Department of Computer Science
Permanent URI for this communityhttps://repository.cuilahore.edu.pk/handle/123456789/16
Browse
9 results
Search Results
Item Developing a Speech Recognition System for the Urdu Language(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2021) Muhammad Ibtesam Arshad; FA17-BSE-061; Dr. Rao Muhammad Adeel Nawab; LHR TP 7015Automatic Speech Recognition Systems (ASR) are used to convert the acoustic signals which are caught through the microphones into the sequence of words. The Automatic Speech Recognition Systems (ASR) empower the machines to react correctly, reliably, and effectively to human speech or voice and offer helpful and important services to the users. The interaction with computers or gadgets is faster, simpler, and easier, through voice or speech instead of typing through the keyboard or console, so the people will prefer Speech Recognition Systems. Speech Recognition Systems will facilitate the users in workplaces, in marketing, in banking, in health care, in language learning, and in the education field also. This project aims to design and the implementation a Speech Recognition System for the Urdu Language. In this first phase, we have collected data for the Urdu Language. To build a general Speech Recognizer, a huge amount of data is needed. In the second phase, we have implemented our System. In the final phase, we have trained and tested our Model. This model is Similar to Deep Speech 2 Models. The evaluation is carried out using Word Error Rate and Character Error Rate.Item Cross-genre Multi-label Emotion Classification on Mono-lingual and Code-mixed Texts(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Ahmad Mahmood; LHR TP 8472; FA21-RCS-008; Dr. Rao Muhammad Adeel NawabThis research focuses on This research focuses on Emotion Classification (EC) in textual data and explores the problem of Cross-genre Multi-Label Emotion Classification (CGMLEC). EC involves detecting the emotional attitude conveyed by text, while Multi-Label Emotion Classification aims to assign relevant emotion labels that accurately reflect the author's state of mind. In previous studies most of the work has been done using Same-genre Multi-label Emotion Classification [2], [14]–[16]. Whereas there is only one article that solves the problem of CGMLEC, but the language is the same. However, the problem of CGMLEC has not explored using mono-lingual (English tweets) and code-mixed (SMS messages). Secondly, this research study has Developed, Applied, Evaluated, and Compared Classical Machine Learning, Deep Learning and Transfer Learning based methods. Whereas in Transfer Learning methods we have implied 5 distinct sentence transformers (all-distilroberta-v1, all-MiniLM-L12-v2, all-mpnet-base-v2, facebook-dpr-question_encoder-single-nq-base, and LaBSE), Also the Feature Fussion based approach is applied which combines the features of all the transformers and then the Machine Learning models are applied. The proposed transformers-based approach outperforms with an F1 score of 0.3333 using all-distilroberta-v1 transformer. ) in textual data and explores the problem of Cross-genre Multi-Label Emotion Classification (CGMLEC). EC involves detecting the emotional attitude conveyed by text, while Multi-Label Emotion Classification aims to assign relevant emotion labels that accurately reflect the author's state of mind. In previous studies most of the work has been done using Same-genre Multi-label Emotion Classification [2], [14]–[16]. Whereas there is only one article that solves the problem of CGMLEC, but the language is the same. However, the problem of CGMLEC has not explored using mono-lingual (English tweets) and code-mixed (SMS messages). Secondly, this research study has Developed, Applied, Evaluated, and Compared Classical Machine Learning, Deep Learning and Transfer Learning based methods. Whereas in Transfer Learning methods we have implied 5 distinct sentence transformers (all-distilroberta-v1, all-MiniLM-L12-v2, all-mpnet-base-v2, facebook-dpr-question_encoder-single-nq-base, and LaBSE), Also the Feature Fussion based approach is applied which combines the features of all the transformers and then the Machine Learning models are applied. The proposed transformers-based approach outperforms with an F1 score of 0.3333 using all-distilroberta-v1 transformer.Item Text Reuse Detection for the Urdu Language at Lexical Level(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2023) Ayesha Noreen; FA21-RCS-026; LHR TP 8461; Dr. Rao Muhammad Adeel NawabText reuse is a process of creating new text(s) from pre-existing text(s). In recent years, Urdu Text Reuse Detection (UTRD) has gained the attention of researchers because the text is readily available in digital format all over the internet and can be copied or paraphrased from another source without proper attribution, which makes it easier to reuse but hard to detect. In previous studies, the problem of UTRD has been explored at the sentence level, sentence/passage level, and document level, along with benchmark corpora and methods. However, the problem of UTRD has not been explored at the Lexical level with respect to corpora and methods. To fulfill this research gap, this research study has developed a large benchmark manually annotated corpus of 22,184 text pairs at two levels of rewrite: 1) Derived = 8,660 and 2) Non-Derived = 13,524. Secondly, this research study has developed, applied, evaluated, and compared baseline approaches (Uni-gram Overlap and Word Embedding-based approaches) with proposed transformer-based approaches on the proposed UTRD-Lex-22 corpus. The proposed transformers-based approach (using a combination of 7 different Sentence Transformers including: distiluse-base-multilingual-cased-v1, LaBSE, paraphrase-multilingual-MiniLM-L12-v2, paraphrase-multilingual-mpnet-base-v2, xlm-r-100langs-bert-base-nli-mean-tokens, xlm-r-bert- base-nli-stsb-mean-tokens, xlm-r-distilroberta-base-paraphrase-v1) outperforms with an F1 score of 0.69 compared to the best results obtained using feature fusion approach (F1 = 0.70).Item Cross-Lingual Abstractive Text Summarization for English-Urdu Language Pair(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2021) Duaa-e-Nadeem; SP20-RCS-001; LHR TP 7600; Dr. Rao Muhammad Adeel NawabCross-Lingual Abstractive Text Summarization (CLATS) is a process to generate an abstractive summary in Language B (target language) of the given document in Language A (source language). CLATS has the potential to allow cross-lingual information access for millions of people around the world who do not speak or understand languages with a broad online presence and it makes documents originally written in native languages available to a wide audience. The popularity of digital documents has resulted in a rapid increase in the amount of information available online. This massive volume of data, however, cannot be manually analyzed. A summary of a big written document allows individuals to quickly grasp the topic's material without reading the entire document. CLATS plays a vital role in news summarization, literature summarization, timeline/event summarization and summarization of Wikipedia or any information source articles for better understanding. This research aims at exploring the problem of CLATS. Previously, CLATS systems have been developed for different language pair like Chinese-English, German-English, Arabic-English and vice versa but this problem is not explored for English-Urdu language pair, even though Urdu is a widely spoken South Asian language. The twofold aim of this research is to fill the gap: (1) develop a large benchmark corpus CLATS for English-Urdu language pair and (2) develop, apply, evaluate and compare state-of the-art deep learning techniques on proposed corpus for CLATS for English-Urdu language pair. For developing CLATS-EU corpus, News Summary and All the News are used as base corpus. The following steps were taken to create the proposed corpus: (1) automatic translation of base corpus (2) manual inspection and correction of testing data of base corpus (3) corpus standardization After development of CLATS-EU corpus, Deep learning sequence to sequence techniques (RNN, LSTM, GRU, BRNN, BLSTM and BGRU) are developed by using CLATS-EU corpus. Then RNN, LSTM and GRU models are implemented by applying attention mechanism. ROUGE1, ROUGE2 and ROUGEL are used as evaluation x measures to evaluate the performance of the techniques that we have developed. For CLATS, GRU gave the best results on our proposed CLATS-EU corpusItem Developing a Urdu Lemmatizer Using Dictionary Look-up Approach(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2020) Saima Shaukat; FA18-RCS-009; LHR TP 5987; Dr. Rao Muhammad Adeel NawabLemmatization is the process of finding the root form of a word. Lemmatization is used as preprocessing step for most of the NLP applications. Lemmatization is im portant part of NLP applications which involve the use of root form instead of using derivational forms of words. Text mining, information retrieval systems, document clustering, search engines and text categorizations are few of the applications where lemmatization process is used. Urdu is a widely spoken language in the world, but very less work has been done on developing basic NLP tools for this language, one of them is Urdu lemmatizer. Since Urdu is a morphologically rich language and has words with many inflectional and derivational forms, development of an efficient lemmatizer is a challenging task and it will be useful for many Urdu NLP applications. These challenges help us to choose the customized and more appropriate approach for Urdu lemmatization which is dictionary-based approach. The main aim of this research work is to develop an Urdu lemmatizer. To develop the efficient Urdu lemmatizer, we used dictionary look-up approach. Majority of exist ing lemmatizers are of in English and other European languages. Most of the existing lemmatizers are rule based and also the information about parts of speech tags is ig nored. Development of large dictionaries for Urdu Language is also missing in existing work. To fulfill this gap, the main goals of the study are: (1) Develop a large bench mark corpus for the Urdu language and (2) Explore the relationship between parts of speech tag and lemmatizer (3) Developed the standard approach for Urdu lemmatizer. The first major contribution of this study is the development of a large benchmark Urdu dictionary. The proposed corpus was developed in the following steps: (1) In the first step we collected data from two different sources: urmono corpus and Wikipedia dump (2) Preprocess and tokenize the collected data (3) extract the frequency of each word using python script (4) Selection of most frequent words (5) Assign parts of speech ix tags to selected words (6) manually annotate data (assign lemma to each word) (7) De velop a benchmark corpus and save in a standard csv format. Second major contribution of this research work is exploring the relationship be tween POS tag and lemma of a word. Train the POS tagger on collected Wikipedia dump and then used in the development of Urdu lemmatizer because user can enter input with or without POS tags. If user will enter input without POS tag, tagger will assign most frequent used tag to the word before generating lemma of a word. As a third major contribution of this research work, proposed a dictionary-based approach for Urdu lemmatizer. As the Urdu is morphological rich language; it also includes many Persian, Turkish and Arabic words. So, for the efficient lemmatizer, dictionary-based approach is more suitable. The evaluation was carried out using two techniques (1) With-POS-DLA (2) Without-POS-DLA. We achieved accuracy of 66.79% and 76.44% respectively. The best result is achieved by using Without-POS DLA technique.Item Author Region Identification for Urdu Language(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2020) Rehman Akram Chughtai; FA18-RCS-007; LHR TP 6407; Dr. Rao Muhammad Adeel NawabAuthor region identification is a subtask of author profiling which intends to identify the region of the author, based on the written text. Identification of author region may improve content recommendation, security, reduction in cyber-crimes due to its different applications in fake profile detection, content recommendation, sales and marketing, and forensic linguistics. Lot of work has been done on several author profiling tasks in English, Arabic and other European languages but no significant work found on author region identification task using Urdu language. To fill this gap, we proposed a large benchmark corpus and methods to predict author’s region using Urdu language articles. The proposed corpus consists of 12,100 Urdu articles written by authors from different regions. These articles are categorized into three different regions i.e. Pakistan, India and others. There are 5000 Urdu articles for Pakistan region, 5000 Urdu articles for India region and 2100 articles for Others category. For each article, article web links, author name and domain of each article is mentioned for future research purposes. To demonstrate how our proposed corpus can be used for the development and evaluation of author region identification for Urdu, I applied different classical machine learning and deep learning approaches. For classical machine learning, I used word n-grams technique for feature extractions and then applied classifiers i.e. Logistic Regression, SVM, Naïve Bayes, Random Forest, Adaboost and Gradient boosting for multi classification task. For deep learning algorithms, I applied CNN, RNN, LSTM, Bi-LSTM, and GRU on proposed corpus. Evaluation was carried out using accuracy, weighted precision, weighted recall and weighted F1-measure as evaluation measures. Experiments show that the best results with accuracy of 85 % is achieved by classical machine learning approach using SVM classifier on 3000 features with word n-gram [n ranges from 1 to 3], while for deep learning approach CNN with 1 layer outclass other classifiers with accuracy of 96 %. Deep learning CNN method performed best among all classical machine learning and deep learning methods for author region identification task on proposed Urdu articles corpus.Item Development of Large Scale English-Urdu Machine Translation Corpus for Statistical and Neural Machine Translation Systems(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2019) Moodser Hussain; FA17-RCS-005; LHR TP 5623; Dr. Rao Muhammad Adeel NawabThe process of automatically converting the text of one natural language into an other is known as Machine Translation (MT). The two state-of-the-art machine translation techniques are Statistical Machine Translation (SMT) and Neural Ma chine Translation (NMT). In both MT techniques, a large-scale parallel aligned corpus of source & target language text is the most basic requirement to develop quality models. For the Urdu language, there are minimal parallel resources de veloped until now, and these existing resources are not enough to produce some quality translation systems. This research study is majorly focusing on: (1) the development of a large-scale semi-automatically aligned English-Urdu parallel corpus of more than 150k sen tences by collecting the data from different domains. (2) the compilation of large sized monolingual data for English and Urdu language by merging existing cor pora. (3) the development of a character-separated parallel corpus for supervised transliteration (by using Wikipedia titles) (4) the evaluation of impact generated by language model and monolingual corpus on translation quality (by using differ ent data and word-order in language models training). (5) development of Phrase based SMT (PB-SMT) and Recurrent Neural Network-based NMT (RNN-NMT) systems to evaluate and compare their performance on the proposed parallel cor pus. (6) the usage of supervised transliteration and ‘minimum error rate training’ to enhance the quality of SMT. To evaluate the performance of MT systems, ‘Bilingual Evaluation Understudy (BLEU)’ is used as an evaluation measure. The best-achieved scores for English to-Urdu translations are 37.39 and 32.70 by using PBSMT and RNN-NMT, re spectively. While for Urdu-to-English translations, PBSMT and RNN-NMT yield the best score of 28.20 and 30.63 respectivelyItem Detecting Urdu Semantic Textual Similarity through Word and Sentence Embedding Techniques(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2019) Muhammad Farhat Ullah; FA17-RCS-019; LHR TP 5625; Dr. Rao Muhammad Adeel NawabSemantic Textual Similarity (STS) evaluates the degree to which two sentences or short texts are semantically proportional to one another. STS is one of the significant prob lems in the field of Natural Language Processing (NLP). Text reuse and plagiarism detection are famous examples of STS. STS has three types; (1) Monolingual STS: if the source and suspicious short texts are in the same language. (2) Multilingual STS: if the source is in one language and sus picious short text is more than two languages. (3) Cross-lingual STS: if the source in one language and suspect is in the other language, often translation of each other. STS could be found several levels, for example, word, sentence, paragraph, and document level. Urdu is one of the low resource languages. It’s the National Language of Pakistan, also widely spoken and used in electronic, print media of Pakistan, India, and Bangladesh. The main aim of this thesis is to develop techniques that measure STS as paragraph level for the Urdu language. Our thesis aims to develop and investigate the new feature extraction techniques to ad dress the problem of STS for Urdu. We divide it into three layers, (1) Train Word and Sentence Embedding models on Urdu datasets, (2) Apply these new feature extraction techniques to extract feature from Urdu short text pairs, and (3) Apply machine learn ing classification algorithms for Urdu STS. In the first step we train Word and Sentence Embedding models on Urdu datasets. By using these embedding models, we extract the word, and sentence embedding features from pre-processed Urdu Short Text Reuse Corpus (USTRC) short texts. After that, we find the cosine similarity between these extracted feature vectors, then apply classification algorithms on similarity to classify short texts into verbatim, paraphrased and independently written. In word embedding techniques, we used Word2Vec, GloVe, and FastText with Addition, Average, and Mul tiplication Functions also we explored Smooth Inverse (SI) and Term Frequency (TF) weighted word embedding techniques. In sentence embedding techniques we have used two unsupervised (sent2vec and LASER) and two supervised (InferSent and BERT) ix techniques. We have used seven machine learning algorithms to classify similarity score including, Naive Bayes (NB), Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), K-Nearest Neighbour (KNN), Gradient Boost (GB), Multi-Layer Percep tron (MLP). To evaluate these classifiers, we apply F1 measure. We got best F1 measure = 0.68, 0.75, 0.92, 0.70 by using sent2vec sentence embedding technique with GB and MLP classifiers.Item Urdu Word Sense Disambiguation using Siamese Neural Networks(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2020) Waqar Ashiq; SP19-RCS-013; LHR TP 6419; Dr. Rao Muhammad Adeel NawabThe process of Word Sense Disambiguation (WSD) is to recognize the exact sense of the targeted polysemous terms in the given context. Word sense ambiguities exist in all human languages, and humans can easily solve this ambiguity but it’s a tough task for machines to solve automatically. WSD is a significant long-established task in the area of natural language processing (NLP) with a remarkable contribution in numerous applications. The task of WSD has been categorized into (1) Lexical Sample, and (2) All-Words WSD. In literature, the problem of Urdu WSD has been mainly addressed using feature-based machine learning methods (based on manual feature engineering). A major limitation of manual feature engineering is that it requires a lot of time and endeavors to develop machine learning methods in Urdu WSD. To overcome the limitations of manual feature engineering, deep learning approaches are proposed, and they have produced promising results for various NLP tasks. To fulfill the gap in WSD research, we developed and apply state-of-the-art Siamese Neural Network (SNN) based approaches for both tasks for the Urdu Language. We developed the SNN based approaches for both WSD tasks using six deep learning models (CNN, RNN, LSTM, BiLSTM, GRU, and BiGRU) with the combination of three distance measure functions including Euclidean distance, Manhattan distance, and Cosine similarity. These Siamese network features of all deep learning models are combined for WSD classification using eight machine learning classifiers with two well-known feature selection methods (RFECV and Chi-Square). Additionally, for the All-words task SNN features are combined with n-gram (unigram and bigram) features. The developed SNN based approaches are evaluated on well-known benchmark corpora which include (1) UAW-WSD-18 Corpus for All-words WSD and (2) ULS WSD-18-Corpus for Lexical sample WSD task. After extensive experimentation, the proposed approach produced state-of-the-art results (Accuracy = 64%) in All-words WSD and outperform the baseline results (Accuracy = 57.71%). The results of Lexical sample WSD are not improved because of very few instances of each ambiguous word in corpora for deep learning approaches.