GDELT Summary: Television Comparer

Television Comparer

This service allows you to rapidly compare up four broadcast television news searches, combining them on the same timeline, displaying one as a percent of the other or identifying differences in the language they use. This makes it easy to rapidly compare searches, for example to see how East and West coast television stations differ in their coverage of an event or to see what percent of coverage of Clinton focuses on her emails, etc.

Step 1: Dataset

GDELT Summary supports searching either global online news coverage monitored by GDELT and its global partners or broadcast television news coverage monitored by the Internet Archive. You can use the dropdown below to switch to global online news coverage.

Dataset

Step 2: Output Type

GDELT Summary supports either a basic visual dashboard that summarizes coverage matching your search, or a comparison visualization that compares up to four searches to make it easier to examine differences in how various topics are being covered in the media.

Output Type

Step 3: Enter Search

Enter your search keywords below. You are searching the complete raw closed captioning stream of all news programming monitored by the Internet Archive's Television Archive. Closed captioning is searched as-is as provided by each station, meaning will always be some small amount of captioning error. The Archive has monitored more than 150 stations since 2009, but has not monitored all stations for the full time period. You should carefully check the start/stop monitoring dates for your stations of interest to ensure they were monitored during the full time period of interest. Advanced users with specific needs to precisely identify any monitoring disruptions should look for the normalization baseline link on the results page above the volume timeline display or can also download the inventory files for the given days. Only programming determined by the Internet Archive to be primarily "news" in nature is monitored - all other shows, such as purely comedy or entertainment programming are excluded.

Previous versions of the Television Explorer divided broadcasts into discrete sentences and reported the number of matching sentences, but the use of sentences was difficult for many users to conceptualize and integrate with other time-based measures. Thus, the new Television Explorer now uses airtime. Each broadcast is divided into a series of sequential 15 second clips and we display the percent of 15 second clips that matched your search. Thus, a 30 minute broadcast will be divided into 120 separate 15 second clips. Note that most commercials are not closed captioned and thus are excluded from searches and so most stations will have less than 5,760 clips per day. Searches for phrases that span the boundary between two clips are counted for the first clip, ensuring there is no double counting.

All words/phrases should be in the language of the station(s) you are searching (for example, to search Univision stations you should provide Spanish keywords/phrases). This means you must conduct separate searches if you wish to search across stations in different languages. Phrases should be enclosed in quote marks and are limited to a maximum of five words. If you include multiple words/phrases all of them will be required to appear somewhere in the 15 second clip for it to match. Only the exact keyword(s) entered are searched (searching for "russia" does NOT match "russian" or "russians"). You can perform limited boolean "OR" searches by enclosing a group of OR'd terms inside a set of parantheses - for example, to search for "syria" appearing near a set of russian-related terms you might search for "syria (russia OR russians OR russian OR kremlin OR putin)". You can also put a "-" in front of a given word or phrase to exclude articles that contain it.

You can also specify one or more keywords/phrases that must appear near your main keywords/phrases, but at a greater distance than an ordinary keyword. For example, you might want to see how often "emails" appeared near "clinton" on each station. If you search for "clinton email" you will get all clips that mention both terms in the same 15 second clip. This will miss cases where "clinton" appeared in one clip and "emails" appeared in the following clip. To allow for these kinds of "context" searches, you can add context:"yourkeyword" to search for a given keyword either in the same clip as your main keyword(s) or in the 15 second clips immediately before and after the matching clip. Thus, a search for clinton (context:"email" OR context:"emails" OR context:"server") would return all clips that contained "clinton" and which also contained either "email", "emails", or "server" either in the same clip or the immediately preceeding or following clip.

Keyword(s) #1

Stations #1

Keyword(s) #2

Stations #2

Keyword(s) #3

Stations #3

Keyword(s) #4

Stations #4

TIME PERIOD: By default June 2009 to present is searched. You can narrow the timeframe to a shorter window to examine a particular time period. (You can also select a "zoomable" volume timeline display in the next section that allows you to interactively narrow your search timeframe.)

Time Period

Start Date

End Date

COMBINE/SEP: By default each station's results are reported separately in the resulting displays. You can combine the output of all of the stations into a single result to make it easier to examine the aggregate coverage of a set of networks, especially for comparison purposes.

Combine/Separate

NORM/RAW: By default, timeline and barchart results are reported as a percent of the total airtime monitored from each station (the number of matching 15 second blocks divided by the total number of monitored 15 second blocks), rather than as a raw count of 15 second blocks. This is because different stations may have different amounts of news programming (strictly entertainment shows are not monitored) and advertising (some stations may have no advertisements while others may devote a substantial amount of airtime to ads), while for some stations the Internet Archive only monitors selections of their programming. There can also be sporadic technical glitches that might cause a brief outage in monitoring. Using normalized counts ensures that the results you see are less impacted by these issues, but for some specific situations you may need to view the raw number of matching 15s blocks. Note that when using raw counts you might want to click on the "Export" button at the top right of the timeline graph and choose "View Normalization Graph" to ensure there were no outages or other issues in the stations in question during the time period of interest.

Normalized/Raw

DATE RESOLUTION: By default, timespans of less than 7 days will be displayed at hourly resolution, those under 3 years will be shown at daily resolution and spans longer than 3 years will be shown at monthly resolution. Hourly resolution is only available for periods of 7 days or less and will generate an error for longer timespans. Otherwise, you can manually select the desired date resolution below. This is especially helpful for rare search terms or those with high burstiness, as trends are often more visible at weekly, monthly and yearly aggregation levels. Note that the start and end date/times of your query are automatically adjusted for each resolution. This means that if you select Weekly resolution below, your start and stop dates are adjusted forward/backward as needed to encompass complete weeks.

Date Resolution

TIMEZONE: By default, all dates/times are expressed in UTC. Sometimes it is necessary to convert to a specific timezone to localize results, especially for precision analysis of sudden onset breaking news events or when aligning news coverage with external timelines of the events they describe. Note that changing the timezone below will change both how the start/stop date/times above are interpreted and how all results are returned.

Timezone

SMOOTHING: By default, timeline visualizations will report exact values, but this can lead to a noisy graph that makes macro-level patterns more difficult to discern. To address this, you can enable smoothing that computes a moving window average to smooth the results and make patterns more apparent. Remember that moving window averages slide trends to the right and under heavy smoothing peaks can appear days to weeks later on the timeline.

Smoothing

Step 4: Displays

Your Television Comparer dashboard is built up by appending together a series of displays and content sections. Use the dropdowns below to select which you'd like to include in your TV Comparer summary.

TIMELINE MERGE: Merges the volume timelines of up to four separate searches and displays them on the same combined results timeline, making it easy to compare coverage across topics.

Timeline Merge

TIMELINE PERCENT: The results timeline for the first keyword is divided by the results timeline for the second keyword, resulting in a percentage/density timeline. Ie, if you choose "clinton context:"email"" as Keywords #1 and "clinton" as Keywords #2, you will get a timeline showing the percent of all clips mentioning Clinton each day that also mentioned her emails.

Timeline Percent

WORDCLOUD RAW: This computes a tabular wordcloud of the top terms appearing in the top 500 most relevant clips for each search and displays the percent of clips that contain each word.

WordCloud Raw

WORDCLOUD TF-IDF: This computes a tabular wordcloud of the top terms appearing in the top 500 most relevant clips for each search and displays the words with the highest modified TF-IDF scores. In essence, instead of displaying the most common terms, this wordcloud decreases the score of words that appear frequently across each of your searches, surfacing the terms that are unique to each search. Thus, if you search for "clinton" and "clinton emails", the word "clinton" will be scored lower in this wordcloud, since it appears frequently in both searches and thus is not unique to either. Note that this is a modified TF-IDF score that uses the returned percentages instead of actual population counts.

WordCloud TF-IDF