The IMS Open Corpus Workbench (CWB)

Corpus Encoding Tutorial

CWB Version 3.0
Stefan Evert & The OCWB Development Team
http://cwb.sourceforge.net/
5 January 2010

Contents
1 Prerequisites

2 First steps: Encoding and indexing

3 Indexing and compression without CWB/Perl

4 CWB corpora and XML

5 Adding attributes to an encoded corpus

6 Adding XML annotations

7 Accessing frequency information

10

8 Sentence alignment

12

A Registry file format

12

Prerequisites

In order to follow this tutorial, you need to install the IMS Open Corpus Workbench (CWB),
version 3.0 or newer, which can be downloaded from
http://cwb.sourceforge.net/download.php

Corpus Encoding Tutorial

2 FIRST STEPS: ENCODING AND INDEXING

It is easiest to install a pre-compiled binary package, following instructions on the Web page and in
the enclosed README file. You should also install the CWB/Perl interface, which includes the useful
cwb-make and cwb-regedit programs. A data package with all input files needed for the examples
in this tutorial below is available from
http://cwb.sourceforge.net/documentation.php

First steps: Encoding and indexing

The standard CWB input format is one-word-per-line text,1 with the surface form in the first column
and token-level annotations specified as additional TAB-separated columns. XML tags for sentence
boundaries and other structural annotation must appear on separate lines. This file format is also
called verticalized text and has the customary file extension .vrt. An example of the verticalized
text format for a short sentence with part-of-speech and lemma annotations is shown in Figure 1.
This file, as well as all other input files required by the following examples are made available in the
accompanying data package.
<s>
It
was
an
elephant
.
</s>

PP
VBD
DT
NN
SENT

it
be
an
elephant
.

Figure 1: Verticalized text file example.vrt

In order to encode the file as a
Create a data directory where files in the binary CWB format will be stored. Here, we assume
that this directory is called /corpora/data/example. If this directory already exists contains
corpus data (from a previous version), you should delete all files in the directory. NB: You need
a separate data directory for each corpus you want to encode.
Choose a registry directory, where all encoded corpora have to be registered to make them
accessible to the CWB tools. It is recommended that you use the default registry directory
/usr/local/share/cwb/registry.2 Otherwise, you will have to specify the path to your registry directory with a -r flag whenever you invoke one of the CWB tools (or set an appropriate
environment variable, see below). In this tutorial, we assume that you use the standard registry
directory.
The next step is to encode the corpus, i.e. convert the verticalized text to CWB binary format
with the cwb-encode tool. Note that the command below has to be entered on a single line.
1
Or, more precisely, one token per line; i.e., the CWB expects punctuation marks, parentheses, quotes, etc. on separate
lines. The precise tokenization rules depend on your theoretical assumptions and the requirements of annotation
software such as part-of-speech taggers. Note that the CWB does not include any NLP components and has to be
provided with a tokenized and annotated corpus.
2
In previous versions of the CWB, the default registry directory used to be /corpora/c1/registry (for historical
reasons). All binary packages of CWB 3.0 and newer use the new default setting. If you already have a working environment with the old registry path, you may want to compile the CWB source code yourself, selecting the traditional
site configuration.

c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

2 FIRST STEPS: ENCODING AND INDEXING

$ cwb-encode -d /corpora/data/example -f example.vrt

-R /usr/local/share/cwb/registry/example
-P pos -P lemma -S s
(The $ character indicates a command line to be entered in a Unix shell. It is inspired by the
customary input prompt used by the Bourne shells sh and bash.)
The first column of the input file is automatically encoded as the default positional attribute
(p-attribute) named word. -P flags are used to declare additional p-attributes, i.e. token-level annotations. -S flags declare structural attributes (s-attributes), which encode non-recursive XML
tags and whose names must correspond to the XML element names. By convention, all attribute
names must be lowercase (more precisely, they may only contain the characters a-z, 0-9, -, and ,
and may not start with a digit). Therefore, the names of XML elements to be included in the CWB
corpus must not contain any non-ASCII or uppercase letters.
The -R option automatically creates a registry file, whose filename has to be written in lowercase.
Note that it is necessary to specify the full path to the registry file, even if the default registry directory
is used. The CWB name of the corpus (also called the corpus ID) is identical to the name of the
registry file, but is written in uppercase (here it will be EXAMPLE). The CWB name is used to activate
a corpus in the query processor CQP, for instance.
Input files with the extension .gz are assumed to be in gzip format and are automatically decompressed
(provided that the gzip program is installed on your computer). Multiple input files can be specified
by using the -f option repeatedly, and will be read in the order in which they appear on the command
line. Note that shell wildcards (e.g. -f *.txt) do not work (since each file name must be preceded
by -f). However, it is possible to read all files named *.vrt or *.vrt.gz in a given directory using
the -F option (possibly repeated for multiple directories). Input files from the same directory will be
read in alphabetical order.
All options (-d, -f, -R, etc.) must precede the attribute declarations (-P, -S, etc.) on the command
line. It is mandatory to specify a data directory with the -d option.3 This directory should always be
given as an absolute path, so the corpus can be used from any location in the file system.
Before a corpus can be used with CQP and other CWB programs, various index files have to
be built. It is also strongly recommended to compress data files, especially for larger corpora.
The easiest and recommended method for indexing and compression is to use the cwb-make
script that comes with the CWB/Perl interface modules. If you are unable to install the
modules and use this script, refer to Section 3 for a manual procedure.
$ cwb-make -V EXAMPLE
If you do not use the standard registry directory /usr/local/share/cwb/registry, you will
have to specify the path to your registry directory with the -r option. Alternatively, you can
set the environment variable CORPUS_REGISTRY, which will automatically be recognized by all
CWB programs. In a Bourne shell (sh or bash), this is achieved with the command
$ export CORPUS_REGISTRY=/home/stefan/registry
In a C shell (csh or tcsh), the corresponding command is
$ setenv CORPUS_REGISTRY /home/stefan/registry
3

Previous versions of the CWB would default to the current working directory. As a result, simply typing cwb-encode
on the command line would litter this directory with a number of empty data files and then hang, waiting for corpus
data on the standard input.

c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

INDEXING AND COMPRESSION WITHOUT CWB/PERL

It is probably a good idea to add this setting to your login profile (~/.profile or ~/.login).
If you do not want to set the environment variable, you need to invoke cwb-make with
$ cwb-make -r /home/stefan/registry -V EXAMPLE
The following examples assume that you either use the default registry directory or have set the
CORPUS_REGISTRY variable appropriately.
You can also specify multiple registry directories separated by colon characters (:), both in
the CORPUS_REGISTRY environment variable and the -r options of command-line tools. This is
convenient e.g. if some corpora are stored on external hard drives that are not always mounted.
Such optional registry directories may be prefixed by a question mark (?) in order to indicate
that they may not be accessible (otherwise CQP and some other tools will print warnings to
alert you to possible typos in the registry path). For instance, one of the lead CWB developers
has the following registry path in his ~/.bashrc configuration:
$ export CORPUS_REGISTRY=/Corpora/registry:?/Volumes/X/CWB/registry
Note that the built-in default registry directory /usr/local/share/cwb/registry is not automatically appended to this path. If you want to specify additional registry directories but keep
the default one, you will have to include it in the value of CORPUS_REGISTRY.
The -V switch enables additional validation passes when an index is created and when data files are
compressed. It should be omitted when encoding very large corpora (above 50 million tokens), in
order to speed up processing. In this case, it is also advisable to limit memory usage with the -M
option. The amount specified should be somewhat less than the amount of physical RAM available
(depending on the number of users etc.; too little is better than too much). For instance, on a Linux
machine with 128 MB of RAM, -M 64 is a safe choice. Note that the cwb-make utility applies a default
limit of -M 75 if not explicit -M option is given.
Display some information about the encoded corpus (add -s option for details and to ascertain
that all necessary data files have been created).
$ cwb-describe-corpus EXAMPLE

Indexing and compression without CWB/Perl

If you do not have the CWB/Perl inteface installed, by far the best thing you can do is to install
the CWB/Perl modules and the included scripts, and then go back to Section 2. If it is absolutely
impossible to install CWB/Perl or you really want to learn the nitty-gritty of corpus encoding, continue
here.
In the manual procedure, indexing and compression are performed in separate steps by dierent
tools. First, you have to run cwb-makeall in order to build the necessary index files.
cwb-makeall -V EXAMPLE
Note that cwb-makeall accepts the same -V, -M and -r options as cwb-make.
When the index files have been created, the corpus can already be used with CQP and other CWB
tools. However, it is recommended that you compress the binary data files to save disk space and
improve performance, especially for large corpora (above 10 million tokens). Compression is only
supported for p-attributes at the current time.
c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

CWB CORPORA AND XML

For positional attributes, both the token stream data and the index can be compressed. There
are separate tools for compressing the two types of data files.
The token stream can be compressed with the cwb-huffcode tool. Use the -P option to process
a single attribute, or compress all p-attributes with -A.
$ cwb-huffcode -A EXAMPLE
Index files can be compressed with the cwb-compress-rdx tool, which accepts the same options.
$ cwb-compress-rdx -A EXAMPLE
When compression was successful, both tools will display the full pathnames of uncompressed data
files that are now redundant and can be deleted (namely, attrib.corpus after running cwb-huffcode,
as well as attrib.corpus.rev and attrib.corpus.rdx after running cwb-compress-rdx).
If you run cwb-makeall again, it will show now that the p-attributes are compressed. Note that the
compressed data files are validated by default, so it is safe to remove the redundant files. Validation
can be turned o with the -T option, but is less performance-critical than with cwb-makeall.
NB: If you re-encode a corpus, it is important to erase all files in the data directory first. The
cwb-makeall program will not recognize that existing index files or compressed data files are
out of date, and will therefore fail to rebuild them automatically. This is one of the reasons why
the CWB/Perl cwb-make tool should be preferred.

CWB corpora and XML

Nowadays, machine-readable text and linguistic annotations are often provided in XML format.
Version 3.0 of the IMS Open Corpus Workbench oers improved XML support, which is activated
by the following encoding options: -x for XML compatibility mode (recognises default entities and
skips comments as well as an XML declaration), -s to skip blank lines in the input, and -B to strip
whitespace from tokens. The verticalized text format with TAB-separated p-attributes is still required
by cwb-encode, but this format can easily be generated from an arbitrary XML file with the help of
an XSLT stylesheet. Figure 2 shows a typical example of an XML input file for the CWB (note that
this is still a well-formed XML file).
XML elements (i.e. matching pairs of start and end tags) can be encoded as s-attributes, which have
to be declared with -S flags (for the file vss.vrt, the flags -S story -S p -S s would be used). If XML
regions of the same type are nested, encoding will only work correctly if you add :0 to the s-attribute
declaration, which enables a rudimentary XML parser built into cwb-encode. Attribute-value pairs
in XML start tags, such as
<story num="4" title="A Thrilling Experience">
can be stored as a single unparsed text string (num="4" title="A Thrilling Experience") by using
the flag -V instead of -S. This form of encoding is not convenient for CQP queries, though. It
is more desirable to declare XML tag attributes explicitly, which will automatically split the XML
elements into multiple s-attributes. Note that the options -xsB should (almost) always be used and
will automatically ignore the XML declaration and the comment line in Figure 2.
Encode the verticalized XML file vss.vrt as a CWB corpus, with indexing and compression. NB:
The last attribute declaration flag (-0 collection) is a digit zero (for a null attribute, see
below).
c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

CWB CORPORA AND XML

<?xml version="1.0" encoding="ISO-8859-1" standalone="yes" ?>

<!-- A Thrilling Experience -->
<story num="4" title="A Thrilling Experience">
<p>
<s>
Tick
NN
tick
.
SENT
.
</s>
<s>
A
DT
a
clock
NN
clock
.
SENT
.
</s>
<s>
Tick
VB
tick
,
,
,
tick
VB
tick
.
SENT
.
</s>
</p>
...
</story>
Figure 2: Verticalized XML file vss.vrt
$ cwb-encode -d /corpora/data/vss -f vss.vrt
-R /usr/local/share/cwb/registry/vss
-xsB -P pos -P lemma
-S s:0 -S p:0 -S story:0+num+title -0 collection
$ cwb-make -V VSS
If you do not have the cwb-make script available, follow the steps in Section 3.
These commands will encode the corpus VSS and create a registry file, including the s-attributes s, p,
story, story num, and story title. The <story> start tags are parsed and the attribute values are
stored as annotations of the attributes story_num (value: 4) and story_title (value: A Thrilling
Experience). Regions of the story attribute itself will not be annotated. Use -V instead of -S to
store all attribute-value pairs as a single string, which can be useful for displaying and re-exporting
the XML tags.
XML elements with dierent names (such as <s> and <p>) are encoded independently, so they can
nest and overlap in arbitrary ways. The cwb-encode program does not perform any validation or
well-formedness tests. When elements are nested recursively (e.g. a <table> within a <table>), the
embedded elements will be ignored, though. After encoding, cwb-encode prints a summary listing
the number of dropped XML elements. If you want to preserve nested elements, you can specify a
maximal level of embedding instead of :0 in the examples above. For instance, -S table:2 allows two
levels of embedding for <table> elements. Nested elements are automatically renamed to <table1>
and <table2>, respectively, and stored in separate s-attributes.
Sometimes, the input data may contain XML tags that should not be encoded in the corpus. For
instance, the stories in vss.vrt have to be wrapped in a single root element <collection> in order
c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

5 ADDING ATTRIBUTES TO AN ENCODED CORPUS

to obtain a well-formed XML file. Instead of removing such tags during data preparation, they can
directly be filtered out by the cwb-encode tool. For this purpose, they have to be declared with the
flag -0 (digit zero, for null attribute) instead of -S or -V. All start and end tags of these elements
will be ignored completely. There is no need to add :0 or XML attribute declarations. Note that
all XML tags that have not been declared with a -S, -V or -0 flag will be encoded as literal tokens
(without annotations), accompanied by a warning message.
You may have noticed in Figure 2 that the XML file is declared to be in ISO-8859-1 (or Latin1) encoding rather than the standard UTF-8 format. The CWB has been developed exclusively for
ISO-8859-1 data (which is still widely used for German corpus data, e.g. by the TreeTagger and
associated tools). While it is possible to store and query data in other ISO-8859-x encodings or
Unicode data in UTF-8 format some features will not work properly unless the ISO-8859-1 encoding
is used.4 If you need to handle non-Latin-1 data with the CWB, make sure that you are aware of the
precise limitations. It is probably a good idea to look for other users on the CWB mailing list who
have experience with such corpora. Version 3.1 of the CWB is expected to provide basic support for
ISO-8859-x and UTF-8 encodings.

Adding attributes to an encoded corpus

In order to add positional attributes to a corpus that has already been encoded, create input data
in the standard verticalized format, but listing only the new attributes. Figure 3 shows an example of
such an input file, containing WordNet synonyms for the tokens from Figure 1 (without attempting
any form of word sense disambiguation). A corresponding list of synonyms for the complete VSS corpus
can be found in the file syns.vrt.
|
|be|cost|live|work|equal|exist|occur|...|
|
|elephant|
|
Figure 3: WordNet synonyms for the text shown in Figure 1 (excerpt from file syns.vrt)
The special notation seen in Figure 3 indicates that the synonyms for any given word constitute
an unordered set (or feature set in CWB terminology). Vertical bars (|) separate individual set
elements and enclose the entire set; a single bar | denotes the empty set. Feature sets are stored as
plain strings in a CWB-encoded corpus, but the special notation enables the query processor CQP
to test whether a particular string is contained in the set, match all set elements against a regular
expression, and compute the intersection of two sets.
The file syns.vrt is encoded as usual, but the default word attribute has to be suppressed with
the option -p -. It is highly recommended to check that the number of tokens in the new file
(wc -l syns.vrt) is equal to the corpus size (as reported by cwb-lexdecode -S EXAMPLE), so
that the new attribute is properly aligned to the rest of the corpus.
$ cwb-encode -d /corpora/data/vss -f syns.vrt -p - -P syn/
4

In particular, case-insensitive (%c) and accent-insensitive (%d) matching as well as LATEX notation for accented
characters are only supported for ISO-8859-1 data. Regular expressions do not work properly for UTF-8 data and should
only be used to express simple prefix and suffix constraint such as .*able in this case.

c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

ADDING XML ANNOTATIONS

Notice the slash (/) appended to the attribute name syn. This notation indicates that the
new attribute should be treated as a feature set; cwb-encode will automatically validate and
normalise the supplied values, issuing warnings if they are not well-formed feature sets.
The registry file for the corpus VSS (usually /usr/local/share/cwb/registry/vss) has to be edited
in order to declare the new attribute. Add the line
ATTRIBUTE syn
at the bottom of the file. If the CWB/Perl interface has been installed, the registry file can also
be edited from the command line with the cwb-regedit registry editor script:
$ cwb-regedit VSS :add :p syn
This script can also be used to list and delete attributes, and to print basic information about a
corpus (similar to cwb-describe-corpus, but easier for further processing). Type cwb-regedit -h
for further information.
Now you can build index files and compress the new attribute:
$ cwb-make -V VSS
In order to add structural attributes with computed start and end points (corpus positions),
you can use the cwb-s-encode tool. The corresponding start and end positions of existing s-attributes
can be obtained with cwb-s-decode. The following example adds information about sentence length
to the VSS corpus.
The existing s attribute is decoded into a temporary file, then awk is used to compute sentence
lengths, and the resulting annotated regions are encoded with cwb-s-encode.
$ cwb-s-decode VSS -S s > s.list
$ awk BEGIN { FS=OFS="\t" } { print $1, $2, $2-$1+1 } s.list > s_len.list
$ cwb-s-encode -d /corpora/data/vss -f s_len.list -V s_len
Note that it is currently not necessary to run cwb-make after adding an s-attribute.
However, the new attribute still has to be declared in the registry file, either by manually adding
STRUCTURE s_len
or from the command line using the registry editor script:
$ cwb-regedit VSS :add :s s_len
Tables of corpus positions as input for cwb-s-encode can also be created from CQP query results
using the dump or tabulate command in a CQP session.

Adding XML annotations

In order to add XML annotations (e.g. <np> and <pp> tags inserted by a chunk parser) to an
existing corpus, the usual strategy is to decode the token stream (and other attributes if necessary) to
a temporary file. A chunk parser will often expect <s> and </s> tags marking sentence boundaries.
Decode token stream (word forms) with start and end tags for <s> regions.
c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

ADDING XML ANNOTATIONS

$ cwb-decode -C VSS -P word -S s > word_s.vrt

We then run the chunk parser on the temporary file, which adds its <np> and <pp> tags to the
token stream, creating the file shown in Figure 4. This file is also provided as part of the data
package for this tutorial.
<s>
<np head="experience>
My
experience
<pp head="of">
of
<np head="life">
life
</np>
</pp>
</np>
did
not
...
</s>
Figure 4: Decoded text with chunk annotations (file chunks.vrt)
It is important that the token stream is left intact when adding XML annotations. In particular,
tokens (as well as XML tags) must remain on separate lines and may not be split or combined.
As a preliminary check, make sure that the number of tokens in chunks.vrt is equal to the corpus
size.
$ grep -v ^< chunks.vrt | wc -l
Now we can use cwb-encode to encode the XML annotations as structural attributes. The start and
end points of regions are automatically computed from the token stream. Since we do not want to
overwrite the word attribute, we specify -p -. With no p-attributes declared, all lines in the input file
except for the XML tags will be ignored. Recall that -0 s (digit zero) instructs cwb-encode to ignore
<s> and </s> tags (without -S s they would otherwise be interpreted as literal tokens and mess up
the token stream).
Encode <np> and <pp> regions in chunks.vrt as new s-attributes.
$ cwb-encode -d /corpora/data/vss -f chunks.vrt
-p - -0 s -S np:0+head -S pp:0+head
In this example, cwb-encode will issue warnings about nested regions being dropped. As can be seen
from Figure 4, <np> (as well as <pp>) regions may be embedded recursively. In order to preserve such
nested regions, change the :0 modifier to :2, allowing up to two levels of embedding (separately for
each element type, i.e. <np> regions embedded in larger <np> regions, etc.). In general, :n allows up
to n levels of embedding. The embedded regions will automatically be renamed to np1, np2, pp1, and
pp2, respectively.
Encode chunks.vrt, allowing up to two levels of embedding for <np> and <pp> regions.
c 20052010 Stefan Evert & The OCWB Development Team

Corpus Encoding Tutorial

ACCESSING FREQUENCY INFORMATION

$ cwb-encode -d /corpora/data/vss -f chunks.vrt

-p - -0 s -S np:2+head -S pp:2+head
The full list of s-attributes created by this command is np, np1, np2, np head, np head1,
np head2, pp, pp1, pp2, pp head, pp head1, and pp head2. They have to be declared in the
registry file of the corpus VSS, either by adding the appropriate entries manually, or with the
registry editor script:
$ cwb-regedit VSS :add :s np np1 np2 np_head np_head1 np_head2
$ cwb-regedit VSS :add :s pp pp1 pp2 pp_head pp_head1 pp_head2
Attribute-value pairs in XML start tags may also contain feature sets. For instance, the German chunk parser YAC5 uses this notation to represent partially disambiguated morphological
features of NPs and PPs (see the CQP Query Language Tutorial for more information and
examples). XML tags of the form
<np agr="|Nom:F:Sg|Acc:F:Sg|" head="Wiese">
might be encoded with the declaration -S np:2+agr/+head, where the slash / indicates that
agr values are feature sets. Since head is not followed by a slash, the corresponding values are
not treated as feature sets.

Accessing frequency information

The cwb-lexdecode tool procides access to the lexicon of positional attributes, i.e. lists of all word
forms or annotation strings (types) with their corpus frequencies. The -S option prints the size of
corpus (tokens) and lexicon (types) only, -P selects the desired p-attribute, -f shows corpus frequencies,
and -s lists the lexicon entries alphabetically (according to the internal sort order). In order to sort
the lexicon by frequency, an external program (e.g. sort) has to be used.
$ cwb-lexdecode -S
-P lemma VSS
$ cwb-lexdecode -f -s -P lemma VSS | tail -20
$ cwb-lexdecode -f
-P lemma VSS | sort -nr -k 1 | head -20
It is also possible to annotate strings from a file (called tags.txt here) with corpus frequencies. The
file must be in one-word-per-line format. -0 (digit zero) prints a frequency of 0 for unknown strings
rather than issuing a warning message; it can be combined with -f to the mnemonic form -f0.
$ cwb-lexdecode -f0 -P pos -F tags.txt VSS
With the -p option, word forms or annotations matching a regular expression can be extracted. Caseinsensitive and accent-insensitive matching is selected with -c and -d, respectively.6 The example
below is similar to the CQP query [lemma = "over.+" %c]; but may be considerably faster on a
large corpus.
$ cwb-lexdecode -f -P lemma -p over.+ -c VSS
An entire corpus or selected attributes from a corpus can be printed in various formats with the
cwb-decode tool. Note that options and switches must appear before the corpus name, and the flags
used to select attributes after the corpus name. Use -P to select p-attributes and -S for s-attributes.
With the -s and -e options, a part of the corpus (identified by start and end corpus position) can be
printed.
5
6

http://www.ims.uni-stuttgart.de/~kermes/YAC/YAC.shtml
Recall that these flags are only guaranteed to work correctly for a corpus in ISO-8859-1 (Latin-1) encoding.

c 20052010 Stefan Evert & The OCWB Development Team

10

Corpus Encoding Tutorial

$ cwb-decode

-C -s 7299 -e 7303

VSS

ACCESSING FREQUENCY INFORMATION

-P word -P pos -S s

-C refers to the compact one-word-per-line format expected by cwb-encode. For a full textual copy
of a CWB corpus, use -ALL to select all positional and structural attributes.
$ cwb-decode

-C

VSS

-ALL

> vss-corpus.vrt

The resulting file vss-corpus.vrt can be re-encoded with cwb-encode (using appropriate flags) to give
an exact copy of the VSS corpus. -Cx is almost identical to the compact format, but changes some
details in order to generate a well-formed XML document (unless there are overlapping regions or
s-attributes with simple annotations).7
$ cwb-decode -Cx VSS -ALL > vss-corpus.xml
$ xmllint --noout vss-corpus.xml # not well-formed :-(
This output format can reliably be re-encoded when the -xsB options are used. Finally, -X produces a
native XML output format (following a fixed DTD), which can be post-processed and formatted with
XSLT stylesheets.
$ cwb-decode

-X -s 7299 -e 7303

VSS

-P word -P pos -S s -S np_head

Note that the regions of s-attributes are not translated into XML regions. Instead, the start and end
tags are represented by special empty <tag> elements.
The cwb-scan-corpus computes combinatorial frequency tables for an encoded corpus. Similar
to the group command in CQP, it is a faster and more memory-efficient alternative for the extraction
of simple structures from large corpora, and is not restricted to singletons and pairs. The output
of cwb-scan-corpus is an unordered list of n-tuples and their frequencies, which have to be postprocessed and sorted with external tools. The simple example below prints the twenty most frequent
(lemma, pos) pairs in the VSS corpus, using the -C option to filter punctuation and noise from the list
of lemmata (note that -C applies to all selected attributes).
$ cwb-scan-corpus -C VSS lemma pos | sort -nr -k 1 | head -20
A non-negative oset can be added to each field key in order to collect bigrams, trigrams, etc. The
following example derives a simple language model in the form of all sequences of three consecutive
part-of-speech tags together with their occurrence counts. Only the twenty most frequent sequences
are displayed.
$ cwb-scan-corpus VSS pos+0 pos+1 pos+2 | sort -nr -k 1 | head -20
For a large corpus such as the BNC, the scan results can directly be written to a file with the -o switch.
If the filename ends in .gz (such as the file language-model.gz in the example below), the output file
is automatically compressed (using gzip).
$ cwb-scan-corpus -o language-model.gz BNC pos+0 pos+1 pos+2
7

In order to re-create the original input file vss.vrt as a well-formed XML document, it would have been necessary to
store the full strings of attribute-value pairs from XML start tags by using -V flags instead of -S in cwb-encode attribute
declarations (e.g. -V story:0+num+title). In the cwb-decode call, problematic s-attributes created by auto-splitting of
these attribute-value pairs (story_num, story_title, s_len, np_head, . . . ) can then be omitted. The specification -S
story would print the full attribute-value pairs in <story> tags, etc.

c 20052010 Stefan Evert & The OCWB Development Team

11

Corpus Encoding Tutorial

REGISTRY FILE FORMAT

The values of the selected p-attributes can also be filtered with regular expressions. The following
command identifies part-of-speech sequences at the end of sentences (indicated by the tag SENT =
sentence-ending punctuation).
$ cwb-scan-corpus VSS pos+0 pos+1 pos+2=/SENT/ | sort -nr -k 1 | head -20
Since the third key is used only for filtering, we can suppress it in the output by marking it as a
constraint key with the ? character. Note that it may be necessary to enclose more complex keys
(containing shell metacharacters) in single quotes.
$ cwb-scan-corpus VSS pos+0 pos+1 ?pos+2=/SENT/ | sort -nr -k 1 | head -20
Note that cwb-scan-corpus can operate both on p-attributes and on s-attributes with annotated
values. To obtain by-story frequency lists for the VSS corpus, use the following command:
$ cwb-scan-corpus -o freq-by-story.tbl VSS lemma+0 story_title+0
As a special case, s-attributes without annotated values can be used to restrict the corpus scan to
regions of a particular type. For instance, the constraint key ?footnote would only scan <footnote>
regions. Keep in mind that such special constraints must not include a regular expression part.
The final example extracts pairs of adjacent adjectives and nouns from the VSS corpus, e.g. as candidate
data for adjective-noun collocations. Constraint keys are used to identify adjectives and nouns, and
only nouns starting with a vowel are accepted here. Note the c and d modifiers (case- and diacriticsinsensitive matching) on this regular expression.
$ cwb-scan-corpus -C VSS lemma+0 ?pos+0=/JJ.*/ lemma+1=/[aeiou].+/cd ?pos+1=/NN.*/
Except for the -C option, this command line is equivalent to the following CQP commands, but it will
execute much faster on a large corpus.
> A = [pos = "JJ.*"] [pos = "NN.*" & lemma = "[aeiou].+" %cd];
> group A matchend lemma by match lemma;
The cwb-scan-corpus command is limited to relatively simple constraints on tokens and it can only
match patterns with fixed osets (but not e.g. determiner and noun separated by an arbitrary number
of adjectives). To obtain frequency tables for more complex patterns, use CQP queries in combination
with the tabulate function. The resulting data tables can be saved to disk files and loaded into a
relational database or processed with a software package for statistical analysis.

Sentence alignment

Describe sentence aligner, encoding of alignment attributes, and import of pre-existing alignment

TODO

Registry file format

Document standard registry file format

TODO

CWB traditionally had a more flexible registry file format (which is still accepted for backward
compatibility), but standard format is recommended for new corpora (and is enforced by the
CWB/Perl scripts)

c 20052010 Stefan Evert & The OCWB Development Team

12

Corpus Encoding Tutorial

REGISTRY FILE FORMAT

note that directory and file paths in HOME and INFO entries have to be double-quoted if they
contain blanks or other non-standard characters (ASCII letters, digits, -, _, / and . are ok, as
long as path does not begin with .); in double-quoted path, " must be escaped as \" and the
backslash \ as \\; cwb-encode and cwb-regedit should always create valid entries, quoted as
necessary

c 20052010 Stefan Evert & The OCWB Development Team

13