Public school teachers in the United States are often constrained in terms of their ability to express their moral views on issues that afect their schools, classrooms, students, and teaching practices, but are able to express their ideas, concerns, and frustrations as private citizens using social media. Previously we developed the Tweet Capture and Clustering System (TCCS) in order to explore how teachers use Twitter, looking at word usage among a group of teacher tweeters, and attempting to find clusters of teachers who have similar patterns of word usage in their tweets. In the work reported here, we look at teacher tweeters across the 12 months of 2016, seeking to understand how the clusters and the words used in these clusters vary from month to month. In this initial look at the dynamics of the system, we see some evidence of word usage changing across the 12-month period. This initial work suggests that extending TCCS to have temporal topic tracing as a core capability will be a meaningful addition to of the system.
• Information systems → Data extraction and integration; • Computing methodologies → Cluster analysis;
The #TeacherTweets project [
In previous work [
In this paper, we explore the utility of an extension to TCCS that would allow us to see how these clusters and signatures change over time. This exploration traces teacher tweeters across the 2016 calendar year, and seeks to understand the extent that word usage within each cluster changes from month to month. We would also like to explore how word usage during each month difers from the overall years usage. Finally, we would like to prepare for the exploration of how clusters change or are static from month to month.
In the remainder of this paper, Section 2 explains relevant background, including an overview of the goals of the larger study of educational philosophy on Twitter, relevant computational approaches, and our previous investigation. Section 3 describes the system design of TCCS. Section 4 explains our research methodology. Section 5 presents the results; Section 6 discusses the implications of the results, and Section 7 describes future work. 2.1
In many cases, U.S. teachers are constrained in their expression of
concerns about their work in classrooms, the well-being of their
students, and education policy. For example, the New York Times
recently describes a New York City public school’s explicit request
that teachers not critique state tests [
Twitter is a popular social media platform that provides methods
to engage, develop communities, and engage in professional
learning [
Twitter provides several mechanisims to create and maintain
dynamic and persistent communities, teacher educators in our
research, including “hashtags”, following, and follow-lists. Topical
communities connect through the use of “hashtags” that link
otherwise independent tweets and ideas together creating a threaded
conversation. Twitter also allows for “retweeting”, repeating
another account’s tweet with attribution, which propagates the
original tweeter’s thought through the re-tweeter’s network of
followers. Twitter users can also “follow” other accounts and develop
lists of accounts they follow for others to use in finding
“thoughtleaders” and to build information sharing networks. These
afordances break down traditional geographical, social, and political
boundaries. They allow teachers to engage other teachers and
others outside their profession about their profession [
The clustering approach used here is AutoClass, developed by NASA
[
The AutoClass algorithm starts by dividing the data at random into the suggested number of clusters. It then repeatedly creates a probabilistic description of each cluster, removes the data from each cluster, and places each data point in the cluster that best describes that data. When the data no longer moves and the cluster descriptions are fixed (or a pre-specified maximum number of iterations has elapsed), the search is completed; AutoClass refers to each completed search as a “try”. Since random numbers are used in the initial cluster definitions, the search is necessarily repeated for a specified number of tries. AutoClass then reports the best clustering found across the set of tries, meaning again, the clusters that have the greatest similarity within each cluster and the greatest diferences across clusters. 1
The output from AutoClass includes a set of cluster descriptions that include the expected means and variance for each attribute as well as the importance or weight of that attribute in defining the cluster. The output from AutoClass also includes a listing for each data point specifying the probability of that data point being in each of one or more clusters. 2.3
In an earlier investigation [
In that work, we found five distinct clusters of teacher tweeters.
Interestingly, the five clusters were quite distinct, without fuzzy
boundaries. That is, each teacher tweeter belonged to a cluster
with 100% probability. Exploration of the intraclass relationships,
by careful examination of word usage and a human sampling of
the tweets in each cluster, allowed us to assign an understandable
definition or label to each cluster with three of them clearly making
moral statements [
These results indicated that educators are using Twitter to talk
about their work and what it means to them. They are making
moral claims about care in the classroom and justice (in at least two
distinct ways). And lastly, they appear to form distinctive moral
communities [
TCCS, as shown in Figure 1, consists of four major components; capture, extraction and translation, clustering, and post processing. Each of these components is distinct from the others in the sense that there are clear inputs and outputs separating it from the modules up and downstream and the module’s purpose.
The capture module is responsible for monitoring Twitter and
capturing tweets from selected accounts. It runs as two processes;
one to collect tweets using Twitter’s streaming interface [
AutoClass C report parsing and extraction extraction and translation clustering post-processing twitter account list
capture word lists one to parse and normalize the raw data for later use (extraction, clustering, etc.). It uses a MySQL database for both intermediate and long-term storage.
The extraction and translation module is responsible for
extracting the parsed tweets collected by the capture module and creating
a “linguistic signature” [
The clustering module is AutoClass, described in Section 2.
Specifically, the AutoClass C implementation [
The post-processing module parses the generated reports and generates comma-separated value (CSV) files that are common, flexible, and importable into a variety of analysis software. During this Twitter stream raw JSON parsed tweets lingustic signatures signatures word counts, clusters Over the entire collection period (calendar year 2016) there were 956,118 tweets among the 494 accounts with a total of 12,437,458 individual tokens (words). The linguistic signature consisted of 123 root words (regular expressions) in the three cagegories of “moral words”, “general words”, and “hashtags” relevant to education and teaching philosophy. Table 2 shows the number of tweets, accounts, and tokens collected for each month of the collection. The number of accounts listed for each month varies as some accounts do not tweet year-round. August
First, we fixed the cluster membership of each individual account to those determined by the full-year run, to investigate how word usage changes over time among each group of users. The five clusters, based on full-year word usage, consisted of 158, 109, 109, 65, and 53 accounts. the 123 words we looked at, sorted according to how often the words were used in the year. While we produced a separate plot for each month, two representative months, August and September, are shown, along with the plot for the full year. Each of the five clusters is represented with a diferent color. In looking at these s t10 n u o C8 d r o6 W g o4 L 2 plots, we get a sense of how the signatures change from month to month. For example, “skills” shows a spike in the red cluster in August, and “protest” and “private.*” show a spike in the blue cluster in September.
#blacklivesmatter
While this form of plotting helps us to understand changes in the cluster signatures over time, it is dificult to see how use of a particular word’s use changes over time. Thus, we have also created plots to trace each individual word over time. Figure 3 shows two words that varied quite a bit over the 12-month period, the hashtag “#blacklivesmatter”, and the root “democra-”.
Figure 4 shows each of the five words most used across the year, and how their usage changes from month to month for each cluster. The five words most influential in distinguishing the clusters turn out to be not heavily used, so the illustrations of these are not interesting and have been omitted due to space concerns. 5.2
To compare monthly data to year data we too two approaches; comparing influential words in creating cluster and examining the Jaccard similarity coeficient of each month’s most frequently used words to the year’s most frequently used words.
To explore how the word use influenced cluster creation each month and over the entire year we examined the words identified by AutoClass as influential in determining cluster membership. For each month of the collection period, the clustering module created a new arrangement of accounts (clusters) and word frequency data about how those clusters were determined.
The influential words were determined by AutoClass in its repeated trial process. They are the words of which their use, or lack of use, were highly influential in determining which cluster an account was placed. While these words were important to understand how a cluster is created, they were not useful in examining how word use or the clusters themselves change over time.
We also utlized the Jaccard similarity coeficient [
May and August stand out as the only months the word “we” is not the most-used word. In May it is second, and in August, third. “we”, “teach.*”, and “school.*” are always the top three frequently used words in every month, which is expected as the accounts captured are educators and partially validates our account selections. The word “we” was considered a moral word, indicating belonging, and confirms the premise that teachers are using Twitter to make moral statements. The appearance of other moral words (“just.*”, “love”, “need”, “know”) further validate this conclusion. teach.* 0 0 0 0 0
Figure 3 shows two specific words that varied quite a bit over the 12-month period, the hashtag “#blacklivesmatter”, and the root “de-mocra.*”. The surge in July in some clusters may reflect reported police killings of African Americans that month and subsequent increases in protests (not a topic for all the teacher tweeters); the surge of “democra.*” towards the end of the year might be a reflection of the U.S. Presidential election.
To understand how the clusters change from month to month, we looked at the influential words used by AutoClass to determine monthly and yearly cluster membership. There was little correlation between the words that influence the creation of the year clusters and the monthly clusters as indicated by the set similarity coeficient and by visual inspection. While these clusterings and inlfuential words were useful for creating clusters and seeing how the monthly clustering changes, they were not useful in understanding how word usage changed across time. That is, they were useful in creating clusters but not in understanding them.
There are several areas where TCCS and this research could be continued. Herein we used pre-selected words of interest, we focused on teachers on Twitter, and we have not traced topics through clusters over time.
Our research used a set of pre-selected terms related to education and statements of morality by educators. TCCS supports alternate word lists and the use of the most frequently used words for building the linguistic signatures. This approach could provide more insight into how the words and language change over time and potentially predict emerging clusters of accounts.
There is no requirement Twitter with TCCS or educator accounts. TCCS supports any form of text input from similar social media or other platforms. While Twitter is heavily used by teachers, we may ifnd in the future that these conversations move to other platforms. TCCS is easily adaptable to accommodate these new technologies.
We fixed the clustering module (AutoClass) to five (5) clusters and 50,000 trials to determine the clustering. These were selected based on empirical study and experience. AutoClass allows for determining the “best” number of clusters automatically. Given the resource requirements of 50,000 trials and an unlimited number of cluster possibilities, we did not attempt this configuration and leave it as a future research question.
There are other methods of looking at word usage and clusters such as Dynamic Topic Modeling and Latent Dirichlet Allocation (LDA) that we did not use in this work. These techniques could be fruitful in validating the manual topic assignment by visual inspection or be useful in other ways. Given our larger research question was about moral statements by educators, we leave these approaches to possible future work.
Lastly, in this research we made no attempt to map the themes of clusters through time. That is we did not classify the language used within a cluster during one month and attempt to find the same thematic language in a cluster in the next month. The data collected and processed via TCCS is clearly useful for this thematic mapping and tracking, we pose here that this is possible, but not yet done. This extension of TCCS, full temporal topic tracing, would add meaningful capabilities for the broader research questions about how the conversation changes over time.