Kathleen M. Carley

Increasingly sources data is available in electronic text form such as email, blogs, news-articles, and web page. For scientific usage, this data has to be converted to a form that can be statistically analyzed. This can be an arduous manual procedure. We present here a semi-automated approach that reduces coding time and enables the extraction from texts of a) ontologically classified networks, b) node attributes, and c) meta-data. We show that this approach makes it possible to combine network analysis and standard statistical analysis in reasoning about the material in the text and to reason both about content and the environment that produced the text. We illustrate this approach using data from two corpi – Enron email and data on political elite. Two software tools are used – AutoMap and ORA.

Previous research suggests that during an organizational crisis some social networks increase in heterogeneity, while the variability or entropy of the actual communication decreases. We investigate what changes in communicative behavior might contribute to this decrease of entropy by analyzing two dimensions of change:  First, we hypothesize that during a crisis people are trying to finger point to sources of rumors and failure, to attribute accountability for the crisis to others, and to establish the belief that responsibility is not allocated with themselves. One indicator for such behavior is the increased usage of anaphora; pronouns that refer back to another social entity. Second, we hypothesize that the intra-organizational discourse drifts towards polarized opinions, and that people try to establish an identity by contrasting points of view, taking a stand on their opinions, and identifying themselves with a particular group view. Indicators for this communication style are the increased occurrence of antonyms within and across messages, an increased co-occurrence of concepts representing particular views and the holders of these views, and an increased use of anaphora referring to groups; e.g. we, they, us. Our methodological contributions are anaphora resolution, which identifies who refers to whom without repeating names, and antonym recognition; both supported by AutoMap. We use anaphora resolution to update the weights of nodes and edges of the underlying graph, and analyze the impact of this enhancement on network analytic measures computed on valued graphs in ORA. We test both hypotheses by extracting meta-matrix networks from the bodies and headers of the Enron email corpus with AutoMap and then processing those data in ORA.

Organizations constantly produce and consume organizational language, and these texts and documents are a primary way that organizations interact with their environment. Each text that an organization produces can be represented as a network of linked concepts. The network analysis technique called ‘map analysis’ allows us to systematically extract and represent the network in a text, and compare networks across texts. When two concepts are linked in a set percentage of the texts in a dataset (i.e. half), those ties are part of that dataset’s ‘central graph’. When there is high consensus between organizations about what a particular type of text should say and how it should say it, the central graph will be relatively dense, but when there is low consensus, the central graph will be relatively sparse. Using three types of organizational language – annual reports, privacy policies, and mission statements – from two types of organizations – universities and corporations – we compare the concepts networks and central graphs within and across datasets. We argue that there will be denser central graphs in a type of text primarily intended to display responsiveness to the environment, and sparser central graphs when the text is primarily a vehicle for asserting a distinct and creative identity.

Map analysis, a type of network analysis, is a technique for systematically extracting, representing, and comparing the networks of ties between concepts in a set of texts. The network of ties is the text’s “map.” Managing issues of self-presentation is a central goal of many different types of texts, and in this paper we present map analysis results that capture the self-presentation strategies authors use in a specific set of texts. The texts we study are a set of applications on behalf of entrepreneurs for an “Entrepreneur of the Year” award. Our research focuses on both interpreting self-presentation strategies from these networks, and comparing the networks created by different coding and data reduction decisions. We use an automated text analysis program (AutoMap©) to extract the concepts in the text, link them into statements based on their proximity in the text, and then into networks of statements within the entire text. The author’s specific strategic intent in the text is reflected in different statements formed from the concepts in the text and the arrangement of those statements. The structures of texts’ concept networks leads us to extract four general self-presentation strategies: the prepared entrepreneur, the driven entrepreneur, the creative niche entrepreneur, and the humble entrepreneur (a single entrepreneur may employ multiple strategies).

Software demonstration of AutoMap, a tool that facilitates the identification and extraction of relational data from unstructured natural language text data. This package has been applied by our lab and others to harvest text collections ranging from a few documents to large amounts of files and data from a variety of topics, domains and languages for network data. We give an overview of the Natural Language Processing and Information Extraction routines implemented in AutoMap, such as Stemming, Parts of Speech Tagging, Anaphora Resolution, Feature Selection, Named Entity Extraction and various filtering techniques. One of the challenges in extracting relational data from texts is locating and classifying relevant instances of node classes according to ontologies or taxonomies that vary across domains and projects. We briefly present Entity Extraction - a methodology based on Conditional Random Fields - that we developed and integrated into AutoMap in order to tackle this problem. We demonstrate how AutoMap can be used to distill, fuse, analyze and represent information provided in headers and bodies of emails from a network analytic point of view. Finally, we show how the relational output from AutoMap can be used as input to other standard packages for analyzing, modeling, visualizing and simulating relational data, and provide a few empirical examples for that.

Once network data on a group is collected, we often ask - what is the structure? That is, what are the subgroups and how do they interact? Within SNA we have a number of partitioning and clustering algorithms for locating groups. A key limitation of these approaches is that they partition the group into distinct nonoverlapping sets. A second limitation is that they do not take non-social network data into account when locating the groups. A third limitation is that, they often only pull out the surface structure of the group - that which is common to all - and don't provide guidance as to the deep structure. In this paper, using data on a large university department, these limitations are illustrated. Then we demonstrate how these limitations can be, at least partially overcome, using data on multiple networks - social and knowledge; segregating overly shared information, and using fuzzy set partitioning.

Network Text Analysis supports analysts in detecting the organizational structure of covert networks from textual data. We have formalized an approach for Network Text Analysis and implemented it into a software package referred to as AutoMap. We will report on the extraction of the organizational structure of three covert networks, which are Hamas, Al-Qaeda and Jamaah Islamiyah, with AutoMap. For each of the three groups we have one corpus with about 550 texts that were collected from a variety of sources such as LexisNexis, trial transcripts and research papers. The network data that we extracted from the corpora is multi-mode, multi-link, and multi-time period, and has attributes of nodes and edges. We will present results of the network analysis of the extracted data such as the identification of critical individuals in the networks and their linkage to knowledge, resources and other organizations, and compare the revealed structures in order to identify idiosyncrasies of each group. The network analysis was performed with ORA, a statistical toolkit for network analysis.

This report describes a study of human and organizational risk within NASA's Team X, a conceptual mission design team. A grounded theory approach was used to develop computational models for risk analysis. Among the major findings in the analysis were identification of critical personnel, risk of turnover and performance tradeoff of differing leadership styles.

Networks, and the nodes within them, are often characterized using a series of metrics. Illustrative graph level metrics are the characteristic path length and the clustering co-efficient. Illustrative node level metrics are degree centrality, betweenness centrality, closeness centrality, and eigenvector centrality. A key issue in using these metrics is how to interpret the values; e.g., is a degree centrality of .2 high? With normalized values, we now that these metrics go between 0 and 1, and while 0 is low and 1 is high, we don't have much other interpretive information. Here we ask, are these values different than what we would expect in a random graph. We report the distributions of these metrics against the behavior of random graphs and we present the 95\% most probable range for each of these metrics. We find that a normal distribution well approximating most metrics, for large slightly dense networks, and that the ranges are centered at the expected mean and the endpoints are two (sample) standard deviations apart from the center.

Abstract: ORA is a network analysis tool that detects risks or vulnerabilities of an organization's design structure. The design structure of an organization is the relationship among its personnel, knowledge, resources, and tasks entities. These entities and relationships are represented by the Meta-Matrix. Measures that take as input a Meta-Matrix are used to analyze the structural properties of an organization for potential risk. ORA contains over 100 measures which are categorized by which type of risk they detect. ...

This technical report provides users and researchers information on the configuration and use of the newest version of Construct, the CASOS dynamic network, agent-based, information and belief diffusion simulation of complex socio-technical systems. The report provides a Quick Start Guide to Construct, a detailed discussion of its configuration, and use through a sample problem and virtual experiment configuration exemplar, and a set of appendices with additional useful information. This document is both an introduction to Construct for casual modelers as well as a reference guide for researchers, modelers, and simulationists.

ORA is a network analysis tool that detects risks or vulnerabilities of an organization's design structure. The design structure of an organization is the relationship among its personnel, knowledge, resources, and tasks entities. These entities and relationships are represented by the Meta-Matrix. Measures that take as input a Meta-Matrix are used to analyze the structural properties of an organization for potential risk. ORA contains over 100 measures which are categorized by which type of risk they detect. Measures are also organized by input requirements and by output. ORA generates formatted reports viewable on screen or in log files, and reads and writes networks in multiple data formats to be interoperable with existing network analysis packages. In addition, it has tools for graphically visualizing Meta-Matrix data and for optimizing a network s design structure. ORA uses a Java interface for ease of use, and a C++ computational backend.

The reality of life is embedded in social networks.  At present, most epidemiological models do not consider the heterogeneity of social networks when predicting disease outbreaks.  One of the challenges in modeling natural and man-made epidemics is understanding how social networks affect disease propagation and how the consequences of disease changes social networks. We describe a simulation system called BIOWAR that uses cognitively realistic agents embedded in social, knowledge and work networks to describe how people interacting in these networks acquire disease, manifest symptoms, seek information and treatment, and recover from illness.  Using a model of disease and symptoms, agents who come in contact with infectious agents through their social and work networks become ill.  These illnesses alter their behavior, changing both the propagation of the disease and the manifestation of the disease on the population. Presently, we have completed a number of simulations that examine the effect of contagious and noncontagious illnesses in high-alert (agents have knowledge of a potential disease outbreak) or low-alert states.  Agents who believe they may be ill and have knowledge of a potential outbreak are more likely to seek care than those who do not.  We have compared results of low-alert states to known influenza epidemics and to data containing emergency room visits, pharmacy purchases and absenteeism.  Although the peak incidence of the simulated outbreak is larger than the peak incidence seen in the population data, the simulation results are temporally similar to those seen in the population data.  Further work to increase the fidelity of both the simulation and the population data is ongoing.  It is hoped that this simulation framework will allow us to ask “what-if” questions regarding appropriate response and detection strategies for both natural and manmade epidemics.

Previous research suggests that the patterns of intra-organizational communication change during crises. Additionally, network-analytic studies indicate that during organizational crises interpersonal communication becomes intensified, diversified, and tends to by-pass formal chains or hierarchies of communication more strongly. However, the connection between the semantics and the morphology of communication networks from organizations in crises is not well understood yet. In our project we investigate this possible relationship by studying e-mail networks. The data set we use is the Enron email corpus. Our research is based on the assumption that communication networks are the place where organizational culture and identity are created through discourse and the circulation of stories. We furthermore assume that the semantic and structural mechanism of this process change during crises. More precisely, for the times of crises we hypothesize that a) The network segmentation and cohesion of network clusters increase, because people engage in strategic alliances and small groups with trusted others. b) The interpersonal usage of antonyms increases, because antonyms are one way or indicator for establishing and distinguishing identity. c) The semantic entropy of communication networks decreases, because the discourse drifts towards polarized ends of themes and issues. In our presentation we report on our findings with respect to these hypotheses.

Enron's recent bankruptcy and the legal fall-out highlight the importance of understanding information transfer among collaborating individuals. Our research is driven by the attempt to establish evidence of knowledge sharing that is independent of explicit records of interaction. In our research, we compare two different strategies for identifying clusters within communication networks among individuals from the Enron email corpus. Our study is based on the assumption that evidence for the existence of clusters can be retrieved from email data. The first strategy relies on explicit relational information ('to' and 'from' ties in the email headers) and traditional social network clustering tactics. We propose and empirically test a second strategy that applies network text analysis to the bodies of the emails in order to identify groups of people (knowledge clusters) who share a 'consensus' measure with respect to their topical mental models. We will report on the clusters identified with both strategies, and the congruence between them.

Extracting and representing the networks of ties between concepts in a set of texts creates a “map” of each text. Using map analysis, a researcher systematically reduces the content of texts, then extracts and compares the networks of ties between concepts. In this paper we will present map analysis results that attempt to capture the self-presentation strategies authors use in their texts. (Managing issues of self-presentation is a central goal of many different types of texts.) Our research focuses on the implications that different coding and data reduction techniques have for interpreting map analysis networks. We use an automated text analysis program (AutoMap) to extract the concepts in the text, link them into statements based on their proximity in the text, and then into networks of statements within the entire text. The texts we study are a set of applications on behalf of entrepreneurs for an “Entrepreneur of the Year” award. The authors use a finite set of concepts in their texts, but arrange them in different combinations depending on the specific strategic intent of the text. Applicants value uniqueness in their application’s content because it sets them apart and demonstrates their worthiness for the award, but the value placed on uniqueness in the structure of their strategic accounts is not as clear. We found that using even a minimally rhetorically informed rule to form statements improves the interpretability of concept networks by eliminating redundancy and creating networks that reflect strong ties between concepts.

When text data pertaining to socio-technical networks are available, these texts are often either analyzed separately from the network data, or are reduced to the fact and frequency of the flow of data or objects between nodes. Examples for the joint availability of text data and network data include answers to open questions in classical network surveys, social media such as emails, blogs, and wikis, and the semantic web. Previous research on the relationship between language and networks suggests an impact of the position of individuals in the network on their motivation and ability to induce innovation and change in socio-technical networks. We present our findings from a study in which we empirically tested this relationship for the case of research proposal that were granted funding by the European Union under the Framework Programmes and a methodology that we developed in order to facilitate this type of studies. This methodology computationally integrates network theory and topic modeling, an unsupervised machine learning technique that reduces the dimensionality of text data to sets of semantically related words, such that network data are enriched through information from text data and vice versa. Our approach is based on prior work that assumes not only texts, but also authors and other types of entities and metadata to have probability distributions over topics (Mimno \& McCallum 2008). We extend this notion by abstracting away from the level of individual authors and collaborators to the structural role level, where the actual role is defined by network theory.

As people adopt life-long learning as a strategy to succeed in modern societies and as traditional forms of face-to-face modes of instruction are supplemented by e-learning opportunities, traditional roles of learners and educators may change. The identification of actor roles and their embeddedness in social systems have a long tradition in social sciences. In education science, learning has traditionally been conceptualized as an adaptive knowledge construction process. This view has started to be extended by also taking the network structure and dynamics of interacting groups of learners and educators into account. We present a case study in which we leverage social network analysis in combination with relational text analysis to investigate emerging roles of actors within the social network of a remote learning community. We analyze the communication infrastructure of tutors and learners in web based learning to find generalizable learning roles. The data comes from e-learning forums that are actively used by eleven universities located in the state of Saxony, Germany. We use the relational text analysis tool AutoMap to examine the flow of information through the network of learners and educators and to represent these data as semantic networks. The semantic networks are then combined with social network data that denote the collaboration between individuals. By performing structural analysis on these rich relational data we identify roles of actors in the given e-learning environment as well as the relationship between network structure and learning processes. With this research we ultimately aim to contribute to a better understanding of the relationship between theories about socio-technical networks, communications, and learning in humans.

Computational analysis is dramatically re-shaping the way we think and reason about society and social processes. Everything from the impact of information technology to the fundamentals of cooperation and altruism are being addressed using computational models. Computational models, often in the form of virtual worlds, are used in social, technological and engineering policy domains to address via what-if analysis, how different technologies, decisions and organizational and government policies influence the performance, effectiveness, flexibility, adaptiveness and survivability of complex social and organizational systems. Computational models are being increasingly used in the classroom to demonstrate social processes and the impact of change to undergraduate and graduate students. New programs are rapidly springing up in which computational modeling and analysis plays a role. ... The relation of computational models to reality is complex. Underlying all the diverse ways in which data can be linked to models is a fundamental tension – accuracy versus simplicity. In this paper, this tension and how it plays out in the computational social and organizational sciences is discussed. Findings from behavioral and cognitive psychology are used to explain the basic way people respond to computational models. On the one hand, there is a belief in simplicity. The basic argument is that, if they are to be explanatory, models should be a reduction of reality. Apply Ocham’s razor and find the simplest explanation. On the other had, there is a belief in accuracy. The basic argument is that, if they are to be accurate, models should provide a match to the real world at a sufficient enough detail for the problem at hand. Apply validations tests and find the satisfactory explanation that enables you to make decisions, set policies, etc., with minimal risk. Immediately, it should be obvious to the reader that the problem is a socio-psychological one – that is simple and satisfactory are in the eye of the beholder. This tension is often played out in terms of arguments over transparency and veridicality.

Texts can be coded and analyzed as networks of concepts often referred to as maps or semantic networks. In such networks, for many texts there are elements of social structure – the connections among people, organizations, and events. Within organizational and social network theory an approach called the meta-matrix is used to describe social structure in terms of the network of connections among people, organizations, knowledge, resources, and tasks. We propose a combined approach using the meta-matrix model, as an ontology, to lend a second level of organization to the networks of concepts recovered from texts. We have formalized and operationalized this approach in an automated tool for text analysis. We demonstrate how this approach enables not only meaning but also social structure to be revealed through text analysis. We illustrate this approach by showing how it can be used to discover the social structure of covert networks – the terrorist groups operating in the West Bank.