Utilizing amortized variational inference for latent-action reinforcement learning (RL) has been shown to be an effective approach in Task-oriented Dialogue (ToD) systems for optimizing dialogue success.Until now, categorical posteriors have been argued to be one of the main drivers of performance. In this work we revisit Gaussian variational posteriors for latent-action RL and show that they can yield even better performance than categoricals. We achieve this by introducing an improved variational inference objective for learning continuous representations without auxiliary learning objectives, which streamlines the training procedure. Moreover, we propose ways to regularize the latent dialogue policy, which helps to retain good response coherence. Using continuous latent representations our model achieves state of the art dialogue success rate on the MultiWOZ benchmark, and also compares well to categorical latent methods in response coherence.

Personalization is a crucial aspect of many online experiences. In particular, content ranking is often a key component in delivering sophisticated personalization results. Commonly, supervised learning-to-rank methods are applied, which suffer from bias introduced during data collection by production systems in charge of producing the ranking. To compensate for this problem, we leverage contextual multi-armed bandits. We propose novel extensions of two well-known algorithms viz. LinUCB and Linear Thompson Sampling to the ranking use-case. To account for the biases in a production environment, we employ the position-based click model. Finally, we show the validity of the proposed algorithms by conducting extensive offline experiments on synthetic datasets as well as customer facing online A/B experiments.

Natural Language Understanding (NLU) models on voice-controlled speakers face several challenges. In particular, music streaming services have large catalogs, often containing millions of songs, artists, and albums and several thousands of custom playlists and stations. In many cases there is ambiguity and little structural difference between carrier phrases and entity names. In this work, we describe how we leveraged multi-armed bandits in combination with implicit customer feedback to improve accuracy and personalization of responses to voice request in the music domain. Our models are tested in a large-scale industrial system containing several other components. In particular, we focused on using this technology to correct errors made by upstream NLU models and personalize responses based on customer preferences and music provider functionality. The models resulted in significant improvement of playback rate for Amazon Music and are deployed in systems serving several countries and languages. We further used the implicit feedback of the customers to generate weakly labeled training data for the NLU models. This improved the experience for customers using other music providers on all Alexa devices.

Text-based knowledge extraction methods for populating knowledge bases have focused on binary facts: relationships between two entities. However, in advanced domains such as health, it is often crucial to consider ternary and higher-arity relations. An example is to capture which drug is used for which disease at which dosage (e.g. 2.5 mg/day) for which kinds of patients (e.g., children vs. adults). In this work, we present an approach to harvest higher-arity facts from textual sources. Our method is distantly supervised by seed facts, and uses the fact-pattern duality principle to gather fact candidates with high recall. For high precision, we devise a constraint-based reasoning method to eliminate false candidates. A major novelty is in coping with the difficulty that higher-arity facts are often expressed only partially in texts and strewn across multiple sources. For example, one sentence may refer to a drug, a disease and a group of patients, whereas another sentence talks about the drug, its dosage and the target group without mentioning the disease. Our methods cope well with such partially observed facts, at both pattern-learning and constraint-reasoning stages. Experiments with health-related documents and with news articles demonstrate the viability of our method.

We demonstrate BioNex, a system to mine, rank and visualize biomedical news events. BioNex takes biomedical queries such as "Ebola virus disease" and retrieves the k most relevant news events for them. To achieve this we first mine the generic news events by clustering them on a daily basis using general named entities and textual features. These clusters are also tagged with disambiguated biomedical entities which aid in biomedical news event exploration. The clusters are then used to compute the importance scores for the event clusters based on a combination of textual, semantic, popularity and historical importance features. BioNex also visualizes the retrieved event clusters to highlight the top news events and corresponding news articles for the given query. The visualization also provides the context for news events using (1) a chain of historically relevant news event clusters, and (2) other non-biomedical events from the same day.

Analyzing and explaining relationships between entities in a knowledge graph is a fundamental problem with many applications. Prior work has been limited to extracting the most informative subgraph connecting two entities of interest. This paper extends and generalizes the state of the art by considering the relationships between two sets of entities given at query time. Our method, coined ESPRESSO, explains the connection between these sets in terms of a small number of relatedness cores: dense sub-graphs that have strong relations with both query sets. The intuition for this model is that the cores correspond to key events in which entities from both sets play a major role. For example, to explain the relationships between US politicians and European politicians, our method identifies events like the PRISM scandal and the Syrian Civil War as relatedness cores. Computing cores of bounded size is NP-hard. This paper presents efficient approximation algorithms. Our experiments with real-life knowledge graphs demonstrate the practical viability of our approach and, through user studies, the superior output quality compared to state-of-the-art baselines.

BACKGROUND: Biomedical knowledge bases (KB's) have become important assets in life sciences. Prior work on KB construction has three major limitations. First, most biomedical KBs are manually built and curated, and cannot keep up with the rate at which new findings are published. Second, for automatic information extraction (IE), the text genre of choice has been scientific publications, neglecting sources like health portals and online communities. Third, most prior work on IE has focused on the molecular level or chemogenomics only, like protein-protein interactions or gene-drug relationships, or solely address highly specific topics such as drug effects. RESULTS: We address these three limitations by a versatile and scalable approach to automatic KB construction. Using a small number of seed facts for distant supervision of pattern-based extraction, we harvest a huge number of facts in an automated manner without requiring any explicit training. We extend previous techniques for pattern-based IE with confidence statistics, and we combine this recall-oriented stage with logical reasoning for consistency constraint checking to achieve high precision. To our knowledge, this is the first method that uses consistency checking for biomedical relations. Our approach can be easily extended to incorporate additional relations and constraints. We ran extensive experiments not only for scientific publications, but also for encyclopedic health portals and online communities, creating different KB's based on different configurations. We assess the size and quality of each KB, in terms of number of facts and precision. The best configured KB, KnowLife, contains more than 500,000 facts at a precision of 93% for 13 relations covering genes, organs, diseases, symptoms, treatments, as well as environmental and lifestyle risk factors. CONCLUSION: KnowLife is a large knowledge base for health and life sciences, automatically constructed from different Web sources. As a unique feature, KnowLife is harvested from different text genres such as scientific publications, health portals, and online communities. Thus, it has the potential to serve as one-stop portal for a wide range of relations and use cases. To showcase the breadth and usefulness, we make the KnowLife KB accessible through the health portal (http://knowlife.mpi-inf.mpg.de).

Knowledge bases (KB's) contribute to advances in semantic search, Web analytics, and smart recommendations. Their coverage of domain-specific knowledge is limited, though. This demo presents the KnowLife portal, a large KB for health and life sciences, automatically constructed from Web sources. Prior work on biomedical ontologies has focused on molecular biology: genes, proteins, and pathways. In contrast, KnowLife is a one-stop portal for a much wider range of relations about diseases, symptoms, causes, risk factors, drugs, side effects, and more. Moreover, while most prior work relies on manually curated sources as input, the KnowLife system taps into scientific literature as well as online communities. KnowLife uses advanced information extraction methods to populate the relations in the KB. This way, it learns patterns for relations, which are in turn used to semantically annotate newly seen documents, thus aiding users in “speed-reading”. We demonstrate the value of the KnowLife KB by various use-cases, supporting both layman and professional users.