Accelerating Sustainability Report Assessment with Natural Language Processing

(1)

UPTEC STS 21029

Examensarbete 30 hp

Juni 2021

Accelerating Sustainability

Report Assessment with Natural

Language Processing

(2)

Teknisk- naturvetenskaplig fakultet UTH-enheten Besöksadress: Ångströmlaboratoriet Lägerhyddsvägen 1 Hus 4, Plan 0 Postadress: Box 536 751 21 Uppsala Telefon: 018 – 471 30 03 Telefax: 018 – 471 30 00 Hemsida: http://www.teknat.uu.se/student

Abstract

Accelerating Sustainability Report Assessment with

Natural Language Processing

Lea Renmarker & Emma Välme

Corporations are expected to be transparent on their sustainability impact and keep their stakeholders informed about how large the impact on the environment is, as well as their work on reducing the impact in question. The transparency is accounted for in a, usually voluntary, sustainability report additional to the already required financial

report. With new regulations for mandatory sustainability reporting in Sweden, comprehensive and complete guidelines for corporations to follow are insufficient and the reports tend to be extensive. The reports are therefore hard to assess in terms of how well the reporting is actually done. The Sustainability Reporting Maturity Grid (SRMG) is an assessment tool introduced by Cöster et al. (2020) used for assessing the quality of sustainability reporting. Today, the assessment is performed manually which has proven to be both time-consuming and resulting in varying assessments, affected by individual interpretation of the content. This thesis is exploring how assessment time and grading with the SRMG can be improved by applying Natural Language Processing (NLP) on sustainability documents, resulting in a compressed assessment method - The Prototype. The Prototype intends to facilitate and speed up the process of

assessment. The first step towards developing the Prototype was to

decide which one of the three Machine Learning models; Naïve Bayes (NB), Support Vector Machines (SVM), or Bidirectional Encoder Representations of Transformers (BERT), is most suitable. This decision was supported by analyzing the accuracy for each model and for respective criteria in the SRMG, where BERT proved a strong classification ability with an average accuracy of 96,8%. Results from the user evaluation of the Prototype indicated that the assessment time can be halved using the Prototype, with an initial average of 40 minutes decreased to 20 minutes. However, the results further showed a decreased average grading and an increased variation in assessment. The results indicate that applying NLP could be successful, but to get a more competitive Prototype, a more nuanced dataset must be developed, giving more space for the model to detect patterns in the data.

(3)

(4)

(5)

(6)

(7)

1. Introduction

Pe e a e c a ded b e ec a i . E ec a i fi i he idea ,

e ec a i e f e , e ec a i a e ca e f hei ed e . The e

e ec a i a e ba ed a c a c a i i h he , hich ha e ched i e f a a i a e e e f he cia c e . We a e c ea ed g i he e idea f

c a i . The ch e a e g he hi ie , i h i bei g g aded a a

ea age de a e hei abi i ie . I he c a e d, hi c a i ha ai e ed a d fi a cia fac , b i ece ea e i dica ha e e e ed he ace f

c a e e i g. The acce e a i f e i e a cha ge Ea h ha b gh e

igh he i a ce f e e e' c ib i a e ai ab e a e . N ,

ec ic a e a cia , a d e i e a fac a a ig ifica e. The e

-fi a cia fac acc ed f i ai abi i e i g, ha e bec e a addi i a

c e f a a c a e e i g. H e e , he e i c e c ea g adi g e

f he a i f c a e ai abi i e i g. T b i g ba a ce c e i i e cie a d be ab e c a e a d e a a e he c a e ai abi i , he e i a eed f hi .

G adi g, c a i g, a d e if i g he c e f ai abi i e i c e

b e a ic a d i e-c i g. The G ba Re i g I i ia i e' ai abi i e i g a da d (GRI-S a da d ), hich a e i e i he a e , ea e a f f

e f-i e e a i . Thi ead i c i e cie i he c e f e (I a a d C e , 2018). S ai abi i e i g ai gai a i igh i h c a ie a e a i g ac i e e ibi i i ai ab e de e e . S ai ab e de e e i defi ed a he

De el me ha mee he eed f he e e i h c m mi i g he abili f f e ge e a i mee hei eed (B d a d, 1987). Wi hi a i a age e ,

ai ab e de e e ca be efe ed a d i g he igh hi g a d d i g i he igh a (C e , Dah i a d I a , 2020). Ba ed he a i a age e a ach a d he

GRI-S a da d , C e e a .(2020) ha e ed a f a e i g ai abi i e ,

he S ai abi i Re Ma i G id (SRMG). A h gh he SRMG ha a c ea c e,

he e a a i f he de de a e ha a e e i h he de ie d high

a iab e e . I i a c ide ed i e-c i g c d c a a e e . T ge a e ie f a e i e i d i he ef e diffic . B ed ci g h a i e e , hi he i ai ed ce b h i e a d a e e a ia i . Machi e ea i g (ML) i a fie d

i hi a ificia i e ige ce (AI), de e ed f c e ea a e f da a. Na a a g age ce i g (NLP) i a bfie d i hi AI, ecifica de e ed i e e

he h a a g age. The e a da a f ai abi i e c ai a ab e i igh

ha c d i e be e ea i acce ed h gh da a ce i g .

Thi he i ai e a i e h NLP a d ML e h d a ied ai abi i e ca he i e he a e e , i h ega d ed ci g i e e e a a i , a d a ia i

i e . The f he diffe e e h d ha bee c i ed i a d c e , ed

(10)

igi a . T e a a e h e he c c ded d c e e e a a b i e, a e e i h

eache a d a i a i hi he ec i f a i a age e a U ala U i e i a

e f ed. T f fi i e, hi he i ai a e he f i g e ea ch e i : Ca i a da a be e ac ed f ai abi i e i h NLP a d ML e h d ? Ca NLP a d ML e h d ed ce he a e e i e f ai abi i e i g? Ca NLP a d ML e h d ed ce he a ia i i g adi g f ai abi i e i g?

1.1 Delimitation

The ai abi i e c ec ed f hi he i a e i PDF f a . Tech i e f

e ac i g e a c e f a PDF diffe f diffe e e f PDF f a . The he i i h de i i ed ha d i g ai abi i e c a ib e i h he ch e e ac i

e h d. F he de i i a i ha bee d e c ed i E g i h. Thi ch ice

i d e ib a ie f NLP bei g i e e ed f he E g i h a g age, h be e e e e e ec ed f hi de i i a i .

1.2 Concepts

AI - A ificia I e ige ce ML - Machi e Lea i g NLP - Na a La g age P ce i g TBL - T i e B Li e, The TBL f a e i c a e h ee di e i f e f a ce: cia , e i e a , a d fi a cia . The 3P; Pe e, P a e a d P fi GHG - G ee h e ga

SRMG- S ai abi i Re i g Ma i G id (C e , Dah i a d I a , 2020) GRI - G ba Re i g I de

1.3 Course of Action

The ai f hi he i a i d ced b Rai e I a , h i a he e i f hi jec . I a , h i a A cia e P fe a he De a e f Ci i a d I d ia E gi ee i g, Q a i Scie ce a U a a U i e i , ha ed he ai abi i f c f

he he i . Hi i i f f e ai abi i e i g ha aid he f da i a d a e i g f a i g a ech ica a ach he a ea. D i g hi he i , ec i g ee i g

i h I a ha e bee c d c ed fi d i b ac e he e e edge a

eeded f he ech ica a ec f he he i . The ee i g ha e a bee ed e e a d di c he ech ica d c a diffe e age f he ce . A he e d f he jec , a

(11)

e . Di c i d i g he e e a i c ib ed e fi a i ha c d be de e ed i he e .

1.4 Structure of Thesis

Thi he i c i f e e cha e . The fi cha e i d ce he a ea f he he i , b e a i e h he a ea h d be e ed, f ed b he ai f he he i a d i

de i i a i . Cha e ide he e e a bac g d i hi he ai abi i

a ea a d he gic behi d he S ai abi i Re Ma i G id (SRMG), hich i he f a e hich hi he i i b i . The hi d cha e i f c gi i g bac g d Na a a g age ce i g (NLP), e e i g e i e ea ch he ic a d a

i d c i he diffe e achi e ea i g (ML) e h d ed i he he i . The f h cha e f he he i e e he e h d a d c e f ac i f he he i . Thi cha e acc f he e f c ec i g a d c ea i g da a, e a a i g he c e ed e i

ed f a e e . Cha e fi e e e he e hich a e f he di c ed i cha e

i , f ed b f e e ea ch. La , a c c i a d e ec e da i a e gi e i

(12)

2. Background

Thi ec i i gi e a bac g d f he i e e a i a ea f hi jec ; ai abi i . I i i d ce e e a bac g d i f a i f he SRMG, he a e i g a i hich

hi he i i ba ed . The ec i i di ided i a ea , S ai abili G al & C ce

a d E al a i f C a e S ai abili . The fi a ea i i d ce i a g a a d

c ce f ai abi i , a d he ec d a ea i e e e i die , he

GRI-S a da d , a d he gic behi d he SRMG.

2.1 Sustainabilit Goals & Concepts

Wi h cha gi g ea he a e , i i g ea e e a d e e e ea he e e , ec ie a e di ed a d i e affec ed. C i a e cha ge i affec i g e e c e e c i e ,

a d ge ac i a d a ai ab e f e a e e i ed (U i ed Na i , 2021).

Acc i hi g a ai ab e f e i e i e e f c ai ab e de e e . The

e e i a ab a defi ed a ead i 1987 a De el me ha

mee he eed f he e e i h c m mi i g he abili f f e ge e a i mee hei eed (B d a d, 1987). Thi ec i i gi e a e ie f he S ai ab e De e e G a (SDG ), hich i he UN' e g a f achie i g ai ab e

de e e . Af e hi , a e e a i f he cie ific c i a e c ce , ch a he a e a b da ie a d he ca b a , i be e e ed. The e c ce ide a cie ific e ie f a e a i e f c i a e i ac a d a ha a e he

GHG- e i i be ed ced.

2.1.1 The Sustainable Development Goals

I 2015, a he UN Me be S a e e d ed he 2030 Age da f S ai able De el me .

Thi e ed i a ha ed i i a d a eg f eace a d e i f b h e e a d he a e , b h a d i he f e (U i ed a i de a e , 2021). The S ai ab e De e e G a , abb e ia ed a SDG , a e he ge e a g a ed i Age da 2030. The e a e 17 e g a a d 169 b-g a hich c i e he SDG . The SDG a e b i

he c ce f T i e B Li e (TBL). The TBL f a e i c a e h ee

di e i f e f a ce: cia , e i e a , a d fi a cia (E i g , 1998). The 17 SDG g a c e a ea ac he TBL a d ai e adica e e e e e , c i a e cha ge a d c ea e eacef a d ec e cie ie b 2030. A e ie f he 17 g a ca be

(13)

Fig e 1: The 17 g al c i i g he S ai able De el me G al . (S ai able De el me G al (SDG ) a d Di abili U i ed Na i E able, 2015)

2.1.2 Planetar Boundaries

T ee Ea h hea h i ab ee i g i ba a ced (R c e al., 2009). E agge a i f

e hi g ca c ea e a i ba a ce hich ca e i a b e f effec . The ce e ha ee he Ea h e ab e a d e i ie ha e bee ide ified b R c e a . (2009), a d he a e ca ed he P a e a B da ie . The P a e a B da ie a e i a i i a f ide if i g cie ifica ha he afe e a i g ace f h a i i . The P a e a B da ie i a a e ide if h h a i ha e a e he i e ide ified

ce e be ab e c ea e hi ai ab e f e f ge e a i c e. The P a e a B da ie i e i e ce e (R c e .a , 2009); C i a e cha ge L f bi he e i eg i La d e cha ge Ni ge a d h h f he bi he e S a he ic e de e i Che ica i a d he e ea e f e e i ie Ocea acidifica i F e h a e c i A he ic ae adi g

(14)

be d (S effe e al., 2015). A e ie f he P a e a B da ie a d he a f he ce e ca be ee i Fig e 2.

Fig e 2. O e ie f he Pla e a B da ie he e he f ce e e ceedi g hei h e h ld a e ma ked i h a ed b . M dified f m S effe e al., 2015

2.1.3 The Carbon Law

I 2015, a i e a i a c i a e a ge a e - he Pa i Ag ee e - hich, a g he hi g , i a ed ha he Ea h' e e a e i c ea e b e ha deg ee . T

ee he -deg ee a ge , R c e a . (2017) ha e ca c a ed ha he ca b di ide e i i be c i ha f e e e ea , hich i a ca ed he Ca b La .

2.2 Evaluation of Corporate Sustainabilit

S ai abi i e i g i ac i g c ea g ide i e f h he e i g h d be d e a d he e a e he ef e diffic ie i b ai i g a c ea ic e f a c a a ga i a i '

ai abi i . The ga i a i Gl bal Re i g I i ia i e (GRI), hich a f ded i 1997, i he ai ide f a c a g age f e i g -fi a cia i ac . GRI i he ga i a i behi d he ide ed ai abi i di c e a da d- he GRI-S a da d (GRI, 2021b). A h gh ga i a i e acc di g he GRI-S a da d

(15)

hich he c a i ac i e. Thi e i dica e ha c a ie ' e ce i f he GRI-S a da d a a d ha he e i i e a g achie e GRI' g a f a c

ai abi i e i g a g age (GRI, 2021b). A he ai abi i e diffe g ea i c e , a e h d f e a a i g he a i f he ai abi i e i g i i i ed f . I

hi ec i , a e a a i f he e c ce f he GRI-S a da d i e e ed. Thi i

f ed b e i e e ed e h d f a e i g ai abi i e i g. La , he S ai abi i Re Ma i G id (SRMG) i be e e ed i a e de ai ed a e . 2.2.1 G R I (GRI) The g e e a ga i a i G ba Re i g I i ia i e (GRI) i he ai d i i g f ce i a da di i g ai abi i e i g (I a a d C e , 2018). GRI i i i c ea e a e ai ab e f e b he i g ga i a i be a a e a d a e

e ibi i f hei i ac he e i e . GRI ai c ea e ibi i ie f

c a i be ee ga i a i ai abi i e b c ea i g a g ba c

a g age f ga i a i e hei i ac (GRI, 2021b). I 2016, GRI G ide i e

e e c e ed S a da d , hich i c e he ai c e f ai abi i e i g

(GRI, 2021b). The GRI-S a da d a e di ided i ai a ea ; he U i e al S a da d e ai i g he c e f ai abi i e i g, a d he T ic-S ecific a da d c e i g

ai abi i i dica hich a e b i he TBL (GRI, 2021b). A e e , he

GRI-S a da d d i a e g ba h ai abi i e h d be c ed. I I a

a d C e ' (2018) e f he SRMG, 37 f 39 a d e ec ed e e e ba ed he GRI-S a da d .

The e a e f e c ce i he GRI-S a da d ; Im ac , Ma e ial ic, D e dilige ce, a d S akeh lde (GRI, 2020). The e c ce a he f da i f ai abi i e i g a d

i h be f he e ai ed ge a de a di g f hei ig ifica i . Im ac efe he effec a ga i a i ha c d ha e he TBL- i c di g ec ic, e i e a ,

a d cia i ac . The i ac ha i c de a ga i a i ' c ib i ai ab e

de e e , b h i i e a d ega i e, ac a a d e ia , a d g ba a e a ca (GRI, 2020). T ecif , he e i e a i ac ha i c de b h i i g a d i i g

a a e , i c di g a e , a he e a e a bi di e i .

The ga i a i ' ide ified i ac a e efe ed a he Ma e ial ic . The e ca be

a a d efe diffe e a ea f he TBL. The GRI-S a da d g ide he a e i ab a d c he i e he a e ia ic b eei g i i a i ie a d c e i g diffe e

ic , e.g. h a d e i i ed bi di e i (GRI, 2020). F i i i a i f he a e ia ic , e e i a d i e ih d a e a e ed. P i i ie a e c e e ed i a a e ia i a a i he e he c a ge he i h i a eh de fig a i e acc f he i i i a i f he ic (GRI, 2020). The a e ia i a a i i ec e ded be

(16)

Fig e 3. Ma e iali a al i - a i al e e e a i f i i i a i f ic . M dified fig e f m GRI,2021a

D e dilige ce i a b ad c ce ha efe he e i e a d h a ga i a i ha d e i ega i e i ac he TBL (GRI, 2020). Thi i c de he e i e ce f h he ide if , e e , i iga e a d acc f h he e effec h d be dea i h. The a c ce i akeh lde , hich GRI defi e a i di id a , g , ga i a i

i h i e e ha a e, a be, affec ed b a ga i a i ' ac i i ie deci i (GRI, 2020). A a eh de a be di ec i di ec affec ed b he i ac f he

ga i a i . I i i a ide if a e ia a eh de , b h he e h a e affec ed , b a he e affec ed i he f e. F e a e, if a ga i a i cha ge

d c i e h d , hich i c ea e hei ic e i i , he i ac a be g ea a e e , b i a affec f e ge e a i (GRI, 2020).

2.2.2 P R E M

Ob e i g e i die , diffe e e h d e a a e he a i f ai abi i e ca be f d. U i g he G ba Re i g I i ia i e S a da d (GRI-S a da d ) a a ba i f

he a e e i a c a i g i (Da b, 2007; Le c a, 2012).

I e i die , ic- ie ed c i g e f diffe e c e a e c ed

e a a e he a i f ai abi i e i g (Da b, 2007; Le c a, 2012). Da b (2007) e e ed a e h d gica a ach, i g b h a a i a i e a d a i a i e a a i f c a e ai abi i . Ba ed he GRI-S a da d f 2000 he e h d g i b i f 33 i di id a c i e ia de f ai ca eg ie (Da b, 2007). Ba ed a

be ch a i g d , i a ee ha e a i a i e e h d ga e addi i a i f a i ; The i f a i f he ai abi i e i g a i f e ced fi i he i age f he

c a i (Da b, 2007). Le c a (2012) ha a a ed he a i f ai abi i

e a d if he ca c ib e ha eh de a e. Le c a a a i c ce ed

(17)

e a a i , a ic- ie ed e h d a ed he e he e a c ed f 0-4 de e di g he e e f c e age f 78 diffe e ic . B c a i g ai abi i

e i g i 2005 i h 2010, Le c a a ig f i e e , e ecia de ai ed

e i g f TBL. Ob e i g e ec f i c i e e , he e e a ce f i f a i , a d e a i , Le c a' e ea ch iced he e e e ea e e . B h Da b (2007) a d Le c a (2012) fi di g c fi ha he g ide i e f ai abi i e i g, he GRI

a da d , a e i ade a e a d ea e a f f c a ie i e e a d i c de i f a i he c ide i a . T e a a e he c e e i e ic e e decide if ha he c a c ide i a c e ha i c ide ed i a f he a e a d e e. F c a ie i di id a i e e ha h d be acc ed f i ea i g f c a ie ea e e i g f hei e i e b i e . G ide i e c i ica e a a e h e i d f e i g a e h de i ab e. 2.2.3 R M G T c i ica e a a e ai abi i e i g, he a e e f e i hi he i i ba ed he S ai abi i Re Ma i G id (SRMG), hich i a e h d e e ed b C e e a . (2020). The SRMG i a a e he a i f ai abi i e a d i ba ed a i a age e i ci e . Ma i efe he a i f e i g a d i a

c c ce i hi cha ge a age e . I a (2019a)e ai ed ha Q ali i

b h ab d i g he igh hi g a d d i g i i he igh a . The e a e he ai f c a ea f he SRMG. The ga i a i ha he ef e e he igh hi g i he igh

a . The igh hi g ee f he e i c de he e i e a e chai a d ide if he a eh de a d he a eh de eed . Re i g i he e i e a e chai i e d ide if he ife c c e f a d c , f c ad e g a e ( c ad e) a d he a eh de ha be ide ified f e e e f he a e chai . The SRMG ide ifie e e a d he a e a

he ai a eh de a d ide if i g he a eh de c ce ed a d hei eed i f

i a ce he e. The igh a ee f he e e e i dica a d a ge ba ed

he i a a eh de eed . Pe f a ce c ce i g a ge h d be e e ed

a d e ed e i e a d i a eade -f ie d a (I a a d C e , 2018). T a e he e a d he he he ga i a i i e i g he igh hi g i he igh a , a a i g id i a ied- he S ai abi i Re Ma i G id (SRMG). The a i g id i a

ef a a ica c ed i hi a i a age e (I a a d C e , 2018). I i i a e ha he SRMG, acc di g C e e a . (2020), i a f e a a i g h ai ab e a c a i i , b a he h e he c a i i e i g f i ai abi i a d a a e e . I a 1_e _{ai ed ha e a a i f} _{ai abi i} _e _{i a he} _{e ai} _{a a} _ed b a a g i g h gh each e . Whe fi i g he SRMG, he ce a

i c de a a a ea ch f d c ide ed i a f he diffe e ic f he g id. P e i e ea ch he SRMG i dica e ha he a e e i c e e i i e

i di id a i e e a i a d ha i , he ef e, ide e a i e a ied a d i c i e

1 _{I a} _{, Rai e; Se i ec} _{e /A} _{cia e P fe} _{a De a} _{e f Ci i a d I d} _{ia E gi ee i g,}

(18)

e (I a a d C e , 2018). I i a i ed ha he e h d i i e-c i g.

F he e ea , i i f i e e d h he e a a i ce ca be i ed, b h

i i e efficie c a d a ia i i g adi g, b a a i g a f he e a a i ce .

The SRMG f a a i f 18 b e c ai i g de c i i f i a ai abi i

cha ac e i ic a e . I c i f i c e e e i g diffe e c i e ia a e , i h he fi h ee c e i g he igh hi g a d he a h ee c e i g he igh a . Each c i e i a e e e f a ce a ca e f 0 5, he e 0 i ie ha he c i e i i

e ed f a d 5 i dica e e fec e i g (C e e a ., 2020). The SRMG ca be b e ed i Fig e 4.

R

The fi h ee c i e ia i he SRMG i he he he c a i c e i g he igh hi g i hei e i g. A e i ed, hi efe h ee a ea ; i c i e e f he a e chai , ide ifica i f a eh de , a d he ide ifica i f he a eh de eed .

The fi c i e i a e i he SRMG i he c e age f he a e chai . The a e chai i defi ed i he GRI-S a da d a "The al e chai c e he f ll a ge f a ga i a i

eam a d d eam ac i i ie , hich e c m a he f ll life c cle f a d c e ice, f m i c ce i i e d e. (GRI, 2020). I he SRMG, he e ha i c de b h a defi i i f he ga i a i ' a e chai a e a a c ea i e e a i

f he a e chai . The i e e a i ca a be b e ed b he e i g e.g. GHG-e i i , a d if he a e c e i g a ce e i hi he ga i a i . Whe

e i g GHG-e i i , h ee e e f Sc e a e f e ed e e e he diffe e ce e f a ga i a i . Sc e 1 i c de a di ec GHG-e i i , he ea Sc e 2

i c de i di ec GHG-e i i f c i f cha ed e ec ici , hea , ea

(G ee h e Ga P c l, 2021). La , Sc e 3 e e he i di ec e i i f

di g ac i i ie , b h a d d ea , he e he c a a ha e e

i f e ce a d c . Thi a i c de e ac i a d d c i f cha ed a e ia a d f e , ced ac i i ie i e a e di a , e c. (G ee h e Ga P c l, 2021). If he

c a c ea defi e a d e f hei c e , i i a ig f c e age a d a a e e f he c a e i e a e chai . Acc di g he SRMG, he e i g f GHG-e i i ha bee ch e a a i dica a iab e hich i be f he e ai ed de The igh a (C e e .a , 2020).

The ec d c i e i c e a eh de a d h e he e a e ide ified. S a eh de e a e i di id a , g , ga i a i ha ha e i e e ha a e a be affec ed b a ga i a i ' ac i i ie deci i . I he SRMG, I a e a . (2015) e ha

he ai a eh de a e e e ( cial) a d a e (e i me al). T c e he igh hi g

i he ide if he a eh de c ce i g e e a d he a e , e.g he a he e, he cea , a d e e i e (I a e al., 2015).

(19)

i a e a d ide if eed c ce i g he i ac he e. I a (2019) e ha he ai eed f he a e ca be ide ified i h he he f he P a e a B da ie a d he UN S ai ab e De e e G a (SDG ). The e e' eed ca , a

e , be ide ified ba ed he SDG . I i a i a f he c a i i i e

a g he ide ified eed decide ha i i a . I a (2019) e ha i e he

i a ce f ide if i g ha i i a . T d hi , he SRMG c ide he Pa e P i ci e hich i a e i ica e c ide ed i hi a i a age e efe i g

ide if i g ca e e i g i aj c e e ce (Ke , 2016). F e a e, e e' eed h d be a e ia i g e i h he highe i i gi e h e i i g i e e e

e (I a , 2019b). Acc di g he GRI-S a da d , he i i i a i f a eh de eed h d be d e h gh a a e ia i a a i (GRI, 2020).

R

The ec d a f he SRMG i a e if he c a i e i g he igh a . Thi ea ha he e i g h d be c ea a d ba ed fac (I a , 2019a). Acc di g GRI (2020); Bala ce, Cla i , a d C m a abili a e i a a ec , hich he SRMG

ha ada ed.Bala ce efe he fac ha ai abi i e ef ec b h i i e

a d ega i e a ec f he c a ' e f a ce. GRI (2020) be ie e ha he

c a h d e e hi i f a i i a a ha he eade ca ee i i e a d ega i e e d i e f a ce f ea ea . The a e e be ef ec ed i

i i a ea . Cla i efe he fac ha ai abi i e be ea i acce ib e

a eh de . Th , he e ha c ai i f a i ha i i e e i g a d e e ia

f he ide ified a eh de . The a g age i he e ha a be i ab e f he a eh de , i.e. he e f ech ica e h d be a ided. C m a abili ea ha

he c a c i e c ec a d e e c i . S a eh de ca

a a e a d c a e he e ed i f a i e i e, efe ab f ea ea . I

he SRMG he e i ha e bee a i ed i he h ee c i e ia Pe f ma ce

I dica , Ta ge f S ai abili , a d Readabili ( e f ma ce c m a ed a ge a e clea l e e ed).

Whe i c e he Pe f ma ce I dica he SRMG f c ea KPI ba ed he

bjec i e ha a e ide ified a i a f he c a , i i i ed i he a e ia i a a i (I a , 2019a). The Ta ge f S ai abili h d be ba ed he ide ified

(20)

Fig e 4. The SRMG i h he mai a ea D i g he igh hi g a d D i g i he igh a a d hei e ec i e b-c i e ia . Each c e d he le el f g adi g a d i cl de e la a i f he e i eme f lfill. M dified fig e f m C e , Dahli a d

(21)

3. Theor

Thi cha e i f c he e ech ica a f hi he i . Fi , a bac g d NLP

a d ML i be e e ed, f ed b e i die de i i g he diffe e ML

de ed i he he i . The e a e he f ed b a i -de h e e a i f he e de .

3.1 AI & Natural Language Processing Methods

The h a b ai i e f he c e achi e i he i e e, e ed ce

a d e i f a i a d ea h ac f e i e e ie ce (De h a de a d

K a , 2018). A ificia I e ige ce (AI) i a ech g de e ed i i a e he h a b ai . Wi h AI, a ge e f da a ca be ce ed ha bef e a d ca h c e e

he h a b ai i fi di g a e a d a i g deci i (De h a de a d K a , 2018). O e e a e he e he c e i f a h a b ai ca be de ec ed i he ce i g

i e a d e a g age. The h a a g age i c e a d i c e ed b b ,

ge e , a d e , addi g addi i a a ce he ea i g f b h i e a d e a g age. Thi c e i a d a ce a e i diffic f a achi e i e e . Na a La g age P ce i g (NLP) i a c ec i e a e f ech i e a d a g i h ed

ce c ed, a a a g age-ba ed da a (Sa a , 2019). Si ce i i d c i i 1954, he ech i e ha e ed. The ech i e ha de e ed be b h g a a ica a d

e a ica c ec b c bi i g a i ica a d i g i ic ech i e (Sa a , 2019). ML i

a b a ch bei g ed i hi NLP ce a ge a f c ed da a. F NLP

a , b h e i ed a d e i ed ML a g i h ca be a ied de e di g

a ai ab e da a a d b e a e e (Tha a i, 2017). Fig e 5 i a i a i f a ificia i e ige ce, ML a d NLP be e de a d h he a e e a ed.

(22)

3.1.1 M L

A he a f a ai ab e da a i i c ea i g, ML ha bee i d ced a a e ac i igh f a ge, c ed, da a. ML ca be ee a edic i e a a ic , he e a e i da a a e ed edic he beha i f he b e ed a e (M e a d G id , 2016). ML ca be di ided i ca eg ie ; e i ed a d e i ed ea i g. The ca eg ie diffe i he a he i a d da a a e ai ed. I e i ed ea i g, he i da a, hich he ML a g i h i ai ed , i abe ed ch ha he i i ai ed i h he de i ed (M e a d G id , 2016). The a g i h i fi d a e i h he i a d a e e a ed a d i d ce edic i ha abe e , ee , i h d be ca eg i ed a . U e i ed ea i g d e ha e a ca eg i a i f he , ea i g ha he gi e i d e ha e a ede e i ed . Si ce he f e i ed ea i g i , he e h d i e diffic e a a e c a ed e i ed ea i g (M e a d G id , 2016).

Wi h i abi i ea f gi e da a, ML ca be ee a a bfie d f a ificia i e ige ce hich i ai i g a a i g c e ac a h a . L i g a h h a ca ea e hi g , i i ai b d i g e i e a e f he a e . B c bi i g diffe e

bfie d f a ificia i e ige ce, ch a ML, a c e ca e f e i i a a h a . A he bfie d f a ificia i e ige ce i Na a La g age P ce i g (NLP), ai i g a a i g he c e ead a d de a d e i e a h a (M e , 2020).

3.2 Text Anal sis on Sustainabilit Reports

The eed ce a ge a f da a ha bec e i c ea i g e e ed i h he g i g acce f da a d e digi a i a i . S ai abi i e i g i c de a ge a

f da a hich ca i dica e h c a ie a e i g hei c i a e i ac a d ai abi i eff . Wi h he e i g bec i g i c ea i g e e ed i i e i e a e acce ib e a f ce i g i c e . ML a d e a a a i ai abi i e i g i e ib e a f i g hi e e . The e ea ch fie d i i e a i e e

a d e i die a i g ML ai abi i e a e i i ed. H e e , he

e i i g die a i dica e he abi i faci i a e he a a ica ce b i g NLP a d ML ech i e .

S ij e (2018) a ied NLP i eg a ed e , i c di g c a e b h fi a cia a d -fi a cia a e e , d h e diffe e ML de e f ed di i g i hi g

age e i g ai abi i . The e d c e a e be b h e e i e a d

c ed, a i g i cha e gi g ead a d di ce he a age f i a ce. B i g

he ML de ; S Vec Machi e a d Na e Ba e , S ij e e ed hei

abi i di i g i h ai abi i - e a ed age f he fi a cia - e a ed age . B h

de ed e f e , i h SVM h i g a e a be e e f a ce. The

(23)

S ij e d i dica e he e ia de ec ai abi i - e a ed age i a i eg a ed

e i g he e e a i e i e ML de .

I a ece a ic e b i hed b L cci i e a . (2020), he a h e ed hei

i e e a i f a a a ica hich c d ide if ai abi i - e a ed a age . The a h a e i g he e a i e e NLP de , BERT (Bidi ec i al E c de

Re e e a i f T a f me ). Thei a ach i ba ed a e i -a e i g e h d,

c c i g e i -a e ai i h e e ce hich a e a f he 14 ecific

e i e e ed b TCFD; a i i ia i e f ed b he Fi a cia S abi i B a d ide c a ie a f a e f hei c i a e- e a ed di c e , i.e. ai abi i

e i g (Ta k F ce Clima e-Rela ed Fi a cial Di cl e , 2021). F he e he ide ified a a ia i i e f a ce f he diffe e e i , i e de e di g he

di e i f a e e f he e i , he ea e e i e e c

a e ed i a e i i a a e .

Shahi e a . (2012) e a a ed diffe e ML de a d hei e f a ce ca eg i i g

e a da a f ai abi i e . The ca eg ie c i ed f he 30 e f a ce

i dica de he GRI e f a ce ec i E i me al Pe f ma ce ec i . F

hei e i d , Shahi e a . (Shahi, I ac a d M da ha a, 2011), e i dica ed a i i a i be ee e f he 30 ca eg ie . F he b e e d , fi e e ca eg ie e e c ea ed i hich he 30 ca eg ie e e a ca ed. A a e he ca eg i a i a , i he a e d , e f ed i e ; e e ca eg i ed he d c e i e f he e ca eg ie , a d e c a ified he d c e i ab e e f a ce i dica be gi g he e ca eg . The e f he b e e d h ha he be e f i g de a Ne a Ne . The de i h e e i e c i g a d e ce i e i e a d he a h he ef e c c de ha Na e Ba e , i h e acc ac b e i e efficie , h d be ed f he e f ed c a ifica i a . The d i dica e ha a a ed c i g a d ca eg i a i f ai abi i d c e i fea ib e i g ML

e h d .

3.3 The Three NLP Models

The e a e a diffe e ech i e f ha d i g e a da a. I hi he i , h ee ML de ha e bee e a a ed; Na e Ba e , S Vec Machi e , a d Bidi ec i a E c de

Re e e a i f T a f e . The de ha e bee e a a ed hei e f a ce i

ca eg i i g ic i hi he a ica i a ea f ai abi i e i g. A h gh he h ee de be g he fie d f ML, hei ech g i a diffe e . I hi ec i , he h ee de a d hei fea e i be e e ed.

3.3.1 N B

The Na e Ba e c a ifie i d a babi i he decide ha c a a bjec bab be g . The c a ifie i ba ed Ba e The e , hich i f a ed he idea ha e i babi i ie f a e e ca be ca c a ed b i g hei e i

(24)

a ib e a e a d he a ge a e, hich i h i i ca ed ai e. The babi i f a a ic a c e i h he d c f babi i ie f each a ib e (I a e al., 2010).

( 1,..., ) = ( ) ( i ) (1)

𝑖 1 𝑛

∏

The c a ifica i i he ade b ch i g he i h highe babi i ,

= a gma ( ) ( i ) (2)

𝑖 1 𝑛

∏

he e ( ) i he c a babi i a d ( i ) i he c di i a babi i . De i e i

i ici , he c a ifie ha e e f e c a ifica i a , e ecia

high-di e i a i .

3.3.2 M

The S Vec Machi e (SVM) i a e i ed ML a g i h ha ea f e a e

da a c ec ca eg i e bjec (N b e, 2006). The ec c a ifica i a g i h i a bi a c a ifica i a ach he e he c a e a e di ided b a i ea b da . The SVM c a ifie i ba ed f ba ic c ce : he Se a a i g H e la e, he

Ma im m-Ma gi H e la e, he S f Ma gi , a d he Ke el F c i (N b e, 2006). T ge a e ie f da a, he SVM i ea i g he da a bjec a i i a high

di e i a ace (N b e, 2006). The idea behi d he SVM c a ifie i ha i ai fi d a h e a e i he high di e i a ace ha i he i ab e a di ide he da a e i c a e , ca ed he Se a a i g H e la e. T decide he e be e a a e he da a e , he SVM e ec he Ma im m-Ma gi H e la e di a ce. Thi i d e b a i g

he di a ce f he e a a i g h e a e he ea e e e i ec a he a gi f he h e a e, a d he a i g he a i (N b e, 2006). Da a i f e e c e di ide a d a aigh i e a e i a e gh i i e e e a i e . A S f

Ma gi a he SVM a e a ce ai be f i a e , e i ca c he

e a a ed h e a e i h affec i g he e (Mi a, 2020). La , e-di e i a da a ca e i e be ic e e i ib e di ide i h e h e a e, ee Fig e 6a. A

(25)

Fig e 6a Fig e 6b

Fig e 6a-6b: Ill a i g f c i ali f he Ke el f c i f he SVM. T he lef he da a

i e e ed i a 1-dime i al ace, he igh he ame da a i ma ed i a

2-dime i al ace. M dified fig e f mN ble, 2006

3.3.3 BER

BERT i a abb e ia i f Bidi ec i a E c de Re e e a i f T a f e a d i

a e- ai ed dee ea i g de f NLP. Wi h BERT bei g a dee ea i g de , hi de i a bi e ad a ced ha he e i de e ai ed. T ge a

de a di g f he c e behi d hi dee ea i g de , hi ec i i be e ha

e f d a d i eed a i e e a e i .

BERT i de ig ed e- ai dee bidi ec i a e e e a i f a abe ed e b j i c di i i g b h ef a d igh c e i a a e (De i e al., 2019). The BERT de ca ada diffe e de i ed b e i e e e e ce- a d e d

edic i e i a e i g. Thi ca be d e b addi g a addi i a a e , a e h d ca ed fi e- i g (De i e al., 2019). Ra icha di a (2021) e ai h he

BERT de ca be di ided i a , e- ai i g a d fi e- i g. The e f

e- ai i g i each he de he a g age, j i e he a chi d ea ea . Thi i d e b ai i g he de abe ed da a e diffe e e- ai i g a . The

fi e- i g i i ead ab ecif i g he de , i e eachi g a chi d a ecific bjec . The e- ai ed de i ed a a ba i i fi e- i g. He e, abe ed da a i ed eadj he e- ai ed a a e e a e he i ab e f he de i ed a ea f e. The f c i f he

de i e de e de fi e- i g. A h gh he e- ai i g i e ac he a e, he a ea f e ca diffe (Ra icha di a , 2021).

The T a f e i a e a chi ec e f NLP a ha a i d ced a a a e c e he i i a i f ec e e a e (RNN), i h i i abi i ca e

(26)

The BERT de i b i f c e -ba ed d e beddi g. I c a i he NLP de , he BERT de ca di i g i h he ea i g f he d gi e i c e

(Ra icha di a , 2021). Ta e f e a e he e e ce I efe li e i g ck m ic a d L k a ha gia ck . The d ck i a h a d gi e he c e , he

ea i g f he d ca be i e e ed. O he e beddi g de ha a e c e -f ee d e bed he d c i he a e a , i i g he c e he d i gi e i

(Ra icha di a , 2021). Wi h BERT bei g bidi ec i a , he de ca e e e he d gi e he c e .

T be e de a d he c e f he BERT de a d he e f-a e i echa i i i

e e a f he i he a f e . The a f e c i f a e c de -dec de

echa i b i i g he e c de . Thi i i ce BERT d e d ce e , b a he edic i g he e e a ce f he e . The e a e i e a e , a ac , f e c de i he a f e . Each ac i ide ica a d c ai a i-head a e i a e a d a feedf a d a e , ee Fig e 7.

Fig e 7: Re e e a i f he a f me i h e c de ack a d he c c i f each e c de . M dified fig e f m Ra icha di a , 2021

The a e i a e , hich ha e aced he ec i g i ha diffe BERT f he NLP de . I i i hi a e he e f-a e i echa i i a ied. The e f-a e i

echa i i b i ch ha he h e e e ce i ce ed i a e , i ead f

ce i g e d a a i e a i ec e de (C deE i , 2020). The

e f-a e i i i g a each d a d h i e a e he e ai i g d i he

e e ce. A ca be ee i Fig e 7, he b c i efe ed a e f-a e i b i ca ed he i-head a e i b c . T ed ce he e a i hi be ee he b e ed d i h i e f i he f e e ce, he e f-a e i i e f ed i e i e , e i g i a a e aged a e i ec f each d. Thi he i fi di g he i e ac i f d a d

(27)

Si ce a c e ca ha d e be , he i c i i g f e fi be

c e ed i be . Thi i d e b a i g d i i a d ha ca be f d i

e- ai ed e beddi g ace . The e beddi g ace i a ec f be ,

e e e i g he i i a i f he d he d . A e i ed ab e, he a e e i g f d ca ha e diffe e ea i g . T ha d e hi , a ech i e ca ed i i a e c di g i a ied he e bedded i . P i i a e c di g ge he i h e e ce e beddi g a d

e e beddi g c i e he i he e c de a e , a i e e e ed i Fig e 8.

Fig e 8: Re e e a i f embedded i f BERT. M dified fig e f m R hma , 2021

The BERT de i e- ai ed diffe e a ge c a, e.g. BERT-ba e ca ed hich ha

bee ai ed a da a e f b i hed b f B C a d he E g i h Wi i edia.

P e ai i g f BERT i d e abe ed da a a d ai he bidi ec i a a f e a d

e ai e e e a i e i ed NLP a a e e f ed; a ed a g age

de i g (MLM) a d e e e ce edic i (NSP) (Sabha a a d Ag a a , 2021).

MLM i a d a i g e i he i i g, a i g he de edic he a ed

d gi e he c e f he i e e ce. NSP i ai i g he de e e ce ai edic if a ec d e e ce be g he fi e e ce. Thi e- ai ed de i h i e- ai ed a a e e i a ai ab e e a d a diffe e a ha i e i e fi e- i g f he de . Fi e- i g f he de ca he be a ied c a ifica i a

(28)

4. Method

The he i ' g a i acce e a e he a e e f ai abi i e . T d hi a

a e a i e e h d f a e i g ai abi i e i h he SRMG ha bee de e ed

a d e a a ed. T de e he a e a i e e h d, ca ed he P e, e e a ce e

ha e bee ca ied . Fi , he e a ea e e c e ed f he SRMG i e hi g

e a gib e. Ba ed hi i f a i , e e a da a e e e c ea ed, e e e i g each f c ic i he SRMG. Th ee diffe e ML de e e he ai ed he da a e a d e a a ed. T i i e he P e, he ML de i h he highe acc ac a e ec ed f f he ai i g, f ed b ca eg i i g ee , ca eg i ed, da a. Ba ed he

da a, he P e a de e ed, e i g i a c e ed e i f he igi a

ai abi i e . The P e a he e ed e a a e he e h d' e f a ce. The e ce e i be e e ed i e de ai i hi cha e . The ce e a e i a ed i Fig e 9 gi e a i i ia e ie f he e de c ibed i he e h d.

(29)

4.1 Specif ing SRMG

The f i g ec i i e e he ce f i e e i g a d ecif i g he SRMG i e hi g e a gib e. Thi a he i i ia e a d c ea i g a da a e .

4.1.1 F

The SRMG i di ided i ai a ea , d i g he igh hi g a d d i g i he igh a . Each a ea he c i f h ee SRMG c i e ia ; Righ hi g: Va e chai , S a eh de ide ifica i , a d S a eh de eed ide ifica i , Righ a : Pe f a ce i dica , Ta ge , a d e f a ce/ eadabi i . The P e ai gi e i f a i f each f he

i c i e ia a d e e he e a a c e ed e i f he e . I he P e, he i c i e ia e e e e e ed b h ee c m e ed a ea ; Va e Chai , S a eh de , a d I dica a d Ta ge . S a eh de - a d a eh de eed ide ifica i , a e a

e f a ce i dica a d a ge , e e c e ed ge he i ce d a d he a g age c ce i g he e a ea e e c ide ed i i a , h i Tab e 1. The c e ed a ea a e a e ed e a a e , acc di g he igi a SRMG, b ba ed he a e age i he c e ed d c e . The a ea f e f a ce/ eadabi i a ee a a i de e de a ea a hi i diffic de ec i he e f a g age. Thi a a he ef e ba ed

he e a i e i f he he a ea d i g a e e , e.g. b b e i g ab e a d g a h he ided age f he e , a d he e a de ig .

(30)

Table 1. A e ie f he SRMG c i e ia a d he ed F c ic a d T ic d . RMG C Va e Cha S a eh de Ide f ca S a eh de Need Ide f ca S a ab Pe f a ce I d ca Ta ge f S a ab C A C I F C C M I

Value chain Scope Stakeholder Materialit Target

Cradle Scope (1-3) Stakeholder engagement Sustainabilit topic Goal Grave Carbon emission Stakeholder

dialogue Focus area SDG

Carbon Planet

Stakeholder

dialogue Goal ... Impact

Emission NGO Priorit area Impact ... Goal

(31)

4.1.2

The d ch e f each f c ic a e ba ed ea ie e ea ch a e he SRMG (I a a d C e , 2018; I a , 2019a; C e , Dah i a d I a , 2020) a d he GRI-S a da d . The egi i ac f each d ha bee di c ed i h e e i he fie d2_.

Si ce d a d c ce ca be ed i e e a c e , ce ai c ce had be

e c ded. F e a e, he d e le i ed i he c e f a eh de , b

a i a e ge e a a .

B

T ge e e a e e ce a i i g each c f he a i , d f i a ce e e

ide ified f each ic. The i a ce f d a i i ia decided ba ed d ha

e e e cei ed c f he ic, b e a a i g e ai abi i e . The i

f d a he e ie ed a d c e ed i c ab a i i h he e i , h e

edge i ai abi i added a ab e i igh a e he i f d a e e a a ib e. Thi a ach i i f e ced b he h a , he ce i i i a be a a e f he bia ha ca affec he e .

Va ia i i e he e a a i g ai abi i e i a ia i f e ced b he ch ice f ea ch d . A e e he ai abi i ic i e i e f he e e a

ea ch d ed i e he e he cab a ed b GRI i bei g ed.

Di c i be ee a h a d e i h c c ded ha f he e f gi i g

e e e he a ifica i e f a e a a i , he bia f ch e d i affec i g he e ega i e . I i a he a a f idi g e e e i g he P e he a e c di i e a a e he e . O he he ha d, i ha bee a e i acc ha he

ch ice f d c d ha e a i ac he e i ce e e a be i g d

ha he P e i b i f ec g i i g.

4.2 Collecting and Creating Data

The e e a d b i di g he P e a c ec e hich c d be ed f

he ce f c ea i g da a e . The ce i be acc ed f i he f i g ec i .

4.2.1 F D

Whe ha d i g NLP a d ML e h d , da a i ece a . Si ce he a ea f i e e a i i

ai abi i e i g, he da a c i ed f e f ai abi i e . The

ai abi i e e e a d e ec ed f he GRI ai abi i e da aba e

(SDD - GRI Da aba e, da e). The e ec ed e e e e i ed be a ed i h he ig i dica i g ha he e f fi ed he GRI-S a da d , a a ce ifica e ha he e i g

a c ide ed c ec e f ed b GRI. The e ec i a ha e bee a ia i ed a

he e ec i a d e a a a d a a e , e e - c a ie a ha e bee

2_{I a} _{, Rai e; Se i ec} _{e /A} _{cia e P fe} _{a De a} _{e f Ci i a d I d} _{ia E gi ee i g,}

(32)

e ea i e ec ed. I a 55 ai abi i e , c ce i g a ied i d ie , e e ch e . 4.2.2 C PDF R Wi h ai abi i e a ce da a, e e a e a e e i ed b ai ef da a. S ai abi i e a e f e e e ed i PDF f a , he ef e he fi e f he e ce i g a e ac he e a c e f he d c e ge e a da a

f he ce . Thi a d e i g a b i -i e c e e f dfmi e (Shi a a,

2019) i h f c e i g a PDF e .

4.2.3 P D

T ce e a d c e , he c e i i ia be c ea ed a d a da di ed bef e

a i igh ca be gai ed (Sa a , 2019). Thi ce i a efe ed a

e ce i g, he e ecia cha ac e a d ece a d a e e ed ge a c ea

e d c e i h. I hi jec , he fi e f he e ce i g a e e a e b ac e a d e a ace a d c e he a e a da a i e ca e a id a i g he a e d a diffe e d . I he fie d f e a a i , e a c e a e c a e . T e a e a e i ha ca diffe i c e, b i e a a i , a e i f e e e e i g a d. T e i a i i he e ce i g f i g e a d c e i a e iece e . C he ce i e i i g a e i e e ce a d each e e ce i d (Sa a , 2019). The e a e e i i g ib a ie f e i a i ha ca be ed i he e i e e ce b ec g i i g e i d a d c a i . S i i g e e ce i d i he a e f ed hi e ace . I hi jec , he i i g ha bee e f ed i

diffe e a de e di g he de . F f he de , NB a d SVM, he e i a i ha bee e f ed hi e ace . The hi d de hich i e a a ed, BERT, a e e e ce a da a i a d he e a da a i he ef e e i ed i e e ce . F he de , NB a d SVM, he e e i a i a e f ed hi e ace , f he e ce i g a c d c ed. O ce a e i e i ed, a he e a e e e ed i a ec i h di e i e a he cab a f he e (Th a , 2020). The ec i i c de d ch a i , a , he , hich a e ef f b i gi g i igh . T ed ce he di e i f he da a, ch d a e c e ed. The e i d f d a e a d a d b i -i g a i g ca be a ied ha e he e ed. The c e ec igh a i c de h g a d ch a ai a d ai i g ,

hich i be e cei ed a diffe e d , e e h gh he efe he a e d. O e a ed ce he di e i f ec i h e ge he e d i h he a e ea i g. Thi ca be d e h gh emmi g a d lemma i a i , ba ed edge f

h g (Th a , 2020). I h g , a d e i he ba e f f a d. Addi g

affi e , ai efi e , a d ffi e , he e i c ea e a e d add e ea i g he e (Sa a , 2019). The ce f e i g he affi e f he d, e di g

(33)

he ai diffe e ce bei g ha a e be a e i i g d, he ea a e a i be

f d i a dic i a (Sa a , 2019). Fig e 10 h a e a e f e he e f i g

e i g a d e a i a i he d ha i e .

Fig e 10: E am le f h emmi g a d lemma i a i k

Af e c ea i g, e i g d , a d e a i i g he da a, he ba ic e ce i g f he e a da a a d e.

4.2.4 C D

T ai he de , a da a e c ai i g e e a abe ed e e ce a eeded. L cci i e a . (2020) acc f a i i a a ach i hei e , he e ha d- abe ed da a f e e ce

a e i g ecific e i c ec ed f ai abi i a a a ed. F hi he i

ch da a a acce ib e, h a i i ic d- ea ch ba ed a i a i de ba ed eg a e e i a ed abe e e a e e ce . The d- ea ch ba ed a i a i de a ch e i ce a c a ic a i a i de b i g he i e i a e . Si ce a e e f ai abi i e i g SRMG f c e ecific ic , he a i a i had be de ig ed i h e ec hi . R E A d- ea ch ba ed a i a i i a e h d fi e ec i f c ide ed i a ce i a e e i e d c e . Reg a e e i a e a a f c c i g ea ch a e . The a e ca he be ed a a a ea ch a ch d a d h a e (Wi dha , 2018). F he d- ea ch ba ed a i a i he e f i g eg a e e i a ec g i e ecific d , i e ec i e f i i f ec i , a d e ac he f e e ce. A eg a e e i i h i e ch ha efi e a d ffi e a e ig ed, a d a i f ec i f he d ca be ide ified. A e a e f h a eg a e e i ca

(34)

Fig e 11: E am le f e l f m diffe e eg la e e i

I a eg a e e i , [a- ] e e e ha he ea ch- d i f ed b a e e f a , a d he f ed b a * he ea ch- d ca be f ed b i e e e . The \b

i dica e he b da f a d a d he f i g e e a ch he d (L e a d

R e , 2014). F Fig e 11 i ca be ee ha \b f ed b ai ead e ai i g bei g a ched. The ef e, i i i a ef ec ha i f ec i f he d i be

f i a ce f he a he f a i g he eg a e e i .

-

-The d- ea ch-ba ed a i a i de a de ig ed ea ch f e e ce

c ai i g e e a d a d h a e , hich e e e e ed a eg a e e i fi d a i f ec i . The be f eg a e e i a ied de e di g he ic f he ea ch a ea, f 3 15, ee Tab e 1. F e a e, f he ic a ea f a e ia i , ma e iali i

a e e ia d. The d- ea ch-ba ed a i a i de he ca e he e e ce

c ai i g ma e iali .

The e e ce ha he a i a i de f d e e a e e he aced i a da a f a e

a d abe ed i h a e. Wi hi ML, a c b e he ai i g a de i ha d e

i ba a ced da a e . T ai i g a i ba a ced da a e i f e i e i ab e b ca ead a bia ed de a d he aj i c a e a e (Fe de e al., 2018). Thi ca ca e

b e f de i g he i i c a e , hich a e f e he e f i e e edic c ec . Ma ML de , i hi bi a c a ifica i , a e de e ed ha d e da a e

ha a e e e e i ba a ced (K a c , 2016). H e e , i ce he da a e a c c ed f c a ch, he deci i a ade c c a ba a ced da a e a id he bia i ca b i g he e . T a e he da a ba a ced he a e a f da a

i / e e ce a a d ic ed f he e a d abe ed i h a e . A e a e i gi e i Tab e 2, e e i g a a f a da a e .

(35)

ea 6000 da a i . The da a e f he f c ic Val e Chai a d Ma e iali had de 1000 da a i hi e he e ai i g h ee f c ic , Sc e, S akeh lde , a d

Ta ge , a had e 3000 da a i . The da a e f abe ed e e ce a he ed ai he h ee NLP de ; NB, SVM, a d BERT.

Table 2: Pa f da a e f f c ic ma e iali (f m Bj B g e ).

Label Sentence

1

main focus areas for sustainabilit at bj rn borg working conditions in factories have been a priorit for bj rn borg for man ears.

1

bj rn borg sustainabilit report 2017 photo: alex povol ashko, unsplash in addition to the focus on social conditions through the amfori bsci code of conduct, we have identified three specific focus areas for our sustainabilit program in coming ears.

1

the roadmap, with its targets and activities, is integrated into the priorit areas (top 10) set for each ear.

0 we expect these kinds of requirements to increase further in the future.

0 we believe that the examination has provided us with sufficient basis for our opinion. 0

our sustainabilit work is integrated in the core of our business and takes a central part in our product development strateg .

0 alwa s strive to be better, never stop innovating. 1

to tackle this, we have carried out a thorough materialit anal sis to identif the emission hot spots along our value chain.

1

the report was reviewed b the highest executive management and external assurance has been performed for selected indicators (clearl marked where applicable, as well as in regards to materialit and stakeholder engagement).

1

process for identif ing the most material topics and their boundaries* during 2016, we aligned our materialit matrix with our new sustainabilit strateg .

1

we developed this strateg in close dialogue with external and internal ke stakeholders and updated our previous materialit matrix with ke takeawa s from this process.

1

for the sake of user friendliness, these were themati- call clustered into 26 focus areas and gathered in a materialit matrix on p. 112. in line with our new strateg , some material topics were re-grouped in these focus areas.

4.3 Classification Modeling

F e i die , i bec e c ea ha he i ab e ch ice f c a ifie a d i e f a ce de e d he ecific a a d i e e i i e . I i he ef e diffic i ad a ce hich c a ifie i gi e he be e . Th , hi he i i i i ia

e i g he e f a ce f h ee diffe e c a ifie decide hich e i i ab e f f he i e iga i ge he be e . A g he e h ee, a e c

c a ifie f NLP a d he hi d i a e ece i d ced c a ifie ha ha h

(36)

4.3.1 C

Na e Ba e (NB) a d S Vec Machi e (SVM) a e c ed c a ifie

f e c a ifica i a . NB ha , de i e i i ici a d ai e a ach, e b h c a i a efficie a d a cce f edic i e e f a ce, hich i h i i a

i hi NLP (Che e al., 2009). The c a ifie a he ef e ch e e a i e h e i

e f ed c a ed i h SVM a d he e ad a ced BERT. SVM i a he a

c a ifie f bi a c a ifica i (Ja e e al., 2013). The c a ifie i i ab e f

high-di e i a da a a d i he ef e i ab e f e a c a ifica i a . The - i ea

c a b da a e he c a ifie e c e ha he e i i ic c a ifie , hich

SVM i a e e i f. The SVM c a ifie i f e c ide ed a e f he be

c a ifica i a g i h (Ja e e al., 2013). Be ide he e c a ifie , hich a e b h

c ed f e c a ifica i , a e ece i d ced a g i h a e a i ed.

Thi a g i h i a BERT (Bidi ec i a E c de Re e e a i f

T a f e ) a d ha h i i g e NLP a . The c a ifie i b i a

dee ea i g a chi ec e i g a f e a d ha e aced e i

a chi ec e a RNN (Rec e Ne a Ne ) f a i NLP a (Ra icha di a ,

2021).

4.3.2 I C

T NB a d SVM, he da a e , c i i g f a CSV fi e i h e e ce a d abe , had be i ed. Thi a d e i g he P h Pa da ib a , c ea i g a c . The c

a he c ea ed, e i ed, a d e a i ed i g i i ie i hi he Na a La g age T i (NLTK) ib a bef e he da a e a a d i i 80% ai i g da a a d 20% e i g da a. B h NB a d SVM e e i e e ed i g he Sci i - ea ib a . F SVM he h e a a e e ke el a d he C- al e, a a he eg a i a i a a e e , e e ed. The C- a e a e 0.1, 1, 10, a d 100 a d f he e e he a e i ea , bf , ig id , a d e e ied. The be e f a ce a gi e i h he C- a e e 1 a d he e e e i ea . NB ha h e a a e e e i de i e he e f a ce. I e e i g BERT a d e i g he P T ch ib a . The Pa da ib a a ed i he da a e he BERT de . P e ce i g f BERT a d e i g he T a f me ib a a d e h d i hi he c a Be T ke i e . I hi a , he e i i a d e i ed i he a e a a he e- ai ed c (H ggi g Face, 2020)3_{. The}

e- ai ed c a e be -ba e- ca ed . Si ce BERT i e- ai ed a c i ad a ce, he fi e- i g f he de i a he fa a d he be f e ch f ai i g d e eed be a ge. The be f e ch a e be ee 2-20 a d a fi a e

4 i ce a highe be f e ch e e iceab e i h de a ed

i e e f e f a ce.

(37)

4.4 The Protot pe

The P e a de e ed e ab e e e a a i . E a a i f ai abi i e

a e a he e ai a a ed b a a g i g h gh each e . Whe fi i g

he SRMG, he ce a i c de a a a ea ch f d c ide ed i a f

he diffe e ic f he g id. Thi e h d ai e e he d ea ch a f he a e e ce . F e a a i g he h ee diffe e de , he e i h he be

c a ifica i e f a ce a ch e b i d he P e. The P e c d c ib e

a e e e i e c e f he c e . I d ce a c e ed, h e e i f

ai abi i e hich ai faci i a e he a e e . The P e a b i ba ed he e e ce hich he ch e de f d e e a a d a di ided i h ee a acc di g he SRMG a d he f c ic .

4.4.1

Whe e a a i g he a i f a ai abi i e , I a 4 _{a ed ha he ch ice f}

d had a i ac he i e i f hei e i g. Whe e a a i g he e i h a a a e h d, hich i bei g ed da , e a ach i ea ch f d c

ed he e i g f each f c ic. O e e a e f hi a ach i a h he c a i e i g f i a e chai . If he e d e e e i c de he d a e chai , i i ib e a e ha he a e acc i g e f hei a e chai

a d he eade f he e . Thi a e hi g I a 5 _{i ed} _{a i} _{a he}

e a a i g he e i ce i gi e he eade h e f he e a a i a i i ia idea f h e he e i g i d e. A a a a i he eade i h hi i f a i , a d ge

a e ie f d ha a e c ide ed i a he e i g he diffe e f c

ic a d if e i i g , a g a hica e e e a i f hi i f a i a ided. Thi

fea e a h gh he eed he e a a i ce i ce he eade ge d

eed d he i i ia ea ch f diffe e d . Fig e 12 i a e a e a e f h he e i e ce f i a d i e e ed he eade .

Q ali Scie ce a U a a U i e i .2021, Mee i g Ma ch 4 h

(38)

Fig e 12: Ill a i f d f e e c f d c ide ed im a he e i g ecific f c a ea . T b-fig e h he d f e e c f he c m e ed a ea alue

chain (f m H dai e ).

4.4.2 H

Wi h a ai abi i e e e i g a -fi a cia i f a i f a c a i , fi di g ec i f e e a ce i e f he c ib i g fac he e a a i ce bei g

i e-c i g. Whe de ig i g a d de e i g he e e a i f he c e ed e i

f he e , he P e, hi a f he e a a i ce a a e i c ide a i . T fi d e e a ec i i a decided high igh e e ce ha he de c a ified a be gi g he ic. The e f high igh i g ecific e e ce a g ide he eade

e e a ec i , b i bei g ab e ge he c e f he age he e he e e ce i

e e ed. Thi a f e e i g he c e ed e i a c ide ed ide he

eade i h e gh i f a i a e a a id e a a i , hich c d e ia be b

bei g e e ed a a i a i f each f c ic.

The high igh i g f c i a b i i g he Fi a d P M PDF ib a ie . The e e ce c a ified a e e a f each f c ic e e ed i a dic i a , i h he e i i g

hich age he e e ce e e f d. B i g h gh he dic i a f e e ce , he f c i a b i ea ch f each e e ce he ecific age, e i g a PDF i h he

(39)

Fig e 13: E am le f a highligh ed age f m he P e (f m S a E e ).

4.4.3 F P

T di ide he a ea f he SRMG, f age f each c i e i e e c ea ed. The e gi e he a e a e ie f he SRMG e e f he ecific a ea. I a ide he a e i f a i f he a ea, e.g. f Va e Chai , gi e a defi i i f he a e chai , a d de c ibe he e e f c e. F each e , he a e a ge a e ie f he d f e e c , a d c , f ha ecific f c a ea.

(40)

4.4.4 L

The e f he P e a e ab e a e e a a i a e a ia i i g adi g

a d i e e i ed e f a a e e i h a c e ed e i . T e ab e

i e e , a ch ice e e ade a e he P e a c e ed a d h a

ib e.

The be f e e a age f d b he ch e de a ied g ea de e di g he f c ic a d e . T ee he be f age a ea ab e e e , c ai e e

e . T ge id f age i i g e e e ce , hich i c i f e a e

ab e f c e he a f a e , a c ai a e he e . De i e

hi i i a i , he age e e c ide ed e . He ce a addi i a e ai a e hich e e ed 3-4 age i c di g he highe a f e e a e e ce f each f c ic. C e e a . (2020) a d he SRMG a e he i a ce f e i g he igh

hi g. Thi a c e ed i he P e f c he age i h he highe be f e e a e e ce . D e hi fac , he P e e e , f each f c ic, he age i dec ea i g de f e e a ce. Page i h he highe be f e e ce a e he ef e

aced fi a d i ch gica de .

4.5 Evaluation

E a a i g a e a i e de i a c cia e e e f a d i ce i a e ea che ceed i h he de ha d ce he i i g e . Thi ec i de c ibe he

ce f e a a i g he ML de a d he P e e e a a i .

4.5.1 C E

C M

F e a a i g he e f a ce f c a ifica i b diffe e ML de i i c

c e ac he c f i a i . The a i e e he e i a, f e , 2 2 ab e

i a i g he edic ed c a e i c a hei ac a c a e . Thi e a a i e h d i

h e f ed e i ed ML de . B i g diffe e ea e , he di ib i f

edic i i he ab e ca be f he a a ed gi e e de a di g f he e . T de a d he ca c a i behi d he ea e ha i be e e ed, he c f i a i i i a ed i Fig e 15. Whe efe i g he diffe e ca e e e e ed i he a i

(41)

Fig e 15: A 2 2 c f i ma i , e e i g edic ed cla agai ac al cla .

F e a a i f diffe e de , he e a i hi f he e f ca e ca be ed

ca c a e eca , eci i , F1- c e, a d acc ac .

(3) 𝑅𝑒𝑐𝑎𝑙𝑙 𝑇𝑃 𝑇𝑃+𝐹𝑁 (4) 𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑇𝑃 𝑇𝑃+𝐹𝑃 (5) 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 𝑇𝑃+𝑇𝑁 𝑇𝑃+𝑇𝑁+𝐹𝑃+𝐹𝑁 (6) 𝐹1 − 𝑠𝑐𝑜𝑟𝑒 2*𝑅𝑒𝑐𝑎𝑙𝑙*𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 𝑅𝑒𝑐𝑎𝑙𝑙+𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛

The eca ea e h e he de i a edic i g he i i e ca e f a he ac a i i e . P eci i i he he ha d a ea e f h a f he ca e edic ed a i i e a e ac a i i e (P e , 2008). P e (2008) e ha i e h b h eci i

a d eca ac i idi g i f a i ab h e he de e f ega i e

ca e a d e he e f i e e ea e de e di g he e f he de . I

ML a , eca i f e ig ed a d e f c i eci i i ce i gi e i f a i h c fide a c a ifie i (P e , 2008). The ea e e f acc ac i c

ed e a a e a de . The ea e e i a i dica f h a edic i e e c ec c a ified, a d h gi e a g d ea e f h e he c a ifie e f . The

a ea e e c c ed f he c f i a i i he F1- c e, hich i a ha ic

ea f eci i a d eca (Ha d, Ch i e a d Ki ie e, 2021). Acc ac i a c ed ea e e f c a ifica i b e , b he F1- c e i i c ea i g ed f c a ifica i b e i h he i ba a ced di ib i f c a e i he da a e (Ha d, Ch i e a d Ki ie e, 2021).

The e i e c i ici agai he e h d f e a a i g a de i h he ea e f he c f i a i , ea i g ha i d e ca e a a ec (P e , 2008). F hi he i ,

he e ea e ha e bee c ide ed ide e gh f da i e a a e hich e f

(42)

F - F

-Whe e a a i g he e f a ce f a de i g he c f i a i i i e e a

decide i ha a he de h d e f be e . I i i a f he de

edic a d c ec c a if i i e ca e , i he abi i c ega i e ca e ha i

e i a ? I he d , i he de e e i i e a d i c ec c a if i g

i i e ega i e ca e . Whe e a a i g he h ee diffe e de ed i hi he i , he

de i be c ide ed e a a d fa e i i e a he ha fa e ega i e . The

de i i he d be c ide ed e e a a d i c ec c a if i g

ega i e ca e a i i e (FP) a he ha i c ec c a if i g i i e ca e a ega i e

(FN).

4.5.2 G E

P e i die f e a a i g he SRMG ha e bee ca ied b g adi g a i e a d a a i g he e . The e a a i ce f hi he i ha bee i i ed b he e a a i ce ed b C e e a . (2020) i hei die he a i g id (SRMG). The e a a i a ca ied 5 i di id a , 2 eache a i a f he c e

S ai able de el me a d CSR , a d 3 c ab a i a i a i he De a e f Ci i a d I d ia E gi ee i g, Q a i Scie ce, b h g affi ia ed i h U a a U i e i , Ca G a d. The e i di id a i be efe ed a he a e . C e e a . a g e i hei a ic e A e The Re i g he Righ Thi g a d A e The D i g I Righ ?

(2020) ha a e de d i g Ma agi g S ai abi i a U a a U i e i a e

c ide ed e e gh edge he bjec f ai abi i a e a

a e e i g he SRMG. The i di id a e f i g he e a a i a e c ide ed e a i i a e e f edge a d, he ef e, c ide ed c edib e f he e a a i . Bef e a i g a e he e a a e had ead he a ic e; A e The Re i g

he Righ Thi g a d A e The D i g I Righ ? A Mea eme Ma i G id f E al a i f S ai abili Re b C e e a . (2020). The he i i ba ed hi a ic e a d a

c ide ed ide he a e i h he e e ia f a e iab e a e e .

F he e a a i , 30 ai abi i e ha e bee e ec ed. O f he e, 20 ca e f a e i e ea ch a ic e he e he SRMG ha bee e a a ed b (I a a d R a , 2020) a d e addi i a e ea ch a e , c d c ed i c ab a i i h B a i ia

e ea che , hich ha e bee b i hed. Be ide he e 20 e , he e a e e

addi i a ai abi i e . The addi i a e e e e ec ed i h he a e e h d a I a a d R a (2020) a ic e e e . The ch ice f c a ie i ba ed he

i f N dic c ie i hi c c i , hich a e e e ed he eb i e La ge

C m a ie 6_{. Of he 30 e ec ed} _{ai abi i e} _{, 15 a e i eg a ed i h he a} _{a e}

a d 15 a e i de e de ai abi i e .

(43)

The be f e ch e f he e e a a i a ba ed a i ica

ec e da i . The ce a i i he e i a i ic i ie ha , gi e a a e i e big

e gh, he a i g di ib i f a a iab e' ea i a i a e a a di ib i (F , 2018). Wha i c ide ed a big e gh a e i e a diffe de e di g he da a e , b 30 i a a he e ab i hed i i he ha d i g a i ic (Ka a d Ra a i ga , 2013; F , 2018). Wi h i i ed acce a e , 30 ai abi i

e e e c ide ed a i ica fficie f he e e a a i .

A 30 ai abi i e e e a e ed f i e , i e i h he c e i a

e i a d i h he c e ed e i . The a e a ed be ee 10-15 e

c e i a a d 10-15 e c e ed. N e f he a e a e ed he a e e

e ha e i e.

F c a i , da a e e c ec ed f each a e a d e . The a e e ed

he i e i f each a e e a d he g adi g f each c i he SRMG f each

a e e . The a e a ga e feedbac he c e ed e i a d ha i a i e a e i h i . Each a e had a e a e i h a e f e i c ce i g he de . t-test T c a e a d a a e he a e e i h he c e i a e h d a d he P e a - e a e f ed. The - e i a e f i fe e ia a i ic ed de e i e if he e' a

ig ifica diffe e ce be ee g ' ea . Thi e f e i c ed f

H he i e i g. (Ha e , 2020)

I hi he i , a ai ed - e a e f ed e a a e he he a diffe e ce c d be ee be ee he a e e e h d , b h i e a d g adi g a ia i .The - e a ca c a ed i G g e hee .

M C

T a a e he g adi g a ia i , he C efficie f Va ia i (CV) a b e ed. The CV i

c ed i a a ica che i a d i a a da di ed a i ica ea e f he

di e i f da a i a d he ea i a da a e ie (Ha e , 2021). I he a e e he e he a ia i i b e ed, he a e i e e a a he a e e i acc di g a ca e, he ef e he CV ha bee dified. The CV e ai i g he e a i be ee he S a da d De ia i ( ) a d he ea ( ) f he eaσ µ e e .

(7)

𝐶𝑉 σ

µ × 100

(44)

(8) 𝐶𝑉 σ 2.5 × 100 The CV gi e he e a i e a e, e e ed i e ce , a d gi e a e c a ab e ea e e ha b e i g he S a da d De ia i (Ha e , 2021). E C F he e e a a i , he f e i e e e e ed b S edi h Re ea ch C ci (2002) he ec i f e a i f a i e e f ed. The f e i e e a e; e i e e f i f a i , he e i e e f c e , he e i e e f c fide ia i , a d e i e e f i i a i .

The e i eme f i f ma i , hich i e d i f a ici a f he ai f he d a d he e f hei c ib i , a f fi ed b idi g a a ici a a

i f a i a d c e . The d c e a i f ed f he a i e f a ici a i g,

ha i e addi i a e f he fi e i e e . Si ce he a ici a i a a ed a a , he e i eme f c e a c ide ed f fi ed a he a ici a c d decide a ici a e , a d he he fi he ided f . Recei i g a fi ed-f a he ef e i e e ed a a a ici a ' gi e c e . The e i eme f

c fide iali i ea ec e a da a. The e a a i d ha bee a , h e a da a ha bee c ec ed. The f h e i e e , he e i eme f

ili a i , i a i g ha c ec ed i f a i h d be ed f he e ha f cie ce. S edi h Re ea ch C ci (2017) di c e he ha i g f i f a i i cie ce. Wi h cie ce bei g e e i e i i e c aged ha e i f a i , a d he S edi h

Re ea ch C ci h ea ha i i a a ib e i e a ici a f a d

ha he c ec ed i f a i i be ed f he e . Thi ha bee a e i

acc a d a ici a e e i f ed ha he a e i be ed f cie ific e ,