Cross-Lingual Word Sense Disambiguation using WordNets and Context - - PowerPoint PPT Presentation

Cross-Lingual Word Sense Disambiguation using WordNets and Context Mapping

Priyank Jaini Ankit Agrawal

{pjaini,ankitag}@iitk.ac.in Department of Mathematics and Statistics IIT Kanpur Advisor: Prof. Amitabha Mukerjee Date: March 21,2013

What is Word Sense Disambiguation(WSD)?

• assigning the correct sense(meaning/context) to a word in a sentence when it can have

multiple meanings

• Example:
• >I was standing on the bank of river Ganga.
• >Mr.Bank owns a bank.

Importance and Motivation

• Machine translation, Lexicography,semantic interpretation,Information retrieval etc
• Hindi lacks resources
• Can be used to create/enrich sense-tagged data

Our Approach

• Parallel Corpus and Alignment
• English WSD
• Synset Mapping
• Transfer to Hindi

Methodology and Algorithms Used

1)Parallel Corpus for Hindi-English (Emille) 2)Alignment of text using Church and Gale Algorithm 3)English WSD on the English text 4)Synset mapping using [10] 5)Transfer senses to Hindi text

Figure taken from [1]

The English Word Sense Disambiguation (Step-3)

• We shall use “WordNet::SenseRelate::AllWords”
• Uses Lesk Algorithm for disambiguation
• After this step, we would have a sense-tagged English text.
• English WordNet would be used for English WSD

Synset Mapping (Step 4)

• Takes an English synset as input and produces as output the best matching Hindi

Synset

• Uses the fact that in WordNet, the first word in a synset best represents the sense of

the synset

• Hypernymy relation is the basis for finding the best match
• In Hypernymy Hierarchies, a weighted formula given in [10] is used to determine the

best synset.

Synset Mapping

• Candidate synsets: obtained by finding the Hindi translations of the first word in the

input synset and then finding the Hindi synsets that contain one or more of these translations in them

• Hypernymy hierarchies of these candidate synsets found. They are called candidate

hierarchies

• Hypernymy hierarchy of the input English synset is also obtained.
• For each synset obtained in the English hypernymy hierarchy, hindi translations of

all the words occuring in it are found.

• These Hindi words are found in the candidate hierarchies. If a match is found,

weight of that candidate synset is increased. Initially, the weights are zero.

• The total weight for each candidate hierarchy is obtained, and the one with the

highest weight is mapped to the English synset.

We are expecting:

• Since a parallel aligned corpus is used we should achieve a better accuracy
• Would give better results for scenarios where:
• An English word is polysemous and it's Hindi equivalent is also polysemous
• An English word is monosemous and it's Hindi equivalent is polysemous

Limitations

• Is valid only for nouns
• Not trained for morphological handling

References

1)Debasri Chakrabarti,Dipak Kumar Narayan,Prabhakar Pandey,Pushpak Bhattacharyya.Experiences in building the Indo Word Net-A WordNet for Hindi. 2)Bahareh Sarrafzadeh, Nikolay Yakovets, Nick Cercone, Aijun An. Cross Lingual Word Sense Disambiguation for Languages with Scarce Resources. 3)Els Lefever and Veronique Hoste. SemEval-2010 Task 3:Cross-Lingual Word Sense Disambiguation. 4) Michael Lesk.Automatic sense disambiguation using machine readable dictionaries:how to tell a pine cone from an ice cream cone. In SIGDOC’86: Proceedings of the 5th annual international conference on Systems documentation,pages 24-26, New York, NY, USA, 1986.ACM 5) Satanjeev Banerjee and Ted Pedersen. Extended gloss overlaps as a measure of semantic

• relatedness. In IJCAI’03, pages 850-810,2003.

6) Els Lefever and Veronique Hoste. Examining the validity of Cross-Lingual Word Sense Disambiguation 7) http://wordnet.princeton.edu/ 8) Roberto Navigli. Word Sense Disambiguation-A Survey. 9) Roberto Navigli and Simone Paolo Ponzetto. BabelNet: The automatic construction, evaluation and application of a wide-coverage multilingual semantic network. 10) Ramanand, Akshay Ukey, Brahm Kiran Singh, and Pushpak Bhattacharyya. Mapping and structural analysis of multi-lingual wordnets.

Thank You!!

Church and Gale Algorithm

• A method for aligning sentences based on a statistical model of character lengths
• Uses the fact that longer/shorter sentences in one language tend to be translated

into longer/shorter sentences in another language.

• The algorithm is a two step process: 1)Paragraph alignment and then 2)Sentence

alignment

• Based on a probabilistic model
• Also, it is language independent, though would have to be tested on Hindi-English.

Ref:A Program for Aligning Sentences in Bilingual Corpora, William A Gale and Kenneth W. Church

English WSD:WordNet::SenseRelate::AllWords

• Each target word is centered in a balanced window whose size is decided by the user.
• The possilble senses of the word are measured for similarity relative to the senses of

the surrounding words present in the window in a pairwise fashion

• The sense of the word that has the highest score after summing up the pair-wise score

is considered the sense of the word.

• For finding similarity it uses the 10 measures of relatedness proposed in

WordNet::Similarity[http://wn-similarity.sourceforge.net]

Lesk Algorithm

• Assigns sense to a word by comparing glosses of the surrounding words with the glosses
• f various senses of the target word.
• The sense whose gloss has most number of overlaps is assigned.
• Extended Lesk Algorithm uses context hierarchy of WordNet to improve the accuracy.