Controlled natural languages (CNL) and computational semantics in general do not address word sense disambiguation, i.e., they tend to interpret only some functional words that are crucial in the construction process of discourse representation structures. We present two alternative frameworks for supporting polysemy in controlled languages. The approaches result in more natural CNLs suitable for description and translation of multi-domain texts.
There are several sophisticated controlled natural languages (CNL), which cover
relatively large subsets of English grammar providing seemingly informal means for
knowledge representation
Two commonly accepted restrictions are used in CNLs to enable construction of unambiguous discourse representation structures (DRS): a set of interpretation rules for potentially ambiguous syntactic constructions, and a monosemous lexicon — words are treated as predicate identifiers whose ‘meaning’ is defined only by FOL formulas derived from the text being analyzed. While the first restriction limits only syntactic sophistication of a language, the second one causes essential communication limitations as the natural language lexicon is inherently polysemous (e.g. “The library[collection] consists of million books”; “The city is constructing a new library[building]”). The polysemy of natural language is not a deficiency, but rather a gateway for referring to the rich background knowledge invoked by the same lexemes in different contexts thus leading to multiple word-senses. In this paper we address the latter limitation.
The root cause of polysemy in a natural language is that at any given moment there
is only a ‘finite’ number of lexemes, which speakers of the given language know and
thus can use for communication. Meanwhile there is a potentially unlimited number
of new concepts (discourses) that might need to be named and communicated about.
Metaphoric reuse of existing lexemes therefore is unavoidable in the natural language,
which can be summed in a saying: language is a graveyard of ‘dead’ metaphors
Internally consistent domain ontologies that follow lexicon-driven naming conventions we call micro-ontologies. Table 1 illustrates WSD problem when the text references different domains that use overlapping and potentially inconsistent terminology.
Micro-ontologies (ontological text)
Axioms in ACE Every building is a construction and has a roof.
Every library is a building.
Every collection is an abstract-entity that contains some items.
Every library is a collection that contains some publications.
Every construction is a physical-entity.
No physical-entity is an abstract-entity.
Assertions (factual text) x o -B There is a library[buildings] that has a green roof.
A The library[collection] contains some valuable publications.
The role of polysemy is most clearly apparent from the multilingual point of a view: it is impossible to avoid interpretation of lexemes when translating a text. In our case, interpretation means selection of the appropriate micro-ontology (see Table 2).
Although grammar constructions (OWL DL mappings) and lexicons for the source and target languages would differ, the interlingua — OWL DL micro-ontologies and their consistent mergers — remain the same. Moreover, by attaching translation equivalents to the ontological concepts, micro-ontologies simultaneously serve as monosemous multi-lingual lexicons facilitating the translation process. The term interlingua we mean in a wider sense: not only in the multi-lingual context, but also for monolingual multi-domain communication.
The problems of WSD and ontology merging are tightly intertwined and, in our view, the lack of definitive success is largely due to addressing these issues separately. Therefore we address both of these problems simultaneously — OWL DL formal semantics can be used to dynamically handle micro-ontologies for WSD over polysemous factual sentences. From all the available micro-ontologies for each sentence (or clause) are selected those that can be invoked (directly or via some merger) by a target lexeme (typically, a predicate) or other syntactically related lexemes (syntactic links are mapped onto ontology properties). In general, more than one micro-ontology can be invoked by an assertion due to the fact that different word senses are not necessarily mutually inconsistent. Selecting the largest micro-ontology merger likely unveils the most specific meaning (and facilitates further reasoning tasks). However, it is not necessary to get rid of the consistent alternatives — in case of later inconsistency they can be used during backtracking. As long as the current discourse remains consistent, its merged ontology is gradually augmented; otherwise additional discourse ontology is created separately.
The proposed concept of micro-ontology is similar to the Cyc concept of
microtheories
The above proposed rather ‘classic’ solution for adding polysemy to CNLs is
theoretically plausible, but it also raises a critical question: is this really how the
natural language works? It is well acknowledged in linguistic and cognitive sciences
that polysemes are etymologically and therefore semantically related, and typically
originate from metaphoric usage
Frame-semantic linguistic theory
The ultimate purpose of a CNL is to build a formal DRS capturing the full semantics of the input text. Although one could try to merge the classic DRS construction techniques rooted in FOL with FrameNet for a more natural polysemous CNL, this would not aid the disambiguation problem. Therefore we propose an alternative discourse representation approach based on PDDL (Planning Domain Description Language) leading to a new kind of CNLs not rooted in FOL anymore.
PDDL
The idea to differentiate two kinds of sentences in natural language — the ontological and the factual ones (T-Box and A-Box in Tables 1 and 2) — turns out to be a helpful principle. Although natural language expressions occasionally might be a mixture of both kinds of sentences, mostly such distinction on sentence level is possible.
Polysemy is less relevant for the background knowledge (ontological sentences), which essentially define language-independent abstract concepts in some, usually monosemous, domain ontology. Meanwhile polysemy support is crucial for the factual communication, which typically does not explicitly reference the background knowledge (T-Box), which needs to be guess-mapped through the polysemous words used in the text. In a CNL the corresponding T-Box has to be introduced explicitly along with an A-Box, usually by manual sharing of ontological sentences among factual texts. This forbids possibility for (inconsistent) polysemy in existing CNLs.
While it is disputable whether a CNL is a more convenient approach for describing ontologies (T-Boxes) than the formal languages and their graphic editors, a CNL is definitely an advantage when describing concrete situations through factual sentences. Vast amounts of such descriptions already exist in a written form: consider, for example, information extraction from a newspaper archive, which is predominantly a factual text.
In contrast to the universal macro-ontologies, micro-ontologies offer several significant advantages: (a) they do not impose a single consistent scheme, allowing many distinct points of view to co-exist; (b) they can be seen as snapshots of some aspects of ‘reality’, supporting non-stable and temporal entities — existing ontologies don’t have to be updated each time the reality changes; alternative ontologies should be introduced instead — it is a task of a reasoner to choose the appropriate ones; and (c) they scale well — things don’t have to be compressed in a restricted number of categories thus avoiding ‘signal losses’; the only restriction is the size of a lexicon.
In Sections 2 and 3 we have proposed two very different approaches for adding polysemy to CNLs — the micro-ontology approach in Section 2 is more traditional and compatible with existing CNLs, while the PDDL approach in Section 3 is more radical.