Body of Secrets: Anatomy of the Ultra-Secret National Security Agency (90 page)

Other
companies taking part in the program included TRW, SAIC, and Lockheed Martin.
The money to finance the Soft Landing contracts comes from funds the agency
saves by retiring senior employees early. By 1998, after two years in
operation, the program had found homes for more than 300 retirees at eight
contractors, saving NSA $25 million along the way.

Born in
the foothills of the Blue Ridge Mountains in Greenville, South Carolina, during
the middle of World War II, Mike McConnell graduated from the local college,
Furman University, with a degree in economics in 1966. Shortly afterward he
joined the Navy and was shipped off to Vietnam as a damage control officer on
the USS
Colleton,
a ship attached to the Mobile Riverine Force in the
Mekong Delta. Having survived the conflict, he went on to counterintelligence
work with the Naval Investigative Service in Yokosuka, Japan, took a liking to
the spy world, attended the Defense Intelligence College, and became an
intelligence specialist.

Assigned
as the operations officer for the Fleet Ocean Surveillance Information Facility
in Rota, Spain, in 1976, McConnell received his initiation into the world of
signals intelligence. "Four Navy chiefs and one NSA civilian took me under
their wing to teach me Sigint," he recalled. "I learned as a young
Navy lieutenant that Sigint is hard; it is complex, esoteric, and difficult to
understand over its depth and breadth. ... It changed my understanding, respect
for, and use of Sigint for the rest of my professional life."

Following
other assignments, including a tour as force intelligence officer aboard the
USS
La Salle
in the Persian Gulf and Indian Ocean, McConnell moved to
NSA, where he headed up the Naval Forces Division. Then he went to Pearl Harbor
as the top naval intelligence officer for the Pacific Fleet, a job that won him
his first star. He earned a second while dealing with such issues as the fall
of the Soviet Union and the war in the Persian Gulf as a key intelligence
staffer to the Chairman of the Joint Chiefs of Staff.

At NSA,
McConnell soon found that it was far easier to eavesdrop than to convert
intercepts into finished, usable intelligence. As always,
codebreaking—"processing"—was the hardest part. "I have three major
problems," McConnell was often heard declaring, "processing, processing,
and processing."

Translation
was also a major problem. "There now exists a world full of 'Navajo Code
Talkers,' in a certain sense," noted McConnell. He was referring to the
Native Americans who during World War II were employed to securely communicate
sensitive messages because their language was unwritten, almost unknown outside
their community, and thus almost impossible for an enemy to translate.
"With the rich diversity of potential intelligence targets owing to
possible U.S. involvement in low-intensity conflict and regional crisis
situations anywhere U.S. interests may be threatened," McConnell
continued, "we are confronted by a linguistic challenge of staggering
proportion."

Down on
the working level, the reductions and changes forced many managers to dig out
their old earphones and go back to being operators. Similarly, those with
language skills now in excess, such as Russian linguists, had to retrain in
another language or develop new skills entirely.

While the
end of the Cold War brought a greater sense of tranquility to most parts of the
country, it created a seismic shift at NSA. Gone were the old traditional
targets, the Soviet Union and Eastern Europe. Taking their place were new
trouble spots that seemed to spring up almost anywhere. In 1980, fully 58
percent of the intelligence community's budget was targeted against the Soviet
Union. Three years later NSA, desperate for Russian linguists, asked fifteen
colleges, including Penn State and Georgetown University, to participate with
the agency in a secrecy-shrouded Russian language internship program.

But by
1993 only 13 percent of the intelligence budget was aimed at Russia, and
Russian linguists were scrambling to find new vocabularies to master. Suddenly
the buzz phrase was "exotic languages."

Exotic
languages have long been NSA's Achilles' heel. In 1985, for example, Libyan
diplomatic messages were intercepted discussing the planning of the terrorist
attack at La Belle discotheque in West Berlin. However, according to
intelligence experts, a shortage of Berber translators led to a critical delay
of several days in reading the dispatches. By then, the deadly bombing had
already taken place.

In 1986,
Bobby Inman had warned a congressional committee that "steadily
deteriorating language training capabilities" presented "a major
hazard to our national security." The message was underscored by the
Pentagon's director of intelligence personnel and training, Craig L. Wilson,
who spoke of the "dismal ignorance," in the Defense Department and
the intelligence community, of Third World languages.

A year
after McConnell arrived, as President Clinton was considering military action
in the former Yugoslavia, NSA began to get worried about finding enough people
who could translate Serbo-Croatian. Thus, on April 23, 1993, a curious
advertisement appeared in
Commerce Business Daily.
Placed by NSA's
military organization, the Army Intelligence and Security Command, it sought
"a group of approximately 125 linguists to provide translation and
interpretation support for U.S. forces in Yugoslavia." The work, said the
ad, "would be in a hostile, harsh environment." And the government
would pick up the cost of "life, dismemberment and medical
insurance."

A similar
crisis at NSA broke the following year, when President Clinton ordered American
troops into Haiti to restore order. "When Haiti blew up a few years
ago," said Deputy Director for Services Terry Thompson, "we looked
around; there were a total of three Haitian Kreyol linguists in the entire cryptologic
system. One in NSA, one in the Navy, and one in the Army, and that was it. So
we had to go outsource— hire a lot of Haitian Kreyol speakers, many of whom
lived in downtown Washington doing menial labor, and put them in a building
over in Columbia [Maryland] and send them the material to transcribe."

One reason
for the shortage of linguists is the tedium of the job. "You sit there
with a pair of headphones, rocking back and forth with your foot on a pedal
trying to figure out what people said," recalled one former NSA Russian
linguist. "It is very repetitious, incredibly boring, and very demanding.
It could drive you crazy." However, it could also be very educational,
said another Russian linguist, who recalled all the Russian curses he learned
while eavesdropping on the walkie-talkie conversations of Soviet troops on
maneuvers in Siberia.

To help
with the language problems, Director McConnell quietly turned to academia.
Several colleges were paid to develop textbooks and teaching materials in
exotic languages as well as to train university and NSA language teachers.
Among the schools chosen was the predominantly black Florida A&M
University, which was given a $1.74 million grant to fund courses in the
difficult African languages of Zulu and Xhosa, spoken largely in South Africa;
Farsi, which is spoken in Iran; and Punjabi and Bengali, from the Indian
subcontinent.

A side
benefit of the grant, agency officials hoped, would be to recruit to the agency
black students who had successfully completed the courses; this would not only
build up the NSA language base but also help increase minority staffing.
Unfortunately, however, many of the students enrolled in the courses had far
more interest in international business than in eavesdropping on communications
networks, and thus never went to work for NSA.

One
solution, which NSA for decades has been trying to perfect, is machine
translation. In the early and mid-1980s, NSA was focusing on a variety of
crises—the Russian invasion of Afghanistan, the fundamentalist Islamic takeover
in Iran, and the civil war in El Salvador. "NSA is faced with the growing
problem of documents in virtually every language and script," said one
agency report. To help find a way to quickly translate the reams of paper
flowing into the agency written in unusual languages with strange alphabets,
NSA turned to the University of Pennsylvania.

The
experimental program, funded on behalf of NSA by the Pentagon, involved
designing optical scanning technology to first identify and then read a number
of exotic languages. The machine was eventually able to translate
Azerbaijani-language newspapers printed in a non-standard version of the
Cyrillic alphabet. A Turkic dialect, Azerbaijani is spoken by several million
people in the former Soviet republic of Azerbaijan and the contiguous areas of
Iran and Afghanistan. Other languages focused on by the project included
Somali, Slovenian, and a Mayan Indian language, Chorti, that is spoken in parts
of Guatemala, Honduras, and El Salvador.

Today, for
more commonplace languages, NSA uses programs such as SYSTRAN that
automatically translate text at up to 750 pages per hour using Russian
dictionaries containing more than half a million words. The program translates
technical texts with better than 90 percent accuracy. On average, human
translation takes forty-five minutes per page. NSA has also developed a
technique that allows analysts with no prior knowledge of a language to quickly
search machine-readable foreign language databases for keywords and topics.

To find
key text quickly within a very large collection of foreign language
documents—such as Chinese or Devanagari (Sanskrit)—one program NSA uses is
Oleada XConcord.

A further
breakthrough in NSA's ability to pick out the right tree in a vast forest of
words came with the development of the software called Semantic Forests.
Semantic Forests allows NSA to sift through printed transcripts of
conversations, faxes, computer transfers, or any other written intercepts and
intelligently come up with the targeted subjects in which the agency is most
interested. The name derives from the software's ability to create a weighted
"tree" of meanings for each word in a document. During lab tests, the
software quickly sifted through an electronic filter large volumes of printed
matter, including transcripts of speech and data from Internet discussion
groups. One of the sample questions in the test was "What have the effects
of the UN sanctions against Iraq been on the Iraqi people, the Iraqi economy,
or world oil prices?" Initial tests proved very successful, increasing the
ability to locate target information from 19 percent to 27 percent in just one
year.

Far more
difficult than machine translation of printed texts is automatic translation
and transcription of voice communications, such as intercepted telephone
conversations in a variety of languages and accents. The ability to
automatically spot targeted words in millions of telephone calls all over the
world has long been a goal of NSA. A recent breakthrough was made by biomedical
engineers at the University of Southern California, who claim to have created
the first machine system that can recognize spoken words better than humans
can. The research was largely funded by the Pentagon, long used as a cover for
NSA contracts.

According
to the university, the system can "instantly produce clean transcripts of
conversations, identifying each of the speakers." Known at NSA as
"Speaker ID," the USC's Berger-Liaw Neural Network Speaker
Independent Speech Recognition System can mimic the way brains process
information. This gives the computer the ability to conduct "word
spotting" in target communications regardless of who or what pronounces
the word.

The new
system is also far better than the human ear at picking out words from vast
amounts of white noise. It can even extract targeted words or conversations
from the background clutter of other voices, such as the hubbub heard during
conference calls, meetings, or cocktail parties. "The system can identify
different speakers of the same word with superhuman acuity," said
university officials.

Despite
such progress, by 2001 there was still far more traffic than there were people
or machines to handle it. "It's a good-size problem," said Hayden.
"It's one that we're paying attention to, but the fixes are not immediate.
There's probably no philosophers' stone here that we can touch and say, 'Oh
yes, now the linguist problem's fixed.' There's probably a whole bunch of
discrete decisions that you make that you begin to reduce the magnitude of the
problem. One aspect of the problem is, just given the nature of our business,
the demands on linguists are higher."

Some NSA
language training takes enormous amounts of time, said Hayden, who himself was
trained as a Bulgarian linguist. "Group Three languages, and I believe
that's Arabic and Hebrew, take eighteen months," he said. "And Group
Four languages take two years. And those are Chinese, Japanese . . . And then
there is a whole other addition there to turn someone who has working knowledge
of the popular language into a cryptolinguist, which is the specialized
vocabulary. . . . It's a long time, these are long-term investments. And you
can see why, then, we have trouble mostly with our military linguists who move
a lot, whereas a civilian you hire for thirty-five years and you make a
front-end investment of five years, you've still got thirty years of return.
You've got a GI going through here on an eighteen-month tour."

Realizing
NSA's personnel plight, the House Intelligence Committee began a major push in
the late 1990s to redirect money away from various fields, ranging from
satellites to support staff, and toward analysis and linguists.