Livrables
The Swedish spoken corpora of email dialogues collected during the design and evaluation of the AthosMail spoken dialogue system consists of 63 calls to the simulated AthosMail application made by a group of 6 people ranging from 11 seconds to 17 minutes and were collected during a Wizard of Oz study. Approximately 6 hours with 1820 user utterances and 2382 system-based utterances were collected. The dialogues have been annotated using the Annotation Graphs notation (in XML format). The annotations include transcriptions and dialogue acts for each turn in the dialogue.
One part of the Finnish spoken dialogue corpora consist of results from a Wizard of evaluation, another from actual calls to the system by expert users, and rest from the systematic evaluation of the final system. Part of the material has been annotated using the Annotation Graphs notation (in XML format).
The English spoken dialogue corpora of email dialogues consists of 18 dialogues, ranging from 3 to 17 minutes, collected during a Wizard of Oz study. The dialogues have been annotated using the Annotation Graphs notation (in XML format). The annotations include transcriptions and dialogue acts for each turn in the dialogue.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The multi-session group scenario for Wizard of Oz studies has two novel features compared with the traditional WOz-studies: instead of doing solo sessions with a static mailbox, the test users communicated with each other in a group of six, and the communication took place over several sessions in a period of five to eight days. The six person experiment was duplicated at two sites, UIAH and SICS.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
Finnish Spoken Group Scenario Dialogue Corpora was collected during the design of the AthosMail-system. The corpora is based on the group scenario setup, and consists of the results from a 5-day session during which the six participants used the simulated AthosMail in interacting with each other according to the predefined scenario. The material has been annotated using the Annotation Graphs notation (in XML format).
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The component suggests the order of importance of messages by using the user's usage history and analysed messages. The user actions (such as which messages the user reads first, which messages she normally ignores or deletes without reading, which messages she moves to folders etc.) are used as cues in determining the importance of messages.
The analysis of the message deals with searching common patterns from message headers and trying to find relations between messages. The component also uses information from the message classification component in order to find common patterns related to topics.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The Dialogue Problem Detector (DPD) is part of the User Model component, and its task is to predict the presence of an ASR error in the user's utterance. It uses seven parameters to train a Naive Bayes classifier.
The precision and recall measures of the class "error" on the test data are 92% and 53%, respectively, and the figures suggest that the DPD might be a useful component, reducing the number of system's misunderstandings and user's errors during the dialogue.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The corpus was collected at Arla Institute (vocational training and development centre for visually impaired and deaf-blind people), and the aim was to verify applicability of the AthosMail in general and to evaluate how well the Design-for-All principles could be achieved in the AthosMail system. The corpus consists of the data from 5 users. It is transcribed and analysed with respect to user errors.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
This is a Windows application, which allows an experimenter running a Wizard-of-Oz simulation of a human-computer dialogue in the e-mail application domain to select responses in real-time from a set of canned responses.
The experimenter has a console in which the contents of the email folder are presented in tabular form, and menus from which to select actions to apply to the selected message (e.g. read it, read metadata about it), or other utterances appropriate to the dialogue state. The application is coupled to a TTS, which produces artificial speech responses.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
This item consists of two tools:
- A tool Web-based editor of dialogue corpora to be marked up in an Annotation-Graph notation.
- An editor to produce classification models for dialogue corpora, marked up in the Annotation Graph notation.
The first item allows the corpus annotator to select attribute values for the turns in a transcribed corpus of spoken language data. (The corpus is assumed to have been previously transcribed with the Annotation Graph toolkit, and aligned with the recorded speech using the PRAAT tool.) The annotator can select one or more dialogue acts to apply to a turn, and can ascribe additional attribute-values to the turn. It is implemented as a Java Servlet, and uses a relational database (currently Access).
The second item facilitates the construction of a classification model for use with the WEKA machine learning toolkit from a dialogue corpus annotated in the Annotation Graph notation. A typical classification model consists of a 'window' of n turns before the one to be classified. The editor allows the user to select n and to specify the attributes to be included in the model. Since the dialogue turns in the DUMAS English WOZ had been parsed using the FDG parser, a facility was added to specify regular expressions that would extract syntactic attribute values from the parser output.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
This is a named entity recogniser (names of persons, organisations, and places) for emails written in Swedish or English.
The recogniser requires to be fed one email at a time, and the email must be formatted according to the mail DTD and analysed with the appropriate FDG. The crucial parts of information are the tags with the corresponding lang-attribute. The agent tags the contents of each tag, using language specific resources depending on the value of the lang attribute.
The agent outputs each email on the same format as it received it, with -elements inserted as a children to the -tags.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The Swedish speech recognition grammar for the email domain was designed for use with the Nuance speech recognition engine. The grammar rules are mainly hand-crafted, whilst certain parts are dynamically generated according to mailbox contents (e.g. names of senders and email subjects).
In Finnish, several grammar formats exists for Philips, Lingsoft and Nuance speech recognition engines. Partly hand-made, partly dynamically generated by the application (for example, names and keywords appearing in the e-mail messages).
English speech recognition grammar for the email domain was designed for use with the Nuance speech recognition engine. The grammar rules are mainly hand-crafted, whilst certain parts are dynamically generated according to mailbox contents (e.g. names of senders and email subjects).
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The AthosCal prototype provides a multimodal interface to a central web-based calendar service and allows users to access it over, for example, a mobile phone or a PDA, using both spoken input and output. The application runs on several platforms including PDAs and desktop computers, and is currently in test usage by a number of Swedish users. Plans are under way to extend its functionality for mobile usage to mobile phone clients. The application was chosen as an advanced prototype due to the possible synergies of a calendar application with an e-mail system. Further, mobile users of an e-mail application may also want to access calendar data.
One of the major goals of the prototype development was to investigate the feasibility of designing speech interfaces for expert and frequent users. These users are expected to be mobile professionals with high requirements on the interface functionality. AthosCal was developed for Swedish and English and has been evaluated separately. For allowing the voice-controlled calendar to be used as a complement to desktop-based calendars, support to the international iCalendar standard was developed.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
AthosMail is an e-mail application that deals with multilingual issues in several forms and environments, and whose functionality can be adapted to different users, different situations and tasks.
The random indexing server for AthosMail implements advanced text processing functionality that uses the Random Indexing vector space technique.
Functionality:
- LexiconExtraction: Extracts the most similar words (default number is 2) to all words in the messages, and returns the lexicon to the calling Agent.
- UpdateMarkProfile: Updates the user's MARK profile.
- EmailClassifier: Computes the correlation to the user's MARK profile, and classifies the message against the user's other profiles. The sever returns the correlation to the MARK profile, and a ranked list of the other profiles together with their correlation scores.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The email summariser implements an n-best sentences selector summariser for emails written in Swedish or English.
The summariser requires to be fed one email at a time, and the email must be formatted according to the mail DTD and analysed with the appropriate FDG. The crucial parts of information are the tags with the corresponding lang-attribute. The summariser tags the contents of each tag, using language specific resources depending on the value of the lang attribute.
The summariser outputs each email on the same format as it received it, with sentences inserted in the -tag.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The project has constructed Athos, a generic and modular framework for multilingual speech-based application development. The Athos architecture is hybrid, utilising components based on different computational paradigms, such as symbolic computation, statistical methods and neural networks.
The Cooperativity Model deals with the systems dialogue control and explicitness of the given information which affect the systems interaction capabilities and naturalness of the dialogue. Cooperativity Model produces recommendations for the appropriate explicitness of the system utterances, depending on the users observed competence levels, and is taken into account by the system's generation components when deciding the system utterances.
The model consists of an offline and an online version, which use somewhat different input parameters, due to their different functionality in the system: the online version reacts to the ongoing interaction, while the offline version records the user's familiarity with the system in general.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
AthosNews is a 'talking newspaper' application aimed at visually-impaired users. Traditional talking newspaper services are based on recording a volunteer human reader in a studio, and this mode of production limits the amount of material that can be recorded in a given period. Using computer-generated speech, this bottleneck can be overcome, and AthosNews is a prototype of such a service, which has been implemented with speech input for control.
The speech input allows users to search or browse for news articles by topic, which is the main feature that distinguishes it from existing services. ASR is done with a Nuance grammar, which is partly dynamically generated from the news article titles, and can be extended by the user via a spelling mode.
The system has been tested by a group of visually-impaired expert computer users, for whom a multi-modal (keypad for within-text control) dialogue would be a basic requirement, but who recognized the benefits of content retrieval via speech. The system prototype could be developed either into a telephone-accessed service or as a desktop application. The system has been described in an article submitted to a journal, currently under review.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
Searcher is an experimental information retrieval system developed at SICS, and currently supports Boolean, Vector Space and structured queries. The document ranking function is a classical combination of term frequency (the importance of a term in a document) and the inverse document frequency (the importance of a term in the collection).
Searcher is publicly released under the MIT license, and the latest version can be retrieved from SICS at http://www.sics.se/humle/socialcomputing/download/
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
The Finnish AthosMail evaluation corpus consists of the data collected during the evaluation of the AthosMail system at UIAH and UTA. A total of 23 users used the system on two consecutive days, and on both days they had two tasks. The tasks were designed to encourage using different functionalities of the system. Each task was a separate dialogue in a separate phone call. The overall number of dialogues is 92.
More information on the DUMAS project can be found at:
http://www.sics.se/dumas/
Recherche de données OpenAIRE...
Une erreur s’est produite lors de la recherche de données OpenAIRE
Aucun résultat disponible