Data processing: speech signal processing – linguistics – language – Linguistics – Natural language
Reexamination Certificate
1999-03-17
2001-09-11
Edouard, Patrick N. (Department: 2644)
Data processing: speech signal processing, linguistics, language
Linguistics
Natural language
C707S793000
Reexamination Certificate
active
06289304
ABSTRACT:
FIELD OF THE INVENTION
The invention relates to automatic text summarization based on part-of-speech.
BACKGROUND AND SUMMARY OF THE INVENTION
A conventional reading machine for the blind or visually impaired allows the user to manually increase the rate at which text in a scanned document is converted into speech, making it possible to generate very rapid speech and thus audibly flip through the document to obtain a sort of summary. Also, the user could manually select samples of the document and generate speech from each sample to obtain another type of summary.
A number of automatic summarization techniques have been proposed in other contexts. According to one such technique, manually derived templates are used to match certain patterns in text. When the templates are filled, a gloss of the template can be produced by the computer. This gloss ignores any item that was not included in the template and thereby reduces the quantity of text. This is the approach used by the participants in the yearly Message Understanding Conference (MUC). A drawback of this technology is that building templates is a long manual process that produces a domain-specific filter that cannot be applied to unrestricted text.
According to another known method it is necessary to read an entire text into memory and calculate statistics of word use, the most frequent terms being deemed to be most important for the sense of the text. Then, the original text is rescanned in memory and entire sentences are scored in terms of position and term importance. The highest scoring sentences are extracted in their entirety as the summary of the text. A disadvantage of this sort of summarization is that it cannot be done on a page by page basis without having to read in an entire document.
Sager, N.,
Natural Language Information Processing—A Computer Grammar of English and Its Applications
, Reading, Mass.: Addison-Wesley, 1981, 7-16 and 253-255, describes a technique for teaching a second language that applies a string excision method starting at the end of a sentence and moving leftward. The method excises one word or a word sequence from the sentence if the residue is again a grammatical sentence; this is repeated for each successive residue until no more excisions are possible. Examples of excisions include removal of a prepositional phrase, reduction of the number of elements in a conjunction, and so forth. The excision analyses of a French sentence and its English translation proved to be remarkably similar.
The invention addresses problems that arise in automatically summarizing text, particularly problems that would affect persons with visual impairment or other persons who cannot view text. For example, a person may be driving a vehicle or performing another activity that precludes looking at text. Or a person may not have time to look at a text or to read the text in its entirety. Or lighting or display conditions may make it impossible to see a text in a printed or displayed form.
The invention addresses the problem of how to automatically summarize text in a way that retains words that are likely to indicate the meaning of the text while retaining very few words that are unlikely to indicate meaning. More specifically, the invention addresses the problem of automatically summarizing short texts, on which no statistical method would be able to work due to lack of sufficient data. Similarly, the invention addresses the problem of how to automatically summarize sentences in a principled manner so that the summarized sentences are shorter than the original ones. The invention also addresses the problem of how to automatically summarize text simply and efficiently, such as in a way that does not require creation of templates and that in principle can be performed in one pass. The invention also addresses the problem of how to automatically summarize text in a way that provides an appropriate level of brevity.
The invention alleviates these problems by providing techniques that use part-of-speech (POS) information in automatically summarizing text. Some of the techniques use the POS information to distinguish, within a group of consecutive tokens, between tokens to be removed and tokens to be retained during automatic summarization. Some of the techniques perform automatic summarization by applying a POS-based criterion selected by a user.
The invention provides a technique for automatically summarizing text in which input text data are used to obtain POS data indicating part of speech for tokens in a text. The POS data are used to obtain group data indicating groups of consecutive tokens and indicating, within each group, any tokens that meet a POS based removal criterion. The group data are then used to obtain a summarized version of the text in which tokens that meet the removal criterion have been removed, thus reducing the number of tokens.
The group data can indicate more than one group type, and each group type can have a respective removal criterion. For example, the group data can indicate first and second word group types, and first and second POS based removal criteria can be applicable to the first and second word group types, respectively. For example, the types can include verb group types, noun group types, prepositional phrase group types, and a subclause group type (which might include other groups), and each group can be preceded and followed by elements indicating the group's type. Within each group of each type, the group data can indicate tokens that meet the applicable removal criterion.
The input text can be tokenized to obtain tokenized sentences, and POS data can be obtained for each tokenized sentence. The sentence's POS data can then be used to obtain group data for the sentence, which can in turn be used to summarize the sentence.
The input text can be obtained by converting image data to machine readable text data representing text matter contained by an image bearing portable medium. The summarized text can be converted to audio data representing the pronunciation of words in the summarized text, and corresponding sounds can be emitted, thus providing an audio summary of the text.
The invention also provides a technique for automatically summarizing text in which a signal from a user input device selects one of a set of POS based removal criteria. The input text data are used to obtain POS data indicating part of speech for tokens in a text, and the POS data are used to obtain a summarized version of the text in which tokens are removed in accordance with the selected POS based criterion, thus reducing the number of tokens.
To obtain the signal selecting the criterion, an image showing the set of POS based removal criteria can be displayed to allow interactive selection or a signal may be obtained based on the position of a manual knob that indicates the criterion. As above, the summarized text can be converted to audio data representing the pronunciation of words in the summarized text, and corresponding sounds can be emitted, thus providing an audio summary of the text.
Each of the above techniques can be implemented in a system that includes input text data and a processor that automatically summarizes text. Furthermore, each technique can be implemented in an article of manufacture that includes instruction data stored by a storage medium, indicating instructions that a system's processor can execute in automatically summarizing text.
The invention provides techniques that are advantageous because they can reduce the length of a text while retaining the meaning, thus reducing the time needed to perform text-to-speech synthesis or other operations that depend on having a summarized version of text. The invention can be implemented with a light syntactic parser to identify which parts of the text can be eliminated. The elimination can be graduated under user control, possibly via a knob, so that more or less of the text is retained. In the extreme case only the important nouns or proper names are retained.
The invention would allow a blind reader to audibly scan
Edouard Patrick N.
Xerox Corporation
LandOfFree
Text summarization using part-of-speech does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Text summarization using part-of-speech, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Text summarization using part-of-speech will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2520024