Language Identification and Modelling in Specialized hardware

1. Neural Networks
2. Parsing

1. Fast
2. Customizable

1. Encoding Grammars

1. Outputs constant to CPU
2. Use stack
   1. Push
   2. Pop
3. Output matched span
4. Halt

1. Naive Bayes model
2. Feature strings are converted to literal regular expressions
   1. Collect feature counts
3. Emulate automata on CPU

1. Using Telescoping series
   1. A telescoping series is any series where nearly every term cancels with a preceeding or following term. For instance, the series

      Annotations:

      • https://en.wikipedia.org/wiki/Telescoping_series
      • http://math.oregonstate.edu/home/programs/undergrad/CalculusQuestStudyGuides/SandS/SeriesTests/telescoping.html
   2. Collapse prob and back off into single function
      1. Preserves sentence-level probabilities
      2. Sends just one value Q per token and not probability and back offs
         1. Saves CPU workload and communications
2. Simplified query
   1. For each word match as much context as possible
      1. Sends just one value Q per token and not probability and back offs

1. Match as much leading context as possible
   1. Scanning until a match is found
      1. Report the longest match
         1. Resume scanning
         2. The longest matching N-gram will be reported
2. Use "greedy" regular expressions

1. More details in the paper
2. tested against some existing models
   1. CLD2
      1. C++
   2. Original
      1. Python
      2. Java

1. KenLM
2. DALM

Annotations:

• https://en.wikipedia.org/wiki/Part-of-speech_tagging