Multilingual document retrieval system and method using semantic vector matching

• US 6,006,221 A
• Filed: 08/14/1996
• Issued: 12/21/1999
• Est. Priority Date: 08/16/1995
• Status: Expired due to Term

• Current Assignee: Syracuse University
• Original Assignee: Syracuse University
• Inventors: Liddy, Elizabeth D., Paik, Woojin, Yu, Edmund S., Li, Ming
• Primary Examiner(s): Choules, Jack M.
• Application Number:
  US08/696,701
• Time in Patent Office: 1,224 Days
• Field of Search: 707/3, 707/4, 707/5, 707/6, 704/2, 704/4, 704/5, 704/8, 704/9
• US Class Current: 1/1
• CPC Class Codes: G06F 16/24526 Internal representations for queries
• CPC Class Codes: G06F 16/313 Selection or weighting of terms for indexing
• CPC Class Codes: G06F 16/33 Querying
• CPC Class Codes: G06F 16/3334 Selection or weighting of terms from queries, including natural language queries
• CPC Class Codes: G06F 16/3337 Translation of the query language, e.g. Chinese to English
• CPC Class Codes: G06F 16/3338 Query expansion
• CPC Class Codes: G06F 16/93 Document management systems
• CPC Class Codes: Y10S 707/99935 Query augmenting and refining, e.g. inexact access

Abstract

A document retrieval system where a user can enter a query, including a natural language query, in a desired one of a plurality of supported languages, and retrieve documents from a database that includes documents in at least one other language of the plurality of supported languages. The user need not have any knowledge of the other languages. Each document in the database is subjected to a set of processing steps to generate a language-independent conceptual representation of the subject content of the document. This is normally done before the query is entered. The query is also subjected to a (possibly different) set of processing steps to generate a language-independent conceptual representation of the subject content of the query. The documents and queries can also be subjected to additional analysis to provide additional term-based representations, such as the extraction of information-rich terms and phrases (such as proper nouns). Documents are matched to queries based on the conceptual-level contents of the document and query, and, optionally, on the basis of the term-based representation. The query'"'"'s representation is then compared to each document'"'"'s representation to generate a measure of relevance of the document to the query.

Claims

1. A method of representing documents in a database that includes documents in a plurality of languages, the method carried out for each document, comprising:

determining a set of potential conceptual-level meanings of at least some words in the document from a language-independent multilingual concept database that reflects the plurality of languages and comprises a collection of concept groups;

mapping the sets of potential conceptual-level meanings, so determined, to respective single language-independent conceptual-level meanings; and

generating a language-independent conceptual representation of the subject content of the document based on the language-independent conceptual-level meanings determined in said mapping step.

2. The method of claim 1, and further comprising, determining the language of the document for at least some documents.

3. The method of claim 1, and further comprising

generating a term-based representation of the document for at least some documents.

4. The method of claim 3 wherein the term-based representation of the document is a representation of a set of proper nouns found in the document.

5. The method of claim 4 wherein the set of proper nouns found in the document are represented as categories from a hierarchical classification scheme.

6. The method of claim 3 wherein the term-based representation of the document is a representation of a set of noun phrases found in the document.

7. The method of claim 1 wherein a given one of said words is polysemous, giving rise to multiple conceptual-level meanings from the multilingual concept database, and said mapping, for the given word, comprises:

disambiguating among the multiple conceptual-level meanings from the multilingual concept database to provide a single multilingual conceptual-level meaning;

mapping the single multilingual conceptual-level meaning to a set of monolingual concept categories in a monolingual concept dictionary; and

if the set of monolingual concept categories from the monolingual concept dictionary contains multiple monolingual concept categories, disambiguating among the multiple monolingual concept categories to provide the single language-independent conceptual-level meaning.

8. The method of claim 7 wherein disambiguating includes:

analyzing local context information to attempt to determine a single meaning;

if a single meaning is not determined from analyzing local context information, analyzing domain knowledge to attempt to determine a single meaning; and

if a single meaning is not determined from analyzing domain knowledge, analyzing global information to attempt to determine a single meaning.

9. The method of claim 7, and further comprising

providing a gloss transliteration using the single multilingual conceptual-level meaning derived from disambiguating among the multiple conceptual-level meanings.

10. The method of claim 1 wherein the collection of concept groups includes words or phrases, from the plurality of languages, that are conceptually synonymous.

11. The method of claim 1, and further comprising:

determining a measure of proximity of the language-independent conceptual representation of each document to the language-independent conceptual representation of the other documents in the plurality; and

clustering the documents in the plurality according to the documents'"'"' respective measures of proximity to each other.

12. A method of retrieving documents in response to a query, the query being in a user-selected language of a plurality of languages, the method comprising:

providing a corpus of documents, each in a language of said plurality of languages, at least one of the documents being in a language other than the user-selected language;

for each document;

determining a set of multilingual concepts of at least some words in the document using a language-independent multilingual concept database that reflects the plurality of languages and comprises a collection of concept groups;

mapping the sets of multilingual concepts, so determined, to respective single language-independent conceptual-level meanings; and

generating a language-independent conceptual representation of the subject content of the document based on the language-independent conceptual-level meanings determined in said mapping;

generating a language-independent conceptual representation of the subject content of the query; and

for each document, generating a measure of relevance of the document to the query using the conceptual representation of the subject content of the document and the conceptual representation of the subject content of the query.

13. The method of claim 12 wherein the query is a natural language query.

14. The method of claim 12 wherein generating a language-independent conceptual representation of the subject content of the document comprises:

generating a conceptual-level vector representing the subject content of the document.

15. The method of claim 12 wherein;

the concept groups include a collection of synonyms and near-synonyms of a given word or phrase in said plurality of languages;

a given word in the document is polysemous, giving rise to multiple conceptual-level meanings from the multilingual concept database; and

for the given word, mapping the sets of multilingual concepts to respective single language-independent conceptual-level meanings comprises;

disambiguating the multiple conceptual-level meanings.

16. The method of claim 15 wherein disambiguating the multiple conceptual-level meanings comprises:

disambiguating among the multiple conceptual-level meanings from the multilingual concept database to provide a single multilingual conceptual-level meaning;

mapping the single multilingual conceptual-level meaning to a set of monolingual concept categories in a monolingual concept dictionary; and

if the set of monolingual concept categories from the monolingual concept dictionary contains multiple monolingual concept categories, disambiguating among the multiple monolingual concept categories to provide the single language-independent conceptual-level meaning.

17. The method of claim 16 wherein disambiguating includes:

analyzing local context information to attempt to determine a single meaning;

if a single meaning is not determined from analyzing local context information, analyzing domain knowledge to attempt to determine a single meaning; and

if a single meaning is not determined from analyzing domain knowledge, analyzing global information to attempt to determine a single meaning.

18. The method of claim 16, and further comprising

providing a gloss transliteration using the single multilingual conceptual-level meaning derived in disambiguating among the multiple conceptual-level meanings.

19. The method of claim 12, and further comprising providing a gloss transliteration of at least some of the words in at least one of the documents.

20. The method of claim 12 wherein said language-independent conceptual representation of the subject content of the document is augmented by a language-dependent statistical index using words in the document'"'"'s language as indexing units.

21. The method of claim 12 wherein said language-independent conceptual representation of the subject content of the document includes a statistical index using N-gram style decomposed words as indexing units.

22. The method of claim 12 wherein generating a language-independent conceptual representation of the subject content of the query comprises:

mapping words or phrases in the query into language-independent concepts; and

generating a conceptual-level vector representing the subject content of the query.

23. The method of claim 12 wherein said language-independent conceptual representation of the subject content of the query includes N-gram style decomposed terms as language-independent query requirements.

24. The method of claim 12 wherein said language-independent conceptual representation of the subject content of the query is augmented using a language-dependent logic requirement, the logic requirement including terms and logical connectives where the terms include the query term and its synonymous terms in the multilingual concept database.

25. The method of claim 12, and further comprising:

providing a list of at least some of the documents;

receiving user input specifying at least one document, or a part thereof, on the list; and

generating a revised query representation based on the original query plus a representation of the specified document or documents, or parts thereof.

26. The method of claim 12, and further providing a relevance-ranked list of at least some of the documents.

27. The method of claim 26, wherein the number of documents in the relevance-ranked list of documents is calculated based on a user-specified level of recall.

28. The method of claim 26, wherein the number of documents in the relevance-ranked list of documents is calculated based on a user-specified level of recall and a user-specified level of confidence in that level of recall.

29. The method of claim 26, wherein retrieved documents in the relevance-ranked list of documents are ranked without regard to the language they are written in.

30. The method of claim 12, and further determining the language of the document before generating a language-independent representation of the document.

31. The method of claim 12 wherein generating a measure of relevance for a given document comprises:

generating conceptual-level vectors for the given document and for the query; and

determining a distance between the vectors, the distance representing the measure of relevance, with a smaller distance representing a higher degree of relevance.

32. The method of claim 12 wherein generating a measure of relevance for a given document comprises:

generating an N-gram decomposed term representation for the given document and for the query; and

determining a degree of overlap between the N-gram decomposed terms, the overlap representing the measure of relevance, with a larger overlap representing a higher degree of relevance.

33. The method of claim 12 wherein generating a measure of relevance for a given document comprises:

generating word representations for the given document and for the query;

organizing words in the query as logical requirements; and

determining a coverage of terms in the documents against the logical requirement of a query, the coverage representing the measure of relevance, with a larger coverage representing a higher degree of relevance.

34. A method of retrieving documents in response to a query in a user-selected language of a plurality of languages, the method comprising:

(a) providing a corpus of documents, each in a language of said plurality of languages, at least some of the documents being in a language other than the user-selected language;

(b) processing each document by

determining the language of the document,

determining conceptual-level meaning of at least some words in the document from a language-independent multilingual concept database comprising a collection of concept groups,

mapping the conceptual-level meanings into language-independent concepts, and

generating a conceptual-level vector representing the subject content of the document;

(c) processing the query by

mapping words or phrases in the query into language-independent concepts, and

generating a conceptual-level vector representing the subject content of the query; and

(d) for each document, determining a measure of relevance to the query.

35. The method of claim 34 wherein mapping the conceptual-level meanings into language-independent concepts comprises

disambiguating multiple senses of polysemous words and phrases to generate the language-independent concepts.