1.4k likes | 1.63k Views
System Architecture Components. Logos Translation Systems Corp. March 12, 1998. Logos Translation System. ALEX (Semantha). LogosClient. Logos Translation Term Search Terminology Management. LogosClient System Architecture. Sentence Memory Word Processor Third Party Applications.
E N D
System Architecture Components Logos Translation Systems Corp.March 12, 1998
Logos Translation System ALEX(Semantha) LogosClient Logos Translation Term Search Terminology Management
LogosClient System Architecture Sentence MemoryWord ProcessorThird Party Applications LogosClient(Java) RMI Logos Job Server JDBC ODBC Database LogosTranslationServer
LogosClient System Architecture Sentence MemoryWord ProcessorThird Party Applications LogosClient(Java) RMI Logos Job Server JDBC ODBC Database LogosTranslationServer
Customer Support Logos Translation Systems Corp.March 1998
Support Services • Status Quo • Strategy for the Future
Support Services • Customer Support • QA • Documentation
Customer Support • Tasks • Customer Expectations • Skillset • Organization
Status Quo Customer Support Tasks • Pre-Sales • Process Consulting • Sample Translations • Sales Training • Post-Sales • Setup and Installation • Training • Troubleshooting (Linguistic and Technical)
Status Quo Customer Expectations “Symbiosis”Resource drain Unlimited support
Skillset Status Quo • Technical • FormattingNetworksOperating SystemsUnix, Windows • Overall Understanding of Software • Linguistic • Translation • Lexicography • Grammatical Concepts • Tools • Logos • Translation Memory • Variety of Terminology Tools • Communication, People Skills
Organization Status Quo • Distribution of Expertise: • First-level Support: Eschborn, Santa Clara. • Second-level Support: Mount Arlington. • CSR Tracking • Mail-based. • Flaws • Problem Administration rather than Solution. • CSR Tracking is Labor intensive and inaccurate.
Strategy for Future • Correct Flaws of Current System • React to New Products and Services
Strategy for Future • Tasks • Organize Service Support • Customer Expectations • Support Level Agreements (Standard vs. Non-Standard support) • Non-Standard Support as Source of Revenue • Skillset • Specialists for specialized tasks: - Distinction between Technical Suppport, - End User Consulting and Service Support • Organization • From mail file to database • From Problem Administration to Problem Solution: Limitations?
Overview of the Logos Model Logos Translation Systems Corp.March 10, 1998
Overview of the Logos Model • Design Goals in 1970 • Design Issues and Motivations • Design Solutions: Internal Language • Design Solutions: Architecture • How this process works
Design Goals of the System • Modular, portable • Language-neutral software • Multi-target • Text-driven • Industrial strength MT: unconstrained text • Fully automatic high quality translation
FAHQT high-quality 3 1 cost-effectivequality Logos 4 PERFORMANCE not costeffective 2 demoquality constrained text only general purpose systems Start COMPETENCE MT Quality/Functionality Matrix
greatest clarity objective to inform mind subjective to move, edify mind + heart formal conventional informal idiosyncratic content style least clarity natural language artificiallanguage contrivedlanguage contracts scientific newspapers essays fiction plays humor poetry legal briefs technical non-fiction letters academic memos logic programming languages experimental avant garde 99% Language Spectrum
Fundamental Design Issues • What kind of knowledge representation is needed? • How do you store that knowledge? • How do you apply that knowledge? • How do you avoid complexity issues?
Issues:What kind of Representation is needed? • Strong semantics, sine qua non • Pragmatic knowledge, unfeasible • “changes in cells produced by digitalis” • “Meaning” that can be managed • “he put the book on Tallyrand on the table”
supporting surface table shelf floor bench stool ledge Typical SAL rule pattern: Prep(from) N(support surface) –> Prep(off of) N(support surface) Second Order Abstraction
Semantico-Syntactic Continuum highchair literal string chair head morpheme support surface SAL subset functional device SAL set concrete object SAL superset noun word class SAL patterns can express elements at any point on the continuum “Message to citizens of Rome and friends” N(info) Prep(to) N(hum) Prep(of) N(place) Conj N(hum)
Rules as Semantico-Syntactic Patterns • “changes in cells produced by digitalis” • “changes in cells effected by x” • “changes in cellsaffected by x” • “changes in cells produced by x” SAL Rule: Pattern: PN Prep NP(~PN) PstP(27) * End of REL Constraint: PP is connominal to PN Action: ((PN(Prep NP))(REL)) Note: verb type 27 = produce, effect, accomplish, complete, etc.
Rules as Semantico-Syntactic Patterns • “He put the book on Tallyrand on the table.” Analyzed to here as: Pro V(63) NP(info) Prep(on) N(Hum) Prep(on) NP(surface) Further Analyzed: • - pattern: V(63) • - action: sets verb type cell to “63” • Note: verb type 63 governs locative preps : put, place, stow, store, etc. • - pattern: NP(info) Prep(“on”) NP(~surface) * Prep(“on”) NP(surface) • - condition: verb cell = 63 • - action: rewrites first Prep(loc) as PP(connominal) and attaches to NP(info) • “on” takes semantics of “concerning” • - pattern: V(63) Prep(loc) * NP(surface) • - action: rewrites Prep(loc) as PP(converbal), i.e. attaches PP to verb
Analogy - a key principle in language acquisition • from floor, chair, bench --> c N • from wall, side, fence --> ot N • from table --> ot N? or c N? • from table --> c N (by analogy) Prep(from) N(supportsurface) --> Prep(off of) N(support) --> c NP Human Mind Uses Second Order Abstractions
Use of Analogy in Human Language Learning • ask John to do it -- demander à Jean qu’il le fasse V NP to V PRO • hire John to do it -- engager Jean pour le faire V NP to V PRO • tell John to do it -- ? (1st construction, by analogy) V (44) NP to V PRO --> V a NP qu’il PRO V (subj) Pattern: V (44) NP * TO V Action: To V is rewritten as INF. COMP. Note: V (44) = tell, ask, request, permit, etc.
Issues: How do you store this knowledge? • Traditional model: lean rulebase, rich lexicon • Radical lexicalism: no rulebase (e.g., “shake and bake”) • Logos Model: lean lexicon, rich rulebase
Source Target Morphology Morphology Source Target Lexicon Lexicon LOGOS SHELL Source Target Rulebase Rulebase Source Target Semantic Semantic Table Table Lean Lexicon, Rich Rulebase
Issues: How do you apply that knowledge? • Relating the knowledge base to the input stream • One-to-many matching: lexical matching • Many-to-one matching: rule matching • Logos Model: one-to-many throughout entire knowledge base • Indexable lexicons • Indexablerulebases
Issues: Complexity • Complexity: the show stopper • Problem with algorithms: complexity • How do humans handle complexity?
“John took the things from the kitchen table and put them away.” n v det start adj halt • Process is controlled by transition diagram (logic) • Transitions must by pre-specified • As knowledge increases, logic becomes complex • Complexity grows logarithmically • To simplify, syntax and semantics are separated Algorithms and Complexity
off ofsupportsurface PP removeoff ofsupportsurface VP semantic fan-incircuit supportsurface syntactic fan-incircuit ledge 2 NP 3 floor supportsurface off of awayfrom geolog.location accept remove etc. out of 1 appro-priate infor-mation take from the kitchen table etc. ambiguityfan-outcircuit ambiguityfan-outcircuits John took the things from the kitchen table and put them away. Mental Model • Processor is associative memory itself and its interconnections • Sentence processing is done in stages (pipeline) • Syntax and semantics are integrated (a continuum) • Employs second-order language abstractions • Memory is content-addressable
Logos Model as an IncrementalPipeline Analyzer-Generator Semantic Dict NL Input Rules Semantic Tables Dict Rules Rules Format Rules Lex Rules Res1 BeginAnalysis Rules Res2 Rules Tran1 Rules Tran2 Tran3 Tran4 S Tran4 BeginGeneration Tran3 Tran2 Tran1 Target Gen Format Lex Lex Lex Rules Rules Rules Rules Rules Rules NLOutput Design Solutions: Architecture
V1 V2 V3 V4 V5 V6 V7 S R1 R2 T1 T2 T3 T4 Solutions: Neural Net Architecture
Output Vector 7 S Hidden Layer T4 LTM = long term memory STM = short term memory Intermediate Vector 6 3 12 9 8 11 Hidden Layer T3 Intermediate Vector 5 11 3 9 10 12 8 Hidden Layer T2 Intermediate Vector 4 3 10 7 9 12 4 8 11 Hidden Layer T1 Intermediate Vector 3 1 2 3 6 7 9 10 12 4 5 8 11 Hidden Layers R1 & R2 1 3 9 12 4 5 8 5 6 Input Vector 1 2 7 10 11 1 3 9 12 4 5 8 6 SAL Representation NL Dictionary A young man in a green sweater is waiting for you outside. 1 2 3 4 5 6 7 8 9 10 11 12 Stratified Net with Hidden LTM Layers Interleavedwith Intermediate STM Vectors
Unresolved Interface Structure V1 Resolved Interface Structure V2 segmentation intoclauses andsubclauses BOS Let me also note that because of the relatively close movement of the Canadian dollar with the US dollar, our currency has declined along with the US dollar against these other currencies this past year , removing much of the exchange rate distortion that was hampering the ability of Canadian firms to compete with producers overseas . uvrbtp d b npdjnpdnnudn pdnnpjnnurpnntvdnpjnvpnu uvrbtp d b npdjnpdnnudn pdnnpjnnurpnntvdnpjnvpnu v v j j Hidden Layers x x v v R1 & R2 d d d d j j v v x x o o b b 2,239,488 paths one path
Interface Structure V3 Interface Structure V4 u v r b v t p n p n p n u n x v p n p n b u v r p n t x v n p n v p n b u BOS Let me also note that because of movement of dollar with dollar, currency has declined along with dollar against currencies (adv)year , removing much of distortion that was hampering ability of firms compete with producers overseas . BOS Let me also note that because of the relatively close movement of the Canadian dollar with the US dollar, our currency has declined along with the US dollar against these other currencies this past year , removing much of the exchange rate distortion that was hampering the ability of Canadian firms to compete with producers overseas . uvrbtp d b npdjnpdnnudn pdnnpjnnurpnntvdnpjnvpnu v j Hidden Layer x v T1 d d j v x o b
Interface Structure V4 Interface Structure V5 u v r v t p n u n x v p n p n b u v n u n x v n n v p n b u v r b v t p n p n p n u n x v p n p n b u v r p n t x v n p n v p n b u BOS Let me also note that because of movement of dollar with dollar, currency has declined along with dollar against currencies (adv)year , removing much of distortion that was hampering ability of firms compete with producers overseas . BOS Let me note that because of movement , currency has declined along with dollar against currencies (adv)year , removing distortion . BEGIN NESTED CL (distortion) was hampering ability END NESTED CL BEGIN NESTED CL (ability) to-compete with producers overseas END NESTED CL . Hidden Layer T2
movement np V6 HiddenLayerT3 V5 V p n n p n n p n n k 5 T2 T2 T2 j j+1 j+2 n p n V4 V4 V4 i i+1 i+2 movement with dollar V6 V5 V4 V3 movement movement with dollar movement of dollar with dollar movement of the Canadian dollar with the US dollar
Interface Structure V5 Interface Structure V6 u v r v t u n x v b u v n u n x v n n v b BOS Let me note that because of movement , currency has declined along with dollar against currencies (adv)year , removing distortion . BEGIN NESTED CL (distortion) was hampering ability END NESTED CL BEGIN NESTED CL (ability) to-compete with producers overseas END NESTED CL . u v r v t p n u n x v p n p n b u v n u n x v n n v p n b BOS Let me note that because of movement , currency has declined along with dollar against currencies (adv)year , removing distortion . BEGIN NESTED CL (distortion) was hampering ability END NESTED CL BEGIN NESTED CL (ability) to-compete with producers overseas END NESTED CL . pp Hidden Layer pp pp T3 pp
Interface Structure V6 Interface Structure V7 BOS Let me note that because of movement , currency has declined along with dollar against currencies (adv)year , removing distortion . BEGIN NESTED CL (distortion) was hampering ability END NESTED CL BEGIN NESTED CL (ability) to-compete with producers overseas END NESTED CL . u v r v t u n x v b u v n u n x v n n v b S MC S1 pp MC Hidden Layer pp pp DC T4 pp
SAL Used for Syntactic DisambiguationUsing SAL to Recognize Governance • ways of cooking lentils “ways” governs verbals “cooking” = verb • types of cooking utensils “types” governs nonverbals, “cooking” = adjective
NOUNS 06 ABSTRACT 40 42 42 225 NON-VERBAL INTANGIBLES(about persons, things) NEGATIVECONCEPTS CONCEPTS anomalyanachronism horrormonstrosity justiceanalogy ideatruth INTRINSICCHARACTERISTICS CONDITIONRELATIONSHIP UNDIFFERENTIATED SOURCE/ORIGIN CLASSIFICATION 40 609 736 731 723 povertystatusfatherhoodinequality cancercircumstancecomaloneliness categoryclasskindmake natureranktype reserveresourcewell-spring baroquecolordesign featureform likeablenessprofiletraitshape 41 VERBAL INTANGIBLES(about agents, processes) 748 733 732 602 655 PROPERTY OFACTION, AGENT(of/in/for v’g) CAUSE, POTENTIALDISPOSITION(to v; of/for v’g METHOD, PROCEDURE(for/of v’g) PURPOSE(of v’g; to v) TIME EVENT modepattern efficiencysuitableness easealacrity instanceendholiday paradebirthday basissensetalentcost motivationreadinessability taskproblem objectivefunction techniquemeans 41 764 765 749 NEGATIVECAUSE/EVENT WEAK VERBALPROCESS/RESULT UNDIFFERENTIATED CONTRARY EVENT dangerthreat risk catastrophedisasteraccident blizzarderror noise abatementheat absorption
SAL Used for Syntactic DisambiguationUsing SAL for detecting semantic symmetry • bank and appliance store - (N1 and (N2 N3) => NP and NP - bank and (appliance store) • computer and TV store - ((N1 (agent) and N2 (agent)) and N3) => NP - (computer and TV) store