Salim Sazzed. Proceedings of the Sixth Workshop on Noisy User-generated Text (W-NUT 2020). 2020.

The resource constraints in many languages have made the multi-lingual sentiment analysis approach a viable alternative for sentiment classification. Although a good amount of research has been conducted using a multi-lingual approach in languages like Chinese, Italian, Romanian, etc. very limited research has been done in Bengali. This paper presents a bilingual approach to sentiment analysis by comparing machine translated Bengali corpus to its original form. We apply multiple machine learning algorithms: Logistic Regression (LR), Ridge Regression (RR), Support Vector Machine (SVM), Random Forest (RF), Extra Randomized Trees (ET) and Long Short-Term Memory (LSTM) to a collection of Bengali corpus and corresponding machine translated English version. The results suggest that using machine translation improves classifiers performance in both datasets. Moreover, the results show that the unigram model performs better than higher-order n-gram model in both datasets due to linguistic variations and presence of misspelled words results from complex typing system of Bengali language; sparseness and noise in the machine translated data, and because of small datasets.

Salim Sazzed. Proceedings of the Fifth Workshop on Computational Approaches to Linguistic Code-Switching. 2021.

The presence of abusive and vulgar language in social media has become an issue of increasing concern in recent years. However, research pertaining to the prevalence and identification of vulgar language has remained largely unexplored in low-resource languages such as Bengali. In this paper, we provide the first comprehensive analysis on the presence of vulgarity in Bengali social media content. We develop two benchmark corpora consisting of 7,245 reviews collected from YouTube and manually annotate them into vulgar and non-vulgar categories. The manual annotation reveals the ubiquity of vulgar and swear words in Bengali social media content (i.e., in two corpora), ranging from 20% to 34%. To automatically identify vulgarity, we employ various approaches, such as classical machine learning (CML) classifiers, Stochastic Gradient Descent (SGD) optimizer, a deep learning (DL) based architecture, and lexicon-based methods. Although small in size, we find that the swear/vulgar lexicon is effective at identifying the vulgar language due to the high presence of some swear terms in Bengali social media. We observe that the performances of machine leanings (ML) classifiers are affected by the class distribution of the dataset. The DL-based BiLSTM (Bidirectional Long Short Term Memory) model yields the highest recall scores for identifying vulgarity in both datasets (i.e., in both original and class-balanced settings). Besides, the analysis reveals that vulgarity is highly correlated with negative sentiment in social media comments.

Although cryo-electron microscopy (cryo-EM) has been successfully used to derive atomic structures for many proteins, it is still challenging to derive atomic structures when the resolution of cryo-EM density maps is in the medium resolution range, such as 5–10 Å. Detection of protein secondary structures, such as helices and β-sheets, from cryo-EM density maps provides constraints for deriving atomic structures from such maps. As more deep learning methodologies are being developed for solving various molecular problems, effective tools are needed for users to access them. We have developed an effective software bundle, DeepSSETracer, for the detection of protein secondary structure from cryo-EM component maps in medium resolution. The bundle contains the network architecture and a U-Net model trained with a curriculum and gradient of episodic memory (GEM). The bundle integrates the deep neural network with the visualization capacity provided in ChimeraX. Using a Linux server that is remotely accessed by Windows users, it takes about 6 s on one CPU and one GPU for the trained deep neural network to detect secondary structures in a cryo-EM component map containing 446 amino acids. A test using 28 chain components of cryo-EM maps shows overall residue-level F1 scores of 0.72 and 0.65 to detect helices and β-sheets, respectively. Although deep learning applications are built on software frameworks, such as PyTorch and Tensorflow, our pioneer work here shows that integration of deep learning applications with ChimeraX is a promising and effective approach. Our experiments show that the F1 score measured at the residue level is an effective evaluation of secondary structure detection for individual classes. The test using 28 cryo-EM component maps shows that DeepSSETracer detects β-sheets more accurately than Emap2sec+, with a weighted average residue-level F1 score of 0.65 and 0.42, respectively. It also shows that Emap2sec+ detects helices more accurately than DeepSSETracer with a weighted average residue-level F1 score of 0.77 and 0.72 respectively.