Data processing: database and file management or data structures – Database design – Data structure types
Reexamination Certificate
1998-03-30
2002-07-09
Black, Thomas (Department: 2171)
Data processing: database and file management or data structures
Database design
Data structure types
C704S257000, C704S009000
Reexamination Certificate
active
06418431
ABSTRACT:
BACKGROUND OF THE INVENTION
The present invention deals with speech recognition and information retrieval. More specifically, the present invention deals with a speech recognition system which employs information retrieval techniques to adapt a language model, and an information retrieval technique which employs speech recognition language models for retrieving relevant documents.
Generally, information retrieval is a process by which a user finds and retrieves information, relevant to the user, from a large store of information. In performing information retrieval, it is important to retrieve all of the information a user needs (i.e., it is important to be complete), and at the same time it is important to limit the irrelevant information that is retrieved for the user (i.e., it is important to be selective). These dimensions are often referred to in terms of recall (completeness) and precision (selectivity). In many information retrieval systems, it is necessary to achieve good performance across both the recall and precision dimensions.
In some current retrieval systems, the amount of information that can be queried and searched is very large. For example, some information retrieval systems are set up to search information on the internet, digital video discs, and other computer data bases in general. These information retrieval systems are typically embodied as, for example, internet search engines, and library catalog search engines.
Many information retrieval techniques are known. A user input query in such techniques is typically presented as either an explicit user generated query, or as an implicit query, such as when a user requests documents or information which is similar to a certain set of existing documents. Typical information retrieval systems then search documents in the large data store at either a single word level, or at a term level. Each of the documents are assigned a relevancy (or similarity) score, and the information retrieval system presents a certain subset of the documents searched to the user, typically that subset which has a relevancy score which exceeds a given threshold.
Some currently known information retrieval techniques or methods include full text scanning, the use of signature files, inversion, vector modeling and clustering, and tf*idf (term frequency*inverse document frequency). In full text scanning, Boolean functions are used in a query to determine whether a document to be searched contains certain letter strings. It is common in such scanning techniques to search each character of a document to see whether it satisfies the search string (i.e., the query) and then move the search one position to the right when a mismatch is found. This system has been adapted to use other ways of preprocessing the query, such as moving more than one position to the right when a mismatch is found.
The use of signature files involves discarding common words from documents to be searched and reducing the non-common words to stems. Each document to be searched yields a bit string (i.e., a signature). The signatures for various documents are stored sequentially in a file separate from the documents themselves.
Inversion techniques involve constructing a list of key words to represent each document. The key words are stored in an index file. For each key word, a list of pointers is maintained which reveals qualifying documents. The query is then advanced against the index and the pointers are used to identify the relevant and qualifying documents.
Vector modeling and clustering involves grouping of similar documents into groups referred to as clusters (this technique can also be applied to terms instead of documents). In order to generate a cluster, an index is formed by removing common words and reducing the remainder of the words to stems (which includes prefix and suffix removal). Synonyms are also commonly placed in a concept class which can have its terms weighted by frequency, specificity, relevancy, etc. The index is used to represent the documents as a point in t-dimensional space. The points are then partitioned into groups with a similarity matrix which is typically developed through an iterative process. In order to search the cluster, a query is represented as a t-dimensional vector and is compared with the cluster centroids. A cluster-to-query similarity function is generated and is used to pull relevant documents. The documents which are pulled (or retrieved) are typically those with a similarity value that exceeds a predetermined threshold value.
Semantic information is used in some information retrieval techniques to capture more information about each document in the information store in order to achieve better performance. In one such system, natural language processing is used to match the semantic content of queries to that of the documents to be searched. Sentences or phrases are used as terms for indexing the documents to be searched. Latent semantic indexing involves forming a term/document matrix in which the number of occurrences of a term in a specific document are plotted on a matrix. Small singular values are typically eliminated and the remaining term frequency vectors are mapped. Queries are also formed of term frequency vectors and are mapped against the matrix which contains the term frequency vectors for the documents. The documents are ranked by using normalized linear products in order to obtain a cosine similarity measure.
Another type of information retrieval technique which uses semantic information is a neural network. Essentially, a thesaurus is constructed, and a node in a hidden layer is created to correspond to each concept in the thesaurus. Spreading activation methods are then used to conduct searches.
Term frequency*inverse document frequency (tf*idf) is another technique used to determine relevancy of documents. First, a term used in a query is measured against the document to determine the frequency of that term in the document. It is believed that the degree to which the document and the term are related increases as the frequency of the term in the document increases. It is also believed that the usefulness of a term in discriminating among documents decreases as the number of documents in which that term appears increases. Therefore, the frequency of the particular term is also measured against the whole data store to determine the frequency level of that term in all of the documents. These two measures are used in determining the relevancy of any given document in the data store being searched.
As the data bases which are accessible to searching become ever more numerous, and as those data bases become larger, the problems associated with information retrieval also become larger. In other words, acceptable performance across the recall and precision dimensions is often more difficult to obtain with larger and more numerous data bases under search.
Speech recognition systems use a combination of the acoustic and linguistic (or language) information contained in an utterance in order to generate a transcript of the meaning of the utterance. The language information used by a recognizer in a speech recognition system is collectively referred to as a language model.
Many current speech recognition systems use language models which are statistical in nature. Such language models are typically generated using known techniques based on a large amount of textual training data which is presented to a language model generator. An N-gram language model may use, for instance, known statistical techniques such as Katz's technique, or the binomial posterior distribution backoff technique. In using these techniques, the language models estimate the probability that a word w(n) will follow a sequence of words w
1
, w
2
, . . . w(n−1). These probability values collectively form the N-gram language model.
There are many known methods which can be used to estimate these probability values from a large text corpus which is presented to the language model generator, and the exact method by which this is done is
Huang Xuedong D.
Mahajan Milind V.
Black Thomas
Kelly Joseph R.
Microsoft Corporation
Rones Charles L.
Westman Champlin & Kelly P.A.
LandOfFree
Information retrieval and speech recognition based on... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Information retrieval and speech recognition based on..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Information retrieval and speech recognition based on... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2901804