Current image captioning methods are usually trained via (penalized) maximum likelihood estimation. However, the log-likelihood score of a caption does not correlate well with human assessments of quality. Standard syntactic evaluation metrics, such as BLEU, METEOR and ROUGE, are also not well correlated. The newer SPICE and CIDEr metrics are better correlated, but have traditionally been hard to optimize for. In this paper, we show how to use a policy gradient (PG) method to directly optimize a linear combination of SPICE and CIDEr (a combination we call SPIDEr): the SPICE score ensures our captions are semantically faithful to the image, while CIDEr score ensures our captions are syntactically fluent. The PG method we propose improves on the prior MIXER approach, by using Monte Carlo rollouts instead of mixing MLE training with PG. We show empirically that our algorithm leads to easier optimization and improved results compared to MIXER. Finally, we show that using our PG method we can optimize any of the metrics, including the proposed SPIDEr metric which results in image captions that are strongly preferred by human raters compared to captions generated by the same model but trained to optimize MLE or the COCO metrics.

No abstract

BackgroundChina’s rapid urbanization over the past decades has exacerbated the problems of environmental degradation and health disparities. However, few studies have analysed the differences between urban and rural residents in relation to how environmental quality impacts health outcomes. This study examines the associations between Chinese people's perceptions of environmental quality and their self-rated health, particularly focusing on differences between rural and urban residents in environment-health relationships.MethodsUsing a logistic regression model and data from the 2013 Chinese General Social Survey (CGSS), a representative sample of data for 3,402 urban residents (46 ± 16 years) and 2,439 rural residents (48 ± 15 years) was analysed. The dependent variable used for the logistic regressions was whether or not respondents reported being healthy. Independent variables included respondents’ evaluations of the living environment, and how frequently they participated in physical activities. Interaction terms were employed to measure the moderating effects of physical exercise on the relationship between perceived environmental quality and health.ResultsThe percentage of healthy urban residents was significantly larger than that of healthy rural respondents (70.87% versus 62.87%). Urban respondents living in areas with sufficient green space were more likely to report good health (OR = 0.749, CI = [0.628, 0.895]), while rural respondents without reliable access to fresh water were more likely to report poor health (OR = 0.762, CI = [0.612, 0.949]). Urban respondents who were exposed to green spaces and exercised frequently were 21.6 per cent more likely to report good health than those who exercised infrequently (OR = 1.216, CI = [1.047, 1.413]). Those who lived in areas with insufficient green space and exercised frequently were 19.1 per cent less likely to report good health than those who exercised infrequently (OR = 0.805, CI = [0.469, 1.381]). No evidence suggested that physical exercise exerted a moderating effect on the relationship between exposure to air pollution and health.ConclusionsOn average, urban residents have better health than rural residents. Among four indicators for low environmental quality (air pollution, lack of green spaces, water pollution, uncertain access to freshwater resources), green space was an important determinant of urban residents’ health status, while unreliable access to fresh water harmed rural residents’ health. Physical exercise moderated the effects of exposure to green spaces on urban residents’ health.

Experimental and theoretical studies show that there are electromagnetic (EM) fields generated by seismic waves with two kinds of conversion mechanisms in a fluid‐saturated, porous medium. Within a homogeneous formation, the seismic wave generates a seismoelectric field that exists only in the area disturbed by the seismic wave and whose apparent velocity is that of the seismic wave. At an interface between differing formation properties, the generated seismoelectric wave is a propagating EM wave that can be detected everywhere. An electrode, used as a receiver on the ground surface, can detect the propagating EM wave generated at an interface, but cannot detect the seismoelectric field generated in a homogeneous formation. When the electrode is in a borehole and close to a porous formation, it can detect both the EM waves and the seismoelectric field. In this paper, electrokinetic measurements are performed with borehole models made of natural rocks or artificial materials. Experimental results show that the Stoneley wave and other acoustic modes, excited by a monopole source in the borehole models, generate seismoelectric fields in fluid‐saturated formations. The electric components of the seismoelectric fields can be detected by an electrode in the borehole or on the borehole wall. The amplitude and frequency of the seismoelectric fields are related not only to the seismic wave, but also to formation properties such as permeability, conductivity, etc. Comparison between the waveforms of the seismoelectric signals and acoustic logging waves suggests that seismoelectric well logging may explore the different properties of the formation. Electroseismic measurements are also performed with these borehole models. The electric pulse through the electrode in the borehole or on the borehole wall induces Stoneley waves in fluid‐saturated models that can be received by a monopole transducer in the same borehole. These measurement methods (seismoelectric logging or electroseismic logging) might directly apply to well logging to investigate formation properties related to the pore fluid flow.

Abstract. When seismic waves generate a relative fluid-solid motion in a fluidsaturated porous medium, the moving charges (streaming current) in the electric double layer induce an electroma. gnetic (EM) field. This paper first experimentally confirms that the coupling between the seismic wave and the electromagnetic field in the kilohertz range is electrokinetic in nature. Seismoelectric signals are measured in homogeneous cylindrical porous rock samples and multilayered models. The seismoelectric signals in homogeneous rock are electric fields that move along with the acoustic wave. The mechanism of the seismoelectric conversion is completely different from the piezoelectric effect of quartz grains. The seismoelectric sensitivity with respect to salinity of the saturant has been experimentally determined. The amplitude of seismoelectric signals increases as the saturant conductivity decreases. The seismoelectric effects are generated by two different mechanisms. Both the EM radiation and the electric potential generated at an interface and within a porous medium, respectively, were measured as the P wave, at ultrasonic frequencies, passes through the layered models. Our experimental results demonstrate that seismoelectric effects exist and are measurable in the kilohertz range. The paper concludes with a comparison of experimental data and modeled data in a threelayer porous model. Seismoelectric measurements could be an effective means of obtaining transport coefficients such as hydraulic permeability and other porous rock properties.