Argumentative texts and clause types Alexis Palmer Leibniz ScienceCampus, University of Heidelberg, IDS (Mannheim) Dagstuhl, April 2016 1
Carlota Smith Argumentative texts and clause types Annemarie Friedrich, Saarland University Alexis Palmer Leibniz ScienceCampus, University of Heidelberg, IDS (Mannheim) Dagstuhl, April 2016 2
Anette Frank Argumentative texts and clause types Maria Becker Alexis Palmer Leibniz ScienceCampus, University of Heidelberg, IDS (Mannheim) Dagstuhl, April 2016 3
Situation entities • Overview: new project annotating argumentative microtexts (Peldszus & Stede) with SE types • Situation entity (SE) type — what kind of situation does the clause evoke in the discourse? • SE type linked to linguistic characteristics of clause - lexical aspect of main verb - genericity of main referent (~ subject) - habituality of clause Alexis Palmer Dagstuhl, April 2016 4
Situation entities • Overview: new project annotating argumentative microtexts (Peldszus & Stede) with SE types • Situation entity (SE) type — what kind of situation does the clause evoke in the discourse? • SE type linked to linguistic characteristics of clause [Friedrich/Palmer/Pinkal, in submission], ~80% accuracy - lexical aspect of main verb [Friedrich/Palmer, ACL14], 66-93% accuracy - genericity of main referent (~ subject) [Friedrich/Pinkal, ACL15], 70-89% accuracy - habituality of clause [Friedrich/Pinkal, EMNLP15], 74-84% accuracy • … and can be modeled computationally and predicted automatically Alexis Palmer Dagstuhl, April 2016 5
Eventualities - states and events (1) The NMFS estimates ST (2) that in 2000, more than 16,000 tons of Chilean seabass were legally caught from an internationally regulated harvest area in the Antarctic Ocean. (3) But more than 32,000 tons may have been taken illegally from those same waters, (4) the fisheries service said. [masc_news_NYTnewswire7_part1.txt] state event state* reporting Alexis Palmer Dagstuhl, April 2016 6
Eventualities - states and events (1) The NMFS estimates (2) that in 2000, more than 16,000 tons of Chilean seabass were EV legally caught from an internationally regulated harvest area in the Antarctic Ocean. (3) But more than 32,000 tons may have been taken illegally from those same waters, (4) the fisheries service said. [masc_news_NYTnewswire7_part1.txt] state event state* reporting Alexis Palmer Dagstuhl, April 2016 7
Eventualities - states and events (1) The NMFS estimates (2) that in 2000, more than 16,000 tons of Chilean seabass were legally caught from an internationally regulated harvest area in the Antarctic Ocean. (3) But more than 32,000 tons may have been taken illegally ST from those same waters, (4) the fisheries service said. [masc_news_NYTnewswire7_part1.txt] state event state* reporting Alexis Palmer Dagstuhl, April 2016 8
Eventualities - states and events (1) The NMFS estimates (2) that in 2000, more than 16,000 tons of Chilean seabass were legally caught from an internationally regulated harvest area in the Antarctic Ocean. (3) But more than 32,000 tons may have been taken illegally from those same waters, (4) the fisheries service said. REP [masc_news_NYTnewswire7_part1.txt] state event state* reporting Alexis Palmer Dagstuhl, April 2016 9
General Statives (5) Blobfish are small fish, typically shorter than 30 cm. G (6) Blobfish are often caught as bycatch in bottom trawling nets. [adapted from wikipedia_wikiGenerics_blobfish.txt] generic generalizing sentence (habitual) Alexis Palmer Dagstuhl, April 2016 10
General Statives (5) Blobfish are small fish, typically shorter than 30 cm. (6) (?) Blobfish are often caught as bycatch in bottom trawling GS nets. [adapted from wikipedia_wikiGenerics_blobfish.txt] generic generalizing sentence (habitual) Alexis Palmer Dagstuhl, April 2016 11
General Statives (5) Blobfish are small fish, typically shorter than 30 cm. (6) (?) Blobfish are often caught as bycatch in bottom trawling GS nets. I have often caught blobfish accidentally on my fishing trips. [adapted from wikipedia_wikiGenerics_blobfish.txt] generic generalizing sentence (habitual) Alexis Palmer Dagstuhl, April 2016 12
Abstract Entities (7a) There is no doubt [state] (7b) that [ Randall enjoys her work ] S . F [masc_journal_VOL15_3_part16.txt] (8a) Mr. Icahn then proposed (8b) that [USAir buy TWA]. [MUC6 data] fact proposition Alexis Palmer Dagstuhl, April 2016 13
Abstract Entities (7a) There is no doubt (7b) that [ Randall enjoys her work ] S . [masc_journal_VOL15_3_part16.txt] (8a) Mr. Icahn then proposed [event] (8b) that [ USAir buy TWA ] S . P [MUC6 data] fact proposition Alexis Palmer Dagstuhl, April 2016 14
Abstract Entities (7a) There is no doubt (7b) that [ Randall enjoys her work ] S . [masc_journal_VOL15_3_part16.txt] (8a) Mr. Icahn then proposed [event] (8b) that [ USAir buy TWA ] S . [MUC6 data] fact proposition PLUS: question imperative Alexis Palmer Dagstuhl, April 2016 15
Abstract Entities (7a) There is no doubt (7b) that [ Randall enjoys her work ] S . [masc_journal_VOL15_3_part16.txt] (8a) Mr. Icahn then proposed [event] (8b) that [ USAir buy TWA ] S . [MUC6 data] fact proposition wait, but why? PLUS: question imperative Alexis Palmer Dagstuhl, April 2016 16
Discourse modes [Smith 2003] ARGUMENT/( NARRATIVE( DESCRIPTION( INFORMATION( REPORT( Discourse COMMENTARY( Modes Related: Werlich’s (1975) typology of texts, Santini (2006), i.a. [images thanks to Kleio Mavridou] Alexis Palmer Dagstuhl, April 2016 17
Patterning of SEs and Genre • Taking genre categories as proxy for DMs, do SE types pattern as the theory predicts? • Manual SE annotations • Manually-Annotated SubCorpus (MASC): 10,270 clauses from news, jokes, essays, and (fund-raising) letters • Penn Discourse TreeBank (PDTB): 2513 clauses from news and essays (following Webber 2009) Alexis Palmer Dagstuhl, April 2016 18
MASC: SEs and Genre % • REPORT texts: greater proportion of States and Events, fewer Generics and Generalizing Sentences • ARGUMENT/COMMENTARY texts: many more Generics and Generalizing Sentences, States over Events [Palmer/Friedrich 2014] Alexis Palmer Dagstuhl, April 2016 19
And now to focus on arguments • Argumentative microtexts offer corpus of “purely argumentative” text passages • *with argument graphs* • Our question: How can SE types be helpful for modeling argumentative regions of text? For mining arguments? For reasoning over argument components? • Annotation project underway - analysis of this data will inform next steps • Next texts: from Potsdam Commentary Corpus Alexis Palmer Dagstuhl, April 2016 20
One example: micro_b005 [e5] Deren W erkzeuge, Daten [e1] Die Geheimdienste müssen [e2] das sollte jedem nach den [e3] Die betreffen zwar vor [e4] aber mit denen arbeiten und Knowhow wird schon lange dringend stärker vom Parlament Enthüllungen von Edward allem die britischen und die Deutschen Dienste zu unserer Überwachung kontrolliert werden, Snowden klar sein. amerikanischen Geheimdienste, bekanntermaßen eng zusammen. genutzt. 5 2 c5 3 4 c4 [e1] Intelligence services must urgently be regulated more tightly [e5] Their tools, data and expertise by parliament; c3 have been used to keep us under surveillance for a long time. c2 [e3] Granted, this concerns primarily the British and American intelligence services, 1 [e2] this should be clear to [e4] but the German services evidently everyone after the disclosures do collaborate with them closely. of Edward Snowden. Alexis Palmer Dagstuhl, April 2016 21
One example: micro_b005 [e5] Deren W erkzeuge, Daten [e1] Die Geheimdienste müssen [e2] das sollte jedem nach den [e3] Die betreffen zwar vor [e4] aber mit denen arbeiten und Knowhow wird schon lange dringend stärker vom Parlament Enthüllungen von Edward allem die britischen und die Deutschen Dienste zu unserer Überwachung kontrolliert werden, Snowden klar sein. amerikanischen Geheimdienste, bekanntermaßen eng zusammen. genutzt. G S S GS GS 5 2 c5 3 4 c4 [e1] Intelligence services must urgently be regulated more tightly [e5] Their tools, data and expertise by parliament; c3 have been used to keep us under surveillance for a long time. c2 [e3] Granted, this concerns primarily the British and American intelligence services, 1 [e2] this should be clear to [e4] but the German services evidently everyone after the disclosures do collaborate with them closely. of Edward Snowden. Alexis Palmer Dagstuhl, April 2016 22
Preliminary (very) observations • Arg. texts are different from non-arg. texts wrt SE types - very few events - many generics and generalizing sentences • Claims tend to be generics or states - when generic, always first segment - when state, usually not first segment (often preceded by counter-claims) • Small tendency for counter-claims to be non-generic vs. generic proponent claims Alexis Palmer Dagstuhl, April 2016 23
Looking ahead… • For what kinds of tasks could it be useful to have clauses labeled with SE types? • Use SE types for discourse mode identification and classification (in progress) - find argumentative regions of texts Alexis Palmer Dagstuhl, April 2016 24
Recommend
More recommend
