\documentclass[fleqn]{article}

\usepackage{haldefs}
\usepackage{notes}
\usepackage{url}
\usepackage{graphicx}

\newcommand{\indep}{\perp\!\!\!\perp}

\begin{document}
\lecture{Artificial Intelligence}{HW10: Assorted Topics}{CS5300, Spring 2009}

% IF YOU'RE USING THIS .TEX FILE AS A TEMPLATE, PLEASE REPLACE
% "CS5300, Spring 2009" WITH YOUR NAME AND UID.

% Hand in at: http://www.cs.utah.edu/~hal/handin.pl?course=cs5300

\section{Speech and Machine Translation}

\begin{enumerate}
\item Statistical approaches to both speech recognition and machine
  translation make use of the notion of \emph{alignments}.  Why are
  alignments important in these tasks?  How do they differ between
  speech and MT?

\item {\bf (Bonus for all)} The ``recognition'' (or ``decoding'') step
  in speech recognition can be performed using HMM technology.  This
  is not immediately true in machine translation.  Why not?

\item Suppose we were to try to automatically induce alignments for
  the following data.  What alignments do you imagine the system would
  settle on?  Based on these, do you think there is enough information
  in this small amount of data to automatically induce alignments?
  What is left unknown?  (Languages are labeled ``A'' and ``B''.)

  \begin{tabular}{|l|l|}
    \hline
    A1. bpm uiv miba                      & A2. bpm uiv ibm                   \\
    B1. pqzc oi bijmzc                    & B2. pqzc oi bijmbi                \\
    \hline
    A3. q ibm zqkm                        & A4. q lzqvs jmmz                  \\
    B3. jwsc oi owpiv ew bijmbi           & B4. jwsc oi jqzc ew vwuc          \\
    \hline
  \end{tabular}
\end{enumerate}

\section{Machine Learning}

Alice decides to build a na\:ive Bayes classifier to distinguish
between emails from Professor Bob and Professor Clarence.  She has
collected the following examples of emails from these two Professors.
She simply uses words as features.  Compute every parameter for the
na\:ive Bayes classifier using maximum likelihood and classify the
final test examples.

\begin{tabular}{|l|l|}
\hline
Bob      & all students did great on this assignment \\
Bob      & students should come to my office \\
Bob      & should you need help talk to the ta \\
Bob      & the ta did great grading this assignment \\
\hline
Clarence & no one did this assignment on time \\
Clarence & all students should fail \\
Clarence & the assignment is graded by the ta  \\
\hline
\end{tabular}

Test example 1: ``you did great''

Test example 2: ``no students should fail''

Did the classifier do what you think it should?  If not, why not?

{\bf (Bonus for all)} Recompute and reclassify the test example using
Laplace smoothing rather than maximum likelihood.  Did the classifier
do what you think it should?  If not, why not?  Did the
classifications change?

\end{document}