Morality in dialogue systems has raised great attention in research recently. A moral dialogue system aligned with users' values could enhance conversation engagement and user connections. In this paper, we propose a framework, MoralDial to train and evaluate moral dialogue systems. In our framework, we first explore the communication mechanisms of morality and resolve expressed morality into three parts, which indicate the roadmap for building a moral dialogue system. Based on that, we design a simple yet effective method: constructing moral discussions between simulated specific users and the dialogue system. The constructed discussions consist of expressing, explaining, revising, and inferring moral views in dialogue exchanges, which makes conversational models learn morality well in a natural manner. Furthermore, we propose a novel evaluation method under the framework. We evaluate the multiple aspects of morality by judging the relation between dialogue responses and human values in discussions, where the multifaceted nature of morality is particularly considered. Automatic and manual experiments demonstrate that our framework is promising to train and evaluate moral dialogue systems.

<iframe src="https://app.sli.do/event/iHwnn5G9xYwcAmSbeKA4Au/embed/polls/aee1b835-b913-4804-b599-ba0027b87aba" width="300" height="400"></iframe>

MoralDial: A Framework to Train and Evaluate Moral Dialogue Systems via Moral Discussions

Linear methods for erasing human-interpretable concepts from neural representations were found tractable and useful.However, the impact of this removal on the behavior of classifiers trained on the modified representations is not fully understood.In this work, we formally define the notion of linear guardedness as the inability of an adversary to predict the concept directly from the representation and study its implications.We show that, in the binary case, an additional linear layer cannot recover the erased concept.However, we demonstrate that multiclass softmax classifiers can be constructed that indirectly recover the concept, pointing to the inherent limitations of linear guardedness as a bias mitigation technique.These findings shed light on the theoretical limitations of linear information removal methods and highlight the need for further research on the connections between intrinsic and extrinsic bias in neural models.

<iframe src="https://app.sli.do/event/eX4PwSmcCc1EvCqjxcYdKB/embed/polls/f44b96ca-a3d7-46ad-bea6-5de84fdfaf93" width="300" height="400"></iframe>

Linear Guardedness and its Implications

We present a study on the integration of time-sensitive information in lexicon-based offensive language detection systems. Our focus is on Offenseval sub-task A, aimed at detecting offensive tweets. We apply a semantic change detection algorithm over a short time span of two years to detect words whose semantics has changed and we focus particularly on those words that acquired or lost an offensive meaning between 2019 and 2020. Using the output of this semantic change detection approach, we train an SVM classifier on the Offenseval 2019 training set. We build on the already competitive SINAI system submitted to Offenseval 2019 by adding new lexical features, including those that capture the change in usage of words and their association with emerging offensive usages. We discuss the challenges, opportunities and limitations of integrating semantic change detection in offensive language detection models. Our work draws attention to an often neglected aspect of offensive language, namely that the meanings of words are constantly evolving and that NLP systems that account for this change can achieve good performance even when not trained on the most recent training data. 

Leveraging time-dependent lexical features for offensive language detection

The record of metamorphic conditions may be highly heterogeneous in spatially close rocks with different rheology and composition. The Cima di Gagnone area in the Central Alps represents such an example, where 1– to 100 s-meter scale ultrahigh–pressure and high–temperature ultramafic lenses are enveloped within amphibolite–facies metasediments. Structural investigation demonstrate that ultramafics and metapelites experienced a common Alpine deformation history, excluding their coupling within a tectonic mélange (Maino et al., 2021). In this contribution, we present new field observations, microstructural and petrological analyses, and thermodynamic modelling results focused on the metasediments, confirming that these rocks generally experienced medium pressure and medium temperature conditions of 1.0–1.2 GPa and 640–700 °C, followed by a retrograde stage around 0.6–0.8 GPa and 600–675 °C. However, a few samples from the immediate proximity of the ultramafic lenses record significantly higher P–T conditions of 1.3–1.7 GPa and 750–850 °C, approaching the high pressure and high temperature conditions of the ultramafic bodies (1.5–3.1 GPa, 650–850 °C). Mineral/bulk chemistry changes during growth of new mineral phases indicate to local melt/fluid interaction (i.e., metasomatism) between metasediments and ultramafics during the high temperature deformation. Preliminary U-Pb LA–ICP–MS dating suggests that zircon grains from the metasomatic reaction zone have been fully re–equilibrated during the early stage of Alpine exhumation (~36 Ma), while the large part of the metasediments records only pre–Alpine ages. Our results documented that, even though metasediments share the same structural evolution with UM, large differences in the local metamorphic conditions (Δ P up to 2 GPa; ΔT up to 160 °C) are recorded and are not randomly distributed but they are in relation to: 1) deformation (micro)structures, and 2) distance from the UM lenses. Moreover, fluid–assisted metasomatism is furthermore documented as being strongly localized at the interface between ultramafic lenses and the metapelitic host throughout all part of the metamorphic evolution, including the HP–HT stage. Therefore, in the Cima di Gagnone type–locality, the interplay between metapelites and ultramafic exerts a crucial first–order control to allow assemblage equilibrium during HT metamorphism and amphibolite–facies retrogression. Furthermore, the new findings exclude that the different metamorphic record may be attributed only to differential preservation during the retrograde path. We finally discuss these new data into the regional P–T–t–D paths and discuss the possible processes that controlled the exhumation of HP rocks.

Local variations of metamorphic record from compositionally heterogeneous rocks: Inferences on exhumation processes of HP rocks from the Adula-Cima Lunga unit

Volcanic ash is the result of extensive magma fragmentation during eruptions. It depends upon a combination of magma properties (rheology, vesicularity, permeability, gas overpressure) and the possible involvement of external fluids during magma ascent. However, fragmentation processes during the ash-dominated, low-to-mid intensity eruptions are still a matter of debate. Combination of morpho-textural data of the erupted ash with geophysical data on the activity can inform about the energy of the activity as well as on the mechanisms of ash production and dispersal. The phase of heavy ash emission occurred during March 2016 at Copahue volcano (Argentina) generated a very low infrasonic amplitude, corresponding to low exit velocity and low magma overpressure. The apparent unbalance between measured geophysical parameters and the intensity of the activity raises a number of questions concerning the links among acoustic pressure, gas overpressure and efficiency of magma fragmentation. Generally, magma-water interaction is supposed to be the main process of magma fragmentation at Copahue. Conversely, our data on ash texture indicate that a process of magma volatile exsolution under low overpressures actively controlled the fragmentation during the investigated eruptive stage. Ash componentry also revealed that a large fraction of deposit consisted of recycled material. Therefore, comparing these results with geophysical information, we interpreted the activity as influenced by a process of energy buffering, operated by the passage of the erupting mixture through a thick layer of granular, ash-bearing crater infilling produced by the ash recycling processes. The importance of this abundant hot, vent-hosted material for the convective stability of the eruptive plume and for the ash dispersal was highlighted, with direct consequences for the assessment of the related hazard and management of eruptive crises,

Mechanisms of ash production and recycling during low-energy, mid-intensity eruptions at Copahue volcano (Argentina)

We investigate the dynamics of regolith on rubble-pile asteroids to explain granular processes observed in reality. In particular, we explain how the appearance of boulders on the surface of asteroid Itokawa could have resulted from a size sorting process in granular media called the Brazil Nut Effect (BNE). The Discrete Element Method (DEM) is implemented to perform numerical simulations of the BNE in a micro-gravity environment caused by inter-particle collisions during seismic vibrations. Firstly, we present the results of how the BNE depends on the magnitude of surface gravity. It is estimated that segregation processes on Itokawa occur over much longer time-scales (in the order of a few hundred years) than the same processes would require in the presence of a strong gravitational field, like on Earth. Secondly, we also find that the size sorting could also result from kinetic sieving encountered during granular avalanches. Finally, we discuss how the void-filling mechanism becomes more efficient when there is a higher relative size difference between the boulders and the surrounding grains. Our model has important implications in understanding the resurfacing of Itokawa by trying to explain one of the many complex geophysical processes that occur in such unique conditions.

Granular segregation on the rubble-pile asteroid Itokawa

Spin rates of minor planets or asteroids are known to have been affected by several agents including but not limited to tidal fly-bys, impacts, and solar radiation. Surface processes like landslides occur as a result of such rotational changes. We study the evolution of landslides on top-shaped rubble pile asteroids like 101955 Bennu and 162173 Ryugu, with the underlying core modeled as two solid cones fused back to back. Using a depth-averaged avalanche theory applicable to granular flows we solve for axisymmetric landslides occurring at various spin rates and regolith friction. Static regions on the surface corresponding to different spin rates are identified from an equilibrium analysis. We then solve for landslides initiated at different latitudes. It is found that landslides equilibrate at lower latitudes as the spin rate is increased. Beyond a critical spin rate regolith is shed from the equator. This critical spin is higher for a lower value of the semi-apex angle of the cone.

Axisymmetric landslides on top-shaped asteroids

Through numerical simulations, we investigate impact generated seismic wave transmission in gran-ular media under extremely low pressure. This mimics the conditions in the interior of asteroids and other smallplanetary bodies. We find a dependency not only on the overburden pressure on the medium, but also on thevelocity of the impact that generates the wave. This is, at extremely low values of overburden pressure, the wavespeed depends no only on the imposed pressure, but also on the increment in pressure created by the passingof the wave. We study crystalline and random packings and find very similar behaviour though with differentwave speeds as expected. We then relate our results to different mission-related events on asteroids.

Seismic waves in the asteroid environment

Constitutive modeling of time-dependent flows in dense frictional suspensions

Abstract Body: RHEED spectra taking during molecular beam epitaxy of GaN and related alloys on sapphire substrates were analyzed using a convolutional neural network. An existing image-classifying convolution neural network (CNN)— AlexNet— was employed for the analysis. The RHEED images during various stages of growth were manually categorized to train the CNN. The CNN only works on images, so video frames of RHEED during growth had to first be extracted from the video. These images were then separated into a training set and a validation set for use in training the network and in validating it. Mobility, deposition temperature, X-ray diffraction, atomic force microscopy, and Hall measurements were taken on the resulting films. After training on the manually categorized images and videos, the CNN achieved a proper categorization rate of 97% on the validation data and employing mobility and various growth stages as categorical outputs. Other machine learning techniques were applied as well to analyze RHEED spectra, such as principal components analysis and k-means clustering. Initial results from growth of ternary (AlxIn1-xN) and quaternary (AlGaInN) alloys will also be presented at the conference.

Machine Learning for Evaluation of RHEED Spectra on III-Nitride Films Grown Using Molecular Beam Epitaxy

Abstract Body: A parametric analysis of growth conditions on the structural properties of N-Polar InGaN films was studied. Three series of N-polar InGaN films were grown on bulk GaN substrates by plasma-assisted molecular beam epitaxy (PAMBE) to study the effect of substrate temperature, Ga, and In fluxes. For all series, the surface morphology was measured by atomic force microscopy (AFM), and the composition and strain were determined by high resolution X-ray diffraction (HRXRD) and XRD reciprocal space mapping (RSM). All the samples in this study were 350 nm thick. First, a series of samples was grown at temperatures varying from 620 oC to 690 oC. The Ga flux for all samples in this series was kept at 3.8E-7 Torr and the In flux was increased by temperature to ensure metal-rich growth conditions. The surface morphology improved significantly by increasing the temperature, and the indium content was reduced as expected. The samples grown at 670 oC and 690 oC showed surprisingly smooth surface morphologies with surface roughness <0.7 nm which has never been reported before on such thick InGaN films. An In content of 16.5% and 13.6% was measured on samples grown at 670 oC and 690 oC, respectively. These temperatures are significantly higher than growth temperatures (<600 oC) typically used for the growth of InGaN films on Ga-polar GaN. In the next step, samples were grown with In fluxes of 3E-7 Torr and 4E-7 Torr. The Ga flux was kept at 3.8E-7 Torr and both samples were grown at a substrate temperature of 670 oC. Both samples showed step flow growth and low surface roughness despite the large difference in In flux. Nonetheless, increasing the In flux from 3E-7 Torr to 4E-7 Torr resulted in an improved surface morphology, and increased In composition and strain relaxation. These results indicate that the In flux can be tuned within an ample window to obtain specific sample properties. For the last series, three samples were grown at 660 oC and with In flux of 4E-7 Torr, while the Ga flux was varied from 3.2E-7 Torr to 3.6E-7 Torr. The surface morphology degraded with decreasing Ga flux. The Indium composition simultaneously increased due to the limited availability of Ga during growth. This indicates that the strain caused by increased In incorporation is relaxed by generating dislocations which in turn results in further increase in In incorporation as relaxation proceeds (the In pulling effect).

Growth of High Quality 350nm-Thick InGaN Films on N-Polar GaN Substrates by Plasma-Assisted Molecular Beam Epitaxy

We present results clearly demonstrating passive gradient-driven droplet motion on coating-free all-metal Al-surfaces, produced via laser-etching method.

Underline digital video library

MoralDial: A Framework to Train and Evaluate Moral Dialogue Systems via Moral Discussions

Linear Guardedness and its Implications

Leveraging time-dependent lexical features for offensive language detection

Local variations of metamorphic record from compositionally heterogeneous rocks: Inferences on exhumation processes of HP rocks from the Adula-Cima Lunga unit

Mechanisms of ash production and recycling during low-energy, mid-intensity eruptions at Copahue volcano (Argentina)

Granular segregation on the rubble-pile asteroid Itokawa

Axisymmetric landslides on top-shaped asteroids

Seismic waves in the asteroid environment

Constitutive modeling of time-dependent flows in dense frictional suspensions

Machine Learning for Evaluation of RHEED Spectra on III-Nitride Films Grown Using Molecular Beam Epitaxy

Growth of High Quality 350nm-Thick InGaN Films on N-Polar GaN Substrates by Plasma-Assisted Molecular Beam Epitaxy

Development of Coating-free Super Water-repellent Micropatterned Aluminium for Spontaneous Droplet Motion

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES