SPELL CORRECTION WITH HIDDEN MARKOV MODELS ON ONLINE SOCIAL NETWORKS

• US 20160299883A1
• Filed: 04/10/2015
• Published: 10/13/2016
• Est. Priority Date: 04/10/2015
• Status: Active Grant

• Current Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Original Assignee: Meta Platforms, Inc. (f/k/a Facebook, Inc.)
• Inventors: Zhu, Hongcheng, Bernhardt, Daniel
• Granted Patent: US 10,049,099 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/3322 using system suggestions G06F16/3325 takes precedence
• CPC Class Codes: G06F 40/232 Orthographic correction, e.g. spell checking or vowelisation
• CPC Class Codes: G06Q 50/01 Social networking
• CPC Class Codes: H04L 67/535 Tracking the activity of the user network monitoring arrangements H04L43/00; recording of computer activity G06F11/34

Abstract

In one embodiment, a method includes receiving a search query including one or more n-grams, where the n-grams include one or more misspelled n-grams, identifying one or more variant-tokens for each misspelled n-gram, calculating a feature value for each identified variant-token based at least on the identified variant-token, the misspelled n-gram, and one or more variant-tokens corresponding to one or more n-grams preceding the misspelled n-gram, generating one or more unique combinations of the n-grams and variant-tokens, calculating a sequence-score for each unique combination based at least in part on the calculated feature values of the variant-tokens of the unique combination, generating one or more corrected queries, where each corrected query includes a unique combination having a sequence-score greater than a threshold sequence-score, and sending one or more of the corrected queries to a user for display.

Claims

1. A method comprising, by one or more computing devices:

receiving, from a client system of a user of an online social network, a search query comprising one or more n-grams, wherein the n-grams comprise one or more misspelled n-grams;

identifying, for each misspelled n-gram, one or more variant-tokens;

calculating, for each identified variant-token of a misspelled n-gram, a feature value based at least on the identified variant-token, the misspelled n-gram, and one or more variant-tokens corresponding to one or more n-grams preceding the misspelled n-gram;

generating one or more unique combinations of the n-grams and variant-tokens, wherein each unique combination comprises a variant-token corresponding to each misspelled n-gram;

calculating a sequence-score for each unique combination based at least in part on the calculated feature values of the variant-tokens of the unique combination; and

generating one or more corrected queries, each corrected query comprising a unique combination having a sequence-score greater than a threshold sequence-score; and

sending, to the client system of the user for display in response to receiving the search query, one or more of the corrected queries.

2. The method of claim 1, further comprising:

receiving from the user a selection of one of the corrected queries;

identifying one or more objects matching the selected query; and

sending, Co the client system of the user, a search-result page responsive to the selected query, the search-results page comprising one or more references to one or more of the identified objects, respectively.

3. The method of claim 1, wherein the calculated feature value for each identified variant-token of a misspelled n-gram comprises a transitional logarithmic probability (p) that the identified variant-token corresponds to the corrected misspelled n-gram, wherein:

p=P (x|y, z), wherein;

x=the identified variant-token corresponding to the corrected misspelled n-gram;

y=the misspelled n-gram entered by the user; and

z=a variant-token corresponding to the n-gram immediately preceding the misspelled n-gram.

4. The method of claim 3, wherein the transitional logarithm probability is determined based at least on a Perceptron training algorithm.

5. The method of claim 3, wherein the transitional logarithm probability is calculated based at least on a contextual speller model.

6. The method of claim 1, wherein the calculated feature value for each identified variant-token of a misspelled n-gram is based at least on one or more language models.

7. The method of claim 1, wherein if there are no n-grams preceding the misspelled n-gram, then the variant-token corresponding to an n-gram preceding the misspelled n-gram is a pre-determined placeholder.

8. The method of claim 1, wherein the calculated sequence-score for each unique combination is based at least in part on a calculated sum of the calculated feature values of the variant-tokens of the unique combination.

9. The method of claim 1, wherein calculating the sequence-score for each unique combination based at least in part on the calculated feature values of the variant-tokens of the unique combination comprises:

weighing each of the calculated feature values of the variant-tokens of the unique combination by a pre-determined scalar value; and

calculating the sequence-score of the unique combination based at least on the weighted calculated feature values of the variant-tokens of the unique combination.

10. The method of claim 9, wherein the pre-determined scalar value is determined based at least on one or more learning algorithms.

11. The method of claim 10, wherein one of the learning algorithms is a Perceptron training algorithm.

12. The method of claim 1, calculating the sequence-score for each unique combination based at least in part on the calculated feature values of the variant-tokens of the unique combination comprises:

determining a feature vector of the unique combination based at least on the calculated feature values of the variant-tokens of the unique combination;

determining a weighted vector based at least on the pre-determined scalar values of the calculated feature values of the variant-tokens of the unique combination; and

calculating the sequence-score for the unique combination based at least on a dot product of the feature vector and the weighted vector of the unique combination.

13. The method of claim 1, wherein one or more of the identified variant-tokens for each misspelled n-gram are uni-grams.

14. The method of claim 1, wherein an identified variant-token for a misspelled n-gram comprises at least two uni-grams.

15. The method of claim 14, wherein calculating a feature value for the identified variant-token for the misspelled n-gram comprising at least two uni-grams comprises:

calculating, for each uni-gram of the identified variant-token, a feature value based at least on the uni-gram, the misspelled n-gram, and a variant-token corresponding to an n-gram substantially preceding the misspelled n-gram; and

determining the feature value for the identified variant-token based at least on an average of the calculated feature values of the at least two uni-grams of the variant-token.

16. The method of claim 1, wherein generating the unique combinations of the n-grams and variant-tokens comprises:

utilizing Viterbi algorithm to generate the unique combinations of the n-grams and variant-tokens.

17. The method of claim 16, wherein utilizing Viterbi algorithm reduces a time for generating the unique combinations of the n-grams and variant-tokens.

18. One or more computer-readable non-transitory storage media embodying software that is operable when executed to:

receive, from a client system of a user of an online social network, a search query comprising one or more n-grams, wherein the n-grams comprise one or more misspelled n-grams;

identify, for each misspelled n-gram, one or more variant-tokens;

calculate, for each identified variant-token of a misspelled n-gram, a feature value based at least on the identified variant-token, the misspelled n-gram, and one or more variant-tokens corresponding to one or more n-grams preceding the misspelled n-gram;

generate one or more unique combinations of the n-grams and variant-tokens, wherein each unique combination comprises a variant-token corresponding to each misspelled n-gram;

calculate a sequence-score for each unique combination based at least in part on the calculated feature values of the variant-tokens of the unique combination; and

generate one or more corrected queries, each corrected query comprising a unique combination having a sequence-score greater than a threshold sequence-score; and

send, to the client system of the user for display in response to receiving the search query, one or more of the corrected queries.

19. A system comprising:

one or more processors; and

a memory coupled to the processors comprising instructions executable by the processors, the processors operable when executing the instructions to;

receive, from a client system of a user of an online social network, a search query comprising one or more n-grams, wherein the n-grams comprise one or more misspelled n-grams;

identify, for each misspelled n-gram, one or more variant-tokens;

calculate, for each identified variant-token of a misspelled n-gram, a feature value based at least on the identified variant-token, the misspelled n-gram, and one or more variant-tokens corresponding to one or more n-grams preceding the misspelled n-gram;

generate one or more unique combinations of the n-grams and variant-tokens, wherein each unique combination comprises a variant-token corresponding to each misspelled n-gram;

calculate a sequence-score for each unique combination based at least in part on the calculated feature values of the variant-tokens of the unique combination; and

generate one or more corrected queries, each corrected query comprising a unique combination having a sequence-score greater than a threshold sequence-score; and

send, to the client system of the user for display in response to receiving the search query, one or more of the corrected queries.