Cross-Language Text &
`Speech Retrieval
`
`Papers from the 1997 AAAI Symposium
`
`David Hull & Doug Oard, Cochairs
`
`March 24—26, Stanford, California
`Technical Report SS-97-05
`
`AAATPress
`MenloPark, California
`
`AOL Ex. 1019
`Page 1 of 7
`
`AOL Ex. 1019
`Page 1 of 7
`
`

`

`Copyright © 1997, AAAI Press
`
`The American Association for Artificial Intelligence
`445 Burgess Drive
`Menlo Park, California 94025
`
`ISBN 1-57735-040-5
`
`SS-97-05
`
`Manufactured in the United States of America
`
`AOL Ex. 1019
`Page 2 of 7
`
`AOL Ex. 1019
`Page 2 of 7
`
`

`

`crosslingual interrogation
`Multilingual database and
`inareal internet application
`
`Architecture and problems of implementation
`
`Christian Fluhr, Dominique Schmit, Faiza Elkateb, Philippe Ortet, Karine Gurtner
`
`Commissariat 4 l'Energie Atomique
`Direction de !'Information Scientifique et Technique
`CEA/Saclay
`91191 Gif sur Yvette cedex
`France
`E_mail : flubr@tabarly.saclay.cea.fr
`
`Abstract
`The EMIR European project demonstrated the
`feasibility of a crosslingual interrogation of fulltext
`databases using a mono and a bilingual general
`reformulation of the query.
`accept
`to
`New developments have been done
`multilingual databases even if there is more than one
`language in the same documents.
`Problems of implementation of full-scale applications
`are discussed.
`
`Background
`
`Our crosslingual text retrieval technology is based on the
`results of the EMIR (Flubr 1994) (European Multilingual
`Information Retrieval) project in the framework of the
`ESPRIT European Program . The project was done on
`three languages: English, French and German, with
`partners from France, Germany and Belgium. The project
`was performed from October 90to April 94.
`
`technological support an already
`The project had as
`existing multimonolingual
`(French and English)
`text
`retrieval software : SPIRIT (Syntactic and Probabilistic
`Indexing and Retrieval of Information in Texts), SPIRIT
`has been marketed since 1980 on IBM mainframe, since
`1985 on various platforms (from PC to mainframes) and
`from 1993 in a client-server architecture.
`
`« The translation and retrieval process are mixed so that
`the fulltext database is used as a semantic knowledge
`for solving translation ambiguities|which are the main
`At this time a part of EMIR results are introduced in the
`problem to solve
`in this
`approach of domain
`SPIRIT system. Applications are especially for the French-
`independent reformulation. One of the main result of
`English couple. A new language has been added : Russian
`this approach is that if an answer to the query exists in
`and the Dutch language is on the way.
`
`Main principles of the approach of
`crosslingual interrogation
`
`The EMIR’s approach is based on the use of a general
`transfer dictionary as a set of reformulation rules. That
`means that all possible translations are proposed as possible
`key for retrieving the documents.
`
`This approach is opposite to the one consisting of a
`translation of the query followed by a monolingual
`interrogation.
`
`There are 3 differcnces :
`
`*
`
`*
`
`the reformulation is based onthe translation of concepts
`in the query and need not to build up a syntactically and
`semantically correct translated query.
`
`the translation of concepts is used to obtain documents,
`if the word inferred in the target language is not the
`tight translation but if a relevant documents is obtained
`we can consider that the system works properly ( for
`example : inference of an hyperonym or a word of a
`different part of speech).
`
`AOL Ex. 1019
`Page 3 of 7
`
`AOL Ex. 1019
`Page 3 of 7
`
`

`

`the database, the system, in most cases, can select the
`right translation by looking in most relevant documents.
`
`Principle of the method
`We suppose that
`the database is monolingual, we will
`discuss in the following paragraphs the problemsspecific to
`multilingual databases.
`
`Database processing
`The database is processed by the morphosyntactic parser.
`The results are normalized single words and compounds.
`Normalization is mainly based on a lemmatization but
`general synonymiescan be taken into account. For example
`« logiciel » and « software » in French are normalized by «
`logiciel », « harbour » and « harbor » in English are
`normalized by « harbor».
`
`As single words we assume really single words and
`idiomatic expressions like « monkey wrench » in GB
`English or « clé anglaise » in French. Compounds are
`words in dependency relations like « multilingual database
`». For each normalized word or compound a semantic
`weight is computed according to the information it brings
`to choosethe relevent documents.
`
`Query processing
`The query is processed by the same morphosyntactic
`parsing than for
`the database. Normalized words and
`compounds are produced with their part of speech.
`
`For each of these units, we try to infer all possible
`translations that agree with the part of speech. For example
`« light » adjective is translated by « léger » adjective in
`French but « light » noun is translated by « lumiére » noun.
`
`The filtering by the database lexicon is not sufficient to
`eliminate all wrong translations. So it is possible to take the
`translations contained in the most relevant documents (that
`means the ones that contain the maximum of the query
`wotds, especially the ones where words have the same
`dependency relations than in the query).
`
`It is necessary before performing this optimization to be
`quite sure that the « best » documents are relevant, that is to
`say that they contain a sufficient number of the most
`weighty words.
`
`If it is decided that the most relevant documents are really
`relevant ones, a feed back can be done on the transfer
`process, In a second step, only words compatible with the
`most relevant documents are proposed.
`
`This process is very strong to increase relevance butit has a
`bad effect on the recall because it can eliminate synonyms
`of the chosen words
`that are only in less relevant
`documents. So it is useful to follow this feed back by a
`monolingual reformulation in the target language. We are
`in the same situation that a well formed query directly in
`the target
`language or a well
`translated query that
`necessitates a monolingual reformulation to have a good
`recall,
`
`Exampleoftranslation and filtering
`
`Query : « spectroscopie de masse par temps de vol » ona
`base of 655000titles ofreports on Energy
`
`transfer rules:
`
`spectroscopie (Noun) : spectroscopy
`masse (Noun) : mass, bulk, ground, sledgehammer
`Compounds can betranslated globally or word for word. In
`temps (Noun) : stroke, tense, beat, time, weather, days
`this last case the word orderis rearrangedtofit the result of
`vol (Noun):flight, flock, theft
`the target
`language normalization of compounds, All
`compoundsthat cannot be translated word for word butit is
`not necessary to consider them as idiomatic expression. A
`compoundlike seat belt is really a belt on a seat and in
`French « ceinture de sécurité » is a belt for security.
`
`Generally, especially for single words, there is a lot of
`translations.
`
`Example : « talon » (French)--a(English) « heel », « crust
`», «spur», « stub », « conterfoil », « talon »
`
`At this level the results of multilingual inference is filtered
`by the database lexicon and a lot of translations that are
`incompatible with the domain are eliminated,
`
`After filtering by the databaselexicon :
`
`spectroscopie (Noun) : spectroscopy
`masse (Noun) : mass, bulk, ground
`temps (Noun): stroke, beat, time, weather, days
`vol (Noun): flight, theft
`
`After filtering by the best documentthat contain : « time of
`flight mass spectroscopy » where the system has recognize
`2 compounds «timeofflight » and « mass spectroscopy »
`
`spectroscopie (Noun) : spectroscopy
`masse (Noun) : mass
`temps (Noun) : time
`vol (Noun) : flight
`
`AOL Ex. 1019
`Page 4 of 7
`
`AOL Ex. 1019
`Page 4 of 7
`
`

`

`that means that the system has dynamically produced that:
`
`tempsde vol : time of flight
`spectroscopy de masse : mass spectroscopy
`
`The rearrangement of words after word for word translation
`has been obtained using rules depending from the couple of
`languages and the syntactic structure of the compounds.
`
`Architecture for multilingual databases on
`INTERNET
`
`The EMIR project has demonstrate that it is possible to
`query a monolingual database in a languagethatis different
`than the database language. But in our countries where we
`use our own language and English for scientific documents,
`the problem is
`that most of the databases contains
`documents in 2 or mores languages. Even in some cases,
`the information about a documentis in several languages.
`For example, in our library catalog, a document can have
`an original title in English, a translated title in French,
`keywords in English and French, a summary in English and
`French.
`
`Generally it is not possible to assume that the information
`is
`redundant between the
`languages. For
`example,
`keywords in English and in French are taken from different
`systems of indexing and they are not translations from each
`other,
`
`The mixed language databases have not been taken into
`account by EMIR andnorin the current version of SPIRIT
`system. That is the reason why,
`the problem has been
`solved through a new architecture based on a more standard
`Webarchitecture,
`
`For permitting the access of SPIRIT server using standard
`INTERNETclients (Netscape, MS explorer, tango,etc. ), a
`WWW-SPIRIT interface has been done at the end of 1995,
`
`This interface is being extended to support distributed
`multibase multilingual databases. That means that
`the
`system can managea logical database composed of several
`physical ones. The problem of mixed language databaseis
`supportedby this architecture in the following way:
`
`into as many
`* The mixed logical database is split
`physical databases as languagesin the logical one..
`
`*
`
`can have parts
`same documents
`the
`monolingual physical databases
`
`in several
`
`*
`
`¢
`
`«
`
`*
`
`during the interrogation the query in one language (for
`example the user’s mother tongue) is sent
`to each
`physical
`database
`composing
`the
`logical mixed
`language one.
`
`receive the query with an
`each database server
`information of what is the query language. The database
`server knowing the language of his database performs
`either a monolingual or a bilingual interrogation and
`sendresults to the interface,
`
`the interface must merge theresults, for that it computes
`a weight of each concept of
`the query on an
`hypothetical database
`composed of
`the
`retrieved
`documents
`ftom the various language parts. This
`processing is necessary because the weight of cach
`concepts computed in each part can be very different
`and cannot represent the weight of the global use of the
`concepts in the multilingual logical database.
`
`In SPIRIT the answers are grouped by intersection
`classes and the classes are sorted according to the
`weight of concepts in the intersection query-document,
`The merging of results can produce new classes and
`suppressothers.
`
`at the end when a user ask to see a document, parts from
`different languages are to be obtained from various
`monolingual physical databases. The location of word
`that is also obtained from various monolingual physical
`databases are used to highlight the words used to extract
`the document. This functionality is specially important
`in the case of crosslingual interrogation to control why a
`documents has beenretrieve.
`
`First applications
`Atthis time our library catalog are put on the web using the
`first version of the SPIRIT/W3 interface which cannot
`manage multilingual databases nor multibase interrogation.
`
`Our documents are mainly in French and English. We are
`obliged to choose only one indexing language, for example
`French. For
`this indexing language, all
`the linguistic
`support is given (lemmatization), for the other language all
`words ate considered as proper nouns and are not
`lemmatized.
`
`It is easy to show that to be sure to have a sufficient answer
`it is necessary to ask a simple query in French, and a query
`with all possible variants of the words in English.
`
`34
`
`AOL Ex. 1019
`Page 5 of 7
`
`AOL Ex. 1019
`Page 5 of 7
`
`

`

`Thanks to god, main of our users are happy with partial
`answers and have not seen the problems. But it is not a
`good idea to expect that this situation will continue.
`
`¢
`
`compound are in different parts of the linguistic data
`considered as simple compound (words in dependency
`relations) or idiomatic expressions and in SPIRIT their
`normalization can bedifferent.
`
`That is the reason why, we decided to implement in an
`operational way the prototype of
`the SPIRIT/W3
`multilingual interface. We aim at beginning the service for
`late march 1997 on our library catalogs, catalogs of reports
`Inalinguistic process which goes through many steps, a
`on energy from the IAEA agency and from ETDE and a
`lack of coherence can brake the inference line. When
`catalog of the publications done by personal belonging to
`managing a huge amount of linguistic data, it is necessary
`our organization. All these applications will be visible from
`employ many people from different origin and even
`the all
`internet. Unfortunately,
`fulltext databases will be
`reserved for intranet use.
`working in different part of the world.
`
`This list is not exhaustive.
`
`So it is not possible to trust only in the human behavior.
`The solution is to implement a control system who can
`verify all that can be verify automatically and suggests to
`the human corrections or additions of information.
`
`Problems to solve
`
`Going from a feasibility prototype to an operational
`software is a hard task especially when the problem to
`solveis crosslingual interrogation.
`
`Acquisition of new compoundsin the transfer
`dictionary
`Such a systems involvesalot of tools and a lot of linguistic
`An other need is the fast adding of new words for new
`data, If somewhere something lacks, the result is strongly
`databases. It is very easy to know what single word lack in
`perturbed, That is the reason why a very strong quality
`the
`source language dictionary and in the transfer
`control
`is necessary to minimize the possibility of
`dictionary. The main problem is what compounds that
`cannot be translated word for word must be introduced into
`discontinuity between the query character string and the
`wordsin an other language to search in the databaseindex.
`the transfer dictionary.
`
`Control quality on linguistic data
`There are many causes ofthe discontinuity in the linguistic
`process:
`
`the word is not in the dictionary of the source language
`
`the word pertain to an idiomatic expression which is not
`in the source languagedictionary.
`
`the word is in the source language dictionary but not
`with all
`the possible part of speech (example ;
`processing as a verb but not as a noun)
`
`the normalized word has no entry in the transfer
`dictionary
`
`the word cannot be translated word for word : a
`translation of compound must be put into the transfer
`dictionary.
`
`*
`
`*
`
`«
`
`*
`
`°
`
`*
`
`Various
`problem :
`
`strategies have been followed to tackle this
`
`*
`
`¢
`
`*
`
`automatic extraction of terminology and treatment by a
`specialist.
`
`treatment of bilingual corpuses in the same domain,
`extraction ofterminology in the two lingual version and
`determination ofthe compounds that are not translatable
`word for word, This way seems promising but cannot be
`followed at
`this time because of the problems of
`consistence in the linguistic data mentioned above. At
`this time too much false detection are obtained to permit
`a use in exploitation
`
`processing of multilingual thesaurus in the domain of
`the database. We have begun to process the INIS
`(IAEA) and ETDE (OECD)thesaurus. The first one is
`on atomic energy, the second one is on all kind of
`energy,
`
`in the source language is
`normalization of words
`different from the normalization ofthe left part of the
`transfer dictionary,
`
`It is difficult at this moment to give general conclusions.
`Our evaluation done during the EMIR project was that even
`
`Conclusion
`
`35
`
`AOL Ex. 1019
`Page 6 of 7
`
`AOL Ex. 1019
`Page 6 of 7
`
`

`

`Conference of the UW Centre for the New Oxford English
`Dictionary and text Research, UW Centre for the New
`OEDand text Research, Waterloo Ontario Canada
`
`Radwan Kh., Foussier F., Flubr C., Multilingual access to
`textual databases. RIAO'91 Conference, April
`1991,
`Barcelona.
`
`Radwan Kh., Fluhr C., Textual database lexicon used as a
`filter
`to resolve semantic
`ambiguity,
`application on
`multilingual information retrieval, 4th annual symposium
`on document analysis and information retrieval, Las Vegas,
`24-26 April 1995.
`
`with a lack of consistency and without adding any word
`into the dictionaries and without translation feed back, we
`had a decreasing of results of about 10 % in comparison
`with a monolingual interrogation with the same system.
`
`In comparison with the actual situation where databases are
`indexed in only one of the languages, the new architecture
`will increase the quality of the answer using only one query
`in one language. What will be the level of the result in
`comparison with really good translation in each languages,
`the evaluation is to be done.
`
`In fact, even if the level of performanceis not the best one,
`according the behavior of our users, (they ask only one
`question in one language),
`the results will be strongly
`ameliorated in comparison with the actual performance of
`the database.
`
`References
`
`Debili F., Fluhr C., Radasoa P., About reformulation in
`fulltext IRS, Conference RIAO 88, MIT Cambridge, mars
`1988, A modified text has been published in Information
`processing and management Vol. 25, N° 6 1989, pp 647-
`657.
`
`Pacific Rim
`Information,
`C., Multilingnal
`Fluhbr
`Intelligence
`on Artificial
`Intemational Conference
`(RICAN),"AI and Large-Scale Information", Nagoya, 14-
`16 November 1990,
`
`lexical
`Fluhr C., Radwan Kh., Fulltext databases as
`semantic knowledge for multilingual
`interrogation and
`machine translation, EWAIC'93 Conference, Moscow, 7-9
`September 1993.
`
`Flubr C., Mordini P., Moulin A., Stegentritt E., EMIR Final
`report, ESPRIT project 5312, DG II, Commission of the
`European Union, October 1994
`
`Fluhr C., Schmit D,, Ortet P., Elkateb F., Gurtner K.,
`Semenova V., Distributed multilingual
`information
`retrieval, MULSAIC Workshop, ECAI96 Conference,
`Budapest, 12-16 August 1996
`
`Gachot D., Lange E., Yang J, Teh SYSTRAN NLP
`Browser:An Application of Machine Translation
`Technology in Multilingual Information Retrieval, Cross-
`Linguistic Information Retrieval Workshop, SIGIR’96,
`August 18-22, Zurich, Switzerland.
`
`Landauer T. K., Littman M. L., Fully Automatic Cross-
`Language Document
`retrieval Using Latent Semantic
`Indexing,
`(1990)
`in Proceedings of the sixth Annual
`
`36
`
`AOL Ex. 1019
`Page 7 of 7
`
`AOL Ex. 1019
`Page 7 of 7
`
`