Document representation

The automated categorization (or classification) of texts into predefined categories has witnessed a booming interest in the last 10 years, due to the increased availability of documents in digital form and the ensuing need to organize them. In the research community the dominant approach to this problem is based on machine learning techniques: a general inductive process automatically builds a classifier by learning, from a set of preclassified documents, the characteristics of the categories. The advantages of this approach over the knowledge engineering approach (consisting in the manual definition of a classifier by domain experts) are a very good effectiveness, considerable savings in terms of expert labor power, and straightforward portability to different domains. This survey discusses the main approaches to text categorization that fall within the machine learning paradigm. We will discuss in detail issues pertaining to three different problems, namely, document representation, classifier construction, and classifier evaluation.

Text categorization -the assignment of natural language texts to one or more predefined categories based on their content -is an important component in many information organization and management tasks. We compare the effectiveness of five different automatic learning algorithms for text categorization in terms of learning speed, realtime classification speed, and classification accuracy. We also examine training set size, and alternative document representations. Very accurate text classifiers can be learned automatically from training examples. Linear Support Vector Machines (SVMs) are particularly promising because they are very accurate, quick to train, and quick to evaluate.

In this paper, an introduction and survey over probabilistic information retrieval (IR) is given. First, the basic concepts of this approach are described: the probability-ranking principle shows that optimum retrieval quality can be achieved under certain assumptions; a conceptual model for IR along with the corresponding event space clarify the interpretation of the probabilistic parameters involved. For the estimation of these parameters, three different learning strategies are distinguished, namely query-related, document-related and description-related learning. As a representative for each of these strategies, a specific model is described. A new approach regards IR as uncertain inference; here, imaging is used as a new technique for estimating the probabilistic parameters, and probabilistic inference networks support more complex forms of inference. Finally, the more general problems of parameter estimations, query expansion and the development of models for advanced document representations are discussed.

This document representation is successivelytransformed into a sequence of sentence plans (to-gether with formatting instructions in a selectabletarget format; SGML, IgTEX , Zmacs and - for screenoutput - formatted ASCII are currently supported),which are handed over to sentence generators. ForEnglish, we use Penman (Penman 1991) and it ssentence planning language (SPL) as input terms.To produce German text, we have implementeda (partial) German version of Penman's grammar(NIGEL),...

Previous researches on advanced representations for document retrieval have shown that statistical state-of-the-art models are not improved by a variety of different linguistic representations. Phrases, word senses and syntactic relations derived by Natural Language Processing (NLP) techniques were observed ineffective to increase retrieval accuracy. For Text Categorization (TC) are available fewer and less definitive studies on the use of advanced document representations as it is a relatively new research area (compared to document retrieval). In this paper, advanced document representations have been investigated. Extensive experimentation on representative classifiers, Rocchio and SVM, as well as a careful analysis of the literature have been carried out to study how some NLP techniques used for indexing impact TC. Cross validation over 4 different corpora in two languages allowed us to gather an overwhelming evidence that complex nominals, proper nouns and word senses are not adequate to improve TC accuracy.

Although we see the positive results of information retrieval research embodied throughout the Internet, on our computer desktops, and in many other aspects of daily life, at the same time we notice that people still have a wide variety of difficulties in finding information that is useful in resolving their problematic situations. This suggests that there still remain substantial challenges for research in IR. Already in 1988, on the occasion of receiving the ACM SIGIR Gerard Salton Award, Karen Spärck Jones suggested that substantial progress in information retrieval was likely only to come through addressing issues associated with users (actual or potential) of IR systems, rather than continuing IR research's almost exclusive focus on document representation and matching and ranking techniques. In recent years it appears that her message has begun to be heard, yet we still have relatively few substantive results that respond to it. In this paper, I identify and discuss a few ...

Drawing on the theory of documents representation (Perfetti, Rouet, & Britt, 1999), we argue that successfully dealing with multiple documents on the WWW requires readers to form documents models, that is, to deal with contents and sources. We present a study in which we tested the assumption that the use of metacognitive strategies is crucial to the formation of documents models. A total of 100 participants with little medical knowledge were asked to conduct an Internet research on a medical topic. Participants were randomly assigned to four experimental groups that either received evaluation prompts, monitoring prompts, both types of prompts, or no prompts. A control group took paper-and-pencil notes. Results showed that laypersons receiving evaluation prompts outperformed controls in terms of knowledge about sources. In addition, laypersons receiving monitoring prompts acquired significantly more knowledge about facts, and performed slightly better on a comprehension test. It is concluded that the results underline the importance of metacognition in dealing with multiple documents.

This study investigates the use of criteria to assess relevant, partially relevant, and not-relevant documents. Study participants identified passages within 20 document representations that they used to make relevance judgments; judged each document representation as a whole to be relevant, partially relevant, or not relevant to their information need; and explained their decisions in an interview. Analysis revealed 29 criteria, discussed positively and negatively, that were used by the participants when selecting passages that contributed or detracted from a document's relevance. These criteria can be grouped into six categories: abstract (e.g.). Results further indicate that multiple criteria are used when making relevant, partially relevant, and not-relevant judgments, and that most criteria can have either a positive or negative contribution to the relevance of a document. The criteria most frequently mentioned by study participants were content, followed by criteria characterizing the full text document. These findings may have implications for relevance feedback in information retrieval systems, suggesting that systems accept and utilize multiple positive and negative relevance criteria from users. Systems designers may want to focus on supporting content criteria followed by full text criteria as these may provide the greatest cost benefit.

In this paper we present a new document representation model based on implicit user feedback obtained from search engine queries. The main objective of this model is to achieve better results in non-supervised tasks, such as clustering and labeling, through the incorporation of usage data obtained from search engine queries. This type of model allows us to discover the motivations of users when visiting a certain document. The terms used in queries can provide a better choice of features, from the user's point of view, for summarizing the Web pages that were clicked from these queries. In this work we extend and formalize as query model an existing but not very well known idea of query view for document representation. Furthermore, we create a novel model based on frequent query patterns called the query-set model. Our evaluation shows that both query-based models outperform the vector-space model when used for clustering and labeling documents in a website. In our experiments, the query-set model reduces by more than 90% the number of features needed to represent a set of documents and improves by over 90% the quality of the results. We believe that this can be explained because our model chooses better features and provides more accurate labels according to the user's expectations.

Wikipedia has been applied as a background knowledge base to various text mining problems, but very few attempts have been made to utilize it for document clustering. In this paper we propose to exploit the semantic knowledge in Wikipedia for clustering, enabling the automatic grouping of documents with similar themes. Although clustering is intrinsically unsupervised, recent research has shown that incorporating supervision improves clustering performance, even when limited supervision is provided. The approach presented in this paper applies supervision using active learning. We first utilize Wikipedia to create a concept-based representation of a text document, with each concept associated to a Wikipedia article. We then exploit the semantic relatedness between Wikipedia concepts to find pair-wise instance-level constraints for supervised clustering, guiding clustering towards the direction indicated by the constraints. We test our approach on three standard text document datasets. Empirical results show that our basic document representation strategy yields comparable performance to previous attempts; and adding constraints improves clustering performance further by up to 20%.

In this paper we describe work relating to classification of web documents using a graph-based model instead of the traditional vector-based model for document representation. We compare the classification accuracy of the vector model approach using the k-Nearest Neighbor (k-NN) algorithm to a novel approach which allows the use of graphs for document representation in the k-NN algorithm. The proposed method is evaluated on three different web document collections using the leave-one-out approach for measuring classification accuracy. The results show that the graph-based k-NN approach can outperform traditional vector-based k-NN methods in terms of both accuracy and execution time.

Scholarly digital libraries increasingly provide analytics to information within documents themselves. This includes information about the logical document structure of use to downstream components, such as search, navigation, and summarization. In this paper, the authors describe SectLabel, a module that further develops existing software to detect the logical structure of a document from existing PDF files, using the formalism of conditional random fields. While previous work has assumed access only to the raw text ...

is becoming the most relevant standardization e ort in the area of document representation through markup languages. Through XML, it is possible to de ne complex documents, containing information at di erent degrees of sensitivity. Moreover, the processes of document exchange and acquisition, which can be very frequent in Web-based informationsystems, are simpli ed and standardized. In this scenario, there is a strong need for policies to control and regulate the access and dissemination of XML documents. In the paper, we discuss main protection requirements posed by XML documents and we present a set of authorization and dissemination policies that enable both a controlled access to XML documents in a given source and the exchange of XML documents across di erent sources.

Document representation using the bag-of-words approach may require bringing the dimensionality of the representation down in order to be able to make effective use of various statistical classification methods. Latent Semantic Indexing (LSI) is one such method that is based on eigendecomposition of the covariance of the document-term matrix. Another often used approach is to select a small number of most important features out of the whole set according to some relevant criterion. This paper points out that LSI ignores discrimination while concentrating on representation. Furthermore, selection methods fail to produce a feature set that jointly optimizes class discrimination. As a remedy, we suggest supervised linear discriminative transforms, and report good classification results applying these to the Reuters-21578 database.

Most existing methods for text categorization employ induction algorithms that use the words appearing in the training documents as features. While they perform well in many categorization tasks, these methods are inherently limited when faced with more complicated tasks where external knowledge is essential. Recently, there have been efforts to augment these basic features with external knowledge, including semi-supervised learning and transfer learning. In this work, we present a new framework for automatic acquisition of world knowledge and methods for incorporating it into the text categorization process. Our approach enhances machine learning algorithms with features generated from domain-specific and common-sense knowledge. This knowledge is represented by ontologies that contain hundreds of thousands of concepts, further enriched through controlled Web crawling. Prior to text categorization, a feature generator analyzes the documents and maps them onto appropriate ontology concepts that augment the bag of words used in simple supervised learning. Feature generation is accomplished through contextual analysis of document text, thus implicitly performing word sense disambiguation. Coupled with the ability to generalize concepts using the ontology, this approach addresses two significant problems in natural language processing-synonymy and polysemy. Categorizing documents with the aid of knowledge-based features leverages information that cannot be deduced from the training documents alone. We applied our methodology using the Open Directory Project, the largest existing Web directory built by over 70,000 human editors. Experimental results over a range of data sets confirm improved performance compared to the bag of words document representation.

The automatic categorisation of web documents is becoming crucial for organising the huge amount of information available in the Internet. We are facing a new challenge due to the fact that web documents have a rich structure and are highly heterogeneous. Two ways to respond to this challenge are (1) using a representation of the content of web documents that captures these two characteristics and (2) using more effective classifiers.

In this paper we describe a type of data fusion involving the combination of evidence derived from multiple document representations. Our aim is to investigate if a composite representation can improve the online detection of novel events in a stream of broadcast news stories. This classification process otherwise known as first story detection FSD (or in the Topic Detection and Tracking pilot study as online new event detection [1]), is one of three main classification tasks defined by the TDT initiative. Our composite document representation consists of a semantic representation (based on the lexical chains derived from a text) and a syntactic representation (using proper nouns). Using the TDT1 evaluation methodology, we evaluate a number of document representation combinations using these document classifiers.

The publication of material in 'electronic form' should ideally preserve, in a unified document representation, all of the richness of the printed document while maintaining enough of its underlying structure to enable searching and other forms of semantic processing. Until recently it has been hard to find a document representation which combined these attributes and which also stood some chance of becoming a de facto multi-platform standard.

The use of the computing with words paradigm for the automatic text documents categorization problem is discussed. This specific problem of information retrieval (IR) becomes more and more important, notably in view of a fast proliferation of textual information available on the Internet. The main issues that have to be addressed here are: document representation and classification. The use of fuzzy logic for both problems has already been quite deeply studied though for the latter, i.e., classification, generally not in an IR context. Our approach is based mainly on the classical calculus of linguistically quantified propositions proposed by Zadeh. Moreover, we employ results related to fuzzy (linguistic) queries in IR, notably various interpretations of the weights of query terms. Some preliminary results on widely adopted text corpora are presented.

Active reading, involving acts such as highlighting, writing notes, etc., is an important part of knowledge workers' activities. Most computer-based active reading support seeks to replicate the affordances of paper, but paper has limitations, being in many ways inflexible. In this paper we introduce LiquidText, a computer-based active reading system that takes a fundamentally different approach, offering a flexible, fluid document representation built on multitouch input, with a range of interaction techniques designed to facilitate the activities of active reading. We report here on our design for LiquidText, its interactions and gesture vocabulary, and our design process, including formative user evaluations which helped shape the final system.

Multiclass support vector machine (SVM) methods are well studied in recent literature. Comparison studies on UCI/statlog multiclass datasets suggest using one-against-one method for multiclass SVM classification. However, in unilabel (multiclass) text categorization with SVMs, no comparison studies exist with one-against-one and other methods, e.g. one-against-all and several well-known improvements to these approaches. In this paper, we bridge this gap by performing empirical comparison of standard one-against-all and one-against-one, together with three improvements to these standard approaches for unilabel text categorization with SVM as base binary learner. We performed all our experiments on three standard text corpuses using two types of document representation. Outcome of our experiments partly support Rifkin and statement that, for small scale unilabel text categorization tasks, if parameters of the classifiers are well tuned, one-against-all will have better performance than one-against-one and other methods.

We present a text document search engine with several new visualization front-ends that aid navigation through the set of documents returned by a query (short "returned documents"). Our methods are based on identifying and selecting keywords on the fly. The choice of keywords depends not only on the frequency of their occurrence within returned documents but also on the specificity of their occurrence just within these returned documents. Keywords are subsequently used to obtain a sparse document representation and to compute document clusters using a variant of the buckshot algorithm. One of the visualization front-ends uses the sparse document representation to obtain keyword clusters. We make use of the clustering to group the documents returned from the search visually, and to label the groups with their most salient keywords. The different front-ends cater for different user needs. Two of them can be employed to browse cluster information as well as to drill down or up in clusters and refine the search using one or more of the suggested keywords.

Modern document collections often contain groups of documents with overlapping or shared content. However, most information retrieval systems process each document separately, causing shared content to be indexed multiple times. In this paper, we describe a new document representation model where related documents are organized as a tree, allowing shared content to be indexed just once. We show how this representation model can be encoded in an inverted index and we describe algorithms for evaluating free-text queries based on this encoding. We also show how our representation model applies to web, email, and newsgroup search. Finally, we present experimental results showing that our methods can provide a significant reduction in the size of an inverted index as well as in the time to build and query it.

Increasingly large text datasets and the high dimensionality associated with natural language create a great challenge in text mining. In this research, a systematic study is conducted, in which three different document representation methods for text are used, together with three Dimension Reduction Techniques (DRT), in the context of the text clustering problem. Several standard benchmark datasets are used. The three Document representation methods considered are based on the vector space model, and they include word, multi-word term, and character N-gram representations. The dimension reduction methods are independent component analysis (ICA), latent semantic indexing (LSI), and a feature selection technique based on Document Frequency (DF). Results are compared in terms of clustering performance, using the k-means clustering algorithm. Experiments show that ICA and LSI are clearly better than DF on all datasets. For word and N-gram representation, ICA generally gives better results compared with LSI. Experiments also show that the word representation gives better clustering results compared to term and N-gram representation. Finally, for the N-gram representation, it is demonstrated that a profile length (before dimensionality reduction) of 2000 is sufficient to capture the information and, in most cases, a 4-gram representation gives better performance than 3-gram representation.

Artificial intelligence models may be used to improve performance of information retrieval (IR) systems and the genetic algorithms (GAs) are an example of such a model. This paper presents an application of GAs as a relevance feedback method aiming to improve the document representation and indexing. In this particular form of GAs, various document descriptions compete with each other and a better collection indexing is sought through reproduction, crossover and mutation operations. In this paradigm, we are searching for the optimal balance between two genetic parameters: the population size and the number of generations. We try to discover the optimal parameter choice both by experiments using the CACM and CISI collections, and by a theoretical analysis providing explanation of the experimental results. The general conclusion tends to be that larger populations have better chance of significantly improving the effectiveness of retrieval.

In the Web context, link-based evidence is most commonly used in conjunction with contentbased evidential information in order to improve retrieval effectiveness. This paper examines the impact the various types of link-based evidence and their combination with content-based evidence have on the retrieval effectiveness for the topic relevance Web task. The inference network model is used in our study, as it supports the combination of multiple document representations and the combination of multiple results produced by different retrieval strategies. Our experiments indicate hardly any improvements in the effectiveness, similarly to previous TREC results for the topic relevance task. However, they allow us to gain an insight into the behaviour of different types of link-based evidence that could be exploited in the context of other retrieval tasks.

Traditional information retrieval systems aim at satisfying most users for most of their searches, leaving aside the context in which the search takes place. We propose to model two main aspects of context: The themes of the user's information need and the specific data the user is looking for to achieve the task that has motivated his search. Both aspects are modeled by means of ontologies. Documents are semantically indexed according to the context representation and the user accesses information by browsing the ontologies. The model has been applied to a case study that has shown the added value of such a semantic representation of context.

This paper presents a new document representation with vectorized multiple features including term frequency and term-connection-frequency. A document is represented by undirected and directed graph, respectively. Then terms and vectorized graph connectionists are extracted from the graphs by employing several feature extraction methods. This hybrid document feature representation more accurately reflects the underlying semantics that are difficult to achieve from the currently used term histograms, and it facilitates the matching of complex graph. In application level, we develop a document retrieval system based on self-organizing map (SOM) to speed up the retrieval process. We perform extensive experimental verification, and the results suggest that the proposed method is computationally efficient and accurate for document retrieval.

This paper provides a thematic frame analysis of Australian newspaper reporting of the outcome and implications of the trial of Rolah McCabe versus British American Tobacco Australasia (BATA). In this trial, a Melbourne woman was awarded A$700 000 damages for smoking-attributable lung cancer when the defendant, BATA, had its case dismissed due to document destruction. In 60 commentaries from Australian national or capital city newspapers between 12 April and 9 May 2002, a total of 79 instances of eight tobacco-related frames were identified. Overall, 43% of the 79 instances were positive for tobacco control, 46% were negative for tobacco control and 11% were neutral. The most common frame that was negative for tobacco control (in 35% of articles) was the conception that smokers exert ‘free will’ in deciding to smoke and should therefore be personally responsible for their smoking and any disease that arises as a result of it.
A related, but less commonly employed frame (in 18% of articles) was the expressed fear of a ‘slippery slope’ of litigation, which portrayed smoking as similar to eating fast food or other ‘vices’. The most common frame that was positive for tobacco control (in 35% of articles) was the notion that the tobacco industry was ‘evil’ and, to a lesser extent, that the government should ‘do more’ to control smoking (15% of articles). These findings provide a sobering public health challenge to improve public communication efforts about the powerful forces that conspire to induce people to start smoking and keep them smoking for decades, despite a strong desire to quit. There is a need to fund public education programs and quit smoking services more adequately to address the complex education task of understanding the nature of addiction to tobacco and the enormity of the health risk.

Effective retrieval of structured documents should exploit the content and structural knowledge associated with the documents. This knowledge can be used to focus retrieval to the best entry points: document components that contain relevant information, and from which users can browse to retrieve further relevant components. To enable this, suitable representation methods must be developed.

In this paper, we explore the effects of data fusion on First Story Detection [1] in a broadcast news domain. The data fusion element of this experiment involves the combination of evidence derived from two distinct representations of document content in a single cluster run. Our composite document representation consists of a concept representation (based on the lexical chains derived

Most web content categorization methods are based on the vector-space model of information retrieval. One of the most important advantages of this representation model is that it can be used by both instance-based and model-based classifiers. However, this popular method of document representation does not capture important structural information, such as the order and proximity of word occurrence or the location of a word within the document. It also makes no use of the mark-up information that can easily be extracted from the web document HTML tags.

Text mining, intelligent text analysis, text data mining and knowledge-discovery in text are generally used aliases to the process of extracting relevant and non-trivial information from text. Some crucial issues arise when trying to solve this problem, such as document representation and deficit of labeled data. This paper addresses these problems by introducing information from unlabeled documents in the training set, using the support vector machine (SVM) separating margin as the differentiating factor. Besides studying the influence of several pre-processing methods and concluding on their relative significance, we also evaluate the benefits of introducing background knowledge in a SVM text classifier. We further evaluate the possibility of actively learning and propose a method for successfully combining background knowledge and active learning. Experimental results show that the proposed techniques, when used alone or combined, present a considerable improvement in classification performance, even when small labeled training sets are available.

In this paper we proposed to automatically classify documents based on the meanings of words and the relationships between groups of meanings or concepts. Our proposed classification algorithm builds on the word structures provided by WordNet, which not only arranges words into groups of synonyms, called Synsets, but also arranges the Synsets into hierarchies representing the relationships between concepts. Most existing methods classify text documents based on the number of occurrences of words and some based on Synsets. Our approach goes one step further by using not only word occurrences and Synsets but also the relationships between Synsets. We also proposed a sense-based document representation based on the semantic hierarchies provided by WordNet. To classify a document, our approach extracts words occurred in the document and uses them to increase the weight of the Synsets corresponding to the words. Words with same meanings will increase the weight of their corresponding Synsets. As a result, we count the occurrences of senses. We also propagate the weight of a Synset upward to its related Synsets in the hierarchies and thus capture the relationships between concepts. In comparing to previous research, our approach increases the classification accuracy by 14%.

Document clustering has the goal of discovering groups with similar documents. The success of the document clustering algorithms depends on the model used for representing these documents. Documents are commonly represented with the vector space model based on words or n-grams. However, these representations have some disadvantages such as high dimensionality and loss of the word sequential order. In this work, we propose a new document representation in which the maximal frequent sequences of words are used as features of the vector space model. The proposed model efficiency is evaluated by clustering different document collections and compared against the vector space model based on words and n-grams, through internal and external measures.

In this paper the problem of indexing heterogeneous structured documents and of retrieving semistructured documents is considered. We propose a flexible paradigm for both indexing such documents and formulating user queries specifying soft constraints on both documents' structure and content. At the indexing level we propose a model that achieves flexibility by constructing personalised document representations based on users' views of the documents. This is obtained by allowing users to specify their preferences on the documents' sections that they estimate to bear the most interesting information, as well as to linguistically quantify the number of sections which determine the global potential interest of the documents. At the query language level, a flexible query language for expressing soft selection conditions on both the documents' structure and content is proposed.

In this paper we describe a solution for incorporating background knowledge into the OntoGen system for semi-automatic ontology construction. This makes it easier for different users to construct different and more personalized ontologies for the same domain. To achieve this we introduce a word weighting schema to be used in the document representation. The weighting schema is learned based on the background knowledge provided by user. It is than used by OntoGen's machine learning and text mining algorithms.

Being able to author a hypermedia document once for presentation under a wide variety of potential circumstances requires that it be stored in a manner that is adaptable to these circumstances. Since the nature of these circumstances is not always known at authoring time, specifying how a document adapts to them must be a process that can be performed separately from its original authoring. These distinctions include the porting of the document to different platforms and formats and the adapting of the document's presentation to suit the needs of the user and of the current state of the presentation environment. In this paper we discuss extensions to our CMIF hypermedia authoring and presentation environment that provide adaptability through this distinction between authoring and presentation specification. This extension includes the use of HyTime for document representation and of DSSSL for presentation specification. We also discuss the Berlage architecture, our extension to HyTime that specifies the encoding of certain hypermedia concepts useful for presentation specification.

Most recent document standards like XML rely on structured representations. On the other hand, current information retrieval systems have been developed for flat document representations and cannot be easily extended to cope with more complex document types. The design of such systems is still an open problem. We present a new model for structured document retrieval which allows computing scores of document parts. This model is based on Bayesian networks whose conditional probabilities are learnt from a labelled collection of structured documents-which is composed of documents, queries and their associated assessments. Training these models is a complex machine learning task and is not standard. This is the focus of the paper: we propose here to train the structured Bayesian Network model using a cross-entropy training criterion. Results are presented on the INEX corpus of XML documents.

The third Cross-Language Evaluation Forum workshop (CLEF-2002) provides the unprecedented opportunity to evaluate retrieval in eight different languages using a uniform set of topics and assessment methodology. This year the Johns Hopkins University Applied Physics Laboratory participated in the monolingual, bilingual, and multilingual retrieval tasks. We contend that information access in a plethora of languages requires approaches that are inexpensive in developer and run-time costs. In this paper we describe a simplified approach that seems suitable for retrieval in many languages; we also show how good retrieval is possible over many languages, even when translation resources are scarce, or when query-time translation is infeasible. In particular, we investigate the use of character n-grams for monolingual retrieval, pre-translation expansion as a technique to mitigate errors due to limited translation resources, and translation of document representations to an interlingua for computationally efficient retrieval against multiple languages.

Most recent document standards rely on structured representations. On the other hand, current information retrieval systems have been developed for flat document representations and cannot be easily extended to cope with more complex document types. Only a few models have been proposed for handling structured documents, and the design of such systems is still an open problem. We present here a new model for structured document retrieval which allows to compute and to combine the scores of document parts. It is based on bayesian networks and allows for learning the model parameters in the presence of incomplete data. We present an application of this model for ad-hoc retrieval and evaluate its performances on a small structured collection. The model can also be extended to cope with other tasks such as interactive navigation in structured documents or corpus.

In Automatic Text Processing tasks, documents are usually represented in the bag-ofwords space. However, this representation does not take into account the possible relations between words. We propose here a review of a family of document density estimation models for representing documents. Inside this family we derive another possible model: the Theme Topic Mixture Model (TTMM). This model assumes two types of relations among textual data. Topics link words to each other and Themes gather documents with particular distribution over the topics. An experiment reports the performance of the different models in this family over a common task.

Most web content classification methods are based on the vectorspace model of information retrieval. One of the important advantages of this representation model is that it can be used by both instance-based and model-based classifiers for categorization. However, this popular method of document representation does not capture important structural information, such as the order and proximity of word occurrence or the location of a word within the document. It also makes no use of the mark-up information that can be easily extracted from the web document HTML tags.