BibleTech2011

Word Sense Detection

and

Word sense Disambiguation

through Data-Mining

Andi Wu & Randall Tan

Asia Bible Society

Outline

� Motivations for word sense identification

� Problems of existing word sense data

� The data-mining approach

� Demo and Discussion

Asia Bible Society 2

Motivations

Addressing the issue of Polysemy

� Bible translation

– Better understanding of every word

– Unification on the basis of senses rather than words

� Bible search

– More refined search results on the basis of senses


Goals

� Word sense detection

For each content word in the Bible, find out how

many senses it has.

� Word sense disambiguation

For each instance of the word, find out which of the senses it has.



ראשית

Sense 1

Sense 1-1: beginning

Sense 1-2: first

Sense 2

Sense 2-1: firstfruits

Sense 2-2: firstborn

Sense 3

Sense 3-1: best

Sense 3-2: choicest

Sense 4

Sense 4-1: foremost

Word sense detectionIdentify the senses of each word :

Identify the sense of each instance:

ים את השמים ואת הארץ׃א ה ברא ראשיתראשיתראשיתראשיתב (Gen1:1)

ח׃ח ניחלריו עלהמזבח לא־ י ־ ואלה תקריבו אתם ליהו ראשיתראשיתראשיתראשיתקרבן (Lev 2:12)

, בגלגל׃א היה יהו ל החרם לזבח יתיתיתיתראש ראש ראש ראש צאן ובקר ל ם מהשל הע קח וי (1Sm 15:21)

ש חרבו׃יג ו העשדרכי־ אל יתיתיתיתראש ראש ראש ראש הוא (Job 40:19)


Sense 1: beginning

Sense 2: firstfruits

Sense 3: best

Sense 4: foremost

Word sense disambiguation

Problems of Existing Data

� No consensus on the number of senses each word has

� No complete data of instance-based sense identification

� Manual identification can be subjective, inconsistent, and time-consuming


The Data-Mining Approach

� Theoretical assumption

� Data for mining

� Machine learning procedures

� Advantages and limitations of the approach

� Tool for sense exploration


Theoretical Assumption

� Translators presumably use different target language words to translate different senses of a word (Translators have done the job of disambiguation sub-consciously and defined each sense with target language words).


Data for Mining

Translations linked word-to-word to the original Hebrew/Greek texts


Basic TaskTake all instances of a word and group the instances into different senses


A Simple and Naive Approach

Look at the words used in a given translation and treat instances with the same translation words as having the same sense.


A Simple and Naive Approach

Problems:

� The translations may not be consistent:

Translators may use different words to translate the

same sense or the same word to translate different senses

� It can be subjective

It only reflects the opinions of a particular translation

� The senses are too fragmentary


The Voting Approach

� Use multiple translations

� Two instances of a word is considered to have the same sense if most of the translations use the same word to translate it.

� Check and balance

� How to define “most”?


The Voting Approach

How many votes to get?

� Maximal agreement:

– Internal consistency within groups

– Too many senses

– Too many unassigned instances

� Minimal agreement:

– Better grouping of senses

– Instances of different senses may be mixed together


Progressive MergingTrying to get the benefits of both maximal and minimal

agreement and avoid their disadvantages

� Start with maximal agreement to get initial sense groups that are internally consistent

� Gradually merge the initial groups with decreasing number of agreements N (N > 0, N < Maximal) and with a variable association rate R (R > 0, R < 1)

� Group B is merged into Group A if A contains B

� A contains B if each instance in B is linked to at least Rof the instances in A by N agreements.

� Pair-wise merge until no further merge can be done


Progressive Merging

Merging two groups: association rate = 0.5


Progressive Merging

Example: Maximal N = 4, R = 70%

� Merge 1: N-1 = 3

B merges into A if each instance in B is linked to at least 70% of the

instances in A by sharing the same translation in at least 3 versions

� Merge 2: N-2 = 2


instances in A by sharing the same translation in at least 2 versions

� Merge 3: N-3 = 1


instances in A by sharing the same translation in at least 1 version


Tuning the Variables

We can get different results by tuning the following variables:

� The translations to use

� The number translations to use

� The number of merges to perform

� The association rate


The “Accents” of Senses

� Senses based on English translations

� Senses based on Chinese translations

� Senses based on both English and Chinese

� The triangulating effect of using different translations


Factors Affecting the Results

� The versions of translations that are used

� The quality of each translation

� The degree of consensus between different translations

� Quality of lemmatization in English

� Surface forms vs. lemmatized forms


Other Features Considered

� Syntactic contexts

– Instances that occur in similar syntactic contexts tend to have the same sense

– Not used because of sparse data problem

� Morphological information

– Verbs with different stems in Hebrew tend to have different senses

– Not used because the stem distinctions do not always correspond well with sense distinctions


Editing Options

The data shown here has not been manually edited, but it can be edited using the tool:

� Merge sense groups

� Split a sense group

� Move an instance from one sense group to another

� Use of manual information in automatic learning


Demo


Applications

� Sense-based translation memory

� Sense-based concordance

� Sense-based consistency check


Advantages of the Current Approach

� Efficiency: a sense dictionary which not only lists the senses but also the specific instances of the sense can be built in a matter of days.

� Objectivity: the results are based on actual data and no pre-conceived subjective categorization is required.

� Flexibility: the granularity of sense divisions can be adjusted by the values of similarity metrics in the clustering process.


Conclusion

A great tool for exploring and studying word senses in biblical texts


BibleTech2011

Documents