NVBG

Overview

The Non-Verbal-Behavior-Generator (NVBG), as the name suggests, is the module that generates behavior other than speech, such as gestures, facial expressions and gazes. The NVBG generates the afore mentioned behavior based on the speech input and other messages that it receives. These behaviors serve to augment and emphasize spoken communication. Based on who speaks and who listens, the NVBG can characterize the NPC as a listener, speaker or a bystander and generate appropriate behavior. These behaviors are configurable using xml files provided as input to NVBG. Using these xml files, we can specify which words or parts-of-speech trigger which animations. We can specify idle animations and idle gazes that get triggered when the character is idle for a specified amount of time.

The NVBG is loosely coupled with a "parser" which gets the input sentence from NVBG and returns the parts of speech in the sentence. NVBG can extract information from the lexical, syntactic, and semantic structure of the surface text that can support the generation of believable nonverbal behaviors.

The NVBG also has a saliency map which it uses to keep track of it's environment, important events and conversational topics. This helps the NVBG generate behaviors that are appropriate to the current scene and helps it prioritize. This saliency map is configurable through xml files too.

The NVBG also allows us to configure facial expressions for a particular character using FACS units and then based on xml tags in the input messages, can trigger those facial animations.

Quick facts

The NVBG is located at https://svn.ict.usc.edu/svn_vh/trunk/core/NVBG
Main Developer: Shridhar Ravikumar

Snapshot of NVBG

Users

Currently the NVBG needs to be setup on a per character basis i.e. each character in the scene needs to have his/her own process of NVBG.The character itself needs to be configured using the various confuguration options that NVBG provides.

In order to configure NVBG for a particular character, you need to have/setup the following:

The rule_input_[culture].xml file, that specifies which rules should be generated for parts of speech in the spoken sentence and also which animations map to which words if any. Assuming you have the rule-input-[char-name].xml file you need to pass it as an argument to NVBG when you run the executable so that it knows which file to use for rules.
The .xslt transform files which process the the intermediate xml generated by the module and generate the final bml to be output. The defaults are available at http://svn.ict.usc.edu/svn_vh/trunk/data/nvbg-common/. The relevant files are NVBG_transform.xsl, NVBG_rules.xsl and NVBG_behavior_description.xsl
The command line parameters to NVBG including the character-name, character-culture, idle posture and some other optional parameters
OPTIONAL - we can also specify a saliency map and a facial-expressions configuration if needed.
1. The saliency map specifies which objects in the scene are of particular importance which allows NVBG to generate idle gazes appropriately.
2. The facial-expressions configuration file allows us to specify which facs units should be triggered with a certain weight so that a facial expression is achieved. This facial expression can later be tagged in the input text as markup to trigger the facial expression.

Command line parameters

Required:

-agent_name: specifies the agent name that the current instance of NVBG corresponds to.

-posture: specifies the current posture that the character is currently in. This helps in generating animations.

-culture: specifies which rule input file to use for behavior e.g. if you specify 'brad', it will use rule_input_brad.xml as the rule input file. If you specify 'Iraqi', then it will use rule_input_Iraqi.xml.

-data_folder_path: specifies the folder path to the xslt files and the rule_input_[culture].xml file.

Optional:

-write_to_file: specifies whether the output bml is to be sent out as a VH message or if it should be written to a file. Default is false.

-write_to_file_path: If above parameter is true, then this file is where the output is written to.

-parsetree_cachefile_path: specifies the file path to the file used for caching the response from the parser used by NVBG. If this doesn't exist, it is created.

-hide_GUI: Hides the GUI if specified as true.

-expressions_file_name: specifies the path to the expressions file if one is being used

-saliency_idle_gaze: specifies whether idle gaze is turned on or not. "false" by default.

-speaker_gaze: specifies whether gazes should be generated while speaking.

-speaker_gesture: specifies whether gestures should be generated while listening.

-listener_gaze: specifies whether gazes should be generated while listening.

-nvbg_POS_rules: specifies whether behavior should be generated based on parts of speech

-saccades: specifies whether saccades should be generated

-storypoint: specifies which story-point should be loaded from the saliency-map xml and the saliency map is updated accordingly with priorities to appropriate pawns etc.

On providing the required command line arguments, it should work with the other VH components and in accordance with the rules specified for it.

Input Messages

NVBG mainly subscribes to vrExpress messages and a few other control messages which allow for setting some options. Below is a list of messages that NVBG subscribes to:

vrExpress

This message is sent to NVBG by the NLU, NPCEditor or similar module. This message can be used to convey information about speech data, posture, status change, emotion change, gaze data etc. as shown below.

Speech
The speech messages are characterized by the speech tag within them. They are interpreted and the corresponding output bml is generated with the speech time marks, animations, head-nods, facial-movements etc. These animations are generated based on the content of the speech tag and the fml tag in the input message.

vrExpress "harmony" "ranger" "harmony221" "<?xml version="1.0" encoding="UTF-8" standalone="no" ?><act>
<participant id="harmony" role="actor"/>
<fml>
<intention>
<object name="A316">
<attribute name="addressee">ranger</attribute>
<attribute name="speech-act">
<object name="A317">
<attribute name="content">
<object name="V28">
<attribute name="modality">
<object name="V29">
<attribute name="conditional">should</attribute>
</object>
</attribute>
<attribute name="polarity">negative</attribute>
<attribute name="attribute">jobAttribute</attribute>
<attribute name="value">bartender-job</attribute>
<attribute name="object-id">utah</attribute>
<attribute name="type">state</attribute>
<attribute name="time">present</attribute>
</object>
</attribute>
<attribute name="motivation">
<object name="V27">
<attribute name="reason">become-sheriff-harmony</attribute>
<attribute name="goal">address-problem</attribute>
</object>
</attribute>
<attribute name="addressee">ranger</attribute>
<attribute name="action">assert</attribute>
<attribute name="actor">harmony</attribute>
</object>
</attribute>
</object>
</intention>
</fml>
<bml>
<speech id="sp1" type="application/ssml+xml">ranger utah cant be bartender if he becomes sheriff</speech>
</bml>
</act>"

vrExpress "utah" "ranger" "utah200" "<?xml version="1.0" encoding="UTF-8" standalone="no" ?><act>
<participant id="utah" role="actor"/>
<fml>
<intention>
<object name="A131">
<attribute name="addressee">ranger</attribute>
<attribute name="speech-act">
<object name="A133">
<attribute name="reason">become-sheriff</attribute>
<attribute name="content">
<object name="V6">
<attribute name="content">
<object name="V7">
<attribute name="q-slot">polarity</attribute>
<attribute name="type">question</attribute>
<attribute name="prop">
<object name="P24">
<attribute name="type">event</attribute>
<attribute name="location">town</attribute>
<attribute name="theme">sheriff-job</attribute>
<attribute name="event">providePublicServices</attribute>
<attribute name="agent">utah</attribute>
<attribute name="time">future</attribute>
</object>
</attribute>
</object>
</attribute>
<attribute name="type">csa</attribute>
<attribute name="action">info-req</attribute>
<attribute name="actor">ranger</attribute>
<attribute name="addressee">utah</attribute>
</object>
</attribute>
<attribute name="motivation">
<object name="V5">
<attribute name="reason">become-sheriff</attribute>
<attribute name="goal">avoid</attribute>
</object>
</attribute>
<attribute name="action">avoid</attribute>
<attribute name="type">backward</attribute>
<attribute name="addressee">ranger</attribute>
<attribute name="actor">utah</attribute>
</object>
</attribute>
</object>
</intention>
</fml>
<bml>
<speech id="sp1" type="application/ssml+xml">im not sure about that</speech>
</bml>
</act>"

Posture change
These messages are characterized by the <body posture=""> tag which allows NVBG to know that there has been a change in posture.
vrExpress "harmony" "None" "??" "<?xml version="1.0" encoding="UTF-8" standalone="no" ?><act>
<participant id="harmony" role="actor" />
<bml>
<body posture="HandsAtSide" />
</bml>
</act>"

Status / request
The idle_behavior and all_behavior attributes within the request tag allows NVBG to keep track of whether or not to generate the corresponding behavior.
vrExpress "harmony" "None" "??" "<?xml version="1.0" encoding="UTF-8" standalone="no" ?><act>
<participant id="harmony" role="actor" />
<fml>
<status type="present" />
<request type="idle_behavior" value="off" />
</fml>
</act>"

Gaze

These gaze tags, if present within the input message are transferred unaltered to the output message.
vrExpress "harmony" "ranger" "constant103" "<?xml version="1.0" encoding="UTF-8" standalone="no" ?><act>
<participant id="harmony" role="actor" />
<fml>
<gaze type="weak-focus" target="ranger" track="1" speed="normal" > "listen_to_speaker" </gaze>
</fml>
</act>"

vrExpress "utah" "ranger" "gaze199" "<?xml version="1.0" encoding="UTF-8" standalone="no" ?><act>
<participant id="utah" role="actor" />
<fml>
<gaze type="avert" target="ranger" track="eyes-offset" speed="slower" > "planning_speech_hold_turn" </gaze>
</fml>
</act>"

Emotion
The affect tag contains data about the emotional state the character is currently in. This can be used to affect output behavior.
vrExpress "harmony" "None" "schererharmony17" "<?xml version="1.0" encoding="UTF-8" standalone="no" ?><act>
<participant id="harmony" role="actor" />
<fml>
<affect type="Fear" STANCE="LEAKED" intensity="110.475"></affect>
</fml>
<bml> </bml>
</act>"

nvbg_set_option

These are control messages to set options for NVBG. They are as shown below:

nvbg_set_option [char-name] all_behavior true/false - sets/unsets flag that allows all behavior generated by NVBG.
nvbg_set_option [char-name] saliency_glance true/false - sets/unsets flag that allows saliency map generated gazes. These gazes are based on content in the speech tag and the information in the saliency map.
nvbg_set_option [char-name] saliency_idle_gaze true/false - sets/unsets flag that allows idle gazes generated by the saliency map. These idle gazes are based on the priority of pawns in the saliency map and are generated when the character is idle.
nvbg_set_option [char-name] speaker_gaze true/false - sets/unsets flag that allows for the character to look at the person he's speaking to.
nvbg_set_option [char-name] speaker_gesture true/false - sets/unsets flag that allows gestures to be generated when speaking.
nvbg_set_option [char-name] listener_gaze true/false - sets/unsets flag that allows the listener to gaze at the speaker when he speaks.
nvbg_set_option [char-name] nvbg_POS_rules true/false - sets/unsets flag that allows behavior to be generated based on parts of speech returned by the parser.

Output

vrSpeak

Below is an example of the vrSpeak message that is output from NVBG. It contains animation/head-nods/facial behavior based on the text that is to be spoken/heard.

vrSpeak brad user 1337363228078-9-1 <?xml version="1.0" encoding="utf-16"?><act>
<participant id="brad" role="actor" />
<bml>
<speech id="sp1" ref="voice_defaultTTS" type="application/ssml+xml">
Depending
on
the
default
voice
you
have
selected
in
Windows,
I
can
sound
pretty
bad.

</speech>
<event message="vrAgentSpeech partial 1337363228078-9-1 T1 Depending " stroke="sp1:T1" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T3 Depending on " stroke="sp1:T3" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T5 Depending on the " stroke="sp1:T5" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T7 Depending on the default " stroke="sp1:T7" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T9 Depending on the default voice " stroke="sp1:T9" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T11 Depending on the default voice you " stroke="sp1:T11" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T13 Depending on the default voice you have " stroke="sp1:T13" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T15 Depending on the default voice you have selected " stroke="sp1:T15" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T17 Depending on the default voice you have selected in " stroke="sp1:T17" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T19 Depending on the default voice you have selected in Windows, " stroke="sp1:T19" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T21 Depending on the default voice you have selected in Windows, I " stroke="sp1:T21" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T23 Depending on the default voice you have selected in Windows, I can " stroke="sp1:T23" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T25 Depending on the default voice you have selected in Windows, I can sound " stroke="sp1:T25" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T27 Depending on the default voice you have selected in Windows, I can sound pretty " stroke="sp1:T27" />
<event message="vrAgentSpeech partial 1337363228078-9-1 T29 Depending on the default voice you have selected in Windows, I can sound pretty bad. " stroke="sp1:T29" />
<gaze target="user" direction="POLAR 0" angle="0" sbm:joint-range="HEAD EYES" xmlns:sbm="http://ict.usc.edu" />
<sbm:event message="vrSpoke brad user 1337363228078-9-1 Depending on the default voice you have selected in Windows, I can sound pretty bad." stroke="sp1:relax" xmlns:xml="http://www.w3.org/XML/1998/namespace" xmlns:sbm="http://ict.usc.edu" />

<animation stroke="sp1:T3" priority="5" name="HandsAtSide_Arms_GestureWhy" />

<head type="NOD" amount="0.10" repeats="1.0" relax="sp1:T5" priority="5" />

<head type="NOD" amount="0.10" repeats="1.0" relax="sp1:T11" priority="5" />

<head type="NOD" amount="0.10" repeats="1.0" relax="sp1:T19" priority="5" />

<head type="NOD" amount="0.10" repeats="1.0" relax="sp1:T21" priority="5" />
</bml>
</act>

Explanation:

The Time-Markers i.e. T0, T1, T2 and so on, are used by the NVBG to mark each individual word in the sentence and later it can use this information to play animations at the start/end of that word.

e.g. in the above output bml, the animation "HandsAtSide_Arms_GestureWhy" is timed to play at sp1:T3, which means it will be played when the word "on" has been spoken. 'sp1' is the id of the speech tag which indicates that this 'T3' is from this particular speech tag. (currently NVBG only supports one speeech tag).

The appropriate "vrAgentSpeech partial" event messages are sent out by smartbody after each word has been spoken by the character. e.g."vrAgentSpeech partial 1337363228078-9-1 T19 Depending on the default voice you have selected in Windows," is sent out after the word Windows is spoken.

Similarly the <head> tag is interpreted by smartbody to generate Nods and other specified behavior.

The "priority" attribute allows NVBG to decide which behavior is to be culled if any of the behaviors overlap. This is done before the final message is sent out.

Developers

Functional Block Diagram

Configuring Rules

The rule-input-[culture].xml file contains the rules that map certain words or parts of speech to animations. You can also specify any idle animations that you want to trigger when the character is idle for a certain period of time.

The xml code below shows how we can map a word to multiple animations which are then picked by random by NVBG when that word is spoken.

<rule keyword="statement_animation" priority="2" >
<pattern>is</pattern>
<pattern>are</pattern>
<pattern>were</pattern>
<pattern>was</pattern>
<pattern>have been</pattern>
<pattern>has been</pattern>
<pattern>at</pattern>
<pattern>stands</pattern>
<pattern>come</pattern>
<pattern>like</pattern>
<animation>
<posture name="CrossedArms">
<clip>CrossedArms_RArm_GestureYou02</clip>
<clip>CrossedArms_RArm_GestureYouPalmUp</clip>
</posture>
<posture name="HandsAtSide">
<clip>HandsAtSide_RArm_GestureOffer</clip>
<clip>HandsAtSide_RArm_LowBeat</clip>
<clip>HandsAtSide_RArm_MidBeat</clip>
<clip>HandsAtSide_Arms_Beat</clip>
<clip>HandsAtSide_Arms_Chop</clip>
<clip>HandsAtSide_RArm_Chop</clip>
<clip>HandsAtSide_RArm_FistsChop</clip>
<clip>HandsAtSide_RArm_LowBeat</clip>
<clip>HandsAtSide_RArm_MidBeat</clip>
</posture>
<posture name="HandsOnHip">
<clip>HandsOnHip_RArm_MidBeat</clip>
</posture>
<posture name="LHandOnHip">
<clip>LHandOnHip_RArm_You</clip>
<clip>LHandOnHip_RArm_GestureOffer</clip>
</posture>
<posture name="Chair">
<clip>Chair_You_Small</clip>
<clip>Chair_You</clip>
</posture>
</animation>
</rule>

The pattern <tag> contains the word that is to be matched and the <clip> tags contain the animations which should be played when the word is uttered. The <posture> tag is chosen based on which posture the character is currently in. Only animations that are in the current characters posture will be played by NVBG. The 'priority' attribute allows NVBG to prioritize between animations when multiple one's overlap.

The rules for parts of speech are similar as shown below.

<rule keyword="first_NP" priority="5" >
<pattern>first_NP</pattern>
</rule>
<rule keyword="noun_phrase" priority="5" >
<pattern>NP</pattern>
</rule>

Notice that in the above case, no animations are specified (although they can). This is because in this case, what the NVBG does is that it checks the spoken sentence for parts of speech (first_NP, noun_phrase etc.) and inserts place-holder xml tags with the keyword as specified in the rule. Later when the NVBG applies the XSL transform to the intermediate BML(with the placeholder tags), it generates output behavior based on what the XSL rules specify (More on this later under the "POS Transform rules" section).

So basically the rule input file is a collection of these rules that the NVBG parses at runtime. Based on whether animations are directly specified in this file as <patterns> or whether they are specified further down the pipeline in the XSL rules, the NVBG generates appropriate behavior.

If you want idle animations to be generated, you can specify idle_animation rules as below


<rule keyword="idle_animation" priority="1" >
<pattern>idle_animation</pattern>
<animation>
<posture name="ChrUtah@IdleHandsAtSide01">
<clip>ChrUtah@IdleHandsAtSide01_FidgetWeightShiftRt01</clip>
<clip>ChrUtah@IdleHandsAtSide01_FidgetWeightShiftLf01</clip>
<clip>ChrUtah@IdleHandsAtSide01_FidgetHead01</clip>
</posture>
</animation>
</rule>

The above idle_animation rule is parsed just like the other rules and is inserted in the output bml when the character is idle.

****************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************************

POS Transform rules

TO BE POPULATED

Saliency map

The saliency map specifies which objects/characters/pawns in the environment are important to the character and what priority they are. These priorities can vary based on what story-point we are at i.e. the priorities can change on loading a new story-point.

Based on the priorities of these objects, the saliency map generates idle gazes and other actions based on spoken sentences and events.

Below is an example of a saliency map for use with a character

<?xml version="1.0" encoding="utf-8" ?>
<storyPoints>
<storyPoint name = "toolkitsession">
<SaliencyMapInit>
<pawn name = "user" recency = "0" primacy = "3"/>
<pawn name = "upright" recency = "0" primacy = "2"/>
<pawn name = "upleft" recency = "0" primacy = "1"/>
</SaliencyMapInit>
<keywordToObjectMap>
<keyword name = "I">
<pawn name = "upleft" primacy = "10"/>
<pawn name = "upright" primacy = "10"/>
<pawn name = "user" primacy = "3"/>
</keyword>
</keywordToObjectMap>
<emotionInit name = "normal"/>
</storyPoint>
</storyPoints>

As we can see, the saliency map can contain multiple story-points each specifying which objects in the scene are of importance, the relation between spoken words and object in the scene and the emotion of the charcter.

The above example contains only one story-point but in general it can contain many. NVBG can be notified of which storypoint it should load so that the appropriate priorities are assigned to the objects in the scene.

Facial-expression file

TO BE POPULATED

Known Issues

Either list of common known issues and/or link to all Jira tickets with that component name.

Related Components

FAQ

What is NVBG?

The Nonverbal Behavior Generator (NVBG) is a tool a that automates the selection and timing of nonverbal behavior for Embodied Conversational Agents (aka Virtual Humans). It uses a rule-based approach that generates behaviors given the information about the agent's cognitive processes but also by inferring communicative functions from a surface text analysis

Where can I find NVBG on svn?

https://svn.ict.usc.edu/svn_vh/trunk/core/NVBG

If I have questions or problems, who should I contact?

vh-support@ict.usc.edu

Page tree

Pages … Home Components Modules NVBG