University of Computer Studies Yangon

Sentence classification is presently utilized in various fields because of its dependability, scalability, and efficiency. The grammatical structure of the Myanmar language presents specific difficulties for sentence classification. The lack of spaces between words complicates the identification of words and syllables during feature extraction. This characteristic significantly affects model performance and preprocessing, unlike in languages such as English. Several challenges must be addressed when performing classification modeling with actual data. The primary objective of sentence classification is to identify the type of subjective information present in the source material. A dataset for sentence classification in the Myanmar language has been developed using data collected from websites and social media platforms. Currently, 112K sentences have been gathered and labeled with class identifiers. The classification models employed for categorizing the input sentences include Feedforward Neural Network (FNN) and Bidirectional Long Short-Term Memory Networks (BiLSTM). To improve the effectiveness of sentence classification, BERT-based features and fastText embeddings are utilized. Both feature types demonstrate strong individual performance and work effectively with FNN and BiLSTM. The effectiveness of these features is also evident when they are combined in both FNN and BiLSTM. This paper focuses on validating the dataset and evaluating the impact of these features on sentence classification in the Myanmar language. Using the combined features, the FNN model reaches an average F1-score of 97%, while the BiLSTM model achieves an average F1-score of 98%.

Teachers and students often employ the use of intrasentential code-switching while teaching Information Technology (IT) and other subjects in universities and colleges. Myanmar automatic speech recognition (ASR) systems still stumble on code-switching because switched utterances blur language borders, mix pronunciations, and broaden word lists-especially when English terms slip through Myanmar sounds. These mismatches push up word-error rates and cause many mixed-language phrases to be misunderstood. This paper presents the MEASR dataset, a spontaneous speech dataset featuring Myanmar-English intra-sentential code-switching collected from real online IT teaching sessions. It addresses the challenges code-switching poses to Myanmar ASR systems - such as language boundary confusion, pronunciation mixing, and vocabulary expansion-which conventional monolingual models struggle with. The MEASR dataset includes around 10 hours of speech recorded at 16 kHz, mono channel, and 16-bit resolution. The paper details the dataset's design and provides an analysis to support improved code-switching in automatic speech recognition (CS-ASR) through bilingual training, pronunciation adaptation, and language identification.

Machine translation (MT) plays a crucial role in bridging linguistic gaps, particularly for underrepresented languages. This paper investigates the feasibility and performance of the mT5-transformer model in machine translation task from Myanmar to Tedim Chin and Tedim Chin to Myanmar language respectively. The experiment conduct on the 10K parallel corpus from UCSY dataset that contain different domain such as greetings, stories (from educational books), communication, leisure (traveling, shopping, nightlife), and resources (numbers, time, and accuracy). In data collection stage, we firstly collect 10K Myanmar sentences from UCSY corpus, and then, translate it into Tedim Chin language and finally create our own Myanmar-Tedim Chin corpus. On the other hand, Myanmar text is written without explicit spaces between words, making segmentation a challenging an essential step. Therefore, we preprocess the dataset using the customized syllable and word level segmentation approaches for Myanmar language and then fine-tune it on the mT5 model using that dataset, and finally evaluate the performance of the proposed model with the BLEU scores and accuracy. From all experiments, we can see that the model performance on the syllable level gets the outperform result than the word level approach. Therefore, comparing the result on both translations on syllable segmentation, the score for the experiment show that the Myanmar-Tedim Chin translation gives better result than the Tedim Chin-Myanmar translation by increasing 0.10% on BLEU score and 0.07% on the accuracy score. Tedim Chin is a low-resource language and the final result shows that the mT5 transformer work pretty well on Tedim Chin and also in Myanmar text.

Reservoir dams play a vital role in the national economy and social development of any country, such as flood control, power generation, irrigation and water supply. The Lower Paunglaung Dam, the largest multipurpose dam in Myanmar, is facing difficulties in managing its water resources. Lower Paunglaung Dam and its weir were mainly constructed to irrigate agricultural land as well as to generate hydropower. Predicting the inflow rate precisely is therefore essential for managing water resources since it aids in planning for floods, droughts, and the production of electricity in the Nay Pyi Taw region. Moreover, climate change impacts are more significantly affecting the safety of dams. In order to prevent flood damage and maintain a steady supply of water from water resources, accurate dam inflow estimates are becoming more and more important in the face of climate change. In order to estimate dam inflow, this study used historical data from the previous 20 years. A deep learning model, Long Short-Term Memory (LSTM) is used in this system to predict the inflow of Lower Paunglaung Dam in the Paunglaung River watershed in Myanmar. Root Mean Square Error (RMSE), accuracy, number of epochs are taken into account for assessing the effectiveness and performance of the suggested inflow prediction model.

Traffic accidents are becoming a major disruptive issue and emergency services play a key role in rescuing injured people from traffic accidents to reduce the loss of life and property. Today's technologies can trace the current place of the traffic victims and send the user's information with that location to the nearest emergency station to get the timely rescue service. The Motion Sensing Method can detect traffic accidents from the values of the sensors in the smartphone, and determine the accident levels. The Difference Angle Method shows the intended destination of the automobile with endless GPS signals. To obtain detailed results, the Modern Positioning Method is used to communicate with the sensors to get the accident location and track the victim. The Object Detection in Fence Algorithm searches for the nearest emergency services to provide medical assistance by sending the accident information. This system can detect the location of injured persons, analyze the severity levels of accidents, send timely messages to rescue stations, and identify the nearest routes.

Many students find it challenging to comprehend the communication mechanism that is a part of TCP/IP because learning about it through textbooks and lectures is not enough.Network simulators are helpful resources for computer network education.Because the complex operations that occur in the computer network during data exchange between network devices are completed in incredibly tiny time slices, the learner is unable to follow and understand them.The main objective of our simulator is to display visually these processes.Our suggested system allows students to play with network activity to help them understand it.Using specific packets, step counts, and the network configuration setting as input, our approach calculates packet transmission step-by-step.Our method enables the students to comprehend network behavior.With our simulator, students may analyze network behavior in a variety of settings without requiring specialist gear.

Myanmar News Retrieval involves searching for and extracting pertinent documents or data depend on a user's query or information need. Neural Information Retrieval (Neural IR) leverages the power of neural networks to enhance the effectiveness of this retrieval process. In a crucial information retrieval (IR) task, neural methods are employed to rank documents in response to a given query. Neural models are utilized in various information retrieval (IR) contexts, including ad-hoc retrieval, recommender systems, multimedia search, and conversational systems aimed at producing responses to natural language queries. The evaluation metrics will be utilized to gauge the effectiveness of the retrieved documents deemed relevant to the given query. This paper introduces a Kernelized Neural Ranking Model (KNRM) designed specifically for ad-hoc retrieval tasks. Through comprehensive analysis of datasets, we analyses that conversational questions exhibit complex phenomena that are absent in current reading comprehension datasets. Our model focuses on adopting a deep architecture specifically at the query term level to support relevance matching.

Nowadays, people have embraced social media for more complex purposes such as communicating for update and reliable news during a disaster. Social media has developed into an essential information channel in recent years that can be utilized to improve disaster response. Due to the high feature dimensionality, categorizing situational awareness tweets based on user posts is a challenging procedure. This system developed a framework for classifying environmental hazards tweets to detect they are situational awareness or not by using machine learning techniques. Three different datasets are used in this system. Feature Extraction of this work has exploited TF-IDF features, linguistic features and psychometric features. Both linguistic and psychometric features were extracted using Linguistic Inquiry Word Count (LIWC). Information Gain approach was used for feature selection. Random Forest, Naïve Bayes and XGBoost are used as based classifiers of ensemble model. This system used two ensemble learning techniques: hard voting and soft voting. The findings suggest that using ensemble learning (hard voting) produced better outcomes than using ensemble learning (soft voting).

Concurrency describes how database management system normally permits several transactions to access the same database simultaneously. Using the optimistic concurrency control approach, this system aims to demonstrate concurrency control. Because of its non-blocking and deadlock-free characteristics, Optimistic concurrency control is particularly appealing for real-time transaction processing. Transactions run concurrently and without the use of locks under optimistic concurrency control. When conflicts arise during the validation phase, the forward validation process in this system restarts to ensure database consistency. This system uses internet purchasing for the MICHAEL KORS bags selling system as a case study. The serializability of the item sales system is highlighted by this system. It avoids interference when two or more users access the database at the same time by utilizing the proposed system.