Interactive Word Cloud for Analyzing Reviews

A five-star quality rating is one of the most widely used systems for evaluating items. However, it has two fundamental limitations: 1) the rating for one item cannot describe crucial information in detail; 2) the rating is not on an absolute scale for comparing to other items. Because of these limitations, users cannot make right decision. In this paper, we introduce our sophisticated approach to extract useful information from user reviews using collapsed dependencies and sentiment analysis. We propose an interactive word cloud that can show grammatical relationships among words, explore reviews efficiently, and display positivity or negativity on a sentence. In addition, we introduce visualization for comparing multiple word clouds and illustrate the usage through test cases. 

Project Athena

Project Athena provides a novel conference-interaction paradigm that allows agendas to be integrated with e-conferences. In this project, I worked on the server and user interfaces for an agenda-based conferencing system.

Emotional Story Generation Based on a Plot Graph

I introduce two novel approaches to retrieving stories that contain emotions based on a given emotional model extracted from another story. Undeniably, stories containing emotions are more compelling than stories without emotions. For instance, if we choose a happy story and create an emotional model based on the story, the generated model is used for extracting happy stories from a plot graph learned from a crowd. Basically, each emotional model is composed of a sequence of emotions generated from sentences in a story, and each sequence generates an emotional score. For instance, a happy-story model may consist of neutral, neutral, happy, and neutral emotions. One story can be composed of different emotions. To retrieve a story with a target emotional story, we propose two approaches: to maximize the emotional score and to compute the similarity between a story in a plot graph and an emotion model. Thus, we can ultimately retrieve an emotional story close to a target emotional story, and ultimately, a virtual agent should able to cope with diverse situations intelligently. 


N-J-N from a sample story
Generated story from a plot graph

N-A-N from a sample story
Generate story from a plot graph

N-J-N-J-N from a sample story

Generate story from a plot graph

Crowd-sourced Story Telling for Long-term Interaction

Long-term interaction is quite hard to machines since they do not have enough resources for the interaction and the repetition of behaviors. Thus, we introduce a method that can gather stories from the crowd and recommend a story to a user based on similarity-attraction. Moreover, we describe a virtual agent that can express the emotion of a story with non-verbal gesture for human-like interaction.

System architecture and data flow

Facial expressions for six emotions and neutral state

(a) GUI of the user client for interaction with a user (b) GUI for assessing a user’s personality and preferences

Conceptual Space for Multi-robot Systems

Symbolic representation and connectionist have various advantages, but they have inevitable disadvantages: frame problem, frame grounding, and difficulty in measuring similarity between objects. Especially, computing similarity is quite important in induction and learning. Thus, we suggest a memory model as well as memory update rules, and the memory model is based on Gädenfors’ the conceptual spaces compensating the problems of classical knowledge representations. The memory model is tested in a simulation environment with a mission that two heterogeneous robots equipped with different sensors search chemical weapons in an indoor environment.

Back to Portfolio

Resuable Q-learning and Q-learning Development Environment

I designed a resuable Q-learning software which can be applied various problems. However, it is not enough for beginners to use it. To reduce the effort to make actions and states of Q-learning for a problem, IDE for Q-learning was implemented. A user can organize states and actions in the IDE and exports the configuration as a XML file. Then, the Q-learning can import a XML file and create states and actions.

Back to Portfolio

TAME 2010

In 2010, Samsung Electronics and Gatech will create machine learning modules in support of affective behavior, specifically:
A)Using reinforcement (Q-)learning as the basis for adjusting the parameters of TAME and the underlying robotic behaviors, based on direct feedback of the user.
B) Using case-based reasoning (CBR) to store situation-specific affective parameters that can support discontinuous switching of TAME behavior based on user needs and situational requirements.
I am in charge of developing reusable Q-learning component in this project.

Back to Portfolio