Design and Implementation of German Legal Decision Corpora

Stefanie Urchs

1 a

, Jelena Mitrovi

2 b

and Michael Granitzer

2 c

Chair of Software Engineering for Business Information Systems, Technical University of Munich, Garching, Germany

Chair of Data Science, University of Passau, Passau, Germany

Keywords:

Datasets, German Law, German Legal Writing Styles, Machine Learning, Natural Language Processing.

Abstract:

Law professionals are wordsmiths, their main tool is language. Therefore, the ﬁeld of law produces a vast

amount of written text. These texts have to be analysed, summarised, and used in the creation of new text,

which is a task that reaches the limits of what is humanly possible. However, it is possible to automate this

analysis by using Natural Language Processing techniques. To perform these techniques (annotated) text

corpora are required. Unfortunately, publicly available (annotated) legal text corpora are rare. Even scarcer is

the availability of (annotated) German legal text corpora.

To meet this need for publicly available German legal text corpora this paper presents two German legal text

corpora. The ﬁrst corpus contains 32,748 decisions from 131 German courts, enriched with metadata. The

second one is a subset of the ﬁrst corpus and consists of 200 randomly chosen judgements. In these judgements

a legal expert annotated the components conclusion, deﬁnition and subsumption of the German legal writing

style Urteilsstil. Furthermore, the paper presents experiments on these corpora.

1 INTRODUCTION

In the ﬁeld of law a vast amount of written text is pro-

duced every day. It is not humanly possible to read

every text and process all these information. There-

fore, law professionals are in need of computational

help. Computers can process thousands of documents

in mere seconds. By utilising machine learning it is

even feasible to go beyond pure digitisation of a docu-

ment. It is possible to analyse its semantics. However,

these machine learning methods need legal text cor-

pora for training purposes. Unfortunately, there are

only few annotated legal corpora publicly available,

even less annotated German legal text corpora. Non-

annotated content is easily available, however, this

content has to be crawled and processed for Natural

Language Processing usage.

In this paper we present two German legal text

corpora. The ﬁrst one consists of 32,748 decisions of

131 German courts that are enriched with meta data.

The second corpus is formed from a subset of these

decisions. It consists of 200 randomly chosen judge-

ments, which are annotated with the components con-

clusion, deﬁnition and subsumption of the German le-

https://orcid.org/0000-0002-1118-4330

https://orcid.org/0000-0003-3220-8749

https://orcid.org/0000-0003-3566-5507

gal writing style Urteilsstil (appraisal style). These

annotations can be utilised to train machine learning

models to automatically detect parts of German legal

writing styles.

The remainder of this paper is structured as fol-

lows. Chapter 2 introduces the related work of this

paper. The ﬁrst corpus is presented in chapter 3, the

annotated judgement corpus in chapter 4. In chapter

5 possible use cases for the corpora are discussed and

chapter 6 concludes the paper.

2 RELATED WORK

This chapter introduces a selection of already avail-

able legal corpora. Furthermore, German legal writ-

ing styles are discussed, especially the Urteilsstil.

2.1 Legal Corpora

In 2006 Reed et al. (Reed, 2006) presented the ﬁrst

corpus containing analysed legal argumentation. This

corpus contains text from various domains, including

judicial summaries and discussion. Due to a security

breach this corpus is no longer available. Researchers

from the same group presented another legal cor-

pus: a collection of 47 documents from the Euro-

Urchs, S., Mitrovi

c, J. and Granitzer, M.

Design and Implementation of German Legal Decision Corpora.

DOI: 10.5220/0010187305150521

In Proceedings of the 13th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2021) - Volume 2, pages 515-521

ISBN: 978-989-758-484-8

515

pean Court of Human Rights (Palau and Moens, 2009;

Mochales and Moens, 2011), annotated with argu-

mentation components and argumentation schemes.

Unfortunately, they did not publish this corpus.

The US courts publish their cases online. Sugath-

adasa et al. (Sugathadasa et al., 2017) compiled

22,776 cases from the United States supreme court

into a corpus and published it in combination with

other legal cases. Judicial decisions from the UK

form the basis of the corpus of Rizzo et al.(Rizzo and

erez, 2012). This corpus was created and published

because most legal corpora are too small to act as nor-

mative reference or are not publicly available. The au-

thors cite the corpus of Favretti et al. (Rossini Favretti

et al., 2001) as one of the few corpora comparable to

their work. It contains European legislation in English

and Italian. A corpus consisting of Italian legislation

in Italian and German is presented by Gamper (Gam-

per, 2000). However, the paper does not mention pub-

lication of this corpus.

The international research group Computer As-

sisted Legal Linguistics (CAL

)

compile legal cor-

pora, containing legislation from around the world.

Unfortunately, these corpora are not publicly avail-

able.

These selected corpora might be valuable for re-

search if they would have been published. To ﬁll the

gap, left open by current research, we publish our cor-

pora on the open science platform zenodo

2.2 German Legal Writing Styles

The legal way of argumentation is the so called le-

gal syllogism (Alexy, 1983). The example in table 1

illustrates a legal syllogism. At ﬁrst a general norm

is cited as ﬁrst premise, subsequently the real world

case is stated and a conclusion is deducted from the

two premises. In other words, general norms/laws are

used to solve concrete problems, by logically deduct-

ing the solution from the norms/laws (Weber, 2018).

The used example shows how similar the thinking

in mathematical logic and legal reasoning is. How-

ever, in the legal reasoning process one encounters

some difﬁculties. For example: What is rain in a

legal sense? It is necessary to deﬁne ”fuzzy” terms

of norms/laws. When rain is deﬁned one has to look

at the concrete problem and ask if that water coming

from the sky is rain in the sense of the common rules.

Here a subsumption is necessary (Weber, 2018).

The German legal education teaches law students

to work on legal problems in two distinct styles. First

https://cal2.eu/index.php

https://zenodo.org/record/3936726#.X1enLIvgomL

Table 1: Example of legal deduction with a syllogistic in-

ferring method (Alexy, 1983).

General

norm / law

Major

premise

A soldier has to say

the truth in ofﬁ-

cial business affairs

(§13 Abs. I Sol-

datenG).

Concrete

problem

Secondary

premise

Mr M. is a soldier.

Final

sentence

Conclusio Mr M. has to say

the truth in ofﬁcial

business affairs.

the Gutachtenstil (appraisal style), which is used al-

most exclusively until the ﬁrst state exam. After the

ﬁrst state exam, the Urteilsstil (judgement style) is

used. Both styles share the components definition

and subsumption. The appraisal style is concluded

in a different way than the judgement style, therefore

this component does not match.

The introduced writing styles might be used in

other countries in a similar way. However, this pa-

per introduces only the German way of formal legal

writing, other countries are not considered.

Urteilsstil. The Urteilsstil begins with the conclu-

sion and proceeds with the reasoning. The most basic

version of the Urteilsstil consists of three stages:

1. the concrete legal consequences

2. followed by the abstract legal facts and conse-

quences (i.e. exact wording of the law)

3. ending in the concrete facts

Between the second and third item a Feststellungssatz

(determination sentence, the result of the subsump-

tion) states if the legal requirements are fulﬁlled or

not (Danger, 2005).

The determination sentence and the concrete facts

together form the subsumption. This leads to the fol-

lowing structure as shown in Table 2.

In practice this basic schema is mostly extended.

If norms refer to other regulations, these must ﬁrst be

deﬁned before a subsumption can be made.

3 GERMAN LEGAL DECISION

CORPUS

This chapter introduces a legal decision corpus, en-

riched with metadata and saved in an easily accessi-

ble JSON format. To show a possible use-case of the

corpus, an experiment on the corpus is presented.

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

516

Table 2: Example of Urteilsstil with named sub-

components according to (Danger, 2005).

Conclusion Overall

Result

The claim is well

justiﬁed. The de-

fendant owes the

plaintiff 1500D in

damages.

Deﬁnition Abstract

Legal

Facts and

Conse-

quences

If an object is

damaged by an

animal,then ac-

cording to §833 S

1 BGB the person

who keeps the

animal is obliged

to compensate

the injured person

for the resulting

damage.

Subsumption Determina-

tion

Sentence

These conditions

are fulﬁlled here.

Concrete

Facts

The defendant’s

dog scratched the

paint of the plaintiff

’s car. The repaint-

ing caused costs of

1500D.

3.1 Corpus Design

The decision corpus is a collection of the decisions

published on the website www.gesetze-bayern.de

At the time of the crawling the website offered 32,748

decisions of 131 Bavarian courts, dating back to 2015.

The decisions are provided from the Bavarian state af-

ter the courts agreed to a publication. All decisions

are processed by the publisher C.H.BECK, commis-

sioned by the Bavarian state. This processing includes

anonymisation, key-wording, and adding of editorial

guidelines to the decisions. Intentionally all decisions

are included to provide a complete corpus of German

law. Normalisation measures can be applied depend-

ing on the task at hand.

The decisions have 22 different types, mostly

resolutions (Beschluss, 16,711 (51%)), judgements

(Urteil, 11,955 (37%)), or end-judgements (En-

durteil, 2,748 (8.3%)). From the 131 courts the most

decisions are made by VG Munich (8,028 (25%)),

VGH Munich (7,141 (22%)) and OLG Munich (2,337

(7%)). The courts of the Bavarian capital are gener-

https://www.gesetze-bayern.de/ accessed and crawled

on 2020-05-13

ally the most busy ones. The VGH Munich is the Ver-

waltungsgerichtshof (Higher Administrative Court)

of Bavaria, the highest instance for all topics regard-

ing the public administration. Below the VGH exist

six VG (Verwaltungsgericht / Administrative Court)

almost all administrative regions have their own VG,

except for lower Bavaria, which fall under the juris-

diction of VG Regensburg. This type of court (VG) is-

sues the most decisions in the corpus (22,155(68%)).

Figure 1 depicts that over the year courts are mostly

consistent in issuing decisions. The ﬁrst three months

are busier that the rest of the year. August, September

and December are the months with the least number

of issued decisions.

Figure 1: Number of decisions issued per month.

Every decision is saved in the following JSON for-

mat:

1 {

2 " meta ": {

3 " m eta _ti tle " : " " ,

4 " c ou rt " : " " ,

5 " d e ci s io n _s t yle " : " " ,

6 " date ": " " ,

7 " f ile _nu mbe r " : "",

8 " t it le " : " " ,

9 " n orm _ch ain s " : ["" , " "],

10 " d e ci s ion _ gu i del ine s " : [ "", ""] ,

11 " k eyw or ds " : " " ,

12 " l owe r_c our t " : ["" , " "],

13 " a d di t ion a l_i nfo r mat ion " : " " ,

14 " d e ci s ion _re f ere nce " : " "

15 } ,

16 " d eci s io n _t e xt " :{

17 " t en or " : [ " " ," " ],

18 " l ega l_f act s " : ["" , " "] ,

19 " d e ci s ion _re aso n s " : [ " " ,"" ]

20 }

21 }

The metadata are provided by the publisher

C.H.BECK. The following is a short description

of the ﬁelds:

• meta title

Title provided by the website, it is used for saving

the decision

• court

Issuing court

• decision style

Design and Implementation of German Legal Decision Corpora

517

Style of the decision

• date

Date when the decision was issued by the court

• file number

Identiﬁcation number used for this decision by the

court

• title

Title provided by the court

• norm chains

Norms related to the decision

• decision guidelines

Short summary of the decision

• keywords

Keywords associated with the decision

• lower court

Court that decided on the decision before

• additional information

Additional Information

• decision reference

References to the location of the decision in beck-

online

• tenor

Designation of the legal consequence ordered by

the court

• legal facts

Facts that form the base for the decision

• decision reasons

In depth explanation of the court decision

The corpus is published on the open research platform

zenodo

3.2 Experiments

As mentioned above, the corpus is intentionally not

normalised towards a speciﬁc task. Therefore, the

corpus can be used for many different research ques-

tions. One of possible question is: “Is it possible to

detect the type of a decision”?

Overall 22 types of decisions are contained in

the corpus. However, 16 of these types are rep-

resented by less than 100 decisions. Furthermore,

the corpus contains nine different kinds of judge-

ments and six different kinds of resolutions. To

form a balanced dataset all judgements get the la-

bel judgement and all kinds of resolutions are la-

belled as resolution, everything else is labelled as

other. Leading to 17,013 (52%) decisions labelled

as resolution, 14,818 (45%) decisions labelled as

https://zenodo.org/record/3936726#.X1enLIvgomL

judgement and 917 (3%) as other. As a feature

for the classiﬁcation term frequency - inverse doc-

ument frequency (tf-idf) is calculated for the deci-

sion reasons of each decision. The scikit learn

Tﬁd-

fVectorizer

is used. With this feature a logistic re-

gression

(LR) and a linear support vector classiﬁ-

cation

(SVC) are trained in a one-vs-the-rest multi-

label scheme. This means that one class is accepted

as positive and all others as negative. For each class

a classiﬁer is trained and ﬁnally the prediction with

the highest probability is returned. Table 3 shows the

results of training on 80% of the data and testing with

the remaining 20%. Although a very simple feature

was used, the classiﬁcations are very good. The SVC

slightly outperforms the LR.

This experiment is only meant to illustrate the use-

fulness of the corpus. For this reason, no further dis-

cussion of how the results of the classiﬁcation were

obtained will be given here, this will be discussed in

future work.

Table 3: Results of training a logistic regression and a sup-

port vector classiﬁcation on the decision corpus. The sup-

port vector classiﬁcation slightly outperformed the logistic

regression.

Feature/

Classiﬁer

Preci-

sion

Recall F1-

Measure

Accur-

acy

tf-idf/ LR 0.96 0.80 0.85 0.96

tf-idf/

SVC

0.97 0.88 0.92 0.97

4 GERMAN LEGAL

JUDGEMENT CORPUS

As stated above, the judgement corpus consist of 200

randomly chosen judgements that are annotated by a

legal expert, who holds a ﬁrst legal state exam. Ac-

cording to (Wissler et al., 2014) multiple experts are

needed to create a gold standard corpus. Due to ﬁ-

nancial, staff and time reasons the presented iteration

of the corpus was only annotated by a single expert.

In a future version several other experts will annotate

the corpus and the inter-annotator agreement will be

calculated.

The decision reasons of German legal judgements

https://scikit-learn.org

https://scikit-learn.org/stable/modules/generated/

sklearn.feature extraction.text.TﬁdfVectorizer.html

https://scikit-learn.org/stable/modules/generated/

sklearn.linear model.LogisticRegression.html

https://scikit-learn.org/stable/modules/generated/

sklearn.svm.LinearSVC.html

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

518

are written in Urteilsstil. Law students are ﬁrst ex-

posed to this writing style after ﬁnishing their ﬁrst

state exam.

In the judgement corpus the components

conclusion, definition and subsumption of the

Urteilsstil are annotated. All other sentences are

labelled as other. The corpus is created as part of

the master thesis of Stefanie Urchs (Urchs, 2020).

The goal of the thesis is to train a model on Urteilsstil

and use it on Gutachtenstil. As outlined in 2.2 the

Urteilsstil and Gutachtenstil share the components

definition and subsumption. Therefore, the

model trained on judgements in Urteilsstil should be

able to withstand a domain transfer to Gutachtenstil

and detect definitions and subsumptions in legal

exercise cases.

This chapter explores the 200 annotated legal

judgements and presents an experiment performed on

the corpus.

4.1 Corpus Design

Overall 25,075 sentences are annotated. 5% (1,202)

of these sentences are marked as conclusion, 21%

(5,328) as definition, 53% (13,322) are marked

as subsumption and the remaining 21% (6,481) as

other. The length of judgements in sentences ranges

from 38 to 862 sentences. The median of judgements

have 97 sentences, the length of most judgements is

on the shorter side.

Figure 2: Number of sentence per label on a judgement ba-

sis. The last part of the ﬁgure depicts the number of sen-

tences per judgement.

Figure 2 shows that the amount of marked

conclusions and definitions is low in all docu-

ments. Furthermore, the length of a judgement only

slightly increases the number of sentences marked as

definition in a judgement. Subsumptions con-

trast this behaviour. The number of sentences la-

belled with subsumption correlated strongly with

the number of sentences in a judgement. The me-

dian of sentences labelled with conclusion is 5 per

judgement, definition is 21 per judgement and for

subsumption 50 per judgement. This shows that in-

dependent of the length of a judgement the conclu-

sion of the overall result and the description of the

abstract law is short. Facts are stated without much

detour. However, on average law professionals use

more than twice as many sentences to subsume the

case. When looking back to chapter 2.2 this seems

reasonable. The subsumption is the place to argue and

combine the pure listing of the law from the deﬁnition

with the real world case. Law professionals can show

their skills in this component. The remaining sen-

tences are labelled as other. These sentences include

text that does not belong to one of the writing style

components. Additionally, the annotator is instructed

to only label sentences as conclusion, definition

or subsumption. If one sentence contains both it is

labelled as other.

As a guiding example the annotation guide for the

annotator includes the example stated in table 2.

Judgements from 22 of the 131 courts are selected

for the corpus. Most judgements originate from the

VG Augsburg (59 / 30%) followed by the VG Ansbach

(39 / 20%) and LSG Munich (33 / 17%). For most

courts one to four judgements are selected. However,

the analysis of the base data already revealed that Ver-

waltungsgerichte (VG) contribute more to the corpus

than other courts. The LSG Munich is in place four for

contributing judgements to the base corpus. There-

fore, the random selection of judgements seems to re-

ﬂect the underlying distribution of decision by courts

well.

29% (58) of all selected judgements are issued in

the year 2016, followed by 22% (44) from the year

2017 and 21% (41) issued in the year 2015. This

selection deviates slightly from the base data where

24.5% (7,663) decisions are issued in 2015, 23.9%

(7,487) in 2016 and 21.7% of the decisions are is-

sued in 2017. The percentages of selected judgements

and decisions issued in 2018 and 2019 are roughly the

same. No judgements from 2020 are selected. How-

ever, decisions of 2020 form only 0.4% of the base

data, which makes not selecting them a valid repre-

sentation of the base data.

The JSON format from 3.1 is extended in the fol-

lowing way:

Design and Implementation of German Legal Decision Corpora

519

1 {

2 " meta ": {

3 " m eta _ti tle " : " " ,

4 " c ou rt " : " " ,

5 " d e ci s io n _s t yle " : " " ,

6 " date ": " " ,

7 " f ile _nu mbe r " : "",

8 " t it le " : " " ,

9 " n orm _ch ain s " : ["" , " "],

10 " d e ci s ion _ gu i del ine s " : [ "", ""] ,

11 " k eyw or ds " : " " ,

12 " l owe r_c our t " : ["" , " "],

13 " a d di t ion a l_i nfo r mat i on " : " " ,

14 " d e ci s ion _re f ere nce " : " "

15 } ,

16 " d eci s io n _t e xt " :{

17 " t en or " : [ " " ," " ],

18 " l ega l_f act s " : ["" , " "] ,

19 " d e ci s ion _re aso n s " : [

20 [[ tex t , l ab el ] ,[ text , l ab el ]]

21 ]

22 }

23 }

The decision reasons are segmented into sentences

using the SoMaJo

tool. SoMaJo is a state of the art

tokenisation and sentence segmentation tool for Ger-

man and English. SoMaJo was chosen after testing

several common segmentation tools, as it is the only

one that performs acceptably on German legal texts.

Each sentence is paired with a label. To preserve

the paragraph structure given by the website, the sen-

tences are saved in a list for every paragraph, resulting

in a list of lists format.

The corpus is published on the open research plat-

form zenodo

4.2 Experiments

To automatically detect conclusion, definition

and subsumption in legal text a logistic regression

and a SVC are trained on the judgement corpus. The

implementation of the logistic regression and the SVC

in scikit learn offers a multi class classiﬁcation op-

tion. The algorithm decomposes a multi-class prob-

lem into a binary problem. The correct prediction is

determined in a one vs. the rest approach.

Two simple features are used in this experiment

all unigrams

of the training data and a tf-idf

for

the vocabulary of the training data.

https://github.com/tsproisl/SoMaJo

https://zenodo.org/record/3936490#.X1ed7ovgomK

https://scikit-learn.org/stable/modules/generated/

sklearn.feature extraction.text.CountVectorizer.html

https://scikit-learn.org/stable/modules/generated/

sklearn.feature extraction.text.TﬁdfVectorizer.html

Table 4 shows the results of a ﬁve fold cross-

validation over logistic regressions and SVM trained

with either unigrams or tf-idf. The baseline is formed

by a decision stump that always predicts the major-

ity class. The baseline is always outperformed, there-

fore, meaningful features are chosen. Interestingly,

LR with unigram features and SVC with tf-idf fea-

tures perform the same. However both are outper-

formed by a LR with tf-idf features in terms of pre-

cision. All three share the same accuracy of 0.77.

Table 4: Results of training a Logistic Regression and a

Support Vector Classiﬁcation on the judgement corpus. The

tf-idf outperforms the Unigram feature independent of the

classiﬁcation method.

Feature/

Classiﬁer

Preci-

sion

Recall F1-

Measure

Accur-

acy

Unigrams/

decision

stump

0.13 0.25 0.17 0.53

Unigrams/

0.74 0.67 0.70 0.77

Unigrams/

SVC

0.67 0.66 0.66 0.74

tf-idf/

decision

stump

0.13 0.25 0.17 0.53

tf-idf/ LR 0.79 0.63 0.68 0.77

tf-idf/

SVC

0.74 0.67 0.70 0.77

5 DISCUSSION

The presented corpora open many different oppor-

tunities to explore German law. Some of the use-

cases are the presented experiments, other possible

use-cases for the decision corpus are the analysis of

topics in courts over time and across courts. Are there

seasonal topics in the courts, that reoccur every year?

Do topics change over the years? When are more

or less decisions issued and does that depend on the

court?

Besides the pure analysis of the meta-data an anal-

ysis of the textual parts might provide other fasci-

nating insights. It might be interesting to explore

whether the authors of decisions can be identiﬁed ac-

cording to their writing style. A sentiment analysis on

the textual part can lead to a deeper insight into what

the author really thought about a case.

Furthermore, the German legal terminology can

be extracted from the corpus and modelled into an up-

per level ontology like SUMO (Mitrovi

c et al., 2019).

Such an ontology is helpful to automatically explore

ICAART 2021 - 13th International Conference on Agents and Artiﬁcial Intelligence

520

unseen German legal documents and analyse their

contents.

The judgement corpus can be used for writ-

ing style detection in free text. Furthermore, it

is possible to look into the inner workings of the

judgement. Which conclusion is connected to

which definition and which definition belongs

to which subsumption. Additionally, judgements

from different courts could be compared in order to

answer whether there are big differences between the

writing styles of the courts.

Based on the work presented in Mitrovi

c et

al. (Mitrovi

c et al., 2017) writing style components

can be represented ontologically, and their persua-

siveness assessed based on the rhetorical elements

contained therein.

6 CONCLUSION

This paper presents two novel German legal corpora

based on Bavarian Court decisions between 2015 and

2020.

The ﬁrst one contains 32,748 decisions from 131

German courts. A model that predicts the type of de-

cision was trained on this corpus. Resulting in a pre-

cision of 0.97.

The second corpus is a subset of the ﬁrst one.

200 judgements were randomly chosen and annotated

with conclusion, definition, subsumption and

other, components of the Urteilsstil. On this cor-

pus several models were trained to predict to which

component a sentence of a judgement belongs. The

baseline is always outperformed, however no clear

best approach could be determined. LR performs well

with unigrams and SCV performs the same with tf-

idf.

Both corpora are published on the open science

platform zenodo.

In future work different legal experts will inspect

the existing labels to ensure the label quality.

ACKNOWLEDGEMENTS

The annotation of the judgement corpus was only pos-

sible due to the help and funding from the research

centre FREDI

The project on which this report is based was

funded by the German Federal Ministry of Educa-

tion and Research (BMBF) under the funding code

01—S20049. The author is responsible for the con-

tent of this publication.

REFERENCES

Alexy, R. (1983). Theorie der juristischen Argumentation.

Suhrkamp Verlag AG.

Danger, D. J. (2005). Urteil und urteilsstil in der zivil-

rechtlichen assessorklausur: eine praktische hilfestel-

lung. Juristische Arbeitsbl

atter, (07).

Gamper, J. (2000). A parallel corpus of italian/german legal

texts. In LREC.

Mitrovi

c, J., O’Reilly, C., Mladenovi

c, M., and Handschuh,

S. (2017). Ontological representations of rhetorical

ﬁgures for argument mining. Argument & Computa-

tion, 8(3):267–287.

Mitrovi

c, J., Pease, A., and Granitzer, M. (2019). Modeling

legal terminology in sumo.

Mochales, R. and Moens, M.-F. (2011). Argumentation

mining. Artiﬁcial Intelligence and Law, 19(1):1–22.

Palau, R. M. and Moens, M.-F. (2009). Argumentation

mining: the detection, classiﬁcation and structure of

arguments in text. In Proceedings of the 12th inter-

national conference on artiﬁcial intelligence and law,

pages 98–107. ACM.

Reed, C. (2006). Preliminary results from an argument cor-

pus. Linguistics in the twenty-ﬁrst century, pages 185–

196.

Rizzo, C. R. and P

erez, M. J. M. (2012). Structure and

design of the british law report corpus (blrc): a legal

corpus of judicial decisions from the uk. Journal of

English Studies, 10:131–145.

Rossini Favretti, R., Tamburini, F., and Martelli, E. (2001).

Words from bononia legal corpus. International jour-

nal of corpus linguistics, 6(1-12):13–34.

Sugathadasa, K., Ayesha, B., de Silva, N., Perera, A. S.,

Jayawardana, V., Lakmal, D., and Perera, M. (2017).

Synergistic union of word2vec and lexicon for do-

main speciﬁc semantic similarity. IEEE International

Conference on Industrial and Information Systems

(ICIIS).

Urchs, S. (2020). Extracting deﬁnition and subsumption

from german law. Master’s thesis.

Weber, D. (2018). Methodik der Fallbearbeitung. Nomos

Verlagsges.MBH + Co.

Wissler, L., Almashraee, M., D

ıaz, D. M., and Paschke, A.

(2014). The gold standard in corpus annotation. In

IEEE GSC.

https://www.jura.uni-passau.de/fakultaet/

forschungseinrichtungen/forschungsstelle-fredi/

Design and Implementation of German Legal Decision Corpora

521