Statistical Language Models based on Neural Networks - Faculty of ...

abilities of n-grams are stored in precomputed tables), reliability coming from simplicity,
and generality (models can be applied to any domain or language effortlessly, as long as
there exists some training data). N-gram models are today still considered as state of the
art not because there are no better techniques, but because those better techniques are
computationally much more complex, and provide just marginal improvements, not critical
for success of given application. Thus, large part of this thesis deals with computational
efficiency and speed-up tricks <strong>based</strong> on simple reliable algorithms.
The weak part of n-grams is slow adaptation rate when only limited amount of in-
domain data is available. The most important weakness is that the number of possible
n-grams increases exponentially with the length of the context, preventing these models
to effectively capture longer context patterns. This is especially painful if large amounts
of training data are available, as much of the patterns from the training data cannot be
effectively represented by n-grams and cannot be thus discovered during training. The idea
of using neural network <strong>based</strong> LMs is <strong>based</strong> on this observation, and tries to overcome the
exponential increase of parameters by sharing parameters among similar events, no longer
requiring exact match of the history H.
2.3 Advanced <strong>Language</strong> Modeling Techniques
Despite the indisputable success of basic n-gram models, it was always obvious that these
models are not powerful enough to describe language at sufficient level. As an introduc-
tion to the advanced techniques, simple examples will be given first to show what n-grams
cannot do. For example, representation of long-context patters is very inefficient, consider
the following example:
THE SKY ABOVE OUR HEADS IS BLUE
In such sentence, the word BLUE directly depends on the previous word SKY. There is
huge number of possible variations of words between these two that would not break such
relationship - for example, THE SKY THIS MORNING WAS BLUE etc. We can even see that
the number of variations can practically increase exponentially with increasing distance of
the two words from each other in the sentence - we can create many similar sentences for
example by adding all days of week in the sentence, such as:
17