Abstract

We examine a variety of dialogue protocols, taking inspiration from two fields: natural language dialogue modelling and multiagent systems. In communicative interaction, one can identify different features that may increase the complexity of the dialogue structure. This motivates a hierarchy of abstract models for protocols that takes as a starting point protocols based on deterministic finite automata. From there, we proceed by looking at particular examples that justify either an enrichment or a restriction of the initial model. ...

Abstract

The belief that humans will be able to interact with computers in conversational speech has long been a favorite subject in science fiction. This reflects the persistent belief that spoken dialogue would be the most natural and powerful user interface to computers. With recent improvements in computer technology and in speech and language processing, such systems are starting to appear feasible. There are significant technical problems that still need to be solved before speech-driven interfaces become truly conversational. This paper describes ...

Note (first note only)

• TRIPS is at the high end of dialog management: planning-based and agent-based DM.
• There are several problem-solving agents, each with planning and reasoning abilities.
• The agents collaborate with the human to solve a problem (it's a collaborative, practical dialog).

The NLU is a best-first bottom-up chart parser, with semantic information. Its output is a sequence of speech-acts, such as "(ASSERT (:ID SA11 :SPEAKER USER :HEARER SYS :CONTENT (NEED...)))"

There are several DM

Abstract

We describe the Degrees of Grounding model, which tracks the extent to which material has reached mutual belief in a dialogue, and conduct experiments in which the model is used to manage grounding behavior in spoken dialogues with a virtual human. We show that the model produces improvements in virtual human performance as measured by post-session questionnaires. 1 ...

Note (first note only)

• Keep, for each piece of information, the degree to which it is understood, e.g. "Just introduced", "Introduced again", "acknowledged", "repeated'.
• Keep, for each piece of information, the degree to which we NEED it to be understood, e.g. sensitive information need higher degree.
• Use this to control grounding actions.

Abstract

We describe the dialogue model for the virtual humans developed at the Institute for Creative Technologies at the University of Southern California. The dialogue model contains a rich set of information state and dialogue moves to allow a wide range of behaviour in multimodal, multiparty interaction. We extend this model to enable non-team negotiation, using ideas from social science literature on negotiation and implemented strategies and dialogue moves for this area. We present a virtual human doctor who uses this model ...

Note (first note only)

Very complex dialog model, with emotions, plans and goals.

Implemented with SOAR.

The agent selects his attitude towards the negotiation: Cooperate, Avoid, or Attack.

Dialog acts: forward acts, backward acts, conversation management, grounding, turn-taking, initiative.

Featuring: Dr. Perez.

See also: The SASO project: http://vhtoolkit.ict.usc.edu/index.php/Projects#SASO .

Note (first note only)

Java-based DM framework.

TODO: contact the authors and ask what is currently used: danilom@stanford.edu, lcavedon@stanford.edu

Abstract

Reinforcement learning (RL) is a promising technique for creating a dialog manager. RL accepts features of the current dialog state and seeks to find the best action given those features. Although it is often easy to posit a large set of potentially useful features, in practice, it is difficult to find the subset which is large enough to contain useful information yet compact enough to reliably learn a good policy. In this paper, we propose a method for RL optimization which ...

Note (first note only)

• Mainly, a method for feature selection in RL.
• Dialog setting: call routing to a named person.
• 9 Possible actions: "AskName", "ConfirmName", etc.

Abstract

We propose a method for learning dialogue management policies from a fixed data set. The method addresses the challenges posed by Information State Update (ISU)-based dialogue systems, which represent the state of a dialogue as a large set of features, resulting in a very large state space and a huge policy space. To address the problem that any fixed data set will only provide information about small portions of these state and policy spaces, we propose a hybrid model that combines ...

Abstract

Spoken dialogue management strategy optimization by means of Reinforcement Learning (RL) is now part of the state of the art. Yet, there is still a clear mismatch between the complexity implied by the required naturalness of dialogue systems and the inability of standard RL algorithms to scale up. Another issue is the sparsity of the data available for training in the dialogue domain which can not ensure convergence of most of RL algorithms. In this paper, we propose to combine a ...

Note (first note only)

Main contribution is efficiency of learning.

• dialog setup: form-filling, tourist info.
• 13 actions: ask-slot*3, explicit-confirm...

Abstract

Spoken Dialogue Systems (SDS) are systems which have the ability to interact with human beings using natural language as the medium of interaction. A dialogue policy plays a crucial role in determining the functioning of the dialogue management module. Handcrafting the dialogue policy is not always an option, considering the complexity of the dialogue task and the stochastic behavior of users. In recent years approaches based on Reinforcement Learning (RL) for policy optimization in dialogue management have been proved to be ...

Note (first note only)

A detailed description and analysis of several dialog system architectures:

• The simple pipeline architecture and its limitations (asynchronous, backchannel...);
• The DARPA hub-and-spoke architecture and several examples - Mitre CommandTalk;
• The OAA architecture and several examples - SRI CommandTalk;
• The TrindiKit architecture (Information-State-Update) and examples - GoDiS, Conversational game player.
• The Verbmobil architecture (human-to-human, pools).
• Some proposed architectures for SIRIDUS.

Abstract

This paper shows how a dialogue system for information-seeking dialogues can be implemented in a type-theory-based syntax editor, originally developed for editing mathematical proofs. The implementation gives a simple logical metatheory to such dialogue systems and also suggests new functions for them, e.g., a local undo operation. The method developed provides a logically based declarative way of implementing simple dialogue systems that is easy to port to new domains. ...

Note (first note only)

DIALOG: Information-seeking.

SEMANTICS: records.

NLU: manually-built GF grammar.

Abstract

We present a new approach for rapidly developing dialogue capabilities for virtual humans. Starting from domain specification, an integrated authoring interface automatically generates dialogue acts with all possible contents. These dialogue acts are linked to example utterances in order to provide training data for natural language understanding and generation. The virtual human dialogue system contains a dialogue manager following the information-state approach, using finite-state machines and SCXML to manage local coherence, as well as explicit modeling of emotions and compliance level and a grounding component based on ...

Note (first note only)

SEMANTIC REPRESENTATION: XML dialog acts, automatically created from the domain description. Example:

hassan.assert <dialogue_act speaker="hassan"> <primitive_dialogue_act> <assertion> <object name="tax"> <attribute name="collector"> <value>hassan</value> </attribute> </object> </assertion> </primitive_dialogue_act> </dialogue_act> Indeed, you might say that I collect the taxes.

NLU: "The NLU uses a statistical language modeling text classification technique (Leuski et al. 2006) to map the text produced by the speech recognition to dialogue acts."

Note (first note only)

PROBLEM: noisy output of an ASR, in an information-seeking dialog.

SOLUTIONS:

• A. Fetch some candidates from the database, and try to select the best one;
• B. Select dialog-management actions, such as implicit/explicit confirmation, etc.

In both cases, the strategies were learned using a WOZ experiment, where the actions of the two most successful wizards were used for machine-learning a model.

Abstract

A field of spoken dialog systems is a rapidly growing research area because the performance improvement of speech technologies motivates the possibility of building systems that a human can easily operate in order to access useful information via spoken languages. Among the components in a spoken dialog system, the dialog management plays major roles such as discourse analysis, database access, error handling, and system action prediction. This survey covers design issues and recent approaches to the dialog management techniques for modeling the dialogs. We also explain the ...

Abstract

This work presents an agenda-based approach to improve the robustness of the dialog manager by using dialog examples and n-best recognition hypotheses. This approach supports n-best hypotheses in the dialog manager and keeps track of the dialog state using a discourse interpretation algorithm with the agenda graph and focus stack. Given the agenda graph and n-best hypotheses, the system can predict the next system actions to maximize multi-level score functions. To evaluate the proposed method, a spoken dialog system for a building guidance robot was developed. Preliminary evaluation shows ...

Abstract

This paper proposes a generic dialog modeling framework for a multi-domain dialog system to simultaneously manage goal-oriented and chat dialogs for both information access and entertainment. We developed a dialog modeling technique using an example-based approach to implement multiple applications such as car navigation, weather information, TV program guidance, and chatbot. Example-based dialog modeling (EBDM) is a simple and effective method for prototyping and deploying of various dialog systems. This paper also introduces the system architecture of multi-domain dialog systems using ...

Note (first note only)

CORPUS: collected from human-human conversations, and manually tagged with semantic tags.

The DM gets as input a semantic representation of the user utterance, the output of the ASR/NLU unit.

The DM searches an SQL database for a similar utterance, and proceeds accordingly.

Abstract

This paper describes the structural annotation of a spoken dialogue corpus. By statistically dealing with the corpus, the automatic acquisition of dialoguestructural rules is achieved. The dialogue structure is expressed as a binary tree and 789 dialogues consisting of 8150 utterances in the CIAIR speech corpus are annotated. To evaluate the scalability of the corpus for creating dialogue-structural rules, a dialogue parsing experiment was conducted. ...

Note (first note only)

CORPUS: CIAIR (in-car conversations for getting information about shops, parking, etc).

Collected from human-human, human-machine and human-WOZ conversations. 789 dialogs, 8150 utterances.

DIALOG MANAGEMENT: using a binary-tree that represents the dialog structure. The tree is learned from examples.

Abstract

This paper introduces a new dialogue management framework for goal-directed conversations. A declarative specification defines the domain-specific elements and guides the dialogue manager, which communicates with the knowledge sources to complete the specified goal. The user is viewed as another knowledge source. The dialogue manager finds the next action by a mixture of rule-based reasoning and a simple statistical model. Implementation in the flight-reservation domain demonstrates that the framework enables the developer to easily build a conversational dialogue system. ...

Note (first note only)

The author writes entities (e.g. flight), constraints (= when an entity is complete), and knowledge sources (e.g. date converter, flight database; or the user).

The dialog manager decides what knowledge source to consult.

Abstract

This paper describes an implementation of some key aspects of a theory of dialogue processing whose main concerns are to provide models of GROUNDING and of the role of DISCOURSE OBLIGATIONS in an agent's deliberation processes. Our system uses the TrindiKit dialogue move engine toolkit, which assumes a model of dialogue in which a participant's knowledge is characterised in terms of INFORMATION STATES which are subject to various kinds of updating mechanisms. ...

Abstract

This work shows how a dialogue model can be represented as a Partially Observable Markov Decision Process (POMDP) with observations composed of a discrete and continuous component. The continuous component enables the model to directly incorporate a confidence score for automated planning. Using a testbed simulated dialogue management problem, we show how recent optimization techniques are able to find a policy for this continuous POMDP which outperforms a traditional MDP approach. Further, we present a method for automatically improving handcrafted dialogue ...

Note (first note only)

• Main contribution: use the confidence level returned by the ASR/NLU component, to improve the performance of the DM, using reinforcement-learning.
• Evaluation: using a testbet simulated DM, in the travel-planning domain.

Abstract

We present the ADAMACH data centric dialog system, that allows to perform on- and offline mining of dialog context, speech recognition results and other system-generated representations, both within and across dialogs. The architecture implements a “fat pipeline” for speech and language processing. We detail how the approach integrates domain knowledge and evolving empirical data, based on a user study in the University Helpdesk domain ...

Note (first note only)

Application: information seeking dialog.

Main novelty: state-less dialog management, based on a database.

"DM becomes a function that maps ASR results and old IS to the TTS and ASR parameters and a new IS".

Abstract

We introduce an architecture and toolkit for building dialogue managers currently being developed in the TRINDI project, based on the notions of information state and dialogue move engine. The aim is to provide a framework for experimenting with implementations of different theories of information state, information state update and dialogue control. A number of dialogue managers are currently being built using the toolkit, and we present overviews of two of them. We believe that this framework ... ...

Abstract

Developing dialogue systems is a complex pro- cess. In particular, designing ecient dialogue management strategies is often dicult as there are no precise guidelines to develop them and no sure test to validate them. Several suggestions have been made recently to use reinforcement learning to search for the optimal management strategy for specic dialogue situations. These approaches have produced interesting results, including applications involving real world dia- logue systems. However, reinforcement learning suers from the fact that it is state based. In other words, the optimal strategy is expressed as a decision table specifying ...

Note (first note only)

• Shows how to generalize the policies learned by Reinforcement Learning, by using

Logic Reasoning.

• Shows how to define a dialog-management system in terms of states and actions, suitable for reinforcement learning.
• The example system is a form-filling question. The actions are: explicit/implicit/no confirmation, general/restricted grammar for the answer, etc.

Abstract

In this paper, we describe RavenClaw, a plan-based, task-independent dialog management framework. RavenClaw isolates the domain-specific aspects of the dialog control logic from domain-independent conversational skills, and in the process facilitates rapid development of mixed-initiative systems operating in complex, task-oriented domains. System developers can focus exclusively on describing the dialog task control logic, while a large number of domain-independent conversational skills such as error handling, timing and turn-taking are transparently supported and enforced by the RavenClaw dialog engine. To date, RavenClaw ...

Note (first note only)

See the RavenClaw page in Olympus wiki: http://www.cs.cmu.edu/~dbohus/ravenclaw-olympus/research.html

See sample dialogs here: http://www.cs.cmu.edu/~dbohus/ravenclaw-olympus/roomline.html

Abstract

Older adults are a challenging user group because their behaviour can be highly variable. To the best of our knowledge, this is the first study where dialogue strategies are learned and evaluated with both simulated younger users and simulated older users. The simulated users were derived from a corpus of interactions with a strict system-initiative spoken dialogue system (SDS). Learning from simulated younger users leads to a policy which is close to one of the dialogue strategies of the underlying SDS, while the simulated older users allow us to learn more ...

Note (first note only)

Training "simulated users" using a WOZ strategy is better when the users are old, because they are more versatile. "Simulated older users allowed us to learn a more flexible version of the strict system-initiative dialogue strategies that were used for creating the original corpus of interactions"!

DOMAIN: scheduling health-care appointments.

IMPLEMENTATION: "The human Wizard took over the function of speech recognition (ASR), language understanding (NLU), and dialogue management components".

CORPUS: "447 dialogues; The older users contributed 232 dialogues, the younger ones 215".

Note (first note only)

In a form-based dialog (such as a tourism information retrieval system), the DM must keep track of what the user wants to know, according to what he said so far. It does this by selecting from an ontology tree of all possible intentions (e.g. all possible venues the user might want to know about - restaurant, Kosher restautrants, etc.).

There is some probability of error in detecting what the user said in a particular utterance. We can improve the accuracy of the