The proliferation of large language models (LLMs) has brought forth unprecedented capabilities in text generation, yet it is often accompanied by the pervasive issue of ”hallucinations,” meaning plausible but factually incorrect or nonsensical outputs. This paper proposes a novel approach for the identification and conceptual reclassification (correction) of such hallucinations using Graph Co-training, a semi-supervised learning paradigm that leverages the inherent relationships within textual data. We present a methodology that constructs a graph representation of text, where nodes represent entities or concepts, and edges signify their semantic relationships. This graph structure facilitates the propagation of labels and the discovery of latent patterns indicative of hallucinatory content. The proposed Graph Co-training framework integrates multiple views of the text (e.g., lexical, semantic, and factual consistency) and iteratively refines a classifier’s understanding of hallucinations. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art machine learning, deep learning, and even existing co-training and Graph Neural Network (GNN) approaches in accurately identifying hallucinatory text segments. Furthermore, we illustrate a conceptual pipeline for reclassifying (correcting) these identified hallucinations, offering a robust solution to enhance the reliability of LLM-generated content.

Text data has grown exponentially due to social media, e-commerce, and online reviews. Recent studies have explored graph neural networks (GNNs) for text classification, since they are not capable of handling complex structures and preserving global information. To overcome these issues, we propose a GNN-based co-training model that creates graphs based on global parameters instead of single parameters It retains global information while eliminating dependency on text and corpus. In addition, we use smaller text windows to extract more local features. Graph neural networks (GNN) have focused on natural language processing in recent years, outperforming existing models while consuming less memory. Co-training enhances the limited amount of labeled data used to train each classifier with pseudo-labeled data derived from predictions from both the classifiers. Moreover, pseudo-labels cannot be controlled by the co-training of the data. Wrongly trained classifiers can generate inaccurate pseudo-labels that cause noticeable errors when used as training data for other classifiers. This can lead to a model overfitting the training data, which can result in poor generalization of the model when applied to new data. To address this issue, we propose a deep multi-label co-training framework for semi-supervised GNNs (MLD-GNN). Two GNNs are trained simultaneously on the same training data. GNNs use original labeled and pseudo-labeled data for training the model. In order to minimize the negative impact of inaccurate pseudo-labels on learning data augmentation, label augmentation strategies make pseudo-labels available earlier during training.

The co-training algorithm is one of the most common methods of semi-supervised learning in machine learning, which allows multiple learners to collaborate to discover the best information in unlabelled data. Co-training works well if the two views satisfy the sufficiency and independence assumptions. As a result of these assumptions of co-training and advancements in data classification algorithms, the performance of the underlying model could be improved. Specifically, view division, correlation between features in each view, domain knowledge, and label confidence estimation are introduced as key steps in improving co-training algorithms in this paper. Furthermore, we discuss the problems with the co-training methods currently being used, suggest some improvements, and speculate at how the algorithm could be improved going forward.

This paper presents a multi-labeled semi-supervised study that includes a small labeled dataset and an unlabeled dataset. With the co-training framework, we combine a two-view semi-supervised learning for text classification by implementing two neural networks. As in the original co-training paradigm, the text is classified according to two independent learner views using two separate classifiers. In order to extend this idea to deep learning, the deep co-training model uses deep neural networks to train on different views of generated samples to calculate similarity in the probability distribution of predicted outcomes. This co-training framework depends on co-trained networks in order to classify text into multiple labels. Furthermore, the method adds noise to keep the classifier from being affected by it during prediction. As a result, such co-trained networks provide more relevant data and improve classification accuracy. To demonstrate the effectiveness of the developed approach, we compare it to state-of-the-art machine learning and deep learning classifiers on a dataset of survivor stories of human rights violations.

Estimating the effect of an intervention (i.e. causal inference) is a fundamental problem in fields like medicine, e-commerce, government (among others). An ideal approach requires experimental data, which is often hard to find or collect. Observational data on the other hand is in abundance, but contains systematic bias posing a great challenge for this task. In this paper, we introduce the use of two methods pioneered in domain adaptation: Euclidean and non-Euclidean correlation alignment in the context of causal inference. We evaluate their performance using widely used datasets and our results perform favourably against the current state of the art.

The use of short text has become widespread in social media like Twitter and Facebook. Typically, users on social media platforms adopt nonstandard format terms when posting. This introduces challenges for Information Retrieval (IR) and Natural Language Processing (NLP) and standard or classical methods tend not to perform well in this domain. In this paper, we have addressed one of the challenges in IR which is Named Entity Recognition (NER). We introduce a novel probabilistic approach which targets entities occurring in an informal (nonstandard) format within short text. The Probabilistic Named Entity Recognition (PNER) model identifies these entities using cooccurrence patterns. These patterns have been detected using the word cooccurrence embeddings of 278.6 million tweets. The results show an enhancement of 7% on two standard methods when used in combination with PNER. The testing dataset has been created using the standard methods in addition to street names and places taken from the Open Street Map (OSM) database.

With the proliferation of social media data reporting on all aspects of human activity, being able to automatically identify events is becoming increasingly important. Identifying events from data streams is very challenging, as those streams are of unbounded size and their data elements arrive in real-time and at an unpredictable rate. Even more, it is not possible to backtrack over past arrived data elements or maintain and review the entire stream history. In this paper, we present an approach for detecting the daily occurring events in social network streams related to major events using a Frequent Pattern Mining method. In addition, we introduce dynamic support definition method to replace the fixed, a priori given one. As the number of text posts streamed each day is not fixed, a dynamic support, can adapt to the nature of data streams and can improve the identification of events. Experiments were employed on two streams related to two different major events. The first stream was related to the UK General Elections 2015, with more than 1. 1M tweets, and the other was related to the Greece Crisis 2015, with more than 150K tweets. Results showed that using a daily dynamic defined support can better identify events from social network streams rather than fixed ones, as it can cope with the dynamic and changing aspect of data streams.

Forecasting time series that are generated in dynamic environments is challenging due to the characteristic of these series: data generated at high speed and/or large amounts of data which contain multiple variables, and have complex seasonality. Forecasting real-time availability of parking spaces can lead to all sorts of benefits including user satisfaction, energy savings and more efficient use of parking spaces. In this paper, we explorer using deep Gated Recurrent Units (GRUs) to forecast multivariate time series in parking lots setting. Predicting the parking availability mainly depends on the multiple seasonal patterns exhibited in the parking lots and by the real-time occupancy data. In particular, we use the heterogeneous data streams that generated from a limited number of existing Internet of Things (IoT) devices which monitor different parking lots, to construct the parking availability data. Then, we employ GRUs on the parking data to provide predictions across a variety of parking lots. Using the proposed approach, we have provided the real-time parking availability information for different time intervals in six parking lots at University of Essex. Although the results indicate the efficiency of applying GRUs to forecast the multivariate time series, these results show that the simple Multilayer Perceptron (MLPs) can perform better than the GRUs models.

In the recent past, many human rights organizations have started using social media to identify, collect and document human rights violations. To manually extract relevant data from the large corpus of this social network data is difficult and time-consuming and expensive. Furthermore, with the advent of technology, the context and significance of the human rights abuses has and will change over time and advice from experts is needed to perform any kind quantitative analysis on this data. There are applications and systems that help structure this data into relevant categories, but detecting underlying latent patterns, finding similar annotated patterns and continuously upgrading the system to perform exploratory analysis requires high maintenance and cost. This paper proposes a solution to address this problem by integrating semi-supervised learning (with Matrix Factorization) and similarity measures algorithms to classify the large unstructured corpus into stories that have been labelled with one or more types of human rights abuses. In the last few decades, recommender systems have come across as powerful machine learning tools to infer from data and provide value-added content. Along the same context, semi-supervised algorithms mitigate situations where there is a relatively small labelled training data, but a large unlabeled data-set. This paper tries to combine both these algorithms to discover patterns in unlabeled victim survivor stories and recommends labels from other similar stories, thus updating the initial labelled set. The efficiency of the algorithm is evaluated using state of art evaluation metrics. Experimental results show a correlation between new and labelled stories. Real-world results show that the algorithm outplays some of in house recommendation algorithms.

The increasing pace of change in languages affects many applications and algorithms for text processing. Researchers in Natural Language Processing (NLP) have been striving for more generalized solutions that can cope with continuous change. This is even more challenging when applied on short text emanating from social media. Furthermore, increasingly social media have been casting a major influence on both the development and the use of language. Our work is motivated by the need to develop NLP techniques that can cope with short informal text as used in social media alongside the massive proliferation of textual data uploaded daily on social media. In this paper, we describe a novel approach for Short Text Topic Modelling using word embeddings and taking into account any informality of words in the social media text with the aim of addressing the challenge of reducing noise in messy text. We present a new algorithm derived from the Term Frequency -Inverse Document Frequency (TF-IDF), named Term Frequency - Inverse Context Term Frequency (TF-ICTF). TF-ICTF relies on a probabilistic relation between words and context with respect to time. Our experimental work shows promising results against other state-of-the-art methods.

Learning patterns from evolving data streams is challenging due to the characteristics of such streams: being continuous, unbounded and high speed data of non-stationary nature, which must be processed on the fly, using minimal computational resources. An additional challenge is imposed by the imbalanced data streams in many real-world applications, this difficulty becomes more prominent in multi-class learning tasks. This paper investigates the multi-class imbalance problem in non-stationary streams and develops a method to exploit real-time stream data and capture the dynamic of patterns from heterogeneous streams. In particular, we seek to extend concept drift adaptation techniques into imbalanced classes' scenarios, and accordingly, we use an adaptive learner to classify multiple streams over a sequence of titled time windows. We include examples of the falsely classified instances in the training set, then we propose using a dynamic support threshold to discover the frequent patterns in these streams. We conduct an experiment on the car parking lots environment of a typical University with three simulated streams from sensors, smart pay stations and a mobile application. The result indicates the efficiency of applying adaptive learner approaches and modifying the training set to cope with the concept drift in multi-class imbalance scenarios, it also shows the merit of using a dynamic threshold to detect the rare patterns from evolving streams.

In the human rights domain, there is need to filter, efficiently classify and prioritize the types of violation endured by victims in order to provide the necessary rehabilitation and support. However, the domain is dominated by unstructured data either from victims' accounts, doctors'/professionals' reports or available on line. Manual classification still prevails in this domain which is extremely time consuming and slow. This is a problem for non-government operated charities. To this end we have explored the application of the co-training algorithm in order to improve the performance of a semi-supervised learning algorithm by incorporating large amounts of unlabeled data into the training data set. However, it remains challenging to apply co-training on the data without two independent and self sufficient views. This paper puts forth a method of randomly dividing the available features to apply matrix factorization so as to discover latent features underlying the interactions between different kinds of entities present in a single view dataset. These labeled views balance the biased information in the dataset, but still satisfy the co-training assumptions. Alongside, the views are constrained such that pairs of labeled views create weak classifiers which in turn increase the prediction accuracy when combined. In the majority of cases, any classification tries to connect a single class to each sample or object. However, in the human rights domain, a victim can be subjected to more than one type of violation or abuse. This is multi-label classification where a sample can be assigned to more than one class. This paper aims to address all these aspects by bringing together a semi supervised classification model that relies on the effectiveness of matrix collaborative filtering in order to classify stories narrated by victims into one or more types of human rights abuses. Experimental results demonstrate the efficiency of this approach when applied on real-world stories from different victims.

Event Detection from data streams is challenging due to the characteristics of such streams, where data elements arrive in real-time and at high velocity, as well as being streams of unbounded size, even more it is not possible to backtrack over the past arrived data elements or review and keep track of the entire history. Financial time-series streams are a source of financial data (tick data) at fine time scales. In this research, we aim to detect events occurring within time-series streams, and our approach utilises the Directional Change Approach, which summarizes price movements based on a given threshold to detect events. In this paper, we propose a dynamic threshold definition method to be used for detecting the directional change events. The threshold is calculated on a daily basis based on previous day price transitions and the current day opening price. An experiment was run for more than 30 weeks to detect the occurring directional change events on a time-series stream, with one minute data flow levels (one minute frequency) to test our threshold definition method against different fixed threshold values. The detected events were evaluated against news headlines published regarding the studied share on the same day the event was found. The results revealed that a daily dynamic calculated threshold more accurately detected events than different static fixed threshold values.

With the rise of the Semantic Web, several public semantic repositories like Knowledge Bases, Ontologies and Taxonomies have been developed in a variety of domains. For specific domains like the biomedical domain they have already formed a huge valuable infrastructure. On the other hand, the development of efficient algorithms for Natural Language Processing gave us access to the massive knowledge hidden in many unstructured resources. Combining and harvesting these two worlds would result into a very productive knowledge fusion applicable in several domains. In this paper, an extensible framework is presented that focuses on accessing and graphically presenting the knowledge coming from all available structured and unstructured resources. An abstraction formalism for representing any type of query based on graphs is the base of this approach. This formalism makes the framework accessible to non-expert users that have no knowledge of constructing queries in any querying language and barely understand what structured and unstructured resources are. The architecture that will allow for the framework to be adaptable to all available resources is described along with a proof of concept implementation in the biomedical domain.

Traders always seek a trading strategy that can increase and maximize their profits. However, given the current challenges in financial time-series streams - data elements (tick prices) arrive in real-time or almost real-time and at high velocity (at finer time scales) - it is difficult to identify and spot the best time and the most profitable price for trading. The Directional Change (DC) is an event-based approach for summarizing price movements based on a fixed given threshold value. An event in the DC approach is detected if the price change between two points satisfies the given threshold value. In this research, we aim to present a trading strategy based on the DC approach and a dynamic threshold to replace the fixed given one. We call this strategy, the Dynamic Threshold Trading Strategy (DT-TS). Thus, once a DC event is detected (a price change is identified) using the defined dynamic threshold, a trading action is triggered as prices continue to increase or decrease depending on the detected DC event. The trading action to be taken (buy or sell) depends on the previous day price transitions. An experiment was conducted on the FTSE 100 minute-by-minute prices stream to evaluate the DT-TS against different fixed threshold values and different trading strategies. Results showed that the DT-TS was the most profitable strategy among different fixed thresholds and all other examined trading strategies.

Detecting events from streams of data is challenging due to the characteristics of such streams: data elements arrive in real-time and at high velocity, and the size of the streams is typically unbounded while it is not possible to backtrack over past data elements or maintain and review the entire history. Social networks are a good source for event identification as they generate huge amount of timely information representing what users are posting and discussing. In this research, we are developing methods for event detection from streams of data. More specifically, we are presenting a framework for detecting the daily occurring events or topics occurring in social network streams related to major events. Our approach utilizes the Frequent Pattern Mining method to detect the daily occurring frequent patterns, which are going to be our detected events. In addition, we propose a dynamic support definition method to replace the fixed given one. An experiment was run on two streams relating to two different major events to examine the detected events and to test our support definition method. The UK General Elections 2015 stream holds more than one million tweets, and the Greece Crisis 2015 stream contains more than 150k tweets. The detected events were evaluated against news headlines published the same day the event was found. The results revealed that the higher the streaming level (bigger window size), the more accurate the detected events. We also show that for too small sized windows, a more strict support definition method is needed to avoid detecting false or insignificant events.

A concept hierarchy created from a document collection can be used for query recommendation on Intranets by ranking terms according to the strength of their links to the query within the hierarchy. A major limitation is that this model produces the same recommendations for identical queries and rebuilding it from scratch periodically can be extremely inefficient due to the high computational costs. We propose to adapt the model by incorporating query refinements from search logs. Our intuition is that the concept hierarchy built from the collection and the search logs provide complementary conceptual views on the same search domain, and their integration should continually improve the effectiveness of recommended terms. Two adaptation approaches using query logs with and without click information are compared. We evaluate the concept hierarchy models (static and adapted versions) built from the Intranet collections of two academic institutions and compare them with a state-of-the-art log-based query recommender, the Query Flow Graph, built from the same logs. Our adaptive model significantly outperforms its static version and the query flow graph when tested over a period of time on data (documents and search logs) from two institutions' Intranets.

Modelling user interests for time sensitive domains like RSS feeds and spontaneous social media has been a vibrant research activity in recent times. Although numerous efforts have been invested in to the personalisation of dynamic web content, the voluminous and diversified production of continuous online information still poses significant research challenges. In this paper, we propose a profiling mechanism that learns the user access patterns in a dynamic environment like RSS feeds. Main goal of the proposed mechanism is the retrieval optimisation and personalised recommendation of RSS feeds in close to real time. The mechanism allocates personalized time windows based on the learned access patterns and tries to minimize the chances of time-sensitive information from being missed by the user using a delay minimisation algorithm based on Non-homogenous Poisson Process. The mechanism acquires implicit feedback from the user interaction to calculate potential recommendations. The experiments conducted prove the significance of the mechanism in terms of optimised information retrieval and faster adaptation process where it clearly outperforms the other mechanisms in the literature on these properties.

A key feature in developing an effective web personalization system is to build and model dynamic user profiles. In this paper, we propose a multi-agent approach for building a dynamic user profile that is effectively capable of learning and adapting to user behaviour. The main goal is to implicitly track user browsing behaviour in order to extract short-term and long-term user interests. User interests are represented as ontological concepts which are constructed by mapping web pages visited by a user to a reference ontology. In this paper, we focus on the learning and the adaptation processes that are essential in modelling a dynamic user profile. Our proposed model has been integrated with a personalized search system and experiments show that our system is able to effectively model a dynamic user profile that is capable of learning and adapting to user behaviour. Experiments also show that our model achieved a higher performance than non-personalized system.

While much research has been performed on query logs collected for major Web search engines, query log analysis to enhance search on smaller and more focused collections has attracted less attention. Our hypothesis is that an intranet search engine can be enhanced by adapting the search system to real users' search behaviour through exploiting its query logs. In this work we describe how a constantly adapting domain model can be used to identify and capture changes in intranet users' search requirements over time. We employ an algorithm that dynamically builds a domain model from query modifications taken from an intranet query log and employs a decay measure, as used in Machine Learning and Optimisation methods, to promote more recent terms. This model is used to suggest query refinements and additions to users and to elevate seasonally relevant terms. A user evaluation using models constructed from a substantial university intranet query log is provided. Statistical evidence demonstrates the system's ability to suggest seasonally relevant terms over three different academic trimesters. We conclude that log files of an intranet search engine are a rich resource to build adaptive domain models, and in our experiments these models significantly outperform sensible baselines.