1.  Ontology Induction (Chen et al., 2013 & 2014) Frame-semantic parsing on ASR results (Das et al., 2013) frame  slot candidate lexical unit  slot filler  1st Issue: How to induce domain-specific concepts? locale_by_use food expensiveness seeking relational_quantity PREP_FOR NN AMOD desiring DOBJ type food pricerange DOBJ AMOD task area PREP_IN Domain: restaurant recommendation in an in- car setting (WER = 37%) o Dialogue slots: addr, area, food, phone, postcode, pricerange, task, type speak on topic addr area food phone part orientational direction locale part inner outer food origin contacting postcode price range task type sending commerce scenario expensiveness range seeking desiring locating locale by use building reference ontology w/ the most frequent dependencies Can a dialogue system automatically learn open domain knowledge and then understand users? Unsupervised Learning and Modeling of Knowledge and Intent for Spoken Dialogue Systems Yun-Nung (Vivian) Chen yvchen@cs.cmu.edu Goal Framework Restaurant Asking Conversations target food price seeking quantity PREP_FOR NN AMOD Domain-Specific Ontology Unlabeled Collection Knowledge Acquisition Ontology Induction Structure Learning Domain-Specific Ontology price=“cheap”, target=“restaurant” behavior=navigation SLU Modeling by MF SLU Component “can i have a cheap restaurant” Semantic Decoding Behavior Prediction Knowledge Acquisition SLU Modeling by Matrix Factorization  Structure Learning (Chen et al., 2015a) Typed syntactic dependencies on ASR https://github.com/yvchen/MRRW/ ccomp amod dobj nsubj det canihave a cheap restaurant capability expensiveness locale_by_use induced ontology  Semantic Decoding (Chen et al., 2015b) concept  semantic slot  Behavior Prediction concept  user behavior 1)Given unlabeled conversations, how can a system automatically induce and organize domain-specific concepts? 2)With the automatically acquired knowledge, how can a system understand utterances? 1 Utterance 1 i would like a cheap restaurant Feature Observation Semantic Concept (Slot / Behavior) Train … … … cheap restaurant food expensiveness 1 target 11 find a restaurant for chinese food Utterance 2 1 1 food 1 1 1 Test 1 1.97.90.95.85.93.92.98.05 Feature Relation Model Concept Relation Model Reasoning with Matrix Factorization Semantic Concept Induction SLU Model Semantic Representation “can I have a cheap restaurant” Ontology Induction Unlabeled Collection SLU Modeling by MF FfFf FsFs Feature Model RfRf RsRs Relation Propagation Model Feature Relation Model Concept Relation Model. Knowledge Acquisition Structure Learning o Relation Propagation Model  Feature Knowledge Graph  Concept Knowledge Graph Assumption: The domain- specific features/concepts have more dependency to each other. Relation matrices allow each node to propagate scores to its neighbors in the knowledge graph, so that domain-specific features/concepts have higher scores during training.  2nd Issue: Hidden semantics cannot be observed but may benefit understanding performance. Good! ? Feature Relation Model Concept Relation Model RfRf RsRs ‧ 1 Feature Concept Train cheap restaurant food expensiveness 1 locale_by_use 1 1 11 food 1 1 1 Test 1 1 Ontology Induction i like 1 1 capability 1 Utterance 1 i would like a cheap restaurant … find a restaurant with chinese food Utterance 2 show me a list of cheap restaurants Test Utterance.90.97.95.85 hidden semantics o Matrix Factorization (MF)  Model implicit feedback  Objective: 1 MF learns a set of well-ranked concepts per utterance. Chen et al., “Unsupervised Induction and Filling of Semantic Slots for Spoken Dialogue Systems Using Frame-Semantic Parsing,” in Proc. of ASRU, 2013. Chen et al., “Leveraging Frame Semantics and Distributional Semantics for Unsupervised Semantic Slot Induction in Spoken Dialogue Systems,” in Proc. of SLT, 2014. Chen et al., “Jointly Modeling Inter-Slot Relations by Random Walk on Knowledge Graphs for Unsupervised Spoken Language Understanding,” in Proc. of NAACL, 2015a. Chen et al., “Matrix Factorization with Knowledge Graph Propagation for Unsupervised Spoken Language Understanding,” in Proc. of ACL-IJCNLP, 2015b.