Text classification, a fundamental technique in information retrieval, categorizes textual data to streamline search operations, significantly enhancing efficiency. This research introduces CNN, a novel hybrid deep learning model that advances text classification by synergistically combining the strengths of transformer models, CNNs, RNNs, and LSTMs. The model stands out with its innovative design, inputting ‘input ids’ and ‘attention mask’ into a pretrained transformer, followed by a series of operations leading to robust feature extraction. CNN outperforms baseline models, demonstrating superior accuracy in text classification. Extensive experimentation on both Chinese and English datasets substantiates this, with the model achieving the highest accuracies of 94.54% and 91.89%, respectively. This paper underscores the effectiveness of the proposed model in enhancing text classification, paving the way for more efficient information retrieval.

This paper presents a rigorous, comparative deep learning benchmark of modern architectures for facial expression recognition (FER) in the wild. We systematically implement and evaluate a representative set of architectures - including residual CNNs, hybrid CNN-Transformer models, Vision Transformers, and contrastive-pretrained ResNet variants - on three challenging datasets: FER2013, RAF-Basic, and RAF-Compound. Our study investigates overfitting, the effectiveness of strong data augmentations, validation/test performance, and architectural trends. We provide open, reproducible code and practical insights for real-world FER deployment. Extensive methodological descriptions, mathematical motivations, and ablation studies are included to make the benchmark both transparent and reusable.

This paper addresses the persistent problem of vocabulary mismatch in information retrieval, a field that aims to locate and retrieve relevant information from a database. The paper introduces the concept of query expansion, a technique that enhances the performance of retrieval models by expanding queries with semantically related terms. The primary goal of this research is to improve the query expansion using Asymmetric Kullback-Leibler Divergence from the randomness framework. This work also presents a comprehensive experimental analysis, implementing both classic and state-of-the-art algorithms like neural reranking models, i.e., Kernel Neural Ranking Model and Vanilla-Bert on two different large real-world datasets. The results demonstrate significant improvements in retrieval performance when using the proposed KL divergence method.

Search engines need to process huge amounts of data and serve many users at the same time. This makes it challenging to implement real-time passage reranking algorithms that can work at a large scale. However, our proposed modified cross encoder architecture provides a good solution by using pre-trained transformer models and a unique scoring method. The two-stage architecture first encodes query-document pairs using a pre-trained transformer model. Then, it uses a novel scoring technique that extracts logits and calculates the cosine similarity between the [CLS] token embeddings of the query and document. During training, it minimizes a combined loss function using sigmoid cross-entropy and cosine similarity. This allows the Modified cross encoder to effectively distinguish between relevant and irrelevant documents, making it better than base models like BM25, SparseBiEncoder, DenseBiEncoder, and cross encoder. Our proposed model demonstrated significant performance enhancements, achieving a higher nDCG@10 accuracy of 0.77 and improving other relevant metrics.

There is a rising need for automated systems that can accurately identify misleading content because the internet is now flooded with inaccurate information. In this work, we examine how well the eight distinct state-of-the-art deep learning models-BiGRU, CNN, RoBERTa, and GraphSAGE-perform on the text-based fake-news classification problem. We evaluate the efficacy, generalization capacity, and limitations of the various models to categorize false information and documents from various writers (i.e., C2L Writers) over time across various forms using a variety of benchmark datasets. Our comprehensive studies demonstrated that models that combine transformers and graphs are far more efficient than conventional neural networks, particularly when it comes to identifying bogus news and comprehending material based on context and how it relates to other stuff. This work emphasize the significance of employing reliable feature extraction and representation learning techniques for scalable, real-time automated misrepresentation identification. Additionally, this study will establish baseline measurements for future development of false-document analysis utilizing automated machine-learning algorithms and deepen our understanding of various machine learning techniques for text-based fake news detection.

Diabetic Retinopathy (DR) is the primary cause of preventable blindness among people with diabetes where early diagnosis is important. It is required to avoid permanent loss of vision impairment. Traditional deep learning-based diagnostic models often depend on centralized repositories of medical images which raises critical concerns about patient privacy, regulatory compliance, and institutional data ownership. Considering these challenges this work presents a Federated Learning (FL) framework for multi-class DR detection, that enables multiple healthcare institutions to collaboratively train a single model without sharing the raw patient data. In the proposed system Gaussian-filtered retinal fundus images resized to 224×224 are used to train local deep learning models across distributed clients. We conduct a comparative analysis through a Custom CNN, MobileNetV2, VGG16, and ResNet50 architectures by only the optimized model parameters are transmitted to a Central server for aggregation using the federated averaging algorithm (FedAvg). Experimental evaluation shows that the ResNet50 architecture achieves the highest diagnostic accuracy (78%), closely tagged by MobileNetV2 (76%), showing that deeper and more robust feature extractors maintain performance superiority even under non-IID FL constraints. The federated model that not only preserves privacy and confidentiality but also achieves clinically reliable diagnostic accuracy with improved generalization across diverse datasets. These results underscore the potential of Federated Learning as secure, scalable, and privacy-preserving solution for early DR screening and real-world large-scale healthcare deployment.

Graph Neural Networks have emerged as powerful tools for analyzing graph-structured data. However, their performance often varies across datasets due to challenges such as noisy edges, sparse connectivity, and over-smoothing in deep layers. To address these limitations, Graph Deformation Network Convolution (GDNConv) is proposed as a novel graph convolution model that incorporates four key innovations: dynamic edge weight learning to filter noisy connections, graph attention deformation to prioritize relevant neighbors, multi-level aggregation to capture multi-scale patterns, and self-regularization to stabilize training. This proposed model demonstrates robustness and scalability, particularly for real-world applications involving complex and noisy graph structures, such as social networks and recommendation systems. It has the ability to dynamically adapt graph topology during training and superior performance on both dense and sparse datasets highlight its potential as a versatile solution for graph-based learning tasks. Additionally, GDNConv’s computational efficiency and self-regularization mechanisms make it suitable for large-scale applications where resource constraints are a concern. The proposed model is evaluated on four benchmark datasets—Cora, CiteSeer, PubMed, and ogbn-arxiv—and compared with several state-of-the-art models, including Graph Convolutional Network, Graph Attention Network, and Graph Sample and Aggregate. The experimental results demonstrate that the proposed model consistently outperforms these baseline approaches, achieving improvements of 4.7% and 4.2% in both accuracy and F1 Score.

Heterogeneous Graph Neural Networks have emerged as powerful tools for analyzing complex, multi-typed graph-structured data. However, their performance often varies across domains due to challenges such as handling diverse node and relation types, capturing dynamic temporal patterns, and providing model explainability. To address these limitations, this paper proposes the Dynamic Meta-Attention Heterogeneous Graph Transformer (DMAHGT) as a novel heterogeneous graph model that incorporates four key innovations: a two-level hierarchical attention mechanism for nuanced message passing; dynamic meta-attention that adapts weights based on node features, relation types, and temporal information; meta-learning capabilities for cross-dataset adaptation via a CrossDatasetMetaLearner; and inherent explainability through importance scoring of nodes and features. The proposed model demonstrates robustness and scalability, particularly for real-world applications involving complex and evolving graph structures, such as academic and social networks. Its ability to dynamically adapt to different relational semantics and temporal contexts highlights its potential as a versatile solution for heterogeneous graph-based learning tasks. The proposed model is evaluated on three benchmark datasets—ACM, IMDB, and DBLP—and is comprehensively compared with several state-of-the-art heterogeneous graph models, including HAN, HGT, HGCN, HeteroGNN, IGNN, and the meta-learning baseline MAML. The experimental results demonstrate that DMAHGT consistently outperforms these baseline approaches, achieving superior performance in node classification, clustering, and similarity search tasks.

Face recognition (FR) has emerged as a vital biometric authentication method due to its non-invasive nature and widespread applications in security, surveillance, and identity verification. However, existing FR systems face significant challenges in preserving user privacy, especially when deployed in cloud environments. Traditional approaches lack robust encryption mechanisms, making sensitive facial data vulnerable to breaches, adversarial attacks, and unauthorised access. This study introduces a novel privacy-preserving face recognition (PPFR) framework that leverages deep learning (DL) models, specifically FaceNet and ArcFace, for extracting discriminative facial embeddings. The framework combines these models with the Cheon–Kim–Kim–Song (CKKS) homomorphic encryption scheme to enable secure computation on encrypted data without requiring decryption. Locality Preserving Projection (LPP) is applied to compress high-dimensional features while preserving the essential data structure, thereby enhancing performance and reducing processing overhead. The proposed system was evaluated on four benchmark datasets: CASIA3D, 105PinsFace, LFW, and Faces94. Among the models, ArcFace showed the best performance, achieving high accuracy and low Equal Error Rates (EER), with results of 99.30% accuracy on CASIA3D, 98.06% on 105PinsFace, 97.07% on LFW, and 99.82% on Faces94 datasets. The experimental results demonstrate that the proposed system is both effective and secure, making it suitable for real-world face recognition applications with enhanced privacy protection.

Early identification of crop diseases is very important. It helps farmers keep their crops healthy and protects the overall farming system. When diseases are found early, farmers can take quick action, reduce damage, and make sure there is good and steady food production. But checking crop diseases by hand is slow and not always correct. This is why we need automated methods that can work fast and give accurate results, even in different field conditions. In this study, we test three modern object detection models - YOLOv8, YOLOv11, and YOLOv12 - using two public datasets. We also check how fine-tuning YOLOv12 with the AdamW optimizer affects its performance. On Dataset D1 (7,982 images with 20 disease types), the fine-tuned YOLOv12 model reached 91.08% accuracy and 92.51% mAP@0.5. This is better than all the other models we tested. On Dataset D2 (3,683 images across 29 classes), the model's performance dropped. This shows that the model was more likely to overfit when there were many classes but fewer images for each class. These results show that we need to choose the right training methods based on the dataset. They also highlight the need for proper regularization when training deep learning models for agricultural tasks. Overall, the results show that YOLO models work well for real-time crop disease detection. They also show that the right fine-tuning can improve how accurately the model detects diseases. The findings also give useful guidance on how to choose and modify YOLO models based on the size and difficulty of crop-disease datasets.

Estimating event-free survival after a hematopoietic cell transplant is not straightforward. Each patient has unique factors that interact and affect the result. Traditional survival models like the Cox proportional hazards model have limitations. They cannot easily capture nonlinear patterns in the data. They also fail to model complex interactions between features. As a result, their predictive power is often restricted. We propose a transformer-based framework for EFS prediction in hematopoietic cell transplantation (HCT) patients. It is built using the Feature Tokenizer-Transformer (FT-Transformer) architecture. The design focuses on structured clinical data. The model processes 35 categorical variables through embedding layers. It also handles 22 continuous features using optimized linear projections. Multi-head self-attention then learns key relationships between the features automatically. In our approach, we obtained a validation concordance index of 0.689. This score is 6.6% higher than traditional models. Survival analysis using Kaplan-Meier curves confirms distinct group separation (p < 0.001). Attention-based explanations reveal influential clinical attributes. As a result, our solution improves model transparency in medical settings. This study demonstrates the value of transformer models for survival analysis. Their ability to provide patient-specific risk estimates makes them useful in transplant medicine.

Financial fraud is now a major threat for both consumers and institutions due to the growth of digital money. The complex and dynamic nature of contemporary fraud schemes makes it difficult for traditional detection methods, which are mostly focused on static rules and human checks, to stay up. This chapter explores how fraud detection procedures in financial ecosystems are being transformed by artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL). AI models, trained on large-scale transactional datasets, can now identify, forecast, and respond to fraudulent behaviors in real time with heightened precision. We explore a variety of AI methodologies, including both supervised and unsupervised learning, and delve into neural network applications for pattern recognition. Practical implementations from major financial organizations such as PayPal, Mastercard, and American Express are presented to illustrate the operational benefits of AI-driven fraud prevention systems. Difficulties like such as data security, algorithmic bias, and false positive management are critically examined. The chapter concludes by highlighting future directions, including the blockchain integration and federated learning to improve privacy and transparency, and the continued importance of human oversight in AI-powered decision-making.

Graph neural networks (GNNs) have emerged as powerful tools for analyzing graph-structured data with applications in social networks, bioinformatics, and recommender systems. However, existing GNNs struggle with (1) rigid edge weighting (e.g., GCN’s fixed normalization), (2) over-smoothing in deep layers, and (3) quadratic attention costs (e.g., GAT). MGCN introduces: (1) adaptive edge weighting to dynamically adjust neighbor influence, (2) residual connections to combat over-smoothing, and (3) a scalable attention mechanism. It also introduces a standardized evaluation framework that incorporates adaptive preprocessing techniques such as feature normalization, edge weighting, and graph augmentation. The proposed model demonstrated superior performance when compared to eight state-of-the-art GNN models such as GraphSAGE, GAT, Graph Transformer, GINConv, GCN, GraphCL, AGCN, and MGCN, across three widely used benchmark datasets: Cora, CiteSeer, and PubMed. All evaluation metrics–including Accuracy, Hit Ratio, Precision, Recall, and F1 Score–are reported as the mean ± standard deviation over 10 independent runs. The experimental results consistently demonstrate the superiority of the proposed MGCN model with approximately 2% improvement on above datasets.

Dehazing is a crucial task in various applications, particularly in Surveillance, where clear visibility is essential for effective monitoring. This work proposes a model, ViTDehazer which leverages the power of Vision Transformer and Fusion Module to effectively remove haze from single images. This work experimentally evaluate our proposed model by comparing with few base models. These experimental results shows that our proposed model performs well compared to the base models in terms of PSNR and SSIM. Proposed model achieves PSNR and SSIM values of 25.76 and 0.85, respectively.

Federated Learning (FL) is decentralized machine learning, which preserves data confidentiality and security but suffers challenges such as high communication overhead, latency, and scalability issues in large scale smart city networks. We propose a Hierarchical Federated Learning (HFL) framework which takes advantage of fog nodes to address these problems. HFL framework ensures that the cost of communication is cut back through the introduction of the multilevel aggregation strategy where local models are aggregated first at fog nodes before they are combined at a central server. Communication costs are reduced while latency is improved and scalability is enhanced by this HFL framework through simulations on real-world smart city datasets. On edge devices, our simulation results with real-world datasets from smart cities show that compared to traditional FL it reduces communication overhead by up to 50%, achieves faster model convergence with similar accuracy, and leads to lower energy consumption. This framework represents a big step forward towards deployment of FL in smart cities making it efficient and scalable in resource-constrained environments.

In the data-driven era, applications that generate vast amounts of data have become a central focus. The exponential growth in data generation poses significant challenges in data analysis. Financial transactions, in particular, have become increasingly complex, necessitating effective methods for detecting anomalies. Unnoticed irregularities can lead to substantial problems for banks and other financial institutions, including financial losses and eroded trust. In many real-world applications, dealing with imbalanced data is a critical concern. While most classification methods focus on two-class data problems, addressing a solution for class-imbalanced scenarios is equally essential. This work proposes a methodology that applies the SMOTE algorithm to various Machine Learning (ML) and Deep Learning (DL) models, aiming to balance the imbalanced data and improve classification performance. In the practical comparison of various ML and DL models with and without the SMOTE technique, this work also experimentally examines and then discusses the challenges in identifying fraudulent transactions over two different financial transactional datasets. Finally, by using SOMTE with various ML and DL models, this work presents 37% - 91% improvement on the banking transactional dataset and 57%-98% improvement on the online shopping transactional dataset in terms of F1-Score.

Face Recognition (FR) and surveillance video analytics are well-defined and solved problems in the applications of Computer Vision. FR aims to identify an already known person in a given image. Surveillance video analytics seeks to identify the occurrence of abnormal events or things in public places. But, recognizing the movements of most wanted criminals or suspects in public areas using FR systems with unclear surveillance video inputs is a very challenging problem. This work analyses the performance of three existing popular machine learning-based FR systems. They are (i) Viola–Jones detector, (ii) HOG-based FR, and (iii) PCA-based FR. This work analyzes the performance of these FR models on two different datasets. One is a benchmark dataset that has only the frontal view of the faces of various subjects. Another dataset we created with 10000 images. These images are collected from 50 subjects. From each subject, 200 images are taken from various angles. This work observes that the above models will improve their performance from 7 to 10% in terms of accuracy by training them on the proposed dataset.

The most common method for determining positive or negative sentiment within a text is sentiment analysis. It is frequently utilized by businesses to grasp customer’s opinion regarding the products. A sentiment analysis examines message discussions and assesses the tone, expectation, and feeling behind each message. It is turning into a significant tool to notice and figure out the sentiment. It naturally analyzes whether a person is happy/frustrated/sad. It deals with the task of determining the difference during a document or a sentence and has gotten loads of consideration lately for national language. With the ascent of social media, a lot of data is available in provisional language other than English. Telugu is one such language with bountiful information accessible in social media, and it is difficult to look out labeled data of sentences for Telugu sentiment analysis. In this project, we differentiate positive and negative sentiment classes based on the polarity of code-mixed sentences, and the metrics are evaluated using machine learning approaches such as Naive Bayes, support vector machine, and recurrent neural network.

The news is the most crucial resource for the general population to learn about what is occurring across the world. Even if newspapers remain a reliable source of news, social media is currently the next frontier in news. Regular individuals may simply alter the news to produce fake news since these social networks are so accessible. These fictitious news stories may be utilized for both political and commercial gains. It may be used as a vehicle to stir up neighborhood animosity, which is detrimental to society. In order to mitigate its impacts, it is crucial to recognize fake news. A platform that can validate and classify news is currently unavailable. In this essay, a technique is presented for figuring out whether or not news is reliable in the present. To train the features that were retrieved from the data using natural language processing techniques, this system makes use of ML classifiers including Decision Tree, Random Forest (RF), and Logistic Regression (LR). We evaluate each classifier’s performance using a variety of parameters. The best classifier will provide the outcome for real-time news prediction.

Natural Language Processing Event Extraction work is crucial. Social media has become increasingly significant in today's world. Using natural language, people can express their views on a wide range of topics on social media. Social Media tools popularized the devices among the masses making the information distribution faster and easier. The exchange of text is the most popular means of communication across social media users. It had become necessary to understand the semantics of messages communicated as the messages had a wide effect across the users. Event extraction means extracting the events across streams of the social media messages. Event extraction helps in taking corrective actions in case of natural calamities and hence possibly save the lives of many people. The major objective of the task is to draw specific knowledge to predict the events(incidents) specified in the code-mixed digital text. This paper proposes a two-step procedure for the extraction of the events. The first phase is by applying a binary classifier to identify the messages containing the event. The second phase is by applying a sequence labeling technique, conditional random fields (CRF), to extract the event from the message. As social media text is a bit noisy, it is a major challenge to develop learning algorithms for these tasks. Parts of Speech (POS) tags and Named Entity Recognition (NER) are used for the words to address some issues in this challenge.

Multiple Instance Learning (MIL) is widely used to locate abnormal video frames in surveillance videos due to its ability to work with weakly-labeled data. On the other hand, graph-based semi-supervised approaches are employed to handle label sparsity-related issues such as Low Degree-of-Supervision (LDoS) and High Class-Imbalance (HCI). However, the application of the MIL paradigm in anomaly classification of surveillance videos faces significant challenges such as (i) LDoS, (ii) HCI, and (iii) A multitude of types of abnormal events and actions instead of a unique pattern. The current work proposes a novel anomaly classifier to address these issues with a new objective function (GssMILP) by leveraging the benefits from graph-based semi-supervised and multiple-instance learning approaches. In this regard, our contributions are six-fold. First, a deep hierarchical architecture (C3D+3DCNN BICLSTM) to extract generic video descriptors. These descriptors represent the multitude of types of abnormal events and actions by incorporating the local–global spatial information and long-short-term temporal information related to the object motion, human actions, and behavioral features. Second, we develop a learning objective (GssMILP) that extends the MIL-based instance-level labeler to graph-based semi-supervising. Third, we use three different regularizer terms, namely graph-based regularizer, non-convex L2-regularizer, and temporal-smoothness regularizer. Fourth, we present an elaborate experimental study comparing the performance of the proposed approach with six similar baseline approaches. The results indicate that at degree-of-supervision 1.0%, the proposed method improves the f1-score by 45.26%, 18.19%, 45.28%, and 9.22% in UCSD-ped1, UCSD-ped2, MED, and RWAD datasets, respectively. Fifth, we present an ablation study on the impact of different features on the performance of an anomaly classifier. Sixth, we provide temporal annotations for a label resource of 1900 videos in the RWAD dataset.

Multiple-instance learning (MIL) is a weakly supervised learning paradigm in which a training dataset contains a set of labeled bags, and each bag contains multiple number of unlabeled instances. Preparation of instance-level labels is resource intensive. Being weakly supervised, MIL is sensitive to several practical issues such as noisy label information and low witness rate. A scenario of high class imbalance and low degree-of-supervision further poses additional challenges. Recent works on graph-based label propagation methods have been shown to be effective in semi-supervised setup to address such issues by propagating the label information over graph-based manifold. Application of such semi-supervised strategies for MIL framework requires the instance-level labeling. Whenever the problem setup contains the three characteristics of high class imbalance, low degree-of-supervision and weak supervision, the state-of-the-art methods of either MIL or graph-based label propagation are inadequate when applied alone. This article proposes a non-convex formulation for instance-level MIL to find the instance-level labels by combining the benefits of both MIL and graph-based label propagation methods. The proposed approach improves the performance of the classifier using density-difference- and distance-based structural smoothness assumptions in the graph structure. This article presents the comparison of the performance of the proposed method to those of several state-of-the-art base-lines in MIL. The experimental results are shown on multiple datasets from four different applications. The proposed method is compared in a total of 616 cases (14 datasets × 11 base-line models × 4 low degree-of-supervision values). The minimum f-score improvements are 15.22%, 1.14%, and 4.25% in DAP, CIR, and ACSV datasets, respectively.

Machine learning algorithms based on semi-supervised strategies have drawn the attention of researchers due to their ability to work with limited labeled data making use of huge number of unlabeled samples. Graph based semi-supervised algorithms make an assumption of similarity of examples in lower dimensional manifold and use an objective that ensures similarity of labels as enforced by the similarity graph. Such methods typically make use of a L2 regularization term to avoid over-fitting. Regularization term further ensures convexity of the overall objective leading to efficient learning algorithms. Addressing the problem of low-supervision and high class imbalance, prior work has shown state-of-the-art results for anomaly detection and other important classification problems by using a convex-concave objective. The current work analyses such performance improvements of convex-concave objective thoroughly. Our study indicates that a KL-Divergence based loss function for semi-supervised learning has performed much better than the convex-concave objective based on L2-Loss. It is also seen that the one-versus-rest setting for multi-class classification using convex-concave objective is performing much weaker compared to the naturally multi-class KL-Divergence based multi-class classification setting.

Social media data has become invaluable component of business analytics. A multitude of nuances of social media text make the job of conventional text analytical tools difficult. Code-mixing of text is a phenomenon prevalent among social media users, wherein words used are borrowed from multiple languages, though written in the commonly understood roman script. All the existing supervised learning methods for tasks such as Parts Of Speech (POS) tagging for code-mixed social media (CMSM) text typically depend on a large amount of training data. Preparation of such large training data is resource-intensive, requiring expertise in multiple languages. Though the preparation of small dataset is possible, the out of vocabulary (OOV) words pose major difficulty, while learning models from CMSM text as the number of different ways of writing non-native words in roman script is huge. POS tagging for code-mixed text is non-trivial, as tagging should deal with syntactic rules of multiple languages. The important research question addressed by this article is whether abundantly available unlabeled data can help in resolving the difficulties posed by code-mixed text for POS tagging. We develop an approach for scraping and building word embeddings for code-mixed text illustrating it for Bengali-English, Hindi-English, and Telugu-English code-mixing scenarios. We used a hierarchical deep recurrent neural network with linear-chain CRF layer on top of it to improve the performance of POS tagging in CMSM text by capturing contextual word features and character-sequence-based information. We prepared a labeled resource for POS tagging of CMSM text by correcting 19% of labels from an existing resource. A detailed analysis of the performance of our approach with varying levels of code-mixing is provided. The results indicate that the F1-score of our approach with custom embeddings is better than the CRF-based baseline by 5.81%, 5.69%, and 6.3% in Bengali, Hindi, and Telugu languages, respectively.

Named entity recognition(NER) is a key task in NLP pipeline useful for various applications such as search engines, question answering systems, sentiment analysis in domains ranging from travel, bio-medical text, newswire text, financial text etc. NER is effectively solved using sequence labeling approaches like HMM and CRF. Though, CRF (being discriminative) shows better performance compared to HMM, it uses discrete features and do not naturally capture semantic features. LSTM based RNNs can address this through their ability to deal with continuous valued features such as Word2Vec, Glove, etc. Another advantage of using LSTM lies in its ability to capture the long and short range dependencies through its novel gating structure. This work presents the deep learning based NER using special type of Recurrent Neural Network(RNN) called Bi-directional Long Short-Term Memory(Bi-LSTM). We use a two stage LSTM based network, one acting at character level capturing the n-gram patterns related to NER. Such features are crucial in NER for Indian languages as suffixes used in Indian languages often carry syntactic information. The character based emebeddings, word2vec embeddings and sequence based bi-LSTM embeddings together carry all the requisite features necessary for the NER prediction problem. We present the experimental results on two test datasets from each Indian language such as hindi, kannada, malayalam, tamil and telugu. The accuracies on test-1 datasets of hindi, kannada, malayalam, tamil and telugu languages are 97.82%, 97.04%, 97.46% 97.41% and 97.54% respectively. These are highest accuracy results given by this model when compared with all other models presented by competitors in this shared task [2]. The accuracies on test-2 datasets of hindi, kannada, malayalam, tamil and telugu languages are 97.82%, 96.79%, 96.58% 96.18% and 97.68% respectively. On test-2 dataset this model stood in first position for hindi language and second position for the remaining four languages. The shared task organizers released F-Scores for test-2 datasets of all languages. This model got 94.0%, 84.55%, 84.78%, 89.55% and 91.44% F-Scores on hindi, kannada, malayalam, tamil and telugu languages respectively. All these F-Scores are in second position compared with other models. In overall average accuracy and F-Score of this model on all these five Indian languages is 97.01% and 86.99% which are in second position.

This working note presents a statistical based classifier for text classification using entity relationship information present in the input text. We observed that parts-of-speech tags and named entities information will help us to predict the relationship between entities. We also presented the procedure for predicting POS tags and named entities, which we considered as the sources of information for entity relationship. These features (POS tags, NE) along with the words, in input text sentence, are used as input features to classify the given input into any one of the predefined relationship class. It also presents the experimental details and performance results of this classifier on five Indian language datasets such as hindi, kannada, malayalam, tamil and telugu.

Social Media tools popularized the digital devices among masses making information dissemination easier and faster. Exchange of text is most popular effective means of communication across social media users. It has become necessity to process, understand the semantics of messages communicated as the messages have wide effect across the users. Event extraction refers to understanding the events across streams of social media messages. Event extraction helps in taking quicker corrective actions in case of natural calamities and hence possibly save lives of people. The main objective of the task is, drawing specific knowledge to predict the events (incidents) specified in digital text. We proposed two step procedure to extract events. First phase consists of applying a binary classifier to identify the messages, containing the event. Second phase consists of applying a sequence labeling technique, conditional random fields(CRF), to extract the event from the message. As social media text is noisy, it is a challenge to develop learning algorithms for these tasks. We use Parts of Speech (POS) tags of the words to address some of the issues in this challenge.

The text exchanged in social media conversations is often noisy with a mixture of stylistic and misspelt variations of original words. Any standard NLP techniques applied on such data such as POS tagging, Named entity recognition sufer because of noisy nature of the input. Usage of mixed script text is also prevalent in social media users. The current work addresses the identification of language at word level in mixed script scenarios, where all the text is written in roman script but the words being used by the users are transliterations of original words in native language into english. The core part of the problem is identifying the language, looking at small fragments of text among a set of languages. We propose a two stage approach for word-level language identification. In the frst stage a mixing language combination is identifed by using character n-grams of the sentence. Secondstage consists of using the previous mixing combination class to make the word level language identification. We apply Conditional Random Fields(CRF) further in second stage to improve the performance of the word level language identification. Such simplification is essential, otherwise the number of states of the model will be huge and resultant model predictions are very noisy. Our methods improve the F-score of word level language identification by over 10% compared to the base-line.