Similarity Search

[captain-pool]

Overview:

• Split the sentence to words.
• return URL for expression.

Tasks: (similarity metric -> 0 for perfect match)

• Precompute maximum length of phrase.
• greedily search for word-combination (max: max-phrase-length) having minimum distance from the centroid.
• each time vectorize the combination.
• find the closest semantically closest word of the new formed phrase / word.
• For proper nouns, check if present in dictionary. else, fingerspell.

NOTE: Don't count the stop words.

API Design Overview:

• class SimSearch: properties-> vectorizer, clusterDB, dictVectDB, MAX_LENGTH_PHRASE (from config)

• simSearch.splitTokens(sentence): iterates over each consecutive word combination and keeps track of minimum distance from the center. stop at MAX_PHRASE_LENGTH / ProperNoun reached, which ever earlier.

• simSearch.tagPOS(word): Helper function for pos tagging.

• simSearch.queryVideoClip(token): Query DB to obtain video clip from token.

• simSearch.fingerspell(character_token): helper function to fingerspell mp4 of the character_token.

• simSearch.query(setence): Stitch everything together and return list of video links. (in-order)

NOTE: Can be made faster by creating parallel jobs to work on splices of length MAX_LENGTH_PHRASE. if stopped before MAX_LENGTH_PHRASE, can be kept in a Queue with a position index in the token list.