• Publications
  • Influence
An empirical evaluation of supervised learning in high dimensions
To the surprise, the method that performs consistently well across all dimensions is random forests, followed by neural nets, boosted trees, and SVMs, and the effect of increasing dimensionality on the performance of the learning algorithms changes. Expand
Multi-Level Structured Models for Document-Level Sentiment Classification
A joint two-level approach for document-level sentiment classification that simultaneously extracts useful (i.e., subjective) sentences and predicts document- level sentiment based on the extracted sentences is proposed. Expand
Computational Approaches to Sentence Completion
It is found that by fusing local and global information, the ability of algorithms to distinguish sense from nonsense based on a variety of sentence-level phenomena can exceed 50% on this task (chance baseline is 20%), and some avenues for further research are suggested. Expand
Compositional Matrix-Space Models for Sentiment Analysis
This paper presents the first such algorithm for learning a matrix-space model for semantic composition, and its experimental results show statistically significant improvements in performance over a bag-of-words model. Expand
Automatically Generating Annotator Rationales to Improve Sentiment Classification
This work explores methods to automatically generate annotator rationales for document-level sentiment classification and finds the automatically generated rationales just as helpful as human rationales. Expand
Cs 674/info 630: Advanced Language Technologies Lecture 7 — September 18 2 Incorporating Term Frequencies
Apart from IDF, term frequencies are also important and we would like to incorporate them into our scoring function. From now on, we will treat Aj as a random variable that denotes the number ofExpand
Exploring Heterogeneous Metadata for Video Recommendation with Two-tower Model
This work shows the feasibility of using two-tower model for recommendations and conducts a series of offline experiments to show its performance for cold-start titles, and explores different types of metadata (categorical features, text description, cover-art image) and an attention layer to fuse them. Expand
Exploiting structure for sentiment classification
This thesis develops computational models that capture useful structure-based intuitions for solving each task, treating the intuitions as latent representations to be discovered and exploited during learning. Expand
Cs 674/info 630: Advanced Language Technologies Lecture 1 — August 28
In the previous lecture we saw two examples of research in the computational analysis of (textual) natural language data. This single area (sometimes divided up into subareas called InformationExpand
Cs 674/info 630: Advanced Language Technologies Lecture 17 — October 30
At the end of the previous lecture we were talking about how to incorporate implicit relevance feedback which came in the form of preferences, i.e. instead of absolute judgments (this document isExpand