Human Language Technology in Thailand

1
Human Language Technology in Thailand

• Virach Sornlertlamvanich
• Information RD Division
• National Electronics and Computer Technology
  Center
• virach_at_nectec.or.th
• 22 March 2002

IEEE Colloquium on Signal Processing
2
Knowledge Information - Data

• KnowledgeAbility in understanding, reasoning and
  problem solving
• InformationEntity to inform the others
• DataEntity observed by intelligent agents

3
Artificial Intelligent Agent
4
Research Theme

• Human Language Technology
• Text
• Speech
• Multimedia and Multimodality
• Intelligent Content
• Knowledge Discovery
• Datamining
• Visualization
• Natural Interaction

5
Text
ADLTSUGINFORMATIONBWGZKTILA ??????????????????????
??????
ADLTSUGINFORMATIONBWGZKTILA ??????????????????????
??????
6
Term Candidate Extraction

• Virach Sornlertlamvanich et. al. (COLING 2000)
• Automatic Corpus-Based Thai Word Extraction with
  the C4.5 Learning Algorithm
• C4.5-trained decision tree for determining
  potential word boundary from MI, Entropy and
  Linguistic information
• Capable of discovering new words in document
  without assistance from static dictionary

7
Mutual Information
y z
x y
x
z
where x is the leftmost character of string
xyz y is the middle substring of xyz z is the
rightmost character of string xyz p( ) is the
probability function.
High mutual information implies that xyz
co-occurs more than expected by chance. If xyz is
a word then its Lm and Rm must be
high.Efunction vs ...Function...
8
Entropy
y
x
y
z
where A is the set of characters x is the
leftmost character of string xyz y is the middle
substring of xyz z is the rightmost character
of string xyz p( ) is the probability function.
Entropy shows the variety of characters before
and after a word. If y is a word then its left
and right entropy must be high....?function...
vs ...?unction...
9
Other Features

• Frequency Words tend to be used more often than
  non-word string sequences.
• Length Short strings are likely to happen by
  chance. The long and short strings should be
  treated differently.
• Functional Words Functional words are used
  mostly in phrases. They are useful to
  disambiguate words and phrases.

Result of subjective test Word
precision 85 Word recall 56
10
Evaluation Result of Word Extraction
RID Royal Institute Dictionary (30,000
words of Thai-Thai dictionary)
11
Dictionary-less Search Engine
???? (common noun in common noun)
...?????????????????????????????????????????????..
. family
...??????????????????...????????...??????????...
kitchen
????? (proper noun in common noun)
...???????????????????????????????????????... PR
...????????????? ????????????????????... proper
noun
???? (common noun in proper noun)
...????????????????????????... proper noun
...??????????????????????????? ????????????... el
ement
12
Sansarn
13
Speech
Input String
Sentence Segmentation
Word Segmentation
Grapheme to Phoneme
Text processing
Prosody / Tone Generation
Signal processing
Demi-syllable Concatenation
14
Sentence Segmentation
Input paragraph
Training POS tagged corpus
Word segmentation andPOS tagging
Winnow(Feature-based ML)
Word sequence with tagged POS
Winnow
Trained network
Paragraph with sentence break
15
Accuracy in Word/Sentence Segmentation

• Word Segmentation
• Longest matching (92)
• Maximal matching (93)
• POS tri-gram (96)
• Machine learning (97)
• Sentence Segmentation
• POS tri-gram (85)
• Machine learning (89)

Supervised approaches
16
Thai Grapheme-to Phoneme
PGLR Table
/som/chaj/
/som4/chaj0/
?????
PGLR parser
Most probable parse tree
G-P Mapping
Tone Generation
CFG Rule
G-P Table
17
PGLR Approach

• Probabilistic Generalize LR parsing
• Advantage in context-sensitivity
• Two levels of context
• Global context - over structures from the CFG
  rules (probability in reduce action)
• Local n-gram context (probability in shift
  action)

18
Prosody Generation

• F0 contour
• Intonation
• Downdrift
• Pitch Range
• Upper limit
• Lower limit

19
Prosody Generation

• Tone concatenation
• Tone Location
• Coarticulation

20
Demi-syllable Concatenation

• Sound Units
• Demi-syllable
• Use original tone 1,505 units
• Tone modification technique 605 units

21
Demi-syllable Concatenation

• Speech Signal Processing
• Tonal modification technique
• Pitch-Synchronous Overlap-Add (PSOLA)
• Spectral smoothing technique
• Line Spectrum Pair (LSP) parameter smoothing

22
Demi-syllable Concatenation

• Cross-syllable coarticulation smoothing
• Tone coarticulation PSOLA
• Waveform interaction LSP parameter smoothing
• Concatenation smoothing
• Intra-syllable smoothing
• Inter-syllable smoothing

23
Text to Speech
?????????????????????????????????
??????????????????????????????
Prosody Generation Syllable Duration F0 Contour

• Text Segmentation
• Sentence Extraction
• Phrasing
• Word Segmentation

?? / ?? / ?????? / ??????? / PhraseBreak / ??? /
?? / ???????? / ???

• Speech Signal SynthesisBoundary Smoothing
• Prosodic waveform modification

Graphem-to-Phoneme conversion
/phom4/kh_at_4/kh_at_p1/khun0/thuk3/than2/PAUSE/thi2
/ma0/jiam2/chom0/ngan0/
Speech Signal
24
Ontologies

• EDR
• Approach Word description as employed in
  dictionaries
• Problem Ambiguities and incomputability
• Wordnet
• Approach Synonym set and simple
  semantic relations to other words
• Problem Ambiguities
• UW
• Approach Headwords and semantic restrictions
• Advantage Computability and no ambiguity

25
Ontologies
Representation of concept tired in different
schemes
EDR Wordnet 1.5 UW
- having or displaying a need for rest-
having lost of interest- lack of imagination
- A1 tired (vs. rested)- A2 bromidic,
commonplace, hackneyed, - V1 tire,
pall, grow weary, fatigue- V2 tire, wear upon,
fag out- V3 run down, exhaust, sap, - V4
bore, tire, ...
- tired- tired(iclgtphysical)-
tired(iclgtmental)
26
Universal Word (UW)

• UW format ltheadwordgt ( ltlist of
  restrictionsgt ) e.g. book (icl gt do, obj gt
  room)
• Headword An English word roughly describes
  the UW sense.
• Restrictions
• Inclusion (icl ) indicates the class of the
  sensee.g. car ( icl gt movable thing)

27
Universal Word (UW)

• Restrictions (continued)
• UNL semantic relationse.g. eat ( agt gt
  volitional thing, obj gt food )The agent of this
  UW is restricted to be volitional thing.The
  object of this UW is restricted to be food.

UW Class Hierarchy
28
Milestone
29
Thailand Knowledge-based Economy

• eContent
• Open Source distributed architecture
• Language resources
• Digital Divide
• Speech-based Internet access
• Language Divide
• Mutilingual access

30
Technology Solution

• Intelligent Terminal
• Knowledge Exchange
• Intelligent System
• Natural Interaction
• eContent