ARGUMENTUM

The term argumentation mining describes technical approaches for the identification and analysis of argumentation structures in electronic texts. This topic has gained importance in recent years and has mostly been investigated considering the field of jurisprudence [4, 5].

In this context, interesting and well-performing text mining approachs for the identification and analysis of argumentation structures based machine learning approaches like Support Vector Machines (SVM) have been proposed [2, 6] and applied using corpora of international jurisprudence [7]. Furthermore, a markup language for argumenation structures – the Argument Markup Language (AML), and software prototypes using AML for the representation of argumentation – the so-called Araucaria system [8] – based on well-known argumentation patterns like the Toulmin Scheme [9] or the schemata by Walton [10] have been proposed as a fundament for argumentation mining.

However, approaches in literature so far typically focus on argumentation in English language and on the so-called Common Law or Case Law jurisdiction [11] as the typical legal system in the Anglosphere. In contrast to that, the ARGUMENTUM concept introduced in the following is supposed to support argumentation mining in the BVerfG decision corpus (German language) and to consider the characteristics of a different legal system – the so-called Civil Law or Codified Law jurisdiction which coins the German legal system. In the following section, the structure and important characteristics of the BVerfG decision corpus are introduced and discussed.

[2] R. Mochales and M.-F. Moens, "Argumentation mining," Artificial Intelligence and Law, vol. 19, pp. 1-22, 2011.

[4] M.-F. Moens, E. Boiy, R. Mochales-Palau, and C. Reed, "Automatic Detection of Arguments in Legal Texts," in Proceedings of the 11th international conference on Artificial Intelligence and Law (ICAIL '07), Stanford, CA, 2007.

[5] A. Wyner, R. Mochales-Palau, M.-F. Moens, and D. Milward, "Approaches to Text Mining Arguments from Legal Cases," in Semantic Processing of Legal Texts, LNCS 6036, E. Frances-coni et al., Eds., Berlin: Springer, 2010, pp. 60-79.

[6] R. Mochales and M.-F. Moens, "Argumentation Mining: The Detection, Classification and Structure of Arguments in Text," in Proceedings of the 12th International Conference on Artificial Intelligence and Law (ICAIL '09), Barcelona 2009.

[7] R. Mochales and A. Ieven, "Creating an argumentation corpus: do theories apply to real arguments?," in Proceedings of the 12th International Conference on Artificial Intelligence and Law (ICAIL '09), Barcelona, 2009.

[8] C. Reed and G. Rowe, "Araucaria: Software for argument ana-lysis, diagramming and representation," International Journal on Artificial Intelligence Tools, vol. 13, pp. 961-979, 2004.

[9] S. E. Toulmin, The Uses of Argument. Cambridge: Cambridge University Press, 1958.

[10] D. N. Walton, Argumentation Schemes for Presumptive Reasoning. Mahwah, NJ: Lawrence Erlbaum Associates, 1996.

[11] R. Mochales and M.-F. Moens, "Study on the Structure of Argumentation in Case Law," in Proceedings of the 2008 conference on Legal Knowledge and Information System (JURIX '08), Amsterdam, Netherlands, 2008, pp. 11-20.

• Literature Review
• Text Mining
• Machine Learning
• Support Vector Machines
• Argumentation Mining
• Common Law
• Case Law
• Argument Markup Language
• AML
• Araucaria
• Walton Schema

Regarding the general structure of BVerfG decisions, it can be concluded that argumentative statements can only appear in the following three places: Guiding principles, Statement of facts, and Reasons for the decision. Since the Guiding principles only present a summarization of the final holding without a justification and since the Statement of facts contains subjective argumentation and the claims of participants of the lawsuit, it is sufficient to focus on the Reasons for the decision (“reasoning”) in order to identify the court’s argumentation. Against this background, the entire decision can be divided in three parts: first the beginning of the document (including Guiding principles, Title of judgment, and Tenor) until the beginning of the reasoning, the reasoning itself, and eventually the end of the document, i. e. the entire text that follows the reasons until the end of the decision. To make use of this three-part separation, the first task entails the automated identification of the reasons. It is possible to simply perform a string match for the heading “Gründe:” (including the colon, German for reasons) which appears uniquely once in each decision and introduces the section containing the reasoning. For the purpose of identifying the end of the decision’s reasoning, we investigated the last sentence of every document in our decision sample. The most expressions allow for a string-based matching.

„This decision is irrefutable.“ („Diese Entscheidung ist unanfechtbar.“)

„[...] reimbursement of expenses [...]“ ( „[…] Auslagenerstattung […]“)

„[...] reimbursement of (necessary) expenses [...]“ ( „[…] Erstattung der (notwendigen) Auslagen […])“

„The decision was issued unanimously.“ ( „Die Entscheidung ist einstimmig ergangen.“)

„[...] necessary expenses were reimbursed [...]“ ( „[…] notwendigen Auslagen erstattet […]“)

In every analyzed decision in the sample, we found that the reasoning is always immediately followed by the names of the judges in charge and, after this listing, potentially some more text. So to circumvent the problem of ambiguous wordings, it seems feasible to utilize a word list that contains the names of all judges at the BVerfG from 1998 until today which can be easily build up by extracting all the names from all decision documents. The occurrence of one of these names followed by another name from this list provides an adequate indicator for the end of reasoning. Assuming that statements contained in the Reasons section are presented in linear order and that they are not scattered across the document (which was the case in every document we analyzed), it is sufficient for argumentation structure analysis to extract the entire text between the identified beginning and the identified end to separate the decision’s argumentation from the text which is not important for argumentation analysis.

• Text Mining
• Data Preparation
• Text Extraction

It is necessary to separate the whole argumentation part (as extracted in phase 1) into argumentative units by considering the inherent paragraph structure of a decision. The paragraph structure in the BVerfG decisions strongly correlates with the structure of argumentative units as regards content – or more precisely – an argumentative unit concerning one topic commonly represents one paragraph. However, in many complex argumentations a hierarchical structure of an argument can be identified. An argumentative unit concerning one topic can be separated into smaller argumentative statements as its constitutional parts. In the BVerfG decisions corpus such relations are often signalized by the usage of different mark-ups and numberings of paragraphs. However, the mark-up of the paragraph structure also depends on the type of decision documents. Important types are judgement ("Urteil"), chamber decision ("Kammerbeschluss") and senate decision ("Senatsbeschluss"). For the initial concept development, we consider the typical structure of a judgement for the separation of argumentative units as this decision type possesses the most comprehensive paragraph structure. The other decision types often contain subsets of this structure.

Table II presents the markups of the hierarchical stucture of judgements.
Level Reasons:
1st A. first level: Latin capital letters + “.”
2nd I. second level: upper Roman numerals + “.”
3rd 1. third level: Arabic numerals + “.”
4th a) fourth level: small Latin letters + “)”
5th aa) fifth level, double small letters + “)”
6th (1) sixth level: Arabic numerals in brackets
7th i) seventh level: small Roman Numerals + “)”
8th α) eighth level: Greek letters + “)”

As the structuring is not totally consistent throughout the whole corpus – especially in the deeper regions of the hierarchy beginning with the 5th level – while the first four levels are mostly consistent, we consider the first four hierarchy levels for our concept. Thus, we separate arguments based on the given numbering concerning the first four levels for our first prototypical implementation. The usefulness and feasibility of this approach as well as the quality of gained results based on this approach have, however, to be investigated later on by means of application and empirical evaluation of the future prototype.

• Argumentative Units
• Data Preparation
• Text Mining

One of the major objectives of ARGUMENTUM is to enable a user to perform a content-based retrieval when looking for arguments, which are “similar” to a given one. With respect to argumentation, the term “similarity” normally involves a deep textual understanding of argumentative relations. However, it can be very useful to narrow down the set of candidate paragraphs by means of their content. For this purpose, we aim at representing each argumentative unit as a vector of relevant index terms in a Vector Space Model [13]. Regarding the specific needs in ARGUMENTUM, we use the following sequence of analysis steps which consider peculiarities of the BVerfG decision corpus and the German language.

1. Assuming that textual keywords which transport im-portant content of a paragraph are nouns or noun phrases, we focus on the detection of these parts of speech. Therefore, PoS tagging is used to identify nouns, followed by Named Entity Recognition (NER) to determine proper nouns [14]. Regarding characteristics of the German language, all words that start with a capital letter and have not been tagged by PoS or NER – except those at the beginning of a sentence – are considered as nouns as well.

2. Next, we determine groups of words (so-called n-grams), i. e. noun phrases in addition to the nouns identified in step 1. Considering n-grams usually yields more significant index terms as the following example shows: the term “act” on its own does not provide a good index term as it is very frequent in common usage. However, the 2-gram “negligent act” pro-vides a much more specific index term. In order to detect n-grams, we start from the nouns that we already identified and consider a maximum of two preceeding words. Groups of words that comprise more than three words are quite unusual in German and often express a very specific (unique) circumstance. Using information about these words’ part of speech, we can also stop an analysis if we encounter words of the classes “verb” or “article” for instance, as a word sequence containing such words can not form a noun phrase.

3. To exclude specific words from being considered as in-dex terms, we use a stop word list. This is useful as the corpus contains several words which occur frequently but which are not very meaningful, e. g. the general term “Gericht” (English court) is often used in references to previous decisions but does not seem to be very important in the context of summarizing a paragraph’s content.

4. The final step applies a stemming algorithm to the remaining words respectively groups of words to reduce them to their basic forms. This causes e. g. plural and singular forms of the same word to be mapped on a single index term. In addition to that, we also consider performing a decomposition of compound nouns, e. g. the German term “Gesellschafterhaftung” (English accessory liability) is decomposed into “Gesellschafter” and “Haftung” thus extending the set of index terms. Finally the TF*IDF method is used to compute a statistical weight for every term [15]. Those weights are used as dimensional values in the vectorial representation of a paragraph.
Executing this four-step sequence results in an individual vector per argumentative unit. The dimensional values of each vector equal the TF*IDF-weight set of index terms of the entire corpus. Determination of argumentative units that resemble other given units in terms of content can be achieved by calculating a similarity measure, e. g. Cosine Similarity or the Euclidean distance for the vectors.

[13] G. Salton, A. Wong, and C. S. Yang, "A Vector Space Model for Automatic Indexing," Communications of the ACM, vol. 18, pp. 613–620, 1975.

[14] D. Jurafsky and J. H. Martin, Speech and Language Processing: an introduction to natural language processing, computational linguistics, and speech recognition, 2 ed., Upper Saddle River, NJ: Prentice Hall, Pearson, 2009.

[15] K. S. Jones, "A statistical interpretation of term specificity and its application in retrieval," Journal of Documentation vol. 28, pp. 11-21, 1972.

Sentiment analysis is an important step towards understanding whether a court’s holding turns out to be pro or contra with respect to the overall decision under discussion. In Phase 4 this analysis is performed for the entire reasoning of a decision (we call this the “absolute sentiment”). Therefore we aim at automatically identifying whether an appeal is admissible or not, and in case it is admissible we look at whether it is also founded or not. Since the German Federal Constitutional Court only uses a few different expressions for classifying a decision, automation can be achieved by conducting a string-based matching for the following expressions:

Inadmissible
„[...] constitutional complaint [...] not accepted for decision [...]“ („[...] Verfassungsbeschwerde […] nicht zur Entscheidung angenommen [...]“)

„The request for remission of provisional order will be discarded.“ („Der Antrag auf Erlass einer einstweiligen Anordnung wird abgelehnt.“)

„The request will be discarded.“ („Die Anträge werden verworfen.“)

Admissible
Founded
„[...] violate the appellant in his right formulated in article [...]“ („[...] verletzen den Beschwerdeführer in seinem Grundrecht aus Artikel [...] des Grundgesetzes [...]")

„[...] is incompatible with article [...] of the constitution.“ („[...] mit Artikel […] des Grundgesetzes unvereinbar […]")

Unfounded
„The constitutional complaint will be discarded.“ („Die Verfassungsbeschwerde wird zurückgewiesen.“)

„[...] is compatible with article [...] of the constitution.“ („[...] mit Artikel […] des Grundgesetzes vereinbar […]")

Aside from the court’s holding, the actual issue under discussion can be extracted at this point as well and can be saved in addition to the results of the sentiment analysis. The same wording in the section Title of judgement always states the issue which can also be automatically extracted: “In the lawsuit concerning the constitutional complaint [... representative/s ...] against [... issue ...] the (x). chamber [...] has decided [...]“. It should be noted that at this point in the analysis it is not possible to answer the question which arguments within the court’s holding argue in favour of or against the issue under discussion. This is what we call “relative sentiment” when analyzing sentiment per argumentative unit in Phase 5.

As a first step towards an automated fine-grained analysis, we aim at classifying each sentence within a paragraph as either a proposition or an explanation. Proposition in that case means a newly declared statement that expresses an opinion held by the court. Explanation denotes the support that is provided to back this proposition. From the manual analysis of the sample’s paragraphs, we draw the following conclusions:

1. Propositions in this context typically comprise only one sentence and only in rare cases more than three consecutive sentences. In contrast to that, explanations are more extensive and typically consist of more than one sentence.

2. Propositions almost always form the beginning of an ar-gumentative unit within a decision, i. e. they formulate its first sentence. Explanations typically follow immediately on propositions but can span several paragraphs and normally end once the next proposition is stated.

3. In several cases, we observed that a proposition and its related explanation are contained in only one paragraph. However, those occurrences seem to be independent of any (sub)-numberings and furthermore, no decisive structure could be identified for when this case appears.

From these conclusions follows that the mere structure of a decision is not sufficient for the identification of propositions and explanations. Therefore, we need to apply methods that work on linguistic features of a decision’s text. Mochales and Moens [6] apply a Support Vector Machine (SVM) along with sophisticated textual features to a similar problem when classifying premises and conclusions according to the Toulmin argumentation scheme [9]. Starting from this approach, we aim at extending their method by the use of linguistic patterns comprising expressions which are common in German legal language. Such patterns can be expected to form a precise means for the identification of specific linguistic characteristics and constitute a powerful SVM feature complementing the ones proposed in [6].
In addition to using a SVM along with linguistic features extracted from the decision text, we will examine the possible implementation of a Context-Free Grammar (CFG) in order to automatically derive the underlying argumentative structure of a decision (argumentative parsing). According to [6], the effectiveness of using a CFG strongly depends on the corpus to be processed: only if it is sufficiently homogenous, production rules that cover the structural text characteristics at hand can be formulated as part of a CFG. In order to determine whether the BVerfG corpus exhibits such homogeneity, further investigation needs to be carried.
As mentioned before in Phase 4: Sentiment analysis, we need to perform a more detailed analysis in addition to the previously determined “absolute sentiment” in order to obtain a “relative sentiment” per paragraph. This sentiment is “relative” as it supports the classification of every proposition as pro or contra regarding the overall issue under discussion and its absolute sentiment. For example, if we determine a specific proposition as pro and the superior discussion is classified as contra, then the pro-proposition supports declining the issue. Basic methods for determining sentiment on a fine-grained level have been proposed in literature, e. g. in [16], and can serve as a starting point for our implementation.

[6] R. Mochales and M.-F. Moens, "Argumentation Mining: The Detection, Classification and Structure of Arguments in Text," in Proceedings of the 12th International Conference on Artificial Intelligence and Law (ICAIL '09), Barcelona 2009.

[9] S. E. Toulmin, The Uses of Argument. Cambridge: Cambridge University Press, 1958.

[16] A. Hogenboom, F. Hogenboom, U. Kaymak, P. Wouters, and F. de Jong, "Mining Economic Sentiment Using Argumen-tation Structures," in Advances in Conceptual Modeling – Applications and Challenges, LNCS 6413, J. Trujillo et al., Eds., Berlin, Heidelberg: Springer 2010, pp. 200-209.

• Argumentation Analysis
• Walton Schema

In the last phase, all developed (meta-)information which have been gathered need to be stored in a structured way. The content-related meta-information is contained in the vectors representing each argumentative unit paragraph and which are based on the extracted index terms. In order to store individual vectors in such a way that they can easily be compared to every other vector in the collection, we plan to use a term-document matrix. This matrix comprises m columns and n rows, where m denotes the number of terms in the whole term collection and n denotes the quantity of documents that constitute the collection. A matrix entry wij represents the TF*IDF weight of term i in document j if it exists, and 0 otherwise. Thus, every row forms a document vector dij in an m-dimensional vector space. Based on this representation, it is possible to directly identify the terms which are shared by two vectors dij and dkl by using adequate underlying similarity measures which considers TF*IDF values. Thus, the content of all argumentative units can be compared and similar arguments as regards content can be identified. Besides the information which are necessary for a content-based retrieval, much more data is generated during the individual phases of the analysis, e. g. during the classification of part of speeches (PoS tagging in phase 3). This information can be saved using a markup language like XML. Markups in XML can be applied in every analysis step in order to save the information produced and it can, furthermore, be used to define and tag hierarchical structures in decisions. For example, paragraphs or sentences could be marked by tags that express their sentiment or classify them as “premise”, “conclusion” etc.

In this step, we have developed the software architecture based on the introduced phase concept for data preparation and text analysis. Each steps of data preparation and analysis shall be performed via one software module.

• software architecture
• modules

In this step, we chose an adequate development framework according to our requirements and fitting the developed technical concept. Fruthermore, existing modules implemeting required functionalities were picked.

A group of two developers implemented the ARGUMENTUM prototype in JAVA. Frutehrmore, a larger group of researchers from DFKI tested the prototype and helped to identify bugs and resolve frutehr existing problems.

• Prototype

The whole project consortium was provided with the implemented prototype and started testing its functionality.

We reviewed new literature sources in the context of argumentation mining and legal information systems research.

• Literature Review

In this step, we integrated the new knowledge into our knowledge base according to the documentation structure from the first iteration.

We identified new requirements, e.g. we chose to implement the second prototype as a web-based information system in order to provide better access. Nevertheless, a similar search speed should be provided to the users of the platform.

We added a new structure of data preparation and improved the data structure for an acceptable performance of the vector space search in the context of a web-based information system. Furthermore, we introduced the idea of a rule-based retrieval of argumentation according to the Walton Schema.

• Walton Schema
• Vector Space Model
• Argumentation Detection

We defined XML-based rule schema in order to provide a rule-based detection of argumentation structures. The rule definition was developed in cooperation with lawyers and argumentation scientists (theoretical philosophers). We provide an example of one argumentation type as an XML file below.

In order to provide the retrieval of argumentation structures independently of the content of an argument, a second vector space for argument types is needed. This second vector space is implemented in the second iteration to enlarge the search functionality.

We implemented the new prototype with a team of three researchers as a web-based information system (search engine).

Functional tests and validity tests were performed by lawyers in order to prove the usefulness of the prototype.

• Prototype
• Argumentation Mining