Anaphora resolution in current computer-processable controlled natural languages relies mainly on syntactic information, accessibility constraints and the distance of the anaphoric expression to its antecedent. This design decision has the advantage that a text can be processed automatically without any additional ontological knowledge, but it has the disadvantage that the author is severely restricted in using anaphoric expressions while writing a text. I will argue that we can allow for a wider range of anaphoric expressions if we consider the anaphora resolution process as an interactive, machine-guided knowledge acquisition process in cases where no suitable antecedent can be found automatically. This knowledge acquisition process relies on the human author who provides additional terminological information for the anaphora resolution algorithm { if required { while a text is written.
Computer-processable controlled natural languages are engineered subsets of natural languages designed to reduce ambiguity and vagueness that is inherent in full natural language [5, 10]. These controlled natural languages look like English but are in fact formal languages that can be translated unambiguously into the input language of an automated reasoner and be used for several reasoning tasks, among them for question answering. Similar to full natural language, these controlled natural languages allow for anaphoric expressions but their form and usage are considerably restricted. An anaphoric expression (= anaphor) is a word or a phrase that points back to an expression (= antecedent) that has been previously introduced in the text (see [8] for an overview). The two most important types of anaphoric expressions that are used in sentences and between sentences in controlled natural languages are pronominal anaphora and de nite noun phrase anaphora. De nite noun phrase anaphora take the form of de nite descriptions and proper names. Computer-processable controlled natural languages use relatively \safe" anaphora resolution algorithms that rely mainly on syntactic information, accessibility constraints and the distance of the anaphor to its antecedent. This makes it easy for the machine to resolve anaphoric expressions automatically but di cult for the human author to remember the approved forms of anaphoric expressions. In the following discussion, I will focus on de nite descriptions, in particular on bridging de nite descriptions [9], and discuss how these anaphoric expressions that require inference and inference-supporting knowledge can be handled in a controlled natural language context.
In the simplest case, a de nite description that is used anaphorically matches
syntactically with its antecedent. For example, the anaphor in (
However, the relation between an anaphor and its antecedent is often more
complex than that of identity. The relation may be a synonymy relation as in (
6. The hard disk is broken and does not spin.
These bridging de nite descriptions point back to a noun phrase antecedent
that has already been introduced in (
On the one hand, WordNet [4] has been used as an approximation of a knowledge base for resolving bridging de nite descriptions in unrestricted texts but this proved to be highly unreliable for the automatic identi cation of correct semantic relations [12]. The situation gets even worse if we work in a domain where a very speci c vocabulary is required.
On the other hand, one can bite the bullet and construct a (linguistically motivated) formal ontology for a particular application domain that contains { among other things { the required terminological knowledge for resolving de nite descriptions, for example terminological statements such as: 6. (define-concept professor (and leader teacher supervisor)) 7. (define-primitive-concept teacher academic staff) 8. (define-concept teaching event
(and event (all has agent teacher) (some has theme course))) 9. (equivalent academic staff faculty member) 10. (define-primitive-concept laptop computer
(some has direct part hard disk))
Note that this terminological knowledge (here expressed in an expressive description logic [1] that follows the KRSS1 notation) can be speci ed directly in a controlled natural language and then be translated automatically into the above target representation: 11. Every professor is de ned as a leader and a teacher and a supervisor. 12. Every teacher is an academic sta . 13. Every teaching event is de ned as an event that has only teachers as an agent and that has a course as a theme. 14. Every academic sta is equivalent to a faculty member. 15. Every laptop computer has a hard disk as a direct part.
In this scenario, the knowledge base can be updated if new factual information becomes available and the semantic relations can be checked automatically by querying the description logic knowledge base. The information in the knowledge base can even be used to guide the writing process in a predictive way (using similar techniques as in [7]) since all background information has been carefully speci ed in advance.
However, there exists another scenario where a domain expert might want
to assert new factual information but the terminological knowledge is not yet
available. For example, the domain expert might want to assert rst the
sentence (
PENG Light is a computer-processable controlled natural language that can be used for knowledge representation [11]. The language processor of the PENG Light system translates texts incrementally into TPTP2 notation. Since the uni cation-based grammar of PENG Light is bidirectional, the language processor can take a syntactically annotated TPTP formula of a sentence as input and produce an output string in controlled natural language. The grammar of PENG Light is written in a DCG-style notation and is processed by a chart parser. The TPTP notation is built up during the parsing process together with a paraphrase that illustrates how a sentence has been interpreted and how anaphoric expressions have been resolved. The grammar maintains a list of accessible noun phrase antecedents during the parsing process whereas accessibility is de ned in a similar way as in Discourse Representation Theory [6].
The proposed anaphora resolution algorithm for de nite descriptions extends the existing algorithm of PENG Light in a systematic way. The existing algorithm resolves an anaphorically used de nite description with the most recent accessible noun phrase antecedent that matches fully or partially with the anaphor and agrees in number and gender with that anaphor. The new algorithm relies 1 http://www.bell-labs.com/user/pfps/papers/krss-spec.ps 2 http://www.cs.miami.edu/~tptp/ on interactivity and allows the author to specify semantic relations between a noun phrase antecedent and a bridging de nite description if this information is not already available in the knowledge base. This solution is compatible with a predictive authoring approach since the anaphora resolution process is machineguided and the author selects among a number of options.
In order to process bridging de nite descriptions the controlled natural language processor of PENG Light communicates with an automated reasoning engine. We have experimented with E-KRHyper [3], a model generator and theorem prover for rst-order logic, with CEL [2], a polynominal-time classi er, and with RacerPro [13], a description logic reasoning system. We currently use RacerPro for processing terminological information since it is the most advanced and versatile description logic system but we search for more expressive alternatives.
The following simpli ed DCG rule for de nite descriptions illustrates how feature structures are used to deal with syntactic, semantic and pragmatic information in PENG Light, and shows that the anaphora resolution algorithm is triggered whenever a de nite description has been processed:
The anaphora resolution algorithm takes a partial logical formula (R2), a list of accessible noun phrase antecedents (A2), a de nite description (D2) and the current paraphrase (P3) as input and returns an updated logical formula (R3), an updated list of antecedents (A3) and an updated paraphrase (P4) as output. If the de nite description is an anaphoric expression, then this expression is replaced by the noun phrase antecedent and this replacement is marked in the output list (P4) of the paraphrase.
The anaphora resolution algorithms rst checks if the de nite description matches fully with the rst of the accessible noun phrase antecedents in the input list A2. The antecedents are ordered and represented as terms that contain syntactic and semantic information, for example: 16. object(X,academic staff)#[third,sg,masc fem]#[academic,staff] The matching is done over the syntactic information, and if no solution can be found, then the algorithm checks for a partial match. If the partial match succeeds, then the algorithm queries the terminological part of the knowledge base for a hypernymy relation using the description logic reasoner RacerPro. RacerPro's allows us to query this semantic relation in a direct way, for example: 17. (concept-subsumes? staff academic staff)
If there is no partial match, then the algorithm checks sequentially for synonymy, hyponymy and meronymy relations (note that the optimal sequence is genre-speci c but we don't know in advance what the author is going to specify): 18. (concept-equivalent? academic staff faculty member) 19. (concept-subsumes? academic staff professor) 20. (concept-subsumes? (some has direct part hard disk) laptop computer) The actual resolution is re ected in a paraphrase and the domain expert can accept or reject the solution. If no solution can be found, then the domain expert has to decide if the noun phrase is a discourse new de nite description or semantically related to one of the accessible antecedents in A1. If the former is the case, then the domain expert simply accepts the new expression and a new discourse referent is introduced. If the latter is the case, then the domain expert has to specify the semantic relation between the antecedent and the anaphoric expression on the interface level by selecting the relevant antecedent and the corresponding semantic relation from a menu. Once this has been done, the knowledge base is updated and the new bridging de nite description is licensed in the text. 5
I argued that anaphora resolution for de nite descriptions in computer-processable controlled natural languages can be interpreted as an interactive knowledge acquisition process in those cases where no suitable noun phrase antecedent can be found by the machine. The presented approach relies on a domain expert who works in collaboration with the machine and who provides the required ontological knowledge while a text is written { this approach brings the human back into the loop.