Entity Graph Extraction from Legal Acts – a Prototype for a Use Case in Policy Design Analysis

Challenges associated with IG annotation could be divided into three groups: issues with the extraction of atomic statements from sentences, issues with atomic statements classification, and problems with transforming a passive voice into an active one in regulative statements. The challenge raised by the extraction of atomic statements comes from the fact that an atomic statement, as IG defines it, can consist of only single instances of Attribute, Aim, and Object (Direct and Indirect). At the same time, one sentence very often regulates several actions of one actor or even various actors. The way an atomic statement is defined leads to a situation when a relatively simple statement: On the basis of proposals submitted by States Parties, and in accordance with criteria to be defined by the Committee and approved by the General Assembly, the Committee shall periodically select and promote national, subregional and regional programmes, projects and activities for the safeguarding of the heritage which it considers best reflect the principles and objectives of this Convention, taking into account the special needs of developing countries. It is divided into 18 atomic statements. This fragmentation is caused by the fact that this sentence indicates two Aims – select and promote – and nine Direct Objects:

• •

  national programmes, subregional programmes, and regional programmes;

• •

  national projects, subregional projects, and regional projects;

• •

  national activities, subregional activities, and regional activities.

At the same time, other IG components are constant:

• •

  Attribute: the Committee;

• •

  Deontic: shall;

• •

  Execution Constrains: "on the basis of proposals submitted by States Parties," "in accordance with criteria to be defined by the Committee and approved by the General Assembly," "periodically," and "taking into account the special needs of developing countries."

This configuration of IG components leads to extracting several atomic statements where Aims and Objects are changing, and other components remain the same. All such atomic statements could be expressed in a schematic one: Based on of proposals submitted by States Parties, and following criteria to be defined by the Committee and approved by the General Assembly, the Committee shall periodically [Aim] [Direct Object], taking into account the unique needs of developing countries.

These types of challenges are raised even higher by enumerations. The presence of enumeration further increases the number of output atomic sentences. They are usually used in sentences that consist of several compressed ones – the first part of the sentence is constant when the second part is written as an enumeration element. The challenge here is to identify individual sentences and, secondly, identify separate statements. As a result, Article 13 of the Convention describes State Parties’ obligations toward safeguarding intangible cultural heritage that consists of four enumeration elements (where the last element consists of three enumeration elements) at the end is divided into 24 atomic statements.

The second challenge associated with IG annotation is statement classification. IG identifies two types of statements: constitutive ones – definitions and statements describing functions of entities – and regulative ones – regulating actions taken by entities. The difficulty of statements classification comes from the similarity of constitutive statements describing the essential role of an entity and regulative statements controlling its behaviour. The institutional statement derived from Article 10 par. 1 – The UNESCO Secretariat shall assist the Committee – could be understood as a statement regulating the action of the UNESCO Secretariat – the body is expected to help the Committee. However, the statement is constitutional and assigns a vital function to the UNESCO Secretariat. The UNESCO Secretariat is a large body composed of many sub-groups, each specialising in assisting different UNESCO conventions and programmes. In the case of the 2003 UNESCO Convention, the sub-group in the Secretariat is labelled the Living Heritage Entity (LHE), which is solely dedicated to serving this Convention. The body was created, among other things, to assist the Committee. However, it does not assist any conventional Committee but the specific organ of the 2003 UNESCO Convention: Intergovernmental Committee for the Safeguarding of the Intangible Cultural Heritage. In this case, as in the case of many other statements, proper classification of the statement is impossible without expert knowledge of UNESCO affairs.

The last IG-related challenge regards the passive voice. One of IG’s assumptions is that regulative statements can only be expressed as an active voice. In the case of each regulative statement, the Attribute – the addressee of the statement – has to be identified. Sometimes, the task is easy. In its original form, Article 1O par. 1 has a form of The Committee shall be assisted by the UNESCO Secretariat. All pieces of information needed to transform this sentence into a statement in active voice are provided by the sentence itself. In other cases, an Attribute has to be found in previous sentences or the broader context of the regulation. However, sometimes, it is challenging to determine who should be an Attribute in a statement. Article 30 par. 2 states that The report shall be brought to the attention of the General Conference of UNESCO. The sentence does not provide information on who is responsible for bringing attention. The statement’s context also does not provide a clear answer to this question – at least two bodies are mentioned in previous sentences that could be responsible for the action (the Committee and the General Assembly). Only based on the domain expert knowledge, it is possible to determine that the addressee of this statement is the UNESCO Secretariat.

Many researchers in the political science domain agree that automated text analysis (e.g., reading, annotating, graph modelling) should become a standard tool for political scientists. However, methodologists and scientists from computer science must contribute new methods, new validation techniques, and adequate technologies with applications Grimmer and Stewart (2013); Long et al. (2019). Many legal and political texts can significantly benefit from the support of automated reasoning. Such support depends on the existence of a logical formalisation of the legal text. Nevertheless, legal documents use a rich language that is not easily accessible to annotate Libal (2020).

The application of IG allows extracting standardised data on rules from legal texts. However, the IG schema is not associated with any popular way of data analysis, which is challenging to utilise in research. Therefore, researchers have been searching for ways in which IG data could be transformed into more popular data formats. The most promising direction is graph theory. Statements can be seen as edges that connect nodes represented by Attributes, Objects, Constituted Entities, and Constituting Properties. In this regard, a set of edges –statements constitutes a graph representing an individual regulation. This approach allows using popular network analysis indicators to characterise the institutional setting developed by the regulation. A legal act is seen here as a list prescribed by the law relations between different objects and actors. Three approaches have been developed to transform IG data into graph data. The first is represented by T. Heikkila and C. Weible Heikkila and Weible (2018). IG is used here to identify actors in a variety of legal documents. They are linked in a single document and across many regulations. The aim is to identify a network of influential actors in a particular policy and use network analysis to characterise their associations. Heikkila and Weible used their method to study regulation on shell gas in Colorado and found that the institutional actors governing this industry create a polycentric system. The second approach to IG and network analysis is represented by research by O. Thomas and E. Schlager Olivier and Schlager ; Olivier (2019a). Here, data transformation from IG to graph allows for comparison intensity with which different aspects of institutional actors’ cooperation are regulated. The authors used this approach to study water management systems in the USA to link IG data with graph representation. It is built around assumptions that network analysis based on IG allows describing institutional processes written into legal documents in great detail. This approach was also used in a study on the climate adaptation of transport infrastructures in the Port of Rotterdam  Mesdaghi et al. (2022).

In contrast, our approach discussed in detail in the use case Section 4 is more actor-centred than the previous ones. We aim to compare the positions of prominent actors in a network of rules and use indicators developed by network analysis to describe the place of institutional actors in the formal design of the UNESCO Convention. We also confront these metrics with IG data to determine if an actor’s position in a rules network correlates with the place in IG statements.

Information structured in the form of a graph, with nodes representing entities and edges representing relationships between entities, can be built manually or using automatic information extraction methods. Statistical models can be used to expand and complete a knowledge graph by inferring missing facts Nickel et al. (2016). There are different technologies and applications of language understanding, knowledge acquisition, and intelligent services in the context of graph theory. This domain still is in statu nascendi and includes the following key research constructs: graph representation and reasoning; information extraction and graph construction; link data, information/knowledge integration and information/knowledge graph storage management; NLP understanding, semantic computing, and information/knowledge graph mining as well as information/knowledge graph application Zhu et al. (2019). Graph theory concepts potentially apply in computer science for many purposes. However, in the last decade, the use of graphs in various fields of social science and human life has become a source of innovative solutions Majeed and Rauf (2020); Spyropoulos et al. (2021); Lüschow (2021).

The collective action approach is the foundation of preparing new scientific research methods and solving current social sciences problems, which affects the structure of stakeholder networks differently in policy settings. Interactions do not usually occur in an institutional vacuum; they are guided and constrained by agreed-on rules. It is also essential to understand the parameters that drive and constrain them. Hence, the design of institutional behaviours can be measured through Networks of Prescribed Interactions (NPIs), capturing patterns of interactions mandated by formal rules Olivier (2019b). Social network analysis offers considerable potential for understanding relational data, which are the contacts, ties, and connections which relate one agent to another and are not reduced to the properties of the individual agents themselves. So, in network analysis, the relations are treated as expressing the linkages between agents. That analysis consists of a body of qualitative measures of network structure. That relational data are central to the principal concerns of social science tradition, emphasising the investigation of the structure of social action. The structures are built from relations, and the structural concerns can be pursued through the collection and analysis of relational data Liu (2011). Our approach uses social network analysis powered by institutional grammar (IG).

The first significant challenge regarding building the above system is a semantic understanding of legal policy texts. To automate the process of legal text analysis, we utilize Institutional Grammar as an intermediate layer for text interpretation and extraction of crucial entities. Thus, in our study, we incorporate IG as a text-semantic bridge. On one side, IG is a tool for describing and analyzing relations between institutional actors in a more systematic and structured way than natural language texts. We used the tagset of IG. However, there are still open issues we currently address with the manual work of policy design researchers. Firstly, there is still a need to define more precisely some IG statements, e.g., context data like Activation Conditions entities. Some entities of IG are not defined precisely and structuralized as a graph structure or formal ontology with classes and logical rules between them. Such structural ordered and precisely defined data (a graph or even an ontology) further can be used, for example, for defining rules of institutions and network analysis and reasoning Olivier et al. (2020). Other issues we have yet to address manually are: atomic statement extractions, their transformations into active voice, and the classification of atomic statements – see Section 2.

The second challenge faced by our prototype is to transform IG format with the extracted entities into graph structure and format, which is a proper data structure for automated analysis. Social science has developed, under the label of "social network analysis", significant expertise in expressing information about particular elements of our reality in computer-readable data structures to analyze interrelation between entities. This task is accomplished using IG as an intermediate layer between a legal text and mainly graph-based metrics expressing interrelations between entities.

Data acquisition and storage is the first part of the workflow with the legal data. Crawling and databases of crawled data are needed to collect and preserve information for further analysis. In this part, the important prototype modules are (1) crawling, (2) document extractor, (3) data quality filtration, and (4) data warehousing / database (see Figure 4). The technologies used in this stage are standard tool such as most popular scrapers²²2https://scrapy.org, PDF parsers³³3https://github.com/py-pdf/PyPDF2,https://pypi.org/project/pdfminer/, and an optical character recognition (OCR) parsers⁴⁴4Tesseract – an OCR tool for Python https://pypi.org/project/pytesseract/. for the documents that are not in a text format.

We extracted hypergraph information from our analyzed documents to get insights into a given document. The hypergraph vertices are the essential entities – actors (e.g. institutions) and objects (e.g. legal documents). The statements (in the analyzed documents) represent connections (edges) in this structure.

Let $H=(V,E)$ be a hypergraph, where verticles $V$ represent actors and objects appearing in the document. $E=(e_{1},...,e_{k}),e_{i}\subset V\forall_{i\in 1,...,k}$ means hyperedges that are created where objects appear together in one statement from the analyzed legal documents. We chose hypergraphs, not graphs, because relations are formed by more than one vertex.

After building the hypergraph, in our analysis, we use different graph measures, e.g., we adopt a hypergraph centrality definition proposed by authors of Aksoy et al. (2020). We can also use other graph metrics to help describe information and relations between entities and define the interrelations and expositions of particular actors.

In this use case, we defined two measures – visibility and centrality. The visibility is based on Institutional Grammar and envisions how entities are exposed in the analyzed legal document. The centrality is based on the built hypergraph and describes the influence and centrality in the legal act formed in the analyzed document.

The aim of our use case was, first of all, to investigate potential tensions between locations of actors in institutional statements and their places in a network that models a whole institutional design of the 2003 Convention. Does the placement of an actor in statements predicate its position in the institutional systems? We were particularly interested in the position of the Secretariat. We hypothesized that the Secretariat is a background actor – it is placed in less important IG components, but it is an essential node in the institutional design of the Convention expressed as a graph.

Based on our analysis, we can state that the Secretariat does not hold any particular interesting position. The Secretariat has the lowest level of centrality among all actors and is placed in the middle of the visibility ranking – see Table 4. Suppose the Secretariat holds a unique position in the 2003 Convention institutional regimes. In that case, it is granted not by the Convention itself but other more operational documents like Directives to the Convention. They should be analyzed in future research.

There are, however, two interesting facts that our analysis exposed. The first one is that all actors, except one, have higher levels of centrality than visibility. They are above the dashed line in Figure 10. They all, including the Secretariat, are "relative" hidden actors – their scores of visibility scores are significantly lower than their centrality scores. The second interesting fact is that State Party is the only overexposed actor in the 2003 Convention. This actor is placed below the dashed line. However, we can only talk about the relative "over-exposition" of the State Party – it has, after all, the highest score of centrality among all actors. Based on this observation, we can tell that states that are parties of the Convention are crucial actors in the institutional setting created by the 2003 UNESCO Convention, but not as vital as their place in institutional statements indicates. This observation probably reflects the fact that international conventions are usually formulated in a manner that, above all, regulates the actions of actors being parties of international conventions - states themselves.