Towards Learning Dialogue Structures from Speech Data and Domain Knowledge: Challenges to Conceptual Clustering using Multiple and Complex Knowledge Source Jens-Uwe Moller Natural Language Systems Division, Dept. of Computer Science, Univ. of Hamburg
Overview Dialog modeling based on a set of units called dialog act Dialog acts from theory doesn’t fit with a specific domain Labeling dialog is time consuming and subjective learn an application specific dialog acts from speech data using conceptual clustering
The learning task Learning dialog acts from turns Unsupervised classification (no prior definition of dialog acts is given) Hierarchy classification with inspectable classifying rules
Features Domain knowledge: structure of task, task knowledge represented by goals and plans Word recognizer: word hypotheses Prosodic data: Pause & Stress mark important unit Lexical semantics Syntax (less important in spoken dialog) Semantics (larger units of lexical semantics)
COWEB Symbolic machine learning algorithm Build a classification tree Distinction between subnodes are made from a function overall attribute Support probabilistic data Support multiple overlapping hierarchies (for ambiguous case) Can handle multiple entries of one attribute (e.g. stream of words)
COWEB (2) Learning from simultaneous events Learn from structure data: Conceptual Graphs. Learn case descriptions from terminological descriptions Subsumption = correclation criterion over structured data. e.g. subsumption of individuals to classes
Metrics for Measuring Domain Independence of Semantic Classes Andrew Pargellis, Eric Fosler-Lussier, Alexandros Potamianos, Chin-Hui Lee Dialogue Systems Research Dept., Bell Labs, Lucent Technologies Murray Hill, NJ, USA
Introduction Employ semantic classes (concepts) from another domain Need to identify domain-independent concepts base on comparison across domain Domain-independent concepts should occur in similar syntactic (lexical) contexts across domains
Comparing concepts across domains Concept-comparison method Concept-projection method
Concept-comparison method Find the similarity between all pairs of concepts across the two domains Two concepts are similar if their respective bigram contexts are similar Use left and right context bigram language models
Kullback-Leibler (KL) distance Compare how san francisco and newark are used in the Travel domain with how comedies and westerns are used in the Movie domain Distance between two concepts
Concept-projection method How well a single concept from one domain is represented in another domain. How the words comedies and westerns are used in both domains Useful for identifying the degree of domain- independence for a particular concept.
Result: Concept-comparison
Result: Concept-projection
Concept Example
Semi-Automatic Acquisition of Domain-Specific Semantic Structures Siu K.C., Meng H.M. Human-Computer Communications Laboratory Department of Systems Engineering and Engineering Management The Chinese University of Hong Kong
Grammar induction Use unannotated corpora Portable across domain & language Output grammar has reasonable coverage of within-domain data and reject out-of-domain data Amenable to interactive refinement by human Support optional injection of prior knowledge
Spatial clustering Use kullback-liebler distance. use left and right context. Consider word with pre-set minimum occurrence. (set to 5) use left and right context. Consider word w1, w2 (later be c1, c2) pair-wise for words that have a least pre-set minimum occurrence. (set to 5)
Temporal clustering Use Mutual Information (MI). N-highest MI pairs are clustered (N=5 in experiment) Do spatial clustering and temporal clustering iteratively Post-process by human
Automatic Concept identification In goal-oriented conversations Ananlada Chotimongkol and Alexander I. Rudnicky Language Technologies Institute Carnegie Mellon University
Concept identification First step towards the goal of automatically inferring domain ontologies Goal-oriented human-human conversation has a clear structure This structure can be used to automatically identify domain topics, e.g. dialog classfication
Clustering algorithm Hierarchical clustering Mutual information based Criterion=minimize the loss of average mutual information Kullback-Lierbler based Criterion=word pair with minimum distance
Evaluation metrics Reference concept from class-based n- gram model Cluster concept=majority concept Precision Recall Singularity score (SS) Quality score (QS)