hparse.tex

隐马尔可夫模型源代码

%/* ----------------------------------------------------------- */
%/*                                                             */
%/*                          ___                                */
%/*                       |_| | |_/   SPEECH                    */
%/*                       | | | | \   RECOGNITION               */
%/*                       =========   SOFTWARE                  */ 
%/*                                                             */
%/*                                                             */
%/* ----------------------------------------------------------- */
%/*         Copyright: Microsoft Corporation                    */
%/*          1995-2000 Redmond, Washington USA                  */
%/*                    http://www.microsoft.com                */
%/*                                                             */
%/*   Use of this software is governed by a License Agreement   */
%/*    ** See the file License for the Conditions of Use  **    */
%/*    **     This banner notice must not be removed      **    */
%/*                                                             */
%/* ----------------------------------------------------------- */
%
% HTKBook - Phil Woodland and Steve Young    24/11/97
%

\newpage
\mysect{HParse}{HParse}

\mysubsect{Function}{HParse-Function}

\index{hparse@\htool{HParse}|(}
The \htool{HParse} program generates word level lattice files (for use
with e.g. \htool{HVite}) from a text file syntax description containing a 
set of rewrite rules based on extended Backus-Naur Form (EBNF). 
The EBNF rules are used to generate an internal
representation of the corresponding finite-state network where \htool{HParse}
network nodes represent the words in the network, and are connected via
sets of links. This \htool{HParse} network is then converted to \HTK\ V2 word
level lattice. The program provides one convenient way of defining such
word level lattices. 

\htool{HParse} also provides a {\em compatibility mode} for use
with \htool{HParse} syntax descriptions used in \HTK\ V1.5 where
the same format was used to define both the word level syntax 
and the dictionary.
In compatibility mode \htool{HParse} will output the word level
portion of such a syntax as an \HTK\ V2 lattice file (via \htool{HNet})
and the pronuciation information as an \HTK\ V2 dictionary file (via 
\htool{HDict}).

The lattice produced by \htool{HParse} will often contain a number
of \texttt{!NULL} nodes in order to reduce the number of arcs in the
lattice. The use of such \texttt{!NULL} nodes can both
reduce size and  increase efficiency when used by recognition programs 
such as \htool{HVite}.

\mysubsect{Network Definition}{HParse-Network Definition}

The syntax rules for the textual definition of the network are 
as follows.  Each node in the network has a \texttt{nodename}.
This node name will normally correspond to a word in the final syntax
network. Additionally, for use in compatibility mode,
each node can also have an external name.
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= \kill
\>        name \> = \> char\{char\} \\
\>        nodename \> = \> name [ "\%" ( "\%" $|$ name ) ]
\end{tabbing}}

\noindent
Here \texttt{char} represents any character except one of the meta chars 
\texttt{\{ \} [ ] $<$ $>$$|$ = \$ ( ) ; $\backslash$ / *}.   The latter may, 
however, be escaped using a backslash.  The first name in a \texttt{nodename}
represents the name of the node (``internal name''), and the second optional name is
the ``external'' name.  This is used only in compatibility mode, and is, by default
the same as the internal name.

Network definitions may also contain variables
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= \kill
\>      variable \> = \> \$name
\end{tabbing}}
\noindent
Variables are identified by a leading \$ character.  They stand for
sub-networks and must be defined before they appear in the RHS of a rule
using the form
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>      subnet \> = \> variable ``='' expr ``;''
\end{tabbing}}
\noindent
An \texttt{expr} consists of a set of alternative sequences representing
parallel branches of the network. 
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>      expr \>  = \> sequence \{``$|$'' sequence\} \\
\>      sequence \> = \> factor\{factor\}
\end{tabbing}}
\noindent
Each sequence is composed of a sequence of factors where a factor
is either a node name, a variable representing some sub-network or
an expression contained within various sorts of brackets.

{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>   factor \> = \> ``('' expr ``)'' \> $|$ \\
\>\>\>            ``\{'' expr ``\}'' \> $|$ \\
\>\>\>            ``$<$'' expr ``$>$'' \> $|$ \\
\>\>\>         ``['' expr ``]'' \>  $|$ \\
\>\>\>             ``$<<$'' expr ``$>>$'' \> $|$ \\
\>\>\>               nodename \> $|$ \\
\>\>\>               variable 
\end{tabbing}}

Ordinary parentheses are used to bracket sub-expressions, curly braces \{ \} denote zero or more repetitions and angle brackets $<>$ denote one or more repetitions.  Square brackets [$\:$] are used to enclose optional items.  The double angle brackets are a special feature included for building context dependent loops and are explained further below.
Finally, the complete network is defined by a list of sub-network
definitions followed by a single expression within parentheses.
{\sf
\begin{tabbing}
++++++ \= ++++++++ \= ++ \= ++++++++++++ \=  \kill
\>    network \> = \> \{subnet\} ``('' expr ``)''
\end{tabbing}}
\noindent
Note that C style comments may be placed anywhere in the text of
the network definition.

As an example, the following network defines a syntax for some
simple edit commands
\begin{verbatim}
   $dir   = up | down | left | right;
   $mvcmd = move $dir | top | bottom;      
   $item  = char | word | line | page;
   $dlcmd = delete [$item];   /* default is char */
   $incmd = insert;
   $encmd = end [insert];
   $cmd = $mvcmd|$dlcmd|$incmd|$encmd;
   ({sil} < $cmd {sil} > quit)
\end{verbatim}

Double angle brackets are used to
construct contextually consistent context-dependent loops such
as a word-pair grammar.\footnote{The expression between 
double angle brackets must be a simple list of alternative node names or
a variable which has such a list as its value}
This function can also be used to generate consistent triphone loops 
for phone recognition\footnote{In \HTK\ V2 it is preferable for
these context-loop expansions to be done automatically via \htool{HNet},
to avoid requiring a dictionary entry for every context-dependent
model}.
The entry and exit conditions to a
context-dependent loop can be controlled by the invisible
pseudo-words TLOOP\_BEGIN and TLOOP\_END.  The right context of TLOOP\_BEGIN
defines the legal loop start nodes, and the left context of TLOOP\_END
defines the legal loop finishers. If TLOOP\_BEGIN/TLOOP\_END are not
present then all models are connected to the entry/exit of the loop.

A word-pair grammar simply defines the legal
set of words that can follow each word in the vocabulary.
To generate a network to represent such a grammar a
right context-dependent loop could be used.
The legal sentence set of sentence start and end words are defined as
above using TLOOP\_BEGIN/TLOOP\_END.

For example, the following lists the legal followers for each