Emergence Group - April 6, 2005

  1. Caveats

    • Emergence?


  2. Computational Linguistics

    • Machine Translation: "My name is Rich" &rarr "Je m'appelle Rich"

    • Speech Recognition: a machine learns to wreck a nice beach


  3. Research interests

    • Computational morpholgy

    • Word-sense disambiguation and translational word-choice

    Solve for arbitrary language pairs, with little or no supervision

  4. Stepping back to Doug's presentation

    • Elman "Learning and development in neural networks: the importance of starting small"

    • SRN given randomly generated sentences according to a simple grammar:
      boys who chase dogs see girls .
      girl who boys who feed cats walk .
      cats chase dogs .
      mary feeds john .
      dogs see boys who cats who marry feeds chase .
      

    • Subject/verb agreement requires long-distance dependency:
      dogs see boys who cats who marry feeds chase .
      

    • Experiment: Generate 10K sentences (number used by Elman) according to grammar. Can interesting insights be discovered?



    • Is this emergence?

  5. A different/bigger/real-data experiment

    • Start with the British National Corpus (115 million words / 6 million sentences / publications since 1960 )

    • Cluster all the verbs by the three following words.

    • What do we expect? All the verbs that take direct objects together?

    • Clustered results: rbtree30.pdf & rbtree100.pdf

    • Note that we could have also clustered on previous there words (results not included).

    • Is this emergence?

    • A note about memory requirements.

      • Storing the co-occurence matrix above requires storing (# verbs) * (# words) integers.

      • In this case, 3470 * 330328 = 1,146,238,160, or about 4.5G of memory.

      • (Though this data is very sparse and only requires about 30M in practice.)

      • The SRN we've seen (input layer = 3470, waist layer = 10, hidden layer = 70, waist layer = 10, output layer = 3470) requires storage of 84,286,300,000 floating point values, or about 337G of memory.

      • And that's not even mentioning anything about the how we will input the non-verb words which could require something like 768,440,705,088,000 floating point values. Do you have 30,736G of memory?

      • We decided the above two arguments were not correct.

      • (Jason Eisner's LSA slides) We didn't talk about this.


  6. What other properties to morphological variants share?

    • Forms are used in a consistent ratio relative to other variants of the same form.

    • All this and we haven't even looked at spelling!

    • some slides if you're still awake


  7. What other fun tricks do I do with these types of techniques?

    • Word-sense disambiguation and translational selection

    • Sentiment classification


Richard Wicentowski