Method and apparatus for generation and augmentation of search terms from external and internal sources

Abstract

A method and apparatus to identify names, personalities, titles, and topics that are present in a repository, and place them into a grammar, and to identify names, personalities, titles, and topics that are not present in the repository, and place them into a grammar, uses information from external data sources, notably the text used in non-speech, text-based searches, to expand the search terms entered into the ASR grammars. The expansion takes place in two forms: (1) finding plausible linguistic variants of existing search terms that are already comprehended in the repository, but that are present under slightly different names; and (2) expanding the existing search term list with items that should be there by virtue of their currency in popular culture, but which for whatever reason have not yet been reflected with content items in the repository.

Description

BACKGROUND OF THE INVENTION

1. Technical Field

The invention relates to speech recognition and speech directed device control. More particularly, the invention relates to a method and apparatus for the generation and augmentation of search terms from external and internal sources, in connection with speech recognition and speech directed device control.

2. Description of the Prior Art

One area of technical innovation is that of navigation of content by spoken and textual command. Such systems typically perform speech recognition by use of a grammar-based ASR (automatic speech recognition) system, where the grammar defines those terms that can be recognized. In such systems, navigated content is comprised of a catalog, content data base, or other repository, for example: currently airing broadcast TV programs, contents of a video-on-demand (VOD) system, a catalog of cell phone ring tones, a catalog of songs, or a catalog of games. Hereafter all of the above sources of content are referred to as a repository.

Content sources are updated and/or expanded on occasion, possibly periodically, possibly as frequently as daily. In some such applications as those described above, content sources are assumed, by both system architects and by system users, to reflect trends and interests in popular culture. However, known recognition systems are limited to recognition of only those phrases that are listed in grammar. Nonetheless, it is desirable to make content sources searchable by names of artists, popular topics, personalities, etc. Yet known ASR systems recognize only those elements that are listed in grammar.

It would be desirable to identify names, personalities, titles, and topics that are present in a repository, and place them into a grammar. It would also be desirable to identify names, personalities, titles, and topics that are not present in the repository, and place them into a grammar; for in this way, such names, personalities, titles and topics may at least be recognized by the ASR system, which can then report that no suitable content is present in the repository.

SUMMARY OF THE INVENTION

The presently preferred embodiment of the invention provides a method and apparatus to identify names, personalities, titles, and topics that are present in a repository, and place them into a grammar. A further embodiment of the invention provides a method and apparatus to identify names, personalities, titles, and topics that are not present in the repository, and place them into a grammar. A key aspect of the invention uses information from external data sources, notably non-speech, text-based searches, to expand the search terms entered into the ASR grammars. The expansion takes place in two forms: (1) finding plausible linguistic variants of existing search terms that are already comprehended in the repository, but that are under slightly different names; and (2) expanding the existing search term list with items that should be there by virtue of their currency in popular culture, but which for whatever reason have not yet been reflected with content items in the repository.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block schematic diagram showing search term generation flow according to the invention.

DETAILED DESCRIPTION OF THE INVENTION

The presently preferred embodiment of the invention provides a method and apparatus to identify names, personalities, titles, and topics that are present in a repository, and place them into a grammar. A further embodiment of the invention provides a method and apparatus to identify names, personalities, titles, and topics that are not present in the repository, and place them into a grammar. A key aspect of the invention uses information from external data sources, notably non-speech, text-based searches, to expand the search terms entered into the ASR grammars. The expansion takes place in two forms: (1) finding plausible linguistic variants of existing search terms that are already comprehended in the repository, but that are present under slightly different names; and (2) expanding the existing search term list with items that should be there by virtue of their currency in popular culture, but which for whatever reason have not yet been reflected with content items in the repository.

An exemplary embodiment of the invention operates as follows:

First, extract search term candidates, also referred to as candidate search terms, from external sources, for instance:

• 1. Published lists of frequent textual searches against popular search engines, e.g. Yahoo “top searches;”
• 2. Published lists of popular artists and songs, e.g. music.aol.com/songs/newsongs “Top 100 Songs;”
• 3. Published lists of popular tags, e.g. ETonline.com “top tags;”
• 4. Published lists of most-emailed stories, e.g. NYtimes.com most emailed stories, ETonline.com most emailed stories; and
• 5. Published news feeds, such as RSS feeds, e.g. NYtimes.com/rss.

Nominally for the first three sources listed above, the candidate search terms are clearly identified as an explicitly marked title, author, artist name, etc. and, hence, processing is purely automatic. For the final two sources listed above, a combination of automatic means, such as named entity extraction (NEE) and/or topic detection and tracking (TDT) methods, and possibly direct human intervention, are applied to the running text or titles to generate candidate search terms. However, human intervention may be used with the first group as well.

Next, extract verified search terms from internal sources, for instance:

• 1. Explicitly marked titles, authors, artist names, etc. that are associated to the content elements in the repository; and/or
• 2. Sources derived by application of named entity extraction (NEE) and/or topic detection and tracking (TDT) methods to descriptive text associated to the content elements in the repository.

EXAMPLES

• • Use of the topic “california fires”, appearing as the tenth-most-popular searched item, as listed in the “MOST POPULAR SEARCHED” section of the website nytimes.com of Oct. 27, 2007.
  • Extraction of the proper name “David Brooks” from the frequently emailed article title “David Brooks: The Outsourced Brain,” appearing as the second-most-popular emailed article, as listed in the “MOST POPULAR EMAILED” section of the website nytimes.com of Oct. 27, 2007.

In the presently preferred embodiment of the invention, typical (although not exclusive) means of NEE and TDT analysis may be found in:

• Foundations of Statistical Natural Language Processing, by Chris Manning and Hinrich Schütze, MIT Press. Cambridge, Mass.: May 1999.
• Proceedings of the DARPA Broadcast News Transcription and Understanding Workshop, Feb. 8-11, 1998, Lansdowne Conference Resort, Lansdowne, Va., available at URL www.nist.gov/speech/publications/darpa98/

Next, match candidate search terms against verified search terms by well-known edit distance techniques, to obtain plausible linguistic variants of verified search terms, used to generate the augmented verified search terms.

• • Example: “Mary J. Blige” is initial verified search term, augmented with “Mary Blige” as a variant.

Finally, by virtue of their high incidence count, repeated appearance in history as either a candidate or verified search term, or other criterion, include in the grammar candidate search terms which do not point to actual content elements, but which the ASR system should nevertheless recognize. We refer to such elements as “null search terms.”

In FIG. 1, a grammar is augmented with regard to external sources 11 and internal sources, e.g. the repository, both as discussed above.

External sources comprise, for example, explicitly marked information 12 and running text 15. Explicitly marked text may be subject to an optional count filtering process 14, providing incidence count information is available, whereby only those instances with sufficiently high incidence count are retained, while running text is processed, as discussed above, with a module 17 that performs, for example, named entity extraction (NEE) or topic detection and tracking (TDT). The data from all external sources is combined by a module 18 and an output, comprising candidate search terms (C[i]) 19 is generated. The combined output from external sources is further processed by a module 22 that performs such functions as incidence counting, low pass filtering, and other functions as desired, and is also passed to an approximate text matching module 33 (discussed below). This module 22 also receives historical information, such as a history of candidate search terms (C[i−1] . . . ) 26, a history of final search terms (S[i−1] . . . ) 21, and verified search terms (discussed in greater detail below). The output of the module 22 is provided to a further module 23, which identifies null search terms (N[i]), as discussed above.

Internal sources comprise, for example, explicitly marked information 27 and running text 28. Explicitly marked text may be subject to an optional count filtering process 29, whereby only those instances with sufficiently high incidence count are retained, while running text is processed, as discussed above, with a module 30 that performs, for example, named entity extraction (NEE) or topic detection and tracking (TDT). The data from all internal sources is combined by a module 31 and an output, comprising verified search terms (V[i]) 32 is generated. The verified search terms are used in connection with the module 22, as discussed above. The verified search terms are also provided to a module 33 for approximate text matching. The module 33 also receives candidate search terms from the module 19 as an input. The output of the module 33 is provided to a module 34 that generates augmented verified search terms (AV[i]).

The processed external sources information that is output by the module 23 and the processed internal sources information that is output by the module 34 are provided as inputs to a combining module 34 to produce final search terms (S[i]) 25, which are output for grammar generation.

Although the invention is described herein with reference to the preferred embodiment, one skilled in the art will readily appreciate that other applications may be substituted for those set forth herein without departing from the spirit and scope of the present invention. Accordingly, the invention should only be limited by the Claims included below.

Claims

1. A method for identifying names, personalities, titles, and topics, whether or not said names, personalities, titles and topics are present in a given repository, and for placing them into a grammar for use in an automatic speech recognition (ASR) system, comprising the steps of: extracting search term candidates from published lists of the text of frequent searches presented to popular text-based search engines, published lists of popular artists and song titles, published lists of most popular tags, published lists of most-emailed stories, and published news feeds, the step of extracting further comprising:automatically identifying explicitly marked candidate search terms from at least one structured published list of content; andextracting candidate search terms from unstructured published content by performing an extraction means selected from among: available named entity extraction (NEE); topic detection and tracking (TDT); direct human intervention; and a combination of NEE, TDT, and direct human intervention;storing said candidate search terms in a historical database of candidate search terms;storing a history of said extracted search term candidates;extracting verified search terms from internal sources of said repository;matching candidate search terms against verified search terms by edit distance techniques to obtain plausible linguistic variants of verified search terms;using said linguistic variants to generate augmented verified search terms;storing a history of said augmented verified search terms;establishing a set of null search terms comprising candidate search terms having a threshold incidence count in said history of said extracted search term candidates and in said history of said augmented verified search terms; andexpanding said grammar by adding said set of null search terms to said grammar of said automatic speech recognition (ASR) system.

2. The method of claim 1, said internal sources comprising explicitly marked titles, authors, artist names, etc. that are associated to content elements in said repository.

3. An apparatus for identifying names, personalities, titles, and topics, whether or not said names, personalities, titles and topics are present in a given repository, and for placing them into a grammar, comprising: a plurality of external data sources, comprising non-speech, published lists of the text of frequent searches presented to popular text-based search engines, published lists of popular artists and song titles, published lists of most popular tags, published lists of most-emailed stories, and published news feeds;means for extracting search term candidates from said external sources, wherein search term candidates are either explicitly marked candidates or extracted candidates, the step of extracting further comprising:automatically identifying explicitly marked candidate search terms from at least one structured published list of content from among a plurality of structured lists of content available over a computer network, wherein said structured published lists of content are organized by an attribute selected from among a group of attributes consisting of: popular search engine search terms, popular artists, popular songs, and popular news feed tags; andextracting candidate search terms from at least one document from among a plurality of documents available from a plurality of sources of unstructured published content available over a computer network, wherein said sources of unstructured published content at least includes sources selected from among a group of sources consisting of published lists of most-emailed stories and published news feeds, and wherein extracting further comprises an automatic extraction means selected from among: named entity extraction (NEE);topic detection and tracking (TDT);direct human intervention; anda combination of NEE, TDT, and direct human intervention;storing said candidate search terms in a historical database of candidate search terms;means for extracting verified search terms from said internal sources;means for expanding search terms entered into one or more automatic speech recognition ASR grammars by using information from said external data sources, said means for expanding search terms comprising means for matching candidate search terms against verified search terms by edit distance techniques to obtain plausible linguistic variants of verified search terms and further comprising any of:means for finding plausible linguistic variants of existing search terms that are already comprehended in the repository, but that are under slightly different names; andmeans for expanding an existing search term list with items that should be in said list by virtue of their currency in popular culture, but which for whatever reason have not yet been reflected with content items in the repository;means for using said linguistic variants to generate augmented verified search terms;means for storing said augmented verified search terms in a historical database of verified search terms;means for establishing a set of null search terms comprising candidate search terms having a high incidence count in said historical database of candidate search terms and in said historical database of verified search terms; andmeans for expanding said grammar by adding said set of null search terms to said grammar of said automatic speech recognition (ASR) system.

4. The apparatus of claim 3, said internal sources comprising any of: explicitly marked titles, authors, artist names, etc. that are associated to content elements in said repository.

5. The method of claim 1, said internal sources comprising: sources obtained by application of named entity extraction (NEE) and/or topic detection and tracking (TDT) methods to descriptive text associated to content elements in said repository.

6. The apparatus of claim 3, said internal sources comprising: sources obtained by application of named entity extraction (NEE) and/or topic detection and tracking (TDT) methods to descriptive text associated to content elements in said repository.

7. A method for identifying names, personalities, titles, and topics, whether or not said names, personalities, titles and topics are present in a given repository and for placing them into a grammar for use in an automatic speech recognition (ASR) system, comprising the steps of: extracting search term candidates from external sources, wherein search term candidates are either explicitly marked candidates or extracted candidates, the step of extracting further comprising:automatically identifying explicitly marked candidate search terms from at least one structured published list of content from among a plurality of structured lists of content available over a computer network, wherein said structured published lists of content are organized by an attribute selected from among a group of attributes consisting of: popular search engine search terms, popular artists, popular songs, and popular news feed tags; andextracting candidate search terms from at least one document from among a plurality of documents available from a plurality of sources of unstructured published content available over a computer network, wherein said sources of unstructured published content at least includes sources selected from among a group of sources consisting of published lists of most-emailed stories and published news feeds, and wherein extracting further comprises an automatic extraction means selected from among: named entity extraction (NEE);topic detection and tracking (TDT);direct human intervention; anda combination of NEE, TDT, and direct human intervention;storing said candidate search terms in a historical database of candidate search terms;extracting verified search terms from internal sources;matching candidate search terms against verified search terms by edit distance techniques to obtain plausible linguistic variants of verified search terms that were extracted from said internal sources;using said linguistic variants to generate augmented verified search terms;storing said augmented verified search terms in a historical database of verified search terms;establishing a set of null search terms comprising candidate search terms having a high incidence count in said historical database of candidate search terms and in said historical database of verified search terms; andexpanding said grammar by adding said set of null search terms to said grammar of said automatic speech recognition (ASR) system.