Man-machine Dialogue System, A Number Of Key Issues For Study

Spoken dialogue system is a research front of natural language processing, ofwhich the performance depends on some key problems, including language structureanalysis, language semantic comprehension and dialogue management. So research ofthis dissertation focus mainly on the key problems described above and someimportant and new findings are summarized as follows:(1) Concept of basic dialogue structure is put forward independently.Base on the requirement of language itself and engineering application, thedissertation defines the concept of basic dialogue structure by four ways and considers:a), the basic dialogue structure is the most basic interactive unit in dialogue structure;b), consisting of a speech act sequence by two or two above different speakers; c), thebasic dialogue structure can be formally represented as a steady structure with initialunits followed by responding ones ; d), having definite semantic intension. Accordingto the descriptions above, the dissertation further brings forward a tag system ofshallow dialogue structures with multiple hierarchies, which can be applied toTSC973 telephone tongue database. The concept of basic dialogue structure putforward by this dissertation usefully extends the research of speech act towards thedirection of dialogue structure, and avoids the difficulty in implementing a wholeHCRC project, which is of notable practice background and realistic signification (asdescribed in detail in chapter 2) .(2) A HHMM model is applied to analysis of dialogue structure for the firsttime.The dissertation shows orderly how to apply Navie Bayesian, HMM and HHMMmodel to shallow dialogue analysis in a manner from simple to complexity. Withrespect to HHMM, this dissertation first illustrates an equivalent expression of PCFGto HHMM, then further shows the HHMM can be decoded by CKY and a detailedalgorithm is listed, finally, a hierarchical solving system comprising two layers andthree parts is put forward to solve the default of high space time cost of HHMM, thetwo layers are the semantic layer and the grammar layer respectively; and the threeparts comprise the structural boundary identification unit, the structural semanticidentification unit and the interior topology identification unit of basic dialoguestructure. With respect to the structural boundary identification unit of basic dialoguestructure, the dissertation brings forward two solving methods based on maximum entropy: Model-Maxent-V1 and Model-Maxent-V2; with respect to the semanticidentification unit of basic dialogue structure, a system exploring both rule methodand statistical method is brought forward; and with respect to the interior topologyidentification unit of basic dialogue structure, the effecting factors are divided intolocation factor, pos-neg factor and pragmatic factor, after that, three factors areprocessed respectively, that is, HMM method is adopted to identify location factor, asfor pos-neg factor and pragmatic factor, because of their good regularity in TSC973telephone tongue database, some simple rule based methods are used. Experimentalresults show that the ultimate F exported by the hierarchical solving system ofHHMM is 39．54%, which is much better than that of HMM model and NavieBayesian model, for example, the ultimate F increases 6．33 percent comparing withthat of HMM model (as described in detail in chapter 2) .(3) Take a probe step to research dialogue structure with an un-supervisedmannerAs an extension and deepening of the supervised learning of dialogue structures,this dissertation takes a probe step to research dialogue structure with anun-supervised manner. Particularly, the studying focus is more on the research in theun-supervised learning of basic dialogue structural boundary. First, the dissertationillustrates the relationship between mutual information distribution and basic dialoguestructural boundary by a mutual information distribution diagram, which to someextent shows the rationality of exploring basic dialogue boundary by use of therelationship degree between speech acts; secondly, a synthetical score criterion is putforward according to the basic dialogue structural unit, which estimates theprobability of a target unit being a basic dialogue structure according to the linkintensity of inner unit and inter-unit respectively, finally, a novel dynamiticprogramming based un-supervised segmentation algorithm is put forward, and themost optimized segmentation is achieved in the range of a whole dialogue.Experimental results show that the system performance value F according to thesynthetical score criterion can reach 69．16%, which is higher almost fifteen percentthan that according to the un-supervised learning framework of MI(as described indetail in chapter 3) .(4) An improved agenda based DM is put forward.The task structure of a dialogue includes process structure and descriptivestructure. To better process a complicated dialogue task having both two typicalstructures, a feature structure is introduced to the traditional agenda based DM, which, as a result, not only the advantage of traditional DM is reserved, but the advantage ofa feature structure , which is suitable for representing complicated objects, is alsocombined into the current dialogue system. All those described effectively extend thesystem capability to cope with complicated dialogue tasks(as described in detail inchapter 4) .( 5 ) A novel shallow parsing algorithm based on random forest is put forward.With respect to the share task of shallow parsing described by CoNLL2000, thedissertation represents a novel algorithm based on random forest, study shows that thealgorithm put forward by the dissertation can reduce the EMS memory requirement ofsystem, and moreover, it can effectively improve system performance by inductingmultiple kinks of random factors such as Bootsrap , Subspace and etc. to enable thewhole performance to form a partial peak value. Experimental results illustrates thatthe peak value F reaches as high as 92．25% in the case of the basic model plusBootstrap with a forest size of 5 trees when in fifteen dimensions, which exceeds 0．46percent than the maximal performance value F of the basic model (as described indetail in chapter 5) .