6.1 INTRODUCTION
Given a pair of arguments or argument components with one referred to as the source and the other as the target, argumentative relation mining involves determining whether a relation holds from the source to the target, and classifying the argumentative function of the relation, e.g., support vs. attack. While some sort of heuristics may be useful to pre-determine source and target components, e.g., relative positions of the components, the general form of the argumentative relation mining problem considers two ordered pairs for each two argument components, i.e., each component is considered as the source source and target in turn. Argumentative relation mining – beyond argument component mining – is perceived as an essential steps towards more fully identifying the argumentative structure of a text (Peldszus and Stede, 2013; Sergeant, 2013; Stab and Gurevych, 2014b). Consider the second paragraph shown in Figure 9. Only detecting the argument components (a claim in sentence 2 and two premises in sentences 3 and 4) does not give a complete picture of the argumentation. By looking for relations between these components, one can also see that the two premises together justify the claim. The argumentation structure of the text in Figure 9is illustrated in Figure 10according to the annotation provided in the first corpus of persuasive essays.
Research on classifying argumentative relations between pairs of arguments or argument components has proposed a variety of features ranging from the superficial level, e.g., word pair, relative position, to the semantic level, e.g., semantic textual similarity, textual entail- ment. Cabrio and Villata(2012);Boltuˇzi´c and ˇSnajder(2014) studied online debate corpora
Essay 73: Is image more powerful than the written word?
... (1)Hence, [I agree only to certain degree that in today’s world, image serves as a
more effective means of communication]M ajorClaim.
... (2)[pictures can influence the way people think ]
Claim. (3)For example, [nowadays
horrendous images are displayed on the cigarette boxes to illustrate the consequences of smoking]P remise. (4)As a result, [statistics show a slight reduction in the number of
smokers, indicating that they realize the effects of the negative habit]P remise...
Figure 9: Excerpt from a student persuasive essay. Sentences are numbered and argument components are tagged.
and aimed at identifying whether user comments support or attack the debate topic. They proposed to use content-rich features including semantic similarity and textual entailment. In principle, they expect the comment text (which is usually longer) to entail the topic phrase (which is usually shorter). Boltuˇzi´c and ˇSnajder (2014) calculated semantic similar- ity between each comment sentence and the topic phrase, and returned the max and mean of sentence-level similarity scores. Despite the fact that user comments are usually long with multiple sentences, both Cabrio and Villata (2012) and Boltuˇzi´c and ˇSnajder (2014) did not consider the discourse structure of the comment as auxiliary information to support the prediction. It has been proposed in (Biran and Rambow,2011) that justifications (e.g., user comment) usually contain discourse structures that characterize argumentation. However, their study made use of only discourse indicators but not the discourse relations. We believe that identifying the discourse structures of justification will give insights to argumentation patterns used by writers to show their stances towards the argument topic.
To illustrate our idea, consider the following excerpt from a persuasive essay in the first corpus:
Essay 26: Prepared food
(1)In addition, cooking is one of arts humans create. (2)The more cooked food we chosen,
the more cooking skills we lose. (3)At the increasing living pace, the majority of people tend
to choose microwave as their unique cooker that help them prepare a dish in five minutes.
(4)But rare people have been aware that this has contributed to a modification of cooking
MajorClaim(1) Claim(2) Premise(4) Premise(3) Premise(6) Support Support Attack Support Support Support
Figure 10: Structure of the argumentation in the excerpt in Figure9. Premises 3 and 4 were annotated for separate relations to Claim 2. Our visualization should not mislead that the two premises are linked or convergent.
(5)In conclusion, although the invention of prepared foods definitely satisfies the demand
of some people who are busy in their work, it is not a good thing.
The excerpt consists of a justification in sentences {1, 2, 3, 4} which supports a claim in sentence 5. Analyzing the discourse structure of the justification, we can see that the writer wanted to prove that “losing cooking skills” is a bad thing, which causes “losing custom and culture”, which consequently shows a stance against the “prepared foods”.
Another example can be taken from Figure9. Without knowing the content “horrendous images are displayed on the cigarette boxes” in sentence 3, one cannot easily tell that “re- duction in the number of smokers” in sentence 4 supports the “pictures can influence” claim in sentence 2. We expect that such content relatedness can be revealed from a discourse analysis, e.g., the appearance of a discourse connective “As a result ”.
Differently from (Cabrio and Villata,2012;Boltuˇzi´c and ˇSnajder,2014),Stab and Gurevych (2014b) aimed at classifying the argumentative relations (i.e., support vs. non-support) be- tween argument components. An argument component in (Stab and Gurevych, 2014b) is a sentence or a clause so it is less content-rich than user comments in (Cabrio and Villata, 2012; Boltuˇzi´c and ˇSnajder, 2014). Stab and Gurevych (2014b) proposed a diverse feature set including features involving information from both components of the pair. e.g., word
pairs, common words, relative positions. However, a limitation of their model is the lack of contextual information as mentioned in their paper. For example, it is hard to determine the support relation between these two argument components: “It helps relieve tension and stress” and “Exercising improves self-esteem and confidence” without knowing that “it ” refers to “Exercising”. Although anaphora resolution may help in this case, other situations could require topic inference to determine the relatedness between texts. While topic in- formation in many writing genres (e.g., scientific publications, Wikipedia articles, student essays) has been used to create features for argument component mining (Teufel and Moens, 2002; Levy et al., 2014; Nguyen and Litman, 2015), topic-based features have been less ex- plored for argumentative relation mining. In the excerpts below, knowing that ‘technology’ and ‘weapons’ in essay 8, and ‘online game’ and ‘computer ’ in essay 24 are topically related might help a model decide support relations between sentences.
Essay 8: Technology cannot solve all the world’s problems
(1)...[there are some serious problems springing from modern technology]
Claim. (2)First,
[deadly and powerful weapons can be a huge threat to the world’s peace]P remise.
Essay 24: Computer has negative effects to children
(1)[People who are addicted to games, especially online games, can eventually bear dangerous
consequences]Claim. (2)Although [it is undeniable that computer is a crucial part of human
life]P remise, [it still has its bad side]M ajorClaim.
Motivated by the discussion above, we propose context-aware argumentative relation mining – a novel approach that makes use of contextual features that are extracted by exploiting context sentence windows and writing topic to improve relation prediction.