Intra-language statistical machine translation

• US 8,615,388 B2
• Filed: 03/28/2008
• Issued: 12/24/2013
• Est. Priority Date: 03/28/2008
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Li, Xiao, Ju, Yun-Cheng, Zweig, Geoffrey, Aero, Alex
• Primary Examiner(s): JACKSON, JAKIEDA R
• Application Number:
  US12/058,328
• Publication Number:
  US 20090248422A1
• Time in Patent Office: 2,097 Days
• Field of Search: 704/2, 704/5
• US Class Current: 704/2
• CPC Class Codes: G06F 40/44 Statistical methods, e.g. probability models
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment

Abstract

Training data may be provided, the training data including pairs of source phrases and target phrases. The pairs may be used to train an intra-language statistical machine translation model, where the intra-language statistical machine translation model, when given an input phrase of text in the human language, can compute probabilities of semantic equivalence of the input phrase to possible translations of the input phrase in the human language. The statistical machine translation model may be used to translate between queries and listings. The queries may be text strings in the human language submitted to a search engine. The listing strings may be text strings of formal names of real world entities that are to be searched by the search engine to find matches for the query strings.

Claims

1. A computer implemented method for intra-language machine translation of phrases in a human language, the method performed by one or more computers comprised of one or more processors and memory, the method comprising:

receiving training data, the training data comprising a list of text queries in the human language submitted to a search engine and a list of text sentences in the human language returned by the search engine when the text queries were submitted to the search engine, and forming pairings of source phrases and target phrases by comparing the text queries and the text sentences to identify text queries that are similar to text sentences and pairing the identified text queries, as the source phrases, with the respective identified text sentences, as the target phrases, and storing the training data in the memory;

processing each pairing by the one or more processors, the processing comprising, for a current pairing being processed, parsing the source phrase of the current pairing into source words and parsing the target phrase of the current pairing into target words, and computing an alignment of the source words and the target words;

using, by the one or more processors, the pairs of training data and their respective alignments to train an n-gram based intra-language statistical machine translation model, where the intra-language statistical machine translation model, when given an input phrase of text in the human language, can identify possible translations of the input phrase in the human language and compute probabilities of semantic equivalence of the input phrase to the possible translations of the input phrase in the human language; and

using the statistical machine translation model to find translations of queries and use the translations to evaluate listings that match the queries, where the queries comprise text strings in the human language submitted to the search engine, where the listing strings comprise text strings of formal names of real world entities that are found by the search engine as matches for the query strings.

2. A method according to claim 1, wherein the using the intra-language statistical translation model comprises:

receiving from the search engine listings that the search engine matched to a user'"'"'s query;

generating query forms of one of the listings by using the translation model to translate the one of the listings to the query forms;

using the translation model to compute similarities of the query forms to the user'"'"'s query, and determining that the listing does not match the user'"'"'s query based on the computed similarities.

3. A method according to claim 1, wherein the using the intra-language statistical translation model comprises:

receiving from the search engine a listing that the search engine matched to the user'"'"'s query;

using the model to find a probability that the listing is a translation of the user'"'"'s query; and

determining whether the listing matches the user'"'"'s query based on the probability.

4. A method according to claim 1, wherein the computing an alignment comprises rearranging either the current source words, the target words, or both, so that semantically equivalent words of the source and target phrases are aligned;

and

iteratively updating the alignment and parameters of the n-gram based intra-language statistical translation model.

5. A method according to claim 1, further comprising using the intra-language statistical translation model to generate a language model of the human language, the language model being capable of determining a likelihood of strings in the human language.

6. A method according to claim 5, wherein the using the intra-language statistical translation model to generate the language model comprises including with the language model translations from the intra-language statistical translation model.

7. A method according to claim 5, further comprising performing automatic speech recognition with the language model.

8. One or more storage devices storing information to enable a computing device to perform a process for translating phrases of a human language to other phrases of the language, the process comprising:

accessing training pairs comprising pairs of phrases in the human language, the training pairs comprising text queries in the human language submitted to a search engine and corresponding text sentences in the human language returned by the search engine when the text queries were submitted to the search engine, the training pairs formed by comparing the text queries and the text sentences to identify text queries that are similar to text sentences and pairing the identified text queries with the respective identified text sentences;

training a statistical machine translation model with the training pairs by computing respective alignments of the training pairs, an alignment mapping words of a phrase with the words the phrase is paired to by inserting null words into the phrase or reordering words of the phrase, the statistical machine translation model being capable of computing probabilities that a target string in the human language is a valid translation of a given source string in the human language;

receiving a text phrase in the human language, decoding the text phrase to different candidate translations of the text phrase in the human language, and using the statistical machine translation model to compute probabilities that the candidate translations are translations of the text phrase; and

based on the probabilities, storing and/or displaying, by computer, one or more of the candidate translations.

9. One or more storage devices according to claim 8, wherein the training pairs comprise query strings submitted by users paired with listings that a search engine matched with the query strings.

10. One or more storage devices according to claim 8, wherein the received text phrase comprises a query string inputted by a user, the query string comprising text in the human language, and the process further comprises using the statistical machine translation model to identify a plurality of probable potential translations of the query string, the potential translations comprising text in the human language.

11. One or more storage devices according to claim 8, wherein the received text phrase comprises a name of an organization or business entity obtained from a search engine for searching listings of business/organization names, the name having been obtained from the search engine according to a user-inputted query, and wherein the process further comprises using the statistical machine translation model to determine a probability that the name is a valid translation of the query string and determining relevancy of the listing to the query based on the probability.

12. One or more storage devices according to claim 8, further comprising using the statistical machine translation model to build a statistical language model of the human language, where the statistical language model provides probabilities of phrases in the human language.

13. One or more storage devices according to claim 8, the process further comprising using the statistical machine translation model to translate text queries recognized by a speech recognizer into display forms.

14. A method performed by one or more computers comprised of one or more processors and physical storage, the method comprising:

providing a statistical machine translation model, stored in the physical storage, and configured to allow the one or more processors to compute probabilities of translations of phrases, wherein the phrases are in a human language and the translations of the phrases are in the same human language, the statistical machine translation model having been trained with training pairs, the training pairs having been computed from a list of text queries in the human language submitted to a search engine and from a list of text sentences in the human language returned by the search engine when the text queries were submitted to the search engine, and by comparing the text queries to the text sentences to identify which of the text queries are similar to which of the text sentences, where text sentences identified as similar to text queries are respectively paired to form the training pairs; and

using, by the processor, the statistical machine translation model to translate between query forms and listing forms of organizations and/or businesses, where the query forms comprise phrases, in the human language, submitted to the search engine, and where the listing forms comprise formal names, in the human language, of organizations and/or businesses searchable by the search engine.

15. A method according to claim 14, wherein the using the statistical machine translation model comprises computing similarity between a query form and a listing form.

16. A method according to claim 14, wherein, given a user query inputted by a user in the human language, given a corresponding listing in the human language that was found by the search engine, and given a set of candidate translations of the listing, the candidate translations also in the human language, the using the statistical machine translation model comprises computing probabilities of the candidate translations.

17. A method according to claim 16, further comprising generating a search result for the given user query based on the computed probabilities.

18. A method according to claim 14, further comprising generating or modifying a search result of the search engine based on probabilities computed by the statistical machine translation model, the search result corresponding to a user-inputted query form.

19. A method according to claim 18, further comprising using the probabilities to rank or eliminate search result.