Rationale

• Need for higher-level of (cultural) understanding

Multi-layered Learning (Supervised Learning using SVMs)

• Emotions

• Category-matching (?)

• Possible Classifications
• By Time (either by season or by daily time)
  
• By the protagonist is the main character, who is not necessarily a hero or a heroine. (Male Adult, Female Adult, Male Child, Female Child, Domestic Animal, Wild Animal, Plant, Group of people, Inanimate object etc). We could also have dynamic generation of classes.
• Try making sentences from the poem and then use POS tagger, WSD algos. Find subjects and objects. Use TF IDF in the poem to figure out the main character.
  
• By the speaker of the poem, who is not necessarily the author (Male Adult, Female Adult, Male Child, Female Child, Animal, Plant, Group of people, Inanimate object etc)

• Word-sense disambiguation & Topic Modelling

• Part 1 - Poetry Language Parsing for existing poem database (Done once every time the database is updated and done only for the poems that have not been parsed):
  

• Step 1: Take a poem from database

• Step 2: Use WSD tool (Sample here) by Prof Ted Pederson from University of Minnesota that is based on a machine readable synset based dictionary - WordNet provided by Princeton University. Use this on the poem in one go.
  

• Step 3: Parse output from the WSD tool to get POS & sense of each word (available in WordNet) for the poem.

• Step 4: Take the poem and its POS/senses sentence by sentence. Run a term frequency of similar nouns (subject and object) in the poem. Use TF-IDF method and for word similarity either use the current disctionary or word similarity module from Ted Pederson to guess the subject of the poem. Assign proportional chances to top 3 subjects. Store that in the database along with their supersense (like school belongs to the supersense location. According to MIT Graduate Notes click here from Approaches to Natural Language Processing there are 28 supersets).
  
• Step 5,6,7: Similarly do a test on verbs, adverbs and adjectives. Assign top v,adv, adj for each of top 3 subjects with a supersense.
• Step 8: Also store subject, object, verb, adjective, adverb combinations for each sentence in the database. (maybe we can use a tree structure like in MIT notes)
  
• Step 9: Simultaneously add similes found to another database (could be useful for future use) and acknowledge them as adjectives.
  
• Step 10: Repeat steps 1-9
•  Part 2 - SMS Parsing:\
• Step 1-3: Do steps 1-3 from Part 1.
• Step 4: Like in Step 8 in Part 1, map the line structure in the form of its POS.
• Step 5: Assign supersense for each noun.
• Step 6: Go look in the poems database for supersense matched items.
• Step 7: Use word similarity module between each of the nouns 
  
•  

Training:

A site, Degree of belief, Weightage

Further Implementations

• Price & Penalty Learning (Reinforcement Learning)

• Degree of belief
  

Other To-Dos:

• Output comparison between old and new systems

• User studies

Things we are working on:

• SVM Integration - Literature (Seb)
  

• WSD & Topic Modelling Classifcation Research (Vid)