Craig Martell

Lexical and Discourse Analysis of Online Chat Dialog

Abstract : One of the goals of natural language processing (NLP) systems is determining the meaning of what is being transmitted. Although much work has been accomplished in traditional written and spoken language domains, little has been performed in the newer computer-mediated communication domain enabled by the Internet, to include text-based chat. This is due in part to the fact that there are no annotated chat corpora available to the broader research community. The purpose of our research is to build a chat corpus, initially tagged with lexical and discourse information. Such a corpus could be used to develop stochastic NLP applications that perform tasks such as conversation thread topic detection, author profiling, entity identification, and social network analysis. During the course of our research, we preserved 477,835 chat posts and associated user profiles in an XML format for future investigation. We privacy-masked 10,567 of those posts and part-of-speech tagged a total of 45,068 tokens. Using the Penn Treebank and annotated chat data, we achieved part-of-speech tagging accuracy of 90.8%. We also annotated each of the privacy-masked corpuss 10,567 posts with a chat dialog act. Using a neural network with 23 input features, we achieved 83.2% dialog act classification accuracy.

Topic Detection and Extraction in Chat

Internet-based Chat environments such as Internet relay Chat and instant messaging pose a challenge for data mining and information retrieval systems due to the multi-threaded, overlapping nature of the dialog and the nonstandard usage of language. In this paper we present preliminary methods of topic detection and topic thread extraction that augment a typical TF-IDF-based vector space model approach with temporal relationship information between posts of the Chat dialog combined with WordNet hypernym augmentation. We show results that promise better performance than using only a TF-IDF bag-of-words vector space model.

Analysis and Defensive Tools for Social-Engineering Attacks on Computer Systems

The weakest link in an information-security chain is often the user because people can be manipulated. Attacking computer systems with information gained from social interactions is one form of social engineering (K. Mitnick, et al. 2002). It can be much easier to do than targeting the complex technological protections of systems (J. McDermott, Social engineering - the weakest link in information security). In an effort to formalize social engineering for cyberspace, we are building models of trust and attack. Models help in understanding the bewildering number of different tactics that can be employed. Social engineering attacks can be complex with multiple ploys and targets; our models function as subroutines that are called multiple times to accomplish attack goals in a coordinated plan. Models enable us to infer good countermeasures to social engineering

Innovations for Requirements Analysis, From Stakeholders' Needs to Formal Designs

14th MontereyWorkshop 2007 Monterey, CA, USA, September 10-13, 2007 Revised Selected Papers

CORPUS-BASED GESTURE ANALYSIS: AN EXTENSION OF THE FORM DATASET FOR THE AUTOMATIC DETECTION OF PHASES IN A GESTURE

We present the results of using an extension of the FORM gesture dataset to predict the mid-level phenomenon of phase. We compare the results of human phase prediction with automated prediction using machine-learning techniques. Specifically, we present the results of hidden Markov model experiments using an extended version of the FORM data to predict phase labels. Additionally, we compare FORM to the currently most accurate method of data gathering in this field — motion capture — by comparing the predictive accuracy of the physical gesture models produced by FORM and by motion capture for phase labeling.

Innovations in Natural Language Document Processing for Requirements Engineering

This paper evaluates the potential contributions of natural language processing to requirements engineering. We present a selective history of the relationship between requirements engineering (RE) and natural-language processing (NLP), and briefly summarize relevant recent trends in NLP. The paper outlines basic issues in RE and how they relate to interactions between a NLP front end and system-development processes. We suggest some improvements to NLP that may be possible in the context of RE and conclude with an assessment of what should be done to improve likelihood of practical impact in this direction.

Tweet! – And I Can Tell How Many Followers You Have

Follower relations are the new currency in the social web. User-generated content plays an important role for the tie formation process. We report an approach to predict the follower counts of Twitter users by looking at a small amount of their tweets. We also found a pattern of textual features that demonstrates the correlation between Twitter specific communication and the number of followers. Our study is a step forward in understanding relations between social behavior and language in online social networks.

MAJIC: A Java Application for Controlling Multiple, Heterogeneous Robotic Agents

When teaching robotics, we have a number of constraints and desires to satisfy. We are limited by the time available to teach a class, so we need a robotic system that our students can get up to speed on quickly and easily. We are limited by robot availability, in the robots that are on hand, but also because manufacturers of inexpensive teaching robots tend to go bankrupt or change focus quickly, making it difficult to purchase new robots with the same interface as previous models. Thus, we desire an interface easily adaptable to new robots. Finally, we have recently become interested in teaching techniques for dealing with teams of possibly heterogeneous robots. All existing systems that we examined fall short in one or more of these areas, prompting our development of the The multi-agent Java interface controller (MAJIC). MAJIC was designed from the bottom up with modern software engineering principles. The interface is easy to use and learn, can be quickly adapted to new robots, and allows control of multiple robots simultaneously. This paper presents the design of this system, highlighting rapid development and clarity compared with other systems.

Online parameter estimation of a robot’s motion model

Simultaneous localization and mapping (SLAM) algorithms rely heavily on a good motion model to provide critical information about the robots current pose. Most of these algorithms assume that the distribution defining a robots motion will remain stationary over the period of operation, and as such use a fixed model for the duration of a trial. This does not easily allow for changes in the robots motion model due to surface changes, wear and tear, and battery life. Also, if new robots of a similar class are to be used, a new motion model may need to be constructed from scratch. In this paper, we introduce a method that allows the robot to automatically learn its motion model, given a rough estimate of its model or the model from a robot of similar class. We validate our method by demonstrating that it learns a new motion model when a robot crosses a threshold onto a different surface. We also demonstrate our method can estimate the motion model for a new robot given the motion model of a robot of similar class.

PROJECTING AWAY THE CLASS IMBALANCE PROBLEM IN AUTHOR ATTRIBUTION

Author identification algorithms attempt to ascribe document to author, with an eye towards diverse application areas including: forensic evidence, authenticating communications, and intelligence gathering. We view author identification as a single label classification problem, where 2000 authors would imply 2000 possible categories to assign to a post. Experiments with a naive Bayes classifier on a blog author identification task demonstrate a remarkable tendency to over-predict the most prolific authors. Literature search confirms that the class imbalance phenomenon is a challenge for author identification as well as other machine learning tasks. We develop a vector projection method to remove this hazard, and achieve a 63% improvement in accuracy over the baseline on the same task. Our method adds no additional asymptotic computational complexity to naive Bayes, and has no free parameters to set. The projection technique will likely prove useful for other natural language tasks exhibiting class imbalance.