[attachment=13274]
Abstract
The precision of the answer is now essential for aquestion answering system, because of the large amount of freetexts on the Internet. Attempting to achieve a high precision,we propose a question answering system supported by casegrammar theory and based on VerbNet frames. It extracts thesyntactic, thematic and semantic information from the questionto filter out unmatched sentences in semantic level and to extractanswer chunk (a phrase or a word that can answer the question)from the answer sentence. VerbNet is applied in our system todetect the verb frames in question and candidate sentences, sothat the syntactic and thematic information as well as semanticinformation can be therefore obtained. Our question answeringsystem works well especially for answering factoid questions.The experiments show that our approach is able to filter outsemantically unmatched sentences effectively and therefore rankthe correct answer (s) higher in the result list.
I. MOTIVATION AND CONTRIBUTION
In the early stages of question answering, the responses toquestions are from the knowledge encoded in preprocesseddatabases. Most of the question answering systems at thattime are structured knowledge-based [1]. As the quantity oftexts on the Internet increased dramatically in the last decades,the other branch of question answering, which is labeled asfree text-based question answering, has become another mainstream.On the other hand, also because the information availableonline has been growing exponentially, locating the requiredinformation resources precisely by an information retrievalsystem (for example, question answering system, web searchengine) becomes more difficult than it was in the past fewdecades. Taking the web search engine for example, for mostof the queries issued by users there are tens of thousands documentsreturned. The ranking policy of the returned results hasbeen a key to search engine user’s satisfaction [2] (for instance,the ranking techniques applied by Google Inc. contributegreatly to its success), which proves that the requirement onprecision is now very strict.As we know, one of the main differences between a websearch engine and a free text-based question answering systemis that the latter takes a complete question sentence as the inputinstead of several words or phrases. It is clear that there aremore syntactic, thematic and semantic information that can beused by a question answering system. How to make use ofthe extra resources to improve the precision is crucial to freetext-based question answering.Issuing the extracted key words of the question sentence forretrieval is now not always sufficient to recognize a correctanswer. For example, there are a question and a sentence asfollows
Q: Who directed the 2008 Beijing Olympics openingceremony?
S: Because Michael directed a documentary on thepreparation of the 2008 Beijing Olympics, he wasoffered a ticket to the opening ceremony.Although all key words of the question are included inthe sentence, however, it is not a correct answer because ofthe incorrect positions of the key words (for example, the2008 Beijing Olympics opening ceremony should be on theobject position of direct). Syntactic, thematic and semanticinformation should be taken into account here to clarify themeaning of the sentence in order to make the decision onwhether it is a correct answer or not.With these in mind, we propose a practically feasible questionanswering system, which extracts syntactic and thematicas well as semantic information from the question and usesthem to filter out incorrect answers. We also perform a set ofcomparative tests, and it is proved by the increased precisionthat our approach is an effective way to make use of thesyntactic, thematic and semantic information in the questionsentence. Our system is a free text-based question answering,therefore not like the structured knowledge-based questionanswering, it is not capable of deep reasoning.The outlines of the paper are as follows. In section II, thefree text-based question answering, the case grammar theory and VerbNet will be introduced. In the next section, we presentthe design of our question answering system in detail. Acomplete example will be run through the section to visualizethe processing flow. After that, we discuss two techniques thatcan probably enhance the system performance. In experimentsection, the experiment setting and results are presented. Therelated work section is where we discuss the differences ofour work and the other previous work. The paper ends withthe conclusion and the future work.
II. BACKGROUND INFORMATION
A. Free Text-based Question Answering
For a generic free text-based question answering system,each time a user issues a question as a query in naturallanguage, the system usually parses and analyzes the questionbefore classifying it. The system then retrieves one or morequeries generated from the question against a large size corpus.From the returned results, one or multiple clear and preciseselected answers, each in the form of a sentence or a pieceof a sentence, are responded to the user. Figure 1 shows thethree phases of a generic free text-based question answeringsystem.B. Introduction to Case GrammarCase grammar is a grammar theory proposed by CharlesJ. Fillmore in 1966 [3]. It is a semantics oriented modelfor syntactic analysis, which pays primary attentions to therelation between the structure and the semantic meaning,focusing on the uniqueness of the surface layer structure andthe generality of the inside structure. In a word, the casegrammar focuses on the relation between the verb and theother parts of the sentence, by which it forms a deep semanticstructure.The case grammar depicts a sentence’s meaning by describingits center predication and the object of the predication. Asentence (S) can be divided into two parts, the Proposition (P)and the Modality (M). A Proposition is the set of relationsbetween the verb and nouns, while the modality is referred tothe way in which the meaning of a sentence or clause may bemodified through the use of a modal verb, such as may, can,will, must.