Mapping of Geographical Entity with Meeting Location from Text for Mobile Kyoungryol Kim
Table of Contents 1.Introduction 2.Background and Related Work 3.The Proposed System 4.Experimentation 5.Conclusion 2
1. Introduction 1) Motivation 2) Problem Definition 3) Contribution
Motivation : IE Techniques on Smartphone 4 Apple iPhone Google Android RIM Blackberry MS Windows Phone Time(Text) Recognition Phone No. Recognition Location(Text) Recognition Adding event by recognized time May 21, 2011 Address Recognition (Captured from Apple iPhone) People start to pay attention to ‘Location Extraction’ technique
Motivation : Characteristics of Mobile Device Memory Issue Android : 16MB heap size limit for each app. iPhone : No memory limit, but totally 512MB of RAM (iPhone4) Speed Issue People who use mobile devices usually feel uncomfortable when it delays. IE System Usually general Information Extraction system consists of many NLP modules which consume more than 1GB memory, at least. Client-Server model Client and server communicating model that every processing is done in server -side. Need internet connection (3G or Wifi). If many clients request to the server at once, there will be overloading delays or the server dies. 5 IE Method Specialized on Mobile Device is Needed
Goal of this Research Mapping Meeting Location text to the Geographical Location and update it to online calendar in mobile device 6 The team meeting for the evaluation of first half of Univcast will be held. Date : July 19 (Sat) PM 2 Location : Myeong-dong Dandelion Territory Directions to Dandelion Territory At Myeong-dong station gate number 8, take a walk following the downtown then there it is on the first floor of YMCA building. Meeting Location NameMyeong-dong Dandelion Territory Address 1-1, Myeong-dong 1-ga, Jung-gu, Seoul, Korea Geocode( , ) Meeting Announcement Extract Meeting Location Update Calendar startTime T14:00 Extract Time
Problem Definition 1. Extract meeting location from meeting announcement 2. Disambiguate the extracted meeting location 7 회의는 오후 5 시 학생회관 101 호에서 열립니다. (Meeting will be held 5 PM at Room 101, Student Union.)
Contribution 8
2. Background and Related Works 1) Information Extraction 2) Geocoding 3) Linked Open Data 4) Local-Grammar Graph
Information Extraction Information Extraction The objective is to construct structured database from free text or semi-stru ctured text (J. H. Kim 2004) Related Work CMU Seminar Announcement Corpus 485 semi-structured seminar announcements Types : stime, etime, location, speaker Focus only on 4 types of information extraction, not on Geocoding. 10 Examples of seminar announcement
Geocoding Geocoding The process of finding associated geographic coordinates, often expressed as l atitude and longitude, from other geographic data such as street addresses or zip codes (Geocoding, Wikipedia) Related Work Geocode from the address (Manov 2003; Jones 2003; Peng 2006; Pouliquen 2006; Volz 2007; Overell 20 07; Goldberg 2007; Kauppinen 2008) The big issue of the research is disambiguation of address (Pouliquen et al. 2006) 1.Multi-referent ambiguity two different geographic locations share the same name, e.g. "Cambridge" is it Cambridge, UK or Cambridge, Massachusetts? 2.Name variant ambiguity the same location has different names, 3.Geoname-Non Geoname ambiguity where a location name could also stand for some other word such as a person name or nouns, e.g. Metro as the city in Indonesia vs. Metro as the subway system Focus only on Geocoding address, not all location entity e.g. "Room 101, Student Union, Hanyang University" 11
Linked Open Data Linked Open Data URL : The project aims to identify data sets that are available under open licenses, re-publi sh these in RDF on the Web and interlink them with each other Geographic Datasets are growing rapidly For only few Korean Geographical data included in LOD, we regard set of open geo graphical data as Linked Data, in this research. 12 March 2009September 2010September 2011
Local-Grammar Graph Local-Grammar Graph The language description model which is to perform automatic analysis an d generation of natural language text, information extraction, using local la nguage information in the form of Finite-State Automata. (J. Nam 2006) Help to increase efficiency and accuracy by lexicalizing the knowledge forming grammar readability by consisting grammar as Directed Acyclic Graphs. Various omission and permutation can be described which cannot be done by rules or specific features. 13 Example of LGG for 176 kinds of French wine un vin rouge de Bordeaux un vin de Bordeaux rouge un rouge de Bordeaux un Bordeaux rouge un Bordeaux un rouge.... du vin d'Alsace blanc du vin blanc d'Alsace du blanc d'Alsace de l'Alsace de l'Alsace blanc du blanc
Finite State Transducer Finite State Transducer for IE 14
3. The Proposed System 1) Preliminaries 2) Overall Architecture 3) Extraction Module 4) Disambiguation Module
Preliminaries Meeting Location Definition of “Meeting Location” : A location where the meeting will be held 16 ::= | | | | ::= | ::= | | | | | ::= | | | | ::= ::= 수도권 | 부산 | 대구 | 광주 | 대전 ::= 인천선 | 분당선 | 중앙선 | 공항철도 | 경의선 | 경춘선 | 1 호선 | 2 호선 | 3 호선 | 4 호선 | 5 호선 | 6 호선 | 7 호선 | 8 호선 | 9 호선
Overall Architecture 17 Extraction Module Disambiguation Module Query Disambiguated Result Mobile Device Server Linked Data Finite-State Transducers INPUT OUTPUT 제목 : 팀장회의 공지 2008 년도의 마지막 팀장회의가 11 월 22 일 토요일 오 후 2 시 종로 토즈에서 열립니다. 재계약 그리고 명함 배 부가 이뤄질 예정이니 팀장님, 그리고 차기팀장님들 모 두 와주시기 바랍니다. 오시는 길 : 종로 종각역 4 번 출구에서 내려서 100m 정 도 걸어오시면 오른쪽에 있습니다. 팀장회의 공지 장소장소 명칭종로 토즈 주소 대한민국 서울특별시 종 로구 종로 가동 84-8 GPS 좌표 ( , ) Template Generator Personal GeoData
Extraction Module (1/2) 1.Construct Local-Grammar Graph (LGG) Find local patterns around meeting location, inductively. Scope of local patterns : Previous/Next/Current sentence including meeting location. Describe local patterns with 110 information types under 7 categories. Location, Time, Title, Actor, Label, Connecting words, Etc. e.g. ‘ 장소 : ‘ is ‘locLbl’ information type under ‘Label’ category. 2.Convert LGG to Finite-State Transducer (FST) 3.Extract Meeting Location by FST 학술대회 일정 : 2003 년 5 월 17 일 ( 토요일 ) 10:30 ~ 16:30 3. 학술대회 장소 : 성공회대학교 피츠버그관 4. 학술대회 순서
Extraction Module (2/2) Category of LGG for Meeting Location 개최장소 1 개 1.1. 장소 장소 장소 1_1 | 장소 1_ 장소 1_1 | 장소 1_2 | 장소 1_ 장소 + 랜드마크 장소 | 랜드마크 장소 1_1 | 장소 1_2 | 랜드마크 장소 | 랜드마크 1 | 랜드마크 장소 + 주소 장소 | 주소 장소 1_1 | 장소 1_2 | 주소 장소 1_1 | 장소 1_2 | 장소 1_3 | 주소 장소 | 랜드마크 | 주소 2. 개최장소 N 개 (N>1) 2.1. 개최장소 2 개 2.2. 개최장소 3 개 2.3. 개최장소 4 개 1. 일시 및 장소 : ( 수 ) 14:00~16:00, 무역협회 중회의실 ( 삼성동 트레이드 타워 51 층 ) 3. 장 소 : 울산광역시 울주군 상북면 등억리 27 번지 먹고쉬었다가 ( )
Disambiguation Module (1/2) Problem Multi-reference ambiguity (Pouliquen et al. 2006) two different geographic locations share the same name e.g. "Cambridge" is it Cambridge, UK or Cambridge, Massachusetts? Disambiguation by Linked Data Personal Geo Data Personalized OpenStreetMap User can map and save geographical location to the ‘meeting location’ (should be applied, consulting by Claus at Leipzig Univ.) Open Geo Data Naver Local Search API Yahoo! POI Search API Seoul Bus-stop DB Disambiguation by applying Ranking algorithm (idea will be borrowed from meta-search researches) disambiguate with 1st ranked geographical location 20
Disambiguation Module (2/2) 21 Personal Geo Data Query : 동측식당 Linked Data Personal Geo Data 동측식당 Naver Local API Yahoo! POI API Seoul Bus-stop Open Geo Data Disambiguation 동측식당 동측식당
4. Experimentation 1) Experiment Data 2) Extraction Module 3) Disambiguation Module
Experiment Data Meeting announcement corpus 1101 meeting announcements Collected from the web, with keyword ‘notice’ Annotation 10 types of term, 13 types of relation 3 human annotators with COAT annotation toolkit 23
Extraction Module Exp1. Extraction speed/memory comparison Baseline system : ML based system Dataset : already gathered corpus (training/test set) Exp2. Extraction performance comparison Baseline system : ML based system Evaluation : Precision/Recall/F-measure Dataset : already gathered corpus (training/test set) newly gathering corpus 24 (Experimentation should be followed)
Disambiguation Module Exp1. Accuracy in distance 6 types of distance : 0≤x≤100m, 100m≤x<1km, 1km≤x<2km, 2km≤x<3km, 3km≤x<5km and 5km ≤x Exp2. Accuracy Improvement with Personal Geo Data Evaluation : hard to show the performance show some scenarios how can it be applied so that it can improve accuracy. Exp3. Performance of Ranking Algorithm comparison Exp4. Disambiguation speed/memory comparison processing and communication speed/memory comparison on Server vs. on Mobile device 25 (Experimentation should be followed)
5. Conclusion 1) Assessment of the Approach 2) Limitation and Future Work
References 27
Abstract 배경 : 연구배경, 문제점, 필요성 (1/2) 논문에서 제시한 해결방안 해결방안의 장점 28