179 research outputs found
Exploring high-level features for detecting cyberpedophilia
No full text[EN] In this paper, we suggest a list of high-level features and study their applicability in detection of cyberpedophiles. We used a corpus of chats downloaded from http://www.perverted-justice.com and two negative datasets of different nature: cybersex logs available online, and the NPS chat corpus. The classification results show that the NPS data and the pedophiles’ conversations can be accurately discriminated from each other with character n-grams, while in the more complicated case of cybersex logs there is need for high-level features to reach good accuracy levels. In this latter setting our results show that features that model behaviour and emotion significantly outperform the low-level ones, and achieve a 97% accuracy.The work of Dasha Bogdanova was partially carried out during the internship at the Universitat Politecnica de Valencia (scholarship of the University of St. Petersburg). Her research was partially supported by Google Research Award. The collaboration with Thamar Solorio was possible thanks to her one-month research visit at the Universitat Politecnica de Valencia (program PAID-PAID-02-11 award n. 1932). The research work of Paolo Rosso was done in the framework of the European Commission WIQ-EI Web Information Quality Evaluation Initiative (IRSES Grant No. 269180) project within the FP 7 Marie Curie People, the DIANA-APPLICATIONS - Finding Hidden Knowledge in Texts: Applications (TIN2012-38603-0O2-01) project, and the VLC/CAMPUS Microcluster on Multimodal Interaction in Intelligent Systems.Bogdanova, D.; Rosso, P.; Solorio, T. (2014). Exploring high-level features for detecting cyberpedophilia. Computer Speech and Language. 28(1):108-120. https://doi.org/10.1016/j.csl.2013.04.007S10812028
The Use of Orthogonal Similarity Relations in the Prediction of Authorship
No full textThe final publication is available at Springer via http://dx.doi.org/10.1007/978-3-642-37256-8_38Recent work on Authorship Attribution (AA) proposes the use of meta characteristics to train author models. The meta characteristics are orthogonal sets of similarity relations between the features from the different candidate authors. In that approach, the features are grouped and processed separately according to the type of information they encode, the so called linguistic modalities. For instance, the syntactic, stylistic and semantic features are each considered different modalities as they represent different aspects of the texts. The assumption is that the independent extraction of meta characteristics results in more informative feature vectors, that in turn result in higher accuracies. In this paper we set out to the task of studying the empirical value of this modality specific process. We experimented with different ways of generating the meta characteristics on different data sets with different numbers of authors and genres. Our results show that by extracting the meta characteristics from splitting features by their linguistic dimension we achieve consistent improvement of prediction accuracy.This research was partially supported by ONR grant N00014-12-1-0217 and by NSF award 1254108. It was also supported in part by the CONACYT grant 134186 and by the European Commission as part of the WIQ-EI project (project no. 269180) within the FP7 People Programme.Sapkota, U.; Solorio, T.; Montes Gómez, M.; Rosso, P. (2013). The use of orthogonal similarity relations in the prediction of authorship. En Computational Linguistics and Intelligent Text Processing. Springer Verlag (Germany). 463-475. https://doi.org/10.1007/978-3-642-37256-8_38S463475Baker, L.D., McCallum, A.: Distributional clustering of words for text classification. In: SIGIR 1998: Proceedings of the 21st Annual International ACM SIGIR, pp. 96–103. ACM, Melbourne (1998)Biber, D.: The multi-dimensional approach to linguistic analyses of genre variation: An overview of methodology and findings. Computers and the Humanities 26, 331–345 (1993)Blum, A., Mitchell, T.: Combining labeled and unlabeled data with co-training. In: Proceedings of the 1998 Conference on Computational Learning Theory (1998)Dhillon, I.S., Mallela, S., Kumar, R.: A divisive information-theoretic feature clsutering algorithm for text classification. Journal of Machine Learning Research 3, 1265–1287 (2003)Escalante, H.J., Montes-y-Gómez, M., Solorio, T.: A weighted profile intersection measure for profile-based authorship attribution. In: Batyrshin, I., Sidorov, G. (eds.) MICAI 2011, Part I. LNCS, vol. 7094, pp. 232–243. Springer, Heidelberg (2011)Escalante, H.J., Solorio, T., Montes-y-Gomez, M.: Local histograms of character n-grams for authorship attribution. In: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, pp. 288–298. Association for Computational Linguistics, Portland (2011)Hayes, J.H.: Authorship attribution: A principal component and linear discriminant analysis of the consistent programmer hypothesis. I. J. Comput. Appl., 79–99 (2008)Houvardas, J., Stamatatos, E.: N-gram feature selection for authorship identification. In: Euzenat, J., Domingue, J. (eds.) AIMSA 2006. LNCS (LNAI), vol. 4183, pp. 77–86. Springer, Heidelberg (2006)Karypis, G.: CLUTO - a clustering toolkit. Tech. Rep. #02-017 (November 2003)Keselj, V., Peng, F., Cercone, N., Thomas, C.: N-gram based author profiles for authorship attribution. In: Proceedings of the Pacific Association for Computational Linguistics, pp. 255–264 (2003)Koppel, M., Schler, J., Argamon, S.: Authorship attribution in the wild. Language Resources and Evaluation 45, 83–94 (2011)Lewis, D.D., Yang, Y., Rose, T.G., Li, F.: Rcv1: A new benchmark collection for text categorization research. Journal of Machine Learning Research 5, 361–397 (2004)Luyckx, K., Daelemans, W.: Authorship attribution and verification with many authors and limited data. In: Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), Manchester, UK, pp. 513–520 (August 2008)Luyckx, K., Daelemans, W.: The effect of author set size and data size in authorship attribution. In: Literary and Linguistic Computing, pp. 1–21 (August 2010)Marneffe, M.D., MacCartney, B., Manning, C.D.: Generating typed dependency parses from phrase structure parses. In: LREC 2006 (2006)Plakias, S., Stamatatos, E.: Tensor space models for authorship identification. In: Darzentas, J., Vouros, G.A., Vosinakis, S., Arnellos, A. (eds.) SETN 2008. LNCS (LNAI), vol. 5138, pp. 239–249. Springer, Heidelberg (2008)Raghavan, S., Kovashka, A., Mooney, R.: Authorship attribution using probabilistic context-free grammars. In: Proceedings of the ACL 2010 Conference Short Papers, pp. 38–42. Association for Computational Linguistics, Uppsala (2010)Slonim, N., Tishby, N.: The power of word clusters for text classification. In: 23rd European Colloquium on Information Retrieval Research, ECIR (2001)Solorio, T., Pillay, S., Raghavan, S., Montes-y-Gómez: Generating metafeatures for authorship attribution on web forum posts. In: Proceedings of the 5th International Joint Conference on Natural Language Processing, IJCNLP 2011, pp. 156–164. AFNLP, Chiang Mai (2011)Stamatatos, E.: Author identification using imbalanced and limited training texts. In: 18th International Workshop on Database and Expert Systems Applications, DEXA 2007, pp. 237–241 (September 2007)Stamatatos, E.: Author identification: Using text sampling to handle the class imbalance problem. Information Processing and Managemement 44, 790–799 (2008)Stamatatos, E.: Plagiarism detection using stopword n-grams. Journal of the American Society for Information Science and Technology 62(12), 2512–2527 (2011)Stamatatos, E.: A survey on modern authorship attribution methods. Journal of the American Society for Information Science and Technology 60(3), 538–556 (2009)Stolcke, A.: SRILM - an extensible language modeling toolkit, pp. 901–904 (2002)Toutanova, K., Klein, D., Manning, C.D., Singer, Y.: Feature-rich part-of-speech tagging with a cyclic dependency network. In: Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology, NAACL 2003, vol. 1, pp. 173–180 (2003)de Vel, O., Anderson, A., Corney, M., Mohay, G.: Multi-topic e-mail authorship attribution forensics. In: Proceedings of the Workshop on Data Mining for Security Applications, 8th ACM Conference on Computer Security (2001)Witten, I.H., Frank, E.: Data Mining: Practical Machine Learning Tools and Techniques, 2nd edn. Morgan Kaufmann (2005
Analyzing Errors of Neural Models in Named Entity Recognition
No full textDespite stellar performance on many NLP tasks, the behavior of neural models like BERT is not properly understood. We attempt to analyze the behavior and recognize patterns in errors for the NER task. We evaluate the predictions and errors generated to gain insight into the model's behavior Our findings show that there are underlying patterns leading to unintended memorization. Future research is required to address these errors and fine-tune the model.Computer Science, Department ofHonors Colleg
Show Some Love to Your n-grams: A Bit of Progress and Stronger n-gram Language Modeling Baselines
No full textIn recent years neural language models (LMs) have set state-of-the-art performance for several benchmarking datasets. While the reasons for their success and their computational demand are well-documented, a comparison between neural models and more recent developments in n-gram models is neglected. In this paper, we examine the recent progress in n-gram literature, running experiments on 50 languages covering all morphological language families. Experimental results illustrate that a simple extension of Modified Kneser-Ney outperforms an LSTM language model on 42 languages while a word-level Bayesian n-gram LM outperforms the character-aware neural model on average across all languages, and its extension which explicitly injects linguistic knowledge on 8 languages. Further experiments on larger Europarl datasets for 3 languages indicate that neural architectures are able to outperform computationally much cheaper n-gram models: n-gram training is up to 15,000 times quicker. Our experiments illustrate that standalone n-gram models lend themselves as natural choices for resource-lean or morphologically rich languages, while the recent progress has significantly improved their accuracy
NextGen AML: distributed deep learning based language technologies to augment anti money laundering Investigation
No full textMost of the current anti money laundering (AML) systems, using handcrafted
rules, are heavily reliant on existing structured databases, which are not capable
of effectively and efficiently identifying
hidden and complex ML activities, especially those with dynamic and timevarying characteristics, resulting in a high
percentage of false positives. Therefore,
analysts1
are engaged for further investigation which significantly increases human capital cost and processing time. To
alleviate these issues, this paper presents
a novel framework for the next generation AML by applying and visualizing deep learning-driven natural language
processing (NLP) technologies in a distributed and scalable manner to augment
AML monitoring and investigation. The
proposed distributed framework performs
news and tweet sentiment analysis, entity recognition, relation extraction, entity linking and link analysis on different data sources (e.g. news articles and
tweets) to provide additional evidence to
human investigators for final decisionmaking. Each NLP module is evaluated
on a task-specific data set, and the overall experiments are performed on synthetic
and real-world datasets. Feedback from
AML practitioners suggests that our system can reduce approximately 30% time
and cost compared to their previous manual approaches of AML investigation
Multimodal representation learning with neural networks
No full textAbstract: Representation learning methods have received a lot of attention by researchers and practitioners because of their successful application to complex problems in areas such as computer vision, speech recognition and text processing [1]. Many of these promising results are due to the development of methods to automatically learn the representation of complex objects directly from large amounts of sample data [2]. These efforts have concentrated on data involving one type of information (images, text, speech, etc.), despite data being naturally multimodal. Multimodality refers to the fact that the same real-world concept can be described by different views or data types. Addressing multimodal automatic analysis faces three main challenges: feature learning and extraction, modeling of relationships between data modalities and scalability to large multimodal collections [3, 4]. This research considers the problem of leveraging multiple sources of information or data modalities in neural networks. It defines a novel model called gated multimodal unit (GMU), designed as an internal unit in a neural network architecture whose purpose is to find an intermediate representation based on a combination of data from different modalities. The GMU learns to decide how modalities influence the activation of the unit using multiplicative gates. The GMU can be used as a building block for different kinds of neural networks and can be seen as a form of intermediate fusion. The model was evaluated on four supervised learning tasks in conjunction with fully-connected and convolutional neural networks. We compare the GMU with other early and late fusion methods, outperforming classification scores in the evaluated datasets. Strategies to understand how the model gives importance to each input were also explored. By measuring correlation between gate activations and predictions, we were able to associate modalities with classes. It was found that some classes were more correlated with some particular modality. Interesting findings in genre prediction show, for instance, that the model associates the visual information with animation movies while textual information is more associated with drama or romance movies. During the development of this project, three new benchmark datasets were built and publicly released. The BCDR-F03 dataset which contains 736 mammography images and serves as benchmark for mass lesion classification. The MM-IMDb dataset containing around 27000 movie plots, poster along with 50 metadata annotations and that motivates new research in multimodal analysis. And the Goodreads dataset, a collection of 1000 books that encourages the research on success prediction based on the book content. This research also facilitates reproducibility of the present work by releasing source code implementation of the proposed methods.Doctorad
Selective attention for context-aware Neural Machine Translation
No full textDespite the progress made in sentence-level NMT, current systems still fall short at achieving fluent, good quality translation for a full document. Recent works in context-aware NMT consider only a few previous sentences as context and may not scale to entire documents. To this end, we propose a novel and scalable top-down approach to hierarchical attention for context-aware NMT which uses sparse attention to selectively focus on relevant sentences in the document context and then attends to key words in those sentences. We also propose single-level attention approaches based on sentence or word-level information in the context. The document-level context representation, produced from these attention modules, is integrated into the encoder or decoder of the Transformer model depending on whether we use monolingual or bilingual context. Our experiments and evaluation on English-German datasets in different document MT settings show that our selective attention approach not only significantly outperforms context-agnostic baselines but also surpasses context-aware baselines in most cases.</p
Stylistically Aware Representations of Books
No full textThe conscious or unconscious choices made by an author to use some language forms constantly over other possible forms constitute the style of the author. Capturing style embedded in documents has a wide range of applications across many domains. In this dissertation, we propose a multitude of hand-crafted lexical, syntactic, and stylistic features together with novel deep learning methods to capture different stylistic markers embedded in documents. The methods are general enough to be applied to any domain. Here, we evaluate on an interesting and important domain: Books. The deeper study of stylistic variations will reveal the dos and don'ts of successful authors, which might help authors in shaping their writings and readers discover new books suited to their taste. We empirically show that traditional hand-crafted features and deep learning methods capture complementary information which upon careful combination yield better performance. Moreover, we find that adding an auxiliary task of genre classification to the primary task of success prediction improves results. Next, we propose a novel multimodal neural architecture that incorporates genre supervision to assign weights to individual feature types. As compared to previous ad-hoc feature combinations, which is time consuming and rigid, this method is capable of dynamically tailoring weights given to feature types based on the characteristics of each book. We then explore the authors' dexterity in use of emotion flow across the entire books to captivate readers. We show that modeling the sequential flow of emotions depicted across entire book performs better than without taking this information into account. Finally, we propose a novel method to learn stylistically aware embeddings for authors by feeding in the stylistic traits from their writings. These embeddings also prove to be assets in predicting books' likability.Computer Science, Department o
Analyzing Notions of Artistic Style Using Computer Vision Techniques
No full textComputer vision has made significant strides in the area of artistic style transfer, and a few attempts have been made to extract and define the style signature of various artists. However, most of these endeavors have been limited by treating a creative task such as painting and critiquing style as a traditional machine learning problem. In this study, we try to shift the viewpoint from machine learning trying to solve an art problem, to one where the art world is using computer vision techniques to fit its purpose. This subtle difference is extremely important because it allows us to build notions of style in a bottom up fashion, rooted in the domain knowledge pertaining to artistic style.
This work aims to take first steps towards building an understanding of similarity in artistic style with the intention of critiquing and valuing art. I begin by constructing a dataset of approximately 14,000 high-resolution paintings, which are part of the Google Art Project, available under the Wikimedia Commons License. I explore the possibility of neural style transfer features being good measures of style similarity, and proceed to develop computationally useful style features for scene composition, color palette, brush strokes, and contours. I conduct a domain-specific study wtih experts to validate the importance of each of the style features. I then present a novel way to normalize and weight these features based on the external study and develop a cumulative measure of artistic style similarity. Finally, I validate the results qualitatively on historically accepted examples in the art community, and quantitatively via a second domain study with experts.Computer Science, Department o
Text Generation from Knowledge Graphs with Graph Transformers
No full textGenerating texts which express complex ideas spanning multiple sentences requires a structured representation of their content (document plan), but these representations are prohibitively expensive to manually produce. In this work, we address the problem of generating coherent multi-sentence texts from the output of an information extraction system, and in particular a knowledge graph. Graphical knowledge representations are ubiquitous in computing, but pose a significant challenge for text generation techniques due to their non-hierarchical nature, collapsing of longdistance dependencies, and structural variety. We introduce a novel graph transforming encoder which can leverage the relational structure of such knowledge graphs without imposing linearization or hierarchical constraints. Incorporated into an encoder-decoder setup, we provide an end-to-end trainable system for graph-to-text generation that we apply to the domain of scientific text. Automatic and human evaluations show that our technique produces more informative texts which exhibit better document structure than competitive encoder-decoder methods
- …
