Process Model Generation from Natural Language Text - Advanced Information Systems Engineering - page 485

Afterwards, each sentence is parsed by the Stanford Parser using the factored
model for English [11]. We utilize the factored model and not the pure proba-
bilistic context free grammar, because it provides better results in determining
the dependencies between markers as “if” or “then”, which are important for
the process model generation. Next, complex sentences are split into individual
phrases. This is accomplished by scanning for sentence tags on the top level of
the Parse Tree and within nested prepositional, adverbial, and noun phrases.
Once the sentence is broken down into individual constituent phrases, actions
can be extracted. First, we determine whether the parsedSentence is in active
or passive voice by searching for the appropriate grammatical relations (Issue
1.1). Then, all Actors and Actions are extracted by analyzing the grammatical
relations. To overcome the problem of example sentences mentioned earlier (Issue
3.2) the actions are also filtered. This filtering method simply checks whether the
sentence contains a word of a stop word list called example indicators . Then, we
extract all objects from the phrase and each Action is combined with each Object.
The same is done with all Actors. This procedure is necessary as an Action
is supposed to be atomic according to the BPMN specification [8] and Issue
2.1. Therefore, a new Action has to be created for each piece of information as
illustrated in the following example sentences. In each sentence the conjunction
relation which causes the extraction of several Actors, Actions or Resources is
highlighted. As a last step, all extracted Actions are added to the World Model.

◦ “Likewise the old supplier creates and sends the final billing to the cus-

tomer.” (Action)

◦ “It is given either by a sales representative or by a pre-sales employee

in case of a more technical presentation.” (Actor)

◦ “At this point, the Assistant Registry Manager puts the receipt and

copied documents into an envelope and posts it to the party.” (Resource)

3.2 Text Level Analysis

This section describes the text level analysis. It analyzes the sentences taking
their relationships into account. The structural overview of this phase is shown
in Figure 3. We use the Stanford Parser and WordNet here, and also an anaphora
resolution algorithm. During each of the five steps, the Actions previously added
to the World Model are augmented with additional information.
An important part of the algorithm presented here is the determination heuris-
tic for resolving relative references within the text (Issue 4.1). Existing libraries
are not seamlessly integrateable with the output provided by the Stanford Parser.
Therefore, we implemented a simple anaphora resolution technique for the reso-
lution of determiner and pronouns. This procedure is described in detail in [25].
An experimental evaluation using our test data set showed that this approach
achieved a good accuracy of 63.06%.
The second step in our analysis is the detection of conditional markers. These
markers can either be a single word like “if”, “then”, “meanwhile” or “other-
wise”, or a short phrase like “in the meantime” or “in parallel”. All of these