Process Model Generation from Natural Language Text - Advanced Information Systems Engineering - page 483

There are also tools available for semantic analysis. They provide semantic
relations on different levels of detail. We use FrameNet [13] and the lexical
database WordNet[14]. WordNet provides various links to synonyms, homonyms,
and hypernyms for a particular class of meaning associated with a synonym-set.
FrameNet defines semantic relations that are expected for specific words. These
relations are useful, e.g., to recognize that a verb “send” would usually go with
a particular object being sent. Syntax parsers and semantic analysis are used in
our transformation approach, augmented with anaphora resolution.

3 Transformation Approach

The most important issue we are facing when trying to build a system for gener-
ating models is the complexity of natural language. We collected issues related
to the structure of natural language texts from the scientific literature and an-
alyzed the test data, which is described in section 4. Thereby, we were able to
identify four broad categories of issues which we have to solve in order to ana-
lyze natural language process descriptions successfully (see Table 1). Syntactic
Leeway relates to the fact that there is a mismatch between the semantic and
syntactic layer of a text. Atomicity deals with the question of how to construct
a proper phrase-activity mapping. Relevance has to check whether parts of the
text might be irrelevant for the generated process model. Finally, Referencing
addresses the question of how to resolve relative references between words and
between sentences.

Table 1. References in the literature to the analyzed issues

Issue Refs. Issue Refs.

1 Syntactic Leeway 3 Relevance

1.1 Active-Passive [15] 3.1 Relative Clause Importance [16]

1.2 Rewording/Order [17,18] 3.2 Example Sentences [19]

1.3 Implicit Conditions [20,21] 3.3 Meta-Sentences [16]

2Atomicity 4 Referencing

2.1 Complex Sentences [16,18] 4.1 Anaphora [22,23]

2.2 Action Split over Sentences [22] 4.2 Textual Links [24]

2.3 Relative Clauses [16] 4.3 End-of-block Recognition [19,16]

Different solution strategies were applied in the works listed in Table 1 to
overcome the stated problems, e.g. by constricting the format of the textual input
[17], but no study considers all mentioned problems and offers a comprehensive
solution strategy. Another interesting fact is that none of the works using a
shallow parser shows how they deal with passive voice [15,22,23,17]. We solved
this problem by using the grammatical relations of the Stanford Parser.
To obtain a structured representation of the knowledge we extract from the
text, we decided to store it in a World Model , as opposed to a direct straight