Headline generation, a crucial task in summarization, aims to summarize an entire article into a concise, single line. Despite the proficiency of sequence-to-sequence encoder-decoder models and transformer-based large language models (LLMs) in text generation and summarization, generating headlines that include numerals representing the numerals in the news body remains a significant challenge. Generating a numeral-aware headline requires the ability of models to solve numerical and mathematical reasoning capabilities to infer relationships between numerals in the news body. Given the challenges in numeral-aware headline generation and numerical reasoning over numerals in news bodies, this study conducts an empirical investigation of LLMs using various strategies, including Pretrained , Few-shot Prompting , and Chain-of-Thought (CoT) Prompting , for numeral-aware headline generation and numerical reasoning for headline generation. Building upon the insights gained from our empirical study on LLMs for numeral-aware headline generation and numerical reasoning, we propose two novel approaches: instruction tuning with LLMs for numeral-aware headline generation and prompt-based masked language modeling for numerical reasoning. We conducted our experiments on the NumHG dataset. We observed that our proposed method outperforms the Pretrained , Few-shot Prompting , and CoT Prompting setups of LLMs, as well as baseline models from the literature, on both numeral-aware headline generation and numerical reasoning tasks. Observations from the experimental results reveal that our proposed Prompt-based Masked Language Modeling significantly improves the performance of small and medium-sized language models on numerical reasoning tasks. We also study the robustness of our proposed models by evaluating their performance in fact-checking numerical claims and performing numerical reasoning for numeral-aware text summarization. Our findings suggest that the proposed Prompt-based Masked Language Modeling approach is also effective for numerical claim verification and numerical reasoning for numeral-aware text summarization.

Abstract With the surge of deep learning framework, various studies have attempted to address the challenges of sentiment analysis of tweets (data sparsity, under-specificity, noise, and multilingual content) through text and network-based representation learning approaches. However, limited studies on combining the benefits of textual and structural (graph) representations for sentiment analysis of tweets have been carried out. This study proposes a multi-view learning framework ( end-to-end and ensemble-based ) that leverages both text-based and graph-based representation learning approaches to enrich the tweet representation for sentiment classification. The efficacy of the proposed framework is evaluated over three datasets using suitable baseline counterparts. From various experimental studies, it is observed that combining both textual and structural views can achieve better performance of sentiment classification tasks than its counterparts.

In this work, we present a novel approach for link prediction on heterogeneous networks – networks that accommodate multiple types of nodes as well as multiple types of relations among them. Specifically, we propose a multi-view network representation learning framework to incorporate structural intuitions from the underlying graph and enrich the relational representations for link prediction. The method relies on the metapath view, the community view, and the subgraph view between a source and target node pair whose linkage is to be predicted. Furthermore, our proposed model leverages a relation-aware attention mechanism to aggregate the candidate contexts in a principled way. Empirically, we demonstrate that the proposed architecture outperforms state-of-the-art transductive and inductive methods in link prediction by a significant margin. A detailed ablation study and attention weight visualizations suggest that the chosen views are complementary and useful to predict links robustly.

Charts are commonly used forms of visualizing scientific observations from research findings or commercial trends. They provide an abstraction of the underlying information in a more understandable way. Over time, different forms of charts are developed. With the increase in the number of scientific documents present on the internet with different types of charts, automatic chart classification is becoming an important task for various applications. There have been several studies on chart classification with methods ranging from traditional machine learning approaches like SVM, KNN, and HMM to recent deep learning models like VGG, ResNet, and Xception. However, inconsistencies in experimental results are evident. This paper evaluates nine of the recently proposed deep learning-based models on three datasets (one curated and annotated by authors, and two publicly available), and systematically studies their performances over various setups to understand the reason for observing inconsistent results.

Multiplex networks are complex graph structures in which a set of entities are connected to each other via multiple types of relations, each relation representing a distinct layer. Such graphs are used to investigate many complex biological, social, and technological systems. In this work, we present a novel semi-supervised approach for structure-aware representation learning on multiplex networks. Our approach relies on maximizing the mutual information between local node-wise patch representations and label correlated structure-aware global graph representations to model the nodes and cluster structures jointly. Specifically, it leverages a novel cluster-aware, node-contextualized global graph summary generation strategy for effective joint-modeling of node and cluster representations across the layers of a multiplex network. Empirically, we demonstrate that the proposed architecture outperforms state-of-the-art methods in a range of tasks: classification, clustering, visualization, and similarity search on seven real-world multiplex networks for various experiment settings.

Recently, social media has been considered the fastest medium for information broadcasting and sharing. Considering the wide range of applications such as viral marketing, political campaigns, social advertisement, and so on, influencing characteristics of users or tweets have attracted several researchers. It is observed from various studies that influential messages or users create a high impact on a social ecosystem. In this study, we assume that public opinion on a social issue on Twitter carries a certain degree of emotion, and there is an emotion flow underneath the Twitter network. In this article, we investigate social dynamics of emotion present in users’ opinions and attempt to understand (i) changing characteristics of users’ emotions toward a social issue over time, (ii) influence of public emotions on individuals’ emotions, (iii) cause of changing opinion by social factors, and so on. We study users’ emotion dynamics over a collection of 17.65M tweets with 69.36K users and observe 63% of the users are likely to change their emotional state against the topic into their subsequent tweets. Tweets were coming from the member community shows higher influencing capability than the other community sources. It is also observed that retweets influence users more than hashtags, mentions, and replies.

Majority of the social network analysis studies for counter-terrorism and homeland security consider homogeneous network. However, a terrorist activity (attack) is often defined by several attributes such as terrorist organisation, time, place, attack type etc. To capture inherent dependency between the attributes, we need to adopt a network which is capable of capturing the dependency between the attributes. In this paper, we define a heterogeneous network to represent a collection of terrorist activities. Further, we propose personalised PageRank (PPR) as a method capable of performing various analytical operations over heterogeneous network just by changing model parameters without changing the underlying model. Using global terrorist data (GTD), behavioural network, and news discussion network, we show various applications of PPR for counter-terrorism over heterogeneous network just by changing the model parameter. In addition we propose heterogeneous version of four local proximity based link prediction methods, namely, Common Neighbour, Adamic-Adar, Jaccard Coefficient, and Resource Allocation.

Existing studies on evolution of social network largely focus on addition of new nodes and links in the network. However, as network evolves, existing relationships degrade and break down, and some nodes go to hibernation or decide not to participate in any kind of activities in the network where it belongs. Such nodes and links, which we refer as "dull", may affect analysis and prediction tasks in networks. This paper formally defines the problem of predicting dull nodes and links at an early stage, and proposes a novel time aware method to solve it. Pruning of such nodes and links is framed as "network data cleaning" task. As the definitions of dull node and link are non-trivial and subjective, a novel scheme to label such nodes and links is also proposed here. Experimental results on two real network datasets demonstrate that the proposed method accurately predicts potential dull nodes and links. This paper further experimentally validates the need for data cleaning by investigating its effect on the well-known "link prediction" problem.

Majority of the studies on modeling the evolution of a social network using spectral graph kernels do not consider temporal effects while estimating the kernel parameters. As a result, such kernels fail to capture structural properties of the evolution over the time. In this paper, we propose temporal spectral graph kernels of four popular graph kernels namely path counting, triangle closing, exponential and neumann. Their responses in predicting future growth of the network have been investigated in detail, using two large datasets namely Facebook and DBLP. It is evident from various experimental setups that the proposed temporal spectral graph kernels outperform all of their non-temporal counterparts in predicting future growth of the networks.

Most of the search engine optimization techniques attempt to predict users interest by learning from the past information collected from different sources. But, a user's current interest often depends on many factors which are not captured in the past information. In this paper, we attempt to identify user's current interest in real time from the information provided by the user in the current query session. By identifying user's interest in real time, the engine could adapt differently to different users in real time. Experimental verification indicates that our approach is encouraging for short queries

Complete set of itemsets can be grouped into non-overlapping clusters identified by closed tidsets. Each cluster has only one closed itemset and is the superset of all itemsets with the same support. Number of closed itemsets is identical to the number of clusters. Therefore, the problem of discovering closed itemsets can be considered as the problem of clustering the complete set of itemsets by closed tidsets. In this paper, we present CloseMiner, a new algorithm for discovering all frequent closed itemsets by grouping the complete set of itemsets into non-overlapping clusters identified by closed tidsets. An extensive experimental evaluation on a number of real and synthetic databases shows that CloseMiner outperforms Apriori and CHARM.