This paper presents MOD-kt, a modular framework designed to bridge the gap between modern language models and key-term-based document clustering. One of the main challenges of using neural language models for key-term-based clustering is the mismatch between the interpretability of the underlying document representation (i.e. document embeddings) and the more intuitive semantic elements that allow the user to guide the clustering process (i.e. key-terms). Our framework acts as a communication layer between word and document models, enabling key-term-based clustering in the context of document and word models with a flexible and adaptable architecture. We report a comparison of the performance of multiple neural language models on clustering, considering a selected range of relevance metrics. Additionally, a qualitative user study was conducted to illustrate the framework's potential for intuitive user-guided quality clustering of document collections.

This study focuses on the importance of well-designed online matching systems for job seekers and employers. We treat resumes and job descriptions as documents. Then, calculate their similarity to determine the suitability of applicants, and rank a set of resumes based on their similarity to a specific job description. We employ Siamese Neural Networks, comprised of identical sub-network components, to evaluate the semantic similarity between documents. Our novel architecture integrates various neural network architectures, where each sub-network incorporates multiple layers such as CNN, LSTM and attention layers to capture sequential, local and global patterns within the data. The LSTM and CNN components are applied concurrently and merged together. The resulting output is then fed into a multi-head attention layer. These layers extract features and capture document representations. The extracted features are then combined to form a unified representation of the document. We leverage pre-trained language models to obtain embeddings for each document, which serve as a lower-dimensional representation of our input data. The model is trained on a private dataset of 268,549 real resumes and 4,198 job descriptions from twelve industry sectors, resulting in a ranked list of matched resumes. We performed a comparative analysis involving our model, Siamese CNN (S-CNNs), Siamese LSTM with Manhattan distance, and a BERT-based sentence transformer model. By combining the power of language models and the novel Siamese architecture, this approach leverages both strengths to improve document ranking accuracy and enhance the matching process between job descriptions and resumes. Our experimental results demonstrate that our model outperforms other models in terms of performance.

Multi-Task Learning (MTL) for text classification takes advantage of the data to train a single shared model with multiple task-specific layers on multiple related classification tasks to improve its generalization performance. We choose pre-trained language models (BERT-family) as the shared part of this architecture. Although they have achieved noticeable performance in different downstream NLP tasks, their performance in an MTL setting for the biomedical domain is not thoroughly investigated. In this work, we investigate the performance of BERT-family models in different MTL settings with Open-I (radiology reports) and OHSUMED (PubMed abstracts) datasets. We introduce the MTLV (Multi-Task Learning Visualizer) library for building Multi-task learning-related architectures which use existing infrastructure (e.g., Hugging Face Transformers and MLflow Tracking). Following previous work in computer vision, we clustered tasks and trained a separate model on each cluster (Grouped Multi-Task Learning (GMTL)). Contextual representation of the class labels (Tasks) and their descriptions was used by the library as features to cluster the tasks. We observed that grouping tasks for training with few models (GMTL) outperforms the MTL also GMTL is computationally more efficient than the STL setting (a separate model is trained for each task).

Short text stream clustering is an important but challenging task since massive amount of text is generated from different sources such as micro-blogging, question-answering, and social news aggregation websites. The two major challenges of clustering such massive amount of text is to cluster them within a reasonable amount of time and to achieve better clustering result. To overcome these two challenges, we propose an efficient short text stream clustering algorithm (called EStream) consisting of two modules: online and offline. The online module of EStream algorithm assigns a text to a cluster one by one as it arrives. To assign a text to a cluster it computes similarity between a text and a selected number of clusters instead of all clusters and thus significantly reduces the running time of the clustering of short text streams. EStream assigns a text to a cluster (new or existing) using the dynamically computed similarity thresholds. Thus EStream efficiently deals with the concept drift problem. The offline module of EStream algorithm enhances the distributions of texts in the clusters obtained by the online module so that the upcoming short texts can be assigned to the appropriate clusters.

Positive and Unlabeled Learning (PUL) uses unlabeled documents and a few positive documents for retrieving a set of "interest" documents from a text collection. Usually, PUL approaches are based on the vector space model. However, when dealing with semi-supervised learning for text classification or information retrieval, graph-based approaches have been proved to outperform vector space model-based approaches. So, in this article, a graph-based approach for PUL is proposed: Label Propagation for Positive and Unlabeled Learning (LP-PUL). The proposed framework consists of three steps: (i) building a similarity graph, (ii) identifying reliable negative documents, and (iii) performing label propagation to classify the remaining unlabeled documents as positive or negative. We carried out experiments to measure the impact of the different choices in each step of the proposed framework. We also demonstrated that the proposal surpasses the classification performance of other PUL (RC-SVM, PU-LP, and PE-PUC) or one-class learning (k-NN-based, k-Means-based, and Dense Autoencoder) algorithms in terms of F1. Considering the best results of any algorithm used in the experimental evaluation, PU-PUL can improve the classification performance from 2%, when using only 1 labeled document, to 28%, when 30 labeled documents are employed.

In this paper we describe a framework to uncover potential causal relations between event mentions from streaming text of news media. This framework relies on a dataset of manually labeled events to train a recurrent neural network for event detection. It then creates a time series of event clusters, where clusters are based on BERT contextual word embedding representations of the identified events. Using these time series dataset, we assess four methods based on Granger causality for inferring causal relations. Granger causality is a statistical concept of causality that is based on forecasting. It states that a cause occurs before the effect, and the cause produces unique changes in the effect, so past values of the cause help predict future values of the effect. The four analyzed methods are the pairwise Granger test, VAR(1), BigVar and SiMoNe. The framework is applied to the New York Times dataset, which covers news for a period of 246 months. This preliminary analysis delivers important insights into the nature of each method, identifies differences and commonalities, and points out some of their strengths and weaknesses.

Short text stream clustering is an important but challenging task since massive amounts of text are generated from different social media. Given streams of texts, the proposed method clusters the streams of texts based on the frequently occurring word pairs (not necessarily consecutive) in texts. It detects outliers in the clusters and reassigns the outliers to appropriate clusters using the semantic similarity between the outliers and the clusters based on the dynamically computed similarity thresholds. Thus the proposed method efficiently deals with the concept drift problem. Experimental results demonstrate that the proposed approach outperforms the state-of-the-art short text stream clustering algorithms by a statistically significant margin on several short text datasets.

Genetic Programming techniques have demonstrated great potential in dealing with the problem of query generation. This work explores different Multi-Objective Genetic Programming strategies for evolving a collection of topic-based Boolean queries. It compares three approaches to build topical Boolean queries: using terms, incorporating Wikipedia semantics (Wikipedia concepts) and a hybrid approach, using a combination of both terms and concepts. In addition, different fitness functions are combined giving rise to seven multi-objective schemes. In particular, we investigate the use of the proposed strategies in conjunction with novel fitness functions aimed at attaining high diversity based on the information-theoretic notion of entropy and Jaccard similarity. Experiments were completed using 25 topics from a dataset consisting of approximately 350,000 webpages classified into 448 topics. The results reveal that the use of Wikipedia concepts does not result in statistically significant improvements in precision, global recall or diversity when compared to the term-based approaches. However, the use of concepts has a positive effect on query interpretability since the use of terms leads to artificial queries that are hard to interpret by humans. In the meantime, concept-based queries contain a smaller number of operands than the term-based ones, hence resulting in better execution times without a loss in retrieval performance.

In this paper, we propose TRIVIR, a novel interactive visualization tool powered by an Information Retrieval (IR) engine that implements an active learning protocol to support IR with high recall. The system integrates multiple graphical views in order to assist the user identifying the relevant documents in a collection, including a content-based similarity map obtained with multidimensional projection techniques. Given representative documents as queries, users can interact with the views to label documents as relevant/not relevant, and this information is used to train a machine learning (ML) algorithm which suggests other potentially relevant documents on demand. TRIVIR offers two major advantages over existing visualization systems for IR. First, it merges the ML algorithm output into the visualization, while supporting several user interactions in order to enhance and speed up its convergence. Second, it tackles the problem of vocabulary mismatch, by providing term's synonyms and a view that conveys how the terms are used within the collection. Besides, TRIVIR has been developed as a flexible front-end interface that can be associated with distinct text representations and multidimensional projection techniques. We describe two use cases conducted with collaborators who are potential users of TRIVIR. Results show that the system simplified the search for relevant documents in large collections, based on the context in which the terms occur.

In this paper, we propose a high recall active document retrieval system for a class of applications involving query documents, as opposed to key terms, and domain-specific document corpora. The output of the model is a list of documents retrieved based on the domain expert feedback collected during training. A modified version of Bag of Word (BoW) representation and a semantic ranking module, based on Google n-grams, are used in the model. The core of the system is a binary document classification model which is trained through a continuous active learning strategy. In general, finding or constructing training data for this type of problem is very difficult due to either confidentiality of the data, or the need for domain expert time to label data. Our experimental results on the retrieval of Call For Papers based on a manuscript demonstrate the efficacy of the system to address this application and its performance compared to other candidate models.

Attention guides computation to focus on important parts of the input data. For pairwise input, existing attention approaches tend to bias towards trivial repetitions (e.g. punctuations and stop words) between two texts, and thus failed to contribute reasonable guidance to model predictions. As a remedy, we suggest taking into account the corpus-level information via global-aware attention. In this paper, we propose an attention mechanism that makes use of intratext, inter-text and global contextual information. We undertake an ablation study on paraphrase identification, and demonstrate that the proposed attention mechanism can obviate the downsides of trivial repetitions and provide interpretable word weightings.

Short text clustering is an important but challenging task. We investigate impact of similarity matrix sparsification on the performance of short text clustering. We show that two sparsification methods (the proposed Similarity Distribution based, and k-nearest neighbors) that aim to retain a prescribed number of similarity elements per text, improve hierarchical clustering quality of short texts for various text similarities. These methods using a word embedding based similarity yield competitive results with state-of-the-art methods for short text clustering especially for general domain, and are faster than the main state-of-the-art baseline.

The worldwide adoption of Twitter turned it into one of the most popular platforms for content analysis as it serves as a gauge of the public’s feeling and opinion on a variety of topics. This is particularly true of political discussions and lawmakers’ actions and initiatives. Yet, one common but unrealistic assumption is that the data of interest for analysis is readily available in a comprehensive and accurate form. Data need to be retrieved, but due to the brevity and noisy nature of Twitter content, it is difficult to formulate user queries that match relevant posts that use different terminology without introducing a considerable volume of unwanted content. This problem is aggravated when the analysis must contemplate multiple and related topics of interest, for which comments are being concurrently posted. This article presents Active Tweet Retrieval Visualization (ATR-Vis), a user-driven visual approach for the retrieval of Twitter content applicable to this scenario. The method proposes a set of active retrieval strategies to involve an analyst in such a way that a major improvement in retrieval coverage and precision is attained with minimal user effort. ATR-Vis enables non-technical users to benefit from the aforementioned active learning strategies by providing visual aids to facilitate the requested supervision. This supports the exploration of the space of potentially relevant tweets, and affords a better understanding of the retrieval results. We evaluate our approach in scenarios in which the task is to retrieve tweets related to multiple parliamentary debates within a specific time span. We collected two Twitter datasets, one associated with debates in the Canadian House of Commons during a particular week in May 2014, and another associated with debates in the Brazilian Federal Senate during a selected week in May 2015. The two use cases illustrate the effectiveness of ATR-Vis for the retrieval of relevant tweets, while quantitative results show that our approach achieves high retrieval quality with a modest amount of supervision. Finally, we evaluated our tool with three external users who perform searching in social media as part of their professional work.

TrWP is a text relatedness measure that computes semantic similarity between words and phrases utilizing aggregated statistics from the Google Web 1T 5-gram corpus. The phrase similarity computation in TrWP is costly in terms of both time and space, making the existing implementation of TrWP impractical for real-world usage. In this work, we present an in-memory computational framework for TrWP, which optimizes the corpus search using perfect hashing and minimizes the required memory cost using variable length encoding. Evaluated using the Google Web 1T 5-gram corpus, we demonstrate that the computational speed of our framework outperforms a file-based implementation by several orders of magnitude.

We introduce a (semi-)automatic OCR post-processing system that utilizes web-scale linguistic corpora in providing high-quality correction. This paper is a comprehensive system overview with the focus on the computational procedures, applied linguistic analysis, and processing optimization.

Radiology reports are one of the most important medical documents that a diagnostician looks into, especially in the emergency context. They provide the emergency physicians with critical information regarding the condition of the patient and help the physicians take immediate action on urgent conditions. However, the reports are in the form of unstructured text, which makes them time consuming for humans to interpret. We have developed a machine learning system to (a) efficiently extract the clinically significant parts and their level of importance in radiology reports, and (b) to classifies the overall report into critical or non-critical categories which help doctors to identify potential high priority reports. As a starting point, the system uses anonymized chest X-RAY reports of adults and provides three levels of importance for medical phrases. We used the Conditional Random Field (CRF) model to identify clinically significant phrases with an average f1-score of 0.75. The proposed system includes a web-based interface which highlights the medical phrases, and their level of importance to the emergency physician. The overall classification of the report is performed using the phrases extracted from the CRF model as features for the classifier. Average accuracy achieved is 85%.

As the spread of rumours has been increasing every day in online social networks (OSNs), it is important to analyze and understand this phenomenon. Damage caused by the spread of rumours is difficult to handle without a full understanding of the dynamics behind it. One of the central steps of understanding rumour spread is to analyze who spread rumours online, why, and how. In this research, we focus on the steps who and why by describing, implementing, and evaluating an approach that studies whether or not a group of users is actively involved in rumour discussions, and assesses rumour-spreading personality types in OSNs. We implement this general approach using Reddit data, and demonstrate its use by determining which users engage with a recurring rumour, and analyzing their comments using qualitative methods. We find that we can reliably classify users into one of three categories: (1) “Generally support a false rumour”, (2) “Generally refute a false rumour”, or (3) “Generally joke about a false rumour”. Combining text mining techniques, such as text classification, sentiment analysis, and social network analysis, we aim to identify and classify those rumour-spreading user categories automatically and provide a more holistic view of rumour spread in OSNs.

Twitter has become one of the most important platforms for gathering information, where users follow breaking news, track ongoing events and learn about their topics of interest. Considering the sheer volume of Twitter data and the ever-growing number of users, it is of great importance to have real-time systems that can monitor and recommend relevant and non-redundant tweets with respect to users' interests. In this paper, we propose a framework using language models as a basis for analyzing strategies and techniques for tweet recommendation based on user interest profiles. Results show that identifying named entities in profiles has a major impact on the accuracy of the recommender. We also performed a thorough comparison to investigate whether state-of-the-art semantic relatedness techniques have a positive impact on the precision of the recommended tweets. The TREC 2015 Microblog track dataset is used for comparison and evaluation throughout this paper.

By reflecting the degree of proximity or remoteness of documents, similarity measure plays the key role in text analytics. Traditional measures, e.g. cosine similarity, assume that documents are represented in an orthogonal space formed by words as dimensions.

The spread of "rumours" in Online Social Networks (OSNs) has grown at an alarming rate. Consequently, there is an increasing need to improve understanding of the social and technological processes behind this trend. The first step in detecting rumours is to identify and extract memes, a unit of information that can be spread from person to person in OSNs. This paper proposes four similarity scores and two novel strategies to combine those similarity scores for detecting the spread of memes in OSNs, with the end goal of helping researchers as well as members of various OSNs to study the phenomenon. The two proposed strategies include: (1) automatically computing the similarity score weighting factors for four elements of a submission and (2) allowing users to engage in the clustering process and filter out outlier submissions, modify submission class labels, or assign different similarity score weight factors for various elements of a submission using a visualization prototype. To validate our approach, we collect submissions on Reddit about five controversial topics and demonstrate that the proposed strategies outperform the baseline.

Measuring document relatedness using unsupervised co-occurrence based word relatedness methods is a processing-time and memory consuming task. This paper introduces the application of compact data structures for efficient computation of word relatedness based on corpus statistics. The data structure is used to efficiently lookup: (1) the corpus statistics for the Common Word Relatedness Approach, (2) the pairwise word relatedness for the Algorithm Specific Word Relatedness Approach. These two approaches significantly accelerate the processing time of word relatedness methods and reduce the space cost of storing co-occurrence statistics in memory, making text mining tasks like classification and clustering based on word relatedness practical.

Technical writing in professional environments, such as user manual authoring for new products, is a task that relies heavily on reuse of content. Therefore, technical content is typically created following a strategy where modular units of text have references to each other. One of the main challenges faced by technical authors is to avoid duplicating existing content, as this adds unnecessary effort, generates undesirable inconsistencies, and dramatically increases maintenance and translation costs. However, there are few computational tools available to support this activity. This paper investigates the use of different similarity methods for the task of identification of reuse opportunities in technical writing. We evaluated our results using existing ground truth as well as feedback from technical authors. Finally, we also propose a tool that combines text similarity algorithms with interactive visualizations to aid authors in understanding differences in a collection of topics and identifying reuse opportunities.

Online social streams such as Twitter timelines, forum discussions and email threads have emerged as important channels for information propagation. Mining transient stories and their correlations implicit in social streams is a challenging task, since these streams are noisy and surge quickly. In this paper, we propose CAST, which is a context-aware story-teller that discovers new stories from social streams and tracks their structural context on the fly to build a vein of stories. More precisely, we model the social stream as a capillary network, and define stories by a new cohesive subgraph type called (k,d)-Core in the capillary network. We propose deterministic and randomized context search to support the iceberg query, which builds the story vein as social streams flow. We perform detailed experimental study on real Twitter streams and the results demonstrate the creativity and value of our approach.

This paper presents P-GTM, a privacy-preserving text similarity algorithm that extends the Google Tri-gram Method (GTM). The Google Tri-gram Method is a high-performance unsupervised semantic text similarity method based on the use of context from the Google Web 1T n-gram dataset. P-GTM computes the semantic similarity between two input bag-of-words documents on public cloud hardware, without disclosing the documents' contents. Like the GTM, P-GTM requires the uni-gram and tri-gram lists from the Google Web 1T n-gram dataset as additional inputs. The need for these additional lists makes private computation of GTM text similarities a challenging problem. P-GTM uses a combination of pre-computation, encryption, and randomized preprocessing to enable private computation of text similarities using the GTM. We discuss the security of the algorithm and quantify its privacy using standard and real life corpora.

No abstract available.

Most text clustering algorithms represent a corpus as a document-term matrix in the bag of words model. The feature values are computed based on term frequencies in documents and no semantic relatedness between terms is considered. Therefore, two semantically similar documents may sit in different clusters if they do not share any terms. One solution to this problem is to enrich the document representation using an external resource like Wikipedia. We propose a new way to integrate Wikipedia concepts in partitional text document clustering in this work. A text corpus is first represented as a document-term matrix and a document-concept matrix. Terms that exist in the corpus are then clustered based on the document-term representation. Given the term clusters, we propose two methods, one based on the document-term representation and the other one based on the document-concept representation, to find two sets of seed documents. The two sets are then used in our text clustering algorithm in an ensemble approach to cluster documents. The experimental results show that even though the document-concept representations do not result in good document clusters per se, integrating them in our ensemble approach improves the quality of document clusters significantly.

Dynamic networks are commonly found in the current web age. In scenarios like social networks and social media, dynamic networks are noisy, are of large-scale and evolve quickly. In this paper, we focus on the cluster evolution tracking problem on highly dynamic networks, with clear application to event evolution tracking. There are several previous works on data stream clustering using a node-by-node approach for maintaining clusters. However, handling of bulk updates, i.e., a subgraph at a time, is critical for achieving acceptable performance over very large highly dynamic networks. We propose a subgraph-by-subgraph incremental tracking framework for cluster evolution in this paper. To effectively illustrate the techniques in our framework, we consider the event evolution tracking task in social streams as an application, where a social stream and an event are modeled as a dynamic post network and a dynamic cluster respectively. By monitoring through a fading time window, we introduce a skeletal graph to summarize the information in the dynamic network, and formalize cluster evolution patterns using a group of primitive evolution operations and their algebra. Two incremental computation algorithms are developed to maintain clusters and track evolution patterns as time rolls on and the network evolves. Our detailed experimental evaluation on large Twitter datasets demonstrates that our framework can effectively track the complete set of cluster evolution patterns from highly dynamic networks on the fly.

We propose an interactive text document method, which is based on term labeling. The algorithm asks the user to cluster the top keyterms associated with document clusters iteratively. The keyterm clusters are used to guide the clustering method. Rather than using standard clustering algorithms, we propose a new text clusterer using term clusters. Terms that exist in a document corpus are clustered. Using a greedy approach, the term clusters are distilled in order to remove non-discriminative general terms. We then present a heuristic approach to extract seed documents associated with each distilled term cluster. These seeds are finally used to cluster all documents. We compared our interactive term labeling to a baseline interactive term selection algorithm on some real standard text datasets. The experiments show that with a comparable amount of user effort, our term labeling is more effective than the baseline term selection method.

It is often desirable to identify the concepts that are present in a corpus. A popular way to deal with this objective is to discover clusters of words or topics, for which many algorithms exist in the literature. Yet most of these methods lack the interpretability that would enable interaction with a user not familiar with their inner workings. The paper proposes a graph-based topic extraction algorithm, which can also be viewed as a soft-clustering of words present in a given corpus. Each topic, in the form of a set of words, represents an underlying concept in the corpus. The method allows easy interpretation of the clustering process, and hence enables the scope of user involvement at various steps. For a quantitative evaluation of the topics extracted, we use them as features to get a compact representation of documents for classification tasks. We compare the classification accuracy achieved by a reduced feature set obtained with our method versus other topic extraction techniques, namely Latent Dirichlet Allocation and Non-negative Matrix Factorization. While the results from all the three algorithms are comparable, the speed and easy interpretability of our algorithm makes it more appropriate to be used interactively by lay users.

We propose a model for assigning Wikipedia Concepts as scientific category labels to scientific documents where their terms are first grouped together using the well-known topic modelling method, Latent Dirichlet Allocation (LDA) and then assigned to Wikipedia Concepts by wikification. We wikify the terms of the topic model of a document to extract related concepts from Wikipedia. We experiment on two different datasets: the abstracts of the documents from the ACM Digital Library and the full papers of the UvT Collection. The ACM dataset includes Computer Science publications whereas UvT includes scientific publications from a range of topics. Domain specific taxonomies are used for evaluation. Results show that our approach is able to assign Wikipedia Concepts to the scientific publications in an automated manner, removing any need for human supervision.

Online social streams such as Twitter/Facebook timelines and forum discussions have emerged as prevalent channels for information dissemination. As these social streams surge quickly, information overload has become a huge problem. Existing keyword search engines on social streams like Twitter Search are not successful in overcoming the problem, because they merely return an overwhelming list of posts, with little aggregation or semantics. In this demo, we provide a new solution called \keysee by grouping posts into events, and track the evolution patterns of events as new posts stream in and old posts fade out. Noise and redundancy problems are effectively addressed in our system. Our demo supports refined keyword query on evolving events by allowing users to specify the time span and designated evolution pattern. For each event result, we provide various analytic views such as frequency curves, word clouds and GPS distributions. We deploy \keysee on real Twitter streams and the results show that our demo outperforms existing keyword search engines on both quality and usability.

The potential for semi-supervised techniques to produce personalized clusters has not been explored. This is due to the fact that semi-supervised clustering algorithms used to be evaluated using oracles based on underlying class labels. Although using oracles allows clustering algorithms to be evaluated quickly and without labor intensive labeling, it has the key disadvantage that oracles always give the same answer for an assignment of a document or a feature. However, different human users might give different assignments of the same document and/or feature because of different but equally valid points of view. In this paper, we conduct a user study in which we ask participants (users) to group the same document collection into clusters according to their own understanding, which are then used to evaluate semi-supervised clustering algorithms for user personalization. Through our user study, we observe that different users have their own personalized organizations of the same collection and a user's organization changes over time. Therefore, we propose that document clustering algorithms should be able to incorporate user input and produce personalized clusters based on the user input. We also confirm that semi-supervised algorithms with noisy user input can still produce better organizations matching user's expectation (personalization) than traditional unsupervised ones. Finally, we demonstrate that labeling keywords for clusters at the same time as labeling documents can improve clustering performance further compared to labeling only documents with respect to user personalization.

Message type or message cluster extraction is an important task in the analysis of system logs in computer networks. Defining these message types automatically facilitates the automatic analysis of system logs. When the message types that exist in a log file are represented explicitly, they can form the basis for carrying out other automatic application log analysis tasks. In this paper, we introduce a novel algorithm for carrying out message type extraction from event log files. IPLoM, which stands for Iterative Partitioning Log Mining, works through a 4-step process. The first three steps hierarchically partition the event log into groups of event log messages or event clusters. In its fourth and final stage, IPLoM produces a message type description or line format for each of the message clusters. IPLoM is able to find clusters in data irrespective of the frequency of its instances in the data, it scales gracefully in the case of long message type patterns and produces message type descriptions at a level of abstraction, which is preferred by a human observer. Evaluations show that IPLoM outperforms similar algorithms statistically significantly.

Despite all of the advantages of tags as an easy and flexible information management approach, tagging is a cumbersome task. A set of descriptive tags has to be manually entered by users whenever they post a resource. This process can be simplified by the use of tag recommendation systems. Their objective is to suggest potentially useful tags to the user. We present a hybrid tag recommendation system together with a scalable, highly efficient system architecture. The system is able to utilize user feedback to tune its parameters to specific characteristics of the underlying tagging system and adapt the recommendation models to newly added content. The evaluation of the system on six real-life datasets demonstrated the system’s ability to combine tags from various sources (e.g., resource content or tags previously used by the user) to achieve the best quality of recommended tags. It also confirmed the importance of parameter tuning and content adaptation. A series of additional experiments allowed us to better understand the characteristics of the system and tagging datasets and to determine the potential areas for further system development.

The objective of a tag recommendation system is to propose a set of tags for a resource to ease the tagging process done manually by a user. Tag recommendation is an interesting and well defined research problem. However, while solving it, it is easy to forget about its practical implications. We discuss the practical aspects of tag recommendation and propose a system that successfully addresses the problem of learning in tag recommendation, without sacrificing efficiency. Learning is realized in two aspects: adaptation to newly added posts and parameter tuning. The content of each added post is used to update the resource and user profiles as well as associations between tags. Parameter tuning allows the system to automatically adjust the way tag sources (e.g., content related tags or user profile tags) are combined to match the characteristics of a specific collaborative tagging system. The evaluation on data from three collaborative tagging systems confirmed the importance of both learning methods. Finally, an architecture based on text indexing makes the system efficient enough to serve in real time collaborative tagging systems with number of posts counted in millions, given limited computing resources.

The importance of event logs, as a source of information in systems and network management cannot be overemphasized. With the ever increasing size and complexity of today's event logs, the task of analyzing event logs has become cumbersome to carry out manually. For this reason recent research has focused on the automatic analysis of these log files. In this paper we present IPLoM (Iterative Partitioning Log Mining), a novel algorithm for the mining of clusters from event logs. Through a 3-Step hierarchical partitioning process IPLoM partitions log data into its respective clusters. In its 4th and final stage IPLoM produces cluster descriptions or line formats for each of the clusters produced. Unlike other similar algorithms IPLoM is not based on the Apriori algorithm and it is able to find clusters in data whether or not its instances appear frequently. Evaluations show that IPLoM outperforms the other algorithms statistically significantly, and it is also able to achieve an average F-Measure performance 78% when the closest other algorithm achieves an F-Measure performance of 10%.

Image-based abstraction (or summarization) of a Web site is the process of extracting the most characteristic (or important) images from it. The criteria for measuring the importance of images in Web sites are based on their frequency of occurrence, characteristics of their content and Web link information. As a case study, this work focuses on logo and trademark images. These are important characteristic signs of corporate Web sites or of products presented there. The proposed method incorporates machine learning for distinguishing logo and trademarks from images of other categories (e.g., landscapes, faces). Because the same logo or trademark may appear many times in various forms within the same Web site, duplicates are detected and only unique logo and trademark images are extracted. These images are then ranked by importance taking frequency of occurrence, image content and Web link information into account. The most important logos and trademarks are finally selected to form the image-based summary of a Web site. Evaluation results of the method on real Web sites are also presented. The method has been implemented and integrated into a fully automated image-based summarization system which is accessible on the Web (www.intelligence.tuc.gr/websummarization)

We present a generative model for simultaneously clustering documents and terms. Our model is a four-level hierarchical Bayesian model, in which each document is modeled as a random mixture of document topics , where each topic is a distribution over some segments of the text. Each of these segments in the document can be modeled as a mixture of word topics where each topic is a distribution over words. We present efficient approximate inference techniques based on Markov Chain Monte Carlo method and a moment-matching algorithm for empirical Bayes parameter estimation. We report results in document modeling, document and term clustering, comparing to other topic models, Clustering and Co-Clustering algorithms including latent Dirichlet allocation (LDA), model-based overlapping clustering (MOC), model-based overlapping co-clustering (MOCC) and information-theoretic co-clustering (ITCC).

Semantic Similarity relates to computing the similarity between conceptually similar but not necessarily lexically similar terms. Typically, semantic similarity is computed by mapping terms to an ontology and by examining their relationships in that ontology. We investigate approaches to computing the semantic similarity between natural language terms (using WordNet as the underlying reference ontology) and between medical terms (using the MeSH ontology of medical and biomedical terms). The most popular semantic similarity methods are implemented and evaluated using WordNet and MeSH. Building upon semantic similarity, we propose the Semantic Similarity based Retrieval Model (SSRM), a novel information retrieval method capable for discovering similarities between documents containing conceptually similar terms. The most effective semantic similarity method is implemented into SSRM. SSRM has been applied in retrieval on OHSUMED (a standard TREC collection available on the Web). The experimental results demonstrated promising performance improvements over classic information retrieval methods utilizing plain lexical matching (e.g., Vector Space Model) and also over state-of-the-art semantic similarity retrieval methods utilizing ontologies.

We propose a novel method for document clustering using character N-grams. In the traditional vector-space model, the documents are represented as vectors, in which each dimension corresponds to a word. We propose a document representation based on the most frequent character N-grams, with window size of up to 10 characters. We derive a new distance measure, which produces uniformly better results when compared to the word-based and term-based methods. The result becomes more significant in the light of the robustness of the N-gram method with no language-dependent preprocessing. Experiments on the performance of a clustering algorithm on a variety of test document corpora demonstrate that the N-gram representation with n=3 outperforms both word and term representations. The comparison between word and term representations depends on the data set and the selected dimensionality.

Image retrieval on the Web requires that important (authoritative) images satisfying the query selection criteria are assigned higher ranking over other relevant images. PicASHOW (Lempel and Soffer, 2002) achieves this goal using link information alone. This work proposes WPicASHOW (weighted PicASHOW), a weighted scheme for co-citation analysis that incorporates within the link analysis method of PicASHOW the text and image content of the queries and of the Web pages. WPicASHOW is implemented and integrated into a fully automated Web retrieval system for logo and trademark images.

A focused crawler is designed to traverse the Web to gather documents on a specific topic. It is not an easy task to predict which links lead to good pages. In this paper, we present a new approach for prediction of the important links to relevant pages based on a learned user model. In particular, we first collect pages that a user visits during a learning session, where the user browses the Web and specifically marks which pages she is interested in. We then examine the semantic content of these pages to construct a concept graph, which is used to learn the dominant content and link structure leading to target pages using a Hidden Markov Model (HMM). Experiments show that with learned HMM from a user's browsing, the crawling performs better than Best-First strategy.