In the last decade the volume of available information on the web has grown exponentially. There are many more sources of information and each one seems to produce much more potentially relevant documents. As an effect of globalization, the news we hear from a remote point of the globe have now gained importance and may influence some aspects of our life. The newspapers, tv and other media spread the news from any event to the whole world.
The average citizen can read the papers and watch some news program on tv. However it's not possible to be aware of all occurrences in the world. In the other hand, most of the information taken in media resources may not be relevant to the end citizen. All this has a special importance to professionals whose activity relies on news analysis, such as stock market brokers, military intelligence or economists.
Nowadays, the main newspapers have an online RSS 1 service where they publish the latest news to all Internet users. Computer based systems can help people, allowing a quick and broader analysis on the available sources. Filtering by date and section (politics, economy) is not enough for today's demands. There must be done some automatic work on the body text of each article in order to capture the expressed semantics in it. This might involve NLP techniques and an inference enabled system. The semantic information captured from a document is stored in a knowledge base. Ontologies allow the definition of class hierarchies, object properties and relation rules, such as, transitivity or functionality. The information extracted from each news document is given a formal representation, associated with an ontology. The resulting knowledge base has the facts list expressed by instances of ontology classes, in a semantic context. Then it will be possible to make some inferences about them. This paper proposes an ontology based methodology for news article processing in order to:
cover a large amount of documents identify the most relevant documents try to automatically understand some information in those documents set a notification or action to do in case a certain 'thing' happens get automatic answers to some simple questions
The initial knowledge base is described in the next section. Section 3 shows the techniques used for news document analysis and knowledge base evolution. In section 4 we present the system features and how they are accomplished. Finally, section 5 describes some related work and in section 6 some conclusions and future work are pointed out.
When we have an isolated sentence it's usually difficult to automatically capture the semantics in it. Our approach uses an ontology as the starting knowledge base with semantic information that helps to perform the sentence analysis and the subsequent inferences and interrogations. Each element found on a sentence is related to the starting knowledge base and also with the previous sentence semantics processed. Senso is still in development and is the sum of a taxonomy of classes and a list of semantic information gathered from several ways. Some influences were taken from previous works ( [2] and [1]) where we had a top-level OWL2 ontology with some basic concepts. This ontology has a hierarchy of concepts used to organize the set of concepts mentioned in Portuguese text documents.
We used OWL because it has the intended semantic features and it is suitable for web publications, allowing us to share parts of our knowledge base in a direct and appropriate manner.
Besides the formal concept definitions and "IsA" relations, there are a few simple facts about everyday life that might be very useful for document analysis. This lead us to ConceptNet [3], which is a freely available common sense knowledge base and natural language processing toolkit. This tool gave us access to a semantic network presently available in two versions: concise (200,000 assertions) and full (1.6 million assertions) about spatial, physical, social, temporal, and psychological aspects of everyday life. ConceptNet is available in English and our work needs Portuguese language. We started by choosing a set of terms that we were interested on. Then we automatically followed their relations in the ConceptNet semantic network for a couple of nodes and collected those too. Having the terms and relations identified we run an automatic translation. Finally we filtered some wrong translations using another Portuguese dictionary. The top-level ontology and the ConceptNet translated terms were merged in a manual form. This work was done by several people using a web application where we could browse the database, insert new classes, update relations between classes, add or remove facts. Our current ontology (Senso Knowledge Base) contains about 2000 concepts and has several relations connecting them: isA, usedFor, locatedAt, capableOf and madeOf. All terms are written in Portuguese and they are not about a specific domain of knowledge. These concepts and relations represent a small common sense knowledge base about places, entities and events. Some of the top-level concepts are: AbstractConcept (the root concept), Event, Time and Entity, as shown in figure 1. Choosing any of those results, with a mouse click will show that term's detail, as illustrated on figure 4 for the term firearm. Those lines are part of that term's detail and they mean that firearm is a kind of weapon that might be used for actions like murder, hunt, shoot something or protect. The next section explains the document analysis performed by the system.
The proposed system might be used with any text documents in Portuguese natural language. Is this case, we focus our work in news articles that are published day by day by the national media. Some popular newspapers like Público 3 or Correio da Manhã 4 have a "last hour" Fig. 4. Senso: some details on firearm term news section in their web site, including an RSS channel. This is suitable for an automatic search for any recently added news article. We used a program to periodically collect the recent news from Público's RSS channel. As we can see in figure 5, each news item has some metadata fields: title, description, author, category, publication date and hour, and of course, the link to the web document containing the information. The category gives us a first simple classification for the document, placing it in Economy, Politics, International or Sports (in Portuguese Desporto -like the item listed in figure 5). The publication date gives the temporal context to the semantic content we find in the document. Later we will see some examples. Each document imported to the system has a text body. That text is processed, following a methodology based on natural language processing tech-niques, namely, a syntactical parser and a semantic analyzer able to obtain a partial interpretation of the document.
The tool used for the syntactical analysis is PALAVRAS [5]. It's a syntactical parser developed by E. Bick in the domain of the VISL Project5 . This parser is based in the Constraint Grammars formalism and it is able to cover a large percentage of the Portuguese language. Because the parser output is in a nonstandard format, it was necessary to transform it into a structured form, like XML and Prolog terms. This was accomplished with the translation tool6 Xtractor [6], that performs the conversion VISL to Prolog and XML. Let us consider a sentence in the above sports news item:
"Marcus Grönholm venceu neste domingo o Rali da Grécia."
(in English: "Marcus Grönholm won the Greece Rally, this Sunday.")
As can be seen in figure 6, the parser identified correctly the subject, the predicate and direct object. The next step is the semantic analysis. The technique used for this process is based on Discourse Representation Structures (DRS) [7]. The partial semantic representation of a sentence is a DRS built with two lists, one with the rewritten sentence and the other with the sentence discourse referents. We are only dealing with a restricted semantic analysis and we are not able to handle every aspect of the semantics: our focus is on the representation on concepts (nouns and verbs) and the correct extraction of its properties (modifiers, agents, objects). In the last section of this paper we point out some possible improvements for this text semantic analysis. The previous news item is stored in the system with the following details: Notice that we store some metadata like the item category and the time of publication. This is sometimes useful, when a sentence has a temporal modifier that might be related to the publication field. Later we shall see it.
All documents retrieved from the newspaper RSS channel are processed using this technique. This produces a knowledge base that we will use for inference processes as we describe in the next section.
Once the news documents are obtained from the source and analyzed they become part of the second knowledge base used by the system. This is called the facts knowledge base, and used with Senso ontology allows the system to perform some inference and then offer interesting features. Three main features are presented in the following subsections.
When we visit the newspaper site we can search for news items from category X or items where some pattern might be found in the text. This is not enough when we want more than pattern match search. Suppose we want to find the items about rally drivers that play some instrument. Simple syntactical searches would probably get many documents with low relevance. Our system can receive a search instruction and retrieve the set of documents that validate that condition. It's possible to search for documents about some entity by the description the user gave for it, in natural language. For the above example, the entity description introduced by the user (in Portuguese) is: 'é piloto de rali' and 'toca um instrumento'. Then, the Prolog syntax for the search is:
Suppose the document D1 has a sentence like: "Marcus Grönholm tocou piano até aos 16 anos de idade."
(in English: "Marcus Grönholm played the piano until 16 years old.")
Our ontology in Senso tells us that piano is an instrument. Now, if there is a sentence in our fact knowledge base saying that Marcus Grönholm is a piloto de rali (rally driver), then the document D will indeed be selected for the returned list of retrieved relevant documents, L. This feature is available in a web form in a high level manner. The prolog query searchItemList/2 is processed in the system backend and the list L is then shown in a nice format to the user.
Sometimes we want to schedule some action in the case of some certain event occurs. Our system allows users to set actions that will be triggered if a news item with a content that validates a defined condition is found. At the moment, these actions are only e-mails with some message sent to the user if any news item text validates a condition. As an example, suppose you are interested in economic transactions where a company buys another company (the Portuguese word is empresa, bellow), but you only want to receive the alert message if the news occur in November. Then you can instruct the system with the following: Action: mail you@di.uevora.pt -s ''ALERT: a Company is buying another company'' ([ metadata(pubDate,after,'2006-11') ],
[ empresa(X), comprar(X,Y), empresa(Y)] ) ]).
When a news item is fetched and processed, it's internal representation is checked for each user defined action condition list. The verified conditions will trigger the execution of the associated action.
The most interesting feature of the system is the Question-Answer feature. This receives a natural language written query, in Portuguese, and searches for answers, based on the collected news documents and the Senso ontology.
The analysis of a natural language query is split in three steps: Syntax, Semantics, and Pragmatics. Each query is processed using the same natural language tools used for the news texts. First, the received interrogation is parsed by the methodology described before. Like we did with the text sentences, each query syntactical structure is translated into a First-Order Logic expression (a DRS). Note that, at present, we are not able to deal with general unrestricted queries nor to translate them from a syntactical into a semantic structure. In fact this is a quite complex NLP problem and we have decided to deal only with specific subsets of the Portuguese language, namely, with interrogatives about specific domains.
The search for an answer is done by a logic-programming based module that performs a pragmatic interpretation of the query DRS over the full system knowledge base (Senso ontology and facts from the news). The inference process is done with the Prolog resolution algorithm, which tries to unify the referent from the query with facts extracted from the documents and expressed in DRS structures.
As an example, we could enter a query like: "Quem ganhou o Rali da Grécia?" (in English: "Who won the Greece Rally?")
The DRS for such query would be:
And the result displayed by the system web interface is given is figure 7. Note that in this case there is only one result. The response given may include zero or more values considered valid as response. For each possible response there is also the link to the news item where the system found the answer. The previous question was answered because the concept vencer is defined as a synonym of ganhar. Another note is that there was only one fact on the knowledge base about someone winning the Greece Rally. If we had a former document with a sentence identifying last year winner then probably we would get two answers. Then we could check the best solution by reading the text. The precise query for this year winner would be:
"Quem venceu o Rali da Grécia no ano de 2006?" (in English: "Who won the Greece Rally in the year 2006?") That would introduce a temporal modifier on the query DRS expression to be checked against the date of publication of the document. Another example of query about time is: "Quando é que Marcus Grönholm venceu o Rali da Grécia?" (in English: "When did Marcus Grönholm won the Greece Rally?")
Once again, the interrogative term quando's referent is matched against the temporal modifier on the sentence DRS: "este domingo" (in English: this Sunday). This information is then related with the news item publication date and we infer the desired date, 2006-06-04. Similar treatment is given to temporal expressions like today, this month, last year and other.
There are other initiatives related to the semantic content search. Ontologies are used in [8] for the specific domain of International Affairs. It has a natural language interface also, but works with RDQL7 instead of the Prolog logic resolution environment we adopted. Other field of research, in this context, is the automatic mapping from an existing large news archive, in XML standards, to OWL ontologies [9]. The semantic archive features provided by [10] include means to annotate news materials and semantic search and browsing capabilities. This system runs inside the newspaper environment and uses a newspaper library specialized ontology, while the system we present works alone and outside the newspaper, allowing the use of many independent news sources, and our ontology is about common sense knowledge and not about a specific domain. The Rich News system [11] helps to annotate radio and television news. It is specialized for the English language and combines automatic speech recognition with information extraction using the KIM8 knowledge management platform.
We proposed a methodology to an ontology supported news reader and questionanswer system. The system is based on the initial knowledge in the Senso ontology and on the semantic content extracted from documents. The system we described has some interesting features. However it exists only in a prototype version and it must now be improved to support a large-scale facts knowledge base.
The accuracy of the results found for automatic question-answer is also a point that needs more work. The system is affected in the inference process by:
the quality of the Senso ontology the precision of the semantic information taken from the text sentences.
The ontology should be manually revised and extended. The semantic analysis can be improved if we add a tool to identify the inter-sentence anaphoric references. Finally, the system needs to be fully evaluated.