Jana Götze

Walk This Way: Spatial Grounding for City Exploration

Recently there has been an interest in spatially aware systems for pedestrian routing and city exploration, due to the proliferation of smartphones with GPS receivers among the general public. Since GPS readings are noisy, giving good and well-timed route instructions to pedestrians is a challenging problem. This paper describes a spoken-dialogue prototype for pedestrian navigation in Stockholm that addresses this problem by using various grounding strategies.

Learning landmark salience models from users’ route instructions

Route instructions for pedestrians are usually better understood if they include references to landmarks, and moreover, these landmarks should be as salient as possible. In this paper, we present an approach for automatically deriving a mathematical model of salience directly from route instructions given by humans. Each possible landmark that a person can refer to in a given situation is modelled as a feature vector, and the salience associated with each landmark can be computed as a weighted sum of these features. We use a ranking SVM method to derive the weights from route instructions given by humans as they are walking the route. The weight vector, representing the person’s personal salience model, determines which landmark(s) are most appropriate to refer to in new situations.

User simulation for the evaluation of bus information systems

In this paper, we describe our contribution to the Spoken Dialog Challenge. We set up a user simulation using the large Lets Go corpus as resource to build our models. Automatic calls were made to all four dialog systems in the SDC, bus information systems that cover the schedule of Pittsburgh, PA. We discuss in detail the architecture and required setup for our system-independent user simulation and report the results and challenges we faced. We also report initial evaluation results.

“Turn Left” Versus “Walk Towards the Café”: When Relative Directions Work Better Than Landmarks

An automatic mechanism that gives verbal navigation instructions to pedestrians in situ needs to take into account a number of factors. Besides giving the instruction at the right time and place, the information needs to be as unambiguous as possible for the user to both choose the correct path and be confident in doing so. Humans make extensive use of landmarks when describing the way to others and are more successful following instructions that include landmarks. We present a study comparing landmark-based instructions with relative direction instructions on pedestrians in a real city environment, measuring both objective and subjective success. We find that at some decision points, relative direction instructions work better. We present a method that uses openly available geographic data to predict which kind of instruction is preferable at a given decision point.

Reference Resolution for Pedestrian Wayfinding Systems

References to objects in our physical environment are common especially in language about wayfinding. Advanced wayfinding systems that interact with the pedestrian by means of (spoken) natural language therefore need to be able to resolve references given by pedestrians (i.e. understand what entity the pedestrian is referring to). The contribution of this paper is a probabilistic approach to reference resolution in a large-scale, real city environment, where the context changes constantly as the pedestrians are moving. The geographic situation, including information about objects’ location and type, is represented using OpenStreetMap data.

Approximating phonotactic input in children’s linguistic environments from orthographic transcripts

This paper presents a new speaker change detection system based on Long Short-Term Memory (LSTM) neural networks using acoustic data and linguistic content. Language modelling is combined with two different Joint Factor Analysis (JFA) acoustic approaches: i-vectors and speaker factors. Both of them are compared with a baseline algorithm that uses cosine distance to detect speaker turn changes. LSTM neural networks with both linguistic and acoustic features have been able to produce a robust speaker segmentation. The experimental results show that our proposal clearly outperforms the baseline system.Child-directed spoken data is the ideal source of support for claims about children’s linguistic environments. However, phonological transcriptions of child-directed speech are scarce,compared to s ...