CS 671 ICT For Development 19 th Sep 2008 Vishal Vachhani CFILT and DIL, IIT Bombay
Agro Explorer A Meaning Based Multilingual Search Engine Vishal Vachhani 2
Web-site for Indian farmers Farmers can submit their problems related to their crops Queries are answered by Agricultural Experts at KVK, Baramati Languages supported: Marathi, Hindi, English Vishal Vachhani 3
Why Need Multilingual Search Vast Amount of Information available ailable on the Web Almost 70% of the Information is in English The Indian rural populace is not English- Literate A Big Language Barrier Information has to be made available to them in their local languages. Vishal Vachhani 4
Why Need Meaning Based Search Most of the current Search Engines are Keyword Based. They do not consider the semantics of fthe query The result set contains a large number of extraneous documents. Search based on the Meaning of the query will help narrow down on the desired information quickly. Vishal Vachhani 5
Query in Hindi search System English Document Marathi Document Result in Hindi English Document Vishal Vachhani 6
Same Keywords Different Semantics Moneylenders Exploit Farmers Farmers Exploit Moneylenders Found 1 Result Found 0 Result Vishal Vachhani 7
Provides both Meaning Based Search Cross-Lingual Information Access Vishal Vachhani 8
System Architecture Vishal Vachhani 9
Vishal Vachhani 10
Vishal Vachhani 11
Vishal Vachhani 12
Vishal Vachhani 13
Vishal Vachhani 14
Conclusion Provides two independent d features Multi-Linguality Meaning Based Search. Because of UNL both multi-lingual and meaning based properties can be incorporated together rather than using separate language translators in search engines. The scheme admits itself to Integration of multiple languages in a seamless, scalable manner. Vishal Vachhani 15
UNL Universal Networking Language Vishal Vachhani 16
Hind i Englis h UNL Frenc h Marath i Tam il Vishal Vachhani 17
Direct translation - translation will be done directly - N*(N-1) translator are needed for N languages translation. Intermediate Language - intermediate language will be used for language translation - Only 2*N translators are required. Vishal Vachhani 18
UNL is an acronym for Universal Networking Language. UNL is a computer language that enables es computers to process information and knowledge across the language barriers. UNL is a language for representing information and knowledge provided by natural languages Unlike natural languages, UNL expressions are unambiguous. Vishal Vachhani 19
Although the UNL is a language for computers, it has all the components of a natural language. It is composed of Universal Words (UWs), Relations, Attributes. Knowledge :semantic graph Nodes concepts Arcs relation between concepts Vishal Vachhani 20
A UW represents simple or compound concepts. There are two classes of UWs: unit concepts compound structures of binary relations grouped together ( indicated with Compound UW-Ids) A UW is made up of a character string (an Englishlanguage word) followed by a list of constraints. <UW>::=<Head Word>[<Constraint List>] example state(icl>express) state(icl>country) Vishal Vachhani 21
A relation label is represented as strings of 3 characters or less. The relations between UWs are binary. rel (UW1, UW2) They have different labels according to the different roles they play. At present, there are 46 relations in UNL For example, agt (agent), ins (instrument), pur (purpose), etc. Vishal Vachhani 22
Attribute labels express additional information about the Universal Words that appear in a sentence. They show what is said from the speaker s point of view; how the speaker views what is said. (time, reference, emphasis, attitude, etc) @entry, @present, @progressive, @topic, etc. Vishal Vachhani 23
Example: Ram eats rice. {unl} agt(eat.@entry.@present, Ram) obj(eat.@entry.@present, rice(icl>eatable)) {/unl} Vishal Vachhani 24
eat plc agt Ram rice Vishal Vachhani 25
Example: The boy who works here went to school. {unl} agt(go(icl>move).@entry.@past, :01) plt(go(icl>occur).@entry.@past,school(icl>institutio n)) agt:01(work(icl>do), boy(icl>person.@entry)) plc:01(work(icl>do),here) {/unl} Vishal Vachhani 26
go agt plt work :01 school plc agt here boy Vishal Vachhani 27
Source language Enconvertor Intermediate Language Deconvertor target language Vishal Vachhani 28
It s alanguage Independent Generator It can deconvert UNL expressions into a variety of native languages, using a number of linguistic data such as Word Dictionary, Grammatical Rules of each language. The DeConverter transforms the sentence represented by a UNL expression into Natural language age sentence. Vishal Vachhani 29
Vishal Vachhani 30
Dictionary Case Marking Rules Morphology Rules Syntax Planning Rules UNL Doc UNL Parser Case Marking Module Morphology Module Syntax Planning Module Hind idoc Language dependent Module Language Independent Module Vishal Vachhani 31
UNL parser module will do following tasks Check input format of UNL document Separate attributes form UWs Separate attributes form dictionary i entries Replace UWs with Hindi root words
Category of morpho-syntactic ti properties which distinguish the various relations that a noun n phrase may bear to a governing head. न, पर,क, स, प,etc. A rule base based on : UNL attributes lexical attributes from dictionary Vishal Vachhani 33
Case marking is implemented using rules. We analyze all UNL as well as dictionary attributes and decide next and previous case marker. Also we use relation with parent to extract the right case mark. Vishal Vachhani 34
agt:null:null:null:न :@past#v:vint:n:null agt:null:null:null:न:@past#v:vint:n:null Structure relname : parent previous case marker: parent next case marker: child previous case marker: child next case marker: the rest four are in form of attr'rel'relationname ti and attr will be separated by # also relation name are separated by # Vishal Vachhani 35
What is Morphology Study of Morphemes Their formation into words, including inflection, derivation and composition Vishal Vachhani 36
Noun, Verb and Adjective Morphology Depends on the phonetic properties of the Hindi word Noun Morphology Depends on gender, number and vowel ending of the noun Adjective Morphology अ छ लडक, अ छ लडक, अ छ लडक adjective अ छ changes, lexical l attribute t AdjA Verb Morphology Depends upon tense, gender, number, person etc. Vishal Vachhani 37
Verbs are categorized by Tense (past,present,future) Gender(male,female) Person (1 st, 2 nd, 3 rd ) Number (sg,pl) Example Ladaka khana kha raha hai. It contains present continuous tense,male, sg, and 3 rd person Vishal Vachhani 38
Arranging word according to the language structure Rule based module It is priority based graph traversal Vishal Vachhani 39
Algorithm for Syntax Planning: 1) Start traversing the UNL graph from the entry node. 2) If node has no children then add this node to final string. 3) If there is more than one child hldof one node then sort children hld based on the priority of the relations. Relation having highest priority will be traversed first. 4) Mark that node as visited node. 5) Repeat steps 3 and 4 until all the children of that node get visited. i 6) If all the children of that node get visited then add that node to final string. 7) Repeat steps 2 to 4 until all the nodes get traversed. Vishal Vachhani 40
Also, spray 5% Neemark solution. obj spray U 3 man obj:17 man:9 mod:5 qua:5 solution also mod mod percent Neemark qua 5 41 Vishal Vachhani
Entry spray Vishal Vachhani 42
Entry spray obj man Vishal Vachhani 43
Entry spray obj:17 man:9 Vishal Vachhani 44
Entry spray obj:17 man:9 solution Vishal Vachhani 45
Entry spray obj:17 man:9 solution mod mod Vishal Vachhani 46
Entry spray obj:17 man:9 solution mod:5 mod:5 Vishal Vachhani 47
Entry spray obj:17 man:9 solution mod:5 mod:5 percent Vishal Vachhani 48
Entry spray obj:17 man:9 solution mod:5 mod:5 percent Vishal Vachhani 49
Entry spray obj:17 man:9 solution mod:5 mod:5 qua:5 percent Vishal Vachhani 50
Entry spray obj:17 man:9 solution mod:5 mod:5 Output : 5 percent qua:5 5 Vishal Vachhani 51
Entry spray obj:17 man:9 solution mod:5 mod:5 percent qua:5 5 Output : 5 percent Vishal Vachhani 52
Entry spray obj:17 man:9 solution mod:5 mod:5 percent qua:5 5 Neemark Output : 5 percent Neemark Vishal Vachhani 53
Entry spray obj:17 man:9 solution mod:5 mod:5 percent qua:5 5 Neemark Output : 5 percent Neemark solution Vishal Vachhani 54
Entry spray obj:17 man:9 solution also mod:5 mod:5 qua:5 percent 5 Neemark Output : 5 percent Neemark Solution also Vishal Vachhani 55
Entry spray obj:17 man:9 solution also mod:5 mod:5 percent qua:5 5 Neemark Output : 5 percent Neemark Solution o also spray Vishal Vachhani 56
Output: 5 percent Neemark solution also spray 5 तशत न मअक घ ल भ छड़क 5 तशत न मअक घ ल भ छड़क Vishal Vachhani 57
Input sentence: Its roots are affected by bacterial infection. Module Input Output Its roots are affected by bacterial infection. UNL parser ज भ वत ज व वक स मण Case marking Morphology Syntax Planning ज भ वत ज व वक स मण स इसक जड़ ज व वक भ वत ह त ह स मण स ज व वक स मण स इसक जड़ भ वत ह त ह Output: ज व वक स मण स इसक जड़ भ वत ह त ह Vishal Vachhani 58
UNL 2005 Specifications: http://www.undl.org/unlsys/unl/unl2005/ S.Singh, M.Dalal, V.Vachhani, P.Bhattacharrya and O.Damani Hindi generation from interlingua MTsummit 2007 (www.cse.iitb.ac.in/~vishalv) in/~vishalv) Mrugank Surve, Sarvjeet Singh, Satish Kagathara, Venkatasivaramasastry K, Sunil Dubey, Gajanan Rane, Jaya Saraswati, Salil Badodekar, Akshay Iyer, Ashish Almeida, Roopali Nikam, Carolina Gallardo Perez, Pushpak Bhattacharyya, AgroExplorer Group: AgroExplorer: a Meaning Based Multilingual Search Engine, International Conference on Digital Libraries (ICDL), New Delhi, India, Feb 2004. Agro Explorer : http://agro.mlasia.iitb.ac.in aaqua : http://www.aaqua.org Vishal Vachhani 59