Human generates responses relying on semantic and functional dependencies, including coreference relation, among dialogue elements and their context. In this paper, we investigate matching a response with its multi-turn context using dependency information based entirely on attention. Our solution is inspired by the recently proposed Transformer in machine translation (Vaswani et al., 2017) and we extend the attention mechanism in two ways. First, we construct representations of text segments at different granularities solely with stacked self-attention. Second, we try to extract the truly matched segment pairs with attention across the context and response. We jointly introduce those two kinds of attention in one uniform neural network. Experiments on two large-scale multi-turn response selection tasks show that our proposed model significantly outperforms the state-of-the-art models. * Equally contributed. † Work done as a visiting scholar at Baidu.
In this paper, we study the task of response selection for multi-turn human-computer conversation. Previous approaches take word as a unit and view context and response as sequences of words. This kind of approaches do not explicitly take each utterance as a unit, therefore it is difficult to catch utterancelevel discourse information and dependencies. In this paper, we propose a multi-view response selection model that integrates information from two different views, i.e., word sequence view and utterance sequence view. We jointly model the two views via deep neural networks. Experimental results on a public corpus for context-sensitive response selection demonstrate the effectiveness of the proposed multi-view model, which significantly outperforms other single-view baselines.
Current fire models are designed to model the spread of a linear fire front in dead, small-diameter fuels. Fires in predominantly living vegetation account for a large proportion of annual burned area in the United States. Prescribed burning is used to manage living fuels; however, prescribed burning is currently conducted under conditions that result in marginal burning. We do not understand quantitatively the relative importance of the fuel and environmental variables that determine spread in live vegetation. To address these weaknesses, laboratory fires have been burned to determine the effects of wind, slope, moisture content and fuel characteristics on fire spread in fuel beds of common chaparral species. Four species (Adenostoma fasciculatum, Ceanothus crassifolius, Quercus berberidifolia, Arctostaphylos parryana), two wind velocities (0 and 2 m s −1 ) and two fuel bed depths (20 and 40 cm) were used. Oven-dry moisture content of fine fuels (<0.63 cm diameter) ranged from 0.09 to 1.06. Seventy of 125 fires successfully propagated the length (2.0 m) of the elevated fuel bed. A logistic model to predict the probability of successful fire spread was developed using stepwise logistic regression. The variables selected to predict propagation were wind velocity, slope percent, moisture content, fuel loading, species and air temperature. Air temperature and species terms were removed from the model for parsimony. The final model correctly classified 94% of the observations. Comparison of results with an empirical decision matrix for prescribed burning in chaparral suggested some agreement between the laboratory data and the empirical tool.
Commercial herbicide atrazine (AT) was first loaded into the lumen of halloysite nanotubes (HNTs) in the amount of 9 wt %, and then the AT-loaded HNTs (HNTs-AT) were further incorporated into poly(vinyl alcohol)/starch composites (PVA/ST, with the weight ratio of 80/20) to construct a dual drug delivery system. AT loaded in nanotubes displayed much slower release from PVA/ST film in water than free AT; for example, the total release amount of AT from PVA/ST film with loaded AT was only 61% after 96 h, while this value reached 97% in PVA/ST film with free AT. The release behavior of AT from PVA/ST film with HNTs-AT was first dominated by the mechanism of matrix erosion and then by the mechanism of Fickian diffusion. In addition, combining HNTs and PVA/ST blends together in the controlled release of herbicide also reduced its leaching through the soil layer, which would be useful for diminishing the environmental pollution caused by pesticide.
Fire suppression tests with ceiling sprinkler protection in a rack storage fuel configuration are simulated using a Computational Fluid Dynamics tool. The fuel is arranged in a double-row, six pallet-load wide and three-tier high (2×6×3) rack storage array. Each pallet load consists of three nested double-wall corrugated cardboard boxes surrounding a metal liner. Two types of ceiling sprinklers are used in this study: a pendent quick response sprinkler designated as K14, and an upright standard response sprinkler designated as K11.2. The tests are simulated using FireFOAM, which couples necessary sub-models for fire growth, sprinkler response, and fire suppression. Numerical results are compared with experiments for both free burn tests under a 20-MW calorimeter and sprinkler suppression tests under a 7.6 m high ceiling. For the free burn case, the model results show good quantitative agreement of heat release rates in all three phases, from ignition to fire growth and steady burning. For the suppression cases, the model reproduces the suppression effectiveness of the two sprinkler protection designs: K14 sprinklers suppress the fire rapidly with only one sprinkler activation, while with K11.2 sprinklers, both in the tests and simulation, the fire spreads to the pallets on the end of the fuel array with multiple sprinkler activations. The modeled sprinkler activation times are within the estimated experimental uncertainty following three repeat tests. Quantitative results characterizing sprinkler suppression performance obtained from the simulations, such as the actual delivered density (ADD) and water evaporation rate, are also reported.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.