Process Extraction from Texts using Semantic Unification
Konstantin Sokolov
1,2
, Dimitri Timofeev
3
and Alexander Samochadin
3
1
Mobile Device Management Systems Lab, Peter the Great St. Petersburg Polytechnic University, St. Petersburg, Russia
2
Department of Mathematical Linguistics, St. Petersburg State University, St. Petersburg, Russia
3
Institute of Computer Science and Technology, Peter the Great St. Petersburg Polytechnic University, St. Petersburg, Russia
Keywords:
Business Process Management, Process Extraction, Text Mining, Natural Language Processing.
Abstract:
In order to extract a process model from natural language sources, process elements such as agents, resources
or actions have to be identified across one or multiple texts. Natural language descriptions tend to provide
only partial and potentially contradictory information. Inability to systematically reconcile partial information
significantly limits capabilities of current systems. We propose to address this problem by means of semantic
unification integrated in the common text processing pipeline.
1 INTRODUCTION
Providing a formal specification of a business process
requires substantial work and high qualification from
the analyst. While the role of human experts in analy-
sis and modeling of business processes is crucial, their
activity may be substantially automated. It is espe-
cially important when business process modeling is
performed in the enterprise environment. Large or-
ganizations have lots of processes but, typically, the
processes are already specified textually in the form
of instructions, user guides, manuals, and standards.
Using text documents as a source of process models
lowers the cost of business process management. It
can also give new insights into the existing processes
by introducing new sources of knowledge, such as
customers claims, reviews, user stories, or other user-
generated content on corporate websites and in social
media.
Our interest in applying process extraction tech-
niques to natural language texts is motivated by the
task of development of process-oriented mobile ser-
vices. This class of services includes:
specialized step-by-step guides that inform users
on actions that should be performed to achieve a
goal (e.g., to obtain a visa, or to apply to the uni-
versity);
enterprise-level business process supporting soft-
ware;
mobile assistants that use procedural knowledge
to answer questions and schedule events.
This research is part of a project aiming to design
and implement a service platform based on a mobile
device management system. The platform provides
a set of process-oriented services, which require de-
velopment of process models. Automated process ex-
traction can greatly simplify service deployment, es-
pecially in organizations like universities, where most
processes are specified in normative documents, but
formal process models are rarely available. Besides
that, users of process-oriented mobile services may
benefit from the ability to share their knowledge of
processes and to use these community contributions.
In this position paper we discuss the limitations
of existing text processing pipelines used for process
extraction, especially with respect to handling multi-
ple documents, and propose an approach to systemati-
cally overcome these limitations. The rest of the paper
is structured as follows. In Section 2 existing methods
of process extraction from text are reviewed. In Sec-
tion 3 we discuss the limitations of current approaches
and propose a way to overcome them by using seman-
tic unification. Section 4 describes the core ideas be-
hind our method. Section 5 discusses how semantic
unification fits into the text processing pipeline, and
provides an example. The section 6 concludes.
2 RELATED WORK
A method of building a process model from a set of
unstructured documents should provide solutions to
the following problems: how to decide what entities
254
Sokolov, K., Timofeev, D. and Samochadin, A..
Process Extraction from Texts using Semantic Unification.
In Proceedings of the 7th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2015) - Volume 3: KMIS, pages 254-259
ISBN: 978-989-758-158-8
Copyright
c
2015 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
should go into the model, how superficially unrelated
data dispersed in the texts should be combined, and
how to obtain a formal model of a process.
The CREWS model (Achour, 1998) is often used
to formulate a conceptual description of a process.
The model specifies two kinds of building blocks, ob-
jects (agents or resources) and actions (atomic actions
or action flows). In a text objects are generally repre-
sented by noun phrases, and actions by verb phrases.
Extraction of entities and relations is a standard topic
of natural language processing, therefore it is rather
straightforward to integrate both technologies in or-
der to match a process with a text that describes it. In
the R-BPD framework prototype (Ghose et al., 2007),
constituents of a process model are extracted with
a combination of template matching, part-of-speech
tagging, and phrase chunking. Other systems take ad-
vantage of more sophisticated natural language pro-
cessing techniques. The use case analysis engine by
Sinha et al. (2008) implements a pipeline that in-
volves lexical processing, parsing, dictionary-based
concepts annotation, anaphora resolution, context an-
notation, and building of process models. A similar
method is implemented by Gonc¸alves et al. (2009,
2011). Both rely on problem-oriented grammars for
shallow parsing. Friedrich et al. (2011) use a wide-
coverage Stanford parser (Manning, 2003), and utilize
WordNet (Miller, 1995) and FrameNet (Baker et al.,
1998) databases to obtain information about semantic
relations, including synonymy. Ackermann and Volz
(2013) describe a prototype system based on recur-
sively defined templates which is able to extract do-
main models from text. They adopt feature structures
to model both the templates and the results of natu-
ral language analysis, and apply a variant of unifica-
tion algorithm to perform the matching between the
two. This approach allows to map text fragments to
elements of an intended domain model and to inter-
actively modify templates based on user feedback. In
general, their system is pattern-based and uses unifi-
cation techniques for comparison of tree structures of
syntactic nature.
3 LIMITATIONS
It should be obvious from the above, that the ability
of such systems to cope with intricacies of natural lan-
guage is limited by the tools that are used to produce
representations of entities and events that comprise a
process. Common approaches to extraction of a pro-
cess model from a document rely on a typical text
processing pipeline, which includes text normaliza-
tion, segmentation, tokenization, morphological anal-
ysis, named entity recognition, and syntactic parsing.
Some systems additionally perform semantic parsing
to produce trees labeled with semantic roles of the
constituents. In order to handle pronouns an anaphora
resolution step may be added to the pipeline. All sys-
tems process each sentence independently.
A serious limitation of current approaches is their
inability to consistently handle multiple descriptions
of the same process. Process descriptions supplied by
users contain errors due to the lack of proof-reading
and tend to describe only those parts of processes that
their authors are aware of or interested in. Documents
by different authors can be inconsistent with respect
to their vocabulary. As a result, these systems are gen-
erally designed to extract a model of a single process
from a single document. Currently, processing multi-
ple sources requires generation of models for each of
the documents with a subsequent merging by a human
expert, which should be automated. The common vo-
cabulary is even more important in an enterprise en-
vironment, where other knowledge-based systems are
available. The process extraction system should be
able to align extracted entities with existing ontolo-
gies.
To create a formal model of a process, a set of
hand-crafted rules is used by each system. This ap-
proach is perfectly valid for building models with a
known structure. However, the facts that do not fit
into the structure are ignored, making such systems
less convenient for exploratory analysis, and restrict-
ing the number of texts that can be processed. Further,
a known limitation of systems that use explicit rules
is the difficulty of rule modification. It is desirable to
have simpler rules that could be learned from corpora,
and to be able to combine them with higher-level pro-
cess analysis algorithms.
In sum, to overcome the limitations of the current
approaches it should be made possible to automati-
cally construct entity and event representations that
could easily integrate in the existing natural language
processing pipelines, support easy alignment with ex-
ternal knowledge bases and ontologies, be expressive
enough to allow for flexible data integration across the
limits of a sentence, and be formally compatible with
existing frameworks of process extraction. To achieve
this we propose a method of semantic unification for
entity resolution and data reconciliation based on fea-
ture structures. To build a process model, we adopt
the methods of process mining (van der Aalst, 2011),
where the text is translated into a sequence of ordered
events, and a process mining algorithm is invoked to
infer a model that fits this sequence.
Process Extraction from Texts using Semantic Unification
255
4 SEMANTIC UNIFICATION
4.1 Feature Structures
The theoretical background of using feature structures
for semantic representation is the situation semantics
of Barwise and Perry (1983). This approach requires
that some extralinguistic knowledge be included in
the context of an utterance for it to be coherent and
non-contradictory. This follows from the fact that in-
terpretation is impossible without relating to contex-
tual information, like time and place of an utterance,
propositional attitude of the speaker, etc. Instead of
linguistic production it is rather the context of obser-
vation of linguistic performance that has to be taken
into account during interpretation. This approach dif-
fers from the traditional model-theoretic view of for-
mal semantics in that the meaning is not related to the
whole information about the world, but instead it is
understood as a structured representation, which only
partially describes the world and consequently is pro-
vided with a partial interpretation. The situation is
thus only a fragment of the model structure, which
should be merged with other such fragments in the
presence of extralinguistic data. Following Pollard
and Sag (1987), the formal description of situations
is given in the form of feature structures, for which an
operation of unification is defined, where unification
serves as a means of obtaining a full description from
partial ones by combining linguistic information and
extralinguistic knowledge as long as it is possible to
do so without falling into contradiction.
Feature structures is a well-known formalism
which has been applied to knowledge representa-
tion and formulation of various grammatical theories
based on unification, including Head-driven Phrase
Structure Grammar (Pollard and Sag, 1994) and Lexi-
cal Functional Grammar (Kaplan and Bresnan, 1982).
Feature structures are recursively defined and can sup-
port internal references. They can be represented
in multiple forms, the traditional attribute-value ma-
trix (AVM) notation of computational linguistics and
the direct rendering of an underlying graph struc-
ture, which is a natural representation for conceptual
graphs, RDF, etc. A number of extensions is com-
monly used, including support for disjunctive features
and typing constraints. Standard definitions of typed
feature structures (TFS) can be found in (Copestake,
2000), and the more extended treatment of formal as-
pects of the underlying theory in (Carpenter, 2005)
and (Francez and Wintner, 2011). Here we limit our-
selves to a few examples.
The following AVM represents an object of type
(or sort) noun, which has two attributes, a number and
an orthography.
noun
NUMBER sg
ORTHOGRAPHY foot
The other AVM represents another possible de-
scription which differs from the first one in that it has
a different set of attributes.
noun
CASE nom
ORTHOGRAPHY foot
Note, however, that the values of common at-
tributes are the same, so the information contained
in both descriptions can be combined (or unified) to
result in a single description.
noun
CASE nom
NUMBER sg
ORTHOGRAPHY foot
The unification process is restricted by the types
that are assigned to particular feature structures. The
types form a pre-order, so that a more specific type
can license combination of a number of more general
partial descriptions of subsuming types. Thus a set of
typed feature structures resembles a system of objects
provided with a mechanism of multiple inheritance.
Feature structures can be regarded as abstract
entities and be given a formal logical description,
which is similar to the knowledge representation ap-
proach of Semantic Web. The earliest proposals along
these lines are the Kasper-Rounds logic (Kasper and
Rounds, 1986), which is close to the constraint-based
system PATR-II, and an approach of Gazdar et al.
(1988). Johnson (1991a,b) uses a decidable frag-
ment of first order logic (Bernays–Sch
¨
onfinkel class),
whereas Blackburn (1993) builds a formalism based
on hybrid logic, an extension of modal logic. The
latter approach was used by Baldridge and Krui-
jff (2002) for Hybrid Logic Dependency Semantics
(HLDS), a semantic representation formalism used in
the OpenCCG
1
system. A connection with descrip-
tion logics is also well established. Blackburn (1993)
notices that both feature logics and description log-
ics are in essence modal logics with superficial differ-
ences of the way they are formulated. What is com-
mon to all these approaches is the property of decid-
ability, that the authors of corresponding logics strive
to maintain.
1
http://openccg.sf.net
KMIS 2015 - 7th International Conference on Knowledge Management and Information Sharing
256
4.2 Semantic Unification
Feature structures can serve as a basic formalism in
the entity resolution task for the following reasons.
The type pre-order of TFS provides the means
of relating feature-based descriptions to an ontology.
Krieger and Sch
¨
afer (2010) describe a method of
translating an OWL ontology into TFS and vice versa.
There an additional knowledge is used to improve re-
call and precision of named entity recognition algo-
rithms. In general, domain knowledge can be used
in the process of text analysis, whereas the results of
entity resolution can be exported to an ontology.
Feature structures allow for incremental construc-
tion of a full description of an object based on its
partial descriptions. Semantic representations based
on feature structures have been used in many ap-
proaches in computational linguistics, including Min-
imal Recursion Semantics (Copestake et al., 2005),
formal representations of lexical semantics in Gen-
erative Lexicon (Pustejovsky, 1991), Hybrid Logic
Dependency Semantics (Baldridge and Kruijff, 2002)
and others. In these approaches the process of unifi-
cation is used to model semantic composition.
A serious limitation of the symbolic unification
approach of the aforementioned proposals is its in-
ability to account for the variations of values present
in the corresponding semantic representations. Mi-
nor modifications of spelling may result in failure
of determining whether multiple realizations of the
same attribute correspond to the same value and
thus block the unification process. Analogously, in
cases of inter- and intrasentential anaphora pronouns
should be linked to the expressions they refer to in
other segments of text, which is problematic to do at
the purely symbolic level. Recently Abramsky and
Sadrzadeh (2014) adopted an approach similar to ours
for anaphora resolution using discourse representa-
tion structures as their semantic formalism. An im-
portant aspect of their work is that instead of main-
taining global coherence of semantic structures they
propose to consider semantic coherence locally, mod-
eling it with sheaf theory, where data integration is
regarded as gluing of semantic representations. This
process generally results in multiple variants of data
combination, each of which can be assigned a prob-
ability based on its realization in a corpus. By intro-
ducing an empirically sound quantitative measure of
appropriateness of particular way of combining par-
tial semantic information it is possible to alleviate the
problems of purely symbolic unification.
An apparent discrepancy between the traditional
feature structure unification and the semantic unifi-
cation approach based on sheaf theory can be made
clear by analogy with the general unification theory
(Baader and Snyder, 2001), where a distinction is
made between purely syntactic and equational unifi-
cation. Whereas the former relies on unconditioned
term substitution, the latter requires that the possibil-
ity of term substitution be controlled by certain ex-
ternal conditions, thus resulting in unification modulo
an equational theory. The semantic unification pro-
cess as applied to entity resolution can be viewed as
an extension of the traditional symbolic unification of
feature structures augmented with a form of empiri-
cally motivated corpus-based equational theory.
5 PROCESS EXTRACTION
To build a description of a business process, we ex-
tract from each text a set of relations that represent
actions and events. In the simplest case these rela-
tions have the form hNP, VPi, where the noun phrase
NP correspond to the agent, and the verb phrase VP
describes the action. More complex relations include
other semantic roles such as theme or goal. For exam-
ple, an argument structure for the sentence “The seller
gives the invoice to the customer” can be described as
follows.
ACTION give
AGENT seller
THEME invoice
RECEPIENT customer
Technically, attribute values are not plain text
segments, but rather references to feature structures
attached to corresponding segments of parsed text.
These structures may contain both semantic and syn-
tactic attributes.
The model is inferred from a log, a sequence of
actions grouped by case. Process logs are built from
the relations extracted during the previous step. To
do so, we take all relations that describe actions, and
sort them by time under the heuristic that the corre-
sponding actions are performed in the order of their
mention in the text, which is a plausible assumption
for descriptions of processes. Explicit time attributes
may change this order according to specified rules.
We use these attributes to generate a set of cases. If
all actions are mandatory, and there are no alterna-
tives, a single linear case is generated. Each optional
action doubles the quantity of the cases, so that one
half of the resulting cases does not contain the action
as it never occurs. Alternatives are handled similarly.
The model may be inferred from the resulting set of
cases using process mining algorithms (currently we
Process Extraction from Texts using Semantic Unification
257
Figure 1: A graph of the inferred process.
use the ProM tool
2
).
As an example, consider a paragraph from a guide
for obtaining the United States visa.
When the visa is approved, you may pay a visa
issuance fee if applicable to your nationality, and will
be informed how your passport with visa will be re-
turned to you. Review the visa processing time, to
learn how soon your passport with visa will generally
be ready for pick-up or delivery by the courier.
Suppose the following actions have been ex-
tracted.
1. Visa is approved (mandatory).
2. Pay a visa issuance fee (optional).
3. Applicant is informed (mandatory).
4. Learn how soon the passport is ready (manda-
tory).
5. Pick up the passport (alternative).
6. Passport is delivered by the courier (alternative).
There are one optional action and one action with
two alternatives, so we get four process cases: (1, 2,
3, 4, 5), (1, 2, 3, 4, 6), (1, 3, 4, 5), and (1, 3, 4, 6).
The inferred process is presented in Figure 1.
6 CONCLUSION
In this paper we described an approach of extraction
of process models from texts. Currently we consider
English texts only. The method itself is not limited
to one language, but linguistic resources, especially
2
http://www.promtools.org
grammars and lexical databases, are required to sup-
port other languages. Our main goal is to handle mul-
tiple documents with partial descriptions of a process.
We propose to add to the processing pipeline an entity
resolution step implemented as the semantic unifica-
tion of typed feature structures. This approach also
allows for integration of text processing algorithms
with domain knowledge represented as ontologies.
The work presented is in the early stage of devel-
opment, and we are currently working on a prototype
and on the evaluation of our method.
ACKNOWLEDGEMENTS
This research is a part of the joint project by IBS
(Moscow, Russia) and Peter the Great St. Petersburg
Polytechnic University (St. Petersburg, Russia). This
work is financially supported by the Ministry of Ed-
ucation and Science of the Russian Federation (state
contract 02.G25.31.0024 from 12.02.2013).
REFERENCES
Abramsky, S. and Sadrzadeh, M. (2014). Semantic unifi-
cation. In Categories and Types in Logic, Language,
and Physics, pages 1–13. Springer.
Achour, C. B. (1998). Guiding scenario authoring. In
European-Japanese Conference on Information Mod-
elling and Knowledge Bases, page 1.
Ackermann, L. and Volz, B. (2013). model [nl] generation:
natural language model extraction. In Proceedings of
the 2013 ACM workshop on Domain-specific model-
ing, pages 45–50. ACM.
Baader, F. and Snyder, W. (2001). Unification theory. Hand-
book of automated reasoning, 1:445–532.
Baker, C. F., Fillmore, C. J., and Lowe, J. B. (1998). The
Berkeley FrameNet project. In COLING-ACL ’98:
Proceedings of the Conference, pages 86–90, Mon-
treal, Canada.
Baldridge, J. and Kruijff, G.-J. M. (2002). Coupling CCG
and Hybrid Logic Dependency Semantics. In Pro-
ceedings of the 40th Annual Meeting of the Associ-
ation for Computational Linguistics (ACL), Philadel-
phia, July 2002, pages 319–326.
Barwise, J. and Perry, J. (1983). Situations and attitudes.
Blackburn, P. (1993). Modal logic and attribute value struc-
tures. Springer.
Carpenter, B. (2005). The logic of typed feature structures:
with applications to unification grammars, logic pro-
grams and constraint resolution, volume 32. Cam-
bridge University Press.
Copestake, A. (2000). Appendix: Definitions of typed
feature structures. Natural Language Engineering,
6(01):109–112.
KMIS 2015 - 7th International Conference on Knowledge Management and Information Sharing
258
Copestake, A., Flickinger, D., Pollard, C., and Sag, I. A.
(2005). Minimal recursion semantics: An introduc-
tion. Research on Language and Computation, 3(2-
3):281–332.
Francez, N. and Wintner, S. (2011). Unification grammars.
Cambridge University Press.
Friedrich, F., Mendling, J., and Puhlmann, F. (2011). Pro-
cess model generation from natural language text.
In Advanced Information Systems Engineering, pages
482–496. Springer.
Gazdar, G., Pullum, G. K., Carpenter, R., Klein, E., Hukari,
T. E., and Levine, R. D. (1988). Category structures.
Computational Linguistics, 14(1):1–19.
Ghose, A., Koliadis, G., and Chueng, A. (2007). Process
discovery from model and text artefacts. In Services,
2007 IEEE Congress on, pages 167–174. IEEE.
Gonc¸alves, J. C. A. R., Santoro, F. M., and Bai
˜
ao, F. A.
(2009). Business process mining from group stories.
In Proceedings of the 2009 13th International Con-
ference on Computer Supported Cooperative Work in
Design.
Gonc¸alves, J. C. A. R., Santoro, F. M., and Bai
˜
ao, F. A.
(2011). Let me tell you a story On how to build
process models. Journal of Universal Computer Sci-
ence, 17(2):276–295.
Johnson, M. (1991a). Features and formulae. Computa-
tional Linguistics, 17(2):131–151.
Johnson, M. (1991b). Logic and feature structures. In IJ-
CAI, pages 992–996.
Kaplan, R. M. and Bresnan, J. (1982). Lexical-functional
grammar: A formal system for grammatical represen-
tation. Formal Issues in Lexical-Functional Grammar,
pages 29–130.
Kasper, R. T. and Rounds, W. C. (1986). A logical se-
mantics for feature structures. In Proceedings of
the 24th annual meeting on Association for Compu-
tational Linguistics, pages 257–266. Association for
Computational Linguistics.
Krieger, H.-U. and Sch
¨
afer, U. (2010). DL meet FL: a bidi-
rectional mapping between ontologies and linguistic
knowledge. In Proceedings of the 23rd International
Conference on Computational Linguistics: Posters,
pages 588–596. Association for Computational Lin-
guistics.
Manning, D. K. C. D. (2003). Natural language parsing. In
Advances in Neural Information Processing Systems
15: Proceedings of the 2002 Conference, volume 15,
page 3. MIT Press.
Miller, G. A. (1995). Wordnet: a lexical database for en-
glish. Communications of the ACM, 38(11):39–41.
Pollard, C. and Sag, I. (1987). Information-based syntax
and semantics, vol. 1.
Pollard, C. and Sag, I. A. (1994). Head-driven phrase struc-
ture grammar. University of Chicago Press.
Pustejovsky, J. (1991). The generative lexicon. Computa-
tional linguistics, 17(4):409–441.
Sinha, A., Paradkar, A., Kumanan, P., and Boguraev, B.
(2008). An analysis engine for dependable elicitation
on natural language use case description and its appli-
cation to industrial use cases. IBM Research Report
RC24712 (W0812-106).
van der Aalst, W. M. P. (2011). Process Mining, Discov-
ery, Conformance and Enhancement of Business Pro-
cesses. Springer-Verlag, Berling Heidelberg.
Process Extraction from Texts using Semantic Unification
259