# Clone this repository
$ git clone https://github.com/artificial-max/polcla.git
# Go into the repository
$ cd polcla/polcla
# Configure paths and properties
Open config/config.properties with a text editor of your choice.
# Install with maven
$ mvn install
# From the pocla/polcla directory, run the jar
java -jar target/polcla-0.0.1-SNAPSHOT-jar-with-dependencies.jar
This system expects the data to be preprocessed by various NLP analyses (i.e. part-of-speech tagging, parsing and named-entity recognition). The specific tools for this kind of preprocessing are not included in the release of this system. However, you may find shellscripts for installing and running these tools in another project here: https://github.com/miwieg/german-opinion-role-extractor
# Example configuration
# Path specifications
TEXT_INPUT=src/main/resources/Input/raw_text.txt
DEPENDENCY_INPUT= src/main/resources/Input/dependency_parse.txt
CONSTITUENCY_INPUT=src/main/resources/Input/constituency_parse.xml
SENTIMENT_LEXICON_INPUT=src/main/resources/dictionaries/germanlex.txt
SHIFTER_LEXICON_INPUT=src/main/resources/dictionaries/shifter_lex.txt
INTENSIFIER_LEXICON_INPUT=src/main/resources/dictionaries/intensifier_lex.txt
OUTPUT=output/salsaResult.xml
###
# Optional path specifications
# If available, the position of subjective expressions in a sentence may be given to the system.
# The xml-file for PRESET_SE_INPUT should be in the same format as the constituency parse file (CONSTITUENCY_INPUT), with additional subjective expression frames.
# The file constituency_parse.preset_ses.xml provides an example.
# If not available, the sentiment lexicon (SENTIMENT_LEXICON_INPUT) is used to identify subjective expressions and their locations.
PRESET_SE_INPUT=src/main/resources/Input/constituency_parse.preset_ses.xml
# Option to use preset subjective expression information (TRUE), or to use
# the sentiment lexicon for subjective expression identification instead (FALSE).
# If no file is specified for PRESET_SE_INPUT, this option only works if a file is specified for PRESET_SE_INPUT.
# Default: TRUE
USE_PRESET_SE_INPUT=TRUE
# Option to also consider neutral entries of SENTIMENT_LEXICON_INPUT (labeled NEU).
# If set to TRUE, neutral expressions will be labeled as polar expressions,
# but won't affect sentence polarity.
# Default: FALSE
INCLUDE_NEUTRAL_EXPRESSIONS=FALSE
###
# Option for dependency parse normalization.
# Turns passive voice into active voice among other things. Should improve detection and classification.
# Default: TRUE
NORMALIZE=TRUE
# Option to consider POS tags for subjective expressions.
# Compares POS tags of words in a sentence with sentiment lexicon entries.
# If a mismatch between the POS tag of a lexicon entry and a corresponding word is found, the word will not be considered to be a subjective
# expression.
# Only meaningful if USE_PRESET_SE_INPUT=FALSE.
# Default: TRUE
POS_LOOKUP_SENTIMENT=TRUE
# Option for pos lookup for shifters.
# The same rules as for POS_LOOKUP_SENTIMENT apply.
# Default: TRUE
POS_LOOKUP_SHIFTER=TRUE
# Option to take shifter orientation as given in the shifter lexicon (SHIFTER_LEXICON_INPUT) into account.
# Example:
# Shifter: "heilen" with orientation "n" (on negative)" will only shift negative subjective expressions if this option is set to
# TRUE.
# Default: FALSE = Higher Recall (and probably higher F-Score). TRUE = Higher Precision.
SHIFTER_ORIENTATION_CHECK=FALSE
###
# Window Baseline Module: Checks for shifter targets to the specified direction of a shifter within a given window-based scope.
# Default: FALSE
WINDOW_BASELINE_MODULE=FALSE
# The window size for shifter targets.
# Default size: 4
BASELINE_WINDOW=4
# The direction in which to search for shifter targets.
# Options: "LEFT", "RIGHT", "BOTH"
# Default: BOTH
BASELINE_DIRECTION=BOTH
# Clause Module: Checks for shifter targets in the clause in which a shifter occurs.
# Default: FALSE
CLAUSE_BASELINE_MODULE=FALSE
###
The system uses a sentiment lexicon (Germanlex), a shifter lexicon, and optionally, an intensifier lexicon.
The employed sentiment lexicon comes from PolArt (Klenner et al., 2009). It contains information about what words express a sentiment, the polarity of the sentiment (positive, negative or neutral) as well as the intensity of the sentiment. And finally, the part of speech of every entry. Germanlex also contains entries about shifters and intensifiers, although these entries are ignored in the sentiment lexicon, as they get their own lexicons instead.
Format:
Word {NEG|POS|NEU|SHI|INT}=PolarityStrength PoS
SHI for Shifters, INT for Intensifiers
INT < 1, e.g. 0.5 is a reduction factor, > 1, e.g. 2 is a gain factor
Example entries:
gut POS=1 adj
Unverschämtheit NEG=1 nomen
anzweifeln NEG=0.7 verben
The contains information about what words should be considered to be shifters, what type of shifter they are, what their scope is, and finally, their part of speech. The scope is specified with dependency labels. A list dependency labels with explanations and examples can be found here: https://github.com/rsennrich/ParZu/blob/master/LABELS.md
Detailed descriptions of label deviations can be looked up in a separate paper: Michael Wiegand et al. "Saarland University’s participation in the German sentiment analysis shared task (GESTALT)." Workshop Proceedings of the 12th KONVENS. 2014 (pp. 174-184).
Format:
<shifter> <type> [<scope>] <pos>
Type {n,p,g}
n: The shifter shifts negative sentiment words
p: The shifter shifts positive sentiment words
g: The shifter shifts all (general) sentiment words
Scope [{attr-rev,objg,obja,objd,objc,obji,s,objp-*,subj,gmod,dependent,governor,clause,objp-ohne}]
If the scope is given as "[subj]", only words that are in a subject relation with the shifter will be considered as the shifter target.
Examples:
stagnieren g [subj] verb
verschlimmern p [objg,obja,objd,objc,obji,s,objp-*,subj] verb
gescheitert g [subj,attr-rev] adj
Abspaltung g [gmod,objp-*] nomen
The functions in the same way as the shifter lexicon. In contrast to shifters, intensifiers are thought to increase the strength of a sentiment without changing its direction from positive to negative or the other way around.
Format:
<intensifier> <type> [<scope>] <pos>
Type {n,p,g}
Scope [{attr-rev,objg,obja,objd,objc,obji,s,objp-*,subj,gmod,dependent,governor,clause,objp-ohne}]
Examples:
ziemlich g [subj,attr-rev] adj
wirklich g [subj,attr-rev] adj
extrem g [subj,attr-rev] adj
Apart from the lexicons, the system expects raw text, a constituency parse and a dependency parse as . Optionally, an xml file specifying the position of subjective expressions can be given. You may find scripts to get the constituency and dependency parses here: https://github.com/miwieg/german-opinion-role-extractor
The format of the xml-file output is the TIGER/SALSA format (Erk & Padó, 2004). This format was designed for semantic role labeling, but also allows for easy modeling of negation. With help of the SALTO tool (Burchardt et al., 2006), quick annotation and a neat visualization of TIGER/SALSA files is possible.
The first screenshot shows an annotated example sentence (translation: The ladies and gentlemen are never satisfied with what they have). In this case, zufrieden is a Polar Expression that is being negated by the shifter Nie. Visualization via the SALTO tool.
To view the sentence polarity, either press CTRL + E, or right click the top left corner in the SALTO window to view the sentence flag marked as INTERESTING (see the circle marked 1 in the screenshot below).
To view the polarity of a Subjective Expression before and after a shift, right click its flag (see 2 in the screenshot).
This software is licensed under the GNU GENERAL PUBLIC LICENSE. Please refer to LICENSE.md for more detail.
If you intend to use this tool for your research, please acknowledge this software by citing M. Wiegand, M. Wolf and J. Ruppenhofer (2017) (Full bibliographic information, see References.)
This work was partially supported by the German Research Foundation (DFG) under grants RU 1873/2-1 and WI 4204/2-1.
This software package also includes the Java API to process the SALSA XML corpora: http://www.coli.uni-saarland.de/projects/salsa/page.php?id=software implemented by members of the SALSA-project at Saarland University. The software package also includes the Zurcher Polart-sentiment lexicon (Klenner et al., 2009). We thank the SALSA-project and the Department for Computational Linguistics at Zurich University (Manfred Klenner) for letting us use and redistribute their resources.
Maximilian Wolf email: M_S_Wolf@web.de Michael Wiegand email: michael.wiegand@lsv.uni-saarland.de
Aljoscha Burchardt et al. "SALTO–a versatile multi-level annotation tool." Proceedings of LREC. 2006.
Katrin Erk, and Sebastian Pado. "A Powerful and Versatile XML Format for Representing Role-semantic Annotation." LREC. 2004.
Manfred Klenner, Angela Fahrni and Stefanos Petrakis "PolArt: A Robust Tool for Sentiment Analysis", in Proceedings of the Nordic Conference on Computational Linguistics (NoDaLiDa), pages 235-238, Odense, Denmark. 2009.
Michael Wiegand, Maximilian Wolf and Josef Ruppenhofer "Negation Modeling for German Polarity Classification", in Proceedings of the German Society for Computational Linguistics and Language Technology (GSCL), Potsdam, Germany. 2017.
Michael Wiegand et al. "Saarland University’s participation in the German sentiment analysis shared task (GESTALT)." , Workshop Proceedings of the 12th KONVENS. 2014.