Word detection

• US 8,463,598 B2
• Filed: 01/28/2011
• Issued: 06/11/2013
• Est. Priority Date: 08/23/2007
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google Inc. (Alphabet Inc.)
• Inventors: Wu, Jun, Liu, Tang Xi, Hong, Feng, Wang, YongGang, Yang, Bo, Zhang, Lei
• Primary Examiner(s): Shah, Paras D
• Application Number:
  US13/016,338
• Publication Number:
  US 20110137642A1
• Time in Patent Office: 865 Days
• Field of Search: None
• US Class Current: 704/10
• CPC Class Codes: G06F 40/129 Handling non-Latin characters, e.g. kana-to-kanji conversion
• CPC Class Codes: G06F 40/216 using statistical methods
• CPC Class Codes: G06F 40/242 Dictionaries
• CPC Class Codes: G06F 40/53 Processing of non-Latin text kana-to-kanji conversion G06F40/129; vowelisation G06F40/232

Abstract

Methods, systems, and apparatus, including computer program products, in which data from web documents are partitioned into a training corpus and a development corpus are provided. First word probabilities for words are determined for the training corpus, and second word probabilities for the words are determined for the development corpus. Uncertainty values based on the word probabilities for the training corpus and the development corpus are compared, and new words are identified based on the comparison.

Claims

1. A computer-implemented method, comprising:

determining, by one or more computers, first word frequencies for existing words and a candidate word in a training corpus, each of the existing words and the candidate word being one or more characters, the candidate word defined by a sequence of characters, wherein the sequence of characters define constituent words that are each an existing word in a dictionary;

determining, by the one or more computers, second word frequencies for the constituent words and the candidate word in a development corpus;

determining, by the one or more computers, a candidate word entropy-related measure based on the second word frequency of the candidate word and the first word frequencies of the constituent words and the candidate word;

determining, by the one or more computers, an existing word entropy-related measure based on the second word frequencies of the constituent words and the first word frequencies of the constituent words and the candidate word; and

determining, by the one or more computers, that the candidate word is a new word when the candidate word entropy-related measure exceeds the existing word entropy-related measure.

2. The method of claim 1, wherein the training corpus and the development corpus comprise web documents.

3. The method of claim 1, wherein:

determining first word frequencies comprises training a language model for probabilities of the existing words and the candidate words in the training corpus; and

wherein determining second word frequencies comprises determining a word count value for each of the constituent words and the candidate word in the development corpus.

4. The method of claim 3, wherein:

determining a candidate word entropy-related measure comprises;

determining a first logarithmic value based on the probabilities of the candidate word and the constituent words; and

determining the candidate word entropy-related measure based on the word count value of the candidate word and the first logarithmic value; and

determining an existing word entropy-related measure comprises;

determining second logarithmic values based on the probabilities of the candidate word and the constituent words; and

determining the existing word entropy-related measure based on the word counts of the constituent words and the second logarithmic values.

5. The method of claim 1, wherein the words each comprise one or more Hanzi characters.

6. A computer-implemented method, comprising:

determining, by one or more computers, first word probabilities for existing words and a candidate word in a first corpus, each of the existing words and the candidate word being one or more characters, the candidate word defined by a sequence of characters, wherein the sequence characters define constituent words that are each an existing word in a dictionary;

determining, by the one or more computers, second word probabilities for the constituent words and the candidate word in the a second corpus;

determining, by the one or more computers, a first entropy-related value based on the second candidate word probability and the first word probabilities of the candidate word and the constituent words;

determining, by the one or more computers, a second entropy-related value based on the second constituent word probabilities and the first word probabilities of the candidate word and the constituent words; and

determining, by the one or more computers, that the candidate word is a new word when the first entropy-related value exceeds the second entropy-related value.

7. The method of claim 6, wherein:

determining first word probabilities comprises training a language model on the first corpus for word probabilities of the existing words and the candidate word in the first corpus; and

determining second word probabilities comprises determining word count values for each of the constituent words and candidate word.

8. The method of claim 7, wherein:

determining a first entropy-related value comprises;

determining a first logarithmic value based on the first word probabilities of the candidate word and the constituent words; and

determining the first entropy-related value based on the word count value of the candidate word and the first logarithmic value; and

determining the second entropy-related value comprises;

determining second logarithmic values based on the first word probabilities of the candidate word and the constituent words; and

determining the second entropy-related value based on the word counts of the constituent words and the second logarithmic values.

9. The method of claim 8, wherein the words each comprise one or more Hanzi characters.

10. A computer-implemented method, comprising:

partitioning, by the one or more computers, a collection of web documents into a training corpus and a development corpus;

training, by the one or more computers, a language model on the training corpus for first word probabilities of words in the training corpus, wherein the words in the training corpus include a candidate word defined by a sequence of two or more characters, the two or more characters defining two or more corresponding words in the training corpus, the two or more corresponding words existing words in a dictionary;

counting, by the one or more computers, occurrences of the candidate word and the two or more corresponding words in the development corpus;

determining, by the one or more computers, a first value based on the occurrences of the candidate word in the development corpus and the first word probabilities;

determining, by the one or more computers, a second valued based on the occurrences of the two or more corresponding words in the development corpus and the first word probabilities; and

comparing, by the one or more computers, the first value to the second value; and

determining, by the one or more computers, whether the candidate word is a new word based on the comparison.

11. The method of claim 10, wherein training a language model on the training corpus for first word probabilities of words in the training corpus comprises training an n-gram language model.

12. The method of claim 11, wherein:

determining a first value based on the occurrences of the candidate word in the development corpus and the first word probabilities corpus comprises;

determining a first logarithmic value based on the first word probability for the candidate word and the first word probabilities of the two or more corresponding words; and

multiplying the first logarithmic value by the counted occurrences of the candidate word; and

wherein determining a second valued based on the occurrences of the two or more corresponding words in the development corpus and the first word probabilities comprises;

determining second logarithmic values based on the first word probability of the candidate word and the first word probabilities of the two or more corresponding words; and

multiplying the second logarithmic values by the counted occurrences of the two or more corresponding words.

13. The method of claim 12, wherein the words each comprise one or more Hanzi characters.

14. An apparatus comprising software stored in a computer readable medium, the software comprising computer readable instructions executable by a computer processing device and that upon such execution cause the computer processing device to:

determine first word frequencies for existing words and a candidate word in a training corpus, each of the existing words and the candidate word being one or more characters, the candidate word defined by a sequence of characters, wherein the sequence of characters define constituent words that are each an existing word in a dictionary;

determine second word frequencies for the constituent words and the candidate word in a development corpus;

determine a candidate word entropy-related measure based on the second word frequency of the candidate word and the first word frequencies of the constituent words and the candidate word;

determine an existing word entropy-related measure based on the second word frequencies of the constituent words and the first word frequencies of the constituent words and the candidate word; and

determine that the candidate word is a new word when the candidate word entropy-related measure exceeds the existing word entropy-related measure.

15. An apparatus comprising software stored in a computer readable medium, the software comprising computer readable instructions executable by a computer processing device and that upon such execution cause the computer processing device to:

determine first word probabilities for existing words and a candidate word in a first corpus, each of the existing words and the candidate word being one or more characters, the candidate word defined by a sequence of characters, wherein the sequence characters define constituent words that are each an existing word in a dictionary;

determine second word probabilities for the constituent words and the candidate word in the a second corpus;

determine a first entropy-related value based on the second candidate word probability and the first word probabilities of the candidate word and the constituent words;

determine a second entropy-related value based on the second constituent word probabilities and the first word probabilities of the candidate word and the constituent words; and

determine that the candidate word is a new word when the first entropy-related value exceeds the second entropy-related value.

16. An apparatus comprising software stored in a computer readable medium, the software comprising computer readable instructions executable by a computer processing device and that upon such execution cause the computer processing device to:

partition a collection of web documents into a training corpus and a development corpus;

train a language model on the training corpus for first word probabilities of words in the training corpus, wherein the words in the training corpus include a candidate word defined by a sequence of two or more characters, the two or more characters defining two or more corresponding words in the training corpus, the two or more corresponding words existing words in a dictionary;

count occurrences of the candidate word and the two or more corresponding words in the development corpus;

determine a first value based on the occurrences of the candidate word in the development corpus and the first word probabilities;

determine a second valued based on the occurrences of the two or more corresponding words in the development corpus and the first word probabilities; and

compare the first value to the second value; and

determine whether the candidate word is a new word based on the comparison.