Due to the growing activity of legislators, lawyers are in need of tools that would allow them to get a better understanding of an ever-growing corpus of legislative materials. Herein we propose a tool that visualizes and clusters thematically similar amending acts, allowing a lawyer to quickly review related provisions, thus giving an insight into a legislative history of a given legal institution. The methods suggested herein (based on TF-IDF, word and paragraph embeddings and PCA as well as k-mean clustering) are evaluated on the provisions of the Polish Civil Code.
This paper describes first steps undertaken in the development of a software solution used for the visualization of legal change, which aims to provide the user with means to efectively explore a database of amending acts. We aim to develop a solution which is able to group together amending acts that are thematically similar, in an unsupervised manner.
The proof-of-concept implementation studied herein has been
tested using the Polish Civil Code and relevant amending acts issued
from its enactment in 1965 up to November 2018. This legal act was
chosen as a basis for experiments for the following reasons:
(i) While the Code was in force, the Polish economy has
undergone transformation from socialism to capitalism and later
its law had be adapted to the law of the European Union. The
processes pertaining to the recognition of the information
and communication technologies in the domain of law were
also reflected in the Code. Turbulent times, in which the
Code existed, made it subject to almost 90 amending acts.
Some of the sections composing the Civil Code were, in fact,
subject to change multiple times - please consult the heat
map (Fig. 1) for a graphical representation of the number
of times a given legal section was amended. Therefore this
research aimed to assess whether modern machine learning
approaches would be able to recognize and discover discrete
categories of changes (not necessarily the three mentioned
hereinbefore), based only on the text of relevant legal
provisions.
(ii) Even though the amending acts should be as straight-forward
to understand and as precise as possible, the legislative
practice does not always live up to this standard. For example,
the titles of the amending acts do not help in the clustering
2
Practising lawyers need tools that would allow them to track legal
changes, especially due to the increasing activity of legislatures. For
example, in the Polish legal system it has been noted a number of
times that currently the legal system is undergoing the process of
"inflation of law". This notion was recognized by theorists [
The problem of orientation in a dynamically changing system of statues can be mitigated to a degree by the introduction of consolidated texts of acts. In practice, in Poland, the process of consolidation of legal texts is two-fold. On the one hand, there are oficial consolidated texts of legal acts published by the authorities. In practice, those are however seldom used. Lawyers routinely use the legal databases and search engines that are developed by private companies (legal information systems) instead. Currently, the market remains split between C.H. Beck, developer of Legalis information system, and Wolters Kluwer Polska, with their Lex system. The editorial ofices of both of these systems carefully analyse every amending act and issue their versions of the consolidated text. Obviously, the consolidated texts published by those privately-owned enterprises do not have a formal force of law, yet the convenience ofered by them makes those closed and paid platforms a go-to solution for professionals. As far as the recognition of amendments goes, both of these systems ofer, inter alia, a clear dif-like view of the legislative history of a given legal provision (Fig. 2).
However, those solutions do not employ any form of graphical
presentation of amendments. In fact, artificial intelligence
methods are used sparsely in those types of software: for example the
consolidated versions of statutes are created mainly by hand [
As far as the analysis of amending acts in the AI and Law
community goes, the focus up to this time was mainly on the automatic
consolidation of legal texts. For example, authors in [
Whilst this research uses word embeddings techniques as well
as older TF-IDF-based methods, the feasibility of using word
embeddings in eDiscovery procedures was in fact already explored.
In [
In pursuit of the aim outlined in the preceding section a pipeline of an existing tools has been created, with all of them instrumented by Python programms. Python 3.6.8 from Anaconda was used for text processing instrumentation, as well as: gensim 3.4.0 for TFIDF and embeddings calculations, scikit-learn 0.20.2 for clustering, pandas 0.24.0 for data manipulation, nltk 3.4 for text processing and matplotlib 3.0.2 for visualization.
Text processing pipeline can be divided into the following phases: • The generation phase involves reading the consolidated versions of a given statute and extracting the diferences between each successive version. In this phase a textual representation of changes, similar to that shown in Fig. 2, is created. The amending acts are not directly processed: this problem, while itself interesting, is out of the scope of this paper. Usable difs can be created using Linux wdiff command. In fact, for the purpose of this study, a number of dif-generating tools were tested, yet wdiff seemed to be best suited for our instant needs, ofering the clearest results (Table 1 can be consulted for examples of diferences between the output of various dif-generating tools).
The extraction of difs allowed the creation of three diferent
bodies of amendments corpora. For their detailed
description and example Table 2 should be consulted. The first
corpus version (C1) consisted of a complete text of given legal
sections after amending; the second version (C2) included
only the words that were inserted into a given legal section.
However, both of these corpora did not include the texts
that were deleted by an amending act. Yet, the provisions
or parts of them that were struck down can carry at least
the same amount of semantic meaning as those that were
left untouched or added by the legislature. Moreover, in
contemporary legal systems, legislative action is not the only
means of changing the statute. For example, in Poland, the
Constitutional Tribunal was called a "negative legislator".
This term means that, in principle, a Tribunal is unable to
amend a given legal act by adding some provisions, yet is
perfectly capable of striking a given provision down. While
this position is overly simplistic (as Tribunal in practice was
able to pass, inter alia, interpretative judgments, in which
it concludes that a given provision is in accordance with
the Constitution as long as its interpretation is in line with
the one put forth by the Tribunal [
The following methods were used to generate document vectors as a basis of clustering: – TF-IDF, which used corpora C1P , C2P , C3P as well as C1¬P ,
C2¬P , C3¬P .
– word2vec, using the same corpora as TF-IDF. We have used
the pretrained word embeddings for this part, which were
generated for Polish by other research groups [
The external evaluation methods compare machine-generated
clusters with some pre-existing evaluation gold standard,
thus allowing the introduction of standard measures of
precision, recall or the F-score. However, the creation of such
metric in the context of this research is not a straightforward
task. Obvious method of such standard creation involves
classifying of existing data by a legal expert. There are however
a number of concerns regarding this method. Firstly, it is
necessarily subjective. Secondly, machine learning methods
are conceived as means to discover latent patterns existing in
the data, that are missable for humans (cf. [
The internal evaluation results of clustering are shown in Table 3. Generally, the word2vec (skip-gram) model achieved the best results as far as the internal evaluation results are concerned and the model worked best when it was run with the C¬P corpus. It scored the best 2 in terms of silhouette coeficient and Calinski-Harabaz index and well in terms of Davies-Bouldin index. Whilst preprocessing was rather detrimental to the quality of internal evaluation of results in case of various word embeddings implementations, in the case of TF-IDF metric it allowed an increase of the aforementioned quality.
In the case of external evaluation, the word2vec(CBOW) model with C P corpus was ranked the highest, even though the internal 1 evaluation results might have not pointed to that. Fig. 3 shows the visualization of the clusters as generated by this model. The results prove that contemporary word embeddings methods should be considered when preparing a clustering legal assistant. The data preprocessing phase does not have to include lemmatization, stemming or stopwords removal. However, the creation of training set and training itself remains a computationally-intensive challenge. 5
We have shown a proof-of-concept system capable of enhancing a lawyer with visual representation of legal change. The work presented herein was concerned with the Civil Code, however other areas of law (e.g. criminal law) should be put under scrutiny as well. Similarly, as far as the created word embeddings are concerned, we should try to create ones that use larger training sets or are more domain-oriented. Whilst this paper used traditional and well-understood methods for clustering and data dimensionality reduction, more modern techniques should be tested as well.
This work has been based on the legal change as caused by the
amending acts. It should be noted that this extremely positivist (or
formalistic) point of view should be supplemented with more
general notions, in which the statutes themselves do not change,
however the practice of oficials (e.g. judges) who apply given laws does.
Two examples of such practices may be given, one stemming from
the practice of Polish legal system, the other based on European
human rights protection system. As for the former, we have already
mentioned that the Polish Constitutional Tribunal sometimes
resorts to pointing out that there exists a certain interpretation of the
statute that makes it compatible with the constitutional provisions.
Secondly, as far as the European Convention on Human Rights is
concerned, the European Court of Human Rights has on a number
of occasions called it a "living instrument" and has stressed that
its provisions, even if unchanged, should always be interpreted in
the light of present circumstances [
This work, which is concerned with the legislative change, should therefore be viewed in the light of a broader subject of legal change. In future work we aim to employ machine learning techniques to discover and visualize changes stemming not only from the actions of the legislature, but also of other legal actors as well.