The significance of interpretation principles seems to be inexorably tied to the moments in time when they are recurrently referenced by scholars or judges. Moreover, these sequences of references may undergo substantial shifts in meaning or contextual usage over time. Based on this phenomenon, our work proposes a network analysis approach to identify and locate such pivotal points. Specifically, we begin by extracting and mapping citations in judicial rulings, focusing on the specific context of fiscal state aids in the case-law of the Court of Justice of the European Union. We then demonstrate how applying network analysis to these citations can serve as a valuable tool for enriching the legal study of CJEU case-law. In detail, we focused on the network of precedents as cited by the Court to verify how the case-law develops new interpretative principles and contributes to the creation of a legal framework for European discipline of fiscal State aids. To retrieve the necessary information on precedent references within a judgment, we utilized the XML representation accessible on the EUR-Lex platform. We then employed regular expressions to parse the text and guarantee the precise and complete extraction of citations. Our research highlights how automated analysis of citation networks can ofer valuable resources to supplement conventional legal methodologies.
for European law more in general, the case-law of the Court of Justice of the European Union (CJEU) plays a key role. To some extent, case-law contributes to defining and realising the single market, which is the ultimate objective of the Treaties. However, from the point of view of the system of sources, they remain a unicum, an undefined legal object which distinguishes the role of EU jurisprudence from the traditions of the member states, both in common law and civil law.
The significance of interpretive principles has become increasingly apparent over time, as the Court of Justice frequently employs them as pivotal points in its exegesis of European law. While literature has emphasized the importance of these principles, it has yet to explore how they achieved such prominence. To investigate the development of citations and ensure the verbatim nature of references to precedents, we opted to integrate quantitative techniques into our analysis1.
In a new case, the Court frequently cites precedents by quoting a specific paragraph (or a few) that encapsulate a significant concept or principle. This implies that citations are typically not intended to refer to the entire judgment, but rather to the specific paragraph in question. For this reason it is common to find citations to case
The structure of the available versions of judgements, and other documents of the CJEU adheres to a set of openly available rules, adopted by the Court of Justice of the European Union.
The method of citing the case-law in particular combines the ECLI with the usual name of the decision and the case number in the register. It has gradually been brought into use by each EU Court/Tribunal since the ifrst half of 2014, and was harmonised as between the Courts of the European Union in 2016. The reference comprises several elements, including the type of decision (i.e., judgment or order), the complete date of the decision, the customary name of the case, the case num- objective is to examine the connections that judges aim to ber in the register (e.g., for the CJEU C-nn/aaaa), the establish when constructing their argumentation frameECLI, and the cited paragraph. work. It is worth noting that, particularly for widely This fixed citation style has many advantages, such as: recognized principles, a judgment citation involves a deliberate selection from numerous precedents, oftentimes • it improves the accessibility of judicial decisions by en- identifying both the most and least recent. This potensuring that references to case-law consistently include tial limit has been overcome with the linking of direct all necessary information to unambiguously identify and indirect citations (see subsection 4.3). We assume, the decision in question. Specifically, each reference and partially demonstrate, that the judgements that the includes all constituent elements, which are reiterated Court decided to refer to are generally the oldest ones, every time the reference is made; until the reaching of a "canonization" of the interpreta• it provides greater linguistic neutrality since the format tive principle, that from then on is the cited case (see of the citation is largely identical in all languages and section 4.3.2). thus contains fewer elements to be translated; We chose to focus on the field of fiscal State aid, since this topic is representative of the creative role played by • it facilitates the automatic insertion of hyperlinks on the CJEU (particularly in the field of fiscal State aids), the ECLI of the cited decision and to the relevant para- and because it is a small enough field that the connecgraph. tions seemed more verifiable in an initial assessment.
By concentrating on a particular field, we are aforded the opportunity to assess the efects of citation network analysis utilizing a methodology that diverges from previous investigations. In particular, we tried to merge the methodological and meta-argumentative paths, connected with the use of precedents by European Judges, with an analysis of the actual impact on specific legal issues.
It is important to underline that the case number in the register identifies all the documents referable to a specific procedure, which means judgement of any grade, opinion of the Advocate General, order, request for a preliminary ruling. Vice versa, the ECLI identifier refers to a single specific document, and allows for the identification of the cited judgement. That means that diferent ECLI may refer to diferent documents related to a single case.
We opted to adopt the Court’s own method of citing
precedents, focusing on the cited paragraphs rather than 3. Related work
the judgment as a whole in our network analysis. This
choice was justified also by the fact that, as mentioned, We are not the first to examine the citation of precedents
the cited paragraphs may pertain to a judgment on a (or other legal authorities) as a means of understanding
diferent topic from the one of the citing document. One the importance of courts, opinions, or judges (see, e.g.,
such example can be found in judgement Case C-322/09 P [
Such connection may appear perplexing if we consider developed with regards to common law systems, which it in the network of judgements on state aid, but is instead rely on the concept of binding precedent rather than of perfectly reasonable if we look at the concepts expressed positive rule. According to this, the importance of precein the two paragraphs. In particular both concern the dents is significant in such systems and many studies issue of the arguments that an appellant is allowed to have focused on it since a long time. put forward, a fact which is equally relevant in the two The approach is slightly diferent when it comes to the judgements. CJEU which is not part of a common law system 2.
Furthermore, the citation of paragraphs, regardless For this reason, there is only a reduced number of studof the topic of the judgements, concurs in giving them ies of citation networks specifically dedicated to CJEU an autonomous value as interpretative principle. As in- case-law, and all of them underline dificulties in workterpretative principles, they may be used in the CJEU argumentation as part of the European interpretative framework, almost comparable to a legal rule.
At this stage, we choose to work only on explicit citations, without looking for the implicit ones. Hence, our
is a hybrid of the two. There is an agreement in literature on
the impossibility of finding any principle about binding precedents.
That fact creates an unbeatable obstacle in considering the European
one as a common law system
ing with judgements. In particular we could recall the
works of Derlen and Lindholm, which developed from
[
According to [
Instead of ranking entire judgments, in [
By learning from these previous works, the dataset we will describe and use is based on single paragraphs as the main nodes, with additional information on the entire judgments.
The categorization of inconsistencies of CJEU citation
policy is analyzed in-depth by [
From this perspective, references to settled cases result in complex changes to the law, which are multidirectional and lack a clear progression. We refers to the concepts of substitution and un-anchoring, trying to deal with them in the analysis of centrality and semantic similarity.
There have been, and are, multiple attempts at the
automated identification and extraction of legal citations,
both from case-law and legislation, with diferent
technological means. In [
In [
Going one step further from the citation extraction
1–10
we have diferent attempts at building graph networks,
mainly focusing on the US legal system, as in [
In [19] the similarity and relevance of legal citations is
then applied to historical cases from the Court of
Friesland, and used to assess the importance of case-law
citations from a historical view. The same idea is applied in
[
used to build the database of case-law citations in the previously defined context.
Often, studies such as the one proposed in this paper sufer from the limited availability of structured data, requiring a long manual and preliminary work. By leveraging the cases made publicly available in declarative markup formats by the ECJ on the EurLex platform, we developed an process that allowed to directly focus on the data analysis of the research.
The objective of this work was to determine if and how an automated network analysis phase could complement the human analysis in navigating and extracting useful insights from complex citation networks [19].
out are all available on the EurLex platform, and accessible through Cellar, the common repository of metadata and content for EurLex 3.
The first step was to verify the availability of cases in the chosen legal domain. This task was carried out by searching the EurLex and Curia databases with the following filters: "judgement" as a type of document; "CJEU" as Court (avoiding judgement of the General Court); "appeal" as procedure; "State Aid" as subject-matter; plus we added the word "tax" in the free text, in order to find cases of fiscal State aid. We then identified what information would need to be extracted.
Most of the cases we analyzed are available in an XML format, Formex4, that is used to add structural information and metadata to case law. In particular, for what was necessary to the definition and analysis of citation networks, the cases contain detailed information on legislative and case-law citations, in particular with the XML
4Formex describes the format for the exchange of data between the Publication Ofice and its contractors. In particular, it defines the logical markup for documents which are published in the diferent series of the Oficial Journal of the European Union. tag "REF.DOC.ECR". The availability of this information made extracting references from cases much easier.
When this XML representation was not available the data extraction relied mainly on regular expressions, although this was more of an issue with older cases. Generally however the style of citation was found to be common enough to make the extraction possible without too many issues.
Having structured cases, either in XML or HTML, made it possible to extract also the cited paragraphs. These were then used to further crawl the network, by parsing the cited paragraph for other citations, as well as to compare the semantic similarity of the cited and citing paragraphs. This is a novel development in the legal citation analysis, and its usefulness and purpose will become clear in the following sections.
For the proper extraction of the paragraph number
and content, the structural information contained in the
XML representation is further enhanced with regular
expressions. Between the two modes (XML and regex
parsing) it is generally possible to extract the correct
paragraph. On the one hand, methods based on regular
expressions are often used in systems for text extraction
and analysis [
From the procedural point of view, the developed tool downloads the XML representation of the case document and identifies the citations. Then, for each cited document it repeats the process recursively, building a the database of cases and citations. In the selected domain, we considered the cases from the Court of Justice, excluding those from the General Court, which in judgements of appeal such as those we are interested in, acts as a Court of First Instance.
sentations of the original XML object, as can be seen in Listing 1, with a subset of the metadata and the struc- In Figure 1 it is possible to see the definition schema ture extracted. In particular, necessary information in of the database, with judgments (in pink), that contain this context is the text of the paragraph containing the paragraphs (in gray). The paragraphs refer to other paracitation, as will be highlighted in Section 4.3.2, and the graphs.
URL pointing to the XML version on EURLex, to recursively repeat the search. Other information is available 4.3. Data analysis but not yet used in the network analysis, while still being captured for future developments (see 5). The database generated can be easily exported to JSON
The main advantage on using a json human readable LD, RDF, cypher, and other languages for further semanrepresentation is that it enables the legal experts involved tic analyses. In our case, we used the cypher lanugage, to directly access the information, for an initial qualita- importing the data in a Neo4J database5. With this data tive evaluation of the output, that can be used to alter 5https://neo4j.com the methodology eficiently.
The general idea is to have the complete set of citations for the cases in our domain, and to proceed with a recursive analysis only for the cited paragraphs. This enables us to analyse from a historical point of view the relevant citations, without broadening the number of citations to analyse more than strictly necessary.
From the initial list of 40 cases extracted from EurLex, we extracted a total of 1435 paragraphs, from 493 judgments. In this dataset there are 1392 relationships between paragraphs. imported, it is possible to visualise the citations as a only of direct citations of paragraphs (case A cites case graph, with arrows going from the citing paragraph to B), but also the indirect citations (where paragraph A the cited one (REFERS_TO), and back from the paragraph cites paragraph C, that in turn cites paragraph B). From a to the containing case (BELONGS_TO). It is then possible legal standpoint this allows us to recollect the historical to query the graph as a database and extract information. evolution of a specific interpretative principle, as well as to have an overall view of the legal precedents referred to in a specific judgement.
To address this, it is possible to collect the citations recursively for each starting node, then summing up the distinct paths. We are interested in the distinct paths to avoid duplication, or counting the same relation more than once.
4.3.1. Centrality algorithms A more detailed analysis on the relevance of citations can be done by using centrality-based algorithms to determine the importance of nodes in the network. The metric used for this evaluation is the Degree Centrality[20], which measures the number of in/out relationships for the diferent nodes. In this case, we are interested in the incoming REFERS_TO relations between Paragraph nodes6, in order to see which are cited more frequently through distinct relationships.
The graph can be analyzed from both a vertical and an horizontal point of view, by either searching for the temporal evolution of a citation, or the frequency of citations. The vertical analysis enables the identification not
occupied in each row. ECLI ECLI:EU:C:1991:142 ECLI:EU:C:1986:22 ECLI:EU:C:1986:22 ECLI:EU:C:1992:381 ECLI:EU:C:1978:36 ECLI:EU:C:1978:36 Paragraph NP0018 NP0037 NP0038 NP0016 NP0018 NP0019
such as PageRank [21], although the use of Degree Centrality already demonstrated to return useful results.
On the basis of this representation, the vertical
analysis of a specific interpretative principle may show how
much the CJEU uses direct repetition and formulas (e.g. it
is settled case law [
tween the judgement which are well-known as "leading case" between scholars and nodes that has a higher number of incoming relations, as is shown with respect to 2021:201 in 3, where the citation ranking is built with the Par 25 13 20
Cit 9 9 9 Name Tubemeuse Banco Ext de Espana Spain v Commission
ECLI EU:C:1990:125 EU:C:1994:100
EU:C:1994:325
One of the efects produced by repetition is what Sadl
[
The most significant outcome of this experimentation is to realize how used the un-anchoring mechanism is within the selected field of fiscal State aid. The most directly cited interpretative principle is par. 59 of EU:C:1994:211, Brazzelli Lualdi et al., while if we consider the aggregate (or re-anchored) data, the most cited one is EU:1992:381, Portuguese Republic and Kingdom of Spain v Council of the European Communities. None of them involve substantial issues on State aid.
The fact that, in our dataset, the most cited paragraphs apparently do not belong to judgement in the field of State aid may have a quite simple explanation. We choose to focus on appeal judgements, since they are closer to judgements in plain litigation, since the CJEU does not merely play the role of an interpretative judge. In these cases thus there may be many procedural issues raised by the parts and discussed by the Court, allowing for many citations of a procedural nature. These kinds of issues do not depend on the substantial object of the controversy. Thus, it is reasonable that judges looked at the precedents that may be relevant for solving the procedural questions. This fact is however not explicitly stated in the citing cases, which is exactly one of the efects that un-anchoring is supposed to have. 4.3.2. Semantic similarity
tion is the semantic comparison of paragraphs.
The idea behind this approach is that a vast network
of citations is not easy to handle manually. Instead, an
automated approach based on diferent similarity metrics
can be used to verify the relevance of citations [
We are interested in a semantic-based similarity approach, to take into account not only the sequence of characters, but also the context of the contained words. • ECLI:EU:C:2021:201, par. 37: "As regards the fundamental freedoms of the internal market, the Court of Justice has held that, given the current state of harmonisation of EU tax law, the Member States are free to establish the system of taxation which they deem most appropriate, meaning that the application of progressive taxation falls within the discretion of each Member State.
The same is true in the field of State aid " • ECLI:EU:C:2020:139, par. 49: "However, it must be recalled that the Member States are free, given the current state of harmonisation of EU tax law, to establish the system of taxation that they deem the most appropriate, and consequently the application of progressive taxation falls within the discretion of each Member State" • ECLI:EU:C:1976:95, par. 9: "Although this provision prevents taxes being levied on the products of other Member States which are higher than the taxes applicable to similar domestic products, it does not however restrict the freedom of each Member State to establish the system of taxation which it considers the most suitable in relation to each product"
can calculate the distance between the first two paragraphs (those that are apparently closest), with and without the citations. This index is calculated by dividing the number of common elements in the two samples by the average number of elements.
In the first case, the similarity seems low (67.60%),
compared with the actual semantic value of the text. If
we remove the references to case law, the value is a higher
(82.20%). This discrepancy could be attributed to what is
stated in [
This method of changing the references has a profound relevance when we focus on a specific field of law, such as fiscal State aids, as in our experiment. Another level of substitution may be the contamination with other fields of law, in which the CJEU expressed an interpretative principle for the first time.
A diferent approach that can be used in assessing the semantic similarity is to use pre-trained Transformer models (e.g. BERT-like) that convert input texts into vectors (embeddings) thus capturing semantic informa- Court also un-anchored the references, since in 2021 any tion, and requiring less manual intervention in remov- referral to the 1976 judgement disappears. ing sections of text before hand. In particular, we calcu- Furthermore, it must be underlined that the sentence lated the Cosine similarity on a set of embeddings of the “The same is true in the field of State aid", starts another paragraphs[22]. To extract the embeddings we used the short chain, in which the same principle may be retrieved Sentence Transformers library, using existing pre-trained in two important cases (ANGED and Gibraltar) pertaining models7. to the field of State aid. However in Gibraltar, which is
In Figure 4 we have mapped the cosine similarity be- a leading case, is referred to as an obiter dictum coming tween three related paragraphs of those in Figure 3. from the judgement of the Court of first instance. Such
The comparison between semantic and non-semantic precedent is weaker that the ones indicated in the main similarity score appears useful especially when the re- chain that we analysed. Consequently, it is possible to sults difer, and it can be useful in the legal analysis (e.g., imagine that Judges chose to un-anchor it, in order to in identifying un-anchoring or substitution). Further- reinforce the case-law support to their statement, which more, discrepancies in semantic and non-semantic re- is fundamental for the solution of the case. sults may signal the presence of an anomaly, that could be interesting to in-depth analyse.
Our example regards the principle of national supremacy on tax matters, which is expressed with regards to competition and State aid (EU:C:2021:201) or to direct taxation and economic freedom (EU:C:2020:139 and EU:C:1976:95). Direct taxation is a sector of exclusive competence of Member States, so an intervention of EU institutions is possible only if it is justified by prevalent interest (like protection of one of the fundamental economic freedoms). It is very diferent when it comes to the field of State aids, which are regulated directly in the Treaties, with a fully European competence. In such a scenario, substitution may be used to reinforce the political statement of the Court of Justice, without any kind of mitigation due to the intersection of diferent subjects.
The first and second paragraphs pertain to similar field, Figure 4: Cosine similarity heatmap since both cases involve progressive turnover taxes in Poland, but they are part of diferent legal procedures, since the first one is an appeal and the latter is a reference A second result of the application of the analysis based for a preliminary ruling. Moreover, the first one refers on semantic similarity is the possibility to identify the to the nature of State aid in turnover taxes, whereas the judgement in which the formulation of an interpretative latter refers to freedom of establishment. principle reach its "canonic" formulation. From that mo
Consequently, the first consideration is that a principle ment on it really became an "autonomous object". Indeed, introduced by the Court in order to find the limit of when an interpretative principle is "canonized" it became applicability of the freedom of establishment moves to a part of the European legal knowledge by itself, detached reasoning about competition and State aid through the from the importance of the case law on which it relies. citation. This diference in the citation may be dificult In figure 5 it is possible to visualize what previously to identify, since the level of semantic similarity of the mentioned on substitution. The figure maps the detext is reported as being quite high. In this case the gree of semantic similarity between paragraphs that cite non-semantic score is an efective signal that the citation one another vertically, from Case ECLI:EU:C:2021:201 to contains diferences. While in itself this is not a reason ECLI:EU:C:1990:1258. for dismissing the reference, it may signify the reference The yellow line in Figure 5 represents, for instance, deserves further analysis. that at par. 38 of the judgement EU:C:2006:197, the well
Such a comparison is much more evident if we look at known Enirisorse case, there is an interpretative turning the second and third paragraph, where the conceptual distance is bigger, since it loses the reference to the iden- 8The list of paragraphs analysed is: ECLI:EU:C:2021:202, par tical tax measure. Furthermore, in this citation chain, the 33, ECLI:EU:C:2021:201, par 27, ECLI:EU:C:2016:981, par 53, ECLI:EU:C:2015:470, par 24, ECLI:EU:C:2015:235, par 7The model used is the all-MiniLM-L6-v2, tuned on a 1B sentence 17, ECLI:EU:C:2010:481, par 39, ECLI:EU:C:2006:197, par pairs dataset from diferent sources (Wikipedia, Reddit, and Stack 38, ECLI:EU:C:2003:415, par 74, ECLI:EU:C:2002:294, par 68, Exchange) ECLI:EU:C:1990:125, par 25, ECLI:EU:C:1994:325, par 20 point. Such paragraph contains the fundamental statement according to which a measure must be classified as aid only in light of the cumulative fulfilment of all four conditions required by the Treaties. It refers to diferent precedents, but it recompiles the definition and becomes a canon of interpretation, with small to no semantic differences from that moment on.
While these requirements are mentioned in previous cases, we can verify that the Enirisorse case is where there is a big diference in similarity between citations.
The diference has to do with the fact that for the first time there is mention of the requirement of cumulative fulfillment. This in turn is derived from other cited precedents. In this way Paragraph 38 of Enirisorse moves from such precedents and creates a new interpretative principle expressed in a general and abstract manner.
The passage from precedents more strictly connected
to the specifics of the cases to this general and abstract Figure 5: Similarity of connected paragraphs
formula marks the essential step to the canonization of
the interpretative principle. Hence, from the subsequent
judgment, EU:C:2010:481 (Deutsche Post AG), there is the EU (including EUR-Lex) and US, although they do not
a significant consolidation of the text, with an average always carry to increases in performance, as described in
similarity score higher than 0.8. [
What that can seems an anomaly, namely the low simi- son between the results of a semantic and non-semantic larity between Enirisorse and Deutsche Post, is explained similarity analysis. To this aim it will be possible to idenby the fact that the latter makes a joint citation of para- tify immediately potentially critical nodes. graphs 38 and 39 of the first. Which means that the An interesting consideration that can be expanded innovative part (the one on cumulative fulfillment) is upon is the diference in performance when using a merged with the descriptive part, the above mentioned more semantically-aware measure (by using word- and chain departing from art. 39 of Enirisorse. So the use of sentence-level embeddings), taking into account word semantic similarity applied to citation network allowed ambiguity and the diferent meanings in specific contexts. us to find the turning point in the building of a fundamen- We would enhance the analysis of both verbatim and tal interpretative principle. Moreover, having identified non-verbatim citations by considering the context, in the critical point through the similarity evaluation, it has order to verify if there is a substitution. Furthermore we been easier to detect which are the steps walked by the wold try to develop a tool that may highlight inconsitenCourt in reaching its result. cies in the cited and citing texts.
Another interesting development has to do with the 5. Future development context of the cited cases, and how adding other information could enhance the analysis. In particular identiWhile at the moment the automated assessment on large fying whether paragraphs that cite one another, irrespecscale is not yet feasible, it is one of the possible future de- tive of the degree of similarity, come from comparable velopments, in order to show statistics on the evolution cases. One significant outcome of this could be identifyof a particular citation over time. This is not a straight- ing groups of homogeneous citation (for type of proceforward task, and a more in depth manual analysis is dure, matter, outcome, etc.) in apparently non-innovative needed to correctly identify the relevant information. judgement (i.e. a judgment that does not introduce any significant interpretative innovation). With the context information added to the analysis, it could be possible to 5.1. Semantic similarty shed light on trends in legal discourse. In future work, the similarity analysis could be further With regard to non verbatim citations, i.e. when the developed through the use of domain-specific word em- citation changes, refining the similarity analysis may be beddings, to see whether their adoption may contribute used to better understand when a principle becomes contowards better results. Indeed, some pre-trained models solidated, and the case in which it is defined becomes already exist for the legal domain, such as the Law2Vec a landmark case, or precedent. In this case a deeper model, which has been trained on legal documents from analysis of the diferent branches, as in Figure 5, could
show when a specific citation becomes a general principle, which is then cited verbatim in newer cases.
Furthermore, the similarity analysis may be used in combination with the community detection method to identify implicit citations, and to compare paragraphs that share only part of the citation chain, or that are indirectly connected, citations that do not directly connect but have a common set of citations.
of the case-law in a specific field, academics are used to reading their impact in the specific context, which means comparing the factual relevant aspects, the state-of-theart of the discipline both in the literature and in practice.
However, this activity is complex and time consuming without the support of automated tools. 5.2. Community detection In this paper, a methodology for the extraction and analysis of citations and their use with a selection of A diferent metric that can assist in the legal analysis cases from the European Court of Justice was presented, is the community detection, an evaluation and identifi- and applied to a limited domain as a practical use case cation of groups (clusters) of nodes and the strength of scenario. In particular, it has been shown that an analysis their grouping. of the citation network can give useful insights for legal
In this case, it is possible to identify clusters of cita- scholars in understanding the vast amount of case law tions and where they connect or split diferent branches. available. In particular, it can be useful i) to visualise how In Figure 6 the two metrics (community and centrality) the Court uses citations and their diferent meanings, have been shown together, with the colors referring to ii) to carry temporal analyses by identifying sequences the clusters and the size of the nodes highlighting their of citations over time, and iii) to capture the relevance importance within the graph. of citations by integrating both semantic and network
The diferent groupings correspond to the vertical di- algorithms and metrics. mension mentioned above, and can be split further. This At present, we worked on a relatively small dataset, feature has been used to assist the human analysis and though obtaining appreciable results, sometimes related interpretation of the graph, but it could be also linked to to issues present in the legal studies and debates, as in some future semantic analysis to better identify landmark the case of the interpretative principle about the limit of cases and concepts. national sovereignty in the field of tax law.
In future developments the community detection could wbeayenohfafilntecreidngwtihthe ocathseesr icno mthpeocnietanttiso,nengarabplihn.g for one a References