A novel approach for explicit song lyrics detection using machine and deep ensemble learning models

Introduction

In recent years, music has become one of the most prevalent forms of entertainment, recreation, and information. The digital music collection is growing rapidly, as artists express themselves through lyrics. However, some lyrics contain content that is not suitable for young listeners, such as references to sexual, violent, or drug-related themes. This can have negative effects on young people (Arnett, 1991; Ballard & Coates, 1995). For the past 40 years, various initiatives and measures have been taken to address this issue and prevent young people from being exposed to inappropriate musical content. Organizations in the music industry, such as the British Phonographic Industry (BPI) and the Recording Industry Association of America (RIAA), have worked to inform parents about the potential risks of certain lyrics and their potential effects on young people (Qamar Bhatti et al., 2018). With the increasing use of technology, particularly mobile devices, internet usage is also on the rise. According to research conducted by The Guardian (2020), most children own a mobile phone by the age of seven. These devices have become an integral part of the lives of young people. With the wide proliferation of mobile phones, screening music content has become even more important for youngsters.

Protecting children from harmful music content is a big challenge for parents. In the 1990s, the music recording industry in the United States introduced parental advisory labels for albums with lyrics that contain references to sex and violence (RIAA, 2021). These labels are intended to alert parents that the content of the album may not be suitable for children. However, the use of these labels is voluntary and not all albums contain them, so it is not always clear which music is appropriate for young listeners. It is important for parents to be aware of the potential risks and to monitor their child’s music consumption (Cole, 2010).

Censoring explicit content in songs is a complex task that raises several issues. One issue is that the results of screening can vary depending on the organization or individuals responsible for the judgment. In the past, there have been many controversies over the assessment of certain songs for broadcast (Myong, 2020). Another issue is that determining whether a song is harmful or suitable requires time and effort. Some music that might be considered inappropriate for broadcast is expressed metaphorically, making it difficult to identify without a thorough review of the lyrics. To address these challenges, it would be useful to have a way to automatically detect inappropriate content in music.

Primarily two ways exist for the automatic detection of inappropriate content in music. The first is by using a profanity dictionary. This method involves comparing the lyrics of a song to a list of words that are considered offensive or inappropriate. If any of these words are found in the lyrics, the song is classified as inappropriate. However, there is not a universal profanity dictionary, so the results of this method may vary between different companies (Bergelid, 2018). Additionally, if a word is not included in the profanity dictionary, the content will not be flagged as inappropriate. This approach also requires regular maintenance to keep the profanity dictionary up-to-date with new offensive language. Another way to automatically detect inappropriate content in music is through the use of machine learning (ML). ML has been used in various ways to filter inappropriate content, particularly in the context of online user-generated content (Xiang et al., 2012; Chen et al., 2012). However, only a few attempts are made to classify inappropriate music lyrics using a machine learning model.

Natural language processing (NLP) and ML are closely related fields and are often used together to improve the capabilities of NLP models. NLP deals with the interaction between computers and human language, while ML provides a set of algorithms and techniques for training models to make predictions or decisions (Mahedero et al., 2005). NLP tasks such as sentiment analysis, language translation, and text summarization can benefit from ML techniques (Joachims, 1998). For example, a machine learning model can be trained on a dataset of labeled text to classify sentiment as positive or negative or to automatically summarize a long document.

Music lyrics are often created to attract more fans and followers, particularly among young people. These lyrics can have a significant impact on the minds of young people. In recent years, lyrics have become increasingly violent and sexual. However, current systems for filtering explicit content in music lyrics are not effective. To address this issue, a new framework is needed that uses machine learning, ensemble learning, and deep learning to classify explicit content in music lyrics. One of the challenges of this task is that song lyrics are written in a poetic notation, which makes it difficult for NLP techniques to analyze them. This study aims at developing an automatic system for filtering explicit content in music lyrics. The key contributions of this study are as follows:

• A framework is proposed that utilizes ELSTM-VC an ensemble of machine learning models extra tree classifier (ETC) and deep learning model long short-term memory (LSTM). The objective is the automatic screening of song lyrics into explicit and non-explicit content.

• Experiments are performed to evaluate the efficacy of the proposed model in comparison to other machine learning and deep learning models. Many machine learning models are employed for this purpose including logistic regression (LR), Bernoulli Naive Bayes (BNB), random forest (RF), stochastic gradient descent (SGD), decision tree (DT), AdaBoost (AB), gradient boosting machine (GBM) and ETC. In addition, LSTM deep learning model and a voting classifier are also tested which combines LR and SGD.

• Performance is evaluated using accuracy, precision, F1-score, and recall. Furthermore, k-fold cross-validation and performance comparison with several bidirectional encoder representations from Transformers (BERT) is also carried out. Empirical results prove that the ELSTM-VC model outperforms other models and is an applicable contribution to the music industry.

The rest of the article is structured as follows. ‘Related Work’ provides a review of related research in this field. ‘Materials and Methods’ presents the methodology and detailed explanation of the models used in the experiment. Results and analysis of the experiment are given in ‘Results and Discussion’. ‘Conclusion’ discusses the implications of the research and potential areas for future works.

Related work

A literature review plays an important role in advancing research by providing a historical perspective on the specific research area. This literature review focuses on the detection of explicit content in song lyrics. The aim is to examine how previous researchers have approached the detection of explicit content in different types of media sources.

The study of using NLP methods to identify offensive content, particularly on social media, is a widely investigated research area. For instance, Davidson et al. (2017) examined the detection of hate speech, which is defined as language that is meant to express hatred towards a specific group or to demean, humiliate, or insult the members of that group. To classify the tweets into three categories the authors used the LR classifier. The system performed well overall, but had difficulty correctly identifying hate speech; 40% of hate speech is misclassified. Corazza et al. (2020) recently suggested a modular neural network design for detecting hate speech that can be applied to multiple languages such as English, Italian, and German. Their approach incorporates a variety of features, including social media network features, text-based features, FASTTEXT word embeddings, and emotion lexica. To classify the toxic comments Carta et al. (2019) proposed a supervised multi-class multi-label word embeddings approach, such as verbal bullying and personal attacks. The dataset is obtained from Wikipedia’s Talk page, and the results show that using sets of word embeddings improves performance compared to the traditional bag-of-words (BOW) model.

A few recent studies have focused on using NLP to automatically detect explicit lyrics. Chin et al. (2018) tackled the problem of detecting offensive content in Korean songs using a corpus of 27,695 Korean song lyrics, 1,024 of which are labeled as explicit (3.7% of the total). The authors established a baseline for detecting explicit content by utilizing a profanity language dictionary from Namu-wiki, a Korean wiki, and achieved a macro F1-score of 0.61 (0.88 for the weighted F1). The authors tested various machine learning classification algorithms, such as AdaBoost and Bagging, and found that Bagging performed the best, with an F1-score of 0.78 and a weighted F1 of 0.96.

Kim & Mun (2019) aimed to identify explicit content in Korean song lyrics by utilizing a corpus of 70,077 lyrics, with 10.7% labeled as explicit. They tested a lexicon-based filtering approach, using a dictionary of explicit words generated automatically. They employed a Hierarchical attention networks (HAN) model and a recurrent neural network (RNN)-based model for processing words sequentially and hierarchically. The results reveal that the combination of HAN and vector representations of explicit words produce from the lexicon performs the best, achieving an F1-score of 0.805. In the study by Rospocher (2020), the effectiveness of FastText word embedding and a classifier is tested on a large dataset of English song lyrics. The dataset consisted of 807,707 songs, with 7.74% of them labeled as explicit. The results show that FastText performs better than other methods such as LR and majority vote, proposing that using more advanced classifiers could improve performance even further.

Fell et al. (2019a) conducted a study that evaluated machine and deep learning techniques for classifying explicit lyrics, using the largest dataset to date with 179,391 English song lyrics, 17,808 of which are marked as explicit. The study compared methods including dictionary-based methods, an LR classifier, a BERT language model, and textual Deconvolution Saliency. The results show that while textual Deconvolution Saliency performs the best with an F1-score of 0.796, deep learning models did not outperform the other simpler methods. Fell et al. (2019b) conducted a follow-up study where they applied an LR classifier with term frequency-inverse document frequency (TF-IDF) BoW vector representations to a dataset of 438,000 song lyrics, achieving an F1-score of 0.773.

Lin, Tseng & Fuh (2003) proposed an explicit content detection algorithm based on a support vector machine (SVM) for classification. SVM is a technique that uses statistical learning theory to analyze, predict, and identify explicit content in images. The proposed model had a low accuracy value for explicit content detection. The study by Kia et al. (2014) focused on identifying pornographic images using a combination of techniques. The authors proposed using two specific features related to the skin region and Fourier descriptors. These features are found to improve the accuracy of identifying explicit images when compared to traditional methods. They also employed a combination of a multilayer perceptron and Neuro Fuzzy using fuzzy integral information. The system is tested and found to have a precision of 93% for true positives and 8% for false positives on the training dataset, and 87% and 5.5% on the testing dataset.

In a nutshell, the literature review shows that the detection of explicit content in song lyrics using NLP techniques is an active area of research. There are multiple applications of machine learning, deep learning, and ensemble learning models in terms of cyber security (Ashraf et al., 2022; Majeed et al., 2021), text classification (Umer et al., 2023; Karim et al., 2022), drinking water quality prediction (Juna et al., 2022; Madni et al., 2023), and digital healthcare problems (Hafeez et al., 2022; Ahmed et al., 2023). For text classification, various approaches have been proposed, including using a lexicon-based filtering approach, machine learning classification algorithms, and deep learning methods such as RNNs and CNNs. These studies have been conducted on different languages and datasets, and have achieved varying levels of performance. Some studies suggest that further improvement may be achieved by using more advanced machine learning classifiers.

Material and Methods

This section describes the architecture used to predict explicit content in music lyrics. The study employs a variety of tools, methods, and techniques to detect explicit content in song lyrics. The pipeline of the adopted methodology is shown in Fig. 1. The first step is the collection of a dataset, followed by preprocessing of the data. Finally, the classifiers used in the experiment are discussed.

Proposed methodology

The proposed methodology in this study combines the use of an ensemble of a deep neural network model and a machine learning model to detect explicit song lyrics. The use of deep neural networks has been gaining popularity among researchers in recent years as they have been shown to increase accuracy compared to traditional classifiers. Thus, a deep neural network, specifically an ELSTM-VC, which combines ETC and LSTM using soft voting, is used in this study. In the soft voting criteria, the final prediction is based on the class with the highest probability, as illustrated in Fig. 2.

Algorithm 1 presents the steps for the proposed ELSTM-VC, which combines ETC and LSTM for explicit song lyrics classification. The algorithm uses ETC and LSTM as two methods and ‘explicit’ and ‘not-explicit’ as the two classes, and the prediction is made using the following equation:

Algorithm 1:

Ensembling of ETC and LSTM (ELSTM-VC).

Input: input data $(x,y{)}_{i=1}^N$

TETC = Trained_ ETC

TLSTM = Trained_ LSTM

1: for $i=1\mathrm{t}\mathrm{o}T$ do

2: if $T_{ETC}\ne 0\mathrm{\&}{T}_{CNN}\ne 0\mathrm{\&}training\mathrm{\_}set\ne 0$ then

3:   $ProbLSTM-Exp=T_{LSTM}.probibility(Exp-class)$

4:   $ProbLSTM-NExp=T_{LSTM}.probibility(NExp-class)$

5:   $ProbETC-Explicit=T_{ETC}.probibility(Exp-class)$

6:   $ProbETC-NExp=T_{ETC}.probibility(NExp-class)$

7:  $\text{Decision}\hspace{0.17em}\text{function}\hspace{0.17em}=\hspace{0.17em}max\left(\frac{1}{N_{classifier}}{\sum }_{classifier}\left(Av{g}_{(ProbLSTM-Exp,ProbETC-Exp)},Av{g}_{(ProbLSTM-NExp,ProbETC-NExp)}\right)\right)$

8: end if

9: Return final label $\hat{p}$

10: end for

DOI: 10.7717/peerj-cs.1469/table-7

(9) $$H_{Prob}=argmax\{Exp{l}_{prob},NEx{p}_{prob}\}.$$

The proposed model, ELSTM-VC uses a combination of ETC and LSTM models to classify explicit song lyrics. The algorithm calculates the highest predicted probability Hprob using the argmax function. The joint probability of the explicit class and the Not-explicit class is represented by Exprob and NExprob, respectively, and are computed using the following equation

(10) $$Ex{p}_{prob}=\frac{ExpPro{b}_{ETC}+ExpPro{b}_{LSTM}}{2}$$

(11) $$NEx{p}_{prob}=\frac{NExpPro{b}_{ETC}+NExpPro{b}_{LSTM}}{2}$$where $ExpPro{b}_{ETC}$, and $ExpPro{b}_{LSTM}$ are the probability for the explicit class by ETC and LSTM respectively, on the other hand, $NExpPro{b}_{ETC}$, and $NExpPro{b}_{LSTM}$ are the non-explicit class probability score for ETC and LSTM. For the detailed description of the voting classifier, the values are obtained from the data sample. ETC and LSTM give the probabilities for the sample are

$ExpPro{b}_{ETC}$ = 0.7

$NExpPro{b}_{ETC}$ = 0.5

$ExpPro{b}_{LSTM}$ = 0.6

$NExpPro{b}_{LSTM}$ = 0.4

$Ex{p}_{Prob}$ and $NEx{p}_{Prob}$ is calculated as

(12) $$Ex{p}_{Prob}=\frac{0.7+0.6}{2}$$

(13) $$NEx{p}_{Prob}=\frac{0.5+0.4}{2}$$

Then probability values are passed to the argmax function, which returns the highest probability class.

(14) $$H_{Prob}=argmax(0.65,0.45)$$

In this case, the highest predicted probability is for the explicit class, so the final prediction will be for the explicit class. The architecture of the proposed approach is illustrated in the above figure. The research also uses various machine learning models that use Word2vec as a feature representation technique as a baseline for comparison with the proposed ELSTM-VC model. The dataset is divided into two parts with a 20% and 80% ratio for testing and training, respectively. The training set is used to train all machine learning models and the proposed ELSTM-VC. The performance of all models is then evaluated using the test data.

Results and discussion

In this research, an ensemble method combining a deep learning model and a machine learning model is proposed. The performance of this ensemble is compared to other base models in terms of efficiency. The classifiers are evaluated using performance metrics such as accuracy, precision, recall, and F1-score. In this research, for the detection of the explicit song lyrics tree-based and regression-based models are used individually and in an ensemble, and deep learning models are also used individually and in an ensemble with machine learning models are used.

Results of machine and deep learning models

Table 4 shows the results of employed models regarding accuracy, precision, recall, and F1-score. Among the machine learning models, RF and ETC achieve the highest accuracy of 91% and perform well in detecting explicit song lyrics.

Table 4:

Comparison of machine learning and deep learning models in terms of precision, recall, F1-score.

Models	Accuracy	Precision	Recall	F1-score
DT	75.34%	80.25%	75.37%	77.41%
AB	75.19%	80.45%	75.67%	77.54%
LR	83.33%	69.49%	83.72%	76.58%
SGD	75.37%	80.29%	75.31%	77.55%
RF	91.17%	92.85%	92.37%	90.35%
GBM	75.36%	80.82%	75.74%	77.69%
ETC	91.70%	92.38%	92.19%	90.52%
BNB	83.45%	69.65%	83.21%	76.44%
CNN	88.34%	82.68%	85.37%	83.62%
LSTM	92.28%	100.00%	89.17%	91.31%
Voting classifier (LR+SGD)	83.74%	69.59%	83.47%	76.34%
BLSTM-VC	91.27%	93.29%	94.55%	94.27%
RLSTM-VC	93.19%	95.68%	96.29%	95.98%
ELSTM-VC	96.92%	100.00%	92.97%	96.79%

DOI: 10.7717/peerj-cs.1469/table-4

Note:

The proposed model is indicated in bold.

Experimental results in Table 4 show that the proposed ELSTM-VC and LSTM models alone achieved a precision of 100%, while the ETC model alone has a precision of 92.38%. The reason for making an ensemble of ETC and LSTM is that they both have the highest accuracy as individual models. In the employed machine learning models, ETC has the best results while deep LSTM has the best results among deep learning models. These models are joined as an ensemble and good accuracy is expected. For further confirmation, other ensemble models are also designed involving the second and third best-performing machine learning models with LSTM but the results are not higher than the ELSTM-VC model. The highest recall value is achieved by the ETC and RF models; both have 92% recall. The LSTM model has the second-highest recall value which is 89%. The LSTM model also has the highest F1-score of 91%, while the RF and ETC models also performed well with an F1-score of 90%. However, the best accuracy and precision are obtained by the proposed model which achieves a 96.92% accuracy which is far better than all other models employed in this study.

Among tree-based algorithms, the decision tree has the worst performance with an accuracy of 75% as shown in Table 4. Decision trees often struggle to handle diverse features and tend to overfit. However, advanced tree-based models like ETC and RF show improved stability and performance. This is why the ETC and RF classifiers perform well in detecting explicit music lyrics.

The deep learning models use a technique called word embedding. LSTM using word embedding has higher results than other machine learning classifiers. This suggests that word embedding has a significant impact on detecting explicit lyrics in songs. The proposed ELSTM-VC outperforms all other models with 96% accuracy, 100% precision, 95% recall, and 97% F1-score. The proposed approach which is an ensemble of a machine learning and deep learning model performs better using Word2vec features which are the highest of all classifiers. A comparison of the results of all machine learning models and the deep learning model is illustrated in Fig. 3.

Comparison with encoding-decoding models

To compare the proposed strategy with the most recent models, this study applied BERT, PHS-BERT, and BioALBERT models. Using a sizable dataset of 3.3 billion words, the success of BERT has been analyzed. BERT is trained on the 800 million words BooksCorpus of Google and the 2.5 billion word Wikipedia. BERT is unique from other conventional models since it can read simultaneously in both directions. Bidirectionality is the term for this capacity, which was made possible with the help of transformers. BERT is outperforming in different 11 NLP tasks. PHS-BERT is proposed for tasks related to public health surveillance and has proved its robustness (Naseem et al., 2022). BioALBERT 1.1 is a large model that is trained on biomedical corpus and outperforms other models on many datasets (Naseem et al., 2021). We also added a comparison with state-of-the-art models like convolutional neural networks (CNN) and hierarchical attention networks (HAN) from Hameed et al. (2021), Umer et al. (2021).

Results of the transformer-based models are given in Table 6 and in Fig. 4. It can be observed that all three models have achieved an accuracy greater than 92%. BioALBERT has attained the highest accuracy with 94.18% which is closest to the proposed model. If we compare both models in terms of complexity, it can be noticed that BioALBERT is a multilayer complex deep learning model, and is trained on millions of parameters that require high computational power. Still, the accuracy of transformer-based models is not greater than the proposed model. Transformer-based models did not perform well because transformer models are computationally expensive and require large amounts of data to train effectively. Owing to comparatively smaller training data, the performance of transformer-based models is not good enough. The transformer model struggles to generalize and performs poorly on new examples. Contrarily, the proposed model is an ensemble-learning model and can be executed on lower-power machines. Thus the proposed model is more suitable for performing explicit song lyrics detection. The performance of CNN is not good but HAN performs second best in terms of accuracy.

Table 6:

Comparison with transformer-based models.

Model	Accuracy	Precision	Recall	F1-score
BERT	92.24%	90.17%	92.34%	91.44%
PHS-BERT	93.65%	91.25%	94.68%	93.67%
BioALBERT	94.18%	92.77%	95.49%	93.19%
CNN	88.34%	82.68%	85.37%	83.62%
HAN	94.08%	94.67%	95.19%	94.86%
ELSTM-VC	96.92%	100.00%	92.97%	96.79%

DOI: 10.7717/peerj-cs.1469/table-6

Note:

The proposed model is indicated in bold.

Conclusion

Explicit song lyrics detection is becoming increasingly important to detect references to sex, violence, and drug-related themes. Automated explicit content detection methods are needed to cope with the large use of mobile phones. This research work presents an ensemble model that utilized ETC and LSTM for detecting explicit lyrics. Experiments are performed on a large dataset employing many machine-learning models. Results reveal that the proposed model ELSTM-VC achieves a 96.0% accuracy which is the highest compared to other machine learning models. The results also indicate that the proposed model performs better than other state-of-the-art models like BERT, PHS-BERT, and BioALBERT for song explicit lyrics detection. However, it is important to note that the computational complexity of the proposed approach is higher than standalone machine learning models which will be a focus of future research. Future work will also involve analyzing important features using dimensional reduction techniques and testing other deep learning models for explicit lyrics detection.