Knowledge-based configuration has been used for numerous applications including natural language processing (NLP). By formalising property grammars as a configuration problem, it has been shown that configuration can provide a flexible, nondeterministic, method of parsing natural language. However, it focuses only on syntactic parsing. In contrast, configuration is usually performed using knowledge about a domain and is semantic in nature. Therefore, we argue that configuration has the potential to be used, not only for syntactic processing, but for the semantic processing of natural language, effectively supporting Natural Language Understanding (NLU). In this paper, we propose an approach to NLP that applies configuration to the (partial) domain model evoked by the processing of a sentence. This has the benefit of ensuring the meaning of the sentence is consistent with the existing domain knowledge. Moreover, it allows the dynamic incorporation of domain knowledge in the configuration model as the text is processed. We demonstrate the approach on a business specification based on the Semantics of Business Vocabulary and Rules.
Knowledge-based configuration has been used in numerous applications. While historically used for configuring physical products, configuration has been applied to other domains such as software services, software product lines, and constraint-based language parsing [Hotz and Wolter, 2013].
In particular, [Estratat and Henocque, 2004; Kleiner et al., 2009] have applied configuration to a translation of property grammars (a constraint-based linguistic formalism). By formalising property grammars as a configuration problem, they show that configuration can provide a flexible, nondeterministic method for parsing natural language. However, these approaches focus on syntactic parsing by using the configuration process to generate a parse tree. In contrast, configuration is usually applied to domain knowledge, that is, semantic processing. Furthermore, in [Kleiner et al., 2009] additional processes are required in order to transform the parse tree into a semantic model to make use of the domain knowledge. This causes some issues in ensuring the consistency and correctness of the domain knowledge.
In this paper we present an approach to parsing natural language that performs semantic processing directly using configuration. Instead of a model of language categories (e.g. noun, verb, noun phrase) and the properties (or constraints) on those categories, such as property grammars, we use a model of domain concepts and the relations between them. As a result, we perform natural language understanding; at least with respect to the semantic model being used.
Our approach maintains the advantages of using configuration for natural language processing, while gaining the following: (1) a simplified lexicon containing minimal lexical information, (2) improved consistency of the domain knowledge as the configuration process ensures its consistency during parsing, and (3) the semantic disambiguation of terms.
As our aim is to support the translation of informal natural language business specifications into formal models, we demonstrate our approach on an example from the business domain. The example business specification is defined using the Semantics of Business Vocabulary and Business Rules (SBVR) [OMG, 2008], which we use as our semantic model. SBVR and the example are discussed in more detail later.
The remainder of this paper is organised as follows: Section 1.1 provides a brief introduction to SBVR and its concepts, Section 2 presents an example that will be used throughout the paper, Section 3 describes our approach to parsing natural language, Section 4 presents experimental results, Section 5 discusses related work, and Section 6 provides insight into future work and concludes the paper. 1.1
The Semantics of Business Vocabulary and Business Rules (SBVR) is a standard developed by the Object Management Group (OMG) to facilitate the transfer of business knowledge between business people and the technical experts responsible for developing information systems for them [OMG, 2008]. It encompasses two aspects: (1) a meta-model for representing vocabularies, facts, and rules in a machine processable format, and (2) a controlled English notation for representing the same vocabularies, facts, and rules more suited to people. Therefore, SBVR supports the exchange of business specifications between both organisations and software tools.
The SBVR meta-model standardises concepts for the definition of business vocabularies (i.e. sets of concepts relevant to a particular organisation or business domain) and rules relating to those vocabularies. It is based on formal logic; primarily first-order predicate logic with an extension in modal logic (necessity, possibility, permissibility, and obligation).
Within the meta-model, vocabularies are defined on the basis of Meanings and Representations. They consist of sets of interrelated object types, individual concepts, and fact types. A distinction is made between a meaning and its representation, allowing a single concept to be represented with multiple words (possibly in different languages), images, or sounds.
The semantic structure of business rules are formed by the Logical Formulations aspect of the meta-model. This includes concepts for first-order logical operators (e.g. conjunction, disjunction, implication), quantification (e.g. universal, existential, exactly n), and modal operators (e.g. necessity, obligation). These concepts allow business people to define structural and operative rules. Structural rules include such rules as cardinality constraints on the relations between concepts and cannot be violated, while operative rules may be violated by a person involved in conducting the business. 2
This section introduces an example that identifies the limitations of existing approaches and that will be used in the remainder of this paper to demonstrate our approach. It is an extract of the EU-Rent business specification included in the SBVR specification[OMG, 2008, Annex E].
EU-Rent is a fictional car rental company with a global presence. The example business specification defines domain specific vocabulary and rules for EU-Rent, its structure, and how it conducts its business. Figure 1 shows a portion of the vocabulary related to the organisational structure of EU-Rent. The following is a structural rule, based on this vocabulary, that defines a cardinality constraint on the part-of relationship between a ‘branch’ and a ‘local area’. (1) Each branch is included in exactly one local area. rental organisation unit
Definition: organisational unit that operates part of EU-Rent’s car rental business rental organisation unit having rental responsibility
Definition: t.h.e. rental organisation unit is responsible for the operation of customer-facing rental business rental organisation unit having area responsibility
Definition: t.h.e. rental organisation unit includes organisation units for which it has the responsibility to coordinate operations and ensure resources local area
Definition: rental organisation unit .t h..a.t has area responsibility branch
Definition:rental organisation unit t.h.a.t. has rental responsibility branch is included in local area
Synonymous Form: local area includes branch
Although quite simple, this example demonstrates a number of important concepts such as nouns, verbs, and quantifiers. The approach of [Kleiner et al., 2009] processes the sentence by executing a series of model transformations that result in a UML model of the text, using SBVR as an intermediate model between the natural language text and UML. The steps up to the creation of the SBVR model are as follows: 1. a text-to-model transformation creates an Ordered Words model that annotates the words with their position in the sentence 2. a model-to-model transformation creates a Labelled Words model by labelling each word with their possible syntactic categories using a lexicon (model) 3. configuration is used to transform the Labelled Words model into a Syntax model, performing syntactic and grammatical analyses, and 4. a model-to-model transformation creates an SBVR model from the Syntax model
Performing this process on the example sentence would result in the Syntax and SBVR models displayed in Figure 2.
This approach provides a flexible, non-deterministic, and extensible method of parsing natural language [Kleiner et al., 2009]; however, it has several issues, chiefly: (1) it is primarily syntactic, (2) it requires a detailed lexicon, and (3) the mapping to SBVR can be problematic, e.g. in the handling of ‘local area’ two correct interpretations can be conceived.
The first is the main issue as, although [Estratat and
Henocque, 2004] suggest that configuration can combine
syntactic and semantic analysis, it is primarily used for
generating syntactic parse trees. As a result, a sentence could be
syntactically correct but not meaningful and, therefore, must be
linked to the semantics somehow
Although [Kleiner et al., 2009] introduce some semantic elements into their model (i.e. each category can be linked to a basic element of the SBVR model1), they do so only to ease the transformation to the SBVR model. The existence of the SBVR elements does not provide any semantic guarantees. Therefore, semantic inconsistencies need to be resolved either during the transformation to SBVR, making it much more complex, or by post-processing of the SBVR model. However, this seems unnecessary if it can be achieved during the configuration process itself.
The requirement of a detailed lexicon is more an issue for our target application area than with the parsing method itself. In the context of businesses creating and maintaining their own sets of domain specific business vocabularies and business rules, we do not see business people defining detailed lexicons with linguistic information such as voice, genre, transitivity, etc. In this application area, the business vocabulary is more like a glossary containing domain specific words and their definitions, like that of Figure 1, rather than a dictionary containing detailed lexical information. Therefore, an approach that requires less linguistic information to be defined is required for our purposes.
Finally, there are some problems with mapping the syntactic tree to the SBVR semantics. Consider how the term ‘local area’ is handled in the above example. For simplicity, ‘local area’ is a single noun, which maps directly to the object type ‘local area’ in SBVR. However, in reality it would be considered a noun phrase, where the term ‘local’ would be used in an adjective sense and ‘area’ would be the noun. This would map to the object type ‘area’ with the characteristic ‘being local’. However, in the vocabulary of EU-Rent, ‘local area’ is a single concept and should not be decomposed in this way.
It seems a simple problem to fix: the term ‘local area’ could be included as a noun in the lexicon, which is what would happen if a business were defining its vocabulary. However, unless the individual terms ‘local’ and ‘area’ were removed, which could affect the processing of sentences in other contexts, it would result in two correct parses of the sentence: one in which ‘local area’ is treated as a simple noun and one treating it as a noun phrase. The transformation to the SBVR model would not resolve this issue either, as both forms have a valid mapping. Therefore, the user would have to select the preferred mapping or the SBVR model would have to be processed to see if either one or the other has already been created. Either way it makes the process more cumbersome, whereas we propose that by configuring the SBVR model directly, this problem would be avoided (as long as only one or the other has been specified in the vocabulary).
It could be argued that this problem is a quirk of the SBVR model as other semantic representations with a structure more similar to the parse tree would have more direct mappings. However, it is an important issue as SBVR has gained traction in our application domain in recent years [Nemuraite et al., 2010; Sukys et al., 2012] and is an important part of the OMG’s Model-Driven Architecture [OMG, 2008]. Moreover, we will show that our proposed approach does not reduce the flexibility with respect to the semantic representation used. 3
In order to overcome these limitations we propose the use of knowledge-based configuration on the semantic representation directly, rather than configuring a syntactical parse tree. In this way we maintain the benefits of parsing using configuration, while ensuring semantic consistency and removing a step from the process. Furthermore, this approach remains agnostic with respect to the semantic representation used; although we utilise the SBVR meta-model in this paper.
In order to avoid complex syntactical analysis, which is a difficult problem in itself, we use an approach inspired by Cognitive Grammar, a theory of grammar from the field of Cognitive Linguistics [Langacker, 2008]. Cognitive Grammar takes a holistic view of language, combining syntax, semantics, and pragmatics into a unified whole. In particular, our approach is based on that of [Holmqvist, 1993], which provides a computational model of Cognitive Grammar.
In Cognitive Grammar, the meaning of an expression is understood by combining the semantic structures evoked by its constituent expression into a unified structure; a process called semantic accommodation in [Holmqvist, 1993]. Evoking the semantic structure of an expression is a relatively simple endeavour as the two are linked; therefore, a semantic structure is evoked by looking up the expression in a lexicon. As a result, our method is able to do away with traditional syntactic analysis for a much simpler model, leaving most of the effort in understanding the expression to be performed by the accommodation process. With the aim of combining semantic structures into a composite structure based on the allowable relationships between them, the accommodation process is analogous to a configuration task.
The syntactic analysis and the accommodation process using configuration are detailed in the following sections. These two processes are performed iteratively with the first using the expectations, or placeholders in our case, to propose possible parses of the sentence, and the second combining the semantic aspects of the suggested parses into a complete structure. The result is a progressively more detailed and complete set of domain knowledge containing the concepts, their definitions, and their associated rules. 3.1
The syntactic analysis of our approach is primarily the evocation of semantic structures from the lexicon, taking into account grammatical dependencies. For example, in the example expression, there are two quantifiers ‘each’ and ‘exactly one’. If only the evoked semantic structures were known, the configurator would not know which quantifier applies to what concept, yet we know that ‘each’ is supposed to apply to the term ‘branch’ and ‘exactly one’ to ‘local area’.
Traditionally, these dependencies are handled by the relationships between categories (as shown in Figure 2a). However, as this leads to complex lexicons that are not suitable for our target application and require complex processing to produce, we account for these dependencies using so called grammatical expectations [Holmqvist, 1993]. Grammatical expectations are a relatively simple model of grammatical dependency that identify locations in an expression where other expressions are “expected” to fill. Moreover, grammatical expectations are a good match for SBVR-based semantics, as they are similar to the placeholders of fact types. In [Holmqvist, 1993], only left and right expectations were introduced, which search to the left or right for another expression. We also introduced internal expectations, which search within the span of the expression, in order to more easily deal with SBVR fact types with more than two placeholders.
This model defines lexicon and lexical entries as follows.
l = (E, LE, lookup) where E is a set of expressions, LE is a set of lexical entries, and lookup : E → 2LE \∅ Definition 2 (Lexical Entry). A lexical entry is a tuple le = (e, ss, GE, typeGE ) where e is an expression of one or more words, ss is its associated semantic structure, GE is a set of grammatical expectations, and typeGE : GE → {lef t, internal, right} assigns a type to each grammatical expectation ge ∈ GE.
This definition is purposefully generic with respect to the form the semantic structure takes. While we use the SBVR meta-model, other semantic representations could be used. As a result, our approach remains flexible in terms of varying the model being configured, as in[Kleiner et al., 2009].
The lexicon is partly predefined. For example, words with explicit semantics in SBVR, such as those for quantifications, logical operators, etc., are explicitly defined as certain expressions, semantic structures, and grammatical expectations. Domain specific terms are provided by the vocabulary of a business specification, e.g. a glossary of terms, which have their semantic structures and grammatical expectations determined by the representation of object types and fact types in SBVR. In our application domain, the vocabulary is provided by business people, which drives our need for a simple lexicon with minimal information.
Using grammatical expectations, the syntactic analysis is
performed incrementally, when an expression is found to fill
an expectation
The suggestion list and suggestions are defined as follows. Definition 3 (Suggestion List). A suggestion list SL is an ordered set of suggestions, where each suggestion s ∈ SL is a tuple s = (SLE, C, be, lc, gc) such that: • SLE is a set of lexical entries or previous suggestions included in s • C is a set of tuples
catch = (sle1 ∈ SLE, ge, sle2 ∈ SLE, cd) that associate a grammatical expectation, ge, of the catching lexical entry or suggestion, sle1, to the lexical entry or suggestion that it catches, sle2, with the catching distance, cd. • be = Pc∈C c.cd is the binding energy of the suggestion, • lc is the local coverage of the suggestion, • gc is the global coverage of the suggestion The order for each s1, s2 ∈ SL is determined by: s1 s2 ⇐⇒ (s1.gc, s1.lc, s1.be) ≤l (s2.gc, s2.lc, s2.be) Based on the preferences that: the best parse should cover the entire sentence; suggestions should have no holes; and, words should be captured at the shortest distance.
Due to the recursive compositional nature of the suggestions, i.e. each x ∈ SLE is a lexical entry or another suggestion, the catching information constitutes a non-traditional parse tree. Figure 3 shows an example of a parse tree produced by our syntactic analysis compared to the traditional parse tree equivalent to Figure 2a. The parse tree and the partial SBVR model (i.e. the semantic structures) of the suggested parse are provided to the semantic accommodation process to be configured into a complete model.
The general algorithm for the syntactic analysis is as follows, for each word in the expression: 1. Retrieve the entry for the current word from the lexicon 2. If the current word has any left placeholders, for each suggested parse in the suggestion list: (a) (b) (a) catch the closest word (or word combination) to the left of the current word (b) add the new suggested parse to the suggestion list 3. Else, for each suggested parse in the suggestion list, if the previous word or combination has any internal placeholders and the current word is within its span: (a) catch the current word with the internal placeholder (b) add the newly suggested parse to the suggestion list 4. Else, for each suggested parse in the suggestion list, if the previous word or combination has any right placeholders: (a) catch the current word with the right placeholder (b) add the newly suggested parse to the suggestion list 5. Update the heuristics, distances between words, order the suggestion list, and cull excess entries 6. Provide newly suggested parses to the semantic accommodation/configuration process
An example of the syntactic analysis after a complete parse of the example sentence is displayed in Figure 4. 3.2
Parses suggested by the syntactic analysis are sent to the semantic accommodation process to determine whether or not they are admissible in the (SBVR) semantics. Rather than the numerous processes for accommodation discussed in [Holmqvist, 1993], we utilise knowledge-based configuration to perform the accommodation. Specifically, componentoriented configuration is used, which combines advantages from connection-, resource-, and structure-based approaches [Soininen et al., 1998; Stumptner, 1997]. Moreover, the object-oriented nature of component-oriented configuration lends itself more easily to that of the SBVR meta-model.
Using the terminology of [Soininen et al., 1998], the SBVR meta-model constitutes the configuration model knowledge of our approach, i.e. it defines the types of entities, properties, and rules that specify the set of correct configurations. It follows that an SBVR (terminal) model constitutes the configuration solution knowledge or (possibly partial) configuration . Lastly, the parse tree created by the lexical analysis constitutes the requirements knowledge, i.e. additional constraints on the configuration that are not strictly part of the configuration model. For example, the SBVR meta-model may allow either quantification to be applied to either object type in the example rule, however, the grammatical dependencies require that ‘each’ be applied to ‘branch’ and ‘exactly one’ to ‘local area’ for the correct interpretation of the sentence.
Since the SBVR (meta-)model is defined using ECore, the Eclipse Modelling Framework 2 implementation of EMOF from the MOF specification of the OMG [OMG, 2006], we first discuss its mapping to the configuration ontology of [Soininen et al., 1998]. The configuration ontology defines standard concepts for representing the different aspects of configuration knowledge including: taxonomy, attributes, structure, topology, and constraints. An example of the mapping for SBVR is displayed in Figure 5. It focuses on the configuration model knowledge, with some example component individuals. For simplicity, port individuals are not included in the figure. The mapping is by no means complete, but provides a link between the meta-model representation and the representation used for the configuration task.
ECore allows classification hierarchies to be defined through the use of the concepts EClass and EObject, and the relations eSuperTypes and eClass. The concept EClass generically represents a type and therefore could be mapped to Configuration Type; however, ECore does not distinguish the sub-types Component Type, Port Type, Resource Type, and Function Type in the same manner as [Soininen et al., 1998]. Therefore, it is more appropriate to map instances of EClass to Component Types. Other ECore concepts map to the other configuration types. Therefore, only component types have a classification hierarchy; the others are made direct subtypes of their appropriate configuration type (i.e.Attribute Type, etc.).
Sub-types and super-types in ECore are represented by eSuperTypes. This relation defines the direct super-types of an EClass and, therefore, maps to the isa relation. Moreover, multiple inheritance is allowed in both representations.
In ECore, an EClass may be abstract, i.e. cannot have any instances. However, in the context of configuration, it makes sense to relax this definition to allow partial information of a configuration, such as in [Soininen et al., 1998]. Therefore, abstract and non-abstract (i.e. concrete) types in ECore are mapped to abstract and concrete classes in the ontology using the appropriate Abstraction Definition.
Instances of EObject represent Individuals from the ontology. In ECore, these are associated to their type by eClass, which maps to is directly of. Moreover, since eClass specifies theEClass of an individual, EObject necessarily maps to Component Individual.
Attributes in ECore are represented by the concept EAttribute, which have a name, a type specified byeAttributeType, and relations for the lower and upper bounds of their cardinality (lowerBound and upperBound, respectively).
An eAttributeType links an EAttribute to its EDataType, which maps to the concept Attribute Type from [Soininen et al., 1998]. It follows that EAttributes map to Attribute Definitions with the appropriate Attribute Name, Attribute Value Type (from eAttributeType), and Necessity Definition. The value of an attribute for a specific EObject is mapped to an Attribute with the respective Attribute Value and Component Individual.
In ECore, attributes can have a zero-to-many cardinality, while Necessity Definitions are restricted to exactly one (necessary) and at most one (optional) attributes. As a result, there exists only a partial mapping to the configuration ontology; however, this is not a problem for the SBVR metamodel as it only includes necessary and optional attributes.
The ontology of [Soininen et al., 1998] differentiates between Part Definitions, which specify the compositional structure of components, and Port Definitions, which specify the topological connections (either physical or logical) between components. Part Definitions constitute a direct has-part relation between components. This relation must be anti-symmetric and anti-reflexive. Moreover, the transitive closure of has-part defines a transitive haspart relation, which must also be anti-symmetric and antireflexive. Port Definitions have no such restriction.
In ECore, both Part Definitions and Port Definitions are represented by the concept EReference. An EReference may be a containment reference (for compositional relationships) and/or it may have an eOpposite for bi-directional relationships.
While it seems intuitive to map containment and unidirectional references to part definitions, and bi-directional relationships to port definitions, this is not possible as ECore does not uphold the anti-symmetric and anti-reflexive requirements of has-part relations. For example, the situation shown in Figure 6, in which the transitive closure of the has-part relations between Meaning, Representation, and Expression are reflexive, is allowable in ECore but not the configuration ontology. To determine those EReferences that could be mapped to part definitions would require analysis of the metamodel; instead, we map all EReferences to ports. As a result, we effectively use ports as a generalised structural relationship similar to that described in [Hotz and Wolter, 2013].
A Port Definition requires a Port Name, a Possible port type set, and a Cardinality. Similar to EAttribute, EReferences have a name, a lowerBound, and an upperBound. Therefore, an EReference maps to a Port Definition, with the name mapping to the Port Name, and the lowerBound and upperBound are mapped to the Cardinality. Port Types and their Compatibility Definitions are created for each EReference to ensure the associated ports can only connect to each other correctly.
When two EObjects are associated to one another through an EReference the appropriate Port Individuals of the equivalent Component Individual are connected-to each other.
Arbitrary constraints in ECore are specified using annotations on model elements, written in some constraint language. These annotations are mapped to Constraint Instances and their corresponding Constraint Expressions. Special cases of constraints, particularly Property Definition and its sub-types (Attribute Definition, etc.), are utilised by previous mappings.
The constraints defined in our configuration model come from the SBVR specification. An example is shown in Figure 6 in that, the ‘has meaning’ relation between an ‘Expression’ and a ‘Meaning’ is allowed if and only if the ‘Meaning’ is connected to the ‘Expression’ through a ‘Representation’ and the ‘has representation’ and ‘has expression’ relations.
In the configurator used by our implementation the different aspects of the configuration knowledge are mapped to a generative CSP (or GCSP) as described in [Stumptner et al., 1998]. This particular approach differs in its definition of some of the previously discussed concepts of [Soininen et al., 1998] in the following respects: • Non-leaf nodes in the taxonomy are assumed to be abstract; therefore, concrete component types that have sub-types are split into two types: an abstract super-type and a concrete sub-type. • Ports are specified as necessary or optional; therefore, Port Definitions with higher cardinalities result in multiple ports that are then grouped into port sets, which allow quantitative reasoning over their members.
By configuring the SBVR models directly, we perform configuration of the concepts in a business specification rather than a syntactic parse tree as in [Kleiner et al., 2009]. This results in an iteratively more detailed domain model, where new domain knowledge is taken into account each time a new sentence is processed. In addition, inconsistencies can be detected more easily than by reprocessing the model after new knowledge is added through a model transformation. Finally, the mapping to SBVR is simplified as the lexicon maps directly to the semantics, rather than an intermediate syntactic model. This solves the issue of multiple parse trees with correct mappings to SBVR as, for example, the parse of ‘Each branch is included in exactly one local area’ where ‘local area’ is considered a noun phrase would be inconsistent with the domain knowledge, while the parse where ‘local area’ is considered a noun would be consistent. 4
We present the results of early experiments on the configuration of domain knowledge for NLP. We performed multiple tests of the example structural rule (1) and gathered statistics on the performance of the configurator in configuring these kinds of models. The results are summarised in Table 1.
The configurations produced were evaluated by hand for correctness. In each case the configuration was correct (corresponding with that shown in Figure 2b). The high number of variable assignments are due to relationships in the SBVR meta-model with cardinalities higher than one producing multiple ports in the configuration model; therefore, most port assignments are to the unconnected state.
It is interesting to note the correlation between the (minimum) number of backtracks and the number of components generated. This is due to the nature of the SBVR meta-model in two respects: (1) it contains a large number of relations with a cardinality of zero-to-many, and (2) it uses reified relations, which means that, in terms of configuration, each relation is represented as a component.
To prevent spurious connections between components, ports representing a zero-to-many relation are first set to the unconnected state; therefore, they are only connected to a component if being unconnected violates some constraint, causing a backtrack. Furthermore, the use of reified relations means that new relation components need to be created, even if it results in connecting two existing (non-relation) components. Therefore, in the optimal search of only connecting existing (non-relation) components, there will always be the same number of backtracks as generated components.
The higher number of backtracks in other configurations of the example are the result of the non-deterministic solver attempting variable assignments in a suboptimal order. This has a negative impact on performance. However, this could be avoided by providing SBVR specific procedural strategies for guiding the search [Stumptner et al., 1998; Hotz and Wolter, 2013]. For example, ordering heuristics can be provided to change the order in which different component types or port types are assigned. Moreover, the search space could be reduced by preventing the creation of certain component types. For example, we assume a sentence is to be interpreted in the context of a provided vocabulary, hence we could prevent the # Constraints # Variable Assignments Min. # Backtracks Max. # Backtracks Ave. # Backtracks # Components Generated
Sentence (1) 197 5162 6 25 17 6 creation of new object types, fact types, and other concepts related to the vocabulary aspects of the SBVR meta-model.
We are in the process of implementing a larger example, a portion of which assigned 4812 variables, generated 5 new components, and took 49 backtracks to do so. This emphasises the need for heuristics to help guide the search. Previous work in using configuration for natural language parsing has translated property grammars into a configuration model [Estratat and Henocque, 2004]. Using this approach, a simple context free-grammar (anbn) and a subset of French were processed. This approach focuses primarily on the syntactic aspect of generating parse trees. Although the possibility of incorporating semantics is suggested in [Estratat and Henocque, 2004], none were incorporated in the processing of the natural language subset.
The approach of [Estratat and Henocque, 2004] was adapted to English and a Model-Driven Engineering environment in [Kleiner et al., 2009]. Although their focus remains on the syntactic aspect of generating parse tress, SBVR semantics are partially taken into account by associating elements of the parse tree with SBVR types. This information is used to simplify the model transformation; however, the domain knowledge included in the SBVR model is not taken into account and, therefore, the process is not truly semantic.
Other approaches have used standard CSP translations of Dependency Grammars [Duchier, 1999] and, more recently, Property Grammars [Duchier et al., 2012] in order to process natural language. However, both of these approaches focus on syntactic parsing, while we aim to incorporate semantics and domain knowledge directly into the parsing process. 6
In this paper we have presented an approach to natural language processing that utilises configuration of domain knowledge to determine the validity of an expression. In effect, this performs natural language understanding, at least in terms of the semantic representation used. Moreover, we have demonstrated how techniques from Cognitive Linguistics can be combined with a translation of the SBVR meta-model (the semantic representation in our case) into a configuration problem in order to achieve this natural language understanding.
Our approach is novel in its combination of techniques from Cognitive Linguistics and configuration, and in that it performs configuration directly on the semantics of the domain knowledge. This is in contrast to previous approaches that use configuration or CSPs for natural language processing, as they tend to focus only on the syntactic aspect of generating a parse tree. As a result, our approach benefits from a simplified lexicon (important to our application in the business domain), improved mapping to the target semantics, and the semantic disambiguation of terms during processing.
The presented experimental results demonstrate the feasibility of our approach. In its current form, however, the configuration of the SBVR model can be inefficient and, therefore, future work will look at providing heuristics in order to ensure better performance in the configuration process. In addition, a more thorough evaluation of the process will be performed over larger examples in order to determine the effect of growing domain knowledge on the process.