Question Answering System automatically answers natural language questions asked by user. Promising results have been achieved for factoid-type questions but there is a need to improve non-factoid question answering. This paper has addressed the issue for automatically answering Why-type questions in Question Answering System. The concept of causality is studied and its important role for di erent modules in developing Why-type Question Answering System. There are various researchers who have used causality as a key component to focus on causes and e ects involved in the sentences and ultimately answering Why-type Questions.
Causality is viewed as a semantic relation that exists between di erent parts of a sentence [1]. It connects two phrases representing as a cause and its e ect. It has its signi cance in various disciplines like Psychology, linguistics, philosophy and even computer science. In computer science, causality plays its major role in understanding the meaning and determines how di erent parts of a sentence are related to each other. The paper tries to explore the importance of causality in understanding natural language text and thus helping to nd candidate answers to a question asked on question answering system.
In English language, semantics deconstruct the words and use sentence structure
to determine the meaning of text. Semantics is understood by three dimensions,
formal semantics uses logic to nd the relation between language and reality,
lexical semantics separates words and phrases to understand the meaning of a
text, and conceptual semantics nds out the basic meaning of the text before any
context and feelings included in it. There are various types of semantic relations
between words represented as (
There are di erent forms of causality contained in a sentence (
The paper is divided into various sections. Section 2 puts light on the importance of causality with its motivation. Section 3 discusses di erent dimensions of causality and its role in di erent elds. Section 4 introduces di erent type of causality involved in a sentence. Section 5 describes various approaches used to identify causal relations in a sentence with future directions. Section 6 addresses di erent issues and challenges involved in identifying causality. Section 7 at last nally concludes the work with future research directions. 2
The Question Answering System (QAS) [5] returns an answer to a question asked by user. The primary requirement of QAS is to understand the need of the user from the question asked by him/her. The research started by identifying adverbial clauses [6] in a relevant document that depicts events occurring in di erent parts of a sentence. The clauses are classi ed by containing various semantic relations e.g. place, manner, cause/reason, purpose, result, condition, contrast, similarity etc. [7] The appropriate answer candidates to a question are assumed to containing such semantic relations. Di erent semantic relations contained in answer candidates depend on their question types. For example, answers to why-question are expected to contain semantic relations `cause/reason' and `purpose', how questions are expected to contain semantic relation `manner', where type questions contain semantic relation `place' and comparative questions are expected to contain relations like `substitution and contrast' and so on. Thus, semantic relations play signi cant role in extracting appropriate answer candidates to a question.
Question Answering System: QAS is an application of Information Retrieval and Natural Language Processing that returns one appropriate answer to a question depending on the user and context of time at which the question is being asked. The process requires identifying semantic relations in document which match the relations contained in the user's question. Depending on the type of question, appropriate semantic relation is identi ed from the documents which are expected to containing an answer to a question.
Query Expansion: Semantic relations play a major role to extract related terms that are associated with existing query terms to use it for query expansion. This is required whenever there is a lexical mismatch between query and answer documents, and the need is to improve query by appending related terms to it. The query expansion process is either carried out automatically or explicitly by user suggesting the terms to be appended [8].
Text Summarization: Text summarization [9] is a process of removing redundant information and including valuable information. It is performed by considering the relations between terms or concepts involved in the text documents. The part of the text which comprises terms from user query as well as cue phrases which determine semantic relations between them are considered while summarizing text documents.
Network Events: Semantic networks for the text are constructed which depict the semantic relationships contained between the events involved in a sentence. Semantic relations are used to deduce meaning from the text by nding semantic roles of the entities involved in di erent events [10]. 4
The table below 1 discusses several parameters used by researchers to classify causal relations. They are used in various di erent scenarios for example, as cue phrases which act as connectives, adverbs, questions etc. [11]
Patterns in- Explicit volved in text
ple Ambiguous Terms not re ecting causation everytime (e.g. since) Non-Ambiguous Terms always re ect causation (e.g. because) Patterns explicitly present in sentence Implicit Explicit patterns are not present but still express causation relation Adverbial connectives Used to link words by referring forward and backward in a sentence (e.g. for this reason,with the result that) Prepositional connec- Prepositions are used to tives link causal and e ect part of a sentence e.g. as a result of, because of, due to etc.
Subordination connec- Conjunctions are used tives to address immediate e ect e.g. so, since, because etc. e.g. lead to, force, generate, cause etc.
Causation verbs Linking verbs (used in explicit causal patterns)
etc. where it results into some other condition Instrumental causatives e.g. poison, hang, clean etc. where instrument is used to perform action Causative ad- Conjunctive adverbs of e.g. because, as, since, verbs (used in cause or reason that, due to, hence and explicit causal many more patterns)
Adverbs referring to ef- e.g. fects quently painful, conse
causation ques- Explicit questions tions Implicit verbs causality to agent (used in implicit causality) e.g. Rahul confessed for not calling Shakshi
tionships
Rela- one cause causality to patient e.g. Rahul admired Shakshi because he like her One cause having multiple e ects one e ect Multiple causes resulting one common e ect chain of causes and ef- A cause prompts an fect e ect, which further prompts another di erent e ect with repeating process causal homeostatis Cycle of cause and e ect relationships where nal e ect has initial cause as its e ect Comprise explicit keywords that determine type of relation in it e.g. cause, e ect, consequence etc.
Semi-explicit questions Includes ambiguous keywords causing dif
culty to determine semantic relations e.g. generate, trigger etc.
Implicit questions No explicit keywords present but cause and e ects are contained in a sentence e.g. kill,sing etc.
This section discusses the usage of causal relations in di erent modules of Questions. Researchers have utilized causal semantic relations to address the development of Why-type Question Answering System. 5.1
Jong Hoon Oh et. al. in 2018 [13] has identi ed causal relations in text to automate the retrieval of Question and Answer pairs for the training data. The process is carried out to improve the answer re-ranking process. The authors extracted sentences containing cue phrases from the text document and identied cause and e ect parts from the sentence. For example, consider the sentence (The sun rises in east) e ect because (the earth rotates around its axis toward the east) cause. The sentence helps to form question and its answer. It has been stated that cause part is regarded as an expected answer to a question which can be automatically extracted from the e ect part. Thus, the cause part 'the earth rotates around its axis toward the east' serves as an answer to the question formulated from the e ect part 'Why sun rises in east?' 5.2
J. Fukumoto in 2007 hand-crafted various extraction and non-extraction patterns for each type of non-factoid questions, whether it is why-type, de nitional or how-type questions [14]. Causal semantic relations are identi ed from the document using Rhetorical Structure theory. Answer candidates are extracted from these retrieved passages containing causal relations in it. If a sentence in document matches extraction patterns, it is considered as an appropriate answer candidate and those containing non-extraction patterns are not considered as appropriate candidate for answers.
In 2013, Jong Hoon Oh et. al. identi ed intra and inter-sentential causal relations for answering Why-type questions [15]. Intra-sentential causal relations contain cue phrases like 'because', 'since', 'causes' etc. and inter-sentential are expected to comprise phrases like 'This causes' and 'As a result' etc. These associations are found using bag-of-word approaches. 5.3
Higashinaka and Isozaki in 2008 addressed the issue of assigning appropriate
weights to the hand-crafted patterns [16]. The authors ranked answer
candidates on the basis of three features contained in them which are (
Jong Hoon Oh et. al. in 2013 trained answer re-ranker on the basis of three features viz. morphosyntactic, semantic word classes and sentiment polarity [21]. Causal relations are accessed through term-matching in which e ect part comprises more than one content word as contained in question, partial-tree matching in which e ect part comprises more than one partial tree as contained in question and excitation- polarity matching in which e ect part and question share one common noun of same polarity. 6 6.1
There are certain phrases which always express causation, and some rarely express causation depending on the context of their appearance in a sentence. It is di cult to understand the nature of sentence which requires semantic interpretations to understand the implications of such cue phrases encountered in a sentence. 6.2
Implicit causations are complex and di cult to address. There are some phrases which are often used consecutively and address cause and e ect but doesn't contain any connectives between them e.g. 'cold tremble' is a phrase which doesn't contain any explicit cue phrase but there is causal relationship between them where cold is the cause and tremble is its e ect. Inference procedures based on semantic knowledge with some background knowledge and common sense reasoning is required to identify such causation within a sentence [22]. 6.3
It has been viewed that causes and e ect may or may not be contained in one sentence. Also there are some cases where there is a chain of causes and their e ects which creates di cult to determine immediate e ect of the causal event involved in a sentence. The procedure requires some background knowledge to identify causes and e ects present in a sentence. 6.4
It is very di cult to identify the events and determine causality relations between two events mentioned in a sentence. Discourse relations help to judge how events are related to each other semantically playing signi cant role in developing question Answering Systems.
The paper discusses a concept of causality with its signi cance in question answering system. A classi cation of causal relations and the approaches used to discern them in a sentence are described. There are some challenges and issues discussed which are needed to get resolved so as to further improve the performance of question answering module. In the future work, other than causal relations, other semantic relations like 'purpose' and 'motivation' can be explored to improve the accuracy of Question Answering System. 16. Higashinaka, R., & Isozaki, H. : Corpus-based question answering for why- questions. In Proceedings of the Third International Joint Conference on Natural Language Processing: Volume-I (2008). 17. Johnson, C. R., Schwarzer-Petruck, M., Baker, C. F., Ellsworth, M., Ruppenhofer,
J., & Fill- more, C. J. Framenet: Theory and practice (2003).
18. Palmer, M.: Semlink: Linking propbank, verbnet and framenet. In Proceedings of
the generative lexicon conference (pp. 9-15). Pisa, Italy: GenLex-09 (2009).
19. Muller, C., & Gurevych, I. : A study on the semantic relatedness of query and
document terms in information retrieval. In Proceedings of the 2009 Conference
on Empirical Methods in Natural Language Processing: Volume 3-Volume 3 (pp.
1338-1347). Association for Computational Linguistics (2009).
20. Yokoi, T. : The EDR electronic dictionary. Communications of the ACM, 38(