際際滷

際際滷Share a Scribd company logo

Present eval

Download as PPTX, PDF
Rodolfo Delmonte
Rodolfo Delmonte

This document describes ITVENSES, a symbolic system for aspect-based sentiment analysis. It walks through an example of how ITVENSES analyzes a sentence in Italian, tagging parts of speech, determining syntax and semantics, identifying aspects and polarities, and outputting an evaluation. ITVENSES uses sieves and rules to analyze sentences at different levels of representation and combine analyses to determine overall sentiment. Performance results on test sets are provided, showing ITVENSES achieves good accuracy for binary and multiclass sentiment classification tasks.

1 of 18
Download to read offline
ITVENSES - A SYMBOLIC SYSTEM FOR ASPECT BASED SENTIMENT ANALYSIS RODOLFO DELMONTE DIPARTIMENTO DI STUDI LINGUISTICI E CULTURALI COMPARATI UNIVERSIT CA FOSCARI EMAIL: DELMONT@UNIVE.IT WEBSITE: RONDELMO.IT 1
OUTLINE Systems Architectures A Walkthrough Example 2
ITGETARUNS 3
ITVENSES To develop the system I used 20% of the dataset and the remaining 80% for testing 4
WALKTHROUGH EXAMPLE TAGGING AND LEMMATIZING opn(1240342904,[0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"M anca il dentifricio in bagno."). 1240342904_1 - ['Manca'-v,il-art,dentifricio-n,in-p,bagno-n,'.'- punto], 1240342904_1-[ i(1,Manca',v,mancare-[sems=intr,mfeats=kl3s]), i(2,il,art,il-[sems=def,mfeats=ms]), i(3,dentifricio,n,dentifricio- [mfeats=ms]), i(4,in,p,in), i(5,bagno,n,bagno- [sems=com,mfeats=ms]), i(6,.',[punto],-) 5
WALKTHROUGH EXAMPLE SYNTAX & SEMANTICS 1240342904_1-[ ibar-[Manca'-v-sn], obj-[il-art-sn,dentifricio-n-sn], obl-[in-p-sp,bagno-n-sn]] 1240342904_1-[ bagno-obl-1-[obl-[in-p-sp,bagno-n-sn]], dentifricio-obj-1-[obj-[il-art-sn,dentifricio-n-sn]], mancare-ibar-1- ibar-[Manca-v-sn] ] 6
WALKTHROUGH EXAMPLE SYNTAX & SEMANTICS pas(1240342904_1, mancare - [ refex(1240342904_1-1, v, 'Manca' - mancare, [sems=intr,mfeats=kl3s], [activ,not_exten]), refex(1240342904_1-3, n, dentifricio-dentifricio, [3,mas,sing], [mfeats=ms], 1, subj / theme), [ i(4,in,p,in,sp,[],1,-), refex(1240342904_1-5, n, bagno - bagno, [def=indef,3,mas,sing], [ [act,agnt,artf,bld,cse,dyn,liqd,locat,med,obj,part], polsem = neut], 4, obl / theme) ] 1 - [ lemma = mancare, disc_m = nil, polsem = negative, subcat = [activ-not_exten], parola = Manca, change = gradual, view = external, factive = factive, moodtense = presente] ) 7
WALKTHROUGH EXAMPLE FROM ITGETARUNS TO ITVENSES Try Match Aspect/s from refexs, i.e. Nouns, Verbs, Adjectives - bagno aspect 2; mancare aspect 3 Try Match Polarity/ies from refexs, i.e. Nouns, Verbs, Adjectives - mancare marked as negative sievesall: recomposes aspects and polarities which can be multiple for every sentence in a text 8
WALKTHROUGH EXAMPLE FROM ITGETARUNS TO ITVENSES sievescheck: invertpols (invert polarities for the current aspect) sievescheck: focalizers (spots focalizers, minimizers, downtoners) sievescheck: checknegpriv (finds negation and its scope) sievescheck: syntax sieves (deletes current aspect assignment identifiers) Ind=2;Ind=3;Ind=6;Ind=7 - bagno Ind=2 (deleted) Ind=3 albergo;hotel;struttura & centro;centrale;a_due_passi Ind=2 camera;moquet;asciugamano;stanza;ambiente;bagno;letto & spazioso;comodo & + pulito Ind=7 strada;piazza & rumoroso Ind=7 arrivare;raggiungere & difficile;distante;scomodo;scarso 9
WALKTHROUGH EXAMPLE collapseall: recovers all clause level analysis of the current sentence both at propositional and at subjective/factivity level and collects them together now each evaluation term is made up by a text index - a set of semantic propositional level representations for that sentence - one aspect assignment - one associated polarity assignment, made up by a positive and a negative slot 10
WALKTHROUGH EXAMPLE AUGMENTED PREDICATE ARGUMENT STR. 1240342904-[ 1240342904_1-mancare(neg,statement,dentifricio-dentifricio-3, bagno-bagno-5)]- [mancare]- Aspect seeds [[],[Manca]]- Polarities: Positive+Negative 3] Aspect Identifier 11
WALKTHROUGH EXAMPLE evalothers: evaluates sentences marked with aspect n.8 and associates semantic representations reduceevals: collapses evaluation terms for the same sentence with identical values othersieve: sieves and modifies aspect value using combinations of aspect assignments present at text level; fires preferences for combined aspect values which modify one or more value 12
WALKTHROUGH EXAMPLE comparevals: sieves and modifies those texts declaring tutto bene or the opposite with an all aspects positive/negative marking checks for texts made up by a couple of aspects each evaluated to the contrary checks for texts which have a semantic propositional level analysis as nonfactual or as negated and marks them with negative polarity - if + double negations 13
WALKTHROUGH EXAMPLE Outputs the resulting 0/1 string 1240342904-[0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]-true www.rondelmo.it 14
PERFORMANCE OF ITVENSES 15 Delayed Results for Test Set After Ablation Experiment
PERFORMANCE OF ITVENSES 16 Results for Development Set
PERFORMANCE OF ITVENSES 17 Published Results for Test Set
ITVENSES FOR IRONITA TASK A: Binary classification TASK B: Multiclass classification 18

More Related Content

Present eval

  • 1. ITVENSES - A SYMBOLIC SYSTEM FOR ASPECT BASED SENTIMENT ANALYSIS RODOLFO DELMONTE DIPARTIMENTO DI STUDI LINGUISTICI E CULTURALI COMPARATI UNIVERSIT CA FOSCARI EMAIL: DELMONT@UNIVE.IT WEBSITE: RONDELMO.IT 1
  • 2. OUTLINE Systems Architectures A Walkthrough Example 2
  • 4. ITVENSES To develop the system I used 20% of the dataset and the remaining 80% for testing 4
  • 5. WALKTHROUGH EXAMPLE TAGGING AND LEMMATIZING opn(1240342904,[0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],"M anca il dentifricio in bagno."). 1240342904_1 - ['Manca'-v,il-art,dentifricio-n,in-p,bagno-n,'.'- punto], 1240342904_1-[ i(1,Manca',v,mancare-[sems=intr,mfeats=kl3s]), i(2,il,art,il-[sems=def,mfeats=ms]), i(3,dentifricio,n,dentifricio- [mfeats=ms]), i(4,in,p,in), i(5,bagno,n,bagno- [sems=com,mfeats=ms]), i(6,.',[punto],-) 5
  • 6. WALKTHROUGH EXAMPLE SYNTAX & SEMANTICS 1240342904_1-[ ibar-[Manca'-v-sn], obj-[il-art-sn,dentifricio-n-sn], obl-[in-p-sp,bagno-n-sn]] 1240342904_1-[ bagno-obl-1-[obl-[in-p-sp,bagno-n-sn]], dentifricio-obj-1-[obj-[il-art-sn,dentifricio-n-sn]], mancare-ibar-1- ibar-[Manca-v-sn] ] 6
  • 7. WALKTHROUGH EXAMPLE SYNTAX & SEMANTICS pas(1240342904_1, mancare - [ refex(1240342904_1-1, v, 'Manca' - mancare, [sems=intr,mfeats=kl3s], [activ,not_exten]), refex(1240342904_1-3, n, dentifricio-dentifricio, [3,mas,sing], [mfeats=ms], 1, subj / theme), [ i(4,in,p,in,sp,[],1,-), refex(1240342904_1-5, n, bagno - bagno, [def=indef,3,mas,sing], [ [act,agnt,artf,bld,cse,dyn,liqd,locat,med,obj,part], polsem = neut], 4, obl / theme) ] 1 - [ lemma = mancare, disc_m = nil, polsem = negative, subcat = [activ-not_exten], parola = Manca, change = gradual, view = external, factive = factive, moodtense = presente] ) 7
  • 8. WALKTHROUGH EXAMPLE FROM ITGETARUNS TO ITVENSES Try Match Aspect/s from refexs, i.e. Nouns, Verbs, Adjectives - bagno aspect 2; mancare aspect 3 Try Match Polarity/ies from refexs, i.e. Nouns, Verbs, Adjectives - mancare marked as negative sievesall: recomposes aspects and polarities which can be multiple for every sentence in a text 8
  • 9. WALKTHROUGH EXAMPLE FROM ITGETARUNS TO ITVENSES sievescheck: invertpols (invert polarities for the current aspect) sievescheck: focalizers (spots focalizers, minimizers, downtoners) sievescheck: checknegpriv (finds negation and its scope) sievescheck: syntax sieves (deletes current aspect assignment identifiers) Ind=2;Ind=3;Ind=6;Ind=7 - bagno Ind=2 (deleted) Ind=3 albergo;hotel;struttura & centro;centrale;a_due_passi Ind=2 camera;moquet;asciugamano;stanza;ambiente;bagno;letto & spazioso;comodo & + pulito Ind=7 strada;piazza & rumoroso Ind=7 arrivare;raggiungere & difficile;distante;scomodo;scarso 9
  • 10. WALKTHROUGH EXAMPLE collapseall: recovers all clause level analysis of the current sentence both at propositional and at subjective/factivity level and collects them together now each evaluation term is made up by a text index - a set of semantic propositional level representations for that sentence - one aspect assignment - one associated polarity assignment, made up by a positive and a negative slot 10
  • 11. WALKTHROUGH EXAMPLE AUGMENTED PREDICATE ARGUMENT STR. 1240342904-[ 1240342904_1-mancare(neg,statement,dentifricio-dentifricio-3, bagno-bagno-5)]- [mancare]- Aspect seeds [[],[Manca]]- Polarities: Positive+Negative 3] Aspect Identifier 11
  • 12. WALKTHROUGH EXAMPLE evalothers: evaluates sentences marked with aspect n.8 and associates semantic representations reduceevals: collapses evaluation terms for the same sentence with identical values othersieve: sieves and modifies aspect value using combinations of aspect assignments present at text level; fires preferences for combined aspect values which modify one or more value 12
  • 13. WALKTHROUGH EXAMPLE comparevals: sieves and modifies those texts declaring tutto bene or the opposite with an all aspects positive/negative marking checks for texts made up by a couple of aspects each evaluated to the contrary checks for texts which have a semantic propositional level analysis as nonfactual or as negated and marks them with negative polarity - if + double negations 13
  • 14. WALKTHROUGH EXAMPLE Outputs the resulting 0/1 string 1240342904-[0,0,0,0,0,0,1,0,1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0]-true www.rondelmo.it 14
  • 15. PERFORMANCE OF ITVENSES 15 Delayed Results for Test Set After Ablation Experiment
  • 16. PERFORMANCE OF ITVENSES 16 Results for Development Set
  • 17. PERFORMANCE OF ITVENSES 17 Published Results for Test Set
  • 18. ITVENSES FOR IRONITA TASK A: Binary classification TASK B: Multiclass classification 18