cwb-ccc 0.9.3

CWB wrapper to extract concordances and collocates

Collocation and Concordance Computation

Introduction

This module is a wrapper around the IMS Open Corpus Workbench (CWB). It requires CWB version 3.4.16 or newer for anchored queries. Main purpose of the module is to extract concordance lines and to calculate collocates, as well as to extract the results of queries with more than two anchors.

Installation

You can install this package from PyPI:

pip3 install cwb-ccc

You can also clone the repository from gitlab.cs.fau.de and use setup.py:

python3 setup.py install

Last but not least, you can just install all requirements specified in setup.py and make sure the ccc subfolder can be found by Python by including it in your PYTHONPATH.

Usage

CWBEngine

All methods rely on the CWBEngine from ccc.cwb, which you first have to initialize with your system specific settings:

from ccc.cwb import CWBEngine
engine = CWBEngine(
	corpus_name="EXAMPLE_CORPUS",
	registry_path="/path/to/your/cwb/registry"
)

NB: this will raise a KeyError if the named corpus is not in the specified registry.

You can use the cqp_bin to point the engine to a specific version of cqp (this is also helpful if cqp is not in your PATH):

engine = CWBEngine(
	corpus_name="EXAMPLE_CORPUS",
	registry_path="/path/to/your/cwb/registry", 
	cqp_bin="/usr/local/cwb-3.4.16/bin/cqp"
)

If you are using macros and wordlists, you have to store them in a separate folder (with subfolders wordlists and macros). Make sure you specify this folder via lib_path when initializing the engine:

engine = CWBEngine(
	corpus_name="EXAMPLE_CORPUS", 
	registry_path="/path/to/your/cwb/registry",
	lib_path="/path/to/your/lib/"
)

Concordancing

You can use the Concordance class from ccc.concordances for concordancing. The concordancer has to be initialized with the engine and accepts valid CQP queries:

from ccc.concordances import Concordance

# initialize the concordancer with the engine
concordance = Concordance(engine)

# extract concordance lines
concordance.query('[lemma="Angela"] [lemma="Merkel"]')

The result will be a dictionary with the cpos of the match as keys and the entries one concordance line each. Each concordance line is formatted as a pandas.DataFrame with the cpos of each token as index:

cpos	word	match	offset
188530363	,	False	-5
188530364	dass	False	-4
188530365	die	False	-3
188530366	Tage	False	-2
188530367	von	False	-1
188530368	Angela	True	0
188530369	Merkel	True	0
188530370	gezählt	False	1
188530371	sind	False	2
188530372	.	False	3

The queries must not end on a "within" clause. If you want to restrict your concordance lines by a structural attribute, use the s_break parameter (defaults to "text"). The default context window is 20 tokens to the left and 20 tokens to the right of the query match and matchend, respectively.

concordance = Concordance(engine, context=50, s_break='s')
concordance.query('[lemma="Angela"] [lemma="Merkel"]')

Further parameters for the Concordance class are order (one of "random", "first", or "last"), cut_off (for the number of concordance lines to extract), and p_show (a list of additional p-attributes besides the primary word layer to show, e.,g. "lemma" or "pos"; these will be added as additional columns).

Anchored Queries

The Concordance class detects anchored queries by default. The following query

concordance.query(
	'@0[lemma="Angela"]? @1[lemma="Merkel"] '
	'[word="\\("] @2[lemma="CDU"] [word="\\)"]'
)

will thus return DataFrames with an additional column indicating the anchor positions:

cpos	word	match	offset	anchor
298906425	auch	False	-5	None
298906426	das	False	-4	None
298906427	Handy	False	-3	None
298906428	von	False	-2	None
298906429	Kanzlerin	False	-1	None
298906430	Angela	True	0	0
298906431	Merkel	True	0	1
298906432	(	True	0	None
298906433	CDU	True	0	2
298906434	)	True	0	None
298906435	sowie	False	1	None
298906436	ihres	False	2	None
298906437	Vorgängers	False	3	None
298906438	Gerhard	False	4	None
298906439	Schröder	False	5	None

Argument Queries

Argument queries are anchored queries with additional information. (1) Each anchor can be modified by an offset (usually used to capture underspecified tokens near an anchor point). (2) Anchors can be mapped to external identifiers for further logical processing, and (3) be given a clear name:

anchor	offset	idx	clear name
0	0	None	None
1	-1	None	None
2	0	None	None
3	-1	None	None

Furthermore, several anchor queries can be combined to form regions, which in turn can be mapped to identifiers and be given a clear name:

start	end	idx	clear name
0	1	"0"	"person X"
2	3	"1"	"person Y"

Example: Given the definition of anchors and regions above, the follwing complex query extracts corpus positions where there's some correlation between "person X" (the region from anchor 0 to anchor 1) and "person Y" (anchor 2 to 3):

query = (
	"<np> []* /ap[]* [lemma = $nouns_similarity] "
	"[]*</np> \"between\" @0:[::](<np>[pos_simple=\"D|A\"]* "
	"([pos_simple=\"Z|P\" | lemma = $nouns_person_common | "
	"lemma = $nouns_person_origin | lemma = $nouns_person_support | "
	"lemma = $nouns_person_negative | "
	"lemma = $nouns_person_profession] |/region[ner])+ "
	"[]*</np>)+@1:[::] \"and\" @2:[::](<np>[pos_simple=\"D|A\"]* "
	"([pos_simple=\"Z|P\" | lemma = $nouns_person_common | "
	"lemma = $nouns_person_origin | lemma = $nouns_person_support | "
	"lemma = $nouns_person_negative | "
	"lemma = $nouns_person_profession] | /region[ner])+ "
	"[]*</np>) (/region[np] | <vp>[lemma!=\"be\"]</vp> | "
	"/region[pp] |/be_ap[])* @3:[::]"
)

NB: the set of identifiers defined in the table of anchors and in the table of regions, respectively, should not overlap.

It is customary to store these queries in json query files such as the example in the repository. You can directly process these files using the process_argmin_file method from ccc.argmin:

from ccc.argmin import process_argmin_file

# process the query file
query_path = "tests/gold/query-example.json"
result = process_argmin_file(engine, query_path)

The result is a dict with the same keys as specified in the query file as well as an entry "result" with the following keys:

• "nr_matches": the number of query matches in the corpus.
• "matches": a list of concordance lines. Each concordance line contains
  • the corpus position of the match (entry "cpos")
  • the actual concordance line as returned from Concordance().query() (see above) converted to a dict (entry "df")
  • a mapping from the idx specified in the anchor and region tables to the tokens or token sequences, respectively (default: lemma layer) (entry "holes")
  • a reconstruction of the concordance line as a sequence of tokens (word layer) (entry "full")
• "holes": a global list of all tokens of the entities specified in the "idx" columns (default: lemma layer).

Acknowledgements

The module relies on several other python modules (see the requirements). Special thanks to Yannick Versley and Jorg Asmussen for the implementation of cwb-python.

This work has been funded by the Deutsche Forschungsgemeinschaft (DFG) within the project "Reconstructing Arguments from Noisy Text", grant number 377333057, as part of the Priority Program "Robust Argumentation Machines (RATIO)" (SPP-1999).