Methodology

By -


Data collection

Firstly we did a problem identification survey via Google Form. Then data collection began for our project to create the models to detect the offensive content in Sinhala, Tamil and English languages. Detecting profanity in audio signals for English was convenient, owing to the presence of sample tools and datasets for English. However, since Sinhala and Tamil languages are rather regional, not many technologies are supported in terms of the specified languages. Therefore, the project demanded most of the NLP components to be created from scratch or to employ alternative approaches to tackle the issue. Since our project concerns Natural Language Processing, social media posts and comments were chosen as the source of raw data, since there was a requirement of data with the presence of heavy colloquial language. This ensured the detection of profanity, regardless of whether the audio file contains formal speech or casual speech form. A list of offensive keywords was formed for all three languages to manually collect the social media posts and comments from mainly Twitter and Facebook. Initially, using the Twitter API to mine the raw data was considered. But, Twitter API requires a premium package to access data older than 7 days. Hence, it was discarded and the team collected the data using the keywords in social media manually.


Preprocessing of data

The comments were collected from social media posts and assigned a class considering the presence of offensive keywords. A different approach was carried out in the data collection for the Multi-class classification model, since it concerns the prediction of the offensive nature of each word in a sentence. Therefore, the data corpus for this purpose was ensured to contain short snippets from social media comments. The data corpora were split into training datasets randomly with 70% of the corpus and the rest was used as testing data set for model evaluation. The data corpus was preprocessed by removing stopwords, numbers, URLs, punctuations, special characters and duplicates. The stopwords list for Sinhala and Tamil was inspired from NLTK’s stopwords for English and it included colloquial stopwords as well.


Building an evaluation model

The preprocessed datasets were used to generate the features for the Machine Learning algorithms. The vectorization and classification tasks were conducted using Scikit-learn library with regards to feature extraction, we tested the datasets with TF-IDF vectorizer and Count vectorizer along with the classifiers. The SVM classifier is known to outperform other classifiers as k-Nearest Neighbors and Naive Bayes. Hence, for the binary classification models, SVM was chosen to proceed with. Meanwhile, for the multi-class classification models, classifiers as Decision Tree, Multinomial Naive Bayes, Logistic Regression and SVM were experimented to determine which offers the most-suitable model.

The secondary filtering model was used to identify the offensive nature of each word in a sentence. This involves a data set of short snippets of text rather than lengthy sentences, where the performance of SVM is relatively lower than that of variants of NB (Naive Bayes). Regarding this, a combination of classifiers and vectorizers were used to identify the most-suitable model.After the models were finalized, the hyperparameters of the classifiers and vectorizers were tuned for maximum accuracy using the pipelining method.The system was created using python GUI library PyQt5. In the backend the system has four main modules: DSP module, Speech recognition module, NLP module, and Audio replacing module.



Finally testing of the system was done with a set of few people of age 25-30 and found overall positive response on their experience with the system.

Comments

Post a Comment

Popular posts from this blog

Conclusion | Automatic Audio Replacement of Objectionable Content for Sri Lankan Locale

By -
Image
When considering Sri Lanka, the government had to temporarily ban a few social media platforms several times from 2018 to 2019 to control the violence between the communities which was accelerating via hate speech and racist fake news spreading through social media. Yet, there is no proper mechanism to specifically filter and replace objectionable content in audio for Sri Lankan locale. In addressing this issue, we proposed a system for Sri Lankan locale, which automatically detects and replaces objectionable content in audio. In this proposed system, to selectively filter out the potentially objectionable audio content, the input audio is first preprocessed and converted into text format. The presence of racist, cursing, and sexist objectionable content are detected along with their corresponding locations and timestamps through the filtering mechanism. Afterwards, the detected objectionable content is seamlessly replaced with predetermined audio input. The model was tested against a ...
Read more

User Interface design

By -
Image
    User Interface gives the users the first impression of the system.  Hence it was developed with a proper understanding of the user goals, skills, preferences, and tendencies. Our Software system will mainly target the editors in the media industry, yet this can be used by whoever wants to automatically replace objectionable content from an audio clip. When we are addressing software that aims the general public,we considered that this software is accessible for almost every one of them. People of different age groups, with different interests, will be using this software.     We created simple UI, along with a decent set of formatting. Our GUIs only contain the most necessary elements. The language used within the UI is simple, and full of clarity, in order to improve the ease of use of the users. When catering to the needs of a greater community of users, it is important to have a good communication system between the user and the software. All the ...
Read more

System Evaluation

By -
Image
After the models were finalized, the hyperparameters of the classifiers and vectorizers were tuned for optimum accuracy using the pipelining method in Scikit-learn. The classification reports of final chosen models for Sinhala, Tamil and English are depicted in TABLE XV. The performance of the finalized binary classification models was evaluated using the Receiver Operating Characteristic (ROC) Curve in Fig 3, Fig 4. and Fig 5, which display the graphical relationships among the metrics shown in TABLE XIV. With the use of ‘Area Under the Curve’ measure, from Fig 1, Fig 2. and Fig 3, it can be observed that the performance of the classifiers is greater than the classifiers with no power. Fig. 1. Receiver Operating Characteristic Curve of the model for Sinhala. Fig. 2. Receiver Operating Characteristic Curve of the model for Tamil.                               Fig. 3.  Receiver Operating Characteristic Curve ...
Read more