WO2006093593A1

WO2006093593A1 - Apparatus and method for generating a personalised content summary

Info

Publication number: WO2006093593A1
Application number: PCT/US2006/002515
Authority: WO
Inventors: Paola Hobson; Michael Brady; Catherine Mary Dolbean
Original assignee: Motorola, Inc.
Priority date: 2005-02-21
Filing date: 2006-01-19
Publication date: 2006-09-08
Also published as: GB2423383A; GB0503503D0

Abstract

An apparatus (100) is provided for generating a personalised content summary for a content item which comprises a plurality of events. The apparatus comprises a rating processor (103) which determines a rating of events of the content item in response to associations between events and content summaries for a first group of training content items. The training content items which are included in the first group are selected by a training data selection processor (107) in response to a user preference profile (109). Specifically, content items matching the user's preferences are selected for the first group. A selection processor (111) selects events for inclusion in the content summary in response to the rating of the events and a summary generator (113) generates the personalised the summary by including summary items associated with the selected events. The invention may allow improved personalisation by biasing training data in response to user preferences.

Description

APPARATUS AND METHOD FOR GENERATING A PERSONALISED

CONTENT SUMMARY

Field of the invention

The invention relates to an apparatus and method for generating a personalised content summary and in particular to automatic generation of personalised content summaries for content items.

Background of the Invention

In recent years, the availability and provision of for example multimedia and entertainment content has increased substantially. For example, the number of available TV and radio channels has increased substantially and the popularity of the Internet has provided new content distribution means. Consequently, users are increasingly provided with a plethora of different types of multimedia content from different sources. In order to identify and select the desired content, the user must typically process large amounts of information which can be very cumbersome and impractical.

Accordingly, significant resources have been invested in research into techniques and algorithms that may provide an improved user experience and assist a user in identifying and selecting content. One such area of research is information filtering. Information filtering tackles the problem of information overload, which is a problem that is becoming more and more pressing as users are confronted with increasing volumes of e.g. multimedia data (including e.g. text, audio and video content items), much of which is either unwanted or irrelevant. Accordingly, information filtering may provide functionality for selecting information which is of particular interest to the user. Furthermore, information filtering may include generation of new information that extracts, condenses or modifies available information in order to provide a more suitable information provision to the user.

Some of the problems that must be addressed by algorithms for information filtering include:

1) How to determine the importance of different pieces of information to the specific user;

2) How to allocate resources (e.g. a summary time limit) such that the user is provided with the most salient information;

3) How to provide a context of the pieces of information that are presented to the user; and

4) How to design the system in such a way as to allow domain-independent summarisation.

Although many algorithms for information filtering have been proposed these tend to be suboptimal in one or more of these areas.

Generation of summaries for content items may be considered a subset of information filtering. Personalisation of summaries is advantageous in many applications and is typically based on content based filtering, wherein filtering is based on properties of the data, and collaborative filtering, wherein the preference-behaviour and qualities of other users are exploited in predicting the preferences of a particular user.

Many algorithms for generating content summaries are based on applying predefined rules, criteria or axioms to the content items. However, such algorithms tend to be very inflexible and are typically not suited for different types of content items or applications. Furthermore, such methods tend to be complex and require accurate models, rules, criteria or axioms to be derived. However, the derivation of such

CML02351EV models is very difficult and time consuming and tends to result in models which are inaccurate for most content items. It furthermore results in suboptimal personalisation as the models or axioms may not be suited for accurate personalisation.

In order to provide an enhanced user experience, it is highly desirable that summaries are personalised for the individual user or group of users. The traditional method for personalising a multimedia summary using content-based filtering is to assign a weight to each of the user's preferences. The weights are then used to vary the scores of the multimedia content entities so that a filtering agent can determine which content should be included in the personalised summary. An example of such an approach is described in A. M. Ferman, J. H. Errico, P. van Beek and M. I. Sezan "Content-based filtering and personalization using structured metadata", Proceedings of the Joint Conference on Digital Libraries, p 393. 13-17 July 2002, Portland, Oregon.

An example of a personalised multi-document summarisation and recommendation system is presented in D. R. Radev, W. Fan and Z. Zhang "WeblnEssence: A Personalized Web-Based Multi-Document Summarization and Recommendation System" Workshop on Automatic Summarization, North American Chapter of the Association for Computational Linguistics, NAACL¹Ol, June 2001, Pittsburgh, USA. This system allows weighted retrieval based on the positions of certain words in a web page. For example, a user can specify in their profile that they would like to give a specific weight for a keyword appearing in the title, another weight for the anchor, and yet another weight for the body. The user profile also contains a field for the type of search the user wants to perform, such as Boolean or Vector Space search. The summary length can be selected as a percentage size of the original document, and the user can choose the ordering of sentences within the document according to various parameters, such as position, time sequence or relevance to the query.

However, such a user profile is complex and cumbersome and will be inconvenient and impractical for most users and applications. For example, it may not be clear to a

CML02351 EV user how the weightings qualitatively affect the personalisation of the content presented to them.

Automated collaborative filtering is a technique often used in recommender systems to match items to users by first matching users to each other and then using statistical algorithms to make recommendations based on correlations between personal preferences. As an example, United States Patent US6687696 describes a system wherein probabilistic latent semantic analysis is used to learn one or more statistical models based on document data and data from previous users and the current user. These models are then used to give each word in a document a score, which is added together to give the document an overall relevance score to the user. However such models are complex and require frequent updating.

Accordingly, known techniques for generating content summaries tend to be suboptimal. For example, most known methods tend to be complex, to provide no or insufficient personalisation and/or to be specific to applications or content domains (e.g. to rely on specific models for the current application or content domain).

Hence, an improved system for generating a personalised content summary would be advantageous and in particular a system allowing improved information selection, improved information presentation, improved personalisation, improved applicability to different content domains, increased flexibility, facilitated implementation, improved performance and/or an improved user experience would be advantageous.

Summary of the Invention

Accordingly, the Invention seeks to preferably mitigate, alleviate or eliminate one or more of the above mentioned disadvantages singly or in any combination.

According to a first aspect of the invention there is provided an apparatus for generating a personalised content summary for a content item comprising a plurality

CML02351EV of events; the apparatus comprising: rating means for determining a rating of events of the content item in response to associations between events and content summaries for a first group of training content items; training data selection means for selecting training content items for the first group of content items in response to a user preference profile; event selection means for selecting events for inclusion in the content summary in response to the rating of the events; and means for generating the summary by including summary items associated with the selected events.

The invention may allow improved performance for a content summary generation apparatus. For example, the invention may allow improved personalisation by biasing training data in response to a user preference profile. The biased training data may be used to implicitly bias the content generation towards the specific preferences of the user(s). The summary generation and/or personalisation may be achieved without requiring a specific model or rules to be defined. Rather, an automatic generation of a content summary based on empirical data may be used.

The invention may e.g. provide improved personalisation of content summaries thereby enhancing the user experience. The approach is e.g. suitable for different content domains thereby allowing the same functionality or algorithm to be used for different content item types and domains

The implementation of a system for generation of personalised content summaries may be facilitated. In particular, the apparatus is well suited for practical implementations and may have low complexity.

An improved personalisation of content summaries in comparison to many automated systems may be achieved.

The training data may include the training content items themselves or may comprise only the data related to the content items and required for generating the content summary.

CML02351 EV The rating means may in particular determine a rating of an event in response to associations between an event class of the event and event classes of the first group of training content items. In the description, the term event may relate to an event instance or an event class.

According to an optional feature of the invention, the training data selection means is arranged to determine a match indication between the user preference profile and a plurality of training content items and to select content items for the first group of content items in response to the match indication. This may provide an efficient and/or low complexity approach to biasing the training data. It may additionally or alternatively result in an improved personalisation of the generated content item.

According to an optional feature of the invention, the training data selection means is arranged to include a content item in the first group of content items if the match indication is above a threshold.

This may provide an efficient and/or low complexity approach to biasing the training data. It may additionally or alternatively result in an improved personalisation of the generated content item.

The match indication may have an increasing value for an increasing match between the preferences of the user preference profile and the contents of the content item. The threshold may be variable and may in particular be determined in response to selection of other content items for the first group of content items. For example, the threshold may be modified to ensure that a desired number of content items are included in the first group.

According to an optional feature of the invention, the training data selection means is arranged to determine the match indication for a first content item in response to an occurrence in the first content item of events matching at least one individual user preference of the user preference profile.

CML02351EV This may provide an efficient, low complexity, accurate and/or high performance approach to selection of content items for the first group. The user preference profile may specifically comprise a plurality of individual user preferences related to specific events and the correlation between these events of the user preference profile and the events in the first content item may provide a suitable indication of the desirability of including the content item.

According to an optional feature of the invention, the training data selection means is arranged to determine the match indication for a first content item in response to a number of occurrences of events matching individual user preferences of the user preference profile. This provides an efficient, high performance and/or low complexity approach to selection of content items for the first group.

According to an optional feature of the invention, the training data selection means is arranged to determine the match indication for a first content item in response to a weighted sum of the number of occurrences of events matching individual user preferences of the user preference profile.

This may provide an efficient, high performance and/or low complexity approach to selection of content items for the first group. All weights may in particular be equal resulting in an improved applicability to different content domains.

According to an optional feature of the invention, the training data selection means is arranged to determine the match indication for the first content item substantially as

N

Match = ∑w_t * frequency (U ) E S)

1=1

where i is an index of individual user preferences of the user preference profile, N is the number of individual user preferences in the user preference profile, Wi is a weight for individual user preference i and U; denotes the i'th individual user preferences and S denotes events of the first content item.

CML02351 EV This provides an efficient, high performance and/or low complexity approach to selection of content items for the first group.

According to an optional feature of the invention, weights for the weighted sum are comprised in the user preference profile. This may allow facilitated implementation.

According to an optional feature of the invention, the training data selection means is arranged to determine the match indication for a first content item in response to an occurrence of events matching at least one individual user preference of the user preference profile in a content summary for the first content item

This may provide an efficient, high performance and/or low complexity approach to selection of content items for the first group.

All the approaches described above for selecting content items for the first group in response to a number of occurrences of events in the content items may equally be applied to selecting content items for the first group in response to a number of occurrences of events in the content summaries.

According to an optional feature of the invention, the training data selection means is arranged to select training content items in response to user preference profiles associated with the training content items. This may improve performance and may in particular allow improved personalisation.

According to an optional feature of the invention, the user preference profile is a multi user preference profile. For example, the user preference profile may relate to a group of users having similar preferences. This may facilitate implementation and/or reduce complexity.

According to an optional feature of the invention, the apparatus further comprises: grouping means for grouping sequential events of the plurality of events into event

CML02351EV groups; and the rating means is arranged to determine a rating for a first event group of the event groups in response to a frequency characteristic of the training data, the frequency characteristic being indicative of a frequency of inclusion of training summary items related to an event sequence of the first event group in training content summaries of the first group of content items.

This may allow an improved generation of a content summary and/or may provide an enhanced user experience. The approach is particularly suitable for personalisation of the content summaries by biasing of the training data.

Specifically, the summary generation may be achieved without requiring a specific model, rule, criterion or axiom to be defined. Rather, an automatic generation of a content item summary based on empirical data may e.g. be achieved. Furthermore, the training data may be biased to implicitly personalise the generated content summary to the preferences of the user preference profile.

The apparatus may for example generate a variable length content summary. A content summary of different lengths may be generated by the same algorithm and may be based on the same training data.

The training content summaries may for example have been generated by manually or semi-automatically selecting event sequences of the training content items for which content summary items are included. For example, the content summaries may have been generated manually for the content items and the invention may allow an automatic generation of a content summary which resembles the characteristics of the manual generation. This may further be achieved without necessitating that criteria or principles of the manual generation are defined or even considered.

According to an optional feature of the invention, the frequency characteristic comprises a frequency indication indicative of a number of times training summary items for the event sequence is included in the training content summaries relative to a number of times the event sequence occurs in the first group of training content items.

CML02351 EV This may e.g. allow an efficient system, improved content summaries and/or facilitated implementation. In particular, a low complexity algorithm may allow content summary generation which corresponds to the characteristics for the training data and in particular the content summary may reflect the underlying principles and criteria used when generating the training content summaries.

According to an optional feature of the invention, the rating means is arranged to determine ratings by applying a Markov chain to event subsets of the event groups. This may e.g. allow a low complexity implementation while achieving accurate performance.

According to an optional feature of the invention, the rating means is arranged to divide at least a first event group into subsets of events and to determine a frequency indication for the first event group in response to frequency indications for each of the subsets.

This may facilitate implementation and reduce the complexity of the apparatus. In particular, it may facilitate the rating of groups as training data frequency information is required only for shorter event sequences rather than for the whole event group.

According to an optional feature of the invention, the grouping means is operable to group events in event groups in response to an event sequence probability.

This may provide an efficient and low complexity implementation. In particular, accurate grouping of associated sequential events suitable for combined decision on whether to include these in the content summary may be enabled or facilitated. The grouping may in particular provide improved context information for the individual events.

According to an optional feature of the invention, the grouping means is arranged to determine the event sequence probability for a first event group by applying a Markov

CML02351EV chain to event subsets of the first event groups. This may e.g. allow a low complexity implementation while achieving accurate performance.

According to an optional feature of the invention, the selection means is arranged to weight the ratings in response to the user preference profile. This may allow an improved personalisation of the content summary.

According to a second aspect of the invention, there is provided a method of generating a personalised content summary for a content item comprising a plurality of events; the method comprising: determining a rating for events of the content item in response to associations between events and content summaries for a first group of training content items; selecting training content items for the first group of content items in response to a user preference profile; selecting event groups for inclusion in the content summary in response to the summary rating of the event groups; and generating the summary by including summary items associated with the selected event groups.

These and other aspects, features and advantages of the invention will be apparent from and elucidated with reference to the embodiment(s) described hereinafter.

Brief Description of the Drawings

Embodiments of the invention will be described, by way of example only, with reference to the drawings, in which

FIG. 1 illustrates an apparatus for generating a personalised content summary in accordance with some embodiments of the invention;

FIG. 2 illustrates a method of generating a personalised content summary for a content item in accordance with some embodiments of the invention; and

CML02351 EV FIG. 3 illustrates an example of an ontology for a soccer content item.

Detailed Description of Embodiments of the Invention

The following description focuses on embodiments of the invention applicable to generation of a personalised summary for an audiovisual content item, such as a TV programme, a sports event, a movie etc. However, it will be appreciated that the invention is not limited to this application but may be applied to many other content items and content item types including for example text based and-or multimedia content items.

FIG. 1 illustrates an apparatus for generating a personalised content summary in accordance with some embodiments of the invention.

In the example of FIG. 1, a content item is provided from a content item source 101. The content item source 101 may be an internal or external content source. In the specific example, the content item is an audiovisual sequence. The content of the audiovisual sequence comprises a number of events which are identified in event information. The event information may be provided integrally with the content item or may be provided separately. For example, the event information may be provided as embedded metadata in the audio visual sequence or may be provided as a separate file comprising event data.

The event information identifies a number of events in the content item. For example, the content item may be a full length sequence of a soccer match and the event information may identify a number of events such as goals, free kicks, penalties, throw ins, bookings etc. The event information may e.g. specify each event by a description, a start time and an end time.

The content item source 101 is coupled to a rating processor 103 which is further coupled to a training data storage 105. The rating processor 103 is arranged to

CML02351 EV determine a rating of the events of the content item in response to training data stored in the training data storage 105. Specifically, the rating processor 103 determines the rating in response to associations between events and content summaries for a first group of training content items.

The training data storage 105 comprises data for the first group of training content items. In the example, the training data comprises data for training content items which, similarly to the content item to be summarised, comprises a number of events. For each training content item, the training data may specifically comprise event information that identifies event sequences in the content item. Each event sequence may comprise one or more events.

The training data furthermore comprises training content summaries and in particular each training content item has an associated training content summary. The training content summaries are made up of training summary items wherein each of the training summary items is associated with one or more of the event sequences.

Typically, the training data comprises data for a large number of training content items for which summaries have been created manually by an operator defining events in the content item and subsequently selecting the most important event sequences. Information for these event sequences may then be included in the content summary by the operator including one or more summary items for the event sequence in the content summary. The training data may e.g. additionally or alternatively comprise previously automatically generated summaries which the user has confirmed via a feedback mechanism that they meet their requirements.

For example, the training data may comprise data for a large number of training content items corresponding to football matches. For each of these, events have manually been defined and a content summary has manually been created. For example, an operator may have identified all free-kicks, goals, penalties etc and may have selected some of these events and included a description in the content summary. For example, the operator may have selected all event sequences which include a goal

CML02351 EV event and may have included information related to these event sequences in the training content summary. The operator may then have selected all free-kicks which result in a booking or a sending off and may have included information related to these event sequences in the training content summary.

It should be noted that training data may be generated specifically for the purpose of creating the training data. However, typically the manual creation of content summaries is performed for other purposes and in particular manual creation of content summaries may be widely used to provide summary information to users allowing them to identify and select between content items. Thus, manually generated content summaries may not only be used for information to a user but may also be used as learning/training data by the apparatus of FIG. 1.

It will be appreciated that in some embodiments, the training data storage 105 may comprise only the data needed for generating the content summary. However, in other embodiments the training data storage 105 may comprise additional information such as for example copies of some or all of the training content items themselves.

In the example of FIG. 1, the rating processor 103 determines the rating for an event by evaluating a frequency characteristic of the training data. In particular, the rating processor 103 evaluates a frequency of the inclusion of training summary items which are related to the event in the training content summaries. Thus, for a given event of the content item, all instances of the same event class occurring in the training data are identified. It is then evaluated how many of these were included in the corresponding content summaries and how many occurred in the corresponding full length sequence.

For example, if an event of the content item is a goal event, all instances of such an event class in the training data are identified. The frequency at which these event classes are referred to in the content summaries is then determined. The probability of inclusion of each event class may be considered as the conditional probability that

CML02351 EV information for the specific event class is included in the summary given that it has occurred in the event class sequence.

However, if the event is a throw-in event, the rating processor 103 will typically find that it is relatively rare that information relating to throw-in events is included in the summary. It will thus determine a much lower frequency value (or probability value).

The rating processor 103 determines the rating for each event group in response to the frequency indication for the event group and may specifically determine an increasing rating for an increasing frequency value. In some embodiments, the frequency (or probability) value may be used directly as a rating.

In the apparatus of FIG. 1, the training data storage 105 is further coupled to a training data selection processor 107. The training data selection processor 107 is arranged to select training content items for the first group of content items in response to a user preference profile 109.

The training data selection processor 107 selects the content items for the first group such that the training data used by the rating processor 103 is biased towards the user preference profile.

It will be appreciated that in some embodiments, the training data selection processor 107 may select the content items of the first group from e.g. an external set of content items and the training data storage 105 may only store data for the first group. However, in other embodiments, the training data storage 105 may alternatively or additionally comprise data for other content items and the training data selection processor 107 may select the first group as the subset of these content items which are to be used by the rating processor 103. For example, the training data storage 105 may comprise a large set of content items and the training data selection processor 107 may select a subset of these to be used by the rating processor 103 when rating the summaries.

CML02351 EV The training data selection processor 107 may for example, evaluate each of the content items stored in the training data storage 105 and may for each content item determine a match between the content item and the user preference profile. The match between the content item and the user preference profile may be determined by comparing the events of the content item itself and/or of the summary of the content item with the individual preferences of the user preference profile.

For example, the user preference profile may comprise N individual preferences and for a given content item all the events may be compared to the individual preferences. The total number of events that match an individual preference may be compared to a threshold and if the number of matches exceeds the threshold, the content item is included in the first group which is used by the rating processor 103. Alternatively or additionally, instead of only comparing the events of the content item with the user preference profile, the comparison may compare events referred to in the content summary with the individual user preferences, and if this results in more than a given number of matches, the content item may be included in the first group. It will be appreciated that in some embodiments both the events of the content item itself and the events of the content summary are included in the evaluation.

The rating processor 103 is coupled to a selection processor 111 which selects event groups for inclusion in the content summary in response to the rating of the events. For example, the selection processor 111 may simply select all events which have a rating above a given threshold thereby ensuring that all events considered sufficiently important to a user (in view of the user preference profile) are included in the content summary. This may provide for a variable length summary which is automatically adapted to characteristics of the content item. For example, the content summary for a goalless soccer match may be shorter than for a high scoring soccer match.

As another example, the selection processor 111 may select events in order of decreasing rating until the content summary reaches a desired size. This may allow relatively fixed size content summaries while ensuring that the most important events are included.

CML02351EV The selection processor may optionally select the events in response to the user preference profile and may in particular weight the ratings in response to the user preference profile. For example, any events which directly correspond to a user preference of the user preference profile may be weighted higher than an event which does not correspond to a user preference.

The selection processor 111 is coupled to a summary generator 113 which generates the content summary by including summary items associated with the selected events. For example, for each event a predefined metadata text may be included. For example, if events were selected corresponding to two goal events and a sending off event, the summary generator 113 may generate the summary by including the text

"first goal" +"second goal"+ "sending off in the content summary.

In other embodiments, the content summary may for example be an audiovisual sequence in itself and the summary generator 113 may for example generate an audiovisual sequence by merging clips of the content item where the clips are those corresponding to the selected events.

The apparatus of FIG. 1 may thus allow an efficient and low complexity generation of a content summary which is highly personalised. The content summary is created without requiring any specific rules or criterion for selection of items for the summary to be defined and in particular the personalisation may be achieved by an implicit biasing of the rating of events.

Furthermore, as the content summary generation does not rely on explicit rules but rather on the training content data, it may be suitable for many content domains. Also, the apparatus is particularly suitable for variable length summary generation as the approach does not require a new model to be built for different summary lengths. Rather a different summary length may be achieved merely by selecting a different

CML02351 EV W

18

number of event groups. The apparatus is furthermore of low complexity and may easily be implemented.

Due to the biasing of the training data to match that of the user preference profile, the rating of the events of the content item for which the personalised summary is to be generated is biased towards the preferences of the user. Thus, a personalisation of the content summary is achieved.

This personalisation may be highly accurate. Indeed, simulations and experiments have shown that the described approach may result in content summaries which are highly accurate and closely resemble those that would typically be obtained by a manual personalisation.

FIG. 2 illustrates a method of generating a personalised content summary for a content item in accordance with some embodiments of the invention. In the following, a more detailed example of embodiments of the invention will be described with reference to the method of FIG. 2. The method is applicable to the apparatus of FIG. 1 and will be described with reference to this.

The method initiates in step 201 wherein event information for a content item comprising a full length sequence of a soccer match is received. The content item comprises event information wherein a number of events are represented by instances of an ontology.

In the example, semantic events are represented as instances of a domain ontology which may be encoded in a standard format (for example RDF (Resource Definition Framework) or OWL (web ontology language)). FIG. 3 illustrates an example of an ontology for a soccer content item. The ontology consists of classes (or event types) and properties or features providing metadata about those events, as well as relationships between them. For example the Goal class has features by, duration, startjime, extra_time, from, taken and resulting_in.

CML02351 EV Similarly, in the example, the training data stored in the training data storage 105 comprises data for a large number of full length content sequences, each consisting of a number of instances of classes in an event ontology, along with summaries which consist of a subset of the events in each full length sequence.

The semantic events may have been entered directly as instances of classes in a standard ontology representation, for example using OWL, or may have been extracted from audio, video or text, using well-known techniques. The ontologies used for the training content item may be the same as the ontology used for the content item or may be a different ontology. In the latter case, the apparatus may comprise means for linking different ontology classes of one ontology to corresponding ontology classes of the other ontology.

The training data may particularly comprise data for a large number of training data content items comprising full length soccer sequences but may also comprise data for content items relating to very different content items and domains.

Step 201 is followed by step 203 wherein training content items are selected for the first group of content items in response to a user preference profile of the user for which the content summary is to be personalised.

Specifically, the apparatus may store a plurality of different user preference profiles and may select the appropriate user preference profile from the store profiles.

A specific example of a user preference profile for a soccer content item is the following:

CML02351 EV

In the example, two instances of example users are also shown, one of whom is more interested in controversial events, and the other in skill and goal-related events. The knowledge base and ontology for the user profiles can again be encoded in RDF or OWL. The instance knowledge may be quite simple and could be elicited from the user when they register for the personalised highlights service. Equally it may be complex and learned from observation of user interactions with the system over a period of time.

In some embodiments, the user preference profile may relate to a plurality of users such as a group of users having similar interests and preferences.

The training data selection processor 107 may proceed to determine a match indication for each content item stored in the training data storage 105. Specifically, for each event of a given content item, the training data selection processor 107 may compare the event with the individual user preferences of the user preference profile. For example, for an event corresponding to a goal scored by Anelka of Manchester city, the training data selection processor 107 may detect three matches. This may be repeated for all events of the content item and the total number of matches may be summed and used as the match indication. If the match indication is above a given

CML02351 EV threshold, the content item may be included in the first group of content items and otherwise it is not included. Thus, only if the content item matches the user's preferences sufficiently closely will the data for that content item be used by the rating processor 103.

In some embodiments, a fixed number of content items are included in the first group and the training data selection processor 107 may simply select the fixed number of content items having the highest match indication.

As a specific example, all the training data stored in the training data storage 105 may be searched for those soccer games which contain the closest content to the user preference profile. This search may be performed using a match indication between the events of the content item itself and/or of the content summaries of the content items. For example, denoting the user preference profile U and the content summary (or the content item) S, the following match indication may be used:

N

Similarity(S ,U) = ^T w,. * frequency{U_t e S)

where i is an index of the user profile properties, N is the number of properties in the user profile and wj is a weighting for each property.

That is, the number of occurrences in the training summary (and/or the content item) S of events with the same class as the favourite event property in the user preference profile, the second favourite event property, and so on, are summed to give the total match indication between content summary (or content item) S and the user with profile U.

If soccer game sequences and highlights are available in the training data which are already personalised to a particular user (for example, a personalised summary might be added to the training set in response to user feedback), user feedback may be

CML02351EV included in the measure. This may increase the importance of the user's previous successful summaries.

In some embodiments, w; = 1 Vi , that is, the weightings are not used, so that the match indication remains domain-independent. However, the weightings could be changed to tailor them to a particular domain where certain user preference profile properties are more important than others. These importance weightings could be included as additional fields in the user profile ontology itself.

In some embodiments a fixed number of the most similar summaries to the user profile are selected, along with their corresponding full length games, to create a biased training set in the form of the first group of content items.

In the example of FIG. 2, the rating processor 103 comprises functionality for grouping individual events of the content item into event groups before the event groups are rated as a whole. This may improve the context information for individual events and provide an improved summary.

Accordingly, step 203 is followed by step 205, wherein the event instances of the content item are grouped into event groups by the rating processor 103.

In the embodiment, the rating processor 103 groups the event groups in response to an event sequence probability. The event sequence probability is indicative of the probability of the events being associated with each other. The event sequence probability is determined in response to the training content items and in particular is determined by evaluating the frequency of which consecutive events corresponding to the event group are included in the training content summaries when they occur in the training content items. Sequential events may in particular be included until the combined probability falls below a given threshold.

In more detail, the events are grouped into causally related event groups using a Markov chain. This allows the system to provide context to the summary so that it

CML02351 EV does not consist solely of disjoint, unrelated events, but makes sense as a whole, and explains to the user, for example, what caused a player to be sent off the pitch, or how a goal came about.

This step uses the assumption that events commonly occurring in sequence in the training set are causally related. In order to reduce complexity and to apply the Markov chain approach, the event groups are divided into subsets comprising event pairs.

A conditional probability matrix, X, may be formed using the training data event sequences. The matrix X represents the probability of an event occurring in a training content item given that another event has just occurred. In the specific example, the matrix may be of size N , where N is the number of event classes (that is, each event E can take on any one of N symbols). E_t denotes the current event at time t, and E_t-1 denotes the previous event at time t- 1.

The elements of matrix X, X;J = P(E_t=i | E_t-1 =j), {i,j } e N are calculated as follows:

where:

frequency (event pairs (E^₁ = j, E_t = ϊ)) r \tL_t = I, t,_t__χ = J) = frequency (all event pairs in training set)

and:

_p , _P _ ._s frequency {event E _t__x = j) frequency (all events in training set )

CML02351EV Thus, each element in the matrix represents the probability of an instance of the same class as the second event in the event pair occurring, given that an instance of the same class as the first event in the event pair occurred.

In order to group the test sequence events into event groups, it is assumed that the Markovian property holds within an event group. That is, all knowledge about the past is assumed to be reflected by the previous state, namely:

P(E₁ I E,__x ,E,_₂...E₂,E₁) = P(E₁ \ E_t__x )

A Markov chain is constructed beginning with the first event in the test sequence E₁, with subsequent events E₂... E _t, to calculate the joint probability of a number of events E_t... E₂, E₁:

P(E_t,E_t__v...,E_ι) = P(E_l \ E_lΛ) *P(E_l__A \ E,_₂)...P(E₂ \ E_ι) *P(E_ι)

When the joint probability falls below a certain threshold (a suitable value may be 0.01) the events are considered to be a complete event group. In the example, there is no normalisation for long chains of events (e.g. by taking the t^root) as it is frequently beneficial to bias against very long sequences of events as these are less likely to be causally related.

It will be appreciated that any suitable method or criterion for grouping events may be used. For example, the grouping of sequential events may be performed manually or semi-manually.

Step 205 is followed by step 207 wherein the event groups are rated by the rating processor 103.

CML02351 EV Specifically, the rating processor 103 evaluates the frequency at which the training content items of the first group comprise the same event class sequences as the current event group and the frequency at which these event sequences are reflected in the training content summaries. A frequency (or probability) indication for the event group may be determined as the ratio between these. Typically, the more frequently information for a given event sequence is included in the summary relative to the frequency of its occurrence in the full length training sequences, the higher the importance of the event sequence and thus the higher the rating. Thus, the frequency or probability value may directly be used as a rating value.

Using the event groups, a second conditional probability matrix, Y, is calculated from the training set. Matrix element Yy represents the frequency or probability of an event E_t = i being included in the summary, given that the previous event E_t-1=j has been included and occurred immediately prior to E_t in the full length sequence. Specifically:

P(E! \E_t'_l) = ^P(E''P

' ' ^M P(E₁ ⁷)

_ frequenc^event pair (E,, E^₁) in both summaries & problem descriptions) frequenctfβvent E_t__λ in summaries

where E_t ¹ denotes the event at time t being included in a summary. Thus, whereas conditional probability matrix X indicated the probability of consecutive events being causally related, conditional probability matrix Y indicates the probability of such consecutive events being included in the training content summaries.

Then, each event group G = {E_t, E_t-1,...E₂, E₁ } is assigned a rating, based on its probability of being included, given that it has occurred:

Priority = P(G¹ \ G°) = ^P(G' 'f ^} ^y ' P(G°)

CML02351 EV The denominator of this equation is calculated as a Markov chain using probabilities of events occurring in a full length sequence. The numerator is also computed as a Markov chain:

P(G', G⁰) = p(E/-° \ E% ) * P(E% \ E_t':° )...P(E₂ ^ιβ \ E[^fi ) *P(β[^β)

where Ε_t ^I>0 denotes an event at time t that is included in the summary and which occurred in the training content item. Obviously, a single event that is included in the summary must have occurred in the training content item, however two events may be included as a pair in the summary, but not have occurred as a pair in the original full length sequence, due to editing points. In order to model these editing points, P(G¹, G°) is replaced by P(G¹), which can be calculated by:

P(G¹) = P(Ej \ EU) *P(EU \ El₂)...P(E₂' \ E{) *P(E()

Therefore, the final rating for each event group G is calculated as:

_{Rating(G) =} P(E₁ ¹ I EU)* P(EU \ El₂)...P(EJ I E/) * P(Ej) ^8K P(Ef \ EU) * P(E^ \ E?_₂)...P(E? \ E°)*P(E?)

These calculations make the assumption that the first order Markovian property holds, not only between events in the problem description, but also over events that are jointly included in the summary. Using this assumption, the problem of not having enough data to calculate probabilities of combinations of events that occur rarely or not at all (the so-called "one-shot learning" problem) may be alleviated substantially. For example, if the probabilities of all 5-event event groups are to be calculated a minimum of 20⁵ = 3200 000 events would be required to get one occurrence of each possible combination of 5 events. However, as the conditional probability matrices are based only on event pair sequences, a much reduced training data volume may be acceptable. This is particularly advantageous for a personalisation using biasing of training sets as the biasing typically reduces the available training data.

CML02351 ΕV Entries in the conditional probability matrix which are zero, because no examples of that particular event pair occur in the training data, may be given a small probability value in order to avoid that an event group is given a zero probability of occurrence which would result in an infinite rating.

Step 207 is followed by step 209 wherein event groups are selected in accordance with their rating. Specifically, the selection processor 111 may select a given number, N, of event groups and these may be selected as the N event groups which have the highest ratings.

As another example, the selection processor 111 may select event groups in response to a size characteristic of the content summary. For example, a fixed or maximum length summary may be required and the selection processor 111 may select event groups in order of decreasing rating until this length has been reached. As yet another example, the selection processor 111 may select all event groups having a rating above a given value resulting in summary a variable length depending on the importance/interest of the event groups.

The ratings of the event groups are implicitly affected by the user preference profile as the biased training set alters the conditional probabilities in the numerator of the above equation. This provides for an efficient, broadly applicable and accurate personalisation.

However, further personalisation may be achieved by weighting the ratings in response to the user preference profile.

For example, if an event in the event group involves a favourite player, the event group may be given additional importance by multiplying the rating for that event group by a weight higher than one. Likewise, the ratings may be modified by a weighting for each 'positive' event in the event group involving the favourite club or secondary clubs, and each 'negative' event involving the opponents of the favourite or

CML02351EV secondary clubs. 'Positive' events are defined to be e.g. Goals, Assists, Saves and Penalties, and 'negative' events are defined to be e.g. a Booking, Sending Off, Foul, Handball, Offside or Substitution. In this way, a fan can enjoy moments when the other side is doing badly which studies have shown is of particular interest to many users.

Step 209 is followed by step 211 wherein a content summary is created by including items in the summary for each of the selected event groups. For example, a semantic description may be included for each of the selected event groups. This may for example be in the form of metadata (such as MPEG-7 metadata).

As another example, if the event descriptions contain time codes and duration information corresponding to the audio visual signal, the summary may be provided in the form of a compilation of the audio visual video clips related to each of the selected event groups. Thus, a summary in the form of a highlights video sequence may be generated.

Simulations and experiments have shown that the described approach may result in personalised content summaries which are highly accurate and closely resemble those that would typically be obtained by a manual summary generation.

Furthermore, the method is applicable to information from any medium (not just text, or video alone). The importance of different pieces of information to the user is determined in a probabilistic manner, which avoids the need for manually specifying reasoning axioms. The method furthermore allows context to be incorporated in the summary, since groups of events that are causally related will all be included in the summary. Also, the approach avoids the requirement for separate models for different length summaries.

It will be appreciated that the above description for clarity has described embodiments of the invention with reference to different functional units and processors. However, it will be apparent that any suitable distribution of functionality between different

CML02351 EV functional units or processors may be used without detracting from the invention. For example, functionality illustrated to be performed by separate processors or controllers may be performed by the same processor or controllers. Hence, references to specific functional units are only to be seen as references to suitable means for providing the described functionality rather than indicative of a strict logical or physical structure or organization.

The invention can be implemented in any suitable form including hardware, software, firmware or any combination of these. The invention may optionally be implemented at least partly as computer software running on one or more data processors and/or digital signal processors. The elements and components of an embodiment of the invention may be physically, functionally and logically implemented in any suitable way. Indeed the functionality may be implemented in a single unit, in a plurality of units or as part of other functional units. As such, the invention may be implemented in a single unit or may be physically and functionally distributed between different units and processors.

Although the present invention has been described in connection with some embodiments, it is not intended to be limited to the specific form set forth herein. Rather, the scope of the present invention is limited only by the accompanying claims. Additionally, although a feature may appear to be described in connection with particular embodiments, one skilled in the art would recognize that various features of the described embodiments may be combined in accordance with the invention. In the claims, the term comprising does not exclude the presence of other elements or steps.

Furthermore, although individually listed, a plurality of means, elements or method steps may be implemented by e.g. a single unit or processor. Additionally, although individual features may be included in different claims, these may possibly be advantageously combined, and the inclusion in different claims does not imply that a combination of features is not feasible and/or advantageous. Also the inclusion of a feature in one category of claims does not imply a limitation to this category but

GML02351 EV rather indicates that the feature is equally applicable to other claim categories as appropriate. Furthermore, the order of features in the claims do not imply any specific order in which the features must be worked and in particular the order of individual steps in a method claim does not imply that the steps must be performed in this order. Rather, the steps may be performed in any suitable order. In addition, singular references do not exclude a plurality. Thus references to "a", "an", "first", "second" etc do not preclude a plurality.

CML02351 EV

Claims

1. An apparatus for generating a personalised content summary for a content item comprising a plurality of events; the apparatus comprising: rating means for determining a rating of events of the content item in response to associations between events and content summaries for a first group of training content items; training data selection means for selecting training content items for the first group of content items in response to a user preference profile; event selection means for selecting events for inclusion in the content summary in response to the rating of the events; and means for generating the summary by including summary items associated with the selected events.

2. The apparatus of claim 1 wherein the training data selection means is arranged to determine a match indication between the user preference profile and a plurality of training content items and to select content items for the first group of content items in response to the match indication.

3. The apparatus of claim 2 wherein the training data selection means is arranged to include a content item in the first group of content items if the match indication is above a threshold.

4. The apparatus of previous claim 2 or 3 wherein the training data selection means is arranged to determine the match indication for a first content item in response to an occurrence in the first content item of events matching at least one individual user preference of the user preference profile.

5. The apparatus of any previous claim 2 to 4 wherein the training data selection means is arranged to determine the match indication for a first content item in response to a number of occurrences of events matching individual user preferences of the user preference profile.

CML02351 EV

6. The apparatus of any previous claim 2 to 5 wherein the training data selection means is arranged to determine the match indication for a first content item in response to a weighted sum of the number of occurrences of events matching individual user preferences of the user preference profile.

7. The apparatus of claim 6 wherein the training data selection means is arranged to determine the match indication for the first content item substantially as

N Match = ^ Σ] W₁ * frequency (C/,. G S)

1=1

where i is an index of individual user preferences of the user preference profile, N is the number of individual user preferences in the user preference profile, w; is a weight for individual user preference i and Uj denotes the i'th individual user preferences and S denotes events of the first content item.

8. The apparatus of claim 6 or 7 wherein weights for the weighted sum are comprised in the user preference profile.

9. The apparatus of any previous claim 2 to 8 wherein the training data selection means is arranged to determine the match indication for a first content item in response to an occurrence of events matching at least one individual user preference of the user preference profile in a content summary for the first content item.

10. The apparatus according to any previous claim wherein the training data selection means is arranged to select content items in response to user preference profiles associated with the training content items.

11. The apparatus of any previous claim wherein the user profile is a multi user preference profile.

CML02351 EV

12. The apparatus of any previous claim further comprising grouping means for grouping sequential events of the plurality of events into event groups; and wherein the rating means is arranged to determine a rating for a first event 5 group of the event groups in response to a frequency characteristic of the training data, the frequency characteristic being indicative of a frequency of inclusion of training summary items related to an event sequence of the first event group in training content summaries of the first group of content items.

10 13. The apparatus claimed in claim 12 wherein the frequency characteristic comprises a frequency indication indicative of a number of times training summary items for the event sequence is included in the training content summaries relative to a number of times the event sequence occurs in the first group of training content items.

15 14. The apparatus claimed in any previous claim wherein the rating means is arranged to determine ratings by applying a Markov chain to event subsets of the event groups.

15. The apparatus of any previous claim 11 to 14 wherein the rating means is 0 arranged to divide at least a first event group into subsets of events and to determine a frequency indication for the first event group in response to frequency indications for each of the subsets.

16. The apparatus claimed in any previous claim 11 to 15 wherein the grouping 5 means is operable to group events in event groups in response to an event sequence probability.

17. The apparatus claimed in claim 16 wherein the grouping means is arranged to determine the event sequence probability for a first event group by applying a Markov 0 chain to event subsets of the first event groups.

CML02351 EV

18. The apparatus claimed in any previous claim wherein the selection means is arranged to weight the ratings in response to the user preference profile.

19. A method of generating a personalised content summary for a content item 5 comprising a plurality of events; the method comprising: determining a rating of events of the content item in response to associations between events and content summaries for a first group of training content items; selecting training content items for the first group of content items in response to a user preference profile;

10 selecting events for inclusion in the content summary in response to the rating of the events; and generating the summary by including summary items associated with the selected events.

15 20. A computer program enabling the carrying out of a method according to claim 19.

CML02351 EV