Structured occurrence nets (SONs) are a Petri net-based formalism for representing the behaviour of complex systems, capturing concurrent events and interactions between subsystems. SONs can be used in modelling different applications such as accidents, crime and cybercrime investigation. Using graphical representations can greatly benefit investigators by revealing the causal relationships within crime events. However, identifying crime-related information that allows such modelling from unstructured resources is a challenging task. This paper proposes integrating SONs with Natural Language Processing (NLP) to extract and model crime events accurately. This is done by creating a new custom Named Entity Recognition (NER) model to identify additional crime-related entities like weapons, location, and transportation. Furthermore, we propose a syntactic pattern-based approach for verb identification aimed at generating more precise event results. This method relies on analysing manually created SON crime-related models extracted from crime documents. Our goal is to identify Acyclic Nets (ANs) and construct patterns that facilitate event extraction. The NER model has shown acceptable performance, achieving a precision rate of 83.95%, a recall of 87.45%, and an F1-score of 84.84%, suggesting its effectiveness in NER tasks.
Structured occurrence Nets (SONs) [
An extension of SONs is communication structured acyclic nets (CSA-nets) [
It can be challenging for investigators to comprehend a crime and make decisions based on the
massive amounts of information acquired during criminal investigations. In crime investigations,
sources like police reports and witness statements are used to gather relevant information for
analysis. Analysis of such unstructured data sources and extraction of crime events can be
facilitated by the use of Natural Language Processing (NLP) techniques. According to [
The remainder of this paper is organised as follows: Section 2 provides an overview of the research background and related work. Section 3 introduces the basic definitions of acyclic nets and CSA-nets. Section 4 discusses the preliminary rules for mapping text to SON components. Section 4.1 explains the design, experimentation and analysis of the modelling process, presenting both manual and potential automatic models derived from the automatic extraction of an example story. Section 5 discusses the evaluation and testing results of the NER model. Finally, Section 6 concludes the paper and outlines future work.
The visual representation of crime events can be helpful for investigation and analysis. [
Authors in [
SON have shown to be effective in accident, criminal and cybercrime investigations. For
instance, [
An acyclic net [
Formally, an acyclic net is a triple acnet = (P, T, F) = (Pacnet, Tacnet, Facnet), where P and T are disjoint sets of places and transitions respectively, and F ⊆ (P × T ) ∪ (T × P) is a flow relation such that F is acyclic, and, for every t ∈ T , there are p, q ∈ P such that pFt and tFq. Moreover, acnet is an occurrence net if, for each place p ∈ P, there is exactly one t ∈ T such that tF p, and exactly one u ∈ T such that pFu.
An acyclic net is well-formed if for every step sequence starting from the default initial marking (i.e., the set of places without incoming arcs), no transition occurs more than once, and the sets of post-places of the transitions which have occurred are disjoint. Note that all occurrence nets are well-formed.
A communication structured acyclic net [
= = =
Pacnet1 ∪ · · · ∪ Tacnet1 ∪ · · · ∪ Facnet1 ∪ · · · ∪
2. Q is a set of buffer places and W ⊆ (Q × Tcsan) ∪ (Tcsan × Q) is a set of arcs adjacent to the buffer places satisfying the following: a) Q ∩ (Pcsan ∪ Tcsan) = ∅. b) For every buffer place q: i. There is at least one transition t such that tW q and at least one transition u such that qWu. ii. If tW q and qWu then transitions t and u belong to different component acyclic nets.
That is, in addition to requiring the disjointness of the component acyclic nets and the buffer places, it is required that buffer places pass tokens between different component acyclic nets. In the step semantics of CSA-nets, the role of the buffer places is special as they can ‘instantaneously’ pass tokens from transitions producing them to transitions needing them. In this way, cycles involving only the buffer places and transitions do not stop steps from being executable.
A CSA-net csan = (acnet1, . . . , acnetn, Q,W ) is a communication structured occurrence net (or CSO-net) if the following hold 1. The component acyclic nets are occurrence nets. 2. For every q ∈ Q, there is exactly one t ∈ Tcsan such that tW q, and exactly one u ∈ Tcsan such that qWu.
3. No place in Pcsan belongs to a cycle in the graph of Fcsan ∪ W .
That is, only cycles involving buffer places are allowed.
All CSO-nets are well-formed in a sense similar to that of well-formed acyclic nets. As a result, they support clear notions of, in particular, causality and concurrency between transitions.
In this paper, we use occurrence nets rather than more general acyclic nets. However, this will change in the future work when we move to the next stages of our work where alternative statements in textual documents are taken into account.
This section outlines steps taken to extract useful information from textual resources for modelling and visualising using CSA-nets. Such a visualisation can help investigators comprehend, e.g., the dynamics of crime incidents and the interactions among involved parties. Since CSA-nets consist of interconnected acyclic nets, our approach involves first identifying individual acyclic nets and their components, followed by the construction of communication links. Each acyclic net (acnet) comprises places (circle nodes) and events/transitions (square nodes), representing the progression of system execution from one state to another. SON models facilitate a visual understanding of the behaviour of complex systems to enhance comprehension.
An example of a complex system is crime investigation, which involves examining numerous variables to assist investigators in decision-making. Extracting information usually involves unstructured data, like written reports, and so presents significant challenges. Investigators often depend on various sources including police reports and witness statements.
CSA-nets provide a method for analysing these types of crimes by representing events and their relationships revealing causal links between them. However, existing CSA-net approaches lack the ability to automatically extract information from written sources and reports. To address this, we leverage statistical Natural Language Processing models by integrating SPACY with CSA-nets to automatically extract useful information from written sources and represent crime events in CSA-nets. In order to enhance the accuracy and consistency of the extraction process, the following rules have been established to guide the linking of natural language with SONs. □ ENTITIES in crime text are the representation of SON acyclic nets. An “entity” is a proper noun that refers to a distinct real-world object or entity, such as a person, organisation, or location. □ VERBS represents SON EVENTS/TRANSITIONS within acyclic nets. □ Shared VERBS by several entities will result in the formation of communication buffer places (communication links).
To illustrate how text is represented in SONs, consider the following phrase: “Allen shot ... ”. Using NLP, the extractor will first identify Allen as an entity and then assign the tag ’PERSON’. Next, the extractor searches for verbs related to the entity Allen. In this case, the verb shot is identified as an event related to Allen. This results in the construction of the occurrence net shown in Figure 1.
Allen p1 shot p2
The aim is to automatically extract CSA-net components (transitions and acyclic nets) from free textual resources to prepare for modelling. Our initial step involved finding a method to comprehend how SON expert users identify transitions and acyclic nets. To achieve this, we engaged expert users to model stories in SONs, which we evaluated to observe the variations in the resulting models. Following a comprehensive assessment of these models, we recognised that there is a mapping from English text to formal models, and concluded that one can represent named entities as acyclic nets and map verbs to events/transitions.
Regarding entities, we decided to create a new custom NER model and introduced new labels for the model to learn. Essential entities such as weapons, family members, and transportation are not identified by the default NER model. Therefore, we developed and trained a new custom NER model on these new labels to generate more precise acyclic nets. Subsequently, when evaluating how expert users identified transitions (verbs), we found it to be a challenging task due to the complexity of language structure. Consequently, identifying verb types was not possible, and so we examined the syntactical verb patterns in verbs manually as identified by SON expert users, resulting in the identification of patterns for recognising transitions.
A customised new NER model was developed to enhance crime entity detection. The model
underwent training using a dataset comprising approximately 256 concise crime stories [
One noteworthy observation was the frequent mention of weapons, transportation, and family members, which were often undetected by default NER models due to the absence of specifically related labels. In response to these findings, we introduced new entity labels to reinforce the model’s accuracy and comprehensiveness. The newly added labels encompass a range of entities including [‘LAW_ENFOR’, ‘WITNESS’, ‘WEAPON’, ‘TRANSPORTATION’, ‘RELATIVE’, ‘PROFESSION’, ‘UNNAMEDPER’]. (Note: ‘UNNAMEDPER’ is a NER tag that we propose for tagging unnamed people in the text. For instance, words like ’one woman’ and ’third employee’ refer to people but are not labelled. Using this tag, we can identify such ambiguous words.) This improvement aims to significantly advance the NER model’s ability to recognise diverse crime-related entities, ensuring a more effective analysis of crime narratives. To train the model on these new entities, we collected approximately 334 crime stories, dividing them into 256 for training data and 78 for evaluation. This process involved several steps, including manual annotation of the testing data, training, evaluation, and finally testing of the custom NER model.
The annotation process involves human intervention for identifying and labelling named entities
(such as people’s names, weapons, locations) in text data according to predefined categories.
Annotators highlight entities and assign the appropriate labels using annotation tools such as
TagEditor [
Another challenge arises from the presence of pronouns. To handle this, we employed
coreferencing models, which involve resolving pronouns and mentions to their original entity. To
address this issue, we integrated NEURALCOREF [
Identifying verbs that accurately represent SON events from crime stories presents a significant challenge. To address this challenge, we adopted a pattern-based approach to identify verbs representing events from manually created models. To ensure precision, we examined various crime-related SON models created by expert users, selecting verbs identified by at least 50% of the modellers. Through dependency parsing, we determined the positional relationships of these verbs within sentences concerning surrounding words. Specifically, we identified the types of words and their dependency relationships preceding and following specified verbs by collecting all verbs representing events from manually created models. We then analysed the part-of-speech tags and syntactical tags of the words before and after the verbs. This enabled us to recognise syntactical patterns for event identification, reflecting the human process observed in manual SON models. This approach led to the identification of seven patterns, outlined in Table 1. For instance, the verb may be preceded by a noun, proper noun, or pronoun subject and followed by a noun, proper noun, or pronoun object. The second column in the table indicates the dependency relationship between the words around the verb and the verb itself. For instance, the noun that falls before the verb is a subject, and the noun after the verb is an object.
The pattern approach employs linguistic features to identify verbs representing events. SPACY processes text through tokenisation and linguistic annotations, assigning each token (word) with part-of-speech (POS) tags and dependency labels. These annotations facilitate the accurate identification of verbs, nouns, and other parts of speech, as well as the determination of syntactic relationships within sentences. By analysing manually selected verbs and their associated patterns from expert users, we compiled a list of patterns. If a verb pattern matches any of the patterns in this list, it is considered a SON event. Furthermore, the evaluation of verb extraction is unnecessary, as it relies more on pattern-rule-based techniques rather than statistical predictive identification.
Part-of-Speech Relationship Dependency Relationship [noun, proper noun, pronoun][EVENT][noun, [nsubj, nsubjpass] [dobj, iobj, pobj] proper noun, pronoun] [proper noun, pronoun][EVENT][verb, adverb] [nsubj, nsubjpass] [ccomp, conj, advcl, xcomp, advmod] [verb][EVENT][verb] [advcl] [advcl] [noun, proper noun, pronoun][EVENT][verb, [dobj, iobj, pobj] [ccomp, conj, advcl, xcomp, adposition] advmod, pcomp, root] [verb, adverb][EVENT][verb, auxiliary] [advcl, advomd, ccomp][advcl, conj, ccomp, root] [pronoun][EVENT][noun] [nsubj] [relcl] [noun][EVENT][verb] [nsubj] [advcl]
We applied the NER model to extract entities, and then, by employing the proposed verb-patternbased identification, we also extracted the verbs. Subsequently, we generated lists for each entity and appended all relevant verbs to each entity list. Algorithm 1 describes the comprehensive extraction process and the assignment of verbs to entities. Figure 2 shows the structure of the output lists (one list for each story), denoted as ExtractedData_Lists. Each list within the ExtractedData_Lists contains information regarding an entity and its corresponding verbs extracted from the text. For each entity and each verb associate with it, three information items are provided: the verb (v), the location of the verb within the sentence (vsent_loc), and the index of the verb within the text (vindex). Following the extraction from each story, a list can contain multiple entities alongside their respective verbs extracted from the text.
ExtractedData_Lists = [︁ [entity1, v1,1, vs1e,1nt_loc, vi1n,1dex, . . . , v1,k1 , vs1e,kn1t_loc, vi1n,kd1ex], . .
.
[entityn, vn,1, vsne,1nt_loc, vinn,1dex, . . . , vn,kn , vsne,knnt_loc, vinn,kdnex]︁]
To transform entity lists into CSA-nets, a sequential acyclic net is constructed to represent each entity list. This involves creating a line-like structure where each verb encountered in the entity list leads to the construction of an event. The construction of acyclic net acnetm starts by creating an initial place pm,0. Then, for every verb vm,i encountered in the entitym list, a corresponding event tm,i is constructed, followed by the insertion of a new place pm,i, as illustrated in Figure 4.
Following the construction of acyclic nets, we use the verb information (i.e., v, vsent_loc, and vindex) to construct the communication link later on. This will involve looping and searching for similar verb information among the entities. If such information is found, a communication link is established.
Communication between different acyclic nets is possible and leads to CSA-nets. Such nets are constructed when transitions in component acyclic nets are connected using buffer places. To accomplish this, we have already extracted verbs along with their corresponding information, such as the specific location within the text and the sentence in which the verb is found. This extracted information enables us to establish communication by identifying identical verbs across different acyclic nets. Upon discovering a match, new buffer places are created to facilitate communication (synchronisation). For instance, if acyclic nets acnetm (modelling entitym) and acnetl (modelling entityl) are such that vm,i = vl, j, then we add two buffer places, btv and bvt , and four arcs between transitions t and v (modelling vm,i and vl, j, respectively) to ensure their synchronicity (the added arcs are: (t, btv), (btv, v), (v, bvt ), and (bvt ,t)). To improve readability, in the diagrams we depict such an addition of places and arcs by showing just a single buffer place linked by dashed edges with t and v, as shown in Figure 4.
Algorithm 1 Entities and Events Extraction Algorithm
To illustrate the applicability of our methodology, consider the following text (STORY A): “Brenton Rhasheem Davis, 24, allegedly shot and killed Tommy Jones III, 19, with a 9mm handgun in the parking lot of a Walgreens in Columbus, Georgia. Witnesses say they saw Davis and Jones enter the store together. When they left toward the parking lot, Jones ran from Davis who then allegedly shot Jones in the leg and back. Davis then walked up to Jones, who had collapsed, and allegedly fired more shots. Jones was shot at least four times. Nine shell casing from the 9mm handgun were found at the scene. A witness added that Davis then picked up a bag Jones had been carrying and then fled the scene. It is unknown what was inside the bag. Jones had gone to the store to meet Davis in order to buy a 45-caliber handgun from him. The two men did not previously know each other. Davis was charged with one count of murder.”
Here, we illustrate the diverse modelling approaches applied to STORY A. Figure 3 shows the manually created models, while Figure 4 illustrates a potential automatically extracted model based on the results of the proposed extraction method. It should be noted that humans tend to abstract and therefore selectively omit verbs they consider irrelevant. Conversely, the proposed verb extraction method aims to identify all verbs present within the pattern, which leads to the generation of larger SON models. However, abstraction may be employed as an option. This could be explored further in future work, with the aim of utilising BSONs as an abstraction method for the produced SONs.
Evaluation focused on the NER model using the 20% of the data (78 stories from [
Then the model was tested on 54 cases related to police shootings obtained from [
A new approach to extracting useful information for SON modelling from crime reports is proposed. Initially, entities and verbs are identified to represent and describe the acyclic nets and events. New named entity labels are introduced, and a custom NER model is trained on these Witness p1 p1 BRD T J enter_store q1 p2 saw q1 enter q2 p2 le f t Store p15
q9 enter p16
Bag p5 p15 p7 killed q7 q5 q6 p6 p14 p16 p7 q7 p17 p5 carrying
p5 Bag
q6 p17 carrying p18 pickedup p20 enter_store le f t_store shoots walks more_shoots takes_bag_ f lees p8 enter_store p9 p10 run p11
p12 collapses p13 q3 p3 q4
p4 (a) Modeller 1 shot walked f ired pickedup f led p3 q3 p4 p5 q4
p6 q5 p8 q8 p9 p10 q10 p1 hadGone p2 saw p3 enter p4 le f t p5 ran p6collapses p7 f ired p8 died p9 0.9mm_handgun
shot p1 saw q3 enter q4 saw saw p15 buy buy labels by annotating crime stories. Additionally, verbs are identified using syntactic patterns generated from manually modelled crime documents using CSA-nets. An algorithm for the identification and extraction process is proposed, and the experiment showed acceptable results.
Future work will focus on structurally and formally representing the extracted data and modelling it using SONCraft. Furthermore, improvements to this approach could involve developing new tagger and NER models to enhance accuracy. Additionally, we will investigate the identification and construction of concurrent and conflicting events within acyclic nets.
Appreciation is extended to the anonymous referees for their comments and suggestions which have led to improvements in the content and presentation of this paper. Furthermore, sincere gratitude is extended to Maciej Koutny and Anirban Bhattacharyya for their invaluable guidance and unwavering help throughout the development of this paper.