forked from ted-strauss-K1/jTRACE
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathjournal.txt
124 lines (77 loc) · 4.96 KB
/
journal.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
*** Up to Friday 22nd of June
Looked at the jTrace code
Started off alternate model
Started off own simulation viewer (comparing directly our model to the original trace)
*** Friday 29th of June
Still building the benchmark interface (lexicon editor)
Results of "locking" the schwa phoneme :
- all the words containing schwa get over-activated (even if schwa isn't in the input, as it still gets activated thanks to other vowels)
- the silence word "-" is recognised when word not recognised, e.g. "tart" -> "-" instead of "tart" -> "bar"
*** Saturday 30th of June
Built a simple Lexicon Editor. Extracted the provided lexicons into "simple" files.
*** Sunday 1st of July
Built Schwa component, which receives activation from schwa phoneme in phonLayer.
Built a lexicon activator (schwa listener) which sends extra activation to the words containing schwa
*** Monday 2nd of July
Built Schwa graph.
Results :
- words containing "^" get a boost in activation
- when the input contains "^", the word is recognized a bit sooner. For other inputs, the results are essentially the same.
*** Tuesday 3rd of July
Activations of the original and modified model are now shown on the same graph, for easier comparison.
*** Thursday 5th of July
Reading on Lexical Stress an full/reduced vowels stuff
*** Friday 6th of July
Removed schwa phoneme from the net.
Trying to get a smooth Schwa curve (pb in my code), brings up the question : should the Schwa component be spread over time ?
*** Wednesday 13th of July
The schwa component is spread over time (as original phoneme was)
*** Tuesday 24th of July
Function words added (some already existed -> lexicon loading makes sure they are unique)
*** Friday 25th of July
Evaluation.java
cleaning the code
*** Tuesday 29th of July
running Evaluations + perl scripts to extract data
*** Friday 3rd of August
removed the new schwa "@" from the phonemes (what a troublemaker)
*** Sunday 5th of August
upgraded to JFreeChart 1.0.12 (to remove legend...)...... downgraded back to 1.0.0 (too many things to change + different behavior)
*** Monday 6th of August
Now schwa is activated by the feature level directly, schwa inhibits other phonemes and activates words which contain it.
Lexical Stress component (re)implemented. Results show that it performs better for bisyllables but worse for mono- and tri-.
TODO
Make an evaluation class manipulable at a high-level. (original/concrete, phoneme extended or not, lexicon extended or not, inhibition/excitation parameters, single/multiwords)
Make separate phoneme sets (original / expanded)
Make sure evaluation is correct (specially if it disadvantages Concrete-TRACE)
Construct multiword strings
Build evaluation of multiword strings
Make sure most of the words in the lexicon have a weak syllable on the schwa
Word segmentation: try pairs of words and see what happens...
Add comments to the code + README
read Ziegler: nested... (augmenting the model)
read Ahissar & Hochstein: magnocellular pathways (some fast pathways draw a sketch of the input and comes back down to meet the rest of the information going up)
no more simulations after the 7th
NOTES
Phonemes which are closest to schwa (in that order) :
i u a ^
*** RESULTS ON RUNNING THE COMPARISON BETWEEN THE 2 MODELS (without Lexical Stress) ***
SAME : 322
BETTER : 259 e.g. r^b^ri
WORSE : 189 e.g. sartr^ (too may "s^..." words get activated at first)
NOOO : 149 e.g. rab->r^b (most due to equivalent with schwa)
interesting words (modified wins, with change in curves) : lukriS^s, prab^bli
*** With lexical stress ***
works wonders with : sal^dli
******* RESULTS ***********
reference1 : original TRACE // NOOOO I don't know what that is
oldmodified1 : first modified model (schwa removed from phoneme set)
oldmodified-Word2SchwaActivated.txt : first modified model with words to schwa connexions
reference.txt = REAL reference (only for words containing schwa, though)
Hypotheses for the dissertation:
- 1st hyp: the heterachical architecture around the concrete universal is viable (doesn't blow up)
- 2nd hyp: the architecture provides a principled way of augmenting the model, e.g. Lexical Stress
- 3rd hyp: the model works best if schwa isn't affected by anything else (interesting because counter-intuitive). e.g activates function words
The exact set of words that have weak forms depends on dialect and speaker; the following is a list of the chief words of this type in Received Pronunciation:
a, am, an, and, are, as, at, be, been, but, can, could, do, does, for, from, had, has, have, he, her, him, his, just, me, must, of, shall, she, should, some, than, that (as conjunction), the, them, there, to, us, was, we, were, who, would, you
In most of the above words the weak form contains schwa, or a syllabic consonant in the case of those ending /l/, /m/ or /n/. However in be, he, me, she, we, been, him the vowel may be the reduced form of /ɪ/, or else [i]; and in do, who, you it may be the reduced form of /ʊ/, or [u]. (For the and to, see above.)