We introduce SampurNER, a fine-grained named entity recognition (FgNER) dataset encompassing all 22 scheduled Indian languages spoken by more than two billion people across various countries. While manual annotation for FgNER resources is often labor-intensive and expensive, distant supervision methods have been employed as a viable solution. However, such datasets are often noisy, with entity mentions tagged with multiple types, requiring computationally intensive noise-aware models for effective FgNER. Moreover, resources for both coarse-grained and fine-grained named entity recognition tasks in Indian languages remain scarce. To address this, we propose an entity-anchored machine translation (EaMaTa) framework that leverages the largest manually annotated English FgNER dataset, FewNERD, to create a large-scale FgNER dataset in 22 languages. On average, the dataset comprises over 153k sentences, 354k entities, and 3.3M tokens in each language. The languages covered are: Assamese (as), Bengali (bn), Bodo (brx), Dogri (doi), Gujarati (gu), Hindi (hi), Kannada (kn), Kashmiri (ks), Konkani (gom), Maithili (mai), Malayalam (ml), Manipuri (mni), Marathi (mr), Nepali (ne), Odia (or), Punjabi (pa), Sanskrit (sa), Santali (sat), Sindhi (sd), Tamil (ta), Telugu (te), and Urdu (ur). Various rigorous analyses and human evaluations confirm the high quality of the dataset and demonstrate the effectiveness of the entity-anchored machine translation framework with up to 9% increase in F1-score against the current state-of-the-art. Additionally, we extend our analysis to zero-shot, multilingual, and cross-lingual settings, investigating the influence of language family and script similarity on cross-lingual FgNER performance.

Argument Mining (AM) involves identifying and extracting Argumentative Components (ACs) and their corresponding Argumentative Relations (ARs). Most of the prior works have broken down these tasks into multiple sub-tasks. Existing end-to-end setups primarily use the dependency parsing approach. This work introduces a generative paradigm-based end-to-end framework argTANL. argTANL frames the argumentative structures into label-augmented text, called Augmented Natural Language (ANL). This framework jointly extracts both ACs and ARs from a given argumentative text. Additionally, this study explores the impact of Argumentative and Discourse markers on enhancing the model’s performance within the proposed framework. Two distinct frameworks, Marker-Enhanced argTANL (ME-argTANL) and argTANL with specialized Marker-Based Fine-Tuning, are proposed to achieve this. Extensive experiments are conducted on three standard AM benchmarks to demonstrate the superior performance of the ME-argTANL.

Argument Mining (AM) helps in automating the extraction of complex argumentative structures such as Argument Components (ACs) like Premise, Claim etc. and Argumentative Relations (ARs) like Support, Attack etc. in an argumentative text. Due to the inherent complexity of reasoning involved with this task, modelling dependencies between ACs and ARs is challenging. Most of the recent approaches formulate this task through a generative paradigm by flattening the argumentative structures. In contrast to that, this study jointly formulates the key tasks of AM in an end-to-end fashion using Autoregressive Argumentative Structure Prediction (AASP) framework. The proposed AASP framework is based on the autoregressive structure prediction framework that has given good performance for several NLP tasks. AASP framework models the argumentative structures as constrained pre-defined sets of actions with the help of a conditional pre-trained language model. These actions build the argumentative structures step-by-step in an autoregressive manner to capture the flow of argumentative reasoning in an efficient way. Extensive experiments conducted on three standard AM benchmarks demonstrate that AASP achieves state-of-the-art (SoTA) results across all AM tasks in two benchmarks and delivers strong results in one benchmark.

Several studies have used distant supervision to create resources for fine-grained entity recognition (FgER) to mitigate the challenges of manual annotation.However, most of these methods are primarily developed for English and cannot be efficiently adapted to many other languages, including Indian languages.Moreover, the emergence of new and unseen entity types deteriorates the performance of the supervised models trained on FgER datasets with different predefined sets of entity types.This work introduces TAFSIL, a taxonomy-adaptable FgER framework to create FgER datasets in six Indian languages.The chosen languages are spoken by more than a billion speakers across various countries.TAFSIL utilizes the high interlink between the knowledge base WikiData and linked corpora Wikipedia through multi-stage heuristics and improves annotation through fuzzy match and quality sentence selection.TAFSIL enables us to create datasets of a total size of around three million samples for six languages Hindi (Hi), Marathi (Mr), Sanskrit (Sa), Tamil (Ta), Telugu (Te), and Urdu (Ur) belonging to two language families Indo-European and Dravidian.We evaluate the robustness of TAFSIL by creating various datasets in four taxonomies FIGER, OntoNotes, HAnDS, and MultiCoNER2.Our extensive experiments suggest the sound quality of the datasets as there is a relative improvement of 83% in average F1 score over zero-shot performance across various FgER state-of-the-art models.The resource is publicly available at https://huggingface.co/datasets/prachuryyaI- ITG/TAFSIL.

Developing Named Entity Recognition (NER) systems for Indian languages has been a long-standing challenge, mainly owing to the requirement of a large amount of annotated clean training instances. This paper proposes an end-to-end framework for NER for Indian languages in a low-resource setting by exploiting parallel corpora of English and Indian languages and an English NER dataset. The proposed framework includes an annotation projection method that combines word alignment score and NER tag prediction confidence score on source language (English) data to generate weakly labeled data in a target Indian language. We employ a variant of the Teacher-Student model and optimize it jointly on the pseudo labels of the Teacher model and predictions on the generated weakly labeled data. We also present manually annotated test sets for three Indian languages: Hindi, Bengali, and Gujarati. We evaluate the performance of the proposed framework on the test sets of the three Indian languages. Empirical results show a minimum 10% performance improvement compared to the zero-shot transfer learning model on all languages. This indicates that weakly labeled data generated using the proposed annotation projection method in target Indian languages can complement well-annotated source language data to enhance performance. Our code is publicly available at https://github.com/aksh555/CL-NERIL.

The large-scale protein-protein interaction (PPI) data has the potential to play a significant role in the endeavor of understanding cellular processes. However, the presence of a considerable fraction of false positives is a bottleneck in realizing this potential. There have been continuous efforts to utilize complementary resources for scoring confidence of PPIs in a manner that false positive interactions get a low confidence score. Gene Ontology (GO), a taxonomy of biological terms to represent the properties of gene products and their relations, has been widely used for this purpose. We utilize GO to introduce a new set of specificity measures: Relative Depth Specificity (RDS), Relative Node-based Specificity (RNS), and Relative Edge-based Specificity (RES), leading to a new family of similarity measures. We use these similarity measures to obtain a confidence score for each PPI. We evaluate the new measures using four different benchmarks. We show that all the three measures are quite effective. Notably, RNS and RES more effectively distinguish true PPIs from false positives than the existing alternatives. RES also shows a robust set-discriminating power and can be useful for protein functional clustering as well.

Reconstruction of time-varying gene regulatory networks underlying a time-series gene expression data is a fundamental challenge in the computational systems biology. The challenge increases multi-fold if the target networks need to be constructed for hundreds to thousands of genes. There have been constant efforts to design an algorithm that can perform the reconstruction task correctly as well as can scale efficiently (with respect to both time and memory) to such a large number of genes. However, the existing algorithms either do not offer time-efficiency, or they offer it at other costs - memory-inefficiency or imposition of a constraint, known as the 'smoothly time-varying assumption'. In this article, two novel algorithms - 'an algorithm for reconstructing Time-varying Gene regulatory networks with Shortlisted candidate regulators - which is Light on memory' (TGS-Lite) and 'TGS-Lite Plus' (TGS-Lite+) - are proposed that are time-efficient, memory-efficient and do not impose the smoothly time-varying assumption. Additionally, they offer state-of-the-art reconstruction correctness as demonstrated with three benchmark datasets. Source Code: https://github.com/sap01/TGS-Lite-supplem/tree/master/sourcecode.

Recent developments in the field of Visual Question Answering (VQA) have witnessed promising improvements in performance through contributions in attention based networks. Most such approaches have focused on unidirectional attention that leverage over attention from textual domain (question) on visual space. These approaches mostly focused on learning high-quality attention in the visual space. In contrast, this work proposes an alternating bi-directional attention framework. First, a question to image attention helps to learn the robust visual space embedding, and second, an image to question attention helps to improve the question embedding. This attention mechanism is realized in an alternating fashion i.e. question-to-image followed by image-to-question and is repeated for maximizing performance. We believe that this process of alternating attention generation helps both the modalities and leads to better representations for the VQA task. This proposal is benchmark on TDIUC dataset and against state-of-art approaches. Our ablation analysis shows that alternate attention is the key to achieve high performance in VOA.

This paper proposes CQ-VQA, a novel two-level hierarchical but end-to-end model to solve the task of visual question answering (VQA). The first level of CQ-VQA, referred to as Question Categorizer (QC), classifies questions to reduce the potential answer search space. The QC uses attended and fused features of the input question and image. The second level, referred to as Answer Predictor (AP), comprises of a set of distinct classifiers corresponding to each question category. Depending on the question category predicted by QC, only one of the classifiers of AP remains active. The loss functions of QC and AP are aggregated together to make it an end-to-end model. The proposed model (CQ-VQA) is evaluated on the TDIUC dataset and is benchmarked against state-of-the-art approaches. Results indicate a competitive or better performance of CQ-VQA.

Rapid advancements in high-throughput technologies have resulted in genome-scale time series datasets. Uncovering the temporal sequence of gene regulatory events, in the form of time-varying gene regulatory networks (GRNs), demands computationally fast, accurate, and scalable algorithms. The existing algorithms can be divided into two categories: ones that are time-intensive and hence unscalable; and others that impose structural constraints to become scalable. In this paper, a novel algorithm, namely 'an algorithm for reconstructing Time-varying Gene regulatory networks with Shortlisted candidate regulators' (TGS), is proposed. TGS is time-efficient and does not impose any structural constraints. Moreover, it provides such flexibility and time-efficiency, without losing its accuracy. TGS consistently outperforms the state-of-the-art algorithms in true positive detection, on three benchmark synthetic datasets. However, TGS does not perform as well in false positive rejection. To mitigate this issue, TGS+ is proposed. TGS+ demonstrates competitive false positive rejection power, while maintaining the superior speed and true positive detection power of TGS. Nevertheless, the main memory requirements of both TGS variants grow exponentially with the number of genes, which they tackle by restricting the maximum number of regulators for each gene. Relaxing this restriction remains a challenge as the actual number of regulators is not known a priori.

The simultaneous administration of multiple drugs increases the probability of interaction among them, as one drug may affect the activities of others. This interaction among drugs may have a positive or negative impact on the therapeutic outcomes. Thus, identification of unknown drug-drug interactions (DDIs) is of significant concern for improving the safety and efficacy of drug consumption. Although multiple DDI resources exist, it is becoming infeasible to maintain these up-to-date manually with the number of biomedical texts growing at a fast pace. Most existing methods model DDI extraction as a classification problem and rely mainly on handcrafted features, and certain features further depend on domain-specific tools. Recently, neural network models using latent features have been demonstrated to yield similar or superior performance compared to existing models. In this study, we present three long short-term memory (LSTM) network models, namely B-LSTM, AB-LSTM, and Joint AB-LSTM. All three models use word and position embedding as latent features; thus, they do not rely on explicit feature engineering. Furthermore, the use of a bidirectional LSTM (Bi-LSTM) network allows for extraction of implicit features from an entire sentence. The two models AB-LSTM and Joint AB-LSTM also apply attentive pooling in the Bi-LSTM layer output in order to assign weights to features. Our experimental results on the SemEval-2013 DDI extraction dataset indicate that the Joint AB-LSTM model produces reasonable performance (F-score: 69.39%) even with the simple architecture.

The task of relation classification in the biomedical domain is complex due to the presence of samples obtained from heterogeneous sources such as research articles, discharge summaries, or electronic health records.It is also a constraint for classifiers which employ manual feature engineering.In this paper, we propose a convolutional recurrent neural network (CRNN) architecture that combines RNNs and CNNs in sequence to solve this problem.The rationale behind our approach is that CNNs can effectively identify coarse-grained local features in a sentence, while RNNs are more suited for long-term dependencies.We compare our CRNN model with several baselines on two biomedical datasets, namely the i2b2-2010 clinical relation extraction challenge dataset, and the SemEval-2013 DDI extraction dataset.We also evaluate an attentive pooling technique and report its performance in comparison with the conventional max pooling method.Our results indicate that the proposed model achieves state-of-the-art performance on both datasets.1

Fine-grained entity type classification (FETC) is the task of classifying an entity mention to a broad set of types.Distant supervision paradigm is extensively used to generate training data for this task.However, generated training data assigns same set of labels to every mention of an entity without considering its local context.Existing FETC systems have two major drawbacks: assuming training data to be noise free and use of hand crafted features.Our work overcomes both drawbacks.We propose a neural network model that jointly learns entity mentions and their context representation to eliminate use of hand crafted features.Our model treats training data as noisy and uses non-parametric variant of hinge loss function.Experiments show that the proposed model outperforms previous stateof-the-art methods on two publicly available datasets, namely FIGER(GOLD) and BBN with an average relative improvement of 2.69% in micro-F1 score.Knowledge learnt by our model on one dataset can be transferred to other datasets while using same model or other FETC systems.These approaches of transferring knowledge further improve the performance of respective models.

Hand-crafted features based on linguistic and domain-knowledge play crucial role in determining the performance of disease name recognition systems.Such methods are further limited by the scope of these features or in other words, their ability to cover the contexts or word dependencies within a sentence.In this work, we focus on reducing such dependencies and propose a domain-invariant framework for the disease name recognition task.In particular, we propose various end-to-end recurrent neural network (RNN) models for the tasks of disease name recognition and their classification into four pre-defined categories.We also utilize convolution neural network (CNN) in cascade of RNN to get character-based embedded features and employ it with word-embedded features in our model.We compare our models with the state-of-the-art results for the two tasks on NCBI disease dataset.Our results for the disease mention recognition task indicate that state-of-the-art performance can be obtained without relying on feature engineering.Further the proposed models obtained improved performance on the classification task of disease names.

Recent studies have shown that incorporation of available biological information often leads to biologically more relevant results. Motivated by such studies, we extend template based clustering algorithm to incorporate functional annotation information available for genes. Functional similarities between two genes are calculated based on their annotation in the Gene Ontology (GO) database. To these end three methods of calculating functional similarity are explored. We have measured the correlation between average pairwise similarity score and average membership function values to check the validity of assumption that biologically and functionally related genes are also similar in their expression profiles as well as in their GO functional annotation. We observe that Jiang and Conrath's measure is highly correlated with average membership function value of genes. So we use this method for further analysis. With the incorporation of functional similarity score, we have more choices for the objective function to find out the best clustering of gene expression data. We have performed a comparative study to find the combination of objective functions that leads to more biologically relevant information. We have found that different choices of the objective function lead to different sets of templates, while some common templates are identified by all of them. Based on the aim of the study we suggest either to use all three objectives or to use the two objectives related to functional similarity and quantization error.

Conventional clustering algorithms based on Euclidean distance or Pearson correlation coefficient are not able to include order information in the distance metric and also unable to distinguish between random and real biological patterns. We present template based clustering algorithm for time series gene expression data. Template profiles are defined based on up-down regulation of genes between consecutive time points. Assignment of genes to templates is based on fuzzy membership function. Multi-objective evolutionary algorithm is used to determine compact clusters with varying number of templates. Statistical significance of each template is determined using permutation based non-parametric test. Statistically significant profiles are further tested for their biological relevance using gene ontology analysis. The algorithm was able to distinguish between real and noisy pattern when tested on artificial and real biological data. The proposed algorithm has shown better or similar performance compared to STEM and better than k-means on a real biological data.

Protein-DNA interactions play a crucial role in transcriptional regulation and other biological processes. Quantitative predictive models of protein-DNA binding affinities can increase our understanding of molecular interaction and help validate putative transcription factor binding sites or other regulatory features. Such predictive models must take into account context-specific features associated with both DNA and proteins. Given the large complexity associated with such features, here we consider only the contextual features of DNA associated with binding affinity. Two types of features are considered in this paper: 1) features accounting for conformational and physico-chemical properties of nucleotide sequence and 2) another set of features accounting for conservation of evolutionary information in the form of position-specific weight matrices. A feature selection approach, named, leave-one-out sequential forward selection (LOOSFS), is presented. The feature selection method employs leave-one-out cross-validation error of the least square support vector machines (LS-SVM) to estimate the test error of quantitative prediction model. The method is used to identify important features possibly responsible for differences in transcriptional activities of 130 DNA sequences. These sequences were obtained by single base substitutions within promoter of the mouse beta-major globin gene. The selected features and predicted activity values correlate well with experimental results

Due to an increasing interest in solving real-world optimization problems using evolutionary algorithms (EAs), researchers have developed a number of real-parameter genetic algorithms (GAs) in the recent past. In such studies, the main research effort is spent on developing an efficient recombination operator. Such recombination operators use probability distributions around the parent solutions to create offspring. Some operators emphasize solutions at the center of mass of parents and some around the parents. We propose a generic parent-centric recombination operator (PCX) and compare its performance with a couple of commonly-used mean-centric recombination operators (UNDX and SPX). With the help of a steady-state, elite-preserving, and computationally fast EA model, simulation results show the superiority of PCX on three test problems.

Due to increasing interest in solving real-world optimization problems using evolutionary algorithms (EAs), researchers have recently developed a number of real-parameter genetic algorithms (GAs). In these studies, the main research effort is spent on developing an efficient recombination operator. Such recombination operators use probability distributions around the parent solutions to create an offspring. Some operators emphasize solutions at the center of mass of parents and some around the parents. In this paper, we propose a generic parent-centric recombination operator (PCX) and a steady-state, elite-preserving, scalable, and computationally fast population-alteration model (we call the G3 model). The performance of the G3 model with the PCX operator is investigated on three commonly used test problems and is compared with a number of evolutionary and classical optimization algorithms including other real-parameter GAs with the unimodal normal distribution crossover (UNDX) and the simplex crossover (SPX) operators, the correlated self-adaptive evolution strategy, the covariance matrix adaptation evolution strategy (CMA-ES), the differential evolution technique, and the quasi-Newton method. The proposed approach is found to consistently and reliably perform better than all other methods used in the study. A scale-up study with problem sizes up to 500 variables shows a polynomial computational complexity of the proposed approach. This extensive study clearly demonstrates the power of the proposed technique in tackling real-parameter optimization problems.