In this work, we tackle the multi-modal irony detection problem and propose a lightweight approach that combines pretrained multi-modal and text-based smaller language models. The proposed approach makes use of generated image captions and original text content to fine-tune a transformer based model for text classification. In the proposed solution, Vision-Language Model (VLM) is used for generating image captions, without any fine-tuning. On two benchmark datasets, the experiments demonstrate the benefit of the proposed method against the baseline solutions.

The rapid dissemination of misinformation through social media increased the importance of automated fact-checking. Furthermore, studies on what deep neural models pay attention to when making predictions have increased in recent years. While significant progress has been made in this field, it has not yet reached a level of reasoning comparable to human reasoning. To address these gaps, we propose a multi-task explainable neural model for misinformation detection. More specifically, this work formulates an explanation generation process of the model’s veracity prediction as a text summarization problem. Additionally, the performance of the proposed model is discussed on publicly available datasets and the findings are evaluated with related studies.

In the big data era, automated index selection and recommendation has been an important research problem to improve the data access efficiency. Previous efforts on artificial intelligence based database index selection have focused on relational databases. In this work, we consider the automated index selection for NoSQL databases and investigate the feasibility of supervised learning and reinforcement learning based solutions. The experiments conducted on the YCSB dataset show that reinforcement learning improves index selection performance as in relational databases, and supervised learning gives promising results and can be considered applicable under sufficient amount of training data.

Named Entity Recognition (NER) is an information extraction task that aims to automatically identify named entities in a given text. Named entities are special types of nouns or noun groups that refer to specific entities including as person, location, organization, date, time, money and percentage. NER also facilitates various other Natural Language Processing (NLP) related tasks such as summarization and question answering. It is a vastly studied problem especially on English texts, however number of NER studies on Turkish is very limited. Being a morphologically rich language, Turkish has an agglutinative structure and hence automated analysis and information extraction performance is generally lower than those on English. The previous studies mostly use conventional supervised learning and sequence tagging methods, such as Conditional Random Fields (CRF). Only few studies use deep neural models for NER problem on Turkish texts. In this work, we particularly focus on the recent neural model, Google BERT, and analyze its performance on Turkish texts. In addition to fully trained BERT model, we investigate the performance of different training levels from fully trained to fully pre-trained.

In this paper, we present a framework for predicting the personality traits by analyzing tweets written in Turkish. The prediction model is constructed with a clustering based approach. Since the model is based on linguistic features, it is language specific. The prediction model uses features applicable to Turkish language and related to writing style of Turkish Twitter users. Our approach uses anonymous BIG5 questionnaire scores of volunteer participants as the ground truth in order to generate personality model from Twitter posts. Experiment results show that constructed model can predict personality traits of Turkish Twitter users with relatively small errors.

Automated Telling Machines (ATM) are one of the prominent services of the banks, which facilitate daily banking operations. Among the offered services, money withdrawal is a very basic functionality of ATMs. For the banks, it is important to manage the amount of money to be loaded in ATMs. Hence, prediction of the withdrawal amount is an important step in the ATM management. In this work, we investigate the performance of deep learning techniques for ATM money withdrawal amount prediction problem. The problem is defined in two ways: predicting the amount of money to be withdrawn, and predicting the class label of the withdrawal amount. Our data set includes daily total withdrawal amounts together with the date information. In our experiments, we further analyzed the effect of additional information, such as size of history window, weather condition, currency rate and location. For numeric value prediction task, we compared the prediction performance with statistical models of ARIMA and SARIMA. The experiments show that the neural architectures are feasible for the task of withdrawal amount prediction. They can especially provide high accuracy results when the problem is modeled as a class label prediction task.

Event detection is a popular research problem, aiming to detect events from online data sources with least possible delay. Most of the previous work focus on analyzing textual content such as social media postings to detect happenings. In this work, we consider event detection as a change detection problem in network structure, and propose a method that detects change in community structure extracted from communication network. We study three versions of the method based on different change models. Experimental analysis on benchmark data set reveals that change in the community can be used as an indication of an event.

A recent “third wave” of neural network (NN) approaches now delivers state-of-the-art performance in many machine learning tasks, spanning speech recognition, computer vision, and natural language processing. Because these modern NNs often comprise multiple interconnected layers, work in this area is often referred to as deep learning . Recent years have witnessed an explosive growth of research into NN-based approaches to information retrieval (IR). A significant body of work has now been created. In this paper, we survey the current landscape of Neural IR research, paying special attention to the use of learned distributed representations of textual units. We highlight the successes of neural IR thus far, catalog obstacles to its wider adoption, and suggest potentially promising directions for future research.

Learning to rank is a supervised learning problem that aims to construct a ranking model. The most common application of learning to rank is to rank a set of documents against a query. In this work, we focus on pointwise approach and compare the performances of four computational methods in developing ranking models using several criteria such as accuracy, stability and robustness. The experimental results show that Multivariate Adaptive Regression Splines (MARS) and Artificial Neural Networks (ANN) are effective methods for learning to rank problem and provide promising results.

This paper presents efficient data structures and a pruning technique in order to improve the efficiency of high utility sequential pattern mining. CRoM (Cumulated Rest of Match) based upper bound, which is a tight upper bound on the utility of the candidates is proposed in order to perform more conservative pruning before candidate pattern generation in comparison to the existing techniques. In addition, an efficient algorithm, HuspExt (High Utility Sequential Pattern Extraction), is presented which calculates the utilities of the child patterns based on that of the parents'. Substantial experiments on both synthetic and real datasets from different domains show that, the solution efficiently discovers high utility sequential patterns under low thresholds.

The pervasiveness of location-acquisition technologies enable location-based social networks (LBSN) to become increasingly popular in recent years. Users are able to check-in their current location and share information with other users through these networks. LBSN check-in data can be used for the benefit of users by providing personalized recommendations. There are several location recommendation algorithms that employ LBSN data in the literature. However, there are few number of proposed activity recommendation algorithms. In this paper, we propose a random walk based context-aware activity recommendation algorithm, namely RWCAR, for LBSNs. RWCAR considers the current context (i.e. social relations, personal preferences, and current location) of the user to provide recommendations. We propose a graph model for representing LBSN data that contains users, locations and activities. We build a graph according to the current context of the user depending on this LBSN model. A random walk approach is employed to predict the recommendation scores of the activities. A list of activities are recommended in decreasing order of calculated recommendation score. In experimental evaluation, we compare RWCAR with friend-based, expert-based and popularity-based activity recommendation algorithms. The proposed algorithm performs better in terms of activity recommendation accuracy in all of the experiments.

High utility sequential pattern mining has been considered as an important research problem and a number of relevant algorithms have been proposed for this topic. The main challenge of high utility sequential pattern mining is that, the search space is large and the efficiency of the solutions is directly affected by the degree at which they can eliminate the candidate patterns. Therefore, the efficiency of any high utility sequential pattern mining solution depends on its ability to reduce this big search space, and as a result, lower the computational complexity of calculating the utilities of the candidate patterns. In this paper, we propose efficient data structures and pruning technique which is based on Cumulated Rest of Match (CRoM) based upper bound. CRoM, by defining a tighter upper bound on the utility of the candidates, allows more conservative pruning before candidate pattern generation in comparison to the existing techniques. In addition, we have developed an efficient algorithm, High Utility Sequential Pattern Extraction (HuspExt), which calculates the utilities of the child patterns based on that of the parents'. Substantial experiments on both synthetic and real datasets from different domains show that, the proposed solution efficiently discovers high utility sequential patterns from large scale datasets with different data characteristics, under low utility thresholds.

Sentiments and opinions expressed in Web pages towards objects, entities, and products constitute an important portion of the textual content available in the Web. In the last decade, the analysis of such content has gained importance due to its high potential for monetization. Despite the vast interest in sentiment analysis, somewhat surprisingly, the discovery of sentimental or opinionated Web content is mostly ignored. This work aims to fill this gap and addresses the problem of quickly discovering and fetching the sentimental content present in the Web. To this end, we design a sentiment-focused Web crawling framework. In particular, we propose different sentiment-focused Web crawling strategies that prioritize discovered URLs based on their predicted sentiment scores. Through simulations, these strategies are shown to achieve considerable performance improvement over general-purpose Web crawling strategies in discovery of sentimental Web content.

This paper aims to enhance event detection methods in a micro-blogging platform, namely Twitter. The enhancement technique we propose is based on lexico-semantic expansion of tweet contents while applying document similarity and clustering algorithms. Considering the length limitations and idiosyncratic spelling in Twitter environment, it is possible to take advantage of word similarities and to enrich texts with similar words. The semantic expansion technique we implement is based on syntagmatic and paradigmatic relationships between words, extracted from their co-occurrence statistics. As our technique does not depend on an existing ontology or a lexicon database such as Word Net, it should be applicable for any language. The proposed technique is applied on a tweet set collected for three days from the users in Turkey. The results indicate earlier detection of events and improvements in accuracy.

The sentiments and opinions that are expressed in web pages towards objects, entities, and products constitute an important portion of the textual content available in the Web. Despite the vast interest in sentiment analysis and opinion mining, somewhat surprisingly, the discovery of the sentimental or opinionated web content is mostly ignored. This work aims to fill this gap and address the problem of quickly discovering and fetching the sentimental content present in the Web. To this end, we design a sentiment-focused web crawling framework for faster discovery and retrieval of such content. In particular, we propose different sentiment-focused web crawling strategies that prioritize discovered URLs based on their predicted sentiment scores. Through simulations, these strategies are shown to achieve considerable performance improvement over general-purpose web crawling strategies in discovering sentimental content.