The task of inferring translations can be achieved by the means of comparable corpora and in this paper we apply explicit topic modelling over comparable corpora to the task of inferring translation candidates. In particular, we use the Orthonormal Explicit Topic Analysis (ONETA) model, which has been shown to be the state-of-the-art explicit topic model through its elimination of correlations between topics. The method proves highly e ective at selecting translations with high precision.
Explicit topic modelling, such as proposed by the Explicit Semantic Analysis
(ESA) [
Translation inference is the task of inferring a translation equivalent between
two languages by means of other bilingual dictionaries in other language pairs.
The principle issue is that the translation graph is not transitive, so by
following a translation pair from English to Spanish, and then a translation pair
from Spanish to French and incorrect translation may be inferred if there are
multiple senses of the Spanish word that is used as a pivot. However, previous
TIAD tasks [
Orthonormal explicit topic analysis follows from explicit semantic analysis by assuming there is a background collection of documents we call B = fb1; : : : ; bng, and in the cross lingual setting it is assumed that there is a paired set of documents B0 = fb01; : : : ; b0ng, with each document being paired with a similar document in a second language. This is most frequently achieved by using Wikipedia, where interlingual links link two articles in di erent languages. It is assumed that we have some language-speci c function that maps a document to a vector in Rn, such that the jth element of the vector is an association with the j (d) with the document bj . In our method, this vector is given by a metric such as TF-IDF such that in our score is: j (d) = tf-idf(b!j)Ttf-idf(!d) (1)
If we consider the application of this method to the background corpus we can construct a matrix X, whose elements are the corresponding TF-IDF values xwj = tf-idfw(bj ), and hence that i(bj ) is the i; jth element of XTX. One of the key assumptions is that we should have that topics that are as distinct as possible in order to reduce the amount of overlap between the topics. This is achieved by assuming that we have some function sim : B ! [0; 1] that has the following property: sim(bi; bj ) = 1 if i = j 0 if i 6= j
This can be though of as maximizing training accuracy as we are ensuring
that the similarity of two di erent topics in our background is zero and the
similarity of the topic with itself is one. In McCrae et al. [
ONETA(d) = (XTX) 1 (d) = X+tf-idf(!d) 1 X+ denotes the Moore-Penrose pseudo-inverse, which satis es X+X = I (2) (3)
For any choice of (d) where (XTX) 1 exists and it is easy to verify that Equation 2 holds as:
I = (XTX) 1XTX = X+X
In practice the computation of this matrix can be time-consuming so instead McCrae et al. proposed rearranging the order of the vocabulary in the background collection to nd a good approximation of the form:
And as it is easy to verify the following equation based on a matrix of this form:2
X '
A B 0 C I =
A+ 0
A+BC+
C+
A B 0 C 0ONETA(d) =
A+ 0
A+BC+
C+ tf-idf(!d) (4) (5) (6) (7)
This leads to a strong approximation of ONETA as follows: 3
Spanish French Portuguese Portuguese
The key purpose of ONETA is to estimate the similarity between documents, and to apply it to the task of inferring the similarity of translation, we make the simple assumption that each term in a translation is a single document consisting of only the term in question. As such we simply apply the system by building two ONETA functions for our source language, s, and target language t and estimate the similarity as: 2 McCrae et al. use the Jacobi preconditioner of C as a further approximation of C+ sim(ws; wt) = cos( sONETA(ws); tONETA(wt)) (8)
In order to construct pairs, we considered only the simple pivot between one language, and as for this task the languages were English, French and Portuguese, there were only two common languages between them namely Spanish and Catalan, and as such we called our two systems ONETA-ES and ONETACA based on the pivot language. We simply considered all possible translations between the two language pairs and then calculated the similarity using the ONETA score. As we found that the distribution of the scores was strongly clustered around zero, we used the following function to provide a more even spread of scores.
sim0(ws; wt) = jsim(ws; wt)j (9)
For our experiments we tuned = 0:3 to provide a reasonable spread of certainty values. As with previous work we used Wikipedia to construct our corpora using the interlingual index to create a comparable corpus for each language pair, the sizes of which are given in Table 1. 4
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.845 0.904 0.902 0.894 0.885 0.884 0.878 0.869 0.866 0.867 0.541 0.237 0.184 0.149 0.119 0.093 0.072 0.053 0.038 0.022
F1 0.659 0.375 0.306 0.255 0.209 0.168 0.133 0.101 0.072 0.043
During development we evaluated on the English to Spanish translations using Catalan as a pivot, as all language pairs are available as part of the training data and the results are presented in Table 2. It should be noted that at the threshold value of 0% the system is basically nothing more than pivot translation and this should be considered a baseline. For higher values of the threshold, ONETA does improve the precision, however the recall also decreases rapidly causing the F-Measure to fall overall. F1
In the o cial results (Table 3), we see a similar outcome where the highest F-Measure is achieved at the trivial threshold of 0% and we see strong gains in precision at the cost of recall. This shows that ONETA can quite e ectively select translations that are very likely to be correct but misses many translations even among those that are generated by a pivot method.
When all systems are compared (Figure 1) at various threshold levels we see that the ONETA-ES system actually reports the strongest F1 Measure (averaged over all language pairs) of any system, however it should be noted that this is a threshold value that we would consider to be a baseline. Even still, we see that at the threshold of 0.1, ONETA still has the second and eighth best result, moreover we have achieved the strongest precision scores across all languages (except for results with a recall that was reported as zero). 5
We have presented the ONETA system and its application to translation, which was the only system to produce a value that beat the baseline, albeit when it is in a mode where it e ectively a baseline itself. The system does show noticeable ability to tune between precision and recall and as such it would likely be e ective for usage in areas where precision is more important than recall, for example in a semi-automated setting where showing annotators too many poor quality translations would waste time. There are two principle aws with the implementation as it stands: rstly, that the recall is limited and even our baseline mode we only achieved a recall of about 20-30%, which needs to be overcome by nding more translations than are just present in the graph. Secondly, the system is not aware of senses, and the selection of multiple document collections likely to show many di erent senses of a word, may help the system to distinguish between translation pairs which do not rely on the most frequent senses of words.
This publication has emanated from research supported in part by a research grant from Science Foundation Ireland (SFI) under Grant Number SFI/12/RC/2289, co-funded by the European Regional Development Fund, and the European Unions Horizon 2020 research and innovation programme under grant agreement No 731015, ELEXIS - European Lexical Infrastructure.