the process of creating a character is iterative in nature. The following steps create an initial character that will probably need to be refined by repeating the steps as required to obtain the desired behavior.
Three actions are required to create the content necessary to drive the natural language component:
The entire information that defines a character, e.g. CakeVendor, for the FLoReS system is defined in a set of files sitting in the directory resources/characters/CakeVendor/
This directory contains three sub-directories that parallel the 3 steps defined above:
Authoring the content consists of editing 2 files. One for the user utterances and one for the system utterances. The file that contains the user utterances is basically the training data for the natural language understanding (NLU) module. The system utterance file instead contains the utterances that the character can say.
The NLU module given an utterance returns the most probably identifying strings. It's based on a maximum entropy multiclass classifier and therefore the user utterance file should list utterances maintaining their natural frequency. That is, the best way to obtain these utterances is by running wizard of oz experiments or role plays. Then annotate the data by assigning to each utterance said by a user during these experiments an identifying string. These identifying strings are sometime called speech acts or dialogue acts (in case more domain specific semantic is attached to the basic speech act). Examples of dialogue acts are: question.age to mark all utterances in which the user is asking about the age of the addressee.
When we design a dialogue policy for the character using this content, whenever we want to wait for the user to say a certain utterance, we will use the string identifier (speech act) associated to that utterance.
These utterances are the one the system can say. Similarly to the user utterances, each utterance has a specific string identifier. When designing the dialogue policy, if we want to say a certain system utterance, we will use the corresponding identifier (also here we call the identifier speech act).
The user and system utterances files use the same Excel spreadsheet format. These files have a number of columns (these are the initial 2 rows of the system utterances file for the character used in the example below):
TTERANCE_ID | VERSION | CHARACTER | STATE | SPEECH_ACT | TEXT |
statement.not-understand | I'm sorry, I didn't understand what you said. Please try to rephrase it. | ||||
greeting.hello | Hello |
The only 2 columns of relevance are SPEECH_ACT and TEXT.
SPEECH_ACT contains the string identifier for the corresponding utterance found in the TEXT column.
The user utterance file needs to be called user-utterances.xlsx and the system utterance file must be called system-utterances.xlsx (these names can be configured, but the default configuration looks for those names in each character available).
The FLoReS (Forward Locking Reward Seeking) dialogue manager is an information state and event driven dialogue manager. That is it does nothing unless an event is received. When an event is received it searches for the best action (i.e. sub-dialogue) that can be executed in the current information state and that achieves the highest expected reward. Once the best action is found it start executing it. Unless:
As mentioned earlier the dialogue manager searches for the best available action every time an event comes in.
Actions are also called sub-dialogues and define dialogue trees. For example this is one sub-dialogue found in the CakeVendor example below:
These sub-dialogue trees define a small self-contained portion of conversation. the criteria to use to decide what should be a sub-dialogue is similar to the criteria used to decide what should be a function or method in a programming language: generality and reusability.
For example, the sub-dialogue above takes care of finding out whether the system can sell to the user a cake with normal sugar or with Xylitol based on collecting information about the user having many cavities or having diabetes. Because the utterances found in this sub-dialogue can happen only in that specific context, then it makes sense to keep them in the same sub-dialogue.
A sub-dialogue can be in 3 states: ACTIVE, INACTIVE and DORMANT.
At any time in the system there is at most 1 active sub-dialogue: the current action. As said above, in some cases there may be no active actions. All actions are normally inactive, unless they have been active and they have been substituted (swapped-out) by another action before their natural termination (that is, at some point the system found a better action and so changed the state of the current action to dormant and made the newly found best action as active). Not all actions that are active and are swapped out for a new best action can become dormant. Some will go back directly to the inactive state. An action, to be allowed to become dormant, must have special entry paths that allow for it to be awoken back to the active state in case it becomes again the best action.
A sub-dialogue has multiple entry paths. The entry paths have a specific order (decided by the author) and each entry path has conditions to regulate when it can be taken and has also a start state. That is when the system during the search for the best available action considers a certain sub-dialogue, it'll considers all the possible entry paths in the order specified. The first that has satisfied conditions will be taken and it'll start the execution of the action at the specified start state in the sub-dialogue tree.
The possible types of entry paths are:
The edges of a sub-dialogue tree are of three types:
Nodes can have effects. There are two types of effects:
Each sub-dialogue is terminated when the execution path reaches a node that has no more outgoing edges.
Each sub-dialogue can be marked final. That means that when the end node of a final sub-dialogue is reached, the conversation ends. When the conversation ends the DM will ignore all events and the user will not be able to interact with the virtual character anymore.
Execution of a sub-dialogue consists of taking a certain entry path (the one that lead to the maximum expected reward) and then at every node, take the first outgoing edge (the order is from left to right and is specified by the author) that can be taken (that is has a satisfied condition) until we reach a waiting point: a user state (i.e. a state with user outgoing edges). At that point the dialogue manager terminates the execution and waits for the next event. If the incoming event is one of the expected events (i.e. the events specified in the user edges) then the execution continues along the first satisfied user edge. If the final node is reached, the sub-dialogue is terminated and becomes inactive and the system searches for a new optimal action to start executing.
The information state is formed by variables and stores the current state of the conversation. Three things can update the information state:
When an event is received, the dialogue manager (DM) checks to see if it is expected by the current action (i.e. the current action is at a user node and one of the user outgoing edges is waiting for the received event). If the current action is waiting for the received event the DM will continue the execution of the current action. Otherwise it'll execute a forward search to find the best action to execute. The forward search simulates possible future conversations. It's a breath first search and it's limited by time and depth (i.e. it'll always return quickly even if the search space is huge). Currently the limits are: 250ms or 10 levels maximum (i.e. the dialogue manager terminates the search for the optimal action after 250ms or if the search graph that represents the possible future conversations reaches a depth of 10 sub-dialogues, that is the search had enough time (i.e. within the 250ms timeout) to explore all possible conversations made up using a sequence of 10 sub-dialogues).
When executing the forward search, two options are possible:
The best action is the one that maximizes the expected reward. More precisely the formula is:
The dialogue policy is composed by several files. The main file that defines it is called policy.xml (also this name can be configured, but this is the default name).
A typical policy.xml file will look like the following:
<policy xmlns:xi="http://www.w3.org/2001/XInclude"> <xi:include href="initKB.xml"/> <xi:include href="goals.xml"/> <stepDiscount value="0.9"/> <include href="textFormat/policy.txt"/> </policy>
line 2 specifies the file used to define all the variables in the information state and to initialize them.
line 3 specifies the file that defines the basic value of the rewards available in this dialogue policy.
line 4 specifies the discount factor alpha mentioned in the expected reward formula.
line 5 includes a file that specifies some operators (actions/sub-dialogues) in a particular text format. One could specify the sub-dialogue trees directly in a xml variant but it's harder and so we prefer to document how to design operators using this special text format. One can have ny number of text format files included so one can organize the operators in multiple files.
CakeVendor.zip contains all is required to define a CakeVendor character that is an extension of the character created in this other tutorial for NPCEditor.