music emotion', 0.014074422805978502),
 ('emotion', 0.013092602664471589),
 ('music features', 0.01308292350877855),
 ('audio data', 0.012708279325535445),
 ('music information', 0.01175624019437015),
 ('emotions', 0.011338641233589263),
 ('features', 0.011109604070071685),
 ('music emotion classi fication', 0.010623139495342838),
 ('musical lyrics', 0.010621753673476389),
 ...

ranker = KeyphrasesRanker()
candidates = ranker.extract_candidate_words(text)
candidates = candidates[0:5]
candidate_features = ranker.extract_candidate_features(candidates, text, abstract, title)
candidate_features = [{k[0]:k[1]} for k in sorted(ranker.candidate_features.items(), key=lambda item: item[1]['term_count'], reverse=True)]
print(json.dumps(ranker.candidate_features,indent=4))
[
    {
        "lyrics": {
            "term_count": 26,
            "term_length": 1,
            "max_word_length": 6,
            "spread": 0.9258158185840708,
            "lexical_cohesion": 0.0,
            "in_excerpt": 1,
            "in_title": 1,
            "abs_first_occurrence": 0.00165929203539823,
            "abs_last_occurrence": 0.9274751106194691
        }
    },
    {
        "features": {
            "term_count": 12,
            "term_length": 1,
            "max_word_length": 8,
            "spread": 0.9496681415929203,
            "lexical_cohesion": 0.0,
            "in_excerpt": 1,
            "in_title": 1,
            "abs_first_occurrence": 0.00283462389380531,
            "abs_last_occurrence": 0.9525027654867256
        }
    },
    {
        "synchronized": {
            "term_count": 11,
            "term_length": 1,
            "max_word_length": 12,
            "spread": 0.8889657079646018,
            "lexical_cohesion": 0.0,
            "in_excerpt": 1,
            "in_title": 1,
            "abs_first_occurrence": 0.0007605088495575221,
            "abs_last_occurrence": 0.8897262168141593
        }
    },
...

textRank = TextRank4Keyword()
textRank.analyze(text, candidate_pos = ['NOUN', 'PROPN'], window_size=4, lower=False)
print(textRank.get_keywords(10))

ranker = EntitiesRanker()
ranks = ranker.main(text)
print(ranks)