Analysis of a 16-state Ergodic HMM

The results of a 16-state Ergodic HMM are shown in Figure  2 . In this example, a lot of state transitions are shown for only one POS. In addition, the initial state value has two states (○ 2,○ 3). The 16-state Ergodic HMM is more refined than the 8-state Ergodic HMM.

図 2: Results of a 16-state Ergodic HMM analysis

The results of this analysis are shown in the following for each POS.

1. Interjection

   Interjections have outputs from (○ 2, ○ 3) which have a high initial value like the 8-state Ergodic HMM. However, the initial value has two state (○ 2, ○ 3), and does not have a self loop of state 2.

2. Common noun

   Common nouns are mainly output from ○ 5○ 6○ 7 to ○ 1○ 9○ 12.

   Common nouns are output from states ○ 6○ 7. Other nouns are output from state ○ 5; this state has many kinds of nouns.

3. Conjugate word

   In the 8-state Ergodic HMM, mainly the conjunctions are grouped, However, in the 16-state Ergodic HMM, the POSs are grouped.

   Verbs are output from ○ 8○ 5, conjoining verbs are output from states ○ 0○ 4, and auxiliary verbs are output from ○ 0○ 11○ 13.

4. Case particle

   Case particles are output from ○ 8○ 9○ 12. Each transition is another word (like the 8-state Ergodic HMM). However, the 16-state Ergodic HMM has more precise clustering. Examples are shown as follows.

   
   
   

   • state ○ 9 $\Rightarrow$○ 7 /no/(93%)
     
   • state ○ 9 $\Rightarrow$○ 5 /kara/(57%) /de/(23%)
     
   • state ○ 8 $\Rightarrow$○ 15 /to/ (63%) /o/ (28%)
     
   • state ○ 0 $\Rightarrow$○ 8 /de/ (86%)
     
   • state ○ 1 $\Rightarrow$○ 8 /ni/ (87%)
     
   • state ○ 12 $\Rightarrow$○ 8 /wo/ (75%)
   
   

   The case particle of /de/ and /ni/ or /ga/ and /wo/ have the same transitions as in the case of the 8-state Ergodic HMM. However, these particles are divided in the 16-state Ergodic HMM.