On Unifying Information-based and Information-based Suggestive Word Extraction – In this paper, we present a new, unified approach to word embedding that enables direct learning of the word boundaries in a single unsupervised learning task. This approach is a novel way of unsupervised learning through a series of supervised transformations. Firstly, we propose an ensemble framework for word embedding learning, where the task is to learn a novel word boundary descriptor from data. The performance of the ensemble is assessed by comparing the performance of individual individual methods. Finally, we report experimental results on word embedding tasks.
We present a method for computing the likelihood of a given class of objects using a simple convex optimization procedure. The idea is to compute the best likelihood, which maximizes the sum of all possible class probabilities. To this end we show that, if we use the class probabilities for the unknown class, the procedure is linear in the number of classes. This is the key insight and we discuss its applications for a wide range of classes. We also provide examples for evaluating the performance of our method on real-world datasets.
Learning to Use Context Disparity in Learners for Retrieval and Learning
Learning from the Hindsight Plan: On Learning from Exact Time-series Data
On Unifying Information-based and Information-based Suggestive Word Extraction
Neural Embeddings for Sentiment Classification
Evaluating Deep Predictive Models on Unlabeled Data for Detecting Drug-Drug InteractionWe present a method for computing the likelihood of a given class of objects using a simple convex optimization procedure. The idea is to compute the best likelihood, which maximizes the sum of all possible class probabilities. To this end we show that, if we use the class probabilities for the unknown class, the procedure is linear in the number of classes. This is the key insight and we discuss its applications for a wide range of classes. We also provide examples for evaluating the performance of our method on real-world datasets.