A Hybrid Definition of Lexical Similarity for Extraction of Meaning from Interlingual Sources – We focus our attention today on the problem of lexical similarity detection (LSC), in which a user identifies a topic, and then uses this task to discover the category of topics for the topic. This paper describes an approach for semantic similarity detection by leveraging a hybrid definition of lexical similarity which takes the topic as input and outputs a dictionary representation of the topic. We give experimental results on the COCO-SOCA benchmark dataset, that show that our proposed method outperforms both other conventional lexical similarity detection methods in terms of accuracy.
In this paper, we propose a new approach to nonlinearly smooth multi-label classification with adaptive clustering that achieves similar performance and generalization guarantees as the classic multi-label learning on both unsupervised and supervised datasets, and significantly improves the generalization error for these two datasets. We then provide a generic algorithm for learning to distinguish between three classes of data points, which we call the three-class class clustering (3CL). A 3CL is a supervised classifier that is not only unbiased but also adaptable to the environment. We show that for 3CL to provide a more accurate classification performance, we need to learn the discriminative class of data points, which is the first step in training a 3CL. Moreover, we show that learning 3CL with regularizing rule improves classification accuracy for the same classification task.
Robust PCA in Speech Recognition: Training with Noise and Frequency Consistency
A Hybrid Definition of Lexical Similarity for Extraction of Meaning from Interlingual Sources
Auxiliary Weight Normalization and Relaxation Paths for Learning Uncertain DataIn this paper, we propose a new approach to nonlinearly smooth multi-label classification with adaptive clustering that achieves similar performance and generalization guarantees as the classic multi-label learning on both unsupervised and supervised datasets, and significantly improves the generalization error for these two datasets. We then provide a generic algorithm for learning to distinguish between three classes of data points, which we call the three-class class clustering (3CL). A 3CL is a supervised classifier that is not only unbiased but also adaptable to the environment. We show that for 3CL to provide a more accurate classification performance, we need to learn the discriminative class of data points, which is the first step in training a 3CL. Moreover, we show that learning 3CL with regularizing rule improves classification accuracy for the same classification task.