Jeong-Won Cha

MMR-based Active Machine Learning for Bio Named Entity Recognition

This paper presents a new active learning paradigm which considers not only the uncertainty of the classifier but also the diversity of the corpus. The two measures for uncertainty and diversity were combined using the MMR (Maximal Marginal Relevance) method to give the sampling scores in our active learning strategy. We incorporated MMR-based active machine-learning idea into the biomedical named-entity recognition system. Our experimental results indicated that our strategies for active-learning based sample selection could significantly reduce the human effort.

Evaluation of User Reputation on YouTube

In the Web 2.0 era, people not only read web contents but upload, view, share and evaluate all contents on the web. This leads us to introduce a new type of social network that is based on user activity and content metadata. Moreover, we can determine the quality of related contents using this new social network. Based on this observation, we introduce a user evaluation algorithm for user-generated video sharing website such as YouTube.

A smart movie recommendation system

We propose a movie recommendation system based on genre correlations. We modify the previous algorithm; we use a list of movies as input instead of genre combinations. We implement a new recommendation algorithm as Android application with additional functions. By combining with existing web services such as Google Movie Showtimes and Open APIs, our system can recommend movies playing in cinemas currently and show the detailed information of movies. Location-based function is also implemented. We utilize GPS information of mobile device and web service provided by Google Maps for recommending suitable cinemas for users with mobile devices.

Predicting Phrase-Level Tags Using Entropy Inspired Discriminative Models

In this paper, we describe a system which predicts phrase-level tags for eojeols in Korean using entropy inspired discriminative probabilistic models such as a conditional random fields. Instead of selecting features by the intuition of user, we use a decision tree and error analysis systematically for selecting the best feature. Once we generate all available features from the corpus, then select features by using decision tree and error analysis iteratively. Experimental results show 93.90% and 49.46% accuracy for eojeols and sentences respectively. This accuracy eventually is able to improve further syntactic analysis results. We find from the results that the better meaningful features using systematic methods is good at raising performance.

Identifying representative reviewers in internet social media

In a society, we have many forms of relations with other people from home, work or school. These relationships give rise to a social network. People in a social network receive, provide and pass lots of information. We often observe that there are a group of people who have high influence to other people. We call these high influence people opinion leaders. Thus, it is important and useful to identify opinion leaders in a social network. In Web 2.0, there are many user participations and we can create a social network from the user activities. We propose a simple yet reliable algorithm that finds opinion leaders in a cyber social network. We consider a social network of users who rate musics and identify representative users of the social network. Then, we verify the correctness of the proposed algorithm by the T-test.

Named entity tagging for korean using DL-CoTrain algorithm

Our approach to solve the problem of Korean named entity classification adopted a co-training method called DL-CoTrain. We use only a part-of-speech tagger and a simple noun phrase chunker instead of a full parser to extract the contextual features of a named entity. We will discuss the linguistic features in Korean which are valuable for named entity classification and experimentally show how large a labeled corpus and which unlabeled corpus is necessary for the better performance and portability of a named entity classifier. With only about a quarter of the labeled corpus, our method can compete with its supervised counterpart.

User Reputation Evaluation Using Co-occurrence Feature and Collective Intelligence

It becomes more difficult to find valuable contents in the Web 2.0 environment since lots of inexperienced users provide many unorganized contents. In the previous researches, people has proved that non-text information such as the number of references, the number of supports, and the length of answers is effective to evaluate answers to a question in a online QnA service site. However, these features can be changed easily by users and cannot reflect social activity of users. In this paper, we propose a new method to evaluate user reputation using co-occurrence features between question and answers, and collective intelligence. If we are able to calculate user reputation, then we can estimate the worth of contents that has small number of reference and small number of support. We compute the user reputation using a modified PageRank algorithm. The experiment results show that our proposed method is effective and useful for identifying such contents.

End-to-end task dependent recurrent entity network for goal-oriented dialog learning

Abstract In this paper, we introduce the Task Dependent Recurrent Entity Network (TDREN) to solve Dialogue System Technology Challenges 6 (DSTC 6) track 1. Traditionally, there have been methods such as collecting the intent of the user in a conversation directly using rules. We design an end-to-end structure that properly models the restaurant pre-related user preferences that appear in the dialogue and gives appropriate responses. We perform experiments on the TDREN and achieved 97.7% at precision 1. We propose a new artificial neural network structure and recurrent cell for modeling user preference information. Then, we show that task-oriented dialogue modeling experiment results using the structure and the recurrent cell.

Reliability of Representative Reviewers on the Web

There are opinion leaders in a society who represent the opinion of general public. The general public accepts information not only by mass media but also by opinion leaders. Since the late 20th century, the number of Internet users has increased fast. Many users interact with each other in an online social network. This makes the Web community similar to the real society. Thus it is a natural task to find influential users in an online society. For example, many online articles posted by influential bloggers are used as marketing tools for companies or political advertisements for parties since these articles have huge influence to other users. We first revisit the previous researches on finding influential users in online society. Then we compose test sets from the GroupLens movie database and identify representative reviewers. Next, we show the validity of the chosen representative reviewers and test the reliability using the 10-fold cross- validation. Finally, we explain applicability of the proposed approach to improve current recommendation systems.

A Joint Statistical Model for Simultaneous Word Spacing and Spelling Error Correction for Korean

This paper presents noisy-channel based Korean preprocessor system, which corrects word spacing and typographical errors. The proposed algorithm corrects both errors simultaneously. Using Eojeol transition pattern dictionary and statistical data such as Eumjeol n-gram and Jaso transition probabilities, the algorithm minimizes the usage of huge word dictionaries.