Technology Research News
Reading the written material that goes
along with museum exhibits is always a little tricky. If you're the type
who has to read every word, you're likely to see the same background information
over and over again, and if you're the type who likes to dip in and out
of the text, you'll probably end up missing at least some of background
Researchers from Europe have built a system designed to tap the powers
of hypertext, information databases, and natural language generation to
allow people to go as deeply or as quickly as they wish through the written
material in museum-type settings without repeating or missing much. "It
occurred to me that... these problems can be addressed by using natural
language generation technology," said Jon Oberlander, a reader in cognitive
science at the University of Edinburgh.
The information can be displayed in several forms in physical places like
museums and virtual spaces like the World
Wide Web. "The same information server and generator can dynamically
supply information to wireless handhelds in a real museum gallery, or
drive synthetic speech over a mobile phone, or build Web pages on-the-fly
to describe a virtual gallery," said Oberlander. The system is also designed
to work with any language.
There are several inherent problems with museum labels, according to Oberlander.
First, they are generally designed to be accessed in any order. This means
they must each represent all the relevant information about their object,
which can mean overly wordy and redundant descriptions. "Small differences
between two objects may be submerged in a sea of similar details," he
said. Using traditional labels, the only way to avoid massive redundancies
is to force visitors to read the descriptions in a certain order "and
that's not great for their sense of freedom," he said.
"Secondly, there's no guarantee that [visitors] will actually find what
they need," he added. In contrast, a live curator can find out what museum-goers
want, present options, and, if necessary, steer them to objects they were
not aware of, said Oberlander.
The researchers' system addresses those problems by generating answers
to visitors' questions on-the-fly. It keeps track of what a visitor has
seen in order to tailor the descriptions appropriately.
Someone visiting via the Web would start from a page of icons showing
a gallery of objects, and when the visitor clicked on a particular icon,
a new page would be generated, with a larger image, a title, a description
and a list of links to related objects. "At this point they can return
to the main page and choose another object, where they can follow one
of the suggested links, or they can ask for more information about the
current object. Either way a new page is generated for the chosen object
[and] the description of the page will take into account what other descriptions
have been generated so far, tailoring both content and form," he said.
Under the hood is software that includes four key components: a content
potential module, a text planner, a surface realizer and a module that
chooses the best presentation for the generated description.
The content potential module keeps track of, and links together, facts
extracted from museum databases and curator interviews. It also places
different values on each fact, depending on how important the curator
judges it to be and how interesting and familiar it is expected to be
to the visitor. This familiarity value changes throughout the course of
When a visitor requests information, the text planner module selects a
subset of facts from the content potential module. "It starts from the...
selected object, and includes all the facts which are nearby and sufficiently
interesting, important and unfamiliar," said Oberlander.
The module takes into consideration the number of facts available for
the current type of user, and organizes the information into a coherent
order that signals explicitly how the facts fit together, Oberlander said.
"The text structure built up this way is still essentially independent
of the language which is used to express the information," he said.
The surface realizer takes this abstract information and chooses the best
way of expressing it using grammatical constructions, words and connectives.
"This is also where the system takes into account the different ways we
refer to objects when we mention them the first time, [than] on subsequent
occasions," said Oberlander. For example, the first time you mention a
designer, you might say 'a British designer named Jessie M. King', then
later refer to her as 'Jessie M. King', 'King', or 'she'.
The final module decides whether to wrap the textual description in HTML
with live links, send it as pure text, or put it through a speech synthesizer.
In theory, the software can work with any language. The researchers are
currently working with English, Italian and Greek. "One of the key challenges
in the current project has been to cleanly separate the parts of the system
that are independent of English, Italian or Greek from the parts that
have to rely on knowledge of the particular language," he said.
In some ways, English is the easy language, Oberlander added. "The sophistication
of the system [had] to be considerably increased for languages with complex
word-information rules like Greek. But once you've done Greek, Italian
is relatively easy," he said. In the end, it shouldn't cost much to add
a new language, he said.
As part of the project, the researchers and a partner, the Foundation
of the Hellenic World in Athens, have constructed an immersive view of
the ancient city of Miletus using the software.
The researchers are also looking to use the software to mine many types
of existing textual information, including online catalogs. "It will work
with almost any kind of online catalog and in customer relationship management,"
said Oberlander. The researchers are also planning on using the system
for tutoring, he said.
The software combines work in several different areas in a very interesting
way, said Paul Aoki, a research scientist at the Xerox Palo Alto Research
Center. "They're able to [make] previous technologies really deployable,"
he said. "You can imagine that typical audio guide content like overviews,
jokes and dramatic stories would be tough to generate on-the-fly, but
something like [this] could be used to weave pre-recorded pieces together
with dynamic factual content."
The overall approach of generating text from a database of descriptive
elements could have many uses, Aoki said. "There are many different...
scenarios where this kind of technology can be applied -- walks through
historic districts, botanic gardens, historic houses. Another example
might be an audio restaurant guide that knows you care about parking and
price... and gives you natural-sounding descriptions that are tailored
to those preferences," he said.
Oberlander's research colleagues were Ion Androutsopoulos and Aggeliki
Dimitromanolaki of the Greek National Center for Scientific Research in
Greece, Vassiliki Kokkinhai of the Foundation of the Hellenic World in
Greece, Jo Calder of the University of Edinburgh, and Elena Not of the
Trentino Cultural Institute in Italy. They presented the research at the
29th Conference on Computer Applications and Quantitative Methods in Archeology
held in Gotland, Sweden, April 25 to 29, 2001. The research was funded
by the European Union.
TRN Categories: Human-Computer Interaction; Databases and
Story Type: News
Related Elements: Technical paper, "Generating Multilingual
Personalize Descriptions of Museum Exhibits -- The M-PIRO Project," presented
29th Conference on Computer Applications and Quantitative Methods in Archeology
in Gotland, Sweden, April 25-29, 2001.
Software guides museum-goers
thread from nanotubes
Brain cells control
heightens Web class divide
One-way heat valve possible
Research News Roundup
Research Watch blog
View from the High Ground Q&A
How It Works
News | Blog
Buy an ad link