Sentiment prediction from textual data

• US 8,682,649 B2
• Filed: 11/12/2009
• Issued: 03/25/2014
• Est. Priority Date: 11/12/2009
• Status: Active Grant

• Current Assignee: Apple Inc.
• Original Assignee: Apple Inc.
• Inventors: Bellegarda, Jerome
• Primary Examiner(s): Godbold, Douglas
• APPLICATION: US 20110112825A1
• Time in patent office: 1594 Days
• Field of Search: 704/1, 704/9, 704/10
• US Class Current: 704/10
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G10L 13/10 Prosody rules derived from text; Stress or intonation

Abstract

A semantically organized domain space is created from a training corpus. Affective data are mapped onto the domain space to generate affective anchors for the domain space. A sentiment associated with an input text is determined based the affective anchors. A speech output may be generated from the input text based on the determined sentiment.

Claims

1. A method to train data for sentiment prediction, comprising:

at a device comprising one or more processors;

creating from a domain corpus a domain space that is a semantic representation of one or more identified areas of information;

creating from affective data an affective space that is a semantic representation of a plurality of sentiments;

generating a plurality of first affective anchors in the affective space, each first affective anchor corresponding to a respective sentiment in the affective space;

mapping the plurality of first affective anchors generated in the affective space onto respective locations in the domain space, wherein the mapping of the first affective anchors is performed by the one or more processors;

based on the respective locations to which the plurality of first affective anchors have been mapped in the domain space, generating a plurality of second affective anchors in the domain space, each second affective anchor corresponding to a respective sentiment in the domain space; and

providing the domain space and the second affective anchors for predicting sentiment of text.

2. The method of claim 1, wherein the domain space is a vector space.

3. The method of claim 1 wherein mapping the plurality of first affective anchors in the affective space onto the respective locations in the domain space comprises:

identifying a plurality of information units associated with emotions that are present in both the domain corpus and the affective data;

determining a transformation that relates respective representations of the plurality of information units in the affective space to respective representations of the plurality of information units in the domain space; and

applying the transformation to the plurality of first affective anchors.

4. The method of claim 3 wherein the plurality of information units comprise a plurality of words.

5. The method of claim 1, wherein the affective space is a vector space.

6. The method of claim 1, wherein the domain space is created using latent semantic mapping (LSM).

7. The method of claim 1, wherein the affective space has a lower dimension than the domain space.

8. A non-transitory computer readable storage medium having instructions stored thereon that, when executed, cause a computer to:

create from a domain corpus a domain space that is a semantic representation of one or more identified areas of information;

create from affective data an affective space that is a semantic representation of a plurality of sentiments;

generate a plurality of first affective anchors in the affective space, each first affective anchor corresponding to a respective sentiment in the affective space;

map the plurality of first affective anchors generated in the affective space onto respective locations in the domain space;

based on the respective locations to which the plurality of first affective anchors have been mapped in the domain space, generating a plurality of second affective anchors in the domain space, each second affective anchor corresponding to a respective sentiment in the domain space; and

provide the domain space and the second affective anchors for predicting sentiment of text.

9. The computer-readable storage medium of claim 8, wherein the domain space is a vector space.

10. The computer readable storage medium of claim 8, wherein mapping the plurality of first affective anchors in the affective space onto the respective locations in the domain space comprises:

identifying a plurality of information units that are present in both the domain corpus and the affective data;

determining a transformation that relates respective representations of the plurality of information units in the affective space to respective representations of the plurality of information units in the domain space; and

;

applying the transformation to the plurality of first affective anchors.

11. The computer readable storage medium of claim 10 wherein the plurality of information units comprise a plurality of words.

12. The computer readable storage medium of claim 8, wherein the affective space is a vector space.

13. The computer readable storage medium of claim 8, wherein the domain space is created using latent semantic mapping (LSM).

14. The computer readable storage medium of claim 8, wherein the affective space has a lower dimension than the domain space.

15. A system for training data for sentiment prediction, comprising:

a memory to store a domain corpus associated with one or more identified areas of information, and affective data associated with a plurality of sentiments;

a processor coupled to the memory;

a latent semantic mapping (LSM) module to create from the domain corpus a domain space that is a semantic representation of the one or more identified areas of information, and to create from the affective data an affective space that is a semantic representation of the plurality of sentiments; and

a mapping module to;

generate a plurality of first affective anchors in the affective space, each first affective anchor corresponding to a respective sentiment in the affective space;

map the plurality of first affective anchors generated in the affective space onto respective locations in the domain space; and

based on the respective locations to which the plurality of first affective anchors have been mapped in the domain space, generate a plurality of second affective anchors in the domain space, each second affective anchor corresponding to a respective sentiment in the domain space.

16. The system of claim 15, wherein the LSM module and the mapping module are stored in the memory wherein the mapping module further comprises:

an embedding module to;

identify a plurality of information units that are present in both the domain corpus and the affective data;

determine a transformation that relates respective representations of the plurality of information units in the affective space to respective representations of the plurality of information units in the domain space; and

;

apply the transformation to the plurality of first affective anchors.

17. The system of claim 16, wherein the plurality of information units comprise a plurality of words.

18. The system of claim 15, wherein the domain space is a first vector space.

19. The system of claim 15, wherein the affective space has a lower dimension than the domain space.

20. A method to predict sentiment, comprising:

at a device comprising one or more processors;

receiving a first text;

determining a representation of the first text in a domain space created from a domain corpus, wherein the domain space has one or more first affective anchors representing one or more sentiment categories, and wherein the first affective anchors are generated by;

generating a plurality of second affective anchors in an affective space, mapping the plurality of second affective anchors generated in the affective space to respective locations in the domain space, and generating the first affective anchors based on the respective locations in the domain space; and

determining a sentiment associated with the first text based on the first affective anchors, wherein determining the sentiment associated with the first text is performed by the one or more processors.

21. The method of claim 20, further comprising:

determining a closeness of the representation of the first text to the one or more first affective anchors in the domain space.

22. The method of claim 20, further comprising:

generating a speech output from the first text based on the sentiment.

23. A non-transitory machine readable storage medium having instructions stored thereon that, when executed, cause a computer to perform operations comprising:

receiving a first text;

determining a representation of the first text in a domain space created from a domain corpus, wherein the domain space has one or more first affective anchors representing sentiment categories, and wherein the first affective anchors are generated by generating a plurality of second affective anchors in an affective space, mapping the plurality of second affective anchors generated in the affective space to respective locations in the domain space, and generating the first affective anchors based on the respective locations in the domain space; and

determining a sentiment associated with the first text based on the one or more first affective anchors in the domain space.

24. The machine-readable storage medium of claim 23, further comprising instructions that cause the computer to perform operations comprising

determining a closeness of the representation of the first text to the one or more first affective anchors in the domain space.

25. The computer-readable storage medium of claim 23, further comprising instructions that cause the computer to perform operations comprising

generating a speech output from the first text based on the sentiment.

26. A system to predict sentiment, comprising:

a processor;

an input device to receive a first text;

a memory coupled to the processor;

a latent semantic mapping (LSM) module, stored in the memory, configured to determine a representation of the first text in a domain space created from a domain corpus, wherein the domain space has one or more first affective anchors representing sentiment categories and wherein the first affective anchors are generated by;

generating a plurality of second affective anchors in an affective space, mapping the plurality of second affective anchors generated in the affective space to respective locations in the domain space, and generating the first affective anchors based on the respective locations in the domain space; and

a sentiment computation module configured to determine a sentiment associated with the first text based on the one or more first affective anchors.

27. The system of claim 26, wherein the LSM module is further configured to determine a closeness of the representation of the first text to the one or more first affective anchors in the domain space.

28. The system of claim 27, further comprising

an output device to generate a speech from the first text based on the sentiment, the output device coupled to the processor.