Identification of Causal Dependencies by using Natural Language
Processing: A Survey
Erika Nazaruka
a
Department of Applied Computer Science, Riga Technical University, Sētas iela 1, Riga, Latvia
Keywords: System Modelling, Knowledge Extraction, Natural Language Processing, Topological Functioning Model.
Abstract: Identification of cause-effect relations in the domain is crucial for construction of its correct model, and
especially for the Topological Functioning Model (TFM). Key elements of the TFM are functional
characteristics of the system and cause-effect relations between them. Natural Language Processing (NLP)
can help in automatic processing of textual descriptions of functionality of the domain. The current research
illustrates results of a survey of research papers on identification and extracting cause-effect relations from
text using NLP and other techniques. The survey shows that expression of cause-effect relations in text can
be very different. Sometimes the same language constructs indicate both causal and non-causal relations.
Hybrid solutions that use machine learning, ontologies, linguistic and syntactic patterns as well as temporal
reasoning show better results in extracting and filtering cause-effect pairs. Multi cause and multi effect
domains still are not very well studied.
1 INTRODUCTION
Models are known from ancient times. Models are
built for a specific purpose, and this determines their
level of abstraction, accuracy, representation means,
scale etc.
Traditional industries use graphical models for
design and experiments to predict how a new product
will function in the real circumstances. Software
development models are mostly textual starting from
requirements specifications and ending with the
source code. Graphical models are used mostly to
simplify understanding of key characteristics of the
product.
The idea of using models as a core element of
software development was met with interest when
The Object Management Group had published their
guide on Model Driven Architecture (MDA) in 2001
(Miller and Mukerji, 2001). MDA suggests using a
chain of model transformations, namely, from a
computation independent model (CIM, mostly
textual) to a platform independent model (PIM,
mostly graphical), then to a platform specific model
(PSM, graphical) and to source code. The weaker
place in this chain of transformations is the CIM, and
its transformations. The CIM is dedicated for
a
https://orcid.org/0000-0002-1731-989X
presentation of software requirements, business
requirements, knowledge about the system domain,
business rules, etc. The main task here is to process
textual descriptions, graphical information, discover
implicit knowledge, or, in other words, to find out all
knowledge about system (software) functioning,
behavior and structure. Analysis of the available
information includes so called causal reasoning
(Khoo et al., 2002). Identified causal dependencies
are those of control flows in the systems and influence
also some structural relations.
In our vision of implementation of MDA
principles, we suggest using a knowledge model
based on the Topological Functioning Model (TFM)
as the CIM to generate code via an intermediary
model – Topological UML model (Osis and Donins,
2017). The TFM elaborated by Janis Osis at Riga
Technical University (Osis, 1969) specifies a
functioning system from three viewpoints –
functional, behavioural and structural. Causal
dependencies are the key element in the Topological
Functioning Model.
Construction of the TFM is based on analysis of
verbal descriptions – instructions, interview
protocols, position descriptions, or other experts’
knowledge expressed in text. At the present we have
Nazaruka, E.
Identification of Causal Dependencies by using Natural Language Processing: A Survey.
DOI: 10.5220/0007842706030613
In Proceedings of the 14th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2019), pages 603-613
ISBN: 978-989-758-375-9
Copyright
c
2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
603
two approaches:
manual processing of the text in the TFM4MDA
(Topological Functioning Model for MDA)
approach (Osis et al., 2007a; Osis and Asnina,
2011b) and
automated processing of steps in use case
scenarios in the IDM (Integrated Domain
Modelling) toolset (Osis and Slihte, 2010; Slihte
et al., 2011).
In practice, preparation of text and manual knowledge
acquisition are too resource-consuming (Elstermann
and Heuser, 2016). It is better either to skip the step
of preparation of the textual description and start from
human analysis of the available information, either to
automate or semi-automate this process.
As metnioned, certain causal dependencies in a
domain form control and message flows in a software
system. This relates not only to the TFM (Nazaruka,
2017), but also to other models used in the
transformations from CIM to (Kardoš and Drozdová,
2010; Kriouile et al., 2013; Kriouile et al., 2014;
Kriouile et al., 2015; Bousetta et al., 2013; Rhazali et
al., 2016; Essebaa and Chantit, 2016).
The key aspect of successful construction of the
domain model is correct and complete identification
of causes and effects. In software models, relations
between causes and effects are implemented as
control flows, message flows, transitions between
states of the system. Not less important it is in case of
the TFM construction, where identification of causal
dependencies (topological relations) between
functional characteristics of the system is crucial. The
open question is how to identify and extract these
relations from textual descriptions in the automated
way that Natural Language Processing (NLP) tools
suggest.
The goal of the given research is to make a survey
on ways and completeness of extracting causal
dependencies from text using NLP, Natural Language
Understanding (NLU) and linguistics techniques.
The paper is organized as follows. Section 2
describes the purpose of the research and enumerates
the research questions. Section 3 presents overview
of research results on identification and extracting
cause-effect relations from text. Section 4 presents a
discussion on findings. Section 5 concludes the paper
with discussion on main results and future research
directions.
2 BACKGROUND AND
RESEARCH QUESTIONS
This section discusses the background on cause-effect
relations in the TFM, i.e. the brief overview of the
TFM as well as formal definitions of cause-effect
relations are presented. At the end of the section
research questions are formulated.
2.1 Cause-effect Relations in the TFM
The TFM is a formal mathematical model that
represents system functionality in a holistic manner.
It describes the functional and structural aspects of
the software system in the form of a directed graph
(X, θ), where a set of vertices X depict functional
characteristics of the system named in human
understandable language, while θ is a set of edges
that depict cause-effect relations (topology) between
them. Such specification is more perceived, precise
and clearer then the large textual descriptions. The
TFM is characterized by the topological and
functioning properties (Osis and Asnina, 2011a). The
topological properties are connectedness,
neighbourhood, closure, and continuous mapping.
The functioning properties are cause-effect relations,
cycle structure, inputs, and outputs. The composition
of the TFM is presented in (Osis and Asnina, 2011b).
Rules of composition and derivation processes
within TFM4MDA from the textual description of
system functionality is provided by examples and
described in detail in (Asnina, 2006; Osis et al.,
2007b; Osis et al., 2008; Osis and Asnina, 2011b).
The TFM can also be generated automatically from
the business use case scenario specifications, which
can be specified in the IDM toolset (Šlihte and Osis,
2014). It also can be manually created in the TFM
Editor from the IDM toolset.
The cause-effect relations are those of causal
dependencies that exist between functional
characteristics of the system and define the cause
from which the triggering of the effect occurs. In fact,
this kind of relations indicates control flow transition
in the system. For instance, termination of execution
of a functional characteristic triggers initiation of
execution of related characteristics (Figure 1).
Formal definitions of a cause-effect relation and a
logical relation among those relations as well as their
incoming and outgoing groups are as follows (Asnina
and Ovchinnikova 2015; Osis and Donins 2017).
MDI4SE 2019 - Special Session on Model-Driven Innovations for Software Engineering
604
Figure 1: Execution of functional feature instances
(Nazaruka et al., 2016).
Formal Definition of a Cause-Effect Relation. A
cause-and-effect relation T
i
is a binary relationship
that relates exactly two functional features X
c
and X
e
.
Both X
c
and X
e
may be the same functional feature in
case of recursion. The synonym for cause-effect
relation is a topological relationship. Each cause-
effect relation is a unique 5-tuple (1):
T
i
= <ID, X
c
, X
e,
N
, S>, whe
r
e (1)
ID is a unique identifier of the relation;
X
c
is a cause functional feature;
X
e
is an effect functional feature;
N is a Boolean value of the necessity of X
c
for
generating X
e
;
S is a Boolean value of the sufficiency of X
c
for
generating X
e
.
Formal Definition of a Logical Relation. A logical
relation L
i
specifies the logical operator conjunction
(AND), disjunction (OR), or exclusive disjunction
(XOR) between two or more cause-effect relations T
i.
The logical relation denotes system execution
behaviour, e.g. decision making, parallel or
sequential actions. Each logical relation is a unique 3-
tuple (2):
L
i
= <ID, T, R
T
>, where (2)
ID is a unique identifier of the relation;
T is a set of cause-effect relations {T
i
, ..., T
n
}
that
participate in this logical relation;
R
T
is a logical operator AND, OR, or XOR over
T; operator OR is a default value.
Formal Definition of Incoming Topological
Relations. A set of logical relations that join cause-
and-effect relations, which go into functional feature
X
i
, is defined as a subset L
in
of set L = {L
i
, ..., L
n
},
where at least one topological relation T
i
such that its
effect functional feature X
e
is equal to X
i
is found in
set T of topological relations in each logical relation
L
i
.
Formal Definition of Outgoing Topological
Relations. A set of logical relations that join cause-
and-effect relations, which go out from functional
feature X
i
, is defined as a subset L
out
of set L = {L
i
,
..., L
n
}, where at least one topological relation T
i
such
that its cause functional feature X
c
is equal to X
i
is
found in set T of topological relations in each logical
relation L
i
.
The connection between a cause and an effect is
represented by a certain conditional expression, the
causal implication. It is characterized by the nature or
business laws (or rules) not by logic rules. In causal
connections “something is allowed to go wrong”,
whereas logical statements allow no exceptions.
Using this property of cause-effect relations a logical
sequence, wherein the execution of the precondition
guarantees the execution of the action, can be
prescinded, this means that even if a cause is
executed, the corresponding effect can be not
generated because of some functional damage.
The human mind applies very sophisticated
mechanism as well as empirical information and
world knowledge to construct “a theory of the causal
mechanism that produced the effect” (Khoo et al.
2002). Trying to discover this “causal mechanism”
they analyse “causal power” of events to generate an
effect.
Since a cause generates an effect, a cause
chronologically precedes an effect. This means that
the cause-effect conditions contain a time dimension.
Causes can be sufficient or necessary (or complete
or partial, correspondingly). A sufficient (complete)
cause generates its effect ever, or in any conditions.
On the other hand, a necessary cause (partial) only
promotes its effect generating, and this effect is
realized only if this partial cause joins other
conditions. However, most cause-effect relations
involve multiple factors. Sometimes there are factors
in series. Sometimes there are factors in parallel. In
case of the TFM, it is assumed that a deal is always
with necessary causes as the functionality of the
system has its known and unknown risks at the time
of analysis.
A cause not only precedes an effect and always is
followed by it, it causes (gives rise to, generates) and
is condition on an effect. The concept of causing
(generating) is necessary to distinguish a cause-and-
effect relation from the simple consequence that is not
causal. The causality is universal. This means that
there is no such a problem domain where no causes
and effects. The person can see nothing, but a cause
or an effect exists.
A structure of cause-effect relations can form a
causal chain. The causal chain begins with the first
cause and follows with series of intermediate actions
Identification of Causal Dependencies by using Natural Language Processing: A Survey
605
or events to some final effect. Though one link may
not be as important or as strong like the other ones,
they are all necessary to the chain. If just one of these
intermediate causes is absent, then the final effect
would not be reached. Additionally, even if you
change something, you cannot remove the effect
without removing or changing the cause.
2.2 Research Questions
Identification of cause-effect relations in a domain is
a key element in the construction of the valid model
of the domain. Sources of information about the
domain can differ, and a human mind works using
visual and audio information as well as its own
background knowledge on the domain and the world.
In case of automation of such analysis, most of
information must be transformed into the verbal form
to be able to use NLP tools.
NLP tools are able to perform the following tasks:
tokenization, part-of-speech (POS) tagging,
chunking, Name Entity Recognition
(NER)/Classification, dependency analysis,
constituency parsing, and coreference resolution. So,
the result of the processing can be analysed further to
identify causes and effects in the text.
The research questions are the following:
RQ1: What natural language constructs for
expressing cause-effect relations in text are used?
RQ2: What models, patterns for identification of
cause-effect relations in text are used?
RQ3: What automatic techniques for extracting
cause-effect relations from text are used?
The aim is to understand what natural language
constructs may be ambiguous for NLP, what pitfalls
exist in discovering cause-effect relations in text and
what issues have been found in application of
extracting tools.
3 CAUSE-EFFECT RELATIONS
EXPRESSED IN TEXT
This section represents discussion on means for
explicit and implicit expressing cause-effect relations
in natural language, what patterns may indicate
cause-effect relations in text at the sentence and
discourse levels, as well as overview of research
papers on automatic discovering cause-effect
relations from text using NLP tools and other
techniques.
3.1 Means for Expressing Cause-effect
Relations in Text
Considering natural language processing and
understanding tasks, researchers noted that causes
and effects usually are states or events that can have
different duration (Khoo et al., 2002; Solstad and
Bott, 2017). The cause-effect relations in text may be
expressed both explicitly and implicitly. And the very
important aspect is similar language constructs used
to express temporal and causal relations, and
sometimes only temporal relations dictate the causal
one (Ning et al., 2018).
3.1.1 Explicitly Expressed Relations
Several authors (Khoo et al., 2002; Solstad and Bott,
2017) mentioned that linguists have identified the
following ways of explicitly expressing causes and
effects:
using causal links to link two phrases, clauses or
sentences,
using causative verbs,
using resultative constructions,
using conditionals (i.e., if…then constructions),
using causative adverbs, adjectives, and
prepositions.
As Khoo et al., (2002) stated, Altenberg (1984) had
classified causal links into four main types:
the adverbial link (e.g., so, hence, therefore). It
can have a reference to the preceding clause or to
the following clause;
the prepositional link (e.g., because of, on account
of). It connects a cause and an effect in the same
clause;
subordination. It can be expressed using a
subordinator (e.g., because, as, since), a
structural link marked by a non-finite -ing clause,
and a correlative comparative construction (e.g.,
so…that);
the clause-integrated link (e.g. that’s why, the
result was). Here they distinguish thematic link¸
when the linking words serve as a subject of the
sentence, and a rhematic link, when the linking
words serve as the complement of the verb.
One may say that causal links include as causal
reasons as temporal reasons (Ning et al., 2018).
Causative verbs are “verbs the meaning of which
include a causal element” (Khoo et al., 2002), e.g.
“register” that in “X registers Y” means that “X
causes Y to be registered”. One of the working
definitions of the causative verbs can be such that
“Causative verbs specify the result of the action,
MDI4SE 2019 - Special Session on Model-Driven Innovations for Software Engineering
606
whereas other action verbs specify the action but not
the result of action” (Khoo et al., 2002).
A resultative construction (Khoo et al., 2002) is “a
sentence in which the object of a verb is followed by
a phrase decribing the state of the object as a result of
the action denoted by the verb”, e.g. “A user marked
records yellow”. A resultative phrase can be an
adjective (the most common kind), a noun phrase, a
prepositional phrase or a particle.
If-then conditionals ofthen indicate that the
antecedent (the if part) causes the consequent (the
then part). However, sometimes they just indicate a
sequence of events not their cause-effect relation
(Khoo et al., 2002).
Causative adverbs, adjectives and prepositions
also can have a causal element in their meaning
(Khoo et al., 2002). Causative adverbs can be
different, the most interesting for us are adverbs that
involve the notion of a result whose properties are
context dependent (e.g. successfully), adverbs that
refer not to causes but to effects (e.g. consequentelly),
and adverbs of means (e.g. mechanically).
Causal adverbs and adjectives are not well studied
(Khoo et al., 2002).
As Khoo et al., (2002) mentioned, prepositions
also can be used to express causality. They can
indicate a cause as proximity, a cause as a source, and
a cause as volume.
3.1.2 Implicitly Expressed Relations
Implicit cause-effect relations usually are inferred by
the reader using information in text and their
background knowledge (Khoo et al. 2002; Solstad
and Bott 2017; Ning et al. 2018). As Khoo et al. stated
(Khoo et al. 2002) implicit causality can be inferred
by several groups of verbs that “have “causal
valence” – they tend to assign causal status to their
subject or object”. The authors referred to Corrigan’s
work (Corrigan 1993; Corrigan and Stevenson 1994),
where the following groups of verbs had been
identified:
Experiential verbs (Experiencer-stimulus and
Stimulus-Experiencer),
Action verbs (Actor verbs and Non-actor
verbs).
Experienced verbs describe someone having a
particular psychological or mental experience.
Therefore, experienced verbs can be skipped in the
system analysis for software development.
Action verbs describe events. The subject of the
verb can take the semantic role agent or actor. The
object of the verb takes the role of patient. Some
verbs give greater causal weight to the subject (actor
verbs), other – to the object (non-actor verbs). At the
moment, causal weigth seems not so important for
domain analysis. However, the interesting thing is
that both verbs have derived adjectives reffering to
the subject or object. This means that some preceding
actions can be expressed in text using not verbs, but
adjectives. Some implicit causative verbs trigger
expectations of explanations to occur in subsequent
discourse (Solstad and Bott, 2017).
3.2 Identification of Cause-effect
Relations in Text
Many theories exist for identification, modeling and
analysis of cause-effect relations in psycholinguistics,
linguistics, psychology and artificial intelligence.
Those of theories attempting to reduce causal
reasoning to a domain-general theory can be grouped
as associative theories, logical theories and
probabilistic theories (Waldmann and Hagmayer
2013).
Waldmann and Hangmayer (2013) stated that
associative theories underestimate aspects of
causality that are important for causal reasoning,
however in some cases causes and effects can be
identified only using associations.
Logical theories model causal reasoning as a
special case of deductive reasoning. Waldmann and
Hangmayer (2013) noted that conditionals do not
distinguish between causes and effects, and
background knowledge can be necessary to
distinguish them as well as some temporal priorities.
Probabilistic theories considers causes as
“difference makers, which raise (generative cause) or
reduce (preventive cause) the probability of the
effect (Waldmann and Hagmayer 2013). However,
as the authors noted, covariation does not necessarily
reflect causation.
All the theories have their limitations in
identification of causes and effects. In case of
processing verbal (written) information to develop
software, causes and effects mostly relate to business,
mechanical and physical domains that certainly make
a task easier for developers. At the present, logical
theories seem to be the most suitable for this task and
domains.
3.2.1 Verbal (Sentence) Domain
Solstad and Bott (2017) stated that verbs, as causative
as action, indicate a cause between two events (3).
[[event1]] CAUSE [[event2]] (3)
Identification of Causal Dependencies by using Natural Language Processing: A Survey
607
Where [[event1]] is when the subject does something,
and [[event2]] is when the object changes its state.
Besides that, it is inferred that the object wasn’t in this
state before the [[event1]] if otherwise is not
mentioned in text.
Causing entities and the manner of the causing
may be introduced using “by” phrases (Solstad and
Bott, 2017).
As Khoo et al. mentioned (Khoo et al., 2002), the
subject of the verb describing the event must be some
agent or actor. It needs not be obligatory an animate
agent, it may be an object, an abstract property, or an
event (Solstad and Bott, 2017). Implicit causality
verbs in most cases express causality between two
animate objects followed by explanation (Solstad and
Bott, 2017).
3.2.2 Discourse Domain
Solstad and Bott (2017) stated that causal relations
such as explanations are crucial for understanding of
discourse. Connections between causal relations may
be expressed explicitly using causal links (Section
3.1) or implicitly. In the latter they must be inferred
by the reader.
Some researchers (Kang et al., 2017; Solstad and
Bott, 2017) indicate that at the level of discourse, the
causal relations differ from thouse of at the sentence
or clause level. At the discourse level they are
supplemented with reasons and explanations. The
authors assumed that the causal relations exist
between entities that are propositional in nature:
[[proposition1]] CAUSE [[proposition2]].
Sometimes, it is hard to understand are they express
parallel or sequential propositions or explanations, as,
for instance, in sentence “The user access was denied.
The hackers taked the control.”
Solstad and Bott stated that in case of implicit
causality verb and discourse domains are mixed,
where causal expressions connect causative verbs
with reasoning and explanations within the same
sentence (Solstad and Bott, 2017).
3.2.3 Conditionals
Conditionals do not express causal relations
explicitly, but they involve causal models in their
evaluation. Solstad and Bott (2017) mentioned that
If…then constructs (i.e., conditionals) may form
constructs hard for NLP analysis – the so-called
counterfactual conditionals. They include such
constructs as might, would, if only. They indicate
possible “state of the world” in case of some “action”
that would be done. For example, as in the sentence
“If librarian would not have ordered the book, a
manager assistant would have”.
From one’s viewpoint, such constructs must be
avoided in the description of system functionality.
However, counterfactual conditions may be used in
expert systems to produce answers to queries of
interest (Pearl, 2019).
3.2.4 Causal and Temporal Reasons
Ning et al., (2018) indicated that many of NLP
research papers focus on the abovementioned
language constructs (causative verb, causal links,
discourse relations, etc.) skipping (or
underestimating) temporal reasons. They believe that
joint consideration of causal models and temporal
models is more valuable for identifying and
extracting cause-effect relations from text. The
authors indicated that starting from 2016 researchers
pay greater attention to this aspect (Mirza, 2014;
Asghar, 2016; Mostafazadeh et al., 2016; Ning et al.,
2018).
The interesting fact is that joint temporal and
causal reasoning correctly identify counterfactual
clauses (Ning et al., 2018).
3.3 Automated Extraction of
Cause-effect Relations using NLP
Cause-effect relations are extracted using so-called
linguistic and syntactic patterns that in most cases are
created manually (at least at the beginning).
Linguistic and syntactic patterns are based on means
for explicit expressing causes and effects, e.g. causal
links and causative verbs for linguistic patterns and
verb phrases and noun phrases for syntactic patterns
(Blanco et al., 2008; Ning et al., 2018; Mirza, 2014;
Blass and Forbus, 2016).
Joint usage of both temporal reason model and
causal model as well as Machine Learning (ML) are
presented by several authors (Mirza, 2014; Ning et
al., 2018).
The temporal model discovers a temporal relation
between two events. The temporal relation can be
annotated as before, after, include, is_included, vague
(Ning et al., 2018), and simultaneous, begins/
begun_by, ends/ended_by, during/during_inv,
identity (Mirza, 2014). Other authors (Mostafazadeh
et al., 2016) use another annotations, i.e., before,
meets, overlaps, finishes, starts, contain and equals.
Their model distinguishes between a precondition
and a cause. The causal models of all the mentioned
authors discover causal relations between events
using linguistic patterns. The authors state that
MDI4SE 2019 - Special Session on Model-Driven Innovations for Software Engineering
608
analysis of both relation types allows extracting
cause-effect relations even if they lack explicit causal
reason.
A set of logical rules and Bayesian inference (in
ML) are used by Sorgente, Vettigli and Mele
(Sorgente et al., 2013; Sorgente et al., 2018).
Bayesian inference is applied to exclude discovered
cause-effect pairs that in essence are non-causal.
Filtering takes into account such features as lexical
features, semantics features (hyponyms and
synonyms) and dependency features.
A comprehensive survey of automatic extraction
of causal relations is presented by Asghar (2016). The
author divided automatic methods into two groups:
approaches that employ linguistic, syntactic and
semantics patter matching, and
techniques based on statistical methods and ML.
The first group started from small text fragments and
evolved till large text corpuses. As Asghar stated the
first group at their beginning was forced to prepare
text fragments manually for automatic processing,
like, for instance, in Blass’ and Forbus’ work (2016).
However, now this group uses NLP techniques to
prepare cause-effect pairs (by using linguistic
patterns) and then filtering them to reduce a number
of non-causal pairs. Starting from the early 2000s,
ML techniques have been used to gain extracting
cause-effect relations. These techniques do not
require a large set of manually predefined linguistic
patterns, however, quality of learning depends on
corpuses used for learning.
4 DISCUSSION
Summarizing results (Table 1), we can conclude that
the larger number of overviewed research papers is
focused on analysis of explicitly expressed cause-
effect relations by using causal links and causative
verbs (Sorgente et al., 2013; Sorgente et al., 2018;
Asghar, 2016; Blanco et al., 2008; Mirza, 2014;
Mostafazadeh et al., 2016). However, a few research
papers pay their attention also to resultative
constructions and causative prepositions. In other
words, causal links, causative verbs and prepositions
are more valuable for creation of linguistic/syntactic
patterns for text processing. The advantage is small
cost of preparation, while the result can be quite
ambiguous.
Causative adverbs and adjectives up to 2018 are
not well studied comparing to the main studies on
causative verbs, causal links and temporal aspects of
events and propositions.
According to survey in (Asghar, 2016), accuracy
of results of extracting if…then conditionals is
satisfactory only using ML techniques
Extracting multiple causes and effects is very
domain-specific task, therefore only a few research
solve it directly (Sorgente et al., 2013; Sorgente et al.,
2018; Mueller and Hüttemann, 2018).
Cause-effect relations implicitly expressed by
action verbs are analysed in the same group of
causative verbs. While automated analysis of
counterfactual conditionals is a quite hard task and
some results are presented just in a few papers (Ning
et al., 2018; Pearl, 2019).
Speaking about techniques used for automated
cause-effect extracting, it could be found from results
in Table 1 that preparation of text corpuses using
predefined syntactic and linguistic patterns is less
costly than using supervised ML techniques.
However, the use of patterns limits discovered types
of cause-effect relation only to these patterns. While
ML enables discovering of much more cause-effect
relations.
Filtering and statistical inferencing may be
considered as a less expensive solution in comparison
with ML techniques. However, some linguistic
constructs may be ambiguous and, thus, results may
differ from the desired ones. But statistical
inferencing requires large datasets.
Ontology banks are also used, but moslty
WordNet. VerbNet, PropBank and FrameNet are
used sparsely.
The more successful results are shown by hybrid
solutions where patterns, temporal reasons, ML and
ontologies are presented. The limitation of the hybrid
solutions is a lack of enough text corpuses for
learning.
TFM construction requires processing verbal
descriptions of the modelled environment. The
descriptions contain information on system
functioning within and interacting with its
environment. Identification and extraction of causes
and effects, as well as their relations, are vital for
correct identification and specification of system’s
functional chacacteristics and causal dependencies
between them.
Most of researh papers investigates cases with one
cause and one (or two) effects. The TFM may have
relationships between causal relations.
So, multi causes and multi effects must be
idenified and extracted from text. However, there is
just a few research papers presenting results on this,
since this is a quite hard task.
It is clear that the starting point for extracting
cause-effect relations from the descriptions of func-
Identification of Causal Dependencies by using Natural Language Processing: A Survey
609
Table 1: Automated extracting cause-effect relations using NLP and other techniques.
Lexical mark / discourse relation Sou
r
ce Pros and cons
Explicit cause-effect relations
Causal links:
- the adverbial link
- the prepositional link
- subordination
- the clause-integrated lin
k
(Sorgente et al., 2013; Sorgente et
al., 2018), survey in (Asghar, 2016);
(Mirza, 2014; Blanco et al., 2008;
Mostafazadeh et al., 2016)
Pros: small costs of preparation
Cons: huge number of potential patterns, ambiguity
Causative verbs
(Sorgente et al., 2013; Sorgente et
al., 2018), survey in (Asghar, 2016;
Mostafazadeh et al., 2016)
Pros: small costs of preparation
Cons: huge number of potential patterns, ambiguity
Resultative constructions survey in (Asgha
r
, 2016)
Causative adverbs
Causative adjectives
Causative prepositions:
- cause as proximity
- cause as source
- cause as volume
(Sorgente et al., 2013; Sorgente et
al., 2018; Blanco et al., 2008)
Pros: small costs of preparation
Cons: huge number of potential patterns, ambiguity
If-then conditionals survey in (Asghar, 2016)
Cons: accuracy is satisfactory only using ML
techniques
Multiple causes and effects
(conjunctions)
(Sorgente et al., 2013; Sorgente et
al., 2018; Mueller and Hüttemann,
2018)
I
mplicit cause-effect relations
Counterfactual conditionals (Ning et al., 2018; Pearl, 2019) Use for prediction
Action verbs Consider as a subset of causative verbs
Techniques
Temporal relations
(Mirza, 2014; Asghar, 2016;
Mostafazadeh et al., 2016; Ning et
al., 2018)
Pros: Analysis of event/proposition pairs where
causality is implicit.
Cons: Some linguistic constructs may be
ambiguous.
Linguistic/syntactic patterns
(Ning et al., 2018), (Sorgente et al.,
2013; Sorgente et al., 2018), survey
(Asghar, 2016), (Blass and Forbus,
2016)
Pros: Does not require large corpuses of text for
learning, domain-independent.
Cons: Limited to the manually predefined set of
patterns and propositions. A use of explicit causal
indicators and in most cases ignoring implicit
causalities.
Filtering (Bayesian inference,
WordNet-based filtering, semantic
filtering based on verb’s senses)
(Sorgente et al., 2013; Sorgente et
al., 2018), survey in (Asghar, 2016)
Pros: reducing a number of non-causal pairs.
Cons: a set of pairs depends on a set of linguistic
patterns
Machine Learning
survey in (Asghar, 2016), (Blanco et
al., 2008; Mirza, 2014; Ning et al.,
2018)
Pros: discovering implicit causality, analysis of
ambiguous constructs, pre-conditions and
postconditions
Cons: requires large corpuses of text, may be
domain-specific.
Statistical inferencing survey in (Asgha
r
, 2016) Cons: requires large datasets
Ontology banks
survey in (Asghar, 2016),
(Mostafazadeh et al., 2016; Kang et
al., 2017)
Pros: WordNet is used frequently
Cons: VerbNet, PropBank and FrameNet are used
sparsely
tionality must be preparation of a corpus of linguistic/
syntactic patterns as well as more thorough analysis
of if…then conditionals.
A use of temporal models, filtering and ontology
banks seems more promising than a use of ML
techniques since each problem domain will certainly
have its own unique characteristics, but at the same
time the diversity in description of functionality is not
so defining than in descriptions of world phenomena.
5 CONCLUSIONS
The results of the survey show that identification of
causes and effects as well as their relations can be
based first on linguistic/syntactic patterns and
temporal reason models. However, main
disadvantage of the patterns is that it is not possible
to identify all patterns for all types of cause-effect
relations. The expression means of the natural
language differ more than any set of predefined rules.
MDI4SE 2019 - Special Session on Model-Driven Innovations for Software Engineering
610
Besides that, one and the same pattern may be applied
for both causal and non-causal relations. As well as
not all linguistic and syntactic patterns have been
researched, e.g. causal adverbs and adjectives.
Filtering can help to solve this issue but is also limited
to the known non-causal constructs. A use of
temporal relation models is more valuable solution of
this issue. However, some discourse descriptions may
be very ambiguous, and there is no a guarantee that
temporal relations will be identified correctly.
The more expensive and more flexible solutions
are those of hybrid using machine learning, ontology
banks and statistics. However, these solutions are
more domain specific. They require a large amount of
text corpuses for supervised learning of models. This
could be a challenge, since not all the domains have
them.
There are two clear trends in cause-effect relation
extraction. The first is increasing the accuracy of the
results using ontology banks, machine learning and
statistical inferring. The second is decreasing the cost
of these activities. The main challenge for
construction of software models is a lack of a large
amount of text corpuses and statistical datasets for
potential problem domains. However, the potential
positive aspect is that source documents for
construction of software models may be limited to
specifications (requirements, scenario, etc.) having
less variability in expressing causality.
The future research direction is related to
implementation of extracting causes and effects from
the description of system functioning. The first step is
to define a list of more frequent (potential)
linguistic/syntactic patterns of causal dependencies in
descriptions of system functioning. One of the very
important aspects here is discovering of multi causes
and multi effects. The accuracy of the obtained results
will lead to the second step, i.e., to finding a solution
that will show acceptable accuracy of results and will
not be very expensive.
REFERENCES
Altenberg, B., 1984. Causal linking in spoken and written
English. Studia Linguistics, 38(1), pp.20–69.
Asghar, N., 2016. Automatic Extraction of Causal
Relations from Natural Language Texts: A
Comprehensive Survey. CoRR, abs/1605.0. Available
at: http://arxiv.org/abs/1605.07895.
Asnina, E., 2006. The Computation Independent
Viewpoint: a Formal Method of Topological
Functioning Model Constructing. Applied computer
systems, 26, pp.21–32.
Asnina, E. and Ovchinnikova, V., 2015. Specification of
Decision-making and Control Flow Branching in
Topological Functioning Models of Systems. In
International Conference on Evaluation of Novel
Approaches to Software Engineering (ENASE), 2015.
Barcelona, Spain: SciTePress, pp. 364–373.
Blanco, E., Castell, N. and Moldovan, D., 2008. Causal
Relation Extraction. In Proceedings of the Sixth
International Conference on Language Resources and
Evaluation (LREC’08). European Language Resources
Association (ELRA), pp. 28–30.
Blass, J.A. and Forbus, K.D., 2016. Natural Language
Instruction for Analogical Reasoning: An Initial
Report. In Workshops Proceedings for the Twenty-
fourth International Conference on Case-Based
Reasoning (ICCBR 2016). pp. 21–30.
Bousetta, B., Beggar el, O. and Gadi, T., 2013. A
methodology for CIM modelling and its transformation
to PIM. Journal of Information Engineering and
Applications, 3(2), pp.1–21.
Corrigan, R., 1993. Causal attributions to the states and
events encoded by different types of verbs. British
Journal of Social Psychology, 32, pp.335–348.
Corrigan, R. and Stevenson, C., 1994. Children’s causal
attribution to states and events described by different
classes of verbs. Cognitive Development, 9, pp.235–
256.
Elstermann, M. and Heuser, T., 2016. Automatic Tool
Support Possibilities for the Text-Based S-BPM
Process Modelling Methodology. In Proceedings of the
8th International Conference on Subject-oriented
Business Process Management - S-BPM ’16. New
York, New York, USA: ACM Press, pp. 1–8.
Essebaa, I. and Chantit, S., 2016. Toward an automatic
approach to get PIM level from CIM level using QVT
rules. In 2016 11th International Conference on
Intelligent Systems: Theories and Applications (SITA).
Mohammedia: IEEE, pp. 1–6.
Kang, D. et al., 2017. Detecting and Explaining Causes
From Text For a Time Series Event. In Proceedings of
the 2017 Conference on Empirical Methods in Natural
Language Processing. The Association for
Computational Linguistics, pp. 2758–2768.
Kardoš, M. and Drozdová, M., 2010. Analytical method of
CIM to PIM transformation in model driven
architecture (MDA). Journal of Information and
Organizational Sciences, 34(1), pp.89–99.
Khoo, C., Chan, S. and Niu, Y., 2002. The Many Facets of
the Cause-Effect Relation. In R. Green, C. A. Bean, and
S. H. Myaeng, eds. The Semantics of Relationships: An
Interdisciplinary Perspective. Dordrecht: Springer
Netherlands, pp. 51–70.
Kriouile, A., Addamssiri, N., Gadi, T. and Balouki, Y.,
2014. Getting the static model of PIM from the CIM. In
2014 Third IEEE International Colloquium in
Information Science and Technology (CIST). Tetouan:
IEEE, pp. 168–173.
Kriouile, A., Addamssiri, N. and Gadi, T., 2015. An MDA
Method for Automatic Transformation of Models from
Identification of Causal Dependencies by using Natural Language Processing: A Survey
611
CIM to PIM. American Journal of Software
Engineering and Applications, 4(1), pp.1–14.
Kriouile, A., Gadi, T. and Balouki, Y., 2013. CIM to PIM
Transformation: A Criteria Based Evaluation. Int. J.
Computer Technology and Applications, 4(4), pp.616–
625.
Miller, J. and Mukerji, J., 2001. Model Driven Architecture
(MDA), Available at: http://www.omg.org/cgi-
bin/doc?ormsc/2001-07-01.
Mirza, P., 2014. Extracting Temporal and Causal Relations
between Events. In Proceedings of the ACL 2014
Student Research Workshop. Baltimore, Maryland,
USA: Association for Computational Linguistics, pp.
10–17.
Mostafazadeh, N., Grealish, A., Chambers, N., Allen, J. and
Vanderwende, L., 2016. CaTeRS: Causal and Temporal
Relation Scheme for Semantic Annotation of Event
Structures. In Proceedings of the Fourth Workshop on
Events. San Diego, California: Association for
Computational Linguistics, pp. 51–61.
Mueller, R. and Hüttemann, S., 2018. Extracting Causal
Claims from Information Systems Papers with Natural
Language Processing for Theory Ontology Learning. In
Proceedings of the 51st Hawaii International
Conference on System Sciences (HICSS). Hawaii, USA:
IEEE Computer Society Press.
Nazaruka, E., 2017. Meaning of Cause-and-effect Relations
of the Topological Functioning Model in the UML
Analysis Model. In Proceedings of the 12th
International Conference on Evaluation of Novel
Approaches to Software Engineering. SCITEPRESS -
Science and Technology Publications, pp. 336–345.
Nazaruka, E., Ovchinnikova, V., Alksnis, G. and
Sukovskis, U., 2016. Verification of BPMN Model
Functional Completeness by using the Topological
Functioning Model. In Proceedings of the 11th
International Conference on Evaluation of Novel
Software Approaches to Software Engineering.
Portugal: SCITEPRESS - Science and and Technology
Publications, pp. 349–358.
Ning, Q., Feng, Z., Wu, H. and Roth, D., 2018. Joint
Reasoning for Temporal and Causal Relations. In
Proceedings of the 56th Annual Meeting of the
Association for Computational Linguistics (Long
Papers). Melbourne, Australia: Association for
Computational Linguistics, pp. 2278–2288.
Osis, J., 1969. Topological Model of System Functioning
(in Russian). Automatics and Computer Science, J. of
Academia of Siences, (6), pp.44–50.
Osis, J. and Asnina, E., 2011a. Is Modeling a Treatment for
the Weakness of Software Engineering? In Model-
Driven Domain Analysis and Software Development.
Hershey, PA: IGI Global, pp. 1–14.
Osis, J. and Asnina, E., 2011b. Topological Modeling for
Model-Driven Domain Analysis and Software
Development: Functions and Architectures. In Model-
Driven Domain Analysis and Software Development:
Architectures and Functions. Hershey, PA: IGI Global,
pp. 15–39.
Osis, J., Asnina, E. and Grave, A., 2007a. Formal
computation independent model of the problem domain
within the MDA. In J. Zendulka, ed. Proceedings of the
10th International Conference on Information System
Implementation and Modeling, Hradec nad Moravicí,
Czech Republic, April 23-25, 2007. Jan Stefan MARQ.,
pp. 47–54.
Osis, J., Asnina, E. and Grave, A., 2007b. MDA oriented
computation independent modeling of the problem
domain. In Proceedings of the 2nd International
Conference on Evaluation of Novel Approaches to
Software Engineering - ENASE 2007. Barcelona:
INSTICC Press, pp. 66–71.
Osis, J., Asnina, E. and Grave, A., 2008. Formal Problem
Domain Modeling within MDA. In J. Filipe et al., eds.
Software and Data Technologies: Second International
Conference, ICSOFT/ENASE 2007, Barcelona, Spain,
July 22-25, 2007, Revised Selected Papers. Berlin,
Heidelberg: Springer Berlin Heidelberg, pp. 387–398.
Osis, J. and Donins, U., 2017. Topological UML modeling:
an improved approach for domain modeling and
software development, Elsevier.
Osis, J. and Slihte, A., 2010. Transforming Textual Use
Cases to a Computation Independent Model. In J. Osis
and O. Nikiforova, eds. Model-Driven Architecture and
Modeling-Driven Software Development: ENASE
2010, 2ndMDAandMTDD Whs. SciTePress, pp. 33–42.
Pearl, J., 2019. The Seven Tools of Causal Inference, with
Reflections on Machine Learning. Communications of
Association for Computing Machinery, 62(3), pp.54–
60.
Rhazali, Y., Hadi, Y. and Mouloudi, A., 2016. CIM to PIM
Transformation in MDA: from Service-Oriented
Business Models to Web-Based Design Models.
International Journal of Software Engineering and Its
Applications, 10(4), pp.125–142.
Šlihte, A. and Osis, J., 2014. The Integrated Domain
Modeling: A Case Study. In Databases and
Information Systems: Proceedings of the 11th
International Baltic Conference (DBandIS 2014).
Tallinn: Tallinn University of Technology Press, pp.
465–470.
Slihte, A., Osis, J. and Donins, U., 2011. Knowledge
Integration for Domain Modeling. In J. Osis and O.
Nikiforova, eds. Model-Driven Architecture and
Modeling-Driven Software Development: ENASE
2011, 3rd Whs. MDAandMDSD. SciTePress, pp. 46–
56.
Solstad, T. and Bott, O., 2017. Causality and causal
reasoning in natural language. In M. R. Waldmann, ed.
The Oxford Handbook of Causal Reasoning. Oxford
University Press.
Sorgente, A., Vettigli, G. and Mele, F., 2018. A Hybrid
Approach for the Automatic Extraction of Causal
Relations from Text. In Emerging Ideas on Information
Filtering and Retrieval. Springer International
Publishing AG, pp. 15–30.
Sorgente, A., Vettigli, G. and Mele, F., 2013. Automatic
extraction of cause-effect relations in Natural Language
Text. In C. Lai, G. Semeraro, and A. Giuliani, eds.
MDI4SE 2019 - Special Session on Model-Driven Innovations for Software Engineering
612
Proceedings of the 7th International Workshop on
Information Filtering and Retrieval co-located with the
13th Conference of the Italian Association for Artificial
Intelligence (AI*IA 2013). pp. 37–48.
Waldmann, M.R. and Hagmayer, Y., 2013. Causal
reasoning. In D. Reisberg, ed. Oxford Handbook of
Cognitive Psychology. New York: Oxford University
Press.
Identification of Causal Dependencies by using Natural Language Processing: A Survey
613