NLP Word Representation

Machine Translation

A knowledge graph for Machine Translation

Posted by Wanxin on Monday, February 15, 2021

Machine Translation (MT)
- Pre- Neural [1950]: rule-based/dictionary based
- statistical [1990s]
  - bayes ruls, learn alignment
- Decoding: search for the best translation
- Nerual Machine Translation [2014]
  - seq2seq
    - two RNNs involved
    - encoding
    - decoding
      - greedy decoding
      - exhaustive search decoding
      - beam search
        
        k host, k^2, k
  - Evaluation
    - BLEU BiLingual Evaluation Understudy
- Attention & seq2seq [2019]
  - core idea on each step of the decoder, use direct connection(dot product) to the encoder to focus on a particular part of the source sentence
  - it’s a general DL technique
  - attention variants
  - concept: context vector = attention output: attention weighted hidden state

Machine Translation (MT)
- Pre- Neural [1950]:
  - rule-based/dictionary based
- statistical [1990s]
  - bayes ruls, learn alignment
- - Decoding:
  - search for the best translation
- Nerual Machine Translation [2014]
  - seq2seq
    - two RNNs involved
    - encoding
    - decoding
      - greedy decoding
      - exhaustive search decoding
      - beam search
        
        k host, k^2, k
  - Evaluation
    - BLEU
      - BiLingual Evaluation Understudy
- Attention & seq2seq [2019]
  - core idea on each step fo the decoderm use direct connection(dot product) to the encoder to focus on a particular part of the source sentence
  - it’s a general DL technique
  - attention variants