Online text and video tutorials are among the most popular and heavily used resources for learning feature-rich software applications (e.g., Photoshop, Weka, AutoCAD, Fusion 360). However, when searching, users can find it difficult to assess whether a tutorial is designed for their level of expertise. TutVis stands for tutorial visualization, which is a Photoshop tutorial browsing system that provides auto-generated information to assist users in the tutorial searching and selection. The provided information is as follows: difficulty level (advanced/beginner), topics covered, length, text complexity, command usage ratio.

The objective is to develop an automated tool, which can find out topics from any article. Different topic modeling techniques are investigated for this topic extraction task. For the investigation, yahoo answer dataset has been chosen. The dataset has ~1.5M question-answer pairs. Only 60,000 samples has been randomly chosen from the data. From the selected data, only the answers are used to train the models. The models are trained using LDA and Top2Vec.

The objective of this project is sentiment analysis (i.e., Positive, Neutral, Negative) from the popular health and fitness app reviews. For this task, app reviews are collected from the google play store. 10 popular health and fitness apps are chosen. All around 12000 most recent reviews are collected. Ratings are considered as the measure/label of positive, negative, and neutral sentiment of the reviews. The collected data are preprocessed and trained using a transformer model.

This project aims to classify ratings of the hotel reviews. There are 5 ratings (i.e., class) in the dataset along with the reviews. The dataset is quite balanced among the 5 classes. The objective is to develop two different classification models i.e., a baseline and a state of the art(SOTA) model to compare the performance of classifying ratings. To uncover interesting insights from the data, different exploratory analysis have been performed.

The objective is to implement different word embeddings and investigate the performance on a dataset. For this task, the chosen dataset includes tweets about disasters, e.g., earthquake, wildfire. Among the tweets, some are fake and some are real. So, there are two labels such as 1:Real tweet and 0: Fake tweet. This data has been chosen so that we can see if the features extracted using different embedding techniques can retain distinguishable information to detect which tweet is fake and which one is real.

The objective is to build a question classification model. The questions have six different categories such as: Description(DESC), Entity(ENTY), Abbreviation(ABBR), Human(HUM), Location(LOC), Numeric Value(NUM). To investigate different approaches, the following data is used (downloaded from https://cogcomp.seas.upenn.edu/Data/QA/QC/): Training set 5(5500 labeled questions) Test set: TREC 10 questions Different data analyses have been performed and four different models are trained. The models are the followings: Tf-Idf + SVM: Tf-Idf is used for vectorizing texts and a linear model (i.

The objective of this project is to classify news category from articles. The input data consist of 2225 news articles from the BBC news website corresponding to stories in 5 topical areas (e.g., business, entertainment, politics, sport, tech). LSTM has been applied in the classification task to categorize articles. TensorFlow 2.0 has been used to train the model. Word embedding is used in feature generation. TSNE is used to visualize the word vectors in 2d space.

The dataset includes tweets about disasters, e.g., earthquake, wildfire. The objective is to detect if the tweet is about a real disaster vs. fake disaster. Different approaches have been performed for data cleaning and training the model. The best model can predict real vs. fake tweets with 89% accuracy using transfer learning (BERT). The following models have been developed for training: BOW Model with Logistic Regression. (accuracy 77%) Tf-Idf with Logistic Regression.

The objective of this project is to detect intent from texts. For this, a benchmark dataset is used, which includes 7 intents (Search Creative Work, get weather, Book Restaurant, Play Music, Add to Playlist, Rate Book, Search Screening Event) and 14 thousand samples. Transfer learning has been leveraged to train a machine learning model. The model takes the raw texts, which are tokenized and vectorized to feed into the pre-trained model.