Language and Cognition Research

My Research in Cognitive Science

Contact: David Reitter, Google.

Among man's most fascinating abilities are the ones to verbalize, communicate and adopt ideas within a vast network of social contacts. Human cognitive capabilities are uniquely suited to communication, and they are crucial to the intelligence emerging from human communities. The cognitive and psycholinguistic mechanisms underlying language comprehension and production are still poorly understood. While recent studies paint a picture of how memory and contextualization help humans comprehend a dialogue partner's ideas and individual language, we do not understand whether human memory has evolved to support team-work and social cognition.

Cognitive modeling and data-driven machine-learning models using deep neural networks have recently sparked interest in understanding how cognition arises from a data-rich environment. I have studied micro-processes at the cognitive level, such as adaptation and information distribution in dialogue. More recently, however, I have been working on understanding how we can learn to use language and common sense using rich, multimodal data. : Individuals adapt their linguistic expressions quickly to their conversation partners, and new communicative conventions may soon spread through a network of connected agents. Cognitive modeling frameworks, validated and refined through careful experimentation, as well as computational tools can now simulate the co-dependency of individual cognition and emergent phenomena in human societies. Networked experimentation platforms facilitate large-scale data-collection. Language resources (corpora) provide data collected in real-life situations that let us test cognitive and psycholinguistic models. Once validated, they will make better predictions and cover broad ranges of human behavior. This combination of broad coverage and large-scale simulation requires new computational tools, new methodologies, new datasets and new experimental designs. We expect that these will lead to advances in the quest for a standard architecture of the human language faculty.

Recent work covers several areas.

1. Dialogue and dialogue systems. Using corpus data, statistical analysis and machine learning, we show how speakers align and increase their task success based on mutual adaptation; cognitive architectures show how the basic underlying process (syntactic priming) can be framed as a memory retrieval effect (with JD Moore and F Keller, e.g., 2006, 2007, 2008, 2011 J.CogSci). More about my work on adaptation in dialogue.

2. Communication of small teams and larger communities. Adaptation between speakers or players of a naming game can be shown to lead to emerging, wide-spread communication systems. Multi-agent cognitive models (using the ACT-R architecture) explain empirical results and make predictions for large groups. In ongoing work, convergence and the effect of communication policy is observed empirically using a multi-player game platform (with C Lebiere and K Sycara, et al., e.g., 2009, 2010, 2011, 2011 J.CSR). More about my work on social cognition.

3. Scalability of cognitive modeling. Cognitive models written in architectures such as ACT-R and SOAR tend to be overly specific compared to the evaluation they receive using reaction times, learning effects and, more recently, neurophysiological data. In a reformulation of ACT-R theory, ACT-UP proposes to underspecify portions of a subject's task strategy. At the same time, the ACT-UP library allows modelers to scale up simulations to thousands of interacting models. ACT-UP is well-validated, well-documented and available for download. (2010 ACT-R, 2010 JAGI)

4. Crowdsourcing for Peace: 
Tracking Worldwide Militarized Conflicts Obtaining accurate and sufficiently detailed, large-scale data on conflicts is vitally important to international political engagement. These data are costly and difficult to obtain. Penn State has, in the past, curated a database of militarized incidents between nation-states; in light of decreased federal funding to political sciences, this database can no longer be maintained by hand. Dr. Reitter has been working to develop an economical way to research international conflicts at a large scale from newswire articles. The dataset we produce contains information about thousands of instances of military action, such as the identification of nations that initiate or are the recipients of aggressive action as well as modes of attack. The crowd-sourcing approach means that the chore of annotating these articles is delegated to remote, anonymous workers who receive limited training. Reitter’s lab has developed machine-learning techniques to combine and error-correct their work. As a result, we obtain highly accurate data for the benefit of peace researchers and policy makers. We now also integrate automated natural-language processing technology to achieve higher accuracy at lower cost. This new, interdisciplinary crowd-sourcing mechanism is intended to pave the way for other applications inside and outside the social science realm. With this approach, we leverage knowledge regarding the differences in how people go about analyzing newswire text. Given this knowledge, we engineer data-driven methods to aggregate their analyses to uncover new global knowledge and exploit patterns within the news text to makes predictions of future interstate events. (with A. Ororbia and Y. Xu, M. Schmierbach, G. Palmer, V. D’Orazio)

5. Games of timing reflect dynamic decision-making under uncertainty, as it applies in many real-world situations, including medical care, safety and security. Rather than making discrete decisions, participants choose one or more points in time that determine the outcome. We study individual's biases and characteristics in games of timing. We examine risk propensity as a personal preference affecting timing decisions and document a bias, \emph{patience}. (with J. Grossklags and M. Ghafurian)

See Publications.