1	eDMV WebMT Project: Spanish Machine Translation for the NYSDMV Web Site Presentation to the NYSFIRM State Webmasters Guild Richard Vang, Internet Language Services Dept. of Information Technology, NYS DMV December 5, 2003
2	The eDMV WebMT Project: Presentation Agenda Why Spanish Web Content? (customer needs, precedents, goals) Translation Methods Overview (human, machine) Controlled Content & Technical Dictionaries Project Processes, Tools & Resources Project Issues & Concerns (technology, processes) Current Status & Future Endeavors Project Metrics (feedback, statistics)
3	DMV Foreign Language Services: Assessing Resources, Setting Goals
4	Foreign Language Services: DMV Language Precedents (in Dec. 2001)
5	Initial Target Customers: Spanish-speaking Population We knew (as of Jan. 02): Total population of NYS (per Census 2000): 18.9 million Total speaking 2^nd language @ home: 5.18 million (27.44%) Total Hispanic/Latino population: 2.86 million (15.1%) (3% increase since 1990) Total speaking Spanish @ home: 2.6 million (13.77%) Total that speak English “less than well”: 1.17 million (6.23%)* >50% of Hispanic adults on-line (increases of 20% past two years) 0.72% of monthly calls to DMV Call Centers are in Spanish Average number of “unique visitors” to DMV web site per month (as of 1/02): 119,399 Therefore: Potential customer translations per month: 860 to 7,438*
6	Finding a Solution: Investigation of Translation Methods Important terms & concepts Translation accuracy Two primary methods for translating content: Human Translation Process Machine Translation Process
7	Investigation of Translation Methods: Important Translation Terms & Concepts Source/Target Language - the original (source) and resulting (target) languages of a translation process Language Pair - the two interacting languages, e.g. English AND Spanish (English – source; Spanish – target) Language Direction - translating from one language to another, e.g. English => Spanish or Spanish => English Translation Gap - length of time it takes to produce a translated product from the original source content Literary Quality – a translation produced as if composed by a native of the target language Gisting – translating just accurately enough to be understood at the first reading
8	Translation Accuracy: A Moving Target What is “accurate” translation? Who judges the accuracy? Nuances of language dialects “It’s about the information, stupid”
9	Human Translation (HT) Process: Current Method for Published Content
10	Human Translation (HT): Advantages & Disadvantages Advantages Literary quality translation Direct, specific control of which content & terms gets translated Disadvantages One language pair per translator (unless multi-lingual) Doubles work load of Webmaster by creating another version of web site (in a language they don’t know) Expensive ($8900 for DMV DM @ 104 pages) Long translation gap (3 months for DM) Updates can be problematic (process, cost, obsolete info) Different translators create different translations due to dialect and inconsistent terminology
11	Machine Translation (MT) Overview: What is “Machine Translation”? One of computing industry’s earliest endeavors Now coming into its own via global market Basically a word for word substitution (gisting) Accuracy enhanced by: Building customized dictionary for specific domains Applying semantic and linguistic rules to dictionaries Available “off the shelf” as a translation tool for single user (CAT – personal solution) Available as real-time, “on the fly” web site translation (WebMT – enterprise solution) Not a substitute for human translation, but a substitute for NO translation
12	The Machine Translation (MT) Process: How MT Works
13	Machine Translation (MT): Advantages & Disadvantages Advantages Decreases HT costs Real time; no translation gap Multiple language pairs leverage same core dictionary Second version of site not created Can leverage already existing translations and dictionaries for use in other DMV projects (i.e., printed materials) Disadvantages Less accurate than HT (gisting) High potential for nonsense translation (but controllable) Less specific control over what gets translated Does not translate graphics
14	SDL International Enterprise Translation Server
15	Project Implementation Phase I: Optimize DMV Site Content for Web MT Review of existing content is recommended by all MT vendors & advocates to achieve best MT results : Use simple grammatical structures Use short, concise sentences Use active verbs (vs. passive) Avoid abbreviations when possible (COD) Use terminology & acronyms consistently Avoid slang Use proper punctuation and accurate spelling Use definite articles where possible Avoid personal and gender-specific pronouns Use a simple format for text layout
16	Project Implementation Phase I: Why Optimize Site Content? Increases accessibility, decreases reading level 19% of NYS pop. >25 has NO high school diploma Brown U. eGov reports ½ of Americans read at 8^th grade level or lower most government sites at 11^th or 12^th grade level (NY: 10.7) Creates a clear, concise source language (simple, not “dumbed-down”) “Translation” of government-speak for less literate or less educated (understanding reduces customer inquiries) Results in cleaner, more accurate MT
17	Project Implementation Phase II: Create Lexicon for the DMV Domain Identify and add DMV-specific terms and phrases to DMV Technical Dictionary) Core dictionary & custom/technical dictionary Develop in accordance and simultaneously with optimization (Plain English) process Leverage existing dictionaries of DMV stakeholders and knowledgeable staff
18	Project Implementation Phase II: Why Create a Technical Lexicon? Has far-reaching and positive consequences which affect many stakeholders & publications Creates an ontology or taxonomy for DMV knowledge management (what is there?) Forces content creators to really think about what they’re saying (Ex: license plates) Creates a consistent language across web site and other publications (Ex: “license” terms) Helps to define how information is organized and presented (content management)
19	Optimization Tool: Smart Communications, Inc. MAXit Controlled English Checker Only software solution that integrated a dictionary building tool, a controlled English writing tool, and training for writers. Created for international airline manufacturing industry Based on “one word, one meaning” Optimizes source content according to MT recommendations Intended to remove ambiguity, create consistent style at easy reading level Identifies common problems in source text before applying MT Maintains a consistent style of writing within a workgroup Forces the writer to use standard terminology and spelling Text Miner: analyzes words for frequency, usage and importance in context, saved us time in developing a technical lexicon Lexicon Manager: Helps create and manage custom dictionaries Export tool to SDL Dictionary Manager
20	Live Demonstrations: Plain English & Dictionary Tools MAXit Controlled English Checker SMART Lexicon Manager SDL International Dictionary Manager
21	Example Handouts: PE-revised Text & Improving MT Accuracy with PE
22	eDMV WebMT Project: Project Processes, Tools and Resources Brief Timeline Project Process (content phases) Project Personnel (resources) Project Management Tools (databases)
23	eDMV Web MT Project: Brief Project Timeline I 12/01 – Initial Research 1/02 – Presentation to DMV Executive Guidance Committee 3/02 – Contracts signed with vendors via ASAP; training provided by vendors 4/02 – Pilot Phase begins with initial content “cluster”
24	Plain English Review Process: Project Phases and Content Clusters Pilot Phase – Driver License cluster Phase 2 – Home Page, various content menu pages, second-level menu pages Phase 3 – RightNow Web FAQs & interfaces Phase 4 – Registration & Title cluster Phase 5 – remaining translatable content Remaining Content – transactions, forms, etc.
25	eDMV WebMT Project: Project Personnel Project Manager Web Site Content Manager Plain English Review Team DMV Webmasters Internet Services Manager Spanish Translation Review Team
26	eDMV WebMT Project: Project Management Tools Project web page on intranet Project Database Custom built in MS Access Provides a variety of tools Accessible by all project personnel via LAN Built into process to provide accurate PM tracking data
27	Spanish Review Process: Improving MT Accuracy Team of volunteers (downstate CS reps) Bought into PE and MT process as a way of possibly lightening their Spanish interactions Two levels of Spanish MT review for approval Project Mgr. initial review (obvious problems, PE) Native speaker expert review (fine tuning, terms) Web page revision, dictionary update after each Review process (no Internet access @ work) PE version on DMV intranet mirror site MT version from Internet saved as HTML on LAN Changes in Word, or hardcopy mark-up
28	eDMV Web MT Project: Brief Project Timeline II 8/02 – Pilot Phase Goes live Site visitor access – home page Welcome page / Disclaimer “Hardcoded menu” of selected content cluster 12/31/02 – “FollowLink” features brings full site content on-line with Spanish MT Site visitor access – home page Welcome page / Disclaimer FollowLink Feature Content not translated
29	eDMV WebMT Project: Live Spanish MT Demonstration www.nysdmv.com in Spanish
30	eDMV WebMT Project: Technological Issues & Concerns MAXit Controlled English Checker government vs. technical language program fixes Word HTML code stripping (CSS) Temporary Dictionary feature Translation of graphic banners not possible without dynamic content suggested <IMG ALT> solution FollowLink Feature not really a site navigation feature bookmarks a problem blocking content from translation visitor trapped in Spanish, can’t get out Translation Issues
31	eDMV WebMT Project: Process Issues & Concerns Required constant vigilance on site updates Fluctuating personnel resources caused delays Budgetary process caused delays Translation accuracy is slave to available resources, not the fault of the technology
32	eDMV WebMT Project: Ongoing Processes (as of 12/03) Still in PE Review Process for Phases 5 and 3 (RNW) Bring RightNow Web on-line with MT (1^st) Still evaluating Forms pages (menu, titles) Still evaluating Transaction pages (PE) Still requires dictionary clean-up & more lexicon building (acronyms)
33	Future Endeavors: Are we “There” yet? Sell rest of agency on benefits of Plain English and controlled vocabulary Get more content creators to use MAXit Coordinate translation services to utilize existing MT translations as drafts for publications Install ETS plug-ins for email and other Office applications
34	eDMV WebMT Project: Project Metrics Plain English review process 94% of targeted web pages completed process (RNW: 36%) Pages being translated 81 web pages currently using MT (some 20+ printed pages) Spanish review process 16% of targeted web pages completed process Dictionary terms PE dictionaries: 6,669 terms; Spanish dictionaries; 2,126 User Stats 7,213 translations/month avg. (in prediction range) 50K translation by 7/03
35	eDMV WebMT Project: Customer & Industry Feedback Well received in MT and CE industries example of PE to soften government-speak (EU) example for international companies (Daimler-Chrysler) example of 1^st time ever that PE and MT applications were used together to create site content (SDL PE) Featured in Information Week (4/9/03) article on MT Compared to users, virtually NO feedback from customers
36	eDMV WebMT Project: Contact Information Project Manager Richard Vang (474-2570) rvang@dmv.state.ny.us DMV Web Site Content Manager George Filieau (486-6596) gfili@dmv.state.ny.us DMV Webmaster Holly New (474-2644) hnew@dmv.state.ny.us