Chinese Named Entity Recognition with Multiple Features

1
Chinese Named Entity Recognition with Multiple
Features
HLT/EMNLP 2005, Vancouver, B.C., Canada, October
6-8, 2005

• Youzheng Wu, Jun Zhao, Bo Xu, Hao Yu
• Institute of Automation, Chinese Academy of
  Sciences
• Fujitsu RD Center Co., Ltd
• yzwu, jzhao,bxu_at_nlpr.ia.ac.cn
• yu_at_frdc.fujitsu.com.cn

2
Outline

• Introduction
• Related Work
• Chinese NER with Multiple Features
• The Hybrid Model (Word Model POS Model)
• Word Model (Word Entity Model Word Context
  Model)
• POS Model (POS Entity Model POS Context Model)
• Heuristic Human Knowledge
• Experiments
• Conclusion

3
Introduction

• Named Entity Recognition is one of the key
  techniques in IE, QA, Parsing, Metadata tagging,
  etc.
• The definitions of NEs by PKU is our focus.
• PN is divided into 5 sub-classes, that is,
  Chinese PN, Japanese PN, Russian PN, Euramerican
  PN, Abbreviated PN.
• LN is divided into 2 sub-classes, that is, CLN,
  ALN.
• ORG, TIM, NUM.

4
Differences Between ENE CNE Recognition

• The main differences between Chinese NE
  Recognition and English NE Recognition include
• Unlike English, Chinese lacks the capitalization
  information which plays very important roles in
  identifying named entities.
• There is no space between words in Chinese, so
  we have to segment the text before NER.
  Consequently, the errors in word segmentation
  will affect the result of NER.

5
Related Work

• Approaches for NER is focused on machine learning
• CoTraining, CoBoosting
• Hidden Markov Model
• Maximum entropy model
• Transformation-based learning
• Typical System of Chinese NER
• Hsin-His Chen, et al. 1997
• Yu, et al. 1998
• CHUA, et al. 2000
• Sun, et al. 2002

6
Chinese NER with Multiple Features

• Features of Chinese NE
• Chinese NEs have very distinct word features in
  their composition and contextual information.
• Data sparseness is very serious when only using
  word features.
• Three Basic Idea
• Combine coarse particle feature (POS Model) with
  fine particle feature (Word Model).
• Introduce heuristic human knowledge into
  statistical model
• Use sub-models to respectively describe
  Japanese, Russian and Euramerican transliterated
  person name and Chinese person name.

7
The Hybrid Model

• The task of NER
• Given a word/pos sequenceto find the optimal
  sequence WC/TC by splitting, combining and
  classifying the given sequence
• We could obtain the optimal sequence WC/TC
  through the following three models
• the Word Model
• the POS Model
• the Hybrid Model

8
Word Model POS Model

• Word Model
• Estimates the probability of generating a NE
  from the viewpoint of word sequence
• The
• POS Model
• Estimates the probability of generating a NE
  from the viewpoint of POS sequence
• The

9
The Hybrid Model

• Combines Word Model with POS Model
• where, factor a 0 is to balance Word Model and
  POS Model.
• The Hybrid Model consists of four sub-models
  word context model P(WC), POS context model
  P(TC), word entity model P(WWC) and POS entity
  model P(TTC).

10
Word Definition in Word Model
11
POS Definition in POS Model

• PKU POS specification is adopted in our system.
• The size of POS set is 48.

12
Context Model

• Word Context Model
• POS Context Model

13
Entity Model

• Class Definition in Entity Model

14
Word Entity Model

• Word Entity Model for PN
• Word Entity Model for LN and ON
• Word Entity Model for ALN

15
POS Entity Model

• POS Entity Model for PN
• Use Word Entity Model for PN to replace the POS
  Entity Model for PN.
• POS Entity Model for LN and ON
• POS Entity Model for ALN

16
Training Corpus for Context Model

• Five-months Peoples Daily tagged with NER tags

17
Training Corpus for Entity Model

• Chinese PN 15.6 million
• Japanese PN 0.15 million
• Euramerican PN 0.4 million
• Russian PN 0.44 million

18
Heuristic Human Knowledge for PN

• Chinese PN surname list (including 476 items)
• Japanese PN surnames list (including 9189 items)
• Russian PN characters lists
• Euramerican PN characters lists
• NE Length restriction

19
Heuristic Human Knowledge for LN

• Location keyword list (including 607 items)
• General word list (such as verbs and
  prepositions) Words in the list usually is
  followed by a location name, such as "?/at",
  "?/go".
• ALN name list (including 407 items)

20
Heuristic Human Knowledge for ON

• Organization keyword list (including 3129 items)
• An organization name template list which is used
  to recognize the missed ONs in the statistical
  model. Some of these templates are as follows.
• ON -- LN D OrgKeyWord
• ON -- PN D OrgKeyWord
• ON -- ON OrgKeyWord

21
Back-off Model to Smooth

• Escape probability to smooth the statistical
  model

22
Experiments

• Three Experiments
• Will the Hybrid Model be more effective than the
  Word Model and the POS Model?
• Will the conclusion from different testing sets
  be consistent?
• Will the performance be improved significantly
  after combining heuristic human knowledge?
• Metrics for evaluations
• Precision, Recall and F-Measure.

23
Experiment I

• a in the Hybrid Model denotes the balancing
  factor of the Word Model and the POS Model
• The largera, the larger contribution of the POS
  Model. The smallera, the larger contribution of
  the Word Model.
• To find the best value of aon the testing corpus
  of one-month's People's Daily.

24
Experiment I (cont.)

• Performance of Recognizing PNs Impacted by a

25
Experiment I (cont.)

• Performance of Recognizing LNs Impacted by a

26
Experiment I (cont.)

• Performance of Recognizing ONs Impacted by a

27
Experiment I (cont.)

• Choose the best value a 2.8 after comparing the
  figures.

28
Experiment I (cont.)

• Conclusion 1
• The Hybrid Model can improve the performance of
  both the Word Model and the POS Model.
• The improvements for PN, LN and ON are different.
  That is, the POS Model has obvious side-effect on
  the recall of ON recognition at all times, while
  the recalls for PN and ON recognition are
  improved in the beginning but decreased in the
  ending with the increasing of a.

29
Experiment II

• Experiments on the MET-2 testing corpus to
  validate the conclusion from Experiment I.
• This table validate that the Hybrid Model is
  better than both the Word Model and the POS
  Model.

30
Experiment II (cont.)

• Conclusion 2
• Our algorithm has consistence on different
  testing, i.e. the Hybrid Model which combining
  the Word Model with the POS Model can achieve
  better performance than either the Word Model or
  the POS Model on different testing sets.

31
Experiment III

• To validate the idea heuristic human knowledge
  can not only reduce the search space, but also
  improve the performance.

Model I Statistical Model without
knowledge Model II Statistical Model with
knowledge
32
Experiment III

• Conclusion 3
• From this experiment, we learn that human
  knowledge can not only reduce the search space,
  but also significantly improve the performance of
  pure statistical model.

33
Conclusion

• Propose a hybrid Chinese NER model which
  combines multiple features
• The main contributions are as follows
• The proposed Hybrid Model emphasizes on
  integrating coarse particle feature (POS Model)
  with fine particle feature (Word Model), so that
  it can make up the disadvantages of each other
• In order to reduce the search space and improve
  the efficiency of model, we incorporate heuristic
  human knowledge into statistical model, which
  could increase the performance of NER
  significantly
• For capturing intrinsic features in different
  types of entities, we design several sub-models
  for different entities. Especially, we divide
  transliterated person name into three sub-classes
  according to their characters set, that is, CPN
  JPN, RPN and EPN.

34

• Thank You!

35
Compare with Sun, et al. 2002(2)