Search

In this paper, we describe the process that we used to create a new corpus of children’s emotional speech. We used a Wizard of Oz (WoZ) setting to induce different emotional reactions in children during speech-based interactions with two robots. We recorded the speech spoken in Mexican Spanish by 174 children (both sexes) between 6 and 11 years of age. The recordings were manually segmented and transcribed. The segments were then labeled with two types of emotional-related paralinguistic information: emotion and attitude. The corpus contained 2093 min of audio recordings (34.88 h) divided into 19,793 speech segments. The Interactive Emotional Children’s Speech Corpus (IESC-Child) can be a valuable resource for researchers studying affective reactions in speech communication during child-computer interactions in Spanish and for creating models to recognize acoustic paralinguistic information. IESC-Child is available to the research community upon request.

Due to the problems presented in current traditional/biometric security systems, the interest to use new security systems, have been increasing. This paper explores the use of brain signals EEG-based during imagined speech in order to use it as a new biometric measure for Subjects identification and thus create a new biometric security system. The main contribution of this paper are two methods for feature extraction, first to improve the signal-to-noise ratio the Common Average Reference was applied. The first method was based on Discrete Wavelet Transform, and the second method was based on statistical features directly from the raw signal. The proposed methods were tested in a dataset of 27 Subjects who performed 33 repetitions of 5 imagined words in Spanish. The results show the feasibility of the task with accurate identification of the Subject, regardless of the imagined word used and using a commercial EEG system (EMOTIV EPOC). In addition, the scope of the method is displayed by decreasing the training data, as well as the number of active sensors for the identification task. Using the proposed method with future improvements and implementing it in a low-cost device can be a new and valuable biometric security system.

An asynchronous Brain--Computer Interface (BCI) based on imagined speech is a tool that allows to control an external device or to emit a message at the moment the user desires to by decoding EEG signals of imagined speech. In order to correctly implement these types of BCI, we must be able to detect from a continuous signal, when the subject starts to imagine words. In this work, five methods of feature extraction based on wavelet decomposition, empirical mode decomposition, frequency energies, fractal dimension and chaos theory features are presented to solve the task of detecting imagined words segments from continuous EEG signals as a preliminary study for a latter implementation of an asynchronous BCI based on imagined speech. These methods are tested in three datasets using four different classifiers and the higher F1 scores obtained are 0.73, 0.79, and 0.68 for each dataset, respectively. This results are promising to build a system that automatizes the segmentation of imagined words segments for latter classification., Comment: 10 pages, 14 figures

This paper summarizes the thesis: ”Author Profiling in Social Media with Multimodal Information.” Our solution uses a multimodal approach to extracting information from written messages and images shared by users. Previous work has shown the existence of useful information for this task in these modalities; however, our proposal goes further, demonstrating the complementarity of the modalities when merging these two sources of information. To do this, we propose to transform images to texts, and with them, to have the same framework of representation for both kinds of information, which allow to achieve their fusion. Our work explores different methods for extracting information either from the text and the images. To represent the extracted information, different distributional term representations approaches were explored in order to identify the topics addressed by the user. For this purpose, an evaluation framework was proposed in order to identify the most appropriate method for this task. The results show that the textual descriptions of the images contain useful information for the author profiling task, and that the fusion of textual information with information extracted from the images increases the accuracy of this task.

The increasing propagation of abusive language in social media is a major concern for supplier companies and governments because of its negative social impact. A large number of methods have been developed for its automatic identification, ranging from dictionary-based methods to sophisticated deep learning approaches. A common problem in all these methods is to distinguish the offensive use of swear words from their everyday and humorous usage. To tackle this particular issue we propose an attention-based neural network architecture that captures the word n-grams importance according to their context. The obtained results in four standard collections from Twitter and Facebook are encouraging, they outperform the \(F_1\) scores from state-of-the-art methods and allow identifying a set of inherently offensive swear words, and others in which its interpretation depends on its context.

E-communication represents a major threat to users who are exposed to a number of risks and potential attacks. Detecting these risks with as much anticipation as possible is crucial for prevention. However, much research so far has focused on forensic tools that can be applied only when an attack has been performed. This paper proposes a novel and effective methodology for the early detection of threats in written social media. The goal is to recognize a potential attack before it is consummated, and using a minimum amount of information. The proposed approach considers the use of profile-based representations (PBRs) for this goal. PBRs have multiple benefits, including non-sparsity, low dimensionality, and a proved discriminative power. Moreover, representations for partial documents can be derived naturally with PBRs, which makes them suitable for the addressed problem. Results include empirical evidence on the usefulness of PBRs in the early recognition setting for two tasks in which anticipation is critical: sexual predator detection and aggressive text identification. These results reveal, on the one hand, that PBRs achieve state of the art performance when using full-length documents (i.e., the classical task), and, on the other hand, that the proposed methodology outperforms previous work on early recognition of sexual predators by a considerable margin, while obtaining state of the art performance in aggressive text identification. To the best of our knowledge, these are the best results reported on early recognition for the approached problems. We foresee this work will pave the way for the development of novel methodologies for the problem and will motivate further research from the intelligent systems and text mining communities.

The concept of polarity is pervasive in natural language. It relates syntax, semantics and pragmatics narrowly (Giannakidou, in: Maienborn, von Heusinger, Portner (eds.), Semantics: an international handbook of natural language meaning, De Gruyter Mouton, Berlin, 2011; Israel in The grammar of polarity: pragmatics, sensitivity, and the logic of scales, Cambridge studies in linguistics, Cambridge University Press, Cambridge, 2014), it refers to items of many syntactic categories such as nouns, verbs and adverbs. Neutral polarity items appear in affirmative and negative sentences, negative polarity items cannot appear in affirmative sentences, and positive polarity items cannot appear in negative sentences. A way of reasoning in Natural Language is through Natural Logic (van Benthem in Essays in logical semantics, vol. 29 of Studies in linguistics and philosophy, Reidel, Dordrecht, 1986; Language in action: categories, lambdas, and dynamic logic, vol. 130 of Studies in logic, Elsevier, Amsterdam, 1991). This logic is based on the concept of polarity in order to make the meaning of a sentence weaker o stronger without changing its truth value. There exist many proposals to compute polarity in the Natural Logic context, the most widely known are the ones by: van Benthem (1986, 1991), Sanchez-Valencia (Studies on natural logic and categorial grammar, Ph.D. thesis, Universiteit van Amsterdam, 1991), Dowty (Proceedings of the 4th conference on semantics and theoretical linguistics, Cornel University, CLC Publications, Rochester, 1994), and van Eijck (in: ten Cate, Zeevat (eds.), 6th international Tbilisi symposium on logic, language, and computation, Batumi, Georgia, Springer, 2007). If Natural Logic is going to be used, as an inferential mechanism between text fragments, in Natural Language Processing applications such as text summarization, question answering, and information extraction, it is a priority to know what the existing relationship among the aforementioned algorithms is; for example, to implement the most general. We show in this paper the equivalence among the analyzed algorithms, filling a gap in Natural Logic research, particularly in computing polarity, and the soundness of their algorithms.

EEG-based Brain–Computer Interfaces (BCI) are an alternative technique that aims to integrate people with severe motor disabilities to their environment. However, they still are not used in everyday life because BCIs are controlled by means of few intuitive electrophysiological sources. To address this problem, work has been carried out with the objective of classifying EEG signals recorded while imagining speech. In this paper, the emerging techniques of sonification and textification are applied to EEG signals, which allows us to characterize EEG signals as either an audio signal or a text document. The aim is to analyze whether these techniques can help to do a better discrimination or to highlight patterns, and assess which of the two methods performs better in order to improve classification results of unspoken words. For proving this we processed signals, from 27 subjects, with and without the application of the proposed methods of sonification and textification. The average accuracy rate using the original EEG signals, the EEG sonified signals and the EEG textified signals are 58.41 ± 12.41%, 63.82 ± 13.44% and 83.34 ± 6.12%, respectively. It can be clearly noticed that the methods of sonification and textification of EEG improve the classification rates obtained using the EEG signals in their original form.

Author Profiling (AP) aims at predicting specific characteristics from a group of authors by analyzing their written documents. Many research has been focused on determining suitable features for modeling writing patterns from authors. Reported results indicate that content-based features continue to be the most relevant and discriminant features for solving this task. Thus, in this paper, we present a thorough analysis regarding the appropriateness of different distributional term representations (DTR) for the AP task. In this regard, we introduce a novel framework for supervised AP using these representations and, supported on it. We approach a comparative analysis of representations such as DOR, TCOR, SSR, and word2vec in the AP problem. We also compare the performance of the DTRs against classic approaches including popular topic-based methods. The obtained results indicate that DTRs are suitable for solving the AP task in social media domains as they achieve competitive results while providing meaningful interpretability.

[EN] Several methods have been proposed for determining plagiarism between pairs of sentences, passages or even full documents. However, the majority of these methods fail to reliably detect paraphrase plagiarism due to the high complexity of the task, even for human beings. Paraphrase plagiarism identi cation consists in automatically recognizing document fragments that contain re-used text, which is intentionally hidden by means of some rewording practices such as semantic equivalences, discursive changes, and morphological or lexical substitutions. Our main hypothesis establishes that the original author's writing style ngerprint prevails in the plagiarized text even when paraphrases occur. Thus, in this paper we propose a novel text representation scheme that gathers both content and style characteristics of texts, represented by means of character-level features. As an additional contribution, we describe the methodology followed for the construction of an appropriate corpus for the task of paraphrase plagiarism identi cation, which represents a new valuable resource to the NLP community for future research work in this field., This work is the result of the collaboration in the framework of the CONACYT Thematic Networks program (RedTTL Language Technologies Network) and the WIQ-EI IRSES project (Grant No. 269180) within the FP7 Marie Curie action. The first author was supported by CONACYT (Scholarship 258345/224483). The second, third, and sixth authors were partially supported by CONACyT (Project Grants 258588 and 2410). The work of the fourth author was partially supported by the SomEMBED TIN2015-71147-C2-1-P MINECO research project and by the Generalitat Valenciana under the Grant ALMAMATER (PrometeoII/2014/030).

Recent works have shown that it is possible to use information extracted from images to address the task of automatic gender identification. These proposals have validated their solutions using monolingual datasets, i.e., collections where images are shared by users having the same mother tongue. This paper aims to test the usefulness of images collected from users who do not share the same language. In principle, these users present cultural differences, which may be reflected in the images they share. However, a cross-cultural image-based approach would be very useful for languages where data is not available or scarce. The experiments presented demonstrate that characteristics obtained from the images, regardless of the users’ mother tongue, can be used for gender prediction. They mainly confirm the usefulness of a cross-cultural image-based approach, showing that culturally different individuals with equivalent profiles traits tend to share similar images.

Deception (the action of deliberately cause someone to believe something that is not true) can have many different repercussions in daily life. However, deception detection is an inherently complex task for humans. Due to this, not only there is uncertainty on which features should be used as cues for automatic deception detection, but labeled data is scarce. In this paper, we explore typical features that can be extracted from videos for affective computing and study their performance for deception detection in videos. Additionally, we perform a study of different multimodal fusion methods meant to improve the results obtained by using the different sets of extracted features separately, including a novel set of methods based on boosting. For this study, high level features are extracted with open automatic tools for the visual, acoustical and textual modalities, respectively. Experiments are conducted using a real-life court trial dataset for deception detection, as well as a novel Mexican deception detection dataset using Spanish as the spoken language.

Depression is a mental disorder with strong social and economic implications. Due to its relevance, recently several researches have explored the analysis of social media content to identify and track depressed users. Most approaches follow a supervised learning strategy supported on the availability of labeled training data. Unfortunately, acquiring such data is very complex and costly. To handle this problem, in this paper we propose a crosslingual approach based on the idea that data already labeled in a specific language can be leveraged to classify depression in other languages. The proposed method is based on a word-level alignment process. Particularly, we propose two representations for the alignment; one of them takes advantage of the psycholinguistic resource LIWC and the other uses bilingual word embeddings. For evaluating the proposed approach, we faced the detection of depression by employing English and Spanish tweets as the source and target data respectively. The results outperformed solutions based on automatic translation of texts, confirming the usefulness of the proposed approach.

It was searched the minimal subset of channels for imagined speech.Channel selection was approached as multi-objective to obtain a Pareto front.A fuzzy system inference was applied to find a promising solution from Pareto front.Channel selection had a statistically similar performance to the use of all channels.It was observed a dependence between features and classes of imagined speech. One of the main purposes of brain-computer interfaces (BCI) is to provide persons of an alternative communication channel. This objective was firstly focused on handicapped subjects but nowadays its scope has increased to healthy persons. Usually, BCIs record brain activity using electroencephalograms (EEG), according to four main neuro-paradigms (slow cortical potentials, motor imagery, P300 component and visual evoked potentials). These analytical paradigms are not intuitive and are difficult to implement. Accordingly, this work researches an alternative neuro-paradigm called imagined speech, which refers to the internal pronunciation of words without emitting sounds or doing facial movements. Specifically, the present research is focused on the recognition of five Spanish words corresponding to the English words "up," "down," "left," "right" and "select", with which a computer cursor could be controlled. We perform an offline computer automatic classification procedure of a dataset of EEG signals from 27 subjects. The method implements a channel selection composed of two stages; the first one obtains a Pareto front and is approached as a multi-objective optimization problem dealing with the error rate and the number of channels; the second stage selects a single solution (channel combination) from the front, applying a fuzzy inference system (FIS). We assess the method's performance through a channel combination and a test set not used to generate the front. Several FIS configurations were explored to evaluate if a FIS is able to select channel combinations that improve or, at least, keep the obtained accuracies using all channels for each subject's data. We found that a FIS configuration, FIS3×3 (three membership functions for both input variables: error rate and the number of channels), obtained the best trade-off between the number of fuzzy rules and its accuracy (68.18% using around 7 channels). Also, the FIS3×3 obtained a similar statistically accuracy compared to the use of all channels (70.33%). Results of our method demonstrate the feasibility of using a FIS to automatically select a solution from the Pareto front to select channels applied to a problem of imagined speech classification. The presented method outperforms previous works in accuracy and showed a dependence relationship between EEG data and imagined words.

Deception detection -identifying when someone is trying to cause someone else to believe something that is not true- is a hard task for humans. The task is even harder for automatic approaches, that must deal with additional problems like the lack of enough labeled data. In this context, transfer learning in the form of cross-domain classification is a task that aims to leverage labeled data from certain domains for which labeled data is available to others for which data is scarce. This paper presents a study on the suitability of linguistic features for cross-domain deception detection on multimodal data. Specifically, we aim to learn models for deception detection across different domains of written texts (one modality) and apply the new knowledge to unrelated topics transcribed from spoken statements (another modality). Experimental results reveal that by using LIWC and POS n-grams we reach a in-modality accuracy of 69.42%, as well as an AUC ROC of 0.7153. When doing transfer learning, we achieve an accuracy of 63.64% and get an AUC ROC of 0.6351.

[EN] Controversial topics are present in the everyday life, and opinions about them can be either truthful or deceptive. Deceptive opinions are emitted to mislead other people in order to gain some advantage. In the most of the cases humans cannot detect whether the opinion is deceptive or truthful, however, computational approaches have been used successfully for this purpose. In this work, we evaluate a representation based on character n-grams features for detecting deceptive opinions. We consider opinions on the following: abortion, death penalty and personal feelings about the best friend; three domains studied in the state of the art. We found character n-grams effective for detecting deception in these controversial domains, even more than using psycholinguistic features. Our results indicate that this representation is able to capture relevant information about style and content useful for this task. This fact allows us to conclude that the proposed one is a competitive text representation with a good trade-off between simplicity and performance., We would like to thank CONACyT for partially supporting this work under grants 613411, CB-2015-01-257383, and FC-2016/2410. The work of the last author was partially funded by the Spanish MINECO under the research project SomEMBED (TIN2015-71147-C2-1-P).

Recognizing Textual Entailment (RTE) is a Natural Language Processing task. It is very important in tasks as Semantic Search and Text Summarization. There are many approaches to RTE, for example, methods based on machine learning, linear programming, probabilistic calculus, optimization, and logic. Unfortunately, no one of them can explain why the entailment is carried on. We can make reasonings, with Natural Logic, from the syntactic part of a natural language expression, and very little semantic information. This paper presents an Automatic Theorem Prover for Natural Logic that allows to know precisely the relationships needed in order to reach the entailment in a class of natural language expressions.

[EN] Paraphrase plagiarism identification represents a very complex task given that plagiarized texts are intentionally modified through several rewording techniques. Accordingly, this paper introduces two new measures for evaluating the relatedness of two given texts: a semantically-informed similarity measure and a semantically-informed edit distance. Both measures are able to extract semantic information from either an external resource or a distributed representation of words, resulting in informative features for training a supervised classifier for detecting paraphrase plagiarism. Obtained results indicate that the proposed metrics are consistently good in detecting different types of paraphrase plagiarism. In addition, results are very competitive against state-of-the art methods having the advantage of representing a much more simple but equally effective solution., This work was partially supported by CONACYT under scholarship 401887, project grants 257383, 258588 and 2016-01-2410 and under the Thematic Networks program (Language Technologies Thematic Network project 281795). The work of the fourth author was partially supported by the SomEMBED TIN2015-71147-C2-1-P MINECO research project and by the Generalitat Valenciana under the grant ALMAMATER (Prometeo II/2014/030).

The goal of Author Profiling (AP) is to identify demographic aspects (e.g., age, gender) from a given set of authors by analyzing their written texts. Recently, the AP task has gained interest in many problems related to computer forensics, psychology, marketing, but specially in those related with social media exploitation. As known, social media data is shared through a wide range of modalities (e.g., text, images and audio), representing valuable information to be exploited for extracting valuable insights from users. Nevertheless, most of the current work in AP using social media data has been devoted to analyze textual information only, and there are very few works that have started exploring the gender identification using visual information. Contrastingly, this paper focuses in exploiting the visual modality to perform both age and gender identification in social media, specifically in Twitter. Our goal is to evaluate the pertinence of using visual information in solving the AP task. Accordingly, we have extended the Twitter corpus from PAN 2014, incorporating posted images from all the users, making a distinction between tweeted and retweeted images. Performed experiments provide interesting evidence on the usefulness of visual information in comparison with traditional textual representations for the AP task.

There is a considerable number of people with some disability they may go from partial limb disability to total incapacity. For these people, technology is an opportunity to bring them back some capabilities. In this work, we present a framework where we envisage a system that can be used by a disabled person who can not move but still possesses eye movements. Therefore, by using electroencephalographic (EEG) signals, we recognize and classify eye movements, which are then translated to control commands. Based on the latter, we developed an application to illustrate how such commands could be used to control a drone that could be used to deliver messages or carry out any other activity that involves the drone having to fly from a start point to a final destination. The results obtained in this study indicate that ocular movements are recognized with an accuracy of 86%, which suggests the feasibility of our approach.

Most of the researches in Electroencephalogram(EEG)-based Brain-Computer Interfaces (BCI) are focused on the use of motor imagery. As an attempt to improve the control of these interfaces, the use of language instead of movement has been recently explored, in the form of imagined speech. This work aims for the discrimination of imagined words in electroencephalogram signals. For this purpose, the analysis of multiple variables of the signal and their relation is considered by means of a multivariate data analysis, i.e., Parallel Factor Analysis (PARAFAC). In previous works, this method has demonstrated to be useful for EEG analysis. Nevertheless, to the best of our knowledge, this is the first attempt to analyze imagined speech signals using this approach. In addition, a novel use of the extracted PARAFAC components is proposed in order to improve the discrimination of the imagined words. The obtained results, besides of higher accuracy rates in comparison with related works, showed lower standard deviation among subjects suggesting the effectiveness and robustness of the proposed method. These results encourage the use of multivariate analysis for BCI applications in combination with imagined speech signals.

The Author Profiling (AP) task aims to reveal as much as possible information from a given author's document (e.g., age, gender, etc.). AP is crucial for several applications, ranging from customized advertising to computer forensics, psychology, and entertainment. Nonetheless, the AP task is far from being solved, particularly in social media domains, where the nature of documents hinder the applicability of state-of-the-art text mining tools (e.g., because of spelling-grammar errors, huge vocabularies, and the presence of many out-of-vocabulary terms). Currently, most of the work in AP for social media has been devoted to the development of descriptive features, which are used under standard representations, such as the Bag-of-Words (BoW). Nevertheless, BoW-like representations have some well known shortcomings, namely: (i) the sparsity and high dimensionality of the representation, and (ii) the failure to capture relationships, other than mere occurrence, among terms. This paper focuses on the study of alternative document representations that can deal with such issues. We propose a representation for documents that capture discriminative and subprofile-specific information of terms. Under the proposed representation, terms are represented in a vector space that captures discriminative information. Then, term representations are aggregated to represent the content of a document. In this manner, documents are represented in a low-dimensional (and discriminative) space which is non-sparse. We evaluate the effectiveness of the proposed representation on several corpora from the social media domain. The proposed representation is compared to the standard BoW representation and a wide variety of state-of-the-art AP approaches. Experimental results reveal that the proposed representation outperforms most of the reference methodologies. Furthermore, we show that the proposed representation is in agreement with previous studies on handcrafted attributes for AP.

An important task in plagiarism detection is determining and measuring similar text portions between a given pair of documents. One of the main difficulties of this task resides on the fact that reused text is commonly modified with the aim of covering or camouflaging the plagiarism. Another difficulty is that not all similar text fragments are examples of plagiarism, since thematic coincidences also tend to produce portions of similar text. In order to tackle these problems, we propose a novel method for detecting likely portions of reused text. This method is able to detect common actions performed by plagiarists such as word deletion, insertion and transposition, allowing to obtain plausible portions of reused text. We also propose representing the identified reused text by means of a set of features that denote its degree of plagiarism, relevance and fragmentation. This new representation aims to facilitate the recognition of plagiarism by considering diverse characteristics of the reused text during the classification phase. Experimental results employing a supervised classification strategy showed that the proposed method is able to outperform traditionally used approaches. 2012 Elsevier Ltd. All rights reserved., This work was done under partial support of CONACyT project Grants: 134186, and Scholarships: 258345/224483. This work is the result of the collaboration in the framework of the WIQEI IRSES project (Grant No. 269180) within the FP 7 Marie Curie. The work of the last author was in the framework of the VLC/CAMPUS Microcluster on Multimodal Interaction in Intelligent Systems.

The vast amount of electronic documents available on the Internet demands for automatic tools that help people finding, organizing and easily accessing to all this information. Although current text classification methods have alleviated some of the above problems, such strategies depend on having a large and reliable set of labeled data. In order to overcome such limitation, this work proposes an alternative approach for semi-supervised text classification, which is based on a new strategy for diminishing the sensitivity to the noise contained on labeled data by means of automatic text summarization. Experimental results showed that our proposed approach outperforms traditional semi-supervised text classification techniques; additionally, our results also indicate that our approach is suitable for learning from only one labeled example per category.

During the last decades the Web has become the greatest repository of digital information. In order to organize all this information, several text categorization methods have been developed, achieving accurate results in most cases and in very different domains. Due to the recent usage of Internet as communication media, short texts such as news, tweets, blogs, and product reviews are more common every day. In this context, there are two main challenges; on the one hand, the length of these documents is short, and therefore, the word frequencies are not informative enough, making text categorization even more difficult than usual. On the other hand, topics are changing constantly at a fast rate, causing the lack of adequate amounts of training data. In order to deal with these two problems we consider a text classification method that is supported on the idea that similar documents may belong to the same category. Mainly, we propose a neighborhood consensus classification method that classifies documents by considering their own information as well as information about the category assigned to other similar documents from the same target collection. In particular, the short texts we used in our evaluation are news titles with an average of 8 words. Experimental results are encouraging; they indicate that leveraging information from similar documents helped to improve classification accuracy and that the proposed method is especially useful when labeled training resources are limited.

In this paper we report the results obtained from experiments with a database of emotional speech in English in order to find the most important acoustic features to estimate Emotion Primitives which determine the emotional content on speech. We are interested in exploiting the potential benefits of continuous emotion models, so in this paper we demonstrate the feasibility of applying this approach to annotation of emotional speech and we explore ways to take advantage of this kind of annotation to improve the automatic classification of basic emotions.

Question answering (QA) is the task of automatically answering a question posed in natural language. Currently, there exists several QA approaches, and, according to recent evaluation results, most of them are complementary. That is, different systems are relevant for different kinds of questions. Somehow, this fact indicates that a pertinent combination of various systems should allow to improve the individual results. This paper focuses on this problem, namely, the selection of the correct answer from a given set of responses corresponding to different QA systems. In particular, it proposes a supervised multi-stream approach that decides about the correctness of answers based on a set of features that describe: (i) the compatibility between question and answer types, (ii) the redundancy of answers across streams, as well as (iii) the overlap and non-overlap information between the question-answer pair and the support text. Experimental results are encouraging; evaluated over a set of 190 questions in Spanish and using answers from 17 different QA systems, our multi-stream QA approach could reach an estimated QA performance of 0.74, significantly outperforming the estimated performance from the best individual system (0.53) as well as the result from best traditional multi-stream QA approach (0.60).

Most current methods for automatic text categorization are based on supervised learning techniques and, therefore, they face the problem of requiring a great number of training instances to construct an accurate classifier. In order to tackle this problem, this paper proposes a new semi-supervised method for text categorization, which considers the automatic extraction of unlabeled examples from the Web and the application of an enriched self-training approach for the construction of the classifier. This method, even though language independent, is more pertinent for scenarios where large sets of labeled resources do not exist. That, for instance, could be the case of several application domains in different non-English languages such as Spanish. The experimental evaluation of the method was carried out in three different tasks and in two different languages. The achieved results demonstrate the applicability and usefulness of the proposed method.

Sequential Pattern Mining is a widely addressed problem in data mining, with applications such as analyzing Web usage, examining purchase behavior, and text mining, among others. Nevertheless, with the dramatic increase in data volume, the current approaches prove inefficient when dealing with large input datasets, a large number of different symbols and low minimum supports. In this paper, we propose a new sequential pattern mining algorithm, which follows a pattern-growth scheme to discover sequential patterns. Unlike most pattern growth algorithms, our approach does not build a data structure to represent the input dataset, but instead accesses the required sequences through pseudo-projection databases, achieving better runtime and reducing memory requirements. Our algorithm traverses the search space in a depth-first fashion and only preserves in memory a pattern node linkage and the pseudo-projections required for the branch being explored at the time. Experimental results show that our new approach, the Node Linkage Depth-First Traversal algorithm (NLDFT), has better performance and scalability in comparison with state of the art algorithms.

This paper proposes a novel representation for Authorship Attribution (AA), based on Concise Semantic Analysis (CSA), which has been successfully used in Text Categorization (TC). Our approach for AA, called Document Author Representation (DAR), builds document vectors in a space of authors, calculating the relationship between textual features and authors. In order to evaluate our approach, we compare the proposed representation with conventional approaches and previous works using the c50 corpus. We found that DAR can be very useful in AA tasks, because it provides good performance on imbalanced data, getting comparable or better accuracy results.

The traditional approach for spoken document retrieval (SDR) uses an automatic speech recognizer (ASR) in combination with a word-based information retrieval method. This approach has only showed limited accuracy, partially because ASR systems tend to produce transcriptions of spontaneous speech with significant word error rate. In order to overcome such limitation we propose a method which uses word and phonetic-code representations in collaboration. The idea of this combination is to reduce the impact of transcription errors in the processing of some (presumably complex) queries by representing words with similar pronunciations through the same phonetic code. Experimental results on the CLEF-CLSR-2007 corpus are encouraging; the proposed hybrid method improved the mean average precision and the number of retrieved relevant documents from the traditional word-based approach by 3% and 7% respectively.

This paper examines the importance of different groups of speech acoustic features in the estimation of emotional primitives which define a three-dimensional continuous model of emotions. A set of proposed features is extracted from a database of German spontaneous emotional speech. This features set tries to represent several aspects of emotional content in speech that have been discussed separately in other works. Features selection and dimensionality reduction techniques are applied to find the features subsets that best estimate the emotional primitives Valence, Activation and Dominance. Finally, the emotional primitives are estimated using different classifiers based on the proposed features getting better results than the ones previously reported on related works.

Reinforcement Learning is commonly used for learning tasks in robotics, however, traditional algorithms can take very long training times. Reward shaping has been recently used to provide domain knowledge with extra rewards to converge faster. The reward shaping functions are normally defined in advance by the user and are static. This paper introduces a dynamic reward shaping approach, in which these extra rewards are not consistently given, can vary with time and may sometimes be contrary to what is needed for achieving a goal. In the experiments, a user provides verbal feedback while a robot is performing a task which is translated into additional rewards. It is shown that we can still guarantee convergence as long as most of the shaping rewards given per state are consistent with the goals and that even with fairly noisy interaction the system can still produce faster convergence times than traditional reinforcement learning techniques.

Crosslingual text classification consists of exploiting labeled documents in a source language to classify documents in a different target language. In addition to the evident translation problem, this task also faces some difficulties caused by the cultural discrepancies manifested in both languages by means of different topic distributions. Such discrepancies make the classifier unreliable for the categorization task. In order to tackle this problem we propose to improve the classification performance by using information embedded in the own target dataset. The central idea of the proposed approach is that similar documents must belong to the same category. Therefore, it classifies the documents by considering not only their own content but also information about the assigned category to other similar documents from the same target dataset. Experimental results using three different languages evidence the appropriateness of the proposed approach.

Current text classification methods are mostly based on a supervised approach, which require a large number of examples to build models accurate. Unfortunately, in several tasks training sets are extremely small and their generation is very expensive. In order to tackle this problem in this paper we propose a new text classification method that takes advantage of the information embedded in the own test set. This method is supported on the idea that similar documents must belong to the same category. Particularly, it classifies the documents by considering not only their own content but also information about the assigned category to other similar documents from the same test set. Experimental results in four data sets of different sizes are encouraging. They indicate that the proposed method is appropriate to be used with small training sets, where it could significantly outperform the results from traditional approaches such as Naive Bayes and Support Vector Machines.

Most common feature selection techniques for document categorization are supervised and require lots of training data in order to accurately capture the descriptive and discriminative information from the defined categories. Considering that training sets are extremely small in many classification tasks, in this paper we explore the use of unsupervised extractive summarization as a feature selection technique for document categorization. Our experiments using training sets of different sizes indicate that text summarization is a competitive approach for feature selection, and show its appropriateness for situations having small training sets, where it could clearly outperform the traditional information gain technique.

Geographic Information Retrieval (GIR) is a specialized Information Retrieval (IR) branch that deals with information related to geographical locations. Traditional IR engines are perfectly able to retrieve the majority of the relevant documents for most geographical queries, but they have severe difficulties generating a pertinent ranking of the retrieved results, which leads to poor performance. A key reason for this ranking problem has been a lack of information. Therefore, previous GIR research has tried to fill this gap using robust geographical resources (i.e. a geographical ontology), while other research with the same aim has used relevant feedback techniques instead. This paper explores the use of Bag of Concepts (BoC; a representation where documents are considered as the union of the meanings of its terms) and Holographic Reduced Representation (HRR; a novel representation for textual structure) as re-ranking mechanisms for GIR. Our results reveal an improvement in mean average precision (MAP) when compared to the traditional vector space model, even if Pseudo Relevance Feedback is employed.

The full deployment of service robots in daily activities will require the robot to adapt to the needs of non-expert users, particularly, to learn how to perform new tasks from "natural" interactions. Reinforcement learning has been widely used in robotics, however, traditional algorithms require long training times, and may have problems with continuous spaces. Programming by demonstration has been used to instruct a robot, but is limited by the quality of the trace provided by the user. In this paper, we introduce a novel approach that can handle continuous spaces, can produce continuous actions and incorporates the user's intervention to quickly learn optimal policies of tasks defined by the user. It is shown how the continuous actions produce smooth trajectories and how the user's intervention allows the robot to learn significantly faster optimal policies. The proposed approach is tested in a simulated robot with very promising results.

Multi-document summarization systems must be able to draw the "best" information from a set of documents.In this paper we propose a novel extractive approach for multidocument summarization based on the detection of locally relevant sentences. Our main hypothesis is that by extracting relevant sentences from each document within a collection, instead of considering all documents at once, the final multi-document summary will be of higher quality. Performed experiments showed that the proposed method is able to outperform conventional baselines as well as traditional approaches by constructing summaries of high quality according to the ROUGE evaluation metrics.

This paper describes experimental results of two approaches to multimedia image retrieval: annotation-based expansion and late fusion of mixed methods. The former formulation consists of expanding manual annotations with labels generated by automatic annotation methods. Experimental results show that the performance of text-based methods can be improved with this strategy, specially, for visual topics; motivating further research in several directions. The second approach consists of combining the outputs of diverse image retrieval models based on different information. Experimental results show that competitive performance, in both retrieval and results diversification, can be obtained with this simple strategy. It is interesting that, contrary to previous work, the best results of the fusion were obtained by assigning a high weight to visual methods. Furthermore, a probabilistic modeling approach to result-diversification is proposed; experimental results reveal that some modifications are needed to achieve satisfactory results with this method.

As any other classification task, Word Sense Disambiguation requires a large number of training examples. These examples, which are easily obtained for most of the tasks, are particularly difficult to obtain for this case. Based on this fact, in this paper we investigate the possibility of using a Web-based approach for determining the correct sense of an ambiguous word based only in its surrounding context. In particular, we propose a semi-supervised method that is specially suited to work with just a few training examples. The method considers the automatic extraction of unlabeled examples from the Web and their iterative integration into the training data set. The experimental results, obtained over a subset of ten nouns from the SemEval lexical sample task, are encouraging. They showed that it is possible to improve the baseline accuracy of classifiers such as Naive Bayes and SVM using some unlabeled examples extracted from the Web.

This year we evaluated our supervised answer validation method at both, the Spanish Answer Validation Exercise (AVE) and the Spanish Question Answering Main Task. This paper describes and analyzes our evaluation results from both tracks. In resume, the F-measure of the proposed method outperformed the baseline result of the AVE 2008 task by more than 100%, and enhanced the performance of our question answering system, showing a gain in accuracy of 22% for answering factoid questions. A detailed analysis of the results shows that the proposed non-overlap features are most discriminative than the traditional overlap ones. In particular, these novel features allowed increasing the F-measure result of our method by 26%.

Cross-language text classification (CLTC) aims to take advantage of existing training data from one language to construct a classifier for another language. In addition to the expected translation issues, CLTC is also complicated by the cultural distance between both languages, which causes that documents belonging to the same category concern very different topics. This paper proposes a re-classification method which purpose is to reduce the errors caused by this phenomenon by considering information from the own target language documents. Experimental results in a news corpus considering three pairs of languages and four categories demonstrated the appropriateness of the proposed method, which could improve the initial classification accuracy by up to 11%.

This paper focuses on the problem of ranking documents for Geographic Information Retrieval. It aims to demonstrate that by using some query-related example texts it is possible to improve the final ranking of the retrieved documents. Experimental results indicated that our approach could improve the MAP of some sets of retrieved documents using only two example texts.

Recent evaluation results from Geographic Information Retrieval (GIR) indicate that current information retrieval methods are effective to retrieve relevant documents for geographic queries, but they have severe difficulties to generate a pertinent ranking of them. Motivated by these results in this paper we present a novel re-ranking method, which employs information obtained through a relevance feedback process to perform a ranking refinement . Performed experiments show that the proposed method allows to improve the generated ranking from a traditional IR machine, as well as results from traditional re-ranking strategies such as query expansion via relevance feedback.

Some recent works have shown that the "perfect" selection of the best IR system per query could lead to a significant improvement on the retrieval performance. Motivated by this fact, in this paper we focus on the automatic selection of the best retrieval result from a given set of results lists generated by different IR systems. In particular, we propose five heuristic measures for evaluating the relative relevance of each result list, which take into account the redundancy and ranking of documents across the lists. Preliminary results in three different data sets, and considering 216 queries, are encouraging. They show that the proposed approach could slightly outperform the results from the best individual IR system in two out of three collections, but that it could significantly improve the average results of individual systems from all data sets. In addition, the achieved results indicate that our approach is a competitive alternative to traditional data fusion methods.

The bag of words representation (BoW), which is widely used in information retrieval (IR), represents documents and queries as word lists that do not express anything about context information. When we look for information, we find that not everything is explicitly stated in a document, so context information is needed to understand its content. This paper proposes the use of bag of concepts (BoC) and Holographic reduced representation (HRR) in IR. These representations go beyond BoW by incorporating context information to document representations. Both HRR and BoC are produced using a vector space methodology known as Random Indexing, and allow expressing additional knowledge from different sources. Our experiments have shown the feasibility of the representations and improved the mean average precision by up to 7% when they are compared with the traditional vector space model.

This paper describes the results of the INAOE's answer va-lidation system evaluated at the Spanish track of the AVE 2007. The system is based on a supervised learning approach that considers two kinds of attributes. On the one hand, some attributes indicating the textual entailment between the given support text and the hypothesis constructed from the question and answer. On the other hand, some new features denoting certain answer restrictions as imposed by the question's type and format. The evaluation results were encouraging; they reached a F-measure of 53% (the best performance in the Spanish track), and outperformed the standard baseline by 15 percentage points.

Nowadays there exist several kinds of question answering systems. According to recent evaluation results, most of these systems are complementary (i.e., each one is better than the others in answering some specific type of questions). This fact indicates that a pertinent combination of various systems may allow improving the best individual result. This paper focuses on this problem. It proposes using an answer validation method to handle this combination. The main advantage of this approach is that it does not rely on internal system's features nor depend on external answer's redundancies. Experimental results confirm the appropriateness of our proposal. They mainly show that it outperforms individual system's results as well as the precision obtained by a redundancy-based combination strategy.