Planning & Language Interest Group

A discussion group for automated planning, language, and related topics

Meeting archive


Instruction giving in virtual worlds

We will discuss the paper:

This paper reports on the results of the working group "Virtual Environments" at the Workshop on Shared Tasks and Comparative Evaluation for NLG.

Alexander Koller will give a brief introduction to the paper to motivate discussion.


What's next for plan-lang?

Original announcement: This meeting will be an open discussion, with a focus on organisational issues. In particular, this meeting will provide an opportunity to reflect on last term's meetings, and the direction we should take this term. Should we continue with the current format of the meetings? Should we try to identify a "shared task" in order to focus our discussions? Should we diversify our reading list?

Update: As a result of this meeting, we have decided to meet on a less regular basis and have more topic-oriented open discussions when we do not have speakers.



Problems and domains (Part V)

We continued our discussion on language problems and the role of uncertainty, incomplete information, and reasoning about knowledge/belief. This was the final meeting of 2007.

Reading list:


Problems and domains (Part IV)

We continued our discussion on problem domains with a focus on different approaches to the problem of planning under conditions of uncertainty and incomplete information.

Reading list:


Problems and domains (Part III)

We continued our discussion on problem domains with a focus on different approaches to the problem of planning under conditions of uncertainty and incomplete information.

Reading list:


Planning and Learning in Dialogue Systems

We'll first briefly survey the main uses of planning in dialogue systems:

  • domain planning, recipes, and plan recognition for input interpretation ("what in the 'world' is the user trying to do?": e.g. COLLAGEN, WITAS, DUDE, BEETLE etc);
  • dialogue planning ("what should I say next?": all dialogue systems);
  • planning for NLG ("how should I say it?": e.g. FLIGHTS).

Different planning techniques are appropriate for each of these tasks, but they must communicate via shared context representations. Focusing on dialogue planning, I'll describe recent advances in decision-theoretic/statistical dialogue planning (e.g. TALK project results) which deal with the issues of noise, uncertainty, and optimization. To illustrate, I'll demonstrate a system (REALL) learning to optimize its dialogue plans under different noise conditions and time constraints, for different types of user. I'll finally attempt to describe the main research questions and directions arising from these techniques.

This will be a joint meeting with the Dialogue Systems Group.


Problems and domains (Part II)

Reading list:


Representing and using assembly plans in cooperative, task-based human-robot dialogue

The JAST human-robot dialogue system is intended to allow a human and a robot to work together to assemble construction toys on a common work area, coordinating their actions through speech, gestures, and facial displays. I will describe how assembly plans are represented in this system and show how information from the assembly plan is used to guide the interaction.


Problems and domains (Part I)

Reading list:


Using Lexicalized Grammars and Headedness for Probabilistic Plan Recognition

This talk will discuss a new algorithm for plan recognition using an action grammar formalism based on Combinatory Categorial Grammar (CCG), that requires a significant shift in thinking about the problem of plan recognition. This approach makes significant use of the concepts of lexicalization and headedness from natural language parsing. It argues that lexicalization of action grammars can help plan recognition address a number of technical and efficiency questions, and that headedness in action grammars is previously unrecognized and an important addition to plan recognition theory in its own right.


AI Models for Automated Problem Solving

AI planners and solvers are described at two levels: the types of problems that they solve (models), and the solution methods (algorithms). For connecting Planning and NLP (be sentence production, language understanding, or story generation), it is probably useful to abstract away the second aspect crucial in planning though) and focus initially on the first: the models. These include Strips Planning Models, the SAT and Constraint Satisfaction (CSP) models, and variations, like Planning with Sensing, Weighted SAT and CSPs, and Conditional Weighted SAT and CSPs, etc. All these models are NP-hard and they are all closely related. Yet having a broader palette of models and their corresponding solvers is good for two reasons. First, once a problem is cast as an instance of a model, the problem can be fed into the corresponding solver. Second, a broader palette of models may be useful for thinking and formulating the problems in the first place.


Sentence Generation as Planning

In my talk, I will sketch my recent work on doing sentence generation by converting it into a planning problem and then running a planner. I will also point out some peculiarities of my system that make it tricky for current planners, and talk about the issues I plan to work on in Edinburgh.


Initial organisational meeting

It has recently become evident that a growing number of researchers within the School of Informatics are either working on, or interested in, the connections between planning and language. As a result, there has been interest in establishing a regular forum where issues related to planning and language could be presented and discussed. In this meeting we will discuss the organisational details and structure of future meetings.