Towards Relation-based Argumentation Mining? Francesca Toni - PowerPoint PPT Presentation

Towards Relation-based Argumentation Mining? Francesca Toni Department of Computing Imperial College London Computational Logic and Argumentation Dagstuhl seminar on Natural Language Argumentation, April 2016

Outline From • Structured argumentation (ABA) and Abstract Argumentation (AA) for rule-based arguments and beyond to • Bipolar argumentation and Quantitative Argumentation Debates (QuADs) supported by and supporting • Mining of attack/support/neither relations amongst Arguments

Structured Argumentation with Conflicting rules

ABA with conflicting rules YOU ARE COVERED FOR : UK and EU Breakdown Assistance for account holder(s) in any private car they are travelling in YOU ARE NOT COVERED FOR : private cars not registered to the account holder(s) unless the account holder(s) are in the vehicle at the time of the breakdown contrary: • Default logic (Reiter 1980) contrary:

ABA/AA semantics with conflicting rules A set of arguments is admissible if it is conflict- free and it attacks every argument that attacks it Mary(friend’s car, in car) covered as travelling in private car attacked by not covered as car not registered to Mary attacked by Mary in car at time of breakdown Dung 1995; Bondarenko, Dung, Kowalski, Toni 1997

ABA for decision making in medicine Mocanu, Fan, Toni, Williams, Chen 2014-16

AA for Smart Electricity Makriyiannis, Lung, Craven, Toni, Kelly 2014-16

AA for Reinforcement Learning robocup Gao, Toni 2012-15

Outline From • Structured argumentation (ABA) and Abstract Argumentation (AA) for rule-based arguments and beyond to • Bipolar argumentation and Quantitative Argumentation Debates (QuADs) supported by and supporting • Mining of attack/support/neither relations amongst Arguments

Baroni, Romano, Toni , Aurisicchio, Bertanza (2015 )

REUSE OF SLUDGE PRODUCED BY WASTEWATER TREATMENT PLANTS

stay exit support

Justified vs Weak/Strong Arguments stay exit weaker stronger if opinions have equal a-priori strength …. …. …. …. ….

www.arganddec.com IBIS (Issue Based Information System, Kunz and Rittel 1970 ) Aurisicchio, Baroni, Pellegrini, Toni (2015)

www.arganddec.com

Computing strength • All opinions have a base score (a-priori strength) in [0,1] • Opinions (arguments) attack or support other opinions (arguments) • Debates are (sets of) trees

Computing strength To compute the strenght SF of argument x the function C combines three elements: the base a factor summarizing a factor summarizing score of x the attackers of x the supporters of x using the sequence of the strengths of the attackers of x Romano, Rago, Baroni, Toni , Aurisicchio, Bertanza 2013-16

Combining all factors: properties • The order of attackers/supporters does not matter = • Adding a supporter will not lower strength s' s 

Combining all factors: more properties • The smaller the strength of an attacker/supporter the smaller its impact s' s s 0 if s 0 is very small then s' is almost the same as s

Outline From • Structured argumentation (ABA) and Abstract Argumentation (AA) for rule-based arguments and beyond to • Bipolar argumentation and Quantitative Argumentation Debates (QuADs) supported by and supporting • Mining of attack/support/neither relations amongst Arguments

Argument mining Carstens, Toni 2015

Outline From • Structured argumentation (ABA) and Abstract Argumentation (AA) for rule-based arguments and beyond to • Bipolar argumentation and Quantitative Argumentation Debates (QuADs) supported by and supporting • Mining of attack/support/neither relations amongst Arguments

Relation-based Argument Mining (RbAM) Classifier (BOW) argument base Strength calculation

Argument base for RbAM 1. Keyword arguments (73) 2. First word arguments (10) 3. Sentiment arguments (3) 4. Similarity arguments (12) 5. Negation arguments (73) e.g. 1. The word “and” appears in Child (supports Sup) 2. The word “and” is the first word in Child (supports the keyword argument)

Corpora for RbAM • Internet Argument Corpus (IAC) from 4forums [Walker et al LREC 2012] • 18 sub-corpora of AIFdb • (new) News articles corpus: • Comprised of 2,274 sentence pairs: • 413 Attack, 456 Support, 1,385 Neither • Fleiss ’ kappa = 0.4287

RbAM: some experiments

Conclusions • Several argumentation frameworks and semantics (for evaluating arguments) • Bipolar argumentation and QuAD for natural language arguments • Quantitative strength as a useful measure of dialectical strength of mined arguments? Goodness of Argument Mining? • Argumentation to help Argument Mining?