NAME
Emotion - discrete emotion annotation processing
SYNOPSIS
empathize <transcript>
empathize --index <transcript> [<transcript> ...]DESCRIPTION
The most difficult part of this module is learning how to annotate free-form text with information about what competition is taking place. If you haven't studied http://ghost-wheel.net then go do that now.
The first part of this document describes the XML representation of 3rd person annotations and the validation rules. The second part describes the perl API.
XML TUTORIAL
The Basics
The minimum situation consists of the type of competition and two opponents.
abbreviation type of competition (situation)
------------ -------------------------------
destroys [-] destroys [-]
steals [-] steals from [+]
uneasy [-] is made uneasy by [0]
exposes [+] exposes [-]
impasse [+] and [+] are at an impasse
admires [+] admires [0]
observes [0] observes [-]
accepts [0] accepts [+]
ready [0] and [0] are at readinessEach opponent is labelled either left or right, indicating on which side of the situation he or she participates. For example:
<observes left="student" right="teacher">
A student watches the teacher make a presentation.
</observes>
<observes right="student" left="teacher">
The teacher watches a student make a presentation.
</observes>Of course a teacher is probably more accustomed to lecturing in front of a class than a student. This information can be captured with the intensity attribute:
<observes left="student" right="teacher" intensity="gentle">
A student watches the teacher make a presentation.
</observes>
<observes right="student" left="teacher" intensity="forceful">
The teacher watches a student make a presentation.
</observes>Now it is clear that the teacher's presentation involves some gentle excitement, but a student presentation is an exciting test of self-control. Since no spin is involved, the emotion is roughly the same regardless of whether we empathize with the presenter or the audience. To contrast, situations involving spin require some additional structure. For example:
<steals left="thief" right="child">
The thief steals the teddy bear from the child.
</steals>Here it is not clear which emotion is intended. Opponents experience the situation very differently. Perhaps the thief is drunk with accomplishment, but the child is probably quite angry. To avoid this ambiguity, situations involving spin follow a transactional structure:
initiator: (1) before ---+---> (3) after
|
v
victim: (2) tensionHere is a hypothetical trace of a steals situation decomposed into the three transaction phases:
<steals left="thief" right="child" before="focused">
The thief eyes the teddy bear maliciously and wrenches it
from the child's passionate grip.
</steals>
<steals left="thief" right="child" tension="focused">
The child becomes furious and starts crying.
</steals>
<steals left="thief" right="child" after="focused">
The thief triumphantly stows the teddy bear in his backpack.
</steals>Perhaps this sequence of events is ghastly, but here the goal is merely to model emotions accurately. In this example, the three phases unambiguously correspond to "anxiety," "anger," and "drunk with accomplishment."
The attributes before, tension, and after describe the emotional tension. Tension modifies a general emotion to fit the precise situation. For example:
<steals left="thief" right="child" before="relaxed">
The thief grabs the teddy bear with abandon.
</steals>
<steals left="thief" right="child" before="stifled">
The thief studies whether the teddy bear can be stolen.
</steals>A tension modifier is either focused, relaxed, or stifled. Emotion correspondance charts can assist in fitting the most accurate modifier to the situation. Of course a thief is not always successful:
<steals left="thief" right="child" before="stifled">
The thief studies whether the teddy bear can be stolen.
</steals>
<impasse left="child" right="thief" tension="focused">
The child grips the teddy bear even more tightly.
</impasse>The four situations involving spin (steals, exposes, admires, and accepts) mostly follow the three phase transaction structure. However, flexibility offered by accepts shows the need to explicitly indicate an initiator. For example:
<accepts left="chef" right="child" before="focused">
I am hungry. Is dinner ready?
</accepts>
<accepts left="chef" right="child" before="focused">
You are hungry!
</accepts>Looking just at the annotation, a reader cannot determine whether the child is making a demand, or the chef is voicing the child's expression of hunger. An initiator attribute eliminates the ambiguity:
<accepts left="chef" right="child" before="focused" initiator="right">
I am hungry. Is dinner ready?
</accepts>
<accepts left="chef" right="child" before="focused" initiator="left">
You are hungry!
</accepts>Now the situations can be distinguish. While the annotation is precise, it is also quite a lot of typing. To reduce the burden on the analyst, the parser is capable of guessing reasonable defaults by consideration of who is talking. For example:
<accepts left="chef" initiator="child" before="focused">
I am hungry. Is dinner ready?
</accepts>Or alternately:
<talk who="child">
<accepts left="chef" before="focused">
I am hungry. Is dinner ready?
</accepts>
</talk>We covered most of the simple variations. Here is a summary:
situation variations
---------- ---------------------------------
steals before|tension|after, initiator
exposes before|tension|after, initiator
admires before|tension|after, initiator
accepts before|tension|after, initiator
impasse tension
observes intensity, [initiator]
uneasy intensity, [initiator]
ready intensity
destroys -The square brackets denote optional attributes. More will be said about this later.
Reaction
Question & answer (Q&A) format is a common way to structure discussion. This section looks at the Q&A structure in comparison with the competition model. It will be shown that there are some conflicts in the corresponce, however, the complementary features can be embraced.
Questions usually end with a question mark. Even so, from the point of view of competition, questions are essentially demands. For example:
<accepts left="child" initiator="parent" before="focused">
Does your teddy bear have brown fur?
</accepts>In answer, the child's choices include:
<accepts initiator="child" right="parent" before="focused">
Yes!
</accepts>Or:
<impasse initiator="child" right="parent" tension="relaxed">
Which teddy bear?
</accepts>Notice that similar replies could be expected from the non-question demand, "I want the teddy bear." For example, accepts ("OK, here it is.") or impasse ("Which teddy bear?"). While the question mark inflects the demand somehow, its presence or absence does not seem particularly significant in the context of competition.
The answer side of Q&A will now be considered. Answers are generally offered in reaction to questions. However, it is easy to demonstrate ambiguity. For example:
<accepts left="child" initiator="parent" before="focused">
Does your teddy bear have brown fur?
</accepts>
<impasse initiator="child" right="parent" tension="relaxed">
Which teddy bear?
</accepts>Is the child's reply a question or an answer? Perhaps both, but clearly the Q&A terminology is imprecise. Despite these problems, some kind of contextual relationship definitely exists. To encode the Q&A relationship precisely, a clear standard is needed to recognize it.
The standard suggested here is that the participants must match but the initiator must flip-flop. These rules were discovered by surveying many Q&A style exchanges and the rules seem to work well in practice. re (meaning "in reaction to") and id attributes serve the purpose. For example:
<accepts id="q1" right="child" initiator="adult" before="focused">
Does your teddy bear have brown fur?
</accepts>
<accepts re="q1" left="parent" initiator="child" before="focused">
Yes!
</accepts>Quality
So far, the chosen markup may seem somewhat arbitrary. For example, why indicate merely accepts?
<accepts id="q1" right="child" initiator="adult" before="focused">
Does your teddy bear have brown fur?
</accepts>Why not more detail, such as:
<accepts id="q1" right="child" initiator="adult" before="focused">
Does your
<admires initiator="child" right="teddy bear" tension="relaxed">
teddy bear
</admires>
have brown fur?
</accepts>In fact, either annotation could be accurate. Part of the problem is that both the story and annotation are being written by the same person. This is a conflict of interest because i can adjust the story to improve my annotations, and visa-versa. Recall that the goal here is to develop the best quality annotations *without* changing the story.
Another problem is that written text is not rich enough as a medium to thoroughly convey life's emotional textual. To contrast, film (or the actual life experience itself) is quite sufficiently expressive. So assuming we are watching a film, what consistitutes best quality markup? For example:
<steals initiator="Yupi" right="Lon" before="focused">
Yupi gives Lon a beautiful and fragrant flower.
</steals>Most likely this markup is wrong. A proposed criteria for judging the quality of markup follows:
Is it detailed enough to isolate single emotions?
Looking at a scene as a sequences of moments, every moment can be consider an *independent* competition event. The question is: into how many intervals must we divide the timeline such that each interval is exactly one emotion?
Does it span the entire duration of the emotion?
Some clarity is lost if a single emotion is sub-divided excessively. A single annotation should span as much time as is practical. The duration criteria is also consided in the next section, "Resolution."
Resolution
The parser tries to keep track of situations which remain unresolved (or pending). For example:
<accepts id="o1" initiator="clerk" left="Joey" before="focused">
What toppings do you want on your pizza?
</accepts>
<impasse id="o2" re="o1" left="Joey" right="cleck" tension="stifled">
Joey puts the phone on hold.
</impasse>In general, stories don't stop abruptly leaving unresolved cliffhangers. Here is a clue that can be used to debug annotations. After processing a transcript, empathize outputs all the situations which do not find closure. In the above example, the situation is pending until Joey picks up the phone or the clerk gives up in disgust or whatever.
Similarly, a warning is triggered when the same situation is resolved more than once. For example:
<accepts id="o3" initiator="clerk" left="Joey" before="focused">
How many pizzas do you want?
</accepts>
<accepts re="o3" initiator="Joey" before="focused">
i want exactly three pizzas.
</accepts>
<accepts re="o3" initiator="Joey" before="focused">
i want exactly two pizzas.
</accepts>Notice the situation o3 is listed in the re attribute twice. This is unnecessarily ambiguous. There are a variety of ways to better model the scenario.
Revoke / Amend
Sometimes i say something which i wish i never said. The revoke attribute reflects this scenario.
<accepts id="q1" initiator="Spike" left="Jeff" before="focused">
What time is it?
</accepts>
<impasse id="a1" re="q1" tension="focused" right="Spike" initiator="Jeff">
No idea!
</observes>
<uneasy id="q2" amend="q1" initiator="Spike" left="Jeff" intensity="gentle">
Spike gives Jeff a threatening look.
</uneasy>
<accepts id="a2" revoke="a1" re="q1" before="focused" left="Spike"
initiator="Jeff">
Eleven hundred hours! Sir!
</accepts>q1 is resolved by a1 until q1 becomes pending again when a1 is revoked by a2. Now that q1 is again pending, the re attribute can be used in a2 to substitute a revised reaction.
To contrast, Spike's demand for the time in q1 is not revoked by his subsequent threatening look in q2. Here, what needs to be reflected is Spike's change in tone, but not change in fact. The amend attribute serves this purpose.
Echo
The echo attribute represents the repeated emphasis of the same situation.
<impasse id="m1" left="child" right="mother" tension="focused"
absent="right">
Where is my mother?
</impasse>
<impasse echo="m1" id="m2" left="child" right="mother" tension="stifled">
i miss my mom! :-(
</impasse>Once echo'd, m1 is no longer pending. Only m2 remains pending.
Absent
Another way to manage resolution is to declare the victim absent. Here is an example:
<talk who="teacher">
<accepts left="students" before="focused" absent="left">
The students listen to the teacher's tedious monologue.
</accepts>
</talk>In fact, the teacher may never know whether the situation was truly accepts, observes, or even merely:
<ready intensity="gentle" left="students" right="teacher">
The students are intermittently aware that the
teacher continues to talk.
</ready>Notice that the absent attribute is unnecessary in this case. A few situations do not trigger any need for resolution, and ready is one of them.
Wildcards
<accepts left="*" initiator="Crawford" before="focused">
Does anyone know the time?
</accepts>Talk about anthropomorphic rivals. XXX
Context
Here is a prior example updated with new markup:
<steals id="p1" initiator="thief" right="child" before="focused">
The thief eyes the teddy bear maliciously and wrenches it
from the child's passionate grip.
</steals>
<steals re="p1" left="thief" initiator="child" tension="focused">
The child becomes furious and starts crying.
</steals>
<steals context="p1" initiator="thief" right="child" after="focused">
The thief triumphantly stows the teddy bear in his backpack.
</steals>This last situation describing the thief gloating over his accomplishment is not really an answer to "p1", but there is certainly some kind of direct contextual connection. The context attribute is available for such cases. context is a catch-all. It is appropriate for any kind of generic relationship for which a more specific classification is not available.
EMOTION LIBRARY API
Emotion::set_transcript($xml_file)
my $dialog_id = Emotion::set_speaker($who)
my @pending = Emotion::unresolved()
Emotion::Atom Methods
my $situation = Emotion::Atom->new($expat, $attr)
my $emotion = $atom->emotion;
SEE ALSO
http://ghost-wheel.net