Ransom K. Winder

Methods for approximating value functions for the Dominion card game

AbstractArtificial neural networks have been successfully used to approximate value functions for tasks involving decision making. In domains where decisions require a shift in judgment as the overall state changes, it is hypothesized here that methods utilizing multiple artificial neural networks are likely to provide a benefit as an approximation of a value function over those that employ a single network. The card game Dominion was chosen as the domain to examine this. This paper compares artificial neural networks generated by multiple machine learning methods successfully applied to other games (such as in TD-Gammon) to a genetic algorithm method for generating two neural networks for different phases of the game along with evolving the transition point. The results demonstrate a greater success ratio with the genetic algorithm applied to two neural networks. This suggests that future work examining more complex neural network configurations and richer evolutionary exploration could apply to Dominion as well as other domains necessitating shifts in strategy.

(Un/Semi-)supervised SMS text message SPAM detection

We address the problem of unsupervised and semi-supervised SMS (Short Message Service) text message SPAM detection. We develop a content-based Bayesian classification approach which is a modest extension of the technique discussed by Resnik and Hardisty in 2010. The approach assumes that the bodies of the SMS messages arise from a probabilistic generative model and estimates the model parameters by Gibbs sampling using an unlabeled, or partially labeled, SMS training message corpus. The approach classifies new SMS messages as SPAM or HAM (non-SPAM) by zero-thresholding their logit estimates. We tested the approach on a publicly available SMS corpora collected from the UK. Used in semi-supervised fashion, the approach clearly outperformed a competing algorithm, Semi-Boost. Used in unsupervised fashion, the approach outperformed a fully supervised classifier, an SVM (Support Vector Machine), when the number of training messages used by the SVM was small and performed comparably otherwise. We believe the approach works well and is a useful tool for SMS SPAM detection.

Person Identification from Text and Speech Genre Samples

In this paper, we describe experiments conducted on identifying a person using a novel unique correlated corpus of text and audio samples of the persons communication in six genres. The text samples include essays, emails, blogs, and chat. Audio samples were collected from individual interviews and group discussions and then transcribed to text. For each genre, samples were collected for six topics. We show that we can identify the communicant with an accuracy of 71% for six fold cross validation using an average of 22,000 words per individual across the six genres. For person identification in a particular genre (train on five genres, test on one), an average accuracy of 82% is achieved. For identification from topics (train on five topics, test on one), an average accuracy of 94% is achieved. We also report results on identifying a persons communication in a genre using text genres only as well as audio genres only.

Generating artificial neural networks for value function approximation in a domain requiring a shifting strategy

Artificial neural networks have been successfully used as approximating value functions for tasks involving decision making. In domains where a shift in judgment for decisions is necessary as the overall state changes, it is hypothesized that multiple neural networks are likely be beneficial as an approximation of a value function over those that employ a single network. For our experiments, the card game Dominion was chosen as the domain. This work compares neural networks generated by various machine learning methods successfully applied to other games (such as in TD-Gammon) to a genetic algorithm method for generating two neural networks for different phases of the game along with evolving a transition point. The results demonstrate a greater success ratio with the method hypothesized. This suggests future work examining more complex multiple neural network configurations could apply to this game domain as well as being applicable to other problems.

A Model-Based Analysis of Semiautomated Data Discovery and Entry Using Automated Content-Extraction

Content extraction systems can automatically extract entities and relations from raw text and use the information to populate knowledge bases, potentially eliminating the need for manual data discovery and entry. Unfortunately, content extraction is not sufficiently accurate for end users who require high trust in the information uploaded to their databases, creating a need for human validation and correction of extracted content. In this article the potential influence of content extraction errors on a prototype semiautomated system that will allow a human reviewer to correct and validate extracted information before uploading it was examined, focusing on the identification and correction of precision errors. Content extraction was applied to 6 different corpora, and a Goals, Operators, Methods, and Selection rules Language (GOMSL) model was used to simulate the activities of a human using the prototype system to review extraction results, correct precision errors, ignore spurious instances, and validate information. The simulated task completion rate of the semiautomated system model was compared with that of a second GOMSL model that simulates the steps required for finding and entering information manually. Results quantify the efficiency advantage of the semiautomated workflow—estimated to be roughly 1.5 to 2 times more efficient than a manual workflow—and illustrate the value of employing multidisciplinary quantitative methods to calculate system-level measures of technology utility.

Dropped personal pronoun recovery in Chinese SMS

In written Chinese, personal pronouns are commonly dropped when they can be inferred from context. This practice is particularly common in informal genres like Short Message Service messages sent via cell phones. Restoring dropped personal pronouns can be a useful preprocessing step for information extraction. Dropped personal pronoun recovery can be divided into two subtasks: (1) detecting dropped personal pronoun slots and (2) determining the identity of the pronoun for each slot. We address a simpler version of restoring dropped personal pronouns wherein only the person numbers are identified. After applying a word segmenter, we used a linear-chain conditional random field to predict which words were at the start of an independent clause. Then, using the independent clause start information, as well as lexical and syntactic information, we applied a conditional random field or a maximum-entropy classifier to predict whether a dropped personal pronoun immediately preceded each word and, if so, the person number of the dropped pronoun. We conducted a series of experiments using a manually annotated corpus of Chinese Short Message Service. Our approaches substantially outperformed a rule-based approach based partially on rules developed by Chung and Gildea ( 2010 , Effects of Empty Categories on Machine Translation. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) . Association for Computational Linguistics. pp. 636–45). Our approaches also outperformed (though by a considerably smaller margin) a machine-learning approach based closely on work by Yang, Liu, and Xue in ( 2015 , Recovering Dropped Pronouns from Chinese Text Messages. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics (ACL) . Association for Computational Linguistics. pp. 309–13). Features derived from parsing largely did not help our approaches. We conclude that, given independent clause start information, the parse information we used was largely superfluous for identifying dropped personal pronouns.

MOSAIC: A Cohesive Method for Orchestrating Discrete Analytics in a Distributed Model

Achieving an HLT analytic architecture that supports easy integration of new and legacy analytics is challenging given the independence of analytic development, the diversity of data modeling, and the need to avoid rework. Our solution is to separate input, artifacts, and results from execution by delineating different subcomponents including an inbound gateway, an executive, an analytic layer, an adapter layer, and a data bus. Using this design philosophy, MOSAIC is an architecture of replaceable subcomponents built to support workflows of loosely-coupled analytics bridged by a common data model.

A Semi-automatic System for Knowledge Base Population

The typical method for transferring key information from unstructured text to knowledge bases is laborious manual entry, but automated information extraction is still at unacceptable accuracies to replace it. A viable alternative is a user interface that allows correction and validation of assertions proposed by the automated extractor for entry into the knowledge base. In this paper, we discuss our system for semi-automatic database population and how issues arising in content extraction and knowledge base population are addressed. The major contributions are detailing challenges in building a semi-automated tool, classifying expected extraction errors, identifying the gaps in current extraction technology with regard to databasing, and designing and developing the FEEDE system that supports human correction of automated content extractors in order to speed up data entry into knowledge bases.