Page 21 - Winter Issue 2018
P. 21

Deep Language Learning
Steven Greenberg How technology enhances language instruction.
_ _ Address: Technology Is Transforming Haw Students
Silicon Speech _
_ Learn a Foreign Language
17270 Greenridge Road TM _ _ , _ . _ _ , _
_ In Star Trek and other science fiction, alien civilizations communicate via a
Hidden Valley Lake, _
, _ universal translator that seamlessly translates from one tongue to another, mak-
California 95467 , , , , ,
USA mg alien—language instruction superfluous. Unfortunately, such flawless universal
translation is unlikely to arrive anytime soon (despite recent advances).
_ _ Ema“: So, at least for now, the most effective path for communicating in a foreign tongue
steven@siliconspeech.com , _ _ , _ ,
is through instruction. Traditional language pedagogy emphasizes classroom and
laboratory practice of vocabulary, grammar, and pronunciation. Lessons are highly
structured, with students practicing language skills in class and laboratory. Feedback
is oifered mostly through exams and drills. However, this classical approach has se-
rious drawbacks, especially when it focuses on declarative knowledge of grammar
and vocabulary to the exclusion of conversational skills and comprehension.
Although the ambitious student might achieve conversational fluency by living
in a foreign community, this option is unavailable to many. Fortunately, curricula
are beginning to incorporate more naturalistic approaches to language learning,
powered by technology. The long-term goal is to emulate real—world learning in
ways that are effective, economical, and enjoyable.
For computer-assisted language learning (CALL), the “holy grail” is courseware
that simulates what a student might experience living in a foreign land. In this
virtual community, the student would converse in the target language and receive
feedback on ways to improve. Although this pedagogical nirvana won’t happen
anytime soon, several advances bring it closer to reality. Among these are

(1) powerful, inexpensive computing residing in the “cloud,” using a multitude
(often thousands) of machines (usually graphical processing units [GPUs])
and abundant memory that mobile devices (e.g., smartphone, tablet) and
computers can access easily;

(2) large amounts of online data to “train” pattern classifiers known as artificial
neural networks (ANNS);

(3) cloud-based “deep learning” neural networks (DNNs). These are especially
powerful ANNs that contain many (often dozens of) hidden layers and in-
tricate connection topologies. A layer is “hidden” if it lies between the input
stage of the ANN and its output (i.e., classifier outcome). Each hidden layer
adds a level of processing that facilitates the “learning” (through adjustment
of activation weights) of features critical for successful classification;

(4) DNN-trained automatic speech recognition (ASR) and speech synthesis
(TTS) that deliver a quality and naturalness close to what humans achieve
in many (though not all) languages. Many companies (e.g., Amazon, Apple,
Google, and Microsoft) use the technology to interact with customers and
clients. The data collected are used to further enhance the technology; and

©2018AcousticaI Society afAmerica. All rights reserved. volume 14, issue 4 | Winter 2018 ] Acuualzica 'n:n:Iay | 1 B














































   19   20   21   22   23