Patent Application
20100318542
The present invention relates generally to classifying content and in particular, to a method and apparatus for classifying electronic content.
Oftentimes electronic content is classified so that a user may determine the type of content. For example, an electronic program guide (EPG) may display a particular program as being classified as a “comedy”. In addition, recommender systems and search engines may exploit content classification in order to match, find, and/or rank relevant content.
In the domain of television programming, where content is often accompanied by EPG metadata, content has often been classified by associations between various aspects of this metadata. One problem with such classification efforts is that textual descriptions of content in metadata are often very sparse yet highly dimensional, and therefore often are of little utility for classifying content. A related problem is that one of the most highly informative aspects of metadata, those which indicate the genre or category, are typically underutilized with respect to other metadata and content features. A third problem is that there is an ever increasing amount of content with little or no metadata, making the classification task more difficult.
There exist a number of previous efforts to resolve these problems in personalization and recommender systems in the prior art, however, none successfully infer an arbitrary amount of additional relationships between metadata by making use of linguistic content. For example, in U.S. Pat. No. 7,243,085 B2, “Hybrid personalization architecture” a probabilistic network is constructed from metadata and linguistic content where nodes in the network are viewed as concepts in an ontology and the edges connecting the nodes are associated with weights indicating the strength of the relationship between the concepts. These edges and weights are in part derived from the relationship between metadata and linguistic content. This approach falls short of solving the general problem of inferring an arbitrary amount of relationships between aspects of metadata, however, because it only allows for pairwise relationships between the concepts. The method identified in the present invention solves this and the three aforementioned problems. Therefore a need exists for a method and apparatus for classifying content that more appropriately utilizes content metadata or aspects of the content itself and provides a more accurate classification of the content.
Skilled artisans will appreciate that elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale. For example, the dimensions and/or relative positioning of some of the elements in the figures may be exaggerated relative to other elements to help to improve understanding of various embodiments of the present invention. Also, common but well-understood elements that are useful or necessary in a commercially feasible embodiment are often not depicted in order to facilitate a less obstructed view of these various embodiments of the present invention. It will further be appreciated that certain actions and/or steps may be described or depicted in a particular order of occurrence while those skilled in the art will understand that such specificity with respect to sequence is not actually required. Those skilled in the art will further recognize that references to specific implementation embodiments such as “circuitry” may equally be accomplished via replacement with software instruction executions either on general purpose computing apparatus (e.g., CPU) or specialized processing apparatus (e.g., DSP). It will also be understood that the terms and expressions used herein have the ordinary technical meaning as is accorded to such terms and expressions by persons skilled in the technical field as set forth above except where different specific meanings have otherwise been set forth herein.
In order to alleviate the above-mentioned need, a method and apparatus for classifying content is provided herein. During operation the natural language existing in metadata and/or in the program content itself is used to infer finer-grained distinctions for television program genres/categories. In accomplishing this task, the occurrences of natural language words are tracked with category labels such as genre (supplied by and/or inferred from the metadata or natural language existing in the content), and then used to produce fine-grained relationships between the genres, to a particular level of precision.
More particularly, natural-language words are associated with each program. The natural-language words are identified from, for example, metadata and/or the actual program itself. Each word identified for a program is associated with the identified genre of the program (from, for example, its tagged metadata). A database is then maintained having a number of occurrences of each word from the multiple programs for each genre. Once the word/genre database is created, subgenres for a particular program can be created by once again using the words identified for that program to rank the most appropriate genres for the words (i.e., the genres having the most occurrences for that word). A program here can be understood to mean any type of content which may contain or be associated with (i.e., via metadata) natural language, such as a television program, movie, video, etc.
The above technique can also be extended such that the words used for ranking the genres are a subset of the words identified for the program, based, for example, on the importance of such words in the program. Similarly, the technique can be extended such that sets of words, rather than single words, are use for the ranking. Further, the technique can be extended such that the items used for the ranking need not be the words from the program, but rather other words or representations that are associated with the words or sets of words from the program. Similarly, the technique can be extended such that criteria other than word frequency are used to rank the most appropriate genre. For example, the ranking criteria could be the probability of the word, the deviation from an expected probability, or the term frequency-inverse document frequency weighting, where different items can alternatively be used as for the basis of document frequency such as programs or genres themselves, etc. Such criteria can be derived with data from the collection of programs and/or from sources external to the programs, such as corpora from related domains. For example, statistics regarding word frequency, genre frequency, and co-occurrence can be derived from additional corpora (e.g., the internet), to aide in the determination of probabilities to be used in the selection of words related to a particular program.
The above technique uses both the textual content in the metadata and/or the linguistic content in the programs themselves to make finer-grained distinctions of program categories/genre. This improved ranking allows for a number of improvements to applications such as better relevancy ranking for search, improved clustering, and better recommendations.
The present invention encompasses a method for classifying content. The method comprises the steps of identifying a particular program in order to determine a genre or category for the program, creating a list of words associated with the program, and accessing a database comprising stored words and their associated genres or categories for each word. The genre or category is then determined for the program based on a comparison of the list of words with the stored words and their associated genres or categories.
The present invention additionally encompasses a method for classifying content. The method comprises the steps of:
creating a database by:
determining a genre or category for a second program by:
The present invention additionally encompasses an apparatus comprising a database comprising stored words and their associated genres or categories for each word, and logic circuitry identifying a particular program in order to determine a genre or category for the program, creating a list of words associated with the program, accessing the database, and determining the genre or category for the program based on a comparison of the list of words with the stored words and their associated genres or categories.
Turning now to the drawings, where like numerals designate like components,
As shown, apparatus 100 comprises an electronic processor 101 and storage 102. Processor 101 comprises logic circuitry such as a digital signal processor (DSP), general purpose microprocessor, a programmable logic device, or application specific integrated circuit (ASIC) and is utilized to create the contents of storage 102 and to classify content. Storage 102 comprises standard random access memory and is used to store information that can be textually searched.
During operation of apparatus 100, program metadata and/or program content is received by processor 101. The metadata and/or program content may be from an electronic program guide, from a textual transcript of the program (e.g., via closed-captioned or automatic speech recognition of natural language components of the program), from an online content or metadata service, and/or any other means for providing content to processor 101. In response, processor 101 will populate storage 102 and output classification results/genres for a particular program. As discussed above, processor 101 will create a database having a number of occurrences of each word from the multiple programs for each genre. Once the word/genre database is created, sub-genres for a particular program can be determined by once again using (a subset of) the words identified for that program to rank the most appropriate genres for the words (i.e., the genres having the most occurrences for that word or genres which are most probable as determined by criteria other than frequency).
As discussed above, a database is created by processor 101 and stored in storage 102. In creating the database, processor 101 utilizes multiple programs and determines their identified genre (e.g., from metadata about each program, or via processing of the natural language taken directly from the content of the program to determine top-level genres for the content when there is no metadata supplied). A list of words is then created for each program by processor 101. As discussed above, the list of words for each program is created by words that are identified from, for example, metadata and/or the actual program itself. More particularly, the list of words (or word sequences) can be optionally preprocessed and normalized using various established techniques from the field of natural language processing, which are helpful in reducing the number of items to consider (thus reducing the dimensionality). These techniques include the removal of certain less informative high-frequency stop words (e.g., “the”, “it); the removal of certain punctuation, dates, symbols, numbers, etc. (depending on the type of application); stemming (the process of reducing various forms of a word to a base or root form); normalizing the case, segmentation, and so on. In addition, the preprocessed and normalized words can be optionally further reduced to a subset of words (or word sequences) which are most representative of the program. This identification of the most representative words can again be done by any of various well-known techniques from the field of natural language processing, such as via keyterm extraction.
Once processor 101 has created a normalized list of words for a given program, processor 101 then counts the occurrence of each word with the program's identified genre and creates, or adds information to a database containing the words/genre and the number of occurrences for each word for the particular genre. This database is illustrated in
As shown in
In order to account for common words, the database may be adjusted for frequency of usage for particular words. The adjustment may be based on (deviations from) expectations of counts as predicted by models of the domain and/or the language in general, etc.
It should be noted that the database held in storage 102 comprises many words from many different programs. The combined results of all analyzed programs are then used to further classify content (described below).
Once the database comprising the word/genre matrix is created, it can then be used to better categorize content such as television shows, internet video, . . . , etc. In order to do so, the content is analyzed by processor 101 to determine a list of words describing the content (as described above). Once a list of words describing the content is determined, the database in storage 102 is accessed by processor 101 in order to determine the different genres associated with each word. For example, referring to
The amount of genres produced may vary. For example, for a given set of words, and a set of, 50 genres, g1, g2, . . . g50, only the top-two ranked genres may be used (e.g., g17 and g31). Alternatively the genre names can be combined together to produce a single, new genre g17_g31 (e.g., outdoors_sporting_events).
Once processor 101 has created a normalized list of words for a given program, processor 101 then counts the occurrence of each word with the program's identified genre (step 307) and creates, or adds information to a database containing the words/genre and the number of occurrences for each word for the particular genre (step 309). Finally, at step 311, processor 101 determines if any other programs need to be analyzed and added to the database, and if so, the logic flow returns to step 301, otherwise the logic flow ends at step 313.
The above technique creates a database that can then be utilized by processor 101 to make finer-grained distinctions of program categories/genre. This improved distinction in categories/genres allows for a number of improvements to applications such as better relevancy ranking for search, improved clustering, and better recommendations.
At step 403 processor 101 creates a list of words most representative of the program. This list may be from metadata associated with the program, or from the actual content of the program itself. As discussed above, the creation of the list may include various preprocessing steps for normalization and the removal of less informative words as discussed above.
Once processor 101 has created a list of words for a given program, processor 101 then accesses storage 102 to determine the different genres associated with each word (step 405). As discussed above, storage 102 comprises a database comprising stored words from multiple programs and their associated genres or categories for each word. At step 407 processor 101 then determines a fine-grained genre or category for the program based on a comparison of the list of words with the stored words and their associated genres or categories. The finer-grained genres or categories may then be output by processor 101.
As discussed, the word/genre combinations having the highest number of occurrences (or ranked highest by a different criteria, as described earlier) are used by processor to determine a listing of genres that identify the program. The determined genre(s) is then output from processor 101, and may be used for a number of improvements to applications such as better relevancy ranking for search, improved clustering, and better recommendations. The amount of genres produced may vary. For example, for a given set of words, and a set of, 50 genres, g1, g2, . . . g50, only the top-two ranked genres may be used (e.g., g17 and g31). Alternatively the genre names can be combined together to produce a single, new genre g17_g31 (e.g., outdoors_sporting_events).
It should be noted that in the flow chart of
While the invention has been particularly shown and described with reference to a particular embodiment, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention. For example, the ‘words’ in the matrix can be extended to any text processing unit or combination thereof, such as sequences, ngrams, POS tags, etc., as well as mapped to smaller or constrained units (e.g., synonyms, etc.), or mapped to units of meaning (e.g., ontologies, etc.) so that generalization can be increased. Additionally, the genre and subgenre relationships learned on content with metadata can be used to infer likely categories (genres and subgenres). Additionally, additional non-linguistic metadata and non-metadata features (e.g., show running time) can be used in conjunction with these genre related features to categorize, cluster, and rank programs. Similarly, the domain of content to be classified can be extended from television or video to any sort of content for which there may be metadata or a natural language representation of the content (e.g., internet content, electronic documents, etc.). Also, the criteria used for determining ranking can be something different from or in addition to simple word frequency, for example, it can take into account the expected frequency of the given word or term in the given domain, based on statistics, the programs themselves, or from other sources. It is intended that such changes come within the scope of the following claims:
