Semantic Role Labeling (SRL) is a well-defined task where the objective is to analyze propositions expressed by the verb. In SRL, each word that bears a semantic role in the sentence has to be identified. There are different types of arguments (also called ’thematic roles’) such as Agent, Patient, Instrument, and also of adjuncts, such as Locative, Temporal, Manner, and Cause. These arguments and adjuncts represent entities participating in the event and give information about the event characteristics.
In the field of SRL, PropBank is one of the studies widely recognized by the computational linguistics communities. PropBank is the bank of propositions where predicate- argument information of the corpora is annotated, and the semantic roles or arguments that each verb can take are posited.
Each verb has a frame file, which contains arguments applicable to that verb. Frame files may include more than one roleset with respect to the senses of the given verb. In the roleset of a verb sense, argument labels Arg0 to Arg5 are described according to the meaning of the verb. For the example below, the predicate is “announce” from PropBank, Arg0 is “announcer”, Arg1 is “entity announced”, and ArgM- TMP is “time attribute”.
[ARG0 Türk Hava Yolları] [ARG1 indirimli satışlarını] [ARGM-TMP bu Pazartesi] [PREDICATE açıkladı].
[ARG0 Turkish Airlines] [PREDICATE announced] [ARG1 its discounted fares] [ARGM-TMP this Monday].
The following Table shows typical semantic role types. Only Arg0 and Arg1 indicate the same thematic roles across different verbs: Arg0 stands for the Agent or Causer and Arg1 is the Patient or Theme. The rest of the thematic roles can vary across different verbs. They can stand for Instrument, Start point, End point, Beneficiary, or Attribute. Moreover, PropBank uses ArgM’s as modifier labels indicating time, location, temporal, goal, cause etc., where the role is not specific to a single verb group; it generalizes over the entire corpus instead.
| Tag | Meaning |
|---|---|
| Arg0 | Agent or Causer |
| ArgM-EXT | Extent |
| Arg1 | Patient or Theme |
| ArgM-LOC | Locatives |
| Arg2 | Instrument, start point, end point, beneficiary, or attribute |
| ArgM-CAU | Cause |
| ArgM-MNR | Manner |
| ArgM-DIS | Discourse |
| ArgM-ADV | Adverbials |
| ArgM-DIR | Directionals |
| ArgM-PNC | Purpose |
| ArgM-TMP | Temporals |
- Collect a set of sentences to annotate.
- Each sentence in the collection must be named as xxxx.yyyyy in increasing order. For example, the first sentence to be annotated will be 0001.train, the second 0002.train, etc.
- Put the sentences in the same folder such as Turkish-Phrase.
- Build the Java project and put the generated sentence-propbank-predicate.jar and sentence-propbank-argument.jar files into another folder such as Program.
- Put Turkish-Phrase and Program folders into a parent folder.
- Open sentence-propbank-predicate.jar file.
- Wait until the data load message is displayed.
- Click Open button in the Project menu.
- Choose a file for annotation from the folder Turkish-Phrase.
- For each predicate word in the sentence, click the word, and choose PREDICATE tag for that word.
- Click one of the next buttons to go to other files.
- Open sentence-propbank-argument.jar file.
- Wait until the data load message is displayed.
- Click Open button in the Project menu.
- Choose a file for annotation from the folder Turkish-Phrase.
- For each word in the sentence, click the word, and choose correct argument tag for that word.
- Click one of the next buttons to go to other files.
After annotating sentences, you can use DataGenerator package to generate classification dataset for the Semantic Role Labeling task.
After generating the classification dataset as above, one can use the Classification package to generate machine learning models for the Semantic Role Labeling task.
You can also see either Python, Java, C++, Js, Swift, or C# repository.
To check if you have a compatible version of Python installed, use the following command:
python -V
You can find the latest version of Python here.
Install the latest version of Git.
pip3 install NlpToolkit-SemanticRoleLabeling-Cy
In order to work on code, create a fork from GitHub page. Use Git for cloning the code to your local or below line for Ubuntu:
git clone <your-fork-git-link>
A directory called DataStructure will be created. Or you can use below link for exploring the code:
git clone https://github.com/starlangsoftware/SemanticRoleLabeling-Cy.git
Steps for opening the cloned project:
- Start IDE
- Select File | Open from main menu
- Choose
SemanticRoleLabeling-Cyfile - Select open as project option
- Couple of seconds, dependencies with Maven will be downloaded.
The first task in Semantic Role Labeling is detecting predicates. In order to detect the predicates of the sentence, we use autoPredicate method of the TurkishSentenceAutoPredicate class.
sentence = ...
turkishAutoPredicate = TurkishSentenceAutoPredicate(FramesetList())
turkishAutoPredicate.autoPredicate(sentence)
Afterwards, one has to annotate the arguments for each predicate. We use autoArgument method of the TurkishSentenceAutoArgument class for that purpose.
turkishAutoArgument.autoArgument(sentence)
@article{tbtkelektrik400987,
journal = {Turkish Journal of Electrical Engineering and Computer Science},
issn = {1300-0632},
eissn = {1303-6203},
address = {},
publisher = {TÜBİTAK},
year = {2018},
volume = {26},
pages = {570 - 581},
doi = {},
title = {Construction of a Turkish proposition bank},
key = {cite},
author = {Ak, Koray and Toprak, Cansu and Esgel, Volkan and Yıldız, Olcay Taner}
}
- Do not forget to set package list. All subfolders should be added to the package list.
packages=['Classification', 'Classification.Model', 'Classification.Model.DecisionTree',
'Classification.Model.Ensemble', 'Classification.Model.NeuralNetwork',
'Classification.Model.NonParametric', 'Classification.Model.Parametric',
'Classification.Filter', 'Classification.DataSet', 'Classification.Instance', 'Classification.Attribute',
'Classification.Parameter', 'Classification.Experiment',
'Classification.Performance', 'Classification.InstanceList', 'Classification.DistanceMetric',
'Classification.StatisticalTest', 'Classification.FeatureSelection'],
- Package name should be lowercase and only may include _ character.
name='nlptoolkit_math',
- Package data should be defined and must ibclude pyx, pxd, c and py files.
package_data={'NGram': ['*.pxd', '*.pyx', '*.c', '*.py']},
- Setup should include ext_modules with compiler directives.
ext_modules=cythonize(["NGram/*.pyx"],
compiler_directives={'language_level': "3"}),
- Define the class variables and class methods in the pxd file.
cdef class DiscreteDistribution(dict):
cdef float __sum
cpdef addItem(self, str item)
cpdef removeItem(self, str item)
cpdef addDistribution(self, DiscreteDistribution distribution)
- For default values in class method declarations, use *.
cpdef list constructIdiomLiterals(self, FsmMorphologicalAnalyzer fsm, MorphologicalParse morphologicalParse1,
MetamorphicParse metaParse1, MorphologicalParse morphologicalParse2,
MetamorphicParse metaParse2, MorphologicalParse morphologicalParse3 = *,
MetamorphicParse metaParse3 = *)
- Define the class name as cdef, class methods as cpdef, and __init__ as def.
cdef class DiscreteDistribution(dict):
def __init__(self, **kwargs):
"""
A constructor of DiscreteDistribution class which calls its super class.
"""
super().__init__(**kwargs)
self.__sum = 0.0
cpdef addItem(self, str item):
- Do not forget to comment each function.
cpdef addItem(self, str item):
"""
The addItem method takes a String item as an input and if this map contains a mapping for the item it puts the
item with given value + 1, else it puts item with value of 1.
PARAMETERS
----------
item : string
String input.
"""
- Function names should follow caml case.
cpdef addItem(self, str item):
- Local variables should follow snake case.
det = 1.0
copy_of_matrix = copy.deepcopy(self)
- Variable types should be defined for function parameters, class variables.
cpdef double getValue(self, int rowNo, int colNo):
- Local variables should be defined with types.
cpdef sortDefinitions(self):
cdef int i, j
cdef str tmp
- For abstract methods, use ABC package and declare them with @abstractmethod.
@abstractmethod
def train(self, train_set: list[Tensor]):
pass
- For private methods, use __ as prefix in their names.
cpdef list __linearRegressionOnCountsOfCounts(self, list countsOfCounts)
- For private class variables, use __ as prefix in their names.
cdef class NGram:
cdef int __N
cdef double __lambda1, __lambda2
cdef bint __interpolated
cdef set __vocabulary
cdef list __probability_of_unseen
- Write __repr__ class methods as toString methods
- Write getter and setter class methods.
cpdef int getN(self)
cpdef setN(self, int N)
- If there are multiple constructors for a class, define them as constructor1, constructor2, ..., then from the original constructor call these methods.
cdef class NGram:
cpdef constructor1(self, int N, list corpus):
cpdef constructor2(self, str fileName):
def __init__(self,
NorFileName,
corpus=None):
if isinstance(NorFileName, int):
self.constructor1(NorFileName, corpus)
else:
self.constructor2(NorFileName)
- Extend test classes from unittest and use separate unit test methods.
class NGramTest(unittest.TestCase):
def test_GetCountSimple(self):
- For undefined types use object as type in the type declarations.
cdef class WordNet:
cdef object __syn_set_list
cdef object __literal_list
- For boolean types use bint as type in the type declarations.
cdef bint is_done
- Enumerated types should be used when necessary as enum classes, and should be declared in py files.
class AttributeType(Enum):
"""
Continuous Attribute
"""
CONTINUOUS = auto()
"""
- Resource files should be taken from pkg_recources package.
fileName = pkg_resources.resource_filename(__name__, 'data/turkish_wordnet.xml')