... training Cyril Goutte February 29, 1996 Abstract. ... (2) In a recent paper, Hansen and Larse... more ... training Cyril Goutte February 29, 1996 Abstract. ... (2) In a recent paper, Hansen and Larsen (Hansen and Larsen, 1995) suggested the use of linear unlearning to overcome the cost of a full cross-validation scheme, and obtain the following estimate: bGLU =1 N N X k=1 ...
2009 IEEE International Workshop on Machine Learning for Signal Processing, 2009
ABSTRACT Detecting and tracking of temporal data is an important task in multiple applications. I... more ABSTRACT Detecting and tracking of temporal data is an important task in multiple applications. In this paper we study temporal text mining methods for Music Information Retrieval. We compare two ways of detecting the temporal latent semantics of a corpus extracted from Wikipedia, using a stepwise Probabilistic Latent Semantic Analysis (PLSA) approach and a global multiway PLSA method. The analysis indicates that the global analysis method is able to identify relevant trends which are difficult to get using a step-by- ...
... learning algorithm. Also [10] combined multi-view learning with active learning in Co-EMT. In... more ... learning algorithm. Also [10] combined multi-view learning with active learning in Co-EMT. In [15 ... 358, 1998. [9]Muslea I., Minton S., and Knoblock C. Active + semi-supervised learning = robust multi-view learning. In Proceedings ...
We propose to use a statistical phrasebased machine translation system in a post-editing task: th... more We propose to use a statistical phrasebased machine translation system in a post-editing task: the system takes as input raw machine translation output (from a commercial rule-based MT system), and produces post-edited target-language text. We report on experiments that were performed on data collected in precisely such a setting: pairs of raw MT output and their manually post-edited versions. In our evaluation, the output of our automatic post-editing (APE) system is not only better quality than the rule-based MT (both in terms of the BLEU and TER metrics), it is also better than the output of a stateof-the-art phrase-based MT system used in standalone translation mode. These results indicate that automatic post-editing constitutes a simple and efficient way of combining rule-based and statistical MT technologies.
We present some on-going research on phrase-based Statistical Machine Translation using flexible ... more We present some on-going research on phrase-based Statistical Machine Translation using flexible phrases that may contain gaps of variable lengths. This allows us to naturally handle various linguistic phenomena such as negations or separable particles. We integrate this within the standard Maximum Entropy model using some dedicated feature functions, and describe a beam-search stack decoder that handles these noncontiguous, elastic phrases. Preliminary experimental results show that the translation performance compares favourably with phrase-based MT using fixed gap size. We expect that future results may allow us to leverage the added flexibility of elastic chunks to further increase translation performance.
In this contribution, we suggest a convenient way to use generalisation error to extract the rele... more In this contribution, we suggest a convenient way to use generalisation error to extract the relevant delays from a timevarying process, i.e. the delays that lead to the best prediction performance. We design a generalisation-based algorithm that takes its inspiration from traditional variable selection, and more precisely stepwise forward selection. The method is compared to other forward selection schemes, as well as to a non-parametric tests aimed at estimating the embedding dimension of time series. The nal application extends these results to the e cient estimation of FIR lters on some real data.
The Journal of Machine Learning Research, Mar 1, 2003
We address the problem of categorising documents using kernel-based methods such as Support Vecto... more We address the problem of categorising documents using kernel-based methods such as Support Vector Machines. Since the work of Joachims (1998), there is ample experimental evidence that SVM using the standard word frequencies as features yield state-of-the-art performance on a number of benchmark problems. Recently, Lodhi et al. (2002) proposed the use of string kernels, a novel way of computing document similarity based of matching non-consecutive subsequences of characters. In this article, we propose the use of this technique with sequences of words rather than characters. This approach has several advantages, in particular it is more efficient computationally and it ties in closely with standard linguistic pre-processing techniques. We present some extensions to sequence kernels dealing with symbol-dependent and match-dependent decay factors, and present empirical evaluations of these extensions on the Reuters-21578 datasets.
Proceedings of the Acm Tenth International Workshop on Data Warehousing and Olap, 2007
Databases and data warehouses contain an overwhelming volume of information that users must wade ... more Databases and data warehouses contain an overwhelming volume of information that users must wade through in order to extract valuable and actionable knowledge to support the decision-making process. This contribution addresses the problem of automatically analyzing large multidimensional tables to get a concise representation of data, identify patterns and provide approximate answers to queries.
Motivation: Searching relevant publications for manual database annotation is a tedious task. In ... more Motivation: Searching relevant publications for manual database annotation is a tedious task. In this paper, we apply a combination of Natural Language Processing (NLP) and probabilistic classification to re-rank documents returned by PubMed according to their relevance to Swiss-Prot annotation, and to identify significant terms in the documents. Results: With a Probabilistic Latent Categoriser (PLC) we obtained 69% recall and 59% precision for relevant documents in a representative query. As the PLC technique provides the relative contribution of each term to the final document score, we used the Kullback-Leibler symmetric divergence to determine the most discriminating words for Swiss-Prot medical annotation. This information should allow curators to understand classification results better. It also has great value for fine-tuning the linguistic preprocessing of documents, which in turn can improve the overall classifier performance. Availability: The medical annotation dataset is available from the authors upon request Contact: [email protected];
Abstract: We propose a new hierarchical generative modelfor textual data, where words may be gene... more Abstract: We propose a new hierarchical generative modelfor textual data, where words may be generated by topicspecific distributions at any level in the hierarchy. Thismodel is naturally well-suited to clustering documents inpreset or automatically generated hierarchies, as well as categorisingnew documents in an existing hierarchy. Furthermore,we present a series of applications that can benefitfrom our model, as well as an experimental evaluation forboth clustering and categorisation on...
... training Cyril Goutte February 29, 1996 Abstract. ... (2) In a recent paper, Hansen and Larse... more ... training Cyril Goutte February 29, 1996 Abstract. ... (2) In a recent paper, Hansen and Larsen (Hansen and Larsen, 1995) suggested the use of linear unlearning to overcome the cost of a full cross-validation scheme, and obtain the following estimate: bGLU =1 N N X k=1 ...
2009 IEEE International Workshop on Machine Learning for Signal Processing, 2009
ABSTRACT Detecting and tracking of temporal data is an important task in multiple applications. I... more ABSTRACT Detecting and tracking of temporal data is an important task in multiple applications. In this paper we study temporal text mining methods for Music Information Retrieval. We compare two ways of detecting the temporal latent semantics of a corpus extracted from Wikipedia, using a stepwise Probabilistic Latent Semantic Analysis (PLSA) approach and a global multiway PLSA method. The analysis indicates that the global analysis method is able to identify relevant trends which are difficult to get using a step-by- ...
... learning algorithm. Also [10] combined multi-view learning with active learning in Co-EMT. In... more ... learning algorithm. Also [10] combined multi-view learning with active learning in Co-EMT. In [15 ... 358, 1998. [9]Muslea I., Minton S., and Knoblock C. Active + semi-supervised learning = robust multi-view learning. In Proceedings ...
We propose to use a statistical phrasebased machine translation system in a post-editing task: th... more We propose to use a statistical phrasebased machine translation system in a post-editing task: the system takes as input raw machine translation output (from a commercial rule-based MT system), and produces post-edited target-language text. We report on experiments that were performed on data collected in precisely such a setting: pairs of raw MT output and their manually post-edited versions. In our evaluation, the output of our automatic post-editing (APE) system is not only better quality than the rule-based MT (both in terms of the BLEU and TER metrics), it is also better than the output of a stateof-the-art phrase-based MT system used in standalone translation mode. These results indicate that automatic post-editing constitutes a simple and efficient way of combining rule-based and statistical MT technologies.
We present some on-going research on phrase-based Statistical Machine Translation using flexible ... more We present some on-going research on phrase-based Statistical Machine Translation using flexible phrases that may contain gaps of variable lengths. This allows us to naturally handle various linguistic phenomena such as negations or separable particles. We integrate this within the standard Maximum Entropy model using some dedicated feature functions, and describe a beam-search stack decoder that handles these noncontiguous, elastic phrases. Preliminary experimental results show that the translation performance compares favourably with phrase-based MT using fixed gap size. We expect that future results may allow us to leverage the added flexibility of elastic chunks to further increase translation performance.
In this contribution, we suggest a convenient way to use generalisation error to extract the rele... more In this contribution, we suggest a convenient way to use generalisation error to extract the relevant delays from a timevarying process, i.e. the delays that lead to the best prediction performance. We design a generalisation-based algorithm that takes its inspiration from traditional variable selection, and more precisely stepwise forward selection. The method is compared to other forward selection schemes, as well as to a non-parametric tests aimed at estimating the embedding dimension of time series. The nal application extends these results to the e cient estimation of FIR lters on some real data.
The Journal of Machine Learning Research, Mar 1, 2003
We address the problem of categorising documents using kernel-based methods such as Support Vecto... more We address the problem of categorising documents using kernel-based methods such as Support Vector Machines. Since the work of Joachims (1998), there is ample experimental evidence that SVM using the standard word frequencies as features yield state-of-the-art performance on a number of benchmark problems. Recently, Lodhi et al. (2002) proposed the use of string kernels, a novel way of computing document similarity based of matching non-consecutive subsequences of characters. In this article, we propose the use of this technique with sequences of words rather than characters. This approach has several advantages, in particular it is more efficient computationally and it ties in closely with standard linguistic pre-processing techniques. We present some extensions to sequence kernels dealing with symbol-dependent and match-dependent decay factors, and present empirical evaluations of these extensions on the Reuters-21578 datasets.
Proceedings of the Acm Tenth International Workshop on Data Warehousing and Olap, 2007
Databases and data warehouses contain an overwhelming volume of information that users must wade ... more Databases and data warehouses contain an overwhelming volume of information that users must wade through in order to extract valuable and actionable knowledge to support the decision-making process. This contribution addresses the problem of automatically analyzing large multidimensional tables to get a concise representation of data, identify patterns and provide approximate answers to queries.
Motivation: Searching relevant publications for manual database annotation is a tedious task. In ... more Motivation: Searching relevant publications for manual database annotation is a tedious task. In this paper, we apply a combination of Natural Language Processing (NLP) and probabilistic classification to re-rank documents returned by PubMed according to their relevance to Swiss-Prot annotation, and to identify significant terms in the documents. Results: With a Probabilistic Latent Categoriser (PLC) we obtained 69% recall and 59% precision for relevant documents in a representative query. As the PLC technique provides the relative contribution of each term to the final document score, we used the Kullback-Leibler symmetric divergence to determine the most discriminating words for Swiss-Prot medical annotation. This information should allow curators to understand classification results better. It also has great value for fine-tuning the linguistic preprocessing of documents, which in turn can improve the overall classifier performance. Availability: The medical annotation dataset is available from the authors upon request Contact: [email protected];
Abstract: We propose a new hierarchical generative modelfor textual data, where words may be gene... more Abstract: We propose a new hierarchical generative modelfor textual data, where words may be generated by topicspecific distributions at any level in the hierarchy. Thismodel is naturally well-suited to clustering documents inpreset or automatically generated hierarchies, as well as categorisingnew documents in an existing hierarchy. Furthermore,we present a series of applications that can benefitfrom our model, as well as an experimental evaluation forboth clustering and categorisation on...
Uploads
Papers by Cyril Goutte