Well File:

Saarland University

Improving Variational Encoder-Decoders in Dialogue Generation

AAAI Conferences

Variational encoder-decoders (VEDs) have shown promising results in dialogue generation. However, the latent variable distributions are usually approximated by a much simpler model than the powerful RNN structure used for encoding and decoding, yielding the KL-vanishing problem and inconsistent training objective. In this paper, we separate the training step into two phases: The first phase learns to autoencode discrete texts into continuous embeddings, from which the second phase learns to generalize latent representations by reconstructing the encoded embedding.  In this case, latent variables are sampled by transforming Gaussian noise through multi-layer perceptrons and are trained with a separate VED model, which has the potential of realizing a much more flexible distribution. We compare our model with current popular models and the experiment demonstrates substantial improvement in both metric-based and human evaluations.

Towards Clause-Learning State Space Search: Learning to Recognize Dead-Ends

AAAI Conferences

We introduce a state space search method that identifies dead-end states, analyzes the reasons for failure, and learns to avoid similar mistakes in the future. Our work is placed in classical planning. The key technique are critical-path heuristics h C , relative to a set C of conjunctions. These recognize a dead-end state s, returning h C (s) = infty, if s has no solution even when allowing to break up conjunctive subgoals into the elements of C. Our key idea is to learn C during search. Starting from a simple initial C, we augment search to identify unrecognized dead-ends s, where h C (s) < infinity. We design methods analyzing the situation at such s, adding new conjunctions into C to obtain h C (s) = infty, thus learning to recognize s as well as similar dead-ends search may encounter in the future. We furthermore learn clauses phi where s' not satisfying phi implies hC(s') = infty, to avoid the prohibitive overhead of computing h C on every search state. Arranging these techniques in a depth-first search, we obtain an algorithm approaching the elegance of clause learning in SAT, learning to refute search subtrees. Our experiments show that this can be quite powerful. On problems where dead-ends abound, the learning reliably reduces the search space by several orders of magnitude.

Red-Black Relaxed Plan Heuristics

AAAI Conferences

Despite its success, the delete relaxation has significant pitfalls. Recent work has devised the red-black planning framework, where red variables take the relaxed semantics (accumulating their values), while black variables take the regular semantics. Provided the red variables are chosen so that red-black plan generation is tractable, one can generate such a plan for every search state, and take its length as the heuristic distance estimate. Previous results were not suitable for this purpose because they identified tractable fragments for red-black plan existence, as opposed to red-black plan generation. We identify a new fragment of red-black planning, that fixes this issue. We devise machinery to efficiently generate red-black plans, and to automatically select the red variables. Experiments show that the resulting heuristics can significantly improve over standard delete relaxation heuristics.