Research On Lattice Based Spoken Document Retrieval

Spoken document retrieval is to search in the mass speech resource and return relevantspoken documents or segments according to the users’ query, which plays an important role ininformation security, speech search engine and management of speech resource. The spokendocument retrieval based on lattice is appearing a booming prospect and receiving more andmore attention, however, the special structure of lattice also brings new problems and challenges.This thesis mainly focuses on the lattice structure improvement, optimal recognition unit andretrieval unit selection and re-ranking of relevant documents, aiming to find effective ways toaccelerate retrieval speed and precision. The main content includes:(1) A lattice structure improvement method integrating phonological feature is proposed tosolve the problem of the ignorance of phonological feature knowledge in the traditional method.Owing to the complementarily of the lattices from different sources, a lattice is firstly generatedby the speech recognition system based on phonological feature detection, which then makeinformation merging with the lattice from automatic speech recognition system. Furthermore, thesegment alignment based on position is employed to compress the scale of merged lattice.Experimental results show that the proposed lattice contains more correct recognition results, theindexing coverage increases from77.83%to80.34%and the lattice error decreases from25.31%to19.66%, making good enhancement of speech retrieval performance.(2) A spoken document indexing method based on subword-based position specificposterior lattices(S-PSPL) is proposed, aiming to overcome the inconsistency between optimalrecognition unit and retrieval unit in the existing Chinese spoken document retrieval methods. Inthis method, the speech document is firstly decoded to generate PSPL by using word asrecognition unit. Then each word in the PSPL is replaced by its constituent subword units.According to the posterior probability relationships between each word and its constituentsubword units, the original PSPL can be converted to corresponding S-PSPL, which can be usedto generate a subword-based index for retrieval. This method achieves the goal of making wordas recognition unit and subword as retrieval unit. Experiment results show that the new methodcan not only utilize the well-trained language information, but also solve the problem ofincorrect segmentation in the word-based recognizer. Better performance is obtained comparedwith the current indexing methods which use word as both recognition unit and retrieval unit.(3) A re-ranking method of relevant documents based on acoustic feature similarity isproposed, solving the problem of the incorrect ranking of relevant documents in the retrievalresults. This method improve the speech document retrieval system by using pseudo relevancefeedback, firstly N pieces of spoken documents with high relevance scores are selected as pseudorelevant documents from the retrieval results, then make acoustic feature similarity comparisonbetween the spoken document and pseudo relevant document during the hit region, at last,integrate the original relevance and acoustic feature similarity to get new relevance scores, whichis the basis to make re-ranking. Experimental results show that the R-Precision increases from69.07%to75.82%, and the retrieval performance is effectively enhanced as the increase of iteration times.