Gerhard Fobe, Univ. of Chemnitz

Exec Summary

Gerhard Fobe is happy to provide a library for C#, released under FreeBSD license, that can support developers to work with EmotionML. With the help of the integrated EmotionML-parser it is possible to create related object instances automatically. Furthermore object instances can be converted to EmotionML too (in DOM and XML mode). Beside a standalone EmotionML document the plug-in version for the inclusion of emotions in other languages is supported.

Marc Schröder, DFKI

Exec Summary

DFKI GmbH is very happy to add support for EmotionML to its expressive speech synthesis system MARY TTS. A standard format for representing the expressivity to be expressed in speech has been missing, so it is good to have it available now. Particularly the support for both categorical and dimensional representations of emotions is important to our system.

Patrick Gebhart, DFKI

Exec Summary

The W3C EmotionML 1.0 Specification is now supported by the ALMA software. ALMA relies on EmotionML for describing appraisal of emotions, the numeric simulation of emotion intensity decay, and mood characteristic changes within the pad dimensions.

Felix Burkhardt, DTAG

Exec Summary

Deutsche Telekom Innovation Laboratories is happy to support EmotionML to deploy it as a standard format in its developments concerning emotion processing services. Working primarily as a system integrator, standardized interfaces are extremely important for us to plug together components from different providers.

Christian Becker-Asano, Univ. of Freiburg

Exec Summary

Dr. C. Becker-Asano of the Foundations of Artificial Intelligence lab at the Department of Computer Science, Freiburg University, is happy to support EmotionML and will use it as a standard format in its developments concerning emotion simulation services. To let the WASABI architecture seamlessly interface with other software modules, a standard interface such as EmotionML is indispensable.

Tim Llewellin, nViso

Exec Summary

nViso 3D Facial Imaging API is an online service for recognition of emotions depicted through facial expressions in still images and videos. The focus of this implementation of EmotionML is on using the media type and uri time for video which are used in the service. Key features used in API which supports EmotionML include defining our own catagory of emotion key words and linking emotions measurements to media analyzed. As the report was produced as part of a commerical implementation, some parts of the specifications were not deemed appropriate for our service, so were not implemented. More information about our service can be found here: www.nviso.ch

Edmon Begoli, Oak Ridge National Laboratory

Exec Summary

EMLPy is a Python based library for generation of EmotionML compliant documents.

This is a second iteration of the initial implementation of EMLPy Python library that is intended as a utility for generation of EmotionML-compliant XML documents. EMLPy generates EmotionML documents by transforming the user specified and populated Python object tree into XML representation. EMLPy performs EmotionML checks covered in assertions while executing this object to XML transformation. From an API perspective, user interacts with an object tree hierarchy that maps directly to EmotionML hierarchy of elements and attributes. User generates an XML document by invoking the .to_xml() function. At this time EMLPy validates the object tree and its properties against the EmotionML schema and specification rules. There are few cases that require special consideration and handling:

The value of the "media-type", if present, MUST be a valid MIME type - we check the MIME types against the http://www.iana.org/assignments/media-types. This is a live check that performs URI resolution for types and page search for sub-types. We pursued this approach in attempt to dynamically validate MIME types against the authoritative repository.

xsd:anyURI types - anyURI checks are syntactic, not semantic. Checks are performed against common URI patterns - http,ftp, structure of IP address, proper domains, etc. We are not performing checks against the actual live URIs to check if the content is there.

Vocabularies - we perform vocabulary checks for those vocabularies defined within EmotionML instance being created. If the dimension, category, appraisal or sets are pointing to external URL we do not fail the generation of the document if the URL defining the set is not valid, but we only issue warning. We have chosen this approach as aligned with Python's dynamic typing philosophy.

Project URL: https://github.com/ebegoli/EMLPy Implementation Team: Edmon Begoli, begolie@ornl.gov, Oak Ridge National Laboratory/University of Tennessee-Knoxville Chelsey Dunnivan, ckd002@morningside.edu, Morningside University

Roddy Cowie, Queens Univ. Belfast

Exec Summary

The HUMAINE centre (School of Psychology, Queen’s University Belfast) welcomes EmotionML and will deploy it as a standard format in its work. That primarily involves use case 1, manual annotation of material involving emotionality, such as annotation of videos, of speech recordings, and of faces, in naturalistic databases. As psychologists, it is important to have a format that we can use, which allows us to use psychologically well-founded descriptors, and know that consumers in the engineering community will be able both to access, and to understand them. The system currently implements tracing for category and dimensional descriptors. We will extend that to appraisal and action tendency descriptors if research confirms that they can be traced reliably, but at present that is not clear.

Abe Kazemzadeh, USC SAIL Lab

Exec Summary

Emotion twenty questions (EMO20Q) is a experimental framework for studying how people describe emotions in language and how computers can simulate this type of verbal behavior. In EMO20Q, the familiar spoken parlor game of twenty questions is restricted to words that players feel refers to emotions.

In this implementation report, we examine the case where a server-side computer agent plays the role of the questioner and a human plays the answerer via a web browser.

The EMO20Q questioner agent can be decomposed into several components, notably a **vocabulary**, **semantic knowledge**, an **episodic buffer**, and a **belief state**. The vocabulary is a list of 110 emotion words and this vocabulary is expected to grow over time as more data is collected, but remains constant during the agent's instantiation. Semantic knowledge is a large object that remains the same across different agent instantiations and states, while the episodic buffer and belief state are smaller objects that vary over time for each interactive session. Because of the size of the semantic knowledge object, serialization of the agent for each session is not possible. Rather, the episodic buffer and belief states are serialized while the semantic knowledge persists as a static object in the server memory. The belief state is represented as a probability vector indexed by items of the vocabulary.

EmotionML is used to implement the questioner agent's vocabulary and belief state. The agent's vocabulary is implemented using the EmotionML ``vocabulary`` idiom and the agent's belief state is represented using the ``emotion`` idiom with a ``category`` as a child and the ``value`` attribute to hold numerical probability values.

The only thing that produced a failure to validate was having a space in the name of a vocabulary item.