Iterative Translation Disambiguation for Cross Language Information Retrieval Christof Monz and Bonnie J. Dorr Institute for Advanced Computer Studies University of Maryland SIGIR 2005

INTRODUCTION Query translation requires access to some form of translation dictionary Use machine translation system to translate the entire query into the target language Use of a dictionary to produce a number of target-language translations for words or phrases in the source language Use of a parallel corpus to estimate the probabilities that w in the source language translates into w ’ in target language

INTRODUCTION An approach that does not require a parallel corpus to induce translation probabilities a machine-readable dictionary (without any rankings or frequency statistics) a monolingual corpus in the target language

TRANSLATION SELECTION Translation ambiguity is very common Apply word sense disambiguation For most languages the appropriate resources do not exist Word-sense disambiguation is a non-trivial enterprise

TRANSLATION SELECTION Our approach uses co-occurrences between terms to modeling context for the problem of word selection Ex. S1=>t11,t21,t31 S2=>t21,t22 S3=>t31

TRANSLATION SELECTION Computing co-occurrence statistics for a larger number of terms induces a data- sparseness issue Use very large corpora (Web) Apply smoothing techniques

ITERATIVE DISAMBIGUATION Only examine pairs of terms in order to gather partial evidence for the likelihood of a translation in a given context

ITERATIVE DISAMBIGUATION Assume that t i1 occurs more frequently with t j1 than any other pair of candidates between a translation for si and sj On the other hand, assume that t i1 and t j1 do not co-occur with tk1 at all, but t i2 and t j2 do Which should be preferred: ti1 and tj1 or ti2 and tj2

ITERATIVE DISAMBIGUATION Associate with each translation candidate a weight (t is a translation candidate for si) Each term weight is recomputed based on two different inputs the weights of the terms that link to the term (W L (t,t ’ )=link weight between t and t ’ )

ITERATIVE DISAMBIGUATION Normalize term weights The iteration stops if the changes in term weights become smaller than some threshold (W T= the vector of all term weights V k =kth element in the vector)

ITERATIVE DISAMBIGUATION There are a number of ways to compute the association strength between two terms MI Dice coefficient log likelihood

ITERATIVE DISAMBIGUATION Example

EXPERIMENT Set-Up Test Data CLEF 2003 English to German bilingual data Contains 60 topics, four of which were removed by the CLEF organizers, as no relevant documents Each topic has a title, a description, and a narrative field, for our experiments, we used only the title field to formulate the queries

EXPERIMENT Set-Up Morphological normalization Since the dictionary only contains base forms, the words in the topics must be mapped to their respective base forms as well Compounds are very frequent in German Instead of de-compounding, we use character 5- grams, an approach that yields almost the same retrieval performance as decompounding

EXPERIMENT Set-Up Ex. Topics Intermediate results of the query formulation process

EXPERIMENT Set-Up Retrieval Model - Lnu.ltc weighting scheme we used sl=0.1,pv=the average number of unique words per document, uw d = refers to the number of unique words in document d, w(i) = weight of term i

Experimental Results

Individual average precision decreases for a number of queries 6% of all English query terms were not in the dictionary Unknown words are treated as proper names, and the original word from the source language is included in the target language query Ex. the word Women is falsely considered a proper noun, although faulty translations of this type affect both the baseline system and the run using term weights, the latter is affected more severely

RELATEDWORK Pirkola ’ s approach does not consider disambiguation at all Jang ’ s approach use MI to re-compute translation probabilities for cross-language retrieval Only considers mutual information between consecutive terms in the query they do not compute the translation probabilities in an iterative fashion

RELATEDWORK Adriani ’ s approach is similar to the approach by Jang does not benefit from using multiple iterations. Gao use a decaying mutual-information score in combination with syntactic dependency relations We did not consider distances between words

RELATEDWORK Maeda compare a number of co-occurrence statistics with respect to their usefulness for improving retrieval effectiveness They consider all pairs of possible translations of words in the query use co-occurrence information to select translations of words from the topic for query formulation, instead of re-weighting them

RELATEDWORK Kikui Only need a dictionary and monolingual resources in the target language Computes the coherence between all possible combinations of translation candidates of the source terms

CONCLUSIONS introduced a new algorithm for computing topic dependent translation probabilities for cross-language information retrieval We experimented with different term association measures, experimental results show Log Likelihood Ratio has the strongest positive impact on retrieval effectiveness

CONCLUSIONS An important advantage of our approach is that it only requires a bilingual dictionary and a monolingual corpus An issue that remains open at this point is the computation of query terms that are not covered by the bilingual dictionary