We investigate different ways of learning structured perceptron models for coreference resolution when using non-local features and beam search. Our experimental results indicate that standard techniques such as early updates or Learning as Search Optimization (LaSO) perform worse than a greedy baseline that only uses local features. By modifying LaSO to delay updates until the end of each instance we obtain significant improvements over the baseline. Our model obtains the best results to date on recent shared task data for Arabic, Chinese, and English.

Abstract. Remote sensing of the gaseous composition of non-eruptive, passively degassing volcanic plumes can be a tool to gain insight into volcano interior processes. Here, we report on a field study in September 2015 that demonstrates the feasibility of remotely measuring the volcanic enhancements of carbon dioxide (CO 2 ), hydrogen fluoride (HF), hydrogen chloride (HCl), sulfur dioxide (SO 2 ), and bromine monoxide (BrO) in the downwind plume of Mt. Etna using portable and rugged spectroscopic instrumentation. To this end, we operated the Fourier transform spectrometer EM27/SUN for the shortwave-infrared (SWIR) spectral range together with a co-mounted UV spectrometer on a mobile platform in direct-sun view at 5 to 10 km distance from the summit craters. The 3 days reported here cover several plume traverses and a sunrise measurement. For all days, intra-plume HF, HCl, SO 2 , and BrO vertical column densities (VCDs) were reliably measured exceeding 5 ×1016 , 2 ×10 17 , 5 ×10 17 , and 1 ×10 14 molec cm −2 , with an estimated precision of 2.2 ×10 15 , 1.3 ×10 16 , 3.6 ×10 16 , and 1.3 ×10 13 molec cm −2 , respectively. Given that CO 2 , unlike the other measured gases, has a large and wellmixed atmospheric background, derivation of volcanic CO 2 VCD enhancements ( CO 2 ) required compensating for changes in altitude of the observing platform and for background concentration variability. The first challenge was met by simultaneously measuring the overhead oxygen (O 2 ) columns and assuming covariation of O 2 and CO 2 with altitude. The atmospheric CO 2 background was found by identifying background soundings via the coemitted volcanic gases. The inferred CO 2 occasionally exceeded 2 × 10 19 molec cm −2 with an estimated precision of 3.7 × 10 18 molec cm −2 given typical atmospheric background VCDs of 7 to 8 × 10 21 molec cm −2 . While the correlations of CO 2 with the other measured volcanic gases confirm the detection of volcanic CO 2 enhancements, correlations were found of variable significance (R 2 ranging between 0.88 and 0.00). The intra-plume VCD ratios CO 2 / SO 2 , SO 2 / HF, SO 2 / HCl, and SO 2 / BrO were in the range 7.1 to 35.4, 5.02 to 21.2, 1.54 to 3.43, and 2.9 × 10 3 to 12.5 × 10 3 , respectively, showing pronounced day-to-day and intra-day variability.

Abstract. The second Cabauw Intercomparison of Nitrogen Dioxide measuring Instruments (CINDI-2) took place in Cabauw (the Netherlands) in September 2016 with the aim of assessing the consistency of multi-axis differential optical absorption spectroscopy (MAX-DOAS) measurements of tropospheric species (NO2, HCHO, O3, HONO, CHOCHO and O4). This was achieved through the coordinated operation of 36 spectrometers operated by 24 groups from all over the world, together with a wide range of supporting reference observations (in situ analysers, balloon sondes, lidars, long-path DOAS, direct-sun DOAS, Sun photometer and meteorological instruments). In the presented study, the retrieved CINDI-2 MAX-DOAS trace gas (NO2, HCHO) and aerosol vertical profiles of 15 participating groups using different inversion algorithms are compared and validated against the colocated supporting observations, with the focus on aerosol optical thicknesses (AOTs), trace gas vertical column densities (VCDs) and trace gas surface concentrations. The algorithms are based on three different techniques: six use the optimal estimation method, two use a parameterized approach and one algorithm relies on simplified radiative transport assumptions and analytical calculations. To assess the agreement among the inversion algorithms independent of inconsistencies in the trace gas slant column density acquisition, participants applied their inversion to a common set of slant columns. Further, important settings like the retrieval grid, profiles of O3, temperature and pressure as well as aerosol optical properties and a priori assumptions (for optimal estimation algorithms) have been prescribed to reduce possible sources of discrepancies. The profiling results were found to be in good qualitative agreement: most participants obtained the same features in the retrieved vertical trace gas and aerosol distributions; however, these are sometimes at different altitudes and of different magnitudes. Under clear-sky conditions, the root-mean-square differences (RMSDs) among the results of individual participants are in the range of 0.01–0.1 for AOTs, (1.5–15) ×1014molec.cm-2 for trace gas (NO2, HCHO) VCDs and (0.3–8)×1010molec.cm-3 for trace gas surface concentrations. These values compare to approximate average optical thicknesses of 0.3, trace gas vertical columns of 90×1014molec.cm-2 and trace gas surface concentrations of 11×1010molec.cm-3 observed over the campaign period. The discrepancies originate from differences in the applied techniques, the exact implementation of the algorithms and the user-defined settings that were not prescribed. For the comparison against supporting observations, the RMSDs increase to a range of 0.02–0.2 against AOTs from the Sun photometer, (11–55)×1014molec.cm-2 against trace gas VCDs from direct-sun DOAS observations and (0.8–9)×1010molec.cm-3 against surface concentrations from the long-path DOAS instrument. This increase in RMSDs is most likely caused by uncertainties in the supporting data, spatiotemporal mismatch among the observations and simplified assumptions particularly on aerosol optical properties made for the MAX-DOAS retrieval. As a side investigation, the comparison was repeated with the participants retrieving profiles from their own differential slant column densities (dSCDs) acquired during the campaign. In this case, the consistency among the participants degrades by about 30 % for AOTs, by 180 % (40 %) for HCHO (NO2) VCDs and by 90 % (20 %) for HCHO (NO2) surface concentrations. In former publications and also during this comparison study, it was found that MAX-DOAS vertically integrated aerosol extinction coefficient profiles systematically underestimate the AOT observed by the Sun photometer. For the first time, it is quantitatively shown that for optimal estimation algorithms this can be largely explained and compensated by considering biases arising from the reduced sensitivity of MAX-DOAS observations to higher altitudes and associated a priori assumptions.