Terminological knowledge extraction

1
Terminological knowledge extraction

• a machine learning approach
• TKE 2005
• Lone Bo Sisseck
• PhD Student
• Dept. of Computational Linguistics

2
Outline

• Overall PhD Project
• Applied method for the machine learning approach
• Description of data
• Results of the experiment
• Conclusion and further research

3
Overall PhD project

• Research into the subject field of linguistic
  signals (Jennifer Person, 1998) that can indicate
  semantic relations between concepts in domain
  specific text
• knowledge probes (Khurshid Ahmad, 1993)
• knowledge patterns (Ingrid Meyer, 2001)
• lexico-syntactic patterns (Marti Hearst, 1992)

4
Overall PhD project

• Example of linguistic signals
• Chrom er et sporstof/generic-specific rel. (ISA)
• chromium is a trace mineral
• Mangel på kobber medfører anæmi/causal rel.
  (CAUSE)
• lack of copper causes anaemia

5
Overall PhD project

• Areas of interest
• Identification of Danish linguistic signals that
  can indicate semantic relations between concepts
  in specialized texts (Lotte Weilgaard, 2002)
• I.e. the generic-specific relation (ISA), the
  causal relation (CAUSE) or the partitive relation
  (PART)
• Manifestation of concepts in writing and in the
  mind
• The complexity of concepts
• Automatic extraction of linguistic signals

6
Applied method for the machine learning approach

• Brill tagging (Eric Brill, 1995). Easy
  accessibility
• Learning from already POS tagged corpus
• All occurrences of a specific semantic relation
  are marked up with the semantic relation tag in
  the training corpus

7
Applied method for the extraction experiment

• For instance
• chrom/N er/V_PRES et/PRON_UBST sporstof/N
• Chromium/N is/V_PRES a/PRON_UBST trace mineral/N
• chrom/N er/V_PRES_ISA et/PRON_UBST sporstof/N
• Chromium/N is/V_PRES_ISA a/PRON_UBST trace
  mineral/N

8
Applied method for the extraction experiment

• An example of a sequence of rules learned for the
  verb
• medføre(cause(s)) indicating the causal
  relation
• V_PRES V_CAUSE PREV1OR2OR3TAG ADJ
• V_INF V_CAUSE PREVBIGRAM N V_PRES
• V_PRES V_CAUSE PREV1OR2TAG PRÆP
• In the extraction process, the contextual rules
  are applied in succession
• The rules are applied to a test corpus in order
  to extract
• linguistic signals that can indicate the causal
  relation

9
Data

• The data consists of
• 3 training corpora/variations of the same corpus
• Nutrition corpus
• 1 test corpus
• Medical corpus

10
Data

• Training corpus 1 92 sentences all containing
  the Danish verb er (is)
• Part of the nutrition corpus of 20.000 words
• POS tagged
• Manually term tagged with unigrams, bigrams and
  trigrams

11
Data

• Manual term mark up in training corpus 1
• N (Noun) where the N is domain-specific
• ADJ (adjective) N (noun) where the two words
  belong together and form a term, e.g. frie
  radikaler (free radicals).
• N (noun) P (preposition) N (noun) when it is
  possible to convert the phrase into a one word
  term, e.g. mangel på A-vitamin (lack of vitamin
  A) A-vitaminmangel (Vitamin A lack).

12
Data

• Training corpus 2 the whole nutrition corpus of
  20.000 words from the OQ project
• POS tagged
• Term tagged with unigrams

13
Data

• Manual term mark up in training corpus 2
• N (Noun) where the N is domain-specific
• A word list from the nutrition corpus was
  compared with a list of Danish general language
  words from the Danish module of INTEX (Tine
  Lassen, 2004).
• Reduced manually

14
Data

• Training corpus 3 nutrition corpus of 20.000
  words - not term tagged
• Test corpus medical corpus of 240.000 words
• Extracted from the Internet, www.netdoktor.dk
• not term tagged

15
Results of the experiments

• Experiment 1
• Based on training corpus 1
• Results from the paper
• Experiment 2
• Based on training corpus 2 and 3
• Comparison of results with and without term mark
  up
• Experiment 3
• Testing contextual rules on test corpus

16
Experiment 1

• Experiment no 1
• Training corpus 1 CORPUS of 92 sentences all
  containing
• the Danish verb er. TERM tagged with unigrams,
  bigrams
• and trigrams

17
Experiment 1

• RULES leaned from mark up of er, is, as
  V_PRES_ISA
• V_PRES V_PRES_ISA NEXT1OR2OR3TAG T
• V_PRES_ISA V_PRES NEXT1OR2TAG PRÆP
• V_PRES_ISA V_PRES NEXT1OR2TAG V_PARTC_PAST
• V_PRES_ISA V_PRES NEXTBIGRAM T ADJ
• V_PRES_ISA V_PRES NEXT1OR2WD ikke

18
Experiment 1

• RULES leaned from mark up of er, is, as
  V_PRES_ISA
• V_PRES V_PRES_ISA NEXT1OR2OR3TAG T
• Chrom er et sporstof
• (Chromium is a trace mineral)
• V_PRES_ISA V_PRES NEXT1OR2TAG PRÆP
• Vitaminer er kendt for at forebygge
  mangelsygdomme.
• (Vitamins are known for preventing illnesses.)
• V_PRES_ISA V_PRES NEXT1OR2TAG V_PARTC_PAST
• I biologien er antioxidanter defineret som enhver
  substans . (Within biology antioxidants have
  been defined as any substance .)

19
Experiment 1

• V_PRES_ISA V_PRES NEXTBIGRAM T ADJ
• Uden behandling med protein, fx mælkepulver, er
  sygdommen dødelig.
• (Without treatment with protein, e.g. milk
  powder, the illness is fatal.)
• V_PRES_ISA V_PRES NEXT1OR2WD ikke
• Hos mennesker er mangelsymptomer ikke
  karakteristiske og . (Symptoms of lack (of
  vitamins) are not characteristic for humans.)
• Extraction of non-ISA senses manually from test
  corpus
• (whole nutrition corpus, 20.000 words). Precision
  87

20
Experiment 2

• Experiment no 2
• Training corpus 2 nutrition corpus of 20.000
  words
• Term tagged with unigrams only

21
Experiment 2

• RULES leaned from mark up of er, is, as
  V_PRES_ISA
• V_PRES V_PRES_ISA SURROUNDTAG N_T PRON_UBST
• Kræft/N i/PRÆP nyrene/N_T er/V_PRES en/PRON_UBEST
  alvorlig/ADJ sygdom/N_T
• Cancer in the kidnies is a serious illness
• V_PRES V_PRES_ISA NEXT1OR2TAG XX
• de/PRON_DEMO mest/ADJ brugte/ADJ kosttilskud/N_T
  er/V_PRES kombinerede/V_PAST vitamin-/XX og/SKONJ
  mineraltilskud/N_T
• the most used dietary supplements are combined
  vitamin and mineral supplements

22
Experiment 2

• V_PRES V_PRES_ISA NEXTWD betacaroten
• det/PRON_DEMO mest/ADJ kendte/ADJ caroten/N_T
  er/V_PRES betacaroten/N_T
• The most known caroten is betacaroten

23
Experiment 2

• Training corpus 3 nutrition corpus of 20.000
  words
• Not term tagged
• RULES leaned from mark up of er, is, as
  V_PRES_ISA
• V_PRES V_PRES_ISA SURROUNDTAG N PRON_UBST
• V_PRES V_PRES_ISA NEXT1OR2TAG XX
• V_PRES V_PRES_ISA NEXTWD betacaroten

24
Experiment 2

• Results of experiment 2
• Extraction of ISA relations from Training Corpus
  2
• Recall 29 Precision 95
• Extraction of ISA relations from Training Corpus
  3
• Recall 29 Precision 85

25
Experiment 3

• Experiment no 3
• Rules for the causal relation ( CAUSE) learned
  from the verb
• medføre/medfører to cause/cause(s)
• V_PRES V_CAUSE PREV1OR2OR3TAG ADJ
• Store/ADJ mængder/N kortisol/N medfører/V_PRES
  alvorlige/ADJ bivirkninger/N
• large amounts of cortisone cause serious adverse
  affects
• V_INF V_CAUSE PREVBIGRAM N V_PRES
• Store/ADJ doser/N kortisol/N vil/V_PRES
  medføre/V_INF bivirkninger/N
• large amounts of cortisone will cause serious
  adverse affects

26
Experiment 3

• V_PRES V_CAUSE PREV1OR2TAG PRÆP
• Mangel/N på/PRÆP ADH/N medfører/V_PRES
  udskillelse/N af/PRÆP urin/N
• lack of ADH causes secretion of urine

27
Experiment 3

• Test corpus medical corpus of 240.000 words
• Extracting ISA relation from Test Corpus
• Relative recall 35 Precision 40
• Extracting CAUSE relation from Test Corpus
• Recall 47 Precision 70

28
Experiment 3

• 5 of the errors are due to errors in the POS
  tagging (ISA)
• Undertiden/N er/V_PRES_ISA man/PRON_UBST nødt/ADJ
  til/PRÆP at/UNIK prøve/V_INF forskellige
  behandlingsformer/N
• Sometimes/N you have to try different treatments
• 3 of the errors are due to errors in the POS
  tagging (CAUSE)
• Højresidigt/ADJ hjertesvigt/ADJ kan/V_PRES
  medføre/V_PRES hævelse/N af/PRÆP benene/N
• Heart disorder/ADJ in the right side can cause
  swelling of the legs

29
Conclusion and future work

• The results with term tagging were better than
  without, but requires adequate term extraction
  tools
• It is useful to continue the extraction process
  without term recognition
• The precision is best when extracting the causal
  relation (CAUSE) from the test corpus
• Future work To make the system perform better by
  adding more data, and perform iterative learning
  processes on the results of the extraction process

30

• Thank you!