Performing semantic analyses of user-generated text content using a lexicon
A method and system for performing semantic analysis for electronic communication using a lexicon is provided. A neural network model is trained with a plurality of annotated text strings, the annotations comprising characteristic tuples that indicate characteristics for the text strings. An unannotated text string is received that comprises a plurality of words from a user. A characteristic matrix for the received text string is generated using a lexicon. The determined characteristic matrix is input into the trained neural network. And a characteristic tuple that indicates a characteristic for the received text string is received as output from the trained neural network.
1. A method for performing semantic analysis for electronic communication using a lexicon, the method comprising:
training a neural network model with a plurality of annotated text strings, the annotations comprising characteristic tuples that indicate characteristics for the text strings;
receiving an unannotated text string comprising a plurality of words from a user;
generating a characteristic matrix for the received text string using a lexicon;
inputting the determined characteristic matrix to the trained neural network; and
receiving as output from the trained neural network a characteristic tuple that indicates a characteristic for the received text string.
2. The method of claim 1 , wherein the annotated text strings and unannotated text string comprise chat text strings.
3. The method of claim 2 , wherein the chat text strings are based on chats between a service provider and a user or simulated chats between a service provider and a user.
4. The method of claim 2 , wherein one of the annotated text strings comprises a plurality of characteristic tuples that correspond to predetermined affect bearing words within the one annotated text string.
5. The method of claim 4 , wherein the characteristic matrix comprises an aggregate of characteristic tuples that correspond to predetermined affect bearing words within the received text string.
6. The method of claim 5 , wherein the characteristic tuples comprise vectors that indicate numerical values for a plurality of emotion categories for the affect bearing word associated with each characteristic tuple.
7. The method of claim 6 , wherein determining a characteristic matrix for the received text string comprises:
identifying the affect bearing words within the received text string;
generating, for each identified affect bearing words, a characteristic tuple by determining numeric values for the emotion categories for the associated affect bearing words based on the lexicon, wherein the lexicon comprises an emotion representative dictionary; and
aggregating the generated characteristic tuples to generate the characteristic matrix.
8. The method of claim 7 , wherein the numeric values for the emotion categories for the associated affect bearing words comprise an average of a pointwise mutual information value for the affect bearing words relative to the words in the emotion representative dictionary.
9. The method of claim 7 , further comprising building the emotion representative dictionary based on the annotated text strings, characteristic tuples corresponding to the annotated text strings, and a plurality of predetermined unannotated text strings.
10. The method of claim 9 , wherein building the emotion representative dictionary comprises:
identifying seed words that comprise numerical values for the emotion categories from the annotated text strings;
identifying affect bearing words in the predetermined text strings; and
computing, for each identified affecting bearing word in the predetermined text strings, numerical values for the emotion categories based on comparisons to the seed words; and
storing the seeds words and identified affect bearings words along with the corresponding numerical values for the emotion categories for these words.
11. The method of claim 10 , wherein identifying the affect bearing words of the predetermined text strings further comprises identifying nouns, adjectives, verbs, and adverbs as the affecting bearing words.
12. The method of claim 11 , wherein comparing the affect bearing words to the seed words comprises:
projecting the identified affect bearing words and the seed words in a same dimension; and
computing a distance between the identified affect bearing words and the seed words in the projected dimension.
13. The method of claim 6 , wherein the emotion categories comprise happiness, assurance, agreement, courteousness, apologetic, unhappiness, disagreement, and no emotion.
14. A system for performing semantic analysis for electronic communication using a lexicon, the system comprising:
a neural network model comprising a processor that is trained with a plurality of annotated text strings, the annotations comprising characteristic tuples that indicate characteristics for the text strings; and
a content analysis module comprising a processor that is configured to:
receive an unannotated text string comprising a plurality of words from a user;
generate a characteristic matrix for the received text string using a lexicon; and
input the determined characteristic matrix to the trained neural network, wherein a characteristic tuple that indicates a characteristic for the received text string is received as output from the trained neural network.
15. The system of claim 14 , wherein the annotated text strings and unannotated text string comprise chat text strings.
16. The system of claim 15 , wherein the characteristic matrix comprises an aggregate of characteristic tuples that correspond to predetermined affect bearing words within the received text string.
17. The system of claim 16 , wherein the characteristic tuples comprise vectors that indicate numerical values for a plurality of emotion categories for the affect bearing word associated with each characteristic tuple.
18. The system of claim 17 , wherein determining a characteristic matrix for the received text string comprises:
identifying the affect bearing words within the received text string;
generating, for each identified affect bearing words, a characteristic tuple by determining numeric values for the emotion categories for the associated affect bearing words based on the lexicon, wherein the lexicon comprises an emotion representative dictionary; and
aggregating the generated characteristic tuples to generate the characteristic matrix.
19. The system of claim 18 , wherein the content analysis module is further configured to build the emotion representative dictionary based on the annotated text strings, characteristic tuples corresponding to the annotated text strings, and a plurality of predetermined unannotated text strings, the building further comprising:
identifying seed words that comprise numerical values for the emotion categories from the annotated text strings;
identifying affect bearing words in the predetermined text strings;
computing, for each identified affecting bearing word in the predetermined text strings, numerical values for the emotion categories based on comparisons to the seed words; and
storing the seeds words and identified affect bearings words along with the corresponding numerical values for the emotion categories for these words to build the emotion representative dictionary.
20. The system of claim 19 , wherein comparing the affect bearing words to the generated seed words comprises:
projecting the identified affect bearing words and the seed words in a same dimension; and
computing a distance between the identified affect bearing words and the seed words in the projected dimension.