1URL:http://www.cyc.com/
Knote (the Planet knowledge editor), edge engineer manually annotates the story using 7. Story retrieval and query answering. The Lois interface supports integrated access to the story archive
describes our suite of tools in order to give some backbelieve it is important to mix the syntactic and semanin this case \murder", in any tense. Trigger words can sentence is also matched using the extra words \by" and Conversely, some grammar constructions (such as dates) ambiguities, for example, ontologies can provide us with \with." visiting a place then this someone should be a person." of objects by syntactic features or by means of the KMi tactically identify some fragments of the sentences. We scribe the use of template-driven IE to populate an ontic. The semantic checking is often necessary to resolve tion 5 gives conclusions and directions for future work. sentences that contain keywords (target-string) encoded information is the subject or the direct object of a verb a dieren t set of keywords. By using this keywords we select only relevant sentences from the whole text. then in our templates. For our specic event type we have tology. The template matching itself is supported seaxioms of common sense knowledge such \if someone is tent.
Planet ontology (dened in section 2). lem of populating the ontology. Section 4 describes the use of ontology to cope with the ambiguity in the idendescribed in 1996 For example, if the targeted [Rilo, ]. an ontology engine to supply the necessary semantic conthen the best trigger word should be the main verb. The based on some of the event types we use. Finally, Secbe reliably identied using linguistic rules lik e the ones we parse this sentences, and nally recognize the t ypes ground. In Section 3 we present a solution to the probmantically by referring to the ontology, but also conclassication process embedded in our system and the tains some lightweight NLP techniques in order to syntication of objects in the story, along with examples can be recognized robustly. Overall, our primary contriFor our IE we use a template-driven approach similar to, and based upon, that of Rilo. We identify the bution is to integrate a template-driven IE engine with In particular, the main aim of this paper is to dewhich are relevant to our institute. The Ontology Licommunication between members of KMi [Domingue journalists. In the case of KMi Planet it contains stories and Motta, 2000]. the ontology library. The Story library contains the text ing activities: brary contains several existing ontologies, in particular of the stories that have been provided to Planet by the nario supported by KMi planet, and supports the followTwo primary components are the story library and the KMi ontology.
KMi Planet is a Web-based news server that facilitates augmented the basic publish/nd sce- PlanetOnto2
the latest stories using a standard Web browser, pre-process the email-text: tokenise the text and join expressions such as dates into a single token a new domain the entire knowledge process of dening Then the system decides to classify the story according For the templates we use a prolog-like notation: the ontology. For example, \visiting-a-place-or-people", \academic-conference", \academic-workshop", \eventa story more than one type of event can be recognised. mining which kind of templates are necessary in our set involving-project", etc. Each event has several templates ple templates. However, problems might occur when in of templates. Currently, in order to apply our system to with the following criteria: how many feature for each sentence is converted to a list of words (or tokens) and keywords, some variables which are allowed to match then matched against a template list containing some We have devised several type of events in KMi planet used to recognize it. So far we have a library of 100 simThe nature of the domain is very important in detertemplates has to be repeated. type were encontered in the story. one or more words, and some underscores \ " that can dened in our KMi planet on tology 1. classify the story according an event type prend the objects in the story using further partial parsing identication 3. produce OCML code 4 Examples of classication and object Figure 1: Email Story on syntactic features, e.g. we assume titles such as ample, project names will be recognized this way. domain this means classify them as person, place, date, etc. This identication is done based partially mantic knowledge obtained by querying the server \Dr." can be used to recognize some people. How\Web-onto" for the KMi planet ontology. For exidentify each of the objects in the story. In our ever, some other classications are based upon sejournalists use slang, break conventions for capitalizaknowledge engineer for conrmation. For each type of snapshot the extracted information is presented to the formation shown in screen snapshot of Figure 2 In this side of the form there is a menu that presents to the event a dieren t form is presented from the information writing in such an archive can vary wildly. All these tion, incorrectly use punctuation, etc. Also, styles of ample, if a word is capitalised and not starting a sentence nalist from KMi sent the story (presented in Figure 1) company. However, as discussed earlier we expect such user with possibles instantiations for each eld. The suge-mail message) and see if it matches any of our domainorganization-being-visited. Therefore, in the right hand Our domain (KMi planet) consists of email message, specic templates. If no extraction template applies to Let us take the following example: Suppose that a jourmeans that irrelevant text can be processed very quickly. Our IE tool will take the story and produce the in\Ltd" then it can be tagged as a proper noun of type The main process in the classication is to tak e each make it diÆcult to apply heuristics to identify proper use templates. sentence in the text (in our case a story written in a a sentence, then no information will be extracted; this to the e-mail server nouns. NLP approaches typically use heuristics, for exstored in the ontology. Figure 2 also has some elds and hence is generally unstructured for example: the then it is a proper name. If a string contains \& Co" or methods to be too fragile for our domain, and so instead that cannot be found in the story such as people-orgested instantiations are obtained from the ontology. system might encounter objects of type: visitor, place instance, for the event \visiting-a-place-or-people" the object’s types. predened objects that should be found in the story . For of the 40 types of events according with the type of the Classication is performed in the follo wing steps: and date. Currently, our system handles 22 dieren t objects that are found. For each event type we have a We classify the story or document as belonging to any 2. [ ,X, , \visited", Y, \from", Z , ] matches the automatically recognized by syntactic features of month, day, and year. & place (Y), but also require \date(Z)"; dates are sentence word list. Again, we demand visitor(X) for matching to the visited templates above:: would also be a match against the template, and extract the same information. In contrast, the misspelled tirely and extract no information. Let us take another example of a candidate sentence initial \Tooo" might cause a complete parser to fail en