“LM-release-v0.2”版本间的差异

2017年1月5日 (四) 06:27的版本

RELEASE TITLE: LM RELEASE
RELEASE VERSION: v0.2
RELEASE TYPE: STEP RELEASE
RELEASE LOCATION: /work5/release/weiy/project/myhexin/lm/v0.2
RELATED BUGDB: 11

1. BACKGROUND:


2. TECHNOLOGY SUMMARY:

利用40G金融语料和64G通用语料，以3G的大小进行分割，用cleaning v0.1进行清洗，以vocab v0.2为词表，用
MultilingualSegmenter.jar进行分词，分别训练，得到36个LM，按照1:1进行插值合并，按照不同剪枝率进行剪枝，
得到2组LM（其中通用领域LM由于体积过大，需先将22个LM合并成2个，再按5gram_1e-9进行剪枝，再将2个LM合并，
再剪枝）。再将2组LM按照1:1进行合并，得到一组混合LM。

3. RELEASE COMPONENT:

LM:  LM RELEASE v0.2

4. TEST RESULT:

==============================================================================
|    wer / ppl      | test_myhexin_20161019 |  test_1000ju  |  test_2000ju  |
==============================================================================
| fin_3gram_1e-7    |      6.33/803.275     | 36.76/1623.97 | 45.17/1687.52 |
| uni_3gram_1e-7    |      8.05/1511.17     | 27.78/379.833 | 38.08/454.279 |
| hybrid_3gram_1e-7 |      6.45/799.42      | 28.89/503.623 | 39.66/590.483 |
| given_finance     |      6.47/1104.9      | 37.51/2362.24 | 45.77/2474.66 |
| given_universe    |     12.23/4277.15     | 20.79/319.08  | 31.77/351.087 |
==============================================================================

Note:
Beam=13
max_active=7000
fin_3gram_1e-7为40G金融语料训练出的LM
uni_3gram_1e-7为64G通用语料训练出的LM
hybrid_3gram_1e-7为混合LM
given_finance为2.9G金融语料训练出的LM
given_universe为海量通用语料训练出的Base LM
发音词典为lexicon v0.2
AM为tonghuashun v0.1 am

5. RELEASE TEAM:

Author: 魏扬
Contributor: 白子薇
Monitor: 赵梦原