Round-trip translation for automated grammatical error correction

• US 9,342,499 B2
• Filed: 03/19/2014
• Issued: 05/17/2016
• Est. Priority Date: 03/19/2013
• Status: Active Grant

• Current Assignee: Educational Testing Service
• Original Assignee: Educational Testing Service
• Inventors: Madnani, Nitin, Tetreault, Joel, Chodorow, Martin
• Primary Examiner(s): Santiago Cordero, Marivelisse
• Assistant Examiner(s): BRINICH, STEPHEN M
• Application Number:
  US14/219,671
• Publication Number:
  US 20140288915A1
• Time in Patent Office: 790 Days
• Field of Search: 704 1- 10
• US Class Current: 1/1
• CPC Class Codes: G06F 40/253 Grammatical analysis; Style critique
• CPC Class Codes: G06F 40/58 Use of machine translation, e.g. for multi-lingual retrieval, for server-side translation for client devices or for real-time translation

Abstract

Systems and methods are provided for correcting a grammatical error in a text sequence. A first text sequence in a first language is received. The first text sequence is translated to a second language to provide a first translated text. The first text sequence is translated to a third language to provide a second translated text. The third language is different from the second language. The first translated text is translated to the first language to provide a first back translation. The second translated text is translated to the first language to provide a second back translation. A plurality of candidate text sequences that include features of the first back translation and the second back translation are determined. The plurality of candidate text sequences include alternative grammatical options for the first text sequence. The plurality of candidate text sequences are scored with the processing system.

Claims

1. A computer-implemented method for correcting a grammatical error in a text sequence, the computer-implemented method comprising:

translating a first text sequence in a first language to a second language with a processing system using a first machine translation algorithm to provide a first translated text;

translating the first text sequence to a third language with the processing system using a second machine translation algorithm to provide a second translated text, the third language being different from the second language;

translating the first translated text back to the first language with the processing system using a third machine translation algorithm to provide a first back translation, the translating of the first translated text back to the first language being performed without reference to the first text sequence;

translating the second translated text back to the first language with the processing system using a fourth machine translation algorithm to provide a second back translation, the translating of the second translated text back to the first language being performed without reference to the first text sequence;

determining a plurality of candidate text sequences that include features of the first back translation and the second back translation with the processing system, the plurality of candidate text sequences including alternative grammatical options for the first text sequence; and

scoring each of the plurality of candidate text sequences using a computer based scoring model with the processing system, the scoring providing a measure of a grammaticality of each of the candidate text sequences, wherein a computer-readable memory is encoded with instructions for commanding the processing system to execute the first, second, third, and fourth machine translation algorithms and to execute the determining of the plurality of candidate text sequences and the scoring of each of the plurality of candidate text sequences.

2. The computer-implemented method of claim 1, wherein the plurality of candidate text sequences include at least one candidate text sequence that is equivalent to neither the first back translation nor the second back translation.

3. The computer-implemented method of claim 1, wherein the determining of the plurality of candidate text sequences includes:

for each unit of text in the first text sequence,

i) determining a first corresponding unit of text in the first back translation, the determining of the first corresponding unit of text comprising a first alignment, and

ii) determining a second corresponding unit of text in the second back translation, the determining of the second corresponding unit of text comprising a second alignment.

4. The computer-implemented method of claim 3, wherein the determining of the plurality of candidate text sequences includes:

constructing a lattice that includes the features of the first back translation and the second back translation, the constructing of the lattice including;

defining a first path of the lattice based on the first text sequence, the first path including nodes for units of text of the first text sequence, and

adding a node to the lattice, wherein the added node provides an alternative unit of text for a corresponding node of the first path, the alternative unit of text being included in the first back translation or the second back translation, and the corresponding node of the first path being determined based on the first alignment or the second alignment; and

determining a plurality of paths through the lattice to determine the plurality of candidate text sequences.

5. The computer-implemented method of claim 4, wherein the determining of the plurality of paths through the lattice includes:

determining a first path through the lattice by selecting at each node of the lattice an outgoing edge with a largest weight;

determining a second path through the lattice that is a shortest path through the lattice; and

determining a third path through the lattice based on a fluency measure, the fluency measure indicating a fluency of a candidate text sequence that corresponds to the third path.

6. The computer-implemented method of claim 1, wherein the candidate text sequences are scored by applying the computer based scoring model to the candidate text sequences, the computer based scoring model being trained based on a plurality of text sequences that are scored by one or more humans or annotated by the one or more humans.

7. The computer-implemented method of claim 1,

wherein the first machine translation algorithm utilizes a first rule set or statistical model in translating the first text sequence from the first language to the second language;

wherein the second machine translation algorithm utilizes a second rule set or statistical model in translating the first text sequence from the first language to the third language;

wherein the third machine translation algorithm utilizes a third rule set or statistical model in translating the first translated text back to the first language; and

wherein the fourth machine translation algorithm utilizes a fourth rule set or statistical model in translating the second translated text back to the first language.

8. A system for correcting a grammatical error in a text sequence, the computer-implemented method comprising:

a processing system; and

computer-readable memory in communication with the processing system encoded with instructions for commanding the processing system to execute steps comprising;

receiving a first text sequence in a first language;

translating the first text sequence to a second language with the processing system using a first machine translation algorithm to provide a first translated text;

translating the first text sequence to a third language with the processing system using a second machine translation algorithm to provide a second translated text, the third language being different from the second language;

translating the first translated text back to the first language with the processing system using a third machine translation algorithm to provide a first back translation, the translating of the first translated text back to the first language being performed without reference to the first text sequence;

translating the second translated text back to the first language with the processing system using a fourth machine translation algorithm to provide a second back translation, the translating of the second translated text back to the first language being performed without reference to the first text sequence;

determining a plurality of candidate text sequences that include features of the first back translation and the second back translation with the processing system, the plurality of candidate text sequences including alternative grammatical options for the first text sequence; and

scoring each of the plurality of candidate text sequences using a computer based scoring model with the processing system, the scoring providing a measure of a grammaticality of each of the candidate text sequences.

9. The system of claim 8, wherein the plurality of candidate text sequences include at least one candidate text sequence that is equivalent to neither the first back translation nor the second back translation.

10. The system of claim 8, wherein the determining of the plurality of candidate text sequences includes:

for each unit of text in the first text sequence,

i) determining a first corresponding unit of text in the first back translation, the determining of the first corresponding unit of text comprising a first alignment, and

ii) determining a second corresponding unit of text in the second back translation, the determining of the second corresponding unit of text comprising a second alignment.

11. The system of claim 10, wherein the determining of the plurality of candidate text sequences includes:

constructing a lattice that includes the features of the first back translation and the second back translation, the constructing of the lattice including;

defining a first path of the lattice based on the first text sequence, the first path including nodes for units of text of the first text sequence, and

adding a node to the lattice, wherein the added node provides an alternative unit of text for a corresponding node of the first path, the alternative unit of text being included in the first back translation or the second back translation, and the corresponding node of the first path being determined based on the first alignment or the second alignment; and

determining a plurality of paths through the lattice to determine the plurality of candidate text sequences.

12. The system of claim 11, wherein the determining of the plurality of paths through the lattice includes:

determining a first path through the lattice by selecting at each node of the lattice an outgoing edge with a largest weight;

determining a second path through the lattice that is a shortest path through the lattice; and

determining a third path through the lattice based on a fluency measure, the fluency measure indicating a fluency of a candidate text sequence that corresponds to the third path.

13. The system of claim 8, wherein the candidate text sequences are scored by applying the computer based scoring model to the candidate text sequences, the computer based scoring model being trained based on a plurality of text sequences that are scored by one or more humans or annotated by the one or more humans.

14. The system of claim 8,

wherein the first machine translation algorithm utilizes a first rule set or statistical model in translating the first text sequence from the first language to the second language;

wherein the second machine translation algorithm utilizes a second rule set or statistical model in translating the first text sequence from the first language to the third language;

wherein the third machine translation algorithm utilizes a third rule set or statistical model in translating the first translated text back to the first language; and

wherein the fourth machine translation algorithm utilizes a fourth rule set or statistical model in translating the second translated text back to the first language.

15. A non-transitory computer-readable storage medium for correcting a grammatical error in a text sequence, the computer-readable storage medium comprising computer executable instructions which, when executed, cause a processing system to execute steps comprising:

receiving a first text sequence in a first language;

translating the first text sequence to a second language with the processing system using a first machine translation algorithm to provide a first translated text;

translating the first text sequence to a third language with the processing system using a second machine translation algorithm to provide a second translated text, the third language being different from the second language;

translating the first translated text back to the first language with the processing system using a third machine translation algorithm to provide a first back translation, the translating of the first translated text back to the first language being performed without reference to the first text sequence;

translating the second translated text back to the first language with the processing system using a fourth machine translation algorithm to provide a second back translation, the translating of the second translated text back to the first language being performed without reference to the first text sequence;

determining a plurality of candidate text sequences that include features of the first back translation and the second back translation with the processing system, the plurality of candidate text sequences including alternative grammatical options for the first text sequence; and

scoring each of the plurality of candidate text sequences using a computer based scoring model with the processing system, the scoring providing a measure of a grammaticality of each of the candidate text sequences.

16. The non-transitory computer-readable storage medium of claim 15, wherein the plurality of candidate text sequences include at least one candidate text sequence that is equivalent to neither the first back translation nor the second back translation.

17. The non-transitory computer-readable storage medium of claim 15, wherein the determining of the plurality of candidate text sequences includes:

for each unit of text in the first text sequence,

i) determining a first corresponding unit of text in the first back translation, the determining of the first corresponding unit of text comprising a first alignment, and

ii) determining a second corresponding unit of text in the second back translation, the determining of the second corresponding unit of text comprising a second alignment.

18. The non-transitory computer-readable storage medium of claim 17, wherein the determining of the plurality of candidate text sequences includes:

constructing a lattice that includes the features of the first back translation and the second back translation, the constructing of the lattice including;

defining a first path of the lattice based on the first text sequence, the first path including nodes for units of text of the first text sequence, and

adding a node to the lattice, wherein the added node provides an alternative unit of text for a corresponding node of the first path, the alternative unit of text being included in the first back translation or the second back translation, and the corresponding node of the first path being determined based on the first alignment or the second alignment; and

determining a plurality of paths through the lattice to determine the plurality of candidate text sequences.

19. The non-transitory computer-readable storage medium of claim 18, wherein the determining of the plurality of paths through the lattice includes:

determining a first path through the lattice by selecting at each node of the lattice an outgoing edge with a largest weight;

determining a second path through the lattice that is a shortest path through the lattice; and

determining a third path through the lattice based on a fluency measure, the fluency measure indicating a fluency of a candidate text sequence that corresponds to the third path.

20. The non-transitory computer-readable storage medium of claim 15, wherein the candidate text sequences are scored by applying the computer based scoring model to the candidate text sequences, the computer based scoring model being trained based on a plurality of text sequences that are scored by one or more humans or annotated by the one or more humans.

21. The non-transitory computer-readable storage medium of claim 15,

wherein the first machine translation algorithm utilizes a first rule set or statistical model in translating the first text sequence from the first language to the second language;

wherein the second machine translation algorithm utilizes a second rule set or statistical model in translating the first text sequence from the first language to the third language;

wherein the third machine translation algorithm utilizes a third rule set or statistical model in translating the first translated text back to the first language; and

wherein the fourth machine translation algorithm utilizes a fourth rule set or statistical model in translating the second translated text back to the first language.