tus_feature_handler

Feature handler for the TUS simulator.

For simplicity, the feature handler supports a single domain unlike the original implementation.

Classes

TUSFeatureHandler

Initializes the feature handler.

Module Contents

class tus_feature_handler.TUSFeatureHandler(domain: usersimcrs.core.simulation_domain.SimulationDomain, max_turn_feature_length: int, context_depth: int, agent_actions: List[str], user_actions: List[str] = ['inform', 'request'])

Bases: usersimcrs.simulator.neural.core.feature_handler.FeatureHandler

Initializes the feature handler.

Parameters:

• domain – Domain knowledge.

• max_turn_feature_length – Maximum length of a turn feature vector.

• context_depth – Number of previous turns to include in the input vector.

• agent_actions – Agent actions.

• user_actions – User actions. Defaults to [“inform”, “request”].

reset() → None

Resets the feature handler.

reset_user_feature_history() → None

Resets the user feature history.

update_action_slots(slots: Iterable[str]) → None

Updates the action slots.

Action slots are slots present in the information need and mentioned during the conversation.

Parameters:

slots – Slots mentioned in an utterance.

get_basic_information_feature(slot: str, information_need: usersimcrs.core.information_need.InformationNeed, state: usersimcrs.dialogue_management.dialogue_state.DialogueState, previous_state: usersimcrs.dialogue_management.dialogue_state.DialogueState) → usersimcrs.simulator.neural.core.feature_handler.FeatureVector

Builds feature vector for basic information.

It concatenates the value in agent state, user state, slot type, completion status, and first mention.

Parameters:

• slot – Slot.

• information_need – Information need.

• state – Current state.

• previous_state – Previous state.

Returns:

Feature vector for basic information.

get_agent_action_feature(agent_dialogue_acts: List[dialoguekit.core.dialogue_act.DialogueAct]) → usersimcrs.simulator.neural.core.feature_handler.FeatureVector

Builds feature vector for agent action.

It concatenates action vectors represented as 3-dimensional vectors that describe whether the slot and value are absent, only slot is present, or both slot and value are present.

Parameters:

agent_dialogue_acts – Agent dialogue acts.

Returns:

Feature vector for agent action.

get_slot_index_feature(slot: str) → usersimcrs.simulator.neural.core.feature_handler.FeatureVector

Builds feature vector for slot index.

Parameters:

slot – Slot.

Returns:

Feature vector for slot index.

get_slot_feature_vector(slot: str, previous_state: usersimcrs.dialogue_management.dialogue_state.DialogueState, state: usersimcrs.dialogue_management.dialogue_state.DialogueState, information_need: usersimcrs.core.information_need.InformationNeed, agent_dialogue_acts: List[dialoguekit.core.dialogue_act.DialogueAct], user_action_vector: torch.Tensor = None) → usersimcrs.simulator.neural.core.feature_handler.FeatureVector

Builds the feature vector for a slot.

It concatenates the basic information, user action, agent action, and slot index feature vectors.

Parameters:

• slot – Slot.

• previous_state – Previous state.

• state – Current state.

• information_need – Information need.

• agent_dialogue_acts – Agent dialogue acts.

• user_action_vector – User action feature vector (output vector for previous turn). Defaults to None.

Returns:

Feature vector for the slot.

get_turn_feature_vectors(agent_dialogue_acts: List[dialoguekit.core.dialogue_act.DialogueAct], previous_state: usersimcrs.dialogue_management.dialogue_state.DialogueState, state: usersimcrs.dialogue_management.dialogue_state.DialogueState, information_need: usersimcrs.core.information_need.InformationNeed, user_action_vectors: Dict[str, torch.Tensor] = {}) → List[usersimcrs.simulator.neural.core.feature_handler.FeatureVector]

Builds the feature vectors for a turn.

It comprises the feature vectors for all slots that in the information need and mentioned during the conversation.

Parameters:

• agent_dialogue_acts – Agent dialogue acts.

• previous_state – Previous state.

• state – Current state.

• information_need – Information need.

• user_action_vectors – User action feature vectors per slot. Defaults to an empty dictionary.

Returns:

Feature vectors for the turn.

build_input_vector(agent_dialogue_acts: List[dialoguekit.core.dialogue_act.DialogueAct], previous_state: usersimcrs.dialogue_management.dialogue_state.DialogueState, state: usersimcrs.dialogue_management.dialogue_state.DialogueState, information_need: usersimcrs.core.information_need.InformationNeed, user_action_vectors: Dict[str, torch.Tensor] = {}) → Tuple[List[usersimcrs.simulator.neural.core.feature_handler.FeatureVector], usersimcrs.simulator.neural.core.feature_handler.FeatureMask]

Builds the input vector $V_{input}$ for a turn.

It concatenates the feature vectors for the last n turns separated by a special token. Note that is inferred from the list of feature vectors provided. The input vector is structured as follows: [CLS] $V^t$ [SEP] $V^{t-1}$ [SEP] … [SEP] $V^{t-n}$ [SEP] {padding}, where {padding} indicates the padding to reach the maximum length.

Parameters:

• agent_dialogue_acts – Agent dialogue acts.

• previous_state – Previous state.

• state – Current state.

• information_need – Information need.

• user_action_vectors – User action feature vectors per slot. Defaults to an empty dictionary.

Returns:

Input vector.

get_label_vector(user_utterance: dialoguekit.core.annotated_utterance.AnnotatedUtterance, current_state: usersimcrs.dialogue_management.dialogue_state.DialogueState, information_need: usersimcrs.core.information_need.InformationNeed) → usersimcrs.simulator.neural.core.feature_handler.FeatureVector

Builds the label vector for a turn.

Parameters:

• user_utterance – User utterance with annotations.

• current_state – Current state.

• information_need – Information need.

Returns:

Label vector for the turn.

save_handler(path: str) → None

Saves the feature handler.

Parameters:

path – Path to the output file.

classmethod load_handler(path: str) → TUSFeatureHandler

Loads a feature handler from a given path.

Parameters:

path – Path to load the feature handler from.

Raises:

FileNotFoundError – If the file is not found.

Returns:

Feature handler.