Explanations in Dialogue Systems through Uncertain RDF Knowledge Bases
3 likes2,257 views
The document discusses the development of dialogue systems that utilize uncertain RDF knowledge bases to enhance human-computer interactions. The authors present a framework for explanation-aware multimodal dialogue systems which integrate complex data for answering user queries, thereby building user trust through transparent data representation. The work emphasizes the need for efficient reasoning in uncertain environments, aiming to provide accurate responses while acknowledging potential inconsistencies in the data.
Explanations in Dialogue Systems through Uncertain RDF Knowledge Bases
- 1. ECAI Exact Workshop Explanations in Dialogue Systems through Uncertain RDF Knowledge Bases Daniel Sonntag (DFKI) and Martin Theobald (MPII) 16/08/2010 “ Where was she born?”
- 2. Outline Introduction Dialogue System Framework URDF Framework Example Dialogue Conclusions |
- 4. Multimodal interaction with large and dynamic data repositories Important topic for the next generation of human-computer interaction systems. Provide new opportunities for semantically-enabled user interfaces. The explicit representation of the meaning of data allows us to transcend traditional keyboard and mouse interaction metaphors, and provide representation structures for more complex, collaborative interaction scenarios with more complex result presentations. Over the last years, we have adhered strictly to the developed rule “No presentation without representation,” -> Explanations Proper Explanations are one main factor that influences the level of user trust in complex (adaptive) AI systems
- 5. New area of IR Advent of structured databases in Semantic Web RDF structures and respective query languages, e.g., SPARQL, RDQL, or SERQL. These languages are based on the notion of RDF triple patterns, which can be connected via several query operators such as “union” or “filter”. For the next generation of human-computer interaction systems, explanation-based inference during data retrieval and uncertain knowledge plays a major role. In order to implement these properties, we use an efficient reasoning framework for graph-based, non-schematic RDF knowledge bases and SPARQL-like queries, Uncertain RDF (URDF).
- 6. Dialogue System Goals Provide a dialogue-based interaction with such a probabilistic advanced database while following the URDF model. Explanation-aware multimodal dialogue system Such dialogue systems can answer complex questions and provide additional multimedia material such as graphs or videos. Dialogue-based question answering (QA) Gain graph-based knowledge for dialogue-based explanations with confidences for inferred knowledge. Derived Explanation knowledge is a unique opportunity for enhancing trust especially in uncertain, inferred answers in human-computer interaction systems.
- 8. Dialogue System Framework: Goals and Requirements Multimodal interaction with the Semantic Web and the Internet of Services Components customisable to different use case scenarios Flexible adaptation to Input and output modalities Interaction devices Knowledge bases Ontologies for world, domain, and dialogue knowledge Related work: translate NL input into structured ontology based representations, cf. NLION, Aqualog, Orlakel.
- 9. Unique requirements and opportunities However, all of them deal with written keywords or simple semantic relations, e.g., X isDefinedAs Y. They do not focus on the much more complex explanation-based answering process while using a dialogue system.
- 10. Dialogue Shell Workflow Speech Interpretation Text Interpretation Gesture Interpretation Graphic Generation Text Generation Speech Interpretation Dialogue and Interaction Management Interactive Semantic Mediator Interactive Service Compo-sition eTFS/SPARQL SPARQL SPARQL SPARQL OWL-API Visualisation Visualisation Service External Information Sources Semantiic (Meta) Services RDF KOIOS (Yago Ontology) Remote Linked Data Services OWL AOIDE (Music Ontology) Text Summarisation Modality Fusion Presentation Planning Personalisation - Domain Model Context Model User Model
- 11. Shared Representation A shared representation and a common knowledge base ease the dataflow within the system and avoid costly and error-prone transformation processes (c.f. “ No presentation without representation ”). More precisely, an ontology-based representation of a user query can be used to create a query that can be posed to a (U)RDF endpoint.
- 12. Dialogue System Architecture Distributed Dialogue System Architecture Major components can be run on different servers. Dialogue system also acts as a middelware. Ontology-based dialogue platform (ODP). Establishes a new session for the client at the dialogue system. Invoke the central RDF/OWL repository.
- 13. Core Answering Workflow The backend access has been extended by addressing URDF. We use an Apache Tomcat server for this purpose. The presentation has to be adapted to the result of the URDF process, i.e., graph-based explanations. Basic processing chain
- 15. URDF Framework |
- 16. Yago and Info Boxes 09.10.10 Reasoning in Uncertain RDF Knowledge Bases IE error-prone and inherently noisy
- 17. Requirements In the next decades: Reasoning under uncertainty/incompleteness/ inconsistency will be a major challenge for data management. Fast Processing: Inference over entire Yago database w/20 Mio facts < 0.25 seconds! Knowledge Extraction fro the Web is incomplete and inconsistent. bornIn(Al_Gore, Washington_D.C.) bornIn(Al_Gore, New_York_City) bornIn(?x,?y) bornIn(?x,?z) ?y=?z KB
- 18. Inference rules vs. Integrity constraints People are not born on two different dates/in two different places, etc. People may live in more than one place, hence no formal inconsistency when more than one answer. People are not married to more than one person (at the same time? , in most countries ?) F C S RDF Triples Soft Rules (Horn) Hard Rules (Horn) -AvBvC BvC -> A
- 19. Soft Rules and Hard Rules
- 20. Dependency / Explanation Graph
- 21. Inference Engine Top-down Datalog (SLD resolution) Polynomial in #facts Exponential in |rules| May run into cycles! First-Order Rules R1: livesIn(?x, ?y) :- marriedTo(?x, ?z), livesIn(?z, ?y) R2: livesIn(?x, ?y) :- represents(?x, ?y) R3: livesIn(?x, ?y) :- governorOf(?x, ?y) livesIn(Bill, ?x) \/ R1 R3 R2 RDF Base Facts F1: marriedTo(Bill, Hillary) F2: represents(Hillary, New York) F3: governorOf(Bill, Arkansas) /\ F1 \/ R2 R3 R1 F2 X F3 … …
- 22. Weighted MaxSat and URDF MAX-SAT problem: “ Given a collection of clauses, we seek a variable assignment that maximizes the number of satisfied clauses. The weighted MAX-SAT problem assigns a weight to each clause, and seeks an assignment that maximizes the sum of the weights of the satisfied clauses. Both of these problems are NP-hard." (Joy, Mitchel and Borchers) Query processing with URDF consists of two phases: lookups of basic query patterns against the knowledge base, which involves both direct lookups of base facts in the knowledge base, but also recursively grounding rules and inferring new facts; an resolving potential inconsistencies by a second reasoning step in the form of a Weighted MAX-SAT solver, which yields the final truth assignments.
- 23. MaxSat Decision Weighted MaxSat decision -> Truth assignment Resolve hard constraints = hard rules Include weights of base facts Include weights of soft rules Probabilistic model on top (for inferred knowledge) Probabilities of base facts No soft rules! Key feature: Capture the resolution steps in form of a DAG over grounded rules (aka data provenance or lineage)
- 25. Interaction Sequence The dialogue concentrates around the questions about the URDF contents, i.e., factoid questions about celebrities, and the multimodal presentation of answer content. 1 U: “Where is Angela Merkel born?” 2 S: Shows corresponding result in a SIE. 3 U: “What do Angela Merkel and Al Gore have in common?” 4 S: Shows corresponding relation graph. 5 U: “Where does he live?” 6 S: Shows corresponding relation graph. 7 S: *Synthesises a summary of the graph’s interpretation.*
- 26. Complex Explanation Graph “ Where does he (Al Gore) live?”
- 27. Conclusions |
- 28. Conclusions and Outlook We have discussed explanations in dialogue systems through Uncertain RDF knowledge bases and presented URDF, a query engine for uncertain and potentially inconsistent RDF knowledge bases. We gained graph-based knowledge for dialogue-based explanations with confidences for inferred knowledge. URDF outlook: Context-sensitive reasoning and explanations URDF provides a new data stream and explanations for future, AI based interaction systems. In our multimodal application scenario, a more complex and tighter integration of the provided result graphs has to be investigated. Outlook: the result graph could be presented in conjunction with the speech synthesis “I think he lives in Washington, D.C., because his wife, Tipper Gore, also lives there.”
- 29. Thank you! Questions? Supported by Acknowledgements: DFKI: Robert Nesselrath, Yajing Zang, Matthieu Deru, Simon Bergweiler, Gerhard Sonnenberg, Norbert Reithinger, Gerd Herzog, Alassane Ndiaye, Tilman Becker, Norbert Pfleger, Alexander Pfalzgraf, Jan Schehl, Jochen Steigner, and Colette Weihrauch MPII: Maximilian Dylla, Saransh Gupta, Javeria Iqbal, Timm Meiser, Ndapa Nakashole, Mohamed Yahya, Christina Teflouidi, Vinay Setty, Yafang Wang. Special guests:Mauro Sozio, Fabian Suchanek