Databases Reference
In-Depth Information
dominant today, but they have been greatly enhanced by increased availability of training data and
improved language modeling techniques, among other advancements.
One of the most recognizable speech recognition systems is
Dragon Naturally Speaking
from
Nuance Communications, Inc. This software allows users to control their personal computer using
voice commands rather than, or in addition to, typed commands. With Nuance's 2010 revenue
surpassing $1 billion, it is clear that there is a burgeoning demand for such voice interfaces
11
.
A common way of evaluating ASR systems is by measuring the percentage of incorrect words,
or the
word error rate
(WER) .The WER can vary hugely depending on the task and the environment.
WER
One of the simplest recognition tasks is to identify digits spoken in isolation. On this task, state-of-
the-art systems feature WERs approaching zero. In contrast, it is much more difficult to recognize
continuous speech coming from multiple participants such as in a meeting environment. We will
introduce meeting datasets where state-of-the-art recognition systems yield slightly greater than
30% WER. A recurring question, then, is what impact ASR errors have on summarization and text
mining tasks.
While Figure
1.2
showed that online conversations are becoming more popular, it is also the
case that professionals still spend a great deal of time in meetings. In 2009, Doodle-acompany
focused on event scheduling - conducted a survey of 2500 administrative and management staff from
across Europe and U.S.
12
and found that on average people are attending 7.1 meetings per week,
that the meetings last a whopping 2.75 hours each, and with 7 participants in attendance. However,
the rise of the Web is also changing the way we meet. Figure
1.3
shows data from the same survey
indicating that only around a quarter of these meetings are face-to-face, with many others being
conducted online or via conference call. In any event, the average professional spends a great deal of
her working life speaking with other people. ASR systems allow us to capture those conversations
and feed them into text mining and summarization systems.
1.2
APPLICATION SCENARIOS
The explosive adoption of the new Internet-based social media indicates that they are extremely
effective in supporting communication and collaboration. However, we argue that, in several situa-
tions, the effectiveness of these new media could be increased considerably by providing users with
tools to mine and summarize both past and ongoing conversations. In this section we describe some
possible application scenarios. By no means do we claim our list to be complete, and it is one of the
goals of this topic to foster the creation of novel applications.
Join an ongoing conversation:
The government in your country just approved a major policy
change. You find an interesting blog/discussion forum about a news article supporting this
change with already 50 comments. You strongly oppose the new policy and you would like to
present your argument. Should you start a new thread? Or should you contribute to one of
Search WWH ::
Custom Search