This is the backend for Korp, a corpus search tool developed by Språkbanken at the University of Gothenburg, Sweden.
The code is distributed under the MIT license.
The Korp backend is a Python 3 WSGI application, acting as a wrapper for Corpus Workbench.
To see what has changed in recent versions, see the CHANGELOG.
To use the basic features of the Korp backend you need the following:
- Python 3.6+
- Corpus Workbench (CWB) 3.4.12 or newer
To use the additional features such as the Word Picture you also need:
For optional (but strongly recommended) caching you need:
These instructions assume you are running a UNIX-like operating system (Linux, macOS, etc).
Download the current stable version of Corpus Workbench. Install by following the
Installing the CWB Core instructions, either by using the provided
packages or building from source. Refer to the included INSTALL
text file for further instructions.
Once CWB is installed, by default you will find it under /usr/local/cwb-X.X.X/bin
(where X.X.X
is the version
number). Confirm that the installation was successful by running:
/usr/local/cwb-X.X.X/bin/cqp -v
CWB needs two directories for storing the corpora. One for the data, and one for the corpus registry. You may create these directories wherever you want, but from here on we will assume that you have created the following two:
/corpora/data
/corpora/registry
Optionally you may set up a virtual Python environment:
$ python3 -m venv venv
$ source venv/bin/activate
Install the required Python modules using pip
with the included requirements.txt.
$ pip3 install -r requirements.txt
The supplied config.py
contains the default configuration. To override the default configuration, make a copy
of config.py
and place it in a directory named instance
in the repo root directory, and edit that copy.
The following variables need to be set for Korp to work:
-
CQP_EXECUTABLE
The absolute path to the CQP binary. By default/usr/local/cwb-X.X.X/bin/cqp
-
CWB_SCAN_EXECUTABLE
The absolute path to the cwb-scan-corpus binary. By default/usr/local/cwb-X.X.X/cwb-scan-corpus
-
CWB_REGISTRY
The absolute path to the CWB registry files. This is the/corpora/registry
folder you created before.
If you are planning on using functionality dependent on a database, you also need to set the following variables:
-
DBNAME
The name of the MySQL database where the corpus data will be stored. -
DBUSER & DBPASSWORD
Username and password for accessing the database.
For caching to work you need to specify both a cache directory (CACHE_DIR
) and a Memcached server address or socket
(MEMCACHED_SERVER
).
To run the backend, simply run run.py
:
python3 run.py
The backend should then be reachable in your web browser on the port you configured in config.py
, for
example http://localhost:1234
.
During development or while testing your configuration, use the flag dev
for automatic reloading.
python3 run.py dev
For deployment, Gunicorn works well.
gunicorn --worker-class gevent --bind 0.0.0.0:1234 --workers 4 --max-requests 250 --limit-request-line 0 'run:create_app()'
Most caching is done using Memcached, except for CWB query results which are temporarily saved to disk to speed up KWIC
pagination.
While Memcached handles removing old cache by itself, you will still have to tell it to invalidate parts of the cache
when one or more corpora are updated or added. This, and cleaning up the disk cache, is easily done by accessing the
/cache
endpoint. It might be a good idea to set up a cronjob or similar to regularly do this, making the cache
maintenance fully automatic.
The API documentation is available as an OpenAPI specification in docs/api.yaml, or online at https://ws.spraakbanken.gu.se/docs/korp.
Korp works as a layer on top of Corpus Workbench for most corpus search functionality. See the CWB corpus encoding tutorial for information regarding encoding corpora. Note that Korp requires your corpora to be encoded in UTF-8. Values of structural attributes may not contain tab characters. Once CWB is aware of your corpora they will be accessible through the Korp API.
For Korp to show the number of sentences and the date when a corpus was last updated, you have to manually add this information. Create a file called ".info" in the directory of the CWB data files for the corpus, and add to it the following lines (editing the values to match your material). Be sure to end the file with a blank line:
Sentences: 12345
Updated: 2019-11-30
FirstDate: 2001-01-16 00:00:00
LastDate: 2001-01-30 23:59:59
Once this file is in place, Korp will be able to access this information.
To use the basic concordance features of Korp there are no particular requirements regarding the markup of your corpora.
To use the Word Picture functionality your corpus must adhere to the following format:
- The structural annotation marking sentences must be named
sentence
. - Every sentence annotation must have an attribute named
id
with a value that is unique within the corpus.
To use the Trend Diagram functionality, your corpus needs to be annotated with date information using
the following four structural attributes: text_datefrom
, text_timefrom
, text_dateto
, text_timeto
.
The date format should be YYYYMMDD, and the time format hhmmss.
A corpus dated 2006 would have the following values:
text_datefrom: 20060101
text_timefrom: 000000
text_dateto: 20061231
text_timeto: 235959
This section describes the database tables needed to use the Word Picture, Lemgram index and Trend Diagram features. If you don't need any of these features, you can skip this section.
The Word Picture data consists of head-relation-dependent triplets and frequencies. For every corpus, you need five database tables. The table structures are as follows:
Table name: relations_CORPUSNAME
Charset: UTF-8
Columns:
id int A unique ID (within this table)
head int Reference to an ID in the strings table (below). The head word in the relation
rel enum(...) The syntactic relation
dep int Reference to an ID in the strings table (below). The dependent in the relation
freq int Frequency of the triplet (head, rel, dep)
bfhead bool True if head is a base form (or lemgram)
bfdep bool True if dep is a base form (or lemgram)
wfhead bool True if head is a word form
wfdep bool True if dep is a word form
Indexes:
(head, wfhead, dep, rel, freq, id)
(dep, wfdep, head, rel, freq, id)
(head, dep, bfhead, bfdep, rel, freq, id)
(dep, head, bfhead, bfdep, rel, freq, id)
Table name: relations_CORPUSNAME_strings
Charset: UTF-8
Columns:
id int A unique ID (within this table)
string varchar(100) The head or dependent string
stringextra varchar(32) Optional preposition for the dependent
pos varchar(5) Part-of-speech for the head or dependent
Indexes:
(string, id, pos, stringextra)
(id, string, pos, stringextra)
Table name: relations_CORPUSNAME_rel
Charset: UTF-8
Columns:
rel enum(...) The syntactic relation
freq int Frequency of the relation
Indexes:
(rel, freq)
Table name: relations_CORPUSNAME_head_rel
Charset: UTF-8
Columns:
head int Reference to an ID in the strings table. The head word in the relation
rel enum(...) The syntactic relation
freq int Frequency of the pair (head, rel)
Indexes:
(head, rel, freq)
Table name: relations_CORPUSNAME_dep_rel
Charset: UTF-8
Columns:
dep int Reference to an ID in the strings table. The dependent in the relation
rel enum(...) The syntactic relation
freq int Frequency of the pair (rel, dep)
Indexes:
(dep, rel, freq)
Table name: relations_CORPUSNAME_sentences
Charset: UTF-8
Columns:
id int An ID from relations_CORPUSNAME
sentence varchar(64) A sentence ID (see the section about corpus structure above)
start int The position of the first word of the relation in the sentence
end int The position of the last word of the relation in the sentence
Indexes:
id
In the main relations_CORPUSNAME
table, each relation should be represented three times. Once with both dependent
and head as base forms, once with dependent as base form and head as word form, and once with dependent as
word form and head as base form. This is to allow searching for both base forms and word forms, giving different
results for different searched word forms, while the results are always displayed as base forms.
If the base form annotation is missing for a dependent, head or both, the word form can be used as both word form and
base form by setting both bfhead/bfdep and wfhead/wfdep to True. In such a case you won't need all three rows for
that relation.
The sentences
table contains sentence IDs for sentences containing the relations, with start and end
values to point out exactly where in the sentences the relations occur (1 being the first word of the sentence).
The lemgram index is an index of every lemgram in every corpus, along with the number of occurrences. This is used by the frontend to grey out auto-completion suggestions which would not give any results in the selected corpora. The lemgram index consists of a single MySQL table, with the following layout:
Table name: lemgram_index
Charset: UTF-8
Columns:
lemgram varchar(64) The lemgram
freq int Number of occurrences
freq_prefix int Number of occurrences as a prefix
freq_suffix int Number of occurrences as a suffix
corpus varchar(64) The corpus name
Indexes:
(lemgram, corpus, freq, freq_prefix, freq_suffix)
For the Trend Diagram, you need to add token-per-time-span data to your database. For tokens without date or time info, use the date 0000-00-00 00:00:00. Use the following table layout:
Table name: timedata
Charset: UTF-8
Columns:
corpus varchar(64) The corpus name
datefrom datetime Full from-date and time
dateto datetime Full to-date and time
tokens int Number of tokens between from-date and (including) to-date
Indexes:
(corpus, datefrom, dateto)
Table name: timedata_date
Charset: UTF-8
Columns:
corpus varchar(64) The corpus name
datefrom date From-date (only date part)
dateto date To-date (only date part)
tokens int Number of tokens between from-date and (including) to-date
Indexes:
(corpus, datefrom, dateto)
The corpus configuration used by the Korp frontend is served by the backend. In config.py
, the variable
CORPUS_CONFIG_DIR
should point to a directory having the following structure:
.
├── attributes
│ ├── positional
│ │ ├── lemma.yaml
│ │ ├── msd.yaml
│ │ ├── ...
│ │ └── pos.yaml
│ └── structural
│ ├── author.yaml
│ ├── title.yaml
│ ├── ...
│ └── year.yaml
├── corpora
│ ├── corpus1.yaml
│ ├── corpus2.yaml
│ ├── ...
│ └── yet-another-corpus.yaml
└── modes
├── default.yaml
├── another.yaml
├── ...
└── other.yaml
- The modes directory contains one YAML file per mode in Korp.
- The corpora directory contains one YAML file per corpus.
- The attributes directory contains two subdirectories: positional and structural, containing optional attribute presets referred to by the corpus configurations.
For some inspiration, here are the config files used by the Korp instance at Språkbanken Text.
Note:
Most settings in these files referring to labels or descriptions can optionally be localized using ISO 639-3 language
codes. For example, a label can look both like this:
label: author
... and like this:
label:
eng: author
swe: författare
At least one mode file is required, and that file must be named default.yaml
. This is the mode that will be loaded
when no mode is explicitly requested.
Required:
- label: The name of the mode, which will be shown in the interface.
Optional:
- description: A description of the mode, shown when first entering it. May include HTML.
- order: A number used for sorting the modes in the interface. Modes without an order will end up last.
- folders: A folder structure for the corpus selector. These folders can then be referenced by individual corpora.
The folder structure can be of any depth, and folders can have any number of sub-folders (using the key
subfolders
). You may use HTML in the descriptions. Example:folders: novels: title: eng: Novels swe: Skönlitteratur description: eng: Corpora consisting of novels. swe: Korpusar bestående av skönlitteratur. subfolders: classics: title: eng: Classics swe: Klassiker scifi: title: Science-Fiction
- preselected_corpora: A list of corpus IDs which will be pre-selected when the user enters the mode. You may also
refer to folders by using the prefix
__
, and dot-notation for refering to subfolders. Example:preselected_corpora: - my-corpus - __novels.scifi
- Other than the above, you can also override almost all the global settings set in the frontend's
config.yaml
. See the documentation for the frontend for a list of available settings.
Corpus configuration files are placed in the corpora
folder, and the filename of each configuration file should
correspond to a corpus ID in lowercase, followed by .yaml
, e.g. mycorpus.yaml
.
Required:
- id: The corpus' system name, same as the configuration file name (minus
.yaml
). - title: Title of the corpus.
- description: Description of the corpus. HTML can be used.
- modes: A list of the modes in which the corpus will be included, optionally specifying a folder. Example:
mode: - name: default folder: novels.classics
Optional:
- within: Use this to override default_within (set in the global or mode config). within is a list of
structural elements to use as boundaries when searching, ordered from smaller to bigger. Example:
within: - label: eng: sentence swe: mening value: sentence - label: eng: paragraph swe: stycke value: paragraph
- context: Use this to override default_context (set in the global or mode config). context is a list of
structural elements that can be used as context in the displaying of the search results, ordered from smaller to
bigger. Example:
context: - label: eng: 1 sentence swe: 1 mening value: 1 sentence - label: eng: 1 paragraph swe: 1 stycke value: 1 paragraph
- attribute_filters: A list of structural attributes on which the user will be able to filter the search results, using menus in both simple and extended search.
- pos_attributes and struct_attributes: Lists of positional and structural attributes. Every item in each list
should be an object with one key. The key should be the ID of the attribute, e.g.
msd
for positional attributes ortext_title
for structural. The value should be either 1) an object with a complete attribute definition, or 2) a string referring to an attribute preset containing such a definition, e.g.msd
to refer toattributes/positional/msd.yaml
. With option 1, you may also refer to a preset by using the keypreset
and then extend/override that preset. The attribute definition is what tells the Korp frontend how to handle each attribute, like how it should be presented in the sidebar and what interface widget to use in extended search. For more information about what options are available for attribute definitions, see the Korp frontend documentation. Example:struct_attributes: - text_title: title - text_type: label: eng: type swe: typ - text_source: preset: url label: eng: source swe: källa
- custom_attributes: See Custom attributes.
- reading_mode: See Reading mode.
- limited_access: Set to
true
to indicate that this corpus requires the user to be logged in and having the right permissions.
See pos_attributes and struct_attributes above.