Regulatory documents are required or provided by authorities in many domains. They commonly point out relevant incidents for speci c scenarios. For those they have to present suitable preventive and reactive measures. We introduce an approach to connect a case-based description of the incidents structure with a case-based description of the according context. This paper shows how to use case-based methods to retrieve, adapt, and reuse incidents descriptions. Subsequently they are used to generate new regulatory documents via case-based reasoning. Case-based reasoning Experience Management Knowledge Management SKOS Semantic Relatedness Natural Language Generation.
A regulatory document describes incidents that are likely to happen in a
certain situation. Preventive measures are elaborated to avoid the occurrence of
relevant incidents and adequate reactions are proposed. Further, harmful
consequences are to be avoided or mitigated. This underlying structure is
represented in the documents structure. Popular examples of regulatory documents
are public events, for the handling of hazardous material or industrial workplace
safety. For a festival, a regulatory document would describe incidents such as
re and relevant measures like the allocation of re-extinguishers. The overall
goal of this work is to support domain experts in writing regulatory documents.
Fundamental considerations have been presented in preceding works [
Copyright c 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
We use natural language processing to connect graph-based and textual knowledge representation. Our goal is to retrieve and adapt passages of documents depicting such regulatory knowledge for usage in another context.
For generating a new document, we need to answer three questions. Which incidents are likely to happen? Which preventive and reactive measures are suitable for each incident under a certain context? How important is each measure? The last question pays attention to the fact, that a limited budget and time does not allow for the implementation of all preventive measures. This sums up to consider a convenient context-based ranking for the incidents and measures. The presented approach is a general framework and easily adaptable to domains providing textual as well as ontological information for the context dependent classi cation, prevention of, and reaction to incidents.
For reasons of simpli cation and consistency we give examples of the domain of public events. Our approach is driven by theoretical and case-based considerations we describe in the rst part of this work. Then we present the experimental setup we used to install a case-study showing practical capabilities of our approach. We nish with related work as well as with discussions and future work. 2
We make the assumption that there exists a corpus of regulatory documents
of a certain domain. The documents are sub-classi ed into passages, that are
connected with incidents or the according measures. Those passages of text are
called information units. The work of identifying these passages was done by
domain experts. All information beyond the textual corpus of available documents
is coded into a knowledge base [
It is very important for the assessment of safety measures to pay respect to the context under that they are applied. For instance to supply rescue boats on a festival F1 besides a river makes sense but for a festival F2 in the forest it is totally senseless. The context are the factual parameters of the environment. If the parameters change, the relevant incidents, the according measures and the importances of both change. Respecting the documentary corpus D, the context is for instance represented by certain parameters whose ful llment is mentioned in the content of each document or the parameters, all documents have in common.
De nition 2. Let KB = (E ; R; D) be the knowledge base. For an entity ei 2 E let Cei E n feig be the context of ei with Cei = fc1; ::; cj g.
For instance, for the previous entities F1 and F2 the context CF1 would be near river and CF2 in the forest. We want to consider relations making entities a preventive or reactive measure to incidents. This means to focus on the chronological order of the execution. In some domains measures are classi ed into before, during and after an incident. We consider the relational classes during and after as uni ed. A measure that is taken before an expected incident is a preventive measure, a measure that is taken during or after an incident is a reactive measure. A measure may be of preventive as well of reactive character. De nition 3. Let RP M C M I be a relation under a context C, indicating which measures are taken in this context C before an incident, making them preventive measures. Let RRM C I M be the analogous relation, indicating which measures are taken after the occurring of an incident making them reactive measures.
Additionally we rely on an importance ranking of incidents and measures under a given context. The importance is quanti ed by assigning a value between 1 for important and 0 for not important.
De nition 4. Let IM P (i; C) 2 ]0; 1] be the importance of an element of I under the context C. Let IM P (m; i; C) 2 ]0; 1] be the importance of a measure m for the incident i under the context C.
For a given context and relevant incident induced by the context the according measures are ordered by importance and classi ed into preventive and reactive measures. Altogether they build a kind of facilitated process snippet we call PIRI (Preventive-Incident-Reactive-Interrelation). The presented model simpli es the real world for facilitation of assessment. Typically there is a cascade of measures that are executed in a speci c order, e.g. in case of re rst evacuate all people, then close the doors and windows. A PIRI-snippet is formally de ned as follows.
De nition 5. Let KB = (E ; R; D) be the knowledge base. Let i 2 I be an incident and C E a context. Then, we de ne a PIRI-snippet P IRI(i; C) = fC; P Gg, where PG = (N,E) is a directed graph. We call PG the PIRI graph with nodes N = C \ M [ fig all measures mentioned by C and the incident i and edges E = fRP M C [ RRM C g \ fig all edges containing the node i. The graph is weighted with node-weights IMP(N,C) and edge weights IMP(E,i,C).
Influence
Preventive Measures
Incident 0,9
0,95 0,85 0,7 I 1
0,8 0,7 0,75
Reactive
Measures
We extend a previously introduced ontology [
secri:Incident
targets
targets
targets
secco:Document
A convenient case-based representation for the so far described scenario
internalizes the document description of incidents and measures. For each incident
mentioned by the regulatory document the preventive and reactive measures are
combined into a PIRI-snippet. We choose a structural case representation using
attributes and their values [
De nition 6. A case c1 = (d1; l1) is de ned by the incident and its context as problem description d1 = fC1; i1g and its solution l1 = f(C1); (m 2 RP M C1 \ i1); (m 2 RRM C1 \ i1)g, the combination of measures targeting the incident i1 under a context C1, separated into preventive and reactive measures.
The problem descriptions and the solutions are conjunctions of elements of the knowledge base. The context may be replaced for a unique identi er naming the context without citing every component. For instance C1 = RegDocument1 then d1 = fRD1; RainStormg and l1 = f(RD1); (W eightT ents^GetF orecast); (CloseLiquidGas ^ LockDoors ^ GetRainCoat ^ Evacuate)g.
De nition 7. The case base CB = fc1; :::; cmg is the collection of all cases ci extracted from available regulatory documents and constructed as described before as PIRI-snippets. A query q to the case base is a conjunct subset of (negated) measures and incidents.
For instance, the query q1 = CloseDoors ^ :LockDoors ^ Evacuation ^ RainStorm retrieves all other PIRI-snippets containing an evacuation and a closing and not locking of doors.
To retrieve cases, we search the case base for similar problem descriptions di for the query q1. To de ne a similarity function, we consider all preventive measures, reactive measures and the incident as individual sub-parts. Each of these parts is then compared by a local similarity measure. With an aggregation function a global similarity measure is composed by weighting with the parameters (!P ; !I ; !R) and summed up as follows: SimPIRI(ck; cl) =
!P SimP (Pk; Pl)+!I SimI (ik; il)+!RSimR(Rk; Rl) =3 (1)
The incidents and measures are classi ed by a taxonomy that was derived from
the connected ontology, building the base for the similarity assessment and
adaptation. The local similarities SimP ; SimI ; SimR are calculated via the taxonomic
order of its elements. The incidents I and the measures P,R are hierarchically
structured. Each element of the hierarchy is assigned with a likelihood
symbolizing the similarity of its sub-elements. The similarity of the leaf-elements is set
to 1 and to 0 for the root element. The similarity increases with depth d of the
element according to for instance simd = 1 1=2d [
!1SimPIRI(ck; cl)+!2SimC (Contextk; Contextl) =2 (2) The context may for instance be the ful llment of a classi cation hierarchy describing the environmental parameters.
For instance, security measures under a context of high consumption of alcoholic beverages are to be considered di erent as under a context of low consumption of alcohol. So SimC is set to the similarity function used in that scenario weighted by the weights !i 2 [0; 1] working as biases. 2.3
The importance ranking can also be used as an order of execution of measures. The most important measures have to be taken rst. But sometimes less important measures have to be taken before other, more important measures. This pays attention to the so called concatenation of circumstances. It is necessary to introduce a (partial) order of measures additionally to the order induced by preventive and reactive and the importance ranking.
De nition 8. For two measures m1 and m2 the constraint m1 m1 should be taken before m2. m2 states that
An obvious problem is as follows. To avoid theft or unauthorized access especially large buildings have to be locked after an evacuation. This can yield people being locked inside the building. In reality it is often too complex or not possible to describe for each incident an order of taking the measures. Additionally in a multi-agent-scenario it is very di cult to execute instructions being too complex or too numerous. We therefore take a simple strategy of providing only rules for pairwise measures, as described before (Evacuate LockDoors).
We exemplify the previous approach by a case study in the domain of public
events. We started with 15 regulatory documents in the domain of public events
that were annotated manually by three di erent domain experts. This corpus
is extended as a basis for the present evaluation. For the annotation process
we developed and evaluated several ontologies. These were used for the classi
cation of public events (OECLA) and the structuring of the according security
documents (OSECCO). The following Table 1 shows the number of ontological
concepts covered by each ontology.
For the ontological description of security incidents we continue the elaboration
of the SECRI ontology (OSECRI ). The SECRI ontology describes the
hierarchical context of security incidents for public events. An excerpt of the ontology can
be seen in Figure 4. In this work we extended the existing ontology by the
capacity of modeling preventive and reactive measures for security incidents in the
domain of public events. All ontologies where implemented using the semantic
wiki KnowWE [
broader broader
For the case-based implementation we made use of myCBR [
To evaluate the similarity assessment induced by the PIRI-strategy we
constructed a post mortem analysis. This means to take every case of the case base
and use it as a query to the same case base. Our similarity measures are
constructed symmetrically, consequently the query is commutative. As context we
use the event classi cation ontology Oecla. The context of each PIRI-snippet is
represented by the factual parameters classifying the event extracted out of the
according regulatory document. The pairwise similarities of the event classi
cation cases are already available due to a post mortem analysis done in previous
work [
In the following we present the strategy for the textual construction of PIRIsnippets and their adaptation for reuse. Figure 6 shows the work ow of breaking documents into reusable information units. Beginning with selected features it shows how they can be put together to form a new document. It presents which methods are used on each level for extraction, retrieval and adaptation. The relevant textual elements were extracted from the corpus and transferred into the ontological structures. The next step is to nd the context dependent information and replace it to make them reusable. We therefore searched for elements of the domain vocabulary. Everything left we considered normal text or context related information that can be abstracted. A strategy for abstraction is to replace words by their class name. For instance a city name is replaced by location data or by the part-of-speech class. The following exemplary text for the incident storm shows, how an according passage of a security document would look in reality. the similarities SimPIRI j SimContext. The value for SimContext was calculated out of SimPIRI and SimECLA which were weighted with 0.5 each. A signi cant change of the retrieval by the incorporated context is marked bold.
Information Units Corpus RD
RD RD
RD RD2
RD
RD1
EXTRACT
SELECT ADAPT AND GENERATE
EXTRACT
SELECT ADAPT
Ontological Concepts Document Similarity Assessment:
ECLA, TF-IDF, SECRI Adaptation: User fills gaps
Sentence Similarity Assessment:
PIRI, Sentence Embeddings Find and adapt similar information units
Concept Similarity Assessment: Word embeddings, Ontologies,
Joint Embeddings "Storm. Get weather information on a regularly basis from the munich weather station. Weight all tents with heavy material or x with ropes. In case of upcoming storm, evacuate the event site using the franz josef avenue and call the re department."
The PIRI-snippet with exemplary importance values for this would be: Preventive(WeightTents(0.9),GetWeatherForecast(0.8))
Incident(Storm)
Reactive(CallFireDepartement(0.9),FullEvacuation(0.8)).
An abstracted information unit for the measure FullEvacuation would be: "[FullEvacuation][StopWord][EventSite][Verb][StopWord][LocationData]" This information unit can be adapted for instance to the measure PartialEvacuation. The ontological concept FullEvacuation is replaced by a retrieved information unit for the new measure. The concept EventSite is for instance replaced by the more speci c concept EventSiteComponent. This information can be retrieved out of other cases because PartialEvacuation is commonly combined with EventSiteComponent. The concept LocationData has to be replaced by the contextual location information which is left to the user. The stop words are inserted and corrected by a natural language generation tool or the user. The generated textual passage before stop word correction and context correction looks as follows:
"[Partial evacuation of the a ected area][the] [EventSiteComponent][using][the][LocationData]"
Figure 7 shows the user interaction and the case-based cycle of natural document extraction and generation. In step (1) a new problem arises. That may be for instance that a new regulatory document is required or an existing document has to be improved as shown in step (2). All features are extracted out of the problem description and the old document at step (3) and queried to the knowledge base at step (4). The retrieved features, phrases and documents are returned in step (5) and adapted in step (6) which requires user interaction. The new regulatory document is used (7) and retained in the corpus enlarging the case base (8). 3.3
The results of the case study for the retrieval of similar information units are very promising even without user support. The incorporation of the contextual paradigm signi cantly improved the simulation of the real world scenario. Regarding the generation of regulatory documents the results were quite good when supported by the user. To answer the initial question, which incidents are likely to happen, the context-based assessment can be used - similar context points to similar incidents. Same holds for the measures that are suitable for an incident. The importance of incidents and measures is made accessible by the percentage of cases covering the incident or measure under a certain context. 4
We started the research for related work to this paper with an overview of state
of the art publications in the domain of natural language generation presented
by Gatt and Krahmer [
There exists some work for the assessment of incidents in di erent domains.
A similar approach we want to mention was presented by Sizov et al. [
The structural integration of context into the case-based assessment was
covered by various authors. Di erent approaches for the incorporation of
context into a case-based decision were proposed by Pla et al. [
In this paper we presented a data structure called PIRI for the representation of a regulatory document describing incidents and according measures. After formally describing it, we transferred the structure into a case-based model. Using this model an approach was shown for the adaptation of similarity measures to di erent context. In a case study the approach was applied to a corpus of regulatory documents of the domain of public events. What we left for future work are strategies for the identi cation of relevant attributes out of existing cases. The application of attribute dependent weights would help to individually adjust the in uence of the context onto the case-based assessment. Additionally, in the eld of document generation the integration of grammar-based natural language generation approaches seems to be promising. To adapt abstracted information units to di erent contexts would help to reduce the needed user support.