CNN 보조 손실을 이용한 차원 기반 감성 분석 (Target-Aspect-Sentiment Joint Detection with CNN Auxiliary Loss for Aspect-Based Sentiment Analysis)
-
- 지능정보연구
- /
- 제27권4호
- /
- pp.1-22
- /
- 2021
텍스트를 바탕으로 한 차원 기반 감성 분석(Aspect-Based Sentiment Analysis)은 다양한 산업에서 유용성을 주목을 받고 있다. 기존의 차원 기반 감성 분석에서는 타깃(Target) 혹은 차원(Aspect)만을 고려하여 감성을 분석하는 연구가 대다수였다. 그러나 동일한 타깃 혹은 차원이더라도 감성이 나뉘는 경우, 또는 타깃이 없지만 감성은 존재하는 경우 분석 결과가 정확하지 않다는 한계가 존재한다. 이러한 문제를 해결하기 위한 방법으로 차원과 타깃을 모두 고려한 감성 분석(Target-Aspect-Sentiment Detection, 이하 TASD) 모델이 제안되었다. 그럼에도 불구하고, TASD 기존 모델의 경우 구(Phrase) 간의 관계인 지역적인 문맥을 잘 포착하지 못하고 초기 학습 속도가 느리다는 문제가 있었다. 본 연구는 TASD 분야 내 기존 모델의 한계를 보완하여 분석 성능을 높이고자 하였다. 이러한 연구 목적을 달성하기 위해 기존 모델에 합성곱(Convolution Neural Network) 계층을 더하여 차원-감성 분류 시 보조 손실(Auxiliary loss)을 추가로 사용하였다. 즉, 학습 시에는 합성곱 계층을 통해 지역적인 문맥을 좀 더 잘 포착하도록 하였으며, 학습 후에는 기존 방식대로 차원-감성 분석을 하도록 모델을 설계하였다. 본 모델의 성능을 평가하기 위해 공개 데이터 집합인 SemEval-2015, SemEval-2016을 사용하였으며, 기존 모델 대비 F1 점수가 최대 55% 증가했다. 특히 기존 모델보다 배치(Batch), 에폭(Epoch)이 적을 때 효과적으로 학습한다는 것을 확인할 수 있었다. 본 연구에서 제시된 모델로 더욱 더 세밀한 차원 기반 감성 분석이 가능하다는 점에서, 기업에서 상품 개발 및 마케팅 전략 수립 등에 다양하게 활용할 수 있으며 소비자의 효율적인 구매 의사결정을 도와줄 수 있을 것으로 보인다.
대한민국 중앙부처, 지방자치단체는 코로나 19가 급속도로 확산하는 팬데믹 상황에서 재난상황 극복을 위해 재난대응에 필요한 정보를 재난문자를 통해 제공하였다. 재난문자는 국민들이 가장 많이 접하는 재난정보 전달수단으로서, 휴대폰에 직접 방송하는 CBS(Cell Broadcast Service) 방식을 채택하고 있어 직접 찾아보는 수고스러움 없이 휴대폰을 통해 쉽게 정보를 접할 수 있다는 장점이 있다. 본 연구는 지난 1년 1개월간(2020년 1월~2021년 1월) 서울특별시에 발송된 재난문자의 특성을 다양한 텍스트마이닝 방법론 등을 통해 도출하고 재난문자에 포함된 다양한 유형의 정보가 국민들의 이동 행태에 어떠한 영향을 미쳤는지를 서울특별시 지역구의 연령별 유동인구의 이동성을 통해 확인하였다. 각 문자에 포함된 주요 단어와 포함된 정보를 분류하는 과정을 거치고 포함된 단어를 기반으로 하는 문서 군집 분석 기법을 적용해 개별 발송 문자를 분석 단위로써 활용할 수 있도록 텍스트 분석을 시행하였다. 이후, 텍스트마이닝을 통해 추출한 재난문자의 특성이 지역별, 연령별 인구이동성에 미친 영향을 규명하였다. 구조화된 모형을 활용하여 재난정보가 인구이동성에 미치는 영향을 기본효과, 누적효과로 구분하여 측정하였다. 지자체가 보유한 재난문자 발송권한으로 인해 재난문자 발송 특성은 지자체별로 상이함을 계량 분석에 활용하였다. 분석 결과 인구이동성에 변화를 유발하는 정보유형은 연령별로 상이함을 확인할 수 있었다. 날짜와 순서에 관련된 정보는 60-70대의 인구이동성을 유의미하게 감소시키는 것을 확인할 수 있었다. 온라인 정보는 20대의 이동성을 감소시켰고, 증상과 관련된 정보는 30대의 인구이동성을 감소시켰다. 한편, 방역 정책 준수를 당부하는 의미를 포함하는 규범적 단어 등은 전 연령의 인구이동성에 유의미한 변화를 불러일으키지 못함을 확인할 수 있었다. 이는 재난대응에 도움이 되는 유의미한 정보들만 재난문자에 포함되어야 함을 의미한다. 한편, 인구이동성에 유의미한 변화를 불러일으키는 정보유형 또한 재난문자가 반복됨에 따라 효과가 상쇄함을 음의 누적효과 추정 결과를 통해 확인할 수 있었다.
The objective of this research is to test the feasibility of developing a statewide truck traffic forecasting methodology for Wisconsin by using Origin-Destination surveys, traffic counts, classification counts, and other data that are routinely collected by the Wisconsin Department of Transportation (WisDOT). Development of a feasible model will permit estimation of future truck traffic for every major link in the network. This will provide the basis for improved estimation of future pavement deterioration. Pavement damage rises exponentially as axle weight increases, and trucks are responsible for most of the traffic-induced damage to pavement. Consequently, forecasts of truck traffic are critical to pavement management systems. The pavement Management Decision Supporting System (PMDSS) prepared by WisDOT in May 1990 combines pavement inventory and performance data with a knowledge base consisting of rules for evaluation, problem identification and rehabilitation recommendation. Without a r.easonable truck traffic forecasting methodology, PMDSS is not able to project pavement performance trends in order to make assessment and recommendations in the future years. However, none of WisDOT's existing forecasting methodologies has been designed specifically for predicting truck movements on a statewide highway network. For this research, the Origin-Destination survey data avaiiable from WisDOT, including two stateline areas, one county, and five cities, are analyzed and the zone-to'||'&'||'not;zone truck trip tables are developed. The resulting Origin-Destination Trip Length Frequency (00 TLF) distributions by trip type are applied to the Gravity Model (GM) for comparison with comparable TLFs from the GM. The gravity model is calibrated to obtain friction factor curves for the three trip types, Internal-Internal (I-I), Internal-External (I-E), and External-External (E-E). ~oth "macro-scale" calibration and "micro-scale" calibration are performed. The comparison of the statewide GM TLF with the 00 TLF for the macro-scale calibration does not provide suitable results because the available 00 survey data do not represent an unbiased sample of statewide truck trips. For the "micro-scale" calibration, "partial" GM trip tables that correspond to the 00 survey trip tables are extracted from the full statewide GM trip table. These "partial" GM trip tables are then merged and a partial GM TLF is created. The GM friction factor curves are adjusted until the partial GM TLF matches the 00 TLF. Three friction factor curves, one for each trip type, resulting from the micro-scale calibration produce a reasonable GM truck trip model. A key methodological issue for GM. calibration involves the use of multiple friction factor curves versus a single friction factor curve for each trip type in order to estimate truck trips with reasonable accuracy. A single friction factor curve for each of the three trip types was found to reproduce the 00 TLFs from the calibration data base. Given the very limited trip generation data available for this research, additional refinement of the gravity model using multiple mction factor curves for each trip type was not warranted. In the traditional urban transportation planning studies, the zonal trip productions and attractions and region-wide OD TLFs are available. However, for this research, the information available for the development .of the GM model is limited to Ground Counts (GC) and a limited set ofOD TLFs. The GM is calibrated using the limited OD data, but the OD data are not adequate to obtain good estimates of truck trip productions and attractions .. Consequently, zonal productions and attractions are estimated using zonal population as a first approximation. Then, Selected Link based (SELINK) analyses are used to adjust the productions and attractions and possibly recalibrate the GM. The SELINK adjustment process involves identifying the origins and destinations of all truck trips that are assigned to a specified "selected link" as the result of a standard traffic assignment. A link adjustment factor is computed as the ratio of the actual volume for the link (ground count) to the total assigned volume. This link adjustment factor is then applied to all of the origin and destination zones of the trips using that "selected link". Selected link based analyses are conducted by using both 16 selected links and 32 selected links. The result of SELINK analysis by u~ing 32 selected links provides the least %RMSE in the screenline volume analysis. In addition, the stability of the GM truck estimating model is preserved by using 32 selected links with three SELINK adjustments, that is, the GM remains calibrated despite substantial changes in the input productions and attractions. The coverage of zones provided by 32 selected links is satisfactory. Increasing the number of repetitions beyond four is not reasonable because the stability of GM model in reproducing the OD TLF reaches its limits. The total volume of truck traffic captured by 32 selected links is 107% of total trip productions. But more importantly, ~ELINK adjustment factors for all of the zones can be computed. Evaluation of the travel demand model resulting from the SELINK adjustments is conducted by using screenline volume analysis, functional class and route specific volume analysis, area specific volume analysis, production and attraction analysis, and Vehicle Miles of Travel (VMT) analysis. Screenline volume analysis by using four screenlines with 28 check points are used for evaluation of the adequacy of the overall model. The total trucks crossing the screenlines are compared to the ground count totals. L V/GC ratios of 0.958 by using 32 selected links and 1.001 by using 16 selected links are obtained. The %RM:SE for the four screenlines is inversely proportional to the average ground count totals by screenline .. The magnitude of %RM:SE for the four screenlines resulting from the fourth and last GM run by using 32 and 16 selected links is 22% and 31 % respectively. These results are similar to the overall %RMSE achieved for the 32 and 16 selected links themselves of 19% and 33% respectively. This implies that the SELINICanalysis results are reasonable for all sections of the state.Functional class and route specific volume analysis is possible by using the available 154 classification count check points. The truck traffic crossing the Interstate highways (ISH) with 37 check points, the US highways (USH) with 50 check points, and the State highways (STH) with 67 check points is compared to the actual ground count totals. The magnitude of the overall link volume to ground count ratio by route does not provide any specific pattern of over or underestimate. However, the %R11SE for the ISH shows the least value while that for the STH shows the largest value. This pattern is consistent with the screenline analysis and the overall relationship between %RMSE and ground count volume groups. Area specific volume analysis provides another broad statewide measure of the performance of the overall model. The truck traffic in the North area with 26 check points, the West area with 36 check points, the East area with 29 check points, and the South area with 64 check points are compared to the actual ground count totals. The four areas show similar results. No specific patterns in the L V/GC ratio by area are found. In addition, the %RMSE is computed for each of the four areas. The %RMSEs for the North, West, East, and South areas are 92%, 49%, 27%, and 35% respectively, whereas, the average ground counts are 481, 1383, 1532, and 3154 respectively. As for the screenline and volume range analyses, the %RMSE is inversely related to average link volume. 'The SELINK adjustments of productions and attractions resulted in a very substantial reduction in the total in-state zonal productions and attractions. The initial in-state zonal trip generation model can now be revised with a new trip production's trip rate (total adjusted productions/total population) and a new trip attraction's trip rate. Revised zonal production and attraction adjustment factors can then be developed that only reflect the impact of the SELINK adjustments that cause mcreases or , decreases from the revised zonal estimate of productions and attractions. Analysis of the revised production adjustment factors is conducted by plotting the factors on the state map. The east area of the state including the counties of Brown, Outagamie, Shawano, Wmnebago, Fond du Lac, Marathon shows comparatively large values of the revised adjustment factors. Overall, both small and large values of the revised adjustment factors are scattered around Wisconsin. This suggests that more independent variables beyond just 226; population are needed for the development of the heavy truck trip generation model. More independent variables including zonal employment data (office employees and manufacturing employees) by industry type, zonal private trucks 226; owned and zonal income data which are not available currently should be considered. A plot of frequency distribution of the in-state zones as a function of the revised production and attraction adjustment factors shows the overall " adjustment resulting from the SELINK analysis process. Overall, the revised SELINK adjustments show that the productions for many zones are reduced by, a factor of 0.5 to 0.8 while the productions for ~ relatively few zones are increased by factors from 1.1 to 4 with most of the factors in the 3.0 range. No obvious explanation for the frequency distribution could be found. The revised SELINK adjustments overall appear to be reasonable. The heavy truck VMT analysis is conducted by comparing the 1990 heavy truck VMT that is forecasted by the GM truck forecasting model, 2.975 billions, with the WisDOT computed data. This gives an estimate that is 18.3% less than the WisDOT computation of 3.642 billions of VMT. The WisDOT estimates are based on the sampling the link volumes for USH, 8TH, and CTH. This implies potential error in sampling the average link volume. The WisDOT estimate of heavy truck VMT cannot be tabulated by the three trip types, I-I, I-E ('||'&'||'pound;-I), and E-E. In contrast, the GM forecasting model shows that the proportion ofE-E VMT out of total VMT is 21.24%. In addition, tabulation of heavy truck VMT by route functional class shows that the proportion of truck traffic traversing the freeways and expressways is 76.5%. Only 14.1% of total freeway truck traffic is I-I trips, while 80% of total collector truck traffic is I-I trips. This implies that freeways are traversed mainly by I-E and E-E truck traffic while collectors are used mainly by I-I truck traffic. Other tabulations such as average heavy truck speed by trip type, average travel distance by trip type and the VMT distribution by trip type, route functional class and travel speed are useful information for highway planners to understand the characteristics of statewide heavy truck trip patternS. Heavy truck volumes for the target year 2010 are forecasted by using the GM truck forecasting model. Four scenarios are used. Fo~ better forecasting, ground count- based segment adjustment factors are developed and applied. ISH 90 '||'&'||' 94 and USH 41 are used as example routes. The forecasting results by using the ground count-based segment adjustment factors are satisfactory for long range planning purposes, but additional ground counts would be useful for USH 41. Sensitivity analysis provides estimates of the impacts of the alternative growth rates including information about changes in the trip types using key routes. The network'||'&'||'not;based GMcan easily model scenarios with different rates of growth in rural versus . . urban areas, small versus large cities, and in-state zones versus external stations. cities, and in-state zones versus external stations.
수도작에 있어서 수량성의 향상과 경영의 합리화률 기하기 위한 집단재배의 기술체계를 확립하고자 지역과 지대별로 임의선정된 18개소에서 집단재배와 개별재배를 한 수도에 대하여 경종 및 생육ㆍ수량구성요소의 변이를 조사연구하였으며 460개소의 집단재배와 개별재배에 대하여 경종ㆍ경영 및 조직운영등을 비교조사하고 집단재배와 개별재배의 경제성을 분석하여 그 결과를 요약하면 아래와 같다. 1. 집단재배와 개별재배의 경종 및 생육ㆍ수량형질에 관하여 연구한 결과는 다음과 같다. (1) 육묘에 있어서 집단재배가 비료의 3요소 특히 인산과 가리의 사용량이 많았고 약제살포의 횟수도 많았다. (2) 본답에 있어서의 경종기술은 집단재배가 개별재배보다
Personalized services directly and indirectly acquire personal data, in part, to provide customers with higher-value services that are specifically context-relevant (such as place and time). Information technologies continue to mature and develop, providing greatly improved performance. Sensory networks and intelligent software can now obtain context data, and that is the cornerstone for providing personalized, context-specific services. Yet, the danger of overflowing personal information is increasing because the data retrieved by the sensors usually contains privacy information. Various technical characteristics of context-aware applications have more troubling implications for information privacy. In parallel with increasing use of context for service personalization, information privacy concerns have also increased such as an unrestricted availability of context information. Those privacy concerns are consistently regarded as a critical issue facing context-aware personalized service success. The entire field of information privacy is growing as an important area of research, with many new definitions and terminologies, because of a need for a better understanding of information privacy concepts. Especially, it requires that the factors of information privacy should be revised according to the characteristics of new technologies. However, previous information privacy factors of context-aware applications have at least two shortcomings. First, there has been little overview of the technology characteristics of context-aware computing. Existing studies have only focused on a small subset of the technical characteristics of context-aware computing. Therefore, there has not been a mutually exclusive set of factors that uniquely and completely describe information privacy on context-aware applications. Second, user survey has been widely used to identify factors of information privacy in most studies despite the limitation of users' knowledge and experiences about context-aware computing technology. To date, since context-aware services have not been widely deployed on a commercial scale yet, only very few people have prior experiences with context-aware personalized services. It is difficult to build users' knowledge about context-aware technology even by increasing their understanding in various ways: scenarios, pictures, flash animation, etc. Nevertheless, conducting a survey, assuming that the participants have sufficient experience or understanding about the technologies shown in the survey, may not be absolutely valid. Moreover, some surveys are based solely on simplifying and hence unrealistic assumptions (e.g., they only consider location information as a context data). A better understanding of information privacy concern in context-aware personalized services is highly needed. Hence, the purpose of this paper is to identify a generic set of factors for elemental information privacy concern in context-aware personalized services and to develop a rank-order list of information privacy concern factors. We consider overall technology characteristics to establish a mutually exclusive set of factors. A Delphi survey, a rigorous data collection method, was deployed to obtain a reliable opinion from the experts and to produce a rank-order list. It, therefore, lends itself well to obtaining a set of universal factors of information privacy concern and its priority. An international panel of researchers and practitioners who have the expertise in privacy and context-aware system fields were involved in our research. Delphi rounds formatting will faithfully follow the procedure for the Delphi study proposed by Okoli and Pawlowski. This will involve three general rounds: (1) brainstorming for important factors; (2) narrowing down the original list to the most important ones; and (3) ranking the list of important factors. For this round only, experts were treated as individuals, not panels. Adapted from Okoli and Pawlowski, we outlined the process of administrating the study. We performed three rounds. In the first and second rounds of the Delphi questionnaire, we gathered a set of exclusive factors for information privacy concern in context-aware personalized services. The respondents were asked to provide at least five main factors for the most appropriate understanding of the information privacy concern in the first round. To do so, some of the main factors found in the literature were presented to the participants. The second round of the questionnaire discussed the main factor provided in the first round, fleshed out with relevant sub-factors. Respondents were then requested to evaluate each sub factor's suitability against the corresponding main factors to determine the final sub-factors from the candidate factors. The sub-factors were found from the literature survey. Final factors selected by over 50% of experts. In the third round, a list of factors with corresponding questions was provided, and the respondents were requested to assess the importance of each main factor and its corresponding sub factors. Finally, we calculated the mean rank of each item to make a final result. While analyzing the data, we focused on group consensus rather than individual insistence. To do so, a concordance analysis, which measures the consistency of the experts' responses over successive rounds of the Delphi, was adopted during the survey process. As a result, experts reported that context data collection and high identifiable level of identical data are the most important factor in the main factors and sub factors, respectively. Additional important sub-factors included diverse types of context data collected, tracking and recording functionalities, and embedded and disappeared sensor devices. The average score of each factor is very useful for future context-aware personalized service development in the view of the information privacy. The final factors have the following differences comparing to those proposed in other studies. First, the concern factors differ from existing studies, which are based on privacy issues that may occur during the lifecycle of acquired user information. However, our study helped to clarify these sometimes vague issues by determining which privacy concern issues are viable based on specific technical characteristics in context-aware personalized services. Since a context-aware service differs in its technical characteristics compared to other services, we selected specific characteristics that had a higher potential to increase user's privacy concerns. Secondly, this study considered privacy issues in terms of service delivery and display that were almost overlooked in existing studies by introducing IPOS as the factor division. Lastly, in each factor, it correlated the level of importance with professionals' opinions as to what extent users have privacy concerns. The reason that it did not select the traditional method questionnaire at that time is that context-aware personalized service considered the absolute lack in understanding and experience of users with new technology. For understanding users' privacy concerns, professionals in the Delphi questionnaire process selected context data collection, tracking and recording, and sensory network as the most important factors among technological characteristics of context-aware personalized services. In the creation of a context-aware personalized services, this study demonstrates the importance and relevance of determining an optimal methodology, and which technologies and in what sequence are needed, to acquire what types of users' context information. Most studies focus on which services and systems should be provided and developed by utilizing context information on the supposition, along with the development of context-aware technology. However, the results in this study show that, in terms of users' privacy, it is necessary to pay greater attention to the activities that acquire context information. To inspect the results in the evaluation of sub factor, additional studies would be necessary for approaches on reducing users' privacy concerns toward technological characteristics such as highly identifiable level of identical data, diverse types of context data collected, tracking and recording functionality, embedded and disappearing sensor devices. The factor ranked the next highest level of importance after input is a context-aware service delivery that is related to output. The results show that delivery and display showing services to users in a context-aware personalized services toward the anywhere-anytime-any device concept have been regarded as even more important than in previous computing environment. Considering the concern factors to develop context aware personalized services will help to increase service success rate and hopefully user acceptance for those services. Our future work will be to adopt these factors for qualifying context aware service development projects such as u-city development projects in terms of service quality and hence user acceptance.