Automatic Semantic Annotation for the Easification of Action Rule Legislative Sentences for Specialist Readers Sherry Maynard The University of the West Indies, Cave Hill Campus, Cave Hill, St. Michael, Barbados Abstract This research has applied automatic semantic annotation to a text easification solution that aids non-legal experts in reading legislation as part of their work. It annotates the modality, actor, action, case and condition concepts within action rule legislative sentences. The research first analyzes the lexical and syntactic compositions of a corpus of legislation commonly read by a group of compliance professionals and then extracts data sets of action rule legislative sentences for annotation. The annotation is rule-based, fully automated and utilizes Tregex patterns and Tsurgeon operations. The resultant easified legislative sentences were confirmed by legal experts as having preserved the semantic integrity of the original sentences. In addition, the professionals who participated in the research, reported lower intrinsic and extraneous cognitive loads when they read the easified version of the legislative sentence, when compared to the loads experienced when they read the original version of the same sentence. Keywords 1 Easification, semantic annotation, specialist readers, cognitive load, intrinsic load, extraneous load, 1. Introduction characterized by technical vocabulary, wordiness, repetition, nominalization and the excessive use of binomial and multinomial expressions [2, 4, 6, This research fully automates the semantic 7]. annotation of five concepts found in action-rule legislative sentences. These concepts include Even legal experts resort to reading the modality, actor, action, case and condition. The explanatory notes that accompany a bill rather semantic annotation is part of a larger goal of than the legislative text itself [8, 9]. Similarly, easifying the legislative sentences to aid the some legislators and government officials have comprehension of specialist readers, i.e. non-legal confessed that they do not understand much of the experts reading legislation as part of their work. bills they vote on [10]. Nonetheless, Specialist readers may include professionals in organisations aiming to reduce cost and looking areas such as compliance, audit, finance, risk, for skills beyond legal expertise, are seeking information security, human resources and health persons with investigative, audit and critical and safety. thinking skills to have primary responsibility for the legal compliance function within their It has long been acknowledged that legal organizations [11-13]. Hence, persons with language is complex both in its construction of training in organizational behavior, finance, and the expression of its ideas. Syntactic accounting and information systems are being contributors to this complexity include the density regarded as ideal candidates for this critical of prepositional phrases, the high degree of responsibility [14]. The legal compliance subordination, syntactic discontinuity and lengthy function is an important part of modern businesses sentences [1-5]. In addition, the language is Proceedings of the Fifth Workshop on Automated Semantic Analysis of Information in Legal Text (ASAIL 2021), June 25, 2021, São Paulo, Brazil. EMAIL: sherry.maynard@cavehill.uwi.edu © 2021 Copyright for this paper by its author. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR ht tp: // ceur-ws. Workshop ISSN1613-0073 Proceedi ngs org CEUR Workshop Proceedings (CEUR-WS.org) as they navigate aggressive regulatory overall average of the corpus being 53 words. environments, unconstrained by geographical This average sentence length significantly boundaries [15], and while the cost of legal exceeds Curtotti et al. (2015) recommendation of compliance is high, the cost of non-compliance is keeping legislative sentence lengths below 30 approximately three times higher [16]. words [20]. Furthermore, it is more than double the average sentence length for English academic 2. The Corpus Analysis articles (26 words) [21] and the recommended length for general text of 15–20 words [22]. Sentence length in legislative writing, could be The Barbados legislation that formed the considered a secondary matter when compared to corpus analyzed in this research are those the benefit gained from having as much related commonly read by forty-five members of a ideas together in a single sentence to mitigate compliance professional association in Barbados. against taking the law out of context [23-25]. Seventy four percent of these participants have no legal training and eighty-four percent experience The corpus has on average three coordinating challenges reading legislation. The challenges conjunctions per sentence. In calculating the reported mirrored those associated with the usage of the coordinating conjunctions, detection syntactic and lexical features of legal language as rules were created to identify when ‘and’ / ‘or’ outlined in the introduction. The Flesch reading were used in binomial or multinomial ease scores of these commonly read Barbados expressions; these usages were deducted from the legislations range from 28.1 – 36.6, i.e. they are total conjunctions prior to calculating the ratio of difficult to very difficult to read [17]. The coordinating conjunction per sentence. upcoming sections detail the syntactic and lexical Therefore, the average represents phrasal or features of the corpus. clausal conjoining. In the corpus, ‘or’, ‘and’ and ‘for’ are the primary conjunctions used, 46.58%, 2.1. Syntactic & Lexical Features 27.16% and 20.62% respectively. On the contrary, the coordinating conjunction ‘but’ that The corpus analyzed is composed of the following marks contrast had only 2.16% presence in the Barbados legislation: corpus. Similarly, ‘nor’ and ‘so’ had only 3.10% and 0.38% usage respectively; ‘yet’ had no • Exempt Insurance Act, 1983 occurrences within the corpus. • Companies Act, 1985 • Proceeds of Crime, 1990 In addition, the corpus had on average two • International Business Companies, 1992 subordinating conjunctions per sentence. Relative clauses are heavily used in the corpus, with • Financial Institutions Act, 1997 relative pronouns making up 53.69% of the total • International Financial Services Act, 2002 subordinating conjunctions identified. As with • Anti-Terrorism Act, 2002 coordinating conjunctions, contrast-type • Money Laundering and Financing of subordinating conjunctions (e.g. while, whereas) Terrorism (Prevention and Control) Act, 2010 are seldom used within the corpus; they make up • Financial Services Commission, 2010 0.06% of the total subordinating conjunctions. In addition, there is one occurrence of the similarity Overall, the corpus contains 192155 tokens type conjunctions i.e. the term ‘likewise’. and 3306 sentences. This size is sufficiently large because the conservative nature of legal discourse Curtotti et al (2015) suggested, for improved does not necessitate a large corpus to determine readability of legislative text, to avoid using more its linguistic features. Bhatia (1983) identified than two conjunctions per sentence [20]. The linguistic patterns in legislative text based on a multiple conjunctions create complex sentence single British Parliament act; these findings were structures and syntactic discontinuities that can later confirmed when similar experiments were make sentences difficult to read and understand. repeated on larger corpuses of European, Hong However, for every negative impact a given Kong and Chinese legislative texts [18, 19]. linguistic feature has on the readability of the legislative text there are corresponding benefits The average sentence length of the legislation for the legal domain. For example, while the in the corpus range from 39 – 66 words, with the intensive use of conjunctions can result in cognitive overload for some readers, they usage Table 1: Concept Definitions serves the legal goals of precision and all- CONCEPT DEFINITIONS inclusiveness [18, 26, 27]. Achieving these goals The auxiliary representing the action’s could mean compacting all relevant information Modality modality into a single, long, complex sentence that aids in The person or class of persons performing minimizing the possibility of loopholes and Actor or prohibited from performing a legal evasions in the law [18, 28, 29] action The rights, privileges, powers, obligations Action A sample of 208 sentences (45 – 115 words) or liabilities was extracted from the corpus and their The circumstances / occasions in which Case dependency distance metric calculated. This the legal action applies metric can be used as an indicator of The prerequisites that must occur before Condition comprehension difficulty and has implication for the legal action becomes operable the utilization of readers’ working memory capacities. A recommended threshold is less than The semantic annotations are rule based and 3 words [30]. The average dependency distance utilize Tregex patterns and Tsurgeon operations metric of the sample sentences is 4 words; the [33]. They are fully automated and require no lowest being 2 words and the highest 9 words. human intervention in pre-processing the Therefore, on average four words separate two sentences. The Stanford CoreNLP [34] pipeline elements that share a syntactic relationship, which was used to perform the typical NLP pre- would typically reside alongside each other in the processing tasks of tokenization, sentence sentence structure. segmentation, part of speech tagging and constituency parsing. The output of the parsed Finally, the use of Latin and Old English terms tree is the primary basis for the annotation rules. in the corpus was assessed. The most commonly Nine Tregex pattern – Tsurgeon operation pairs used archaic terms are “thereof”, “forthwith”, were created to detect the five semantic concepts “thereby” and “thereafter”; i.e. 98, 61, 26 and 22 defined in table 1 above. The upcoming sections occurrences respectively. The most commonly provide an overview of the Tregex rules specified used Latin term was “mutatis mutandis”, which is in table 2 below. used 12 times. However, overall the use of Old English and Latin words in the corpus is Table 2: Rule Specification miniscule: 243 Old English words and 30 Latin words. In a corpus of 192155 words, these usages average less than zero for a term-to-sentence ratio. This lexical occurrences support the findings of a study by Dell’Orletta (2012) which showed no significant differences in the lexicon of a set of EU legislation and the stories from the Wall Street Journal. On the contrary, there was a noticeable difference in the underlying syntactic structure of the writings in the two domains [31]. 3. The Semantic Annotation of Legal Concepts The concepts annotated for the easification of action rule legislative sentences are defined in 3.1. The Modality Concept table 1 below. The concepts were adopted from Coode (1845) specification of the essential and The first rule searches for modal auxiliaries optional elements of action rule legislative within the sentence, primarily those at higher sentences [32]. levels within the tree structure. The rule however is deliberately wide reaching to ensure that it captures the correct modal auxiliary needed for the identification of the ‘Actor’ and ‘Action’ any other characters. This regular expression concepts in subsequent rules. Generally, the detects clauses beginning with terms such as targeted modal auxiliary is sandwiched between ‘Where’, ‘When’, ‘Whence’ and extensions such the ‘Actor’ and ‘Action’ sub-trees. The as ‘Whenever’. annotation rule identifies a modal auxiliary which is dominated by a verb phrase (VP). The verb The condition rule identifies adverbial and phrase (VP) is in turn immediately dominated by prepositional phrases that are immediately either a declarative clause or a subordinate clause dominated by a declarative clause and that is immediately dominated by the root of the immediately dominates an adverb or a preposition parsed tree. respectively. In most instances, the case and condition clauses end with a comma. An 3.2. The Actor Concept additional rule searches for this comma and relocates it inside the case and condition sub-trees. The goal is to ensure that during the easification The actor rule detects the noun phrase that acts as process an orphan comma is not left behind. the subject in the English language sentence structure. Therefore, it is a node that must be immediately dominated by nodes that are at high levels within the parse tree, i.e. clauses 4. Related Works immediately dominated by the root node. The actor noun phrase (NP) is the left sister of the verb Boella et al. (2013) implemented a legal phrase (VP) that dominates the modal auxiliary concept detection mechanism using a Support detected in the modality rule. In addition, the rule Vector Machine binary classifier. They utilized accommodates instances where the connection syntactic dependencies to build triplets to train between the NP and the VP is interrupted by an three classifiers to categorize the concepts of adverbial phrase and makes provisions for active roles, passive roles and objects [35]. They complex sentences joined by coordinating used the Italian TULE parser to create the conjunctions, in which case the conjunction node dependency information for the legislative text acts as the head of the embedded sentence. [36]. The results of their approach showed high precision and recall for the detection of the active 3.3. The Action Concept role (precision 97.2% and recall 92.6%), moderate performance for the passive role (precision 100% and recall 26.8%), and low The legal action within the legislative sentence performance for the object role (precision 59.3% is a verb phrase (VP) who is the right sister of the and recall 31.9%). These results were negatively sub-tree that represents the ‘Actor’ concept” and affected by the accuracy of the POS tagger and the which precedes the ‘Modality’ concept. The syntactic parser. For instance, when the POS ‘Action’ verb phrase represents the predicate of tagger did not recognized a noun, it missed an the sentence and is therefore immediately eligible word for a semantic label and the dominated by high-level nodes in the sentence dependency parser could incorrectly label the tree that have direct connections to the root node. semantic relations associated with that term [35]. The annotation rules covered to this point are the One of the reasons given for the use of the core or mandatory concepts in the action-rule machine learning classifier was to overcome the legislative sentences. need for the sequential execution typically associated with pattern-matching rules. 3.4. The Case & Condition Concepts Sleimi et al. (2018) utilized the traditional ordered set of pattern matching rules to detect a The case rule captures the Wh-clauses in the collection of legal concepts and attained high initial sentence position, which typically represent performance across the varying concepts [37]. the case concept. These clauses are subordinate The purpose for the annotation in this work is to clauses that immediately dominates a ‘Wh- support legal requirements engineering. Sleimi et adverbial phrase, which in turn dominates a ‘Wh- al. (2018) used Tregex patterns to extract ten main adverb’ that begins with a upper case ‘W’ phrase level concepts from constituency and followed by a lower case ‘h’ and ‘e’ and then by dependency parsed trees. They established a set of markers for each concept type based on Contiguous and complete sentences are those with dictionaries and ontologies. These markers a non-bulleted format that end with a full stop and formed part of the pattern matching rules. For not a semicolon. The selective nature of the example, one of the patterns for the “Actor” rule sentences in the experiment were driven primarily (subject dependency and NP < actor marker) was by the easification methodology utilized in the represented as a noun phrase in the subject next stage of the experiment and the limitations of dependency position and one that immediately using a constituency parser not trained on dominates a term from the list of actor markers. legislative text. The accuracy of Sleimi et al. (2018) rule detections had overall precision and recall A hundred development sentences (Dev-Set) measures of 87.4% and 85.5% respectively using were extracted from a set of Barbados intellectual 200 statements from Luxembourg traffic laws property legislation and annotated by the author. [37]. These were used to iteratively test the annotation rules during construction. These legislation The level of accuracy attained in the work of included: Sleimi et al. (2018) may result in part to the use of • Trademark Act, 1985 predefined terms within the relevant concept • Patent Act, 2001 repositories. While this approach simplifies the • Industrial Designs, 1981 rule construction, it requires human pre- • Copyright Act, 1998 processing to identify the terms that represent the • Telecommunications Act, 2001 markers for each concept. This technique was utilized in other tools such as, the Gaius T, [38] An assessment of the syntactic composition of and the NomosT, [39]. It however has some the intellectual property legislations was done and drawbacks, for instance, where the repositories compared against those read by the research are inadequately defined, the performance of the participants to ensure a degree of compatibility. detection rules will be negatively affected. In The use of development sentences from a addition, new markers will need to be added to comparable but different legislative domain from extend the detection capabilities of the annotation those read by the participants was to ensure that rules beyond the initial legislative domain. It is the algorithm only processes sentences from the important to note that the work of Sleimi et al. participants’ domain after the rule development (2018) also suffered challenges associated with was frozen. Two test sets were extracted for the the performance of the parser as with the work of purpose of testing the performance of the Boella et al. (2013). Much of Sleimi et al. (2018) annotation rules. detection errors occurred from the constituency parser’s inaccurate attachments of subordination, The first test set (Test Set A) contained one coordination and prepositional phrases and hence hundred and twenty-one sentences extracted from causing the dependency parser to infer incorrect the legislation read by the participants. These dependency relationships amongst the nodes [37]. legislation were primarily from the financial services sector. The average sentence length for 5. Research Experiment Test-set A was 63 words and the average dependency distance metric was four. The author The semantic annotations were done at a annotated Test-set A to provide a gold standard to sentence level using three data sets containing assessment the performance of the annotation action rule sentences that met the following rules. criteria: The second test set (Test-Set B) consisted of • contiguous and complete; sixty-three sentences extracted from the Barbados • a single legal action Road Traffic Act 1981. The average sentence • simple, complex & compound structures; length for Test-set B was 60 words and the average dependency distance metric was four. • a single or compound subject; Two legal experts independently annotated these • at least one modal auxiliary in the upper level sentences. The author was guided by the of the sentence tree; annotation procedures recommended by Hovy • 40 or more words; and Lvid (2010) [40], these included: • dependency distance metric of 3 or more; • The provision of guidelines that define the Table 3: Annotation Results for Dev-Set concepts and the method of highlighting each Precision % concept within the data set; F Measure Extracted Recall % Perfect Match Truth • Giving the annotators practice sentences to CONCEPT % ensure the annotation process is understood and the instructions are clear; • Using annotators with reasonably similar Modality 118 141 118 83.7 100 91.1 levels of education; Actor 116 103 94 97.9 82.5 89.5 • A minimum use of two annotators and have them act independently; Action 117 98 92 98.9 79.3 88.0 • In the absence of a third adjudicator Case 34 33 27 100 79.4 88.5 annotator, any sentences where the Condition 20 17 17 100 85.0 91.9 annotations differ should be discarded; Overall 405 392 348 93 86.6 89.7 The annotators were two lawyers with equivalent educational training. They used the The rules detected 392 annotations from the text highlight feature in Microsoft Office Word to development set. Of these 348 or 86% were highlight each concept using a specified color perfect matches and 57 were missed or partially scheme. As a way of improving the speed and detected annotation (14%). Annotations were reliability of the annotations, the legal experts missed either because of the wrong text or no text were instructed to annotate one concept at a time being detected for a given concept. across all the sentences; for example, the first round of annotations highlights the actor concepts Once the rule construction was frozen, the only, the second round the actions etc. [40]. Since performance of the semantic annotation rules was two annotators were used in the experiment, the tested using Test-Set A and Test-Set B. The thirteen sentences where their annotations algorithm had not seen any of the sentences in differed were deleted from the test set. Hence 50 these test sets prior to the computation of the sentences remained in Test-Set B, which results shown in table 4 and 5 below. represents a 79% agreement between the annotators. In addition, to maximize the limited Table 4: Annotation Results for Test-Set A time of the legal experts, a trade off was made Precision % where the experts annotated all of the mandatory F Measure Extracted Recall % Perfect Match Truth concepts and the case concept; the optional CONCEPT % condition concept was not annotated. The legal experts did not engaged the author during the annotation process. Modality 142 159 142 89.3 100 94.4 Actor 141 134 129 97.0 91.5 94.2 5.1. Results of the Annotations Action 142 131 124 100 87.3 93.2 Case 47 47 41 100 87.2 93.2 The precision, recall and F measures were computed for the development and the two test Condition 34 30 28 100 82.4 90.2 sets. Both lenient and strict computations were Overall 506 501 464 95.7 91.7 93.6 performed; the lenient computation assigned 0.5 points to partial annotations, while the strict Table 4 shows the detection results for Test-set A; computations assigned no points to partial of the 501 annotations detected, 464 were perfect detections, hence treating them as missed match, i.e. 92%; 42 were missed or partially annotations. The measures were done using detected (8%). As expected, based on the strategy GATE Developer 8.0 [41]. Based on the discussed earlier, the results for the modality application of the semantic annotation to the concept showed a 100% recall. The recall for the easification of sentences within the business condition concept was the lowest at 82.4%. context, the partial detections are unacceptable Alternately, there were 100% precision results for therefore only the strict computations were used. Table 3 below shows the results of the annotation the action, case and condition concepts. The F rules using the Dev-set. measures for all the concepts were above ninety, with the overall precision, recall and F measures The detection rules for the three mandatory being 95.7, 91.7 and 93.6 percentage respectively. components of the action rule legislative These overall percentages are not averages of the sentences have a high degree of dependence. individual concept measures, but rather Hence the risk of an initial failure in detecting the computations based on the detection totals across modality concept can be transferred into failed the concepts. actor and action detections. To mitigate this drawback, the modal detection rule was designed The results presented so far, have been compared to be all-inclusive in nature and in all the test sets against truths annotated by the author. The results had a 100% recall results. for Test Set B are compared against truths annotated by the two legal experts participating in The automated detection rules used in this the research; these are shown in table 5 below. research suffered from similar parser related difficulties experienced in other works [35, 37, 42, Table 5: Annotation Results for Test-Set B 43]. In the case of the Stanford constituency parser, while the support website recommended the most up-to-date version of the parser for the Precision % F Measure Extracted Recall % Perfect Match best performance, that recommendation did not Truth CONCEPT % hold true for the legislative text used in this study. The researcher found that the older probabilistic context free grammar parser generated less Modality 50 60 50 83.3 100 90.9 parsing errors than the newer shift-reduce parser. Actor 51 44 44 100 86.3 92.6 Action 51 41 41 100 80.4 89.1 The increase in the parsing errors was directly Case 21 19 19 100 90.5 95.0 linked to the increase in the complexity in the sentence structures. Repeated errors occurred Overall 173 164 154 93.9 89.0 91.4 when the subject of the sentence had one or more embedded qualifiers, when prepositional phrases Of the 173 annotations detected, 154 were perfect broke the continuity between the modal auxiliary match, i.e. 89%; 19 were missed or partially and the main verb, and where compound detected (11%). The performance results on sentences contained ‘or’ conjunctions. In Test-set B are comparable with those on the Test- addition, some sentences were tagged as set A. The overall precision was 93.9%; a 100% fragments if the typical English sentence structure recall measure for the modality concept and the (subject-verb-object) was not detected. Another ‘case’ concept had a recall of 90.5%. The overall interesting parsing error occurred when the term F-measure was 91.4%. ‘issue’ used in the context “shall issue to the applicant” was tagged as a noun instead of a verb. 6. Discussion This miss tagging of the word ‘issue’ reflected the part-of-speech tagger’s interpretation of ‘issue’ as a topic or problem, instead of the act of Generally, the detection results of the semantic distributing something. This error is likely rooted annotations were good, with values of 83 – 100 % in the differences in the genre of the material used for precision, 80 – 100% for recall and 89 – 94% in the training the part of speech tagger when for the F measure. To ensure the annotations were compared to legislative text. fully automatic and hence eliminating the human pre-processing, the implementation deviated from While the current work showed the the use of concept markers utilized in tools such applicability of the annotation rules across as, the Gaius T, [38], NomosT, [39] and the tool legislation in different domains, an expanded by Sleimi et al (2018) [37]. This made the scope of the action rule sentences would further detection rules more complicated but allows for test the generalizability of the annotation rules. scalability and applicability across multiple Therefore, future work includes the utilizing legislations in varying domains. As illustrated in larger, more diverse datasets to test the annotation the data sets, the annotation rules detection rules. However this will also necessitate the capabilities spanned the intellectual property, employment of techniques to overcome the financial services and road traffic legislations. limitations of the part of speech and constituency parsers. 7. The Semantic Annotation Applied • The demands on working memory occurs from conscious cognitive activities; to Easification • Schematic structures are utilized to store information in long-term memory; The semantic annotation of the legal concepts was a necessary step in the easification process. Cognitive load is the demand placed on the The diagram in figure 1 below shows how the storage and processing resources of working semantic annotation fitted into the overall memory. When the mental demands of the algorithm design. It added computer readable activities in working memory, at a given instance, intelligence to the legislative sentence to facilitate exceed an individual’s cognitive capacity, the the automation of the clarifying cognitive individual experiences cognitive overload [45, structuring easification device. 47]. Miller (1956) estimated that working memory stores approximately, 7 (+/- 2) amount of active information chunks, which decay within 15 – 30 seconds if not actively rehearsed [48]. Other researchers suggested a more precise capacity might be 3 - 5 chunks during information processing [49]. These working memory constraints have implications for sentence processing and comprehension. The capacity theory asserts that sentence parsing and memory processes compete for the same pool of resources. Therefore, if sentence processing demands a substantial amount of resources, the resources dedicated to Figure 1: Semantic Annotation applied to storage would be reassigned to meet the Easification processing demand; the resultant reduction in storage capacity can lead to forgetting part of the The easification of legislative sentences is a viable sentence; i.e. forgetting by displacement [50]. alternative to text simplification and is suitable for The longer and more syntactically complex the specialist readers. Unlike text simplification, it sentence, the more likely readers will lose track of focuses less on modifying the text and more on the structural development of the idea [18]. This aiding the mental processes of the readers to can occur when some of the components succumb facilitate the intake of the idea. Consequently, to working memory decay before integration into easification evades a major risk of text the structure being built [51]. Typically, readers simplification, that of inadvertently altering the are unaware of the intricate resource allocations in meaning of the legislative text. This shift in working memory until they reach near full emphasis from the text to the reader increased the capacity and the resultant trade-offs in working likelihood of easification preserving the semantic memory distribution starts to occur [52]. integrity of the legislative text. For the purpose of this research two types of The easification device, clarifying cognitive cognitive loads were measured, intrinsic load and structuring makes the components, the structure extraneous load. The intrinsic load (IL) is the and relationships of the action rule legislative innate complexity of the information or task. This sentences more apparent to specialist readers. It complexity is determined by element draws on cognitive load theory (CLT) [44], which interactivity, which is the degree of offers insights into the consumption of working interconnectivity amongst elements that memory resources during task performance and necessitates them being processed simultaneous. learning. CLT is built on the following basic ideas Intrinsic load is essential for comprehension [47, about the human cognitive architecture (HCA) 53-57]. The extraneous load (EL) is induced by [45, 46]: the way information is presented and organized. • HCA has a very limited working memory It is considered the ‘bad’ load because it results in storage mechanism and a very large long-term cognitive processing that is unrelated to learning memory storage facility; and could impede learning. EL occurs when there is high element interactivity and suboptimal construct that makes the cause and effect communication. The aim is to minimized relationship more obvious. extraneous load [58, 59]. 7.1. Results of the Application to Easification The easification algorithm performs the following functions utilizing the semantic Figure 3: The Easified version of the Securities Act annotations along with additional annotations. It 2002 318A, s48 (2) searches and extracts the semantic annotated elements; annotates additional lower stratum The output illustrated in figure 3 utilizes the elements, extracts the main legislative idea, following If-Then format proposed by Langton inserts logic indicators and generates output (2005) as an extension to the initial easification formats for the readers. device [61]: IF case(s) Take for example section 48 (2) of the IF condition(s), sub-condition(s) Barbados Securities Act 2002 as shown below: THEN legal actor(s) modal legal action(s) “Where a broker is charged with an offence involving fraud or dishonesty or Four lawyers were asked to evaluate the where it is alleged that he has defaulted similarity in the semantics of four pairs of action in the payment of moneys due to a self- rule legislative sentences; the original-unmodified regulatory organisation or to any other version and the corresponding easified version. market actor, the Commission may, if it There was an overarching agreement amongst the considers that it is in the public interest lawyers that the meanings of the original to do so, suspend the registration of the legislative sentences were retained in the easified broker pending the final determination versions. of the charge or allegation.” [60] An additional experiment was also conducted This legislative sentence has 68 words and a to identify the impact of the easified legislative dependence distance metric of 4.75. The sentence on the cognitive load of sixty-three easification algorithm generates the two outputs professionals that participated in this part of the in figure 2 and 3 from the input sentence above. experiment. A modified version of Leppink, Pass et al (2013) cognitive load measurement instrument was used to capture the perceived intrinsic and extraneous load of the participants [62]. Confirmatory Factor Analysis was performed on the modified measurement instrument and it was found to be valid, reliable Figure 2: The Main Idea of Securities Act 2002 and the data collected showed good model fit. In 318A, s48 (2) the experiment, the control group was given the original version of the legislative sentence and the The main legislative idea shown in figure 2, experimental group was given the easified version consist of 18 words; approximately 74% less than of the same legislative sentence. An independent the amount of words in the full sentence (68 sample t-test showed that the lower means for the words). In addition, the complexity of the intrinsic and extraneous loads of the experimental sentence has been reduced in this transient phase group, when compared to the control group were of the sentence processing. The aim is to give the statistically significant. reader the opportunity to create a mental frame of the legislative idea prior to processing the details. Presenting the research participants with the The output in figure 3 below, adds the details with main idea first, temporarily reduced the element informative component labels and the If-Then interactivity of the legislative sentence. In addition, the use of progressive revelation allowed [4] P.R. Macleod, Latin in Legal Writing: An the participants to add the details incrementally, at Inquiry into the Use of Latin in the Modern their own pace; this further assisted them in Legal World. 39 B.C.L. Rev. 235, 1998. managing their intrinsic load. The mean of the [5] J. Crandall, V.R. Charrow, Linguistic intrinsic load, of the experimental group was 3.33 Aspects of Legal Language. 1990. and the control group is 4.57, with a statistically [6] R. Hyland, A Defense of Legal Writing. significant p value of .01038 and a 95% University of Pennsylvania Law Review, confidence interval. Similarly, the mean 1986. 134(3) 599-626. extraneous load of the experimental group was [7] D. Mellinkoff, The Language of the Law. 4.16 and the control group was 5.43 and was 1963, Eugene, OR: Resource Publications. statistically significant with a p value of .021 at a 526. confidence interval of 95%. [8] J. Sheridan, Legislation.gov.uk and Good Law. Civil Service Quarterly, 2014. [9] R. Heaton, When Laws Become Too 8. Conclusion Complex - A review into the causes of complex legislation. 2013. [10] B.C. Jones, Don't Be Silly: Lawmakers This research assessed the lexical and syntactic 'Rarely' Read Legislation and Oftentimes composition of a corpus of Barbados legislation Don't Understand It . . . But That's Okay. read by compliance professionals working in Penn State Law Review, Penn Statim, 2013. Barbados. This research bridged a gap, and 118(7) 7 - 21. developed a solution for specialist readers [11] Deloitte, The changing role of compliance working in the business context where preserving officers. 2014. the semantic integrity of the legislative text is [12] Ernst & Young, Compliance seeks a path to critical to legal compliance. An algorithm was regulatory readiness, in Insurace CCO successfully developed to easify action rule survey. 2014, Ernst & Young Global: legislative sentences. This included creating London. several semantic annotation rules to detect key [13] J.A. Tabuena, The Chief Compliance Officer legal concepts without requiring any human pre- vs the General Counsel: Friend or foe?, in processing of the text. The algorithm outputted an Compliance & Ethics Magazine. 2006, easified legislative sentence with multiple Society of Corporate Compliance and Ethics: perspectives of the legislative idea. The Minneapolis, MN. pp. 4-7 & 10-15. easification of the action rule legislative sentence [14] A. Gross-Schaefer, C.A. Cueto, Ethics & proved effective in lowering the intrinsic and Compliance: The Game-Changer in the extraneous loads of the specialist readers in the Business World. International Journal of research sample, without compromising the Business and Social Science, 2017. 8(2) 57 - semantic integrity of the legislative sentence. 65. Future work will seek to expand the sample size [15] Sovos, The State of Regulatory Compliance. of the participants and to explore the impact of 2017. informed ratings in the cognitive load tests. [16] Ponemon Institute The True Cost of Compliance with Data Protection 9. References Regulations. 2017. [17] R. Flesch and A.J. Gould, The art of readable [1] E. Mattiello, Nominalization in English and writing. 1949. Italian Normative Legal Texts. SESP Across [18] V.K. Bhatia, Simplification v. Easification— Cultures, 2010. 7 129 - 146. The Case of Legal Texts. Applied [2] P.M. Tiersma, Language of Legal Texts, in Linguistics, 1983. 4(1) 42-54. Encyclopedia of Language & Linguistics, B. [19] V.K. Bhatia, N.M. Langton, J. Lung, Legal Keith, Editor. 2006, Elsevier: Oxford. pp. Discourse: Opportunities and treats for 549-556. corpus linguistics, in Discourses in the [3] R.P. Charrow, V.R. Charrow, Making Legal Professions: Perspectives from Corpus Language Understandable: A Linguistics, U. Connor and T.A. Upton, Psycholinguistic Study of Jury Instructions. Editors. 2004, John Benjamins Netherlands. Columbia Law Review, 1979. 79(7) 1306- pp. 203-232. 1374. [20] M. Curtotti, E. McCreath, T. Bruce, S. Frug, [33] R. Levy, G. Andrew. Tregex and Tsurgeon: W. Waibel, C. Nicolas, Machine learning for Tools for querying and manipulating tree readability of legislative sentences, in data structures. Proceedings of the The Fifth Proceedings of the 15th International International Conference on Language Conference on Artificial Intelligence and Resources and Evaluation (LREC’06). Law. 2015, Association for Computing Genoa, Italy: European Language Resources Machinery: San Diego, California. pp. 53– Association (ELRA). 2006. 62. [34] CoreNLP, S. Using the Stanford CoreNLP [21] G. Zhang, H. Liu, A Quantitative Analysis of API.2014, URL: English Variants Based on Dependency http://stanfordnlp.github.io/CoreNLP/api.ht Treebanks. Glottometrics, 2019. 44 16-33. ml. [22] M. Cutts, Oxford Guide to Plain English. [35] G. Boella, L. Di Caro, L. Robaldo. Semantic 2013, URL: Relation Extraction from Legislative Text http://site.ebrary.com/id/10775452. Using Generalized Syntactic Dependencies [23] C. Williams, Legal English and Plain and Support Vector Machines. Proceedings Language: an introduction. ESP Across of. Berlin, Heidelberg: Springer Berlin Cultures, 2004. 1 111-124. Heidelberg. 2013, pp. 218-225. [24] E. Tanner, Sanctity of the Single Legal [36] L. Lesmo, The Turin University Parser at Rule/Single Sentence Structure. Monash U. Evalita 2009. Proceedings of the 11th L. Rev., 2000. 26. Conference of the Italian Association for [25] C. Renton, Renton Commitee Report on Artificial Intelligence. Itally, 2009. legislation. 1975. [37] A. Sleimi, N. Sannier, M. Sabetzadeh, L. [26] V.K. Bhatia, An Investigation into formal Briand, J. Dann, Automated Extraction of and functional characteristics of Semantic Legal Metadata using Natural qualifications in legislative writing and its Language Processing. Proceedings of the application to English for Academic legal IEEE International Requirements purposes. 1982, Univ. of Aston in Engineering Alberta, Canada 2018, pp. 124- Birmingham. 135. [27] V. Bhatia, J. Engberg, M. Gotti, D. Heller, [38] N. Kiyavitskaya, N. Zeni, T.D. Breaux, A.I. Introduction, in Vagueness in Normative Antón, J.R. Cordy, L. Mich, J. Mylopoulos, Texts - Linguistic Insights Studies in Automating the Extraction of Rights and Language and Communication, V. Bhatia, et Obligations for Regulatory Compliance. al., Editors. 2005, Peter Lang International Proceedings of. Berlin, Heidelberg: Springer Academic Publishers pp. 9-21. Berlin Heidelberg. 2008, pp. 154-168. [28] H.E.S. Mattila, Comparative Legal [39] N. Zeni, E.A. Seid, P. Engiel, J. Mylopoulos, Linguistics 2006: Ashgate Pub Co. 347. Building Large Models of Law with [29] C. Frade, Legal Multinomials: Recovering NómosT. Proceedings of. Cham: Springer Possible Meanings from Vague Tags, in International Publishing. 2016, pp. 233-247. Vagueness in Normative Texts Linguistic [40] E. Hovy, J. Lavid, Towards a 'science' of Insights Studies in Language and corpus annotation: A new methodological Communication, V. Bhatia, et al., Editors. challenge for corpus linguistics. 2005, Peter Lang International Academic Interrnational Journal of Translation, 2010. Publishers pp. 133-153. 22(1) 13-36. [30] H. Liu, Dependency distance as a metric of [41] Sheffield, T.U.o., GATE Developer. 1995, language comprehension difficulty. Journal The Univ. of Sheffield: Sheffield, UK. of Cognitive Science, 2008. 9(2) 159 - 191. [42] F. Dell’Orletta, S. Marchi, S. Montemagni, [31] F. Dell’Orletta, S. Marchi, S. Montemagni, G. Venturi, T. Agnoloni, E. Francesconi, B. Plank, G. Venturi, The SPLeT–2012 Domain adaptation for dependency parsing Shared Task on Dependency Parsing of at evalita 2011. International Workshop on Legal Texts, in 4th Workshop on “Semantic Evaluation of Natural Language and Speech Processing of Legal Texts. 2012: Istanbul, Tool for Italian, pp. 58-69. Springer, Berlin, Turkey, May. pp. 42 -51. Heidelberg, 2012. [32] G. Coode, On legalislative Expression or the [43] G. Boella, L. Humphreys, M. Martin, P. Language of the Written law. 1845, William Rossi, L. van der Torre, Eunomos, a Legal Benning and Co.: London. Document and Knowledge Management System to Build Legal Services. Proceedings [55] J. Sweller, Cognitive load during problem of. Berlin, Heidelberg: Springer Berlin solving: Effects on learning. Cognitive Heidelberg. 2012, pp. 131-146. Science, 1988. 12(2) 257-285. [44] J. Sweller, P. Ayres, S. Kalyuga, Cognitive [56] J. Sweller, P. Chandler, Why Some Material Load Theory. Explorations in the Learning Is Difficult to Learn. Cognition and Sciences, Instructional Systems and Instruction, 1994. 12(3) 185-233. Performance Technologies, ed. M. Spector [57] J. Sweller, J.J.G.v. Merrienboer, F.G.W.C. and S. Lajoie. 2011, New York: Springer. Paas, Cognitive Architecture and [45] J. Sweller, Human Cognitive Architecture. Instructional Design. Educational Instructional Science, 2004. 32(1) 9-31. Psychology Review, 1998. 10(3) 251-296. [46] J. Sweller, S. Sweller, Natural Information [58] S. Kalyuga, Informing: A Cognitive Load Processing Systems. Evolutionary Perspective. Informing Science: the Psychology, 2006. 4(1) International Journal of an Emerging 147470490600400135. Transdiscipline, 2011. 14 33-45. [47] J. Sweller, Cognitive Load Theory, Learning [59] J. Sweller, Element Interactivity and Difficulty, and Instructional Design. Intrinsic, Extraneous, and Germane Learning and Instruction, 1994. 4 295-312. Cognitive Load. Educational Psychology [48] G.A. Miller, The Magical Number Seven, Review, 2010. 22(2) 123-138. Plus or Minus Two Some Limits on Our [60] The Securities Act 2002, 318A, s 48(2). Capacity for Processing Information. 2002, The Government of Barbados. Psychological review, 1956. 63 81 - 97. [61] N.M. Langton, Cleaning up the act:using [49] N. Cowan, The magical number 4 in short- plain English in legislation. Clarity - Journal term memory: A reconsideration of mental of the International Association Promoting storage capacit. Behavioural and Brain Plain Legal Language, 2005(54) 28-33. Science, 2000(24) 87-185. [62] J. Leppink, F. Paas, C.P. Van der Vleuten, T. [50] M.A. Just, P.A. Carpenter, A capacity theory Van Gog, J.J. Van Merriënboer, of comprehension: individual differences in Development of an instrument for measuring working memory. Psychological review, different types of cognitive load. Behavior 1992. 99(1) 122-149. Research Methods, 2013. 45(4) 1058-1072. [51] E. Gibson, The dependency locality theory: A distance-based theory of linguistic complexity, in Image, Language, Brain: Papers from the First Mind Articulation Project Symposium, Y.M. Marantz and W. O'Neil, Editors. 2000, The MIT Press: Cambridge, MA. pp. 95-126. [52] P.A. Carpenter, M.A. Just, The Role of Working Memory in Language Comprehension, in Complex Information Processing: The Impact of Herber A. Simon, D. Klahr and K. Kotovsky, Editors. 1989, Lawrence Erlbaum Associates: Hillsdale, NY. pp. 31-68. [53] K.E. DeLeeuw, R.E. Mayer, A Comparison of Three Measures of Cognitive Load: Evidence for Separable Measures of Intrinsic, Extraneous, and Germane Load. Journal of Educational Psychology 2008. 100(1) 223 - 234. [54] W. Schnotz, C. Kürschner, A Reconsideration of Cognitive Load Theory. Educational Psychology Review, 2007. 19(4) 469-508.