This paper describes the di erent proposed approaches to the TIAD 2019 Shared Task, which consisted in the automatic discovery and generation of dictionaries leveraging multilingual knowledge bases. We present three methods based on graph analysis and neural machine translation and show that we can generate translations without parallel data. 1 { a graph-based approach where loops of length four are searched in the Apertium [2] dictionaries in order to discover new translations; { a path-based graph approach which retrieves the translation candidates based on translation inference using the language paths of the Apertium dictionaries;
The growing amount of semantically structured monolingual, as well as multilingual resources, such as dictionaries or knowledge graphs (KGs), o ers an excellent opportunity to explore, link and to enrich them with possibly missing multilingual knowledge. Since a manual translation of such resources is very time consuming and expensive, this work focuses on the automatic generation of dictionary entries, which is the objective of the Translation Inference Across Dictionaries (TIAD-2019) Shared Task.2 In this task, the participants should submit dictionaries containing pairs of source and target language words or expressions, the part of speech (POS) of the entry and a con dence score.
In this work, we propose several di erent approaches for this task: 1 The datasets used and the trained NMT models are available at server1.nlp.insightcentre.org/tiad2019/
{ a multi-way neural machine translation (NMT) model trained with multiparallel English and Spanish, Italian and Portuguese and French and Romanian corpora and tuned with the dictionaries produced in the other approaches. This approach is further ltered using monolingual dictionaries.
PT
GL IT
CA FR
EO
RO ES
EN
EU We devised a heuristic that focuses on producing high precision entries, even though the recall might su er. The model builds a graph with all the bilingual word-POS pairs in all the Apertium dictionaries that can be used as a pivot between Portuguese, French and English, that is, all the language pairs with thick lines in Figure 1. Whenever we nd a length 4 cycle in the graph, we connect all the nodes in the cycle. All discovered edges for the respective language pair are used as dictionary. Figure 2 shows an example of discovered translations. Similar to the cycle-based method, we use another heuristic technique, which aims to create translation candidates by traversing the paths between the source and the target languages in the Apertium language graph. The candidate translations T are weighted with respect to the path length and the frequency. In this
section, language graph refers to the Apertium dictionary graph (Figure 1) and translation graph refers to a graph where vertices represent a word and edges represent the translations in other languages. Figure 3 illustrates the translation graph of the word spring as a noun in English based on the language path English!Basque!Spanish!French!Esperanto!Catalan!Portuguese.
The basic idea behind pivot-oriented translation inference is the transitivity
assumption of translations. If wp, a pivot word in the dictionary Dp, has the
translation equivalents wi and wj in dictionaries Dp!1 and Dp!2 respectively,
then wi and wj may be equivalents in the D1!2 dictionary. Although the
pivotoriented approach can mostly create accurate translations for monosemous words
(depending on the lexicon completeness), this oversimpli es for polysemous words
leading to incorrect translations [
For the current task, we have considered all the simple paths, i.e. paths without any repeating vertex, starting from and ending with the goal languages of the shared task, of which there are 92, 67 and 58 simple paths between Portuguese!French, French!English and English!Portuguese, respectively. As the language graph is undirected, the paths between the vertices are identical in each direction. For instance, the dictionary paths from English to Portuguese are the same as those from Portuguese to English.
In order to describe the likelihood of a translation being correct, we introduce the following weighting factor: wt = f requency(t) l (1) where wt is the weight of the translation candidate t 2 T , f requency(t) is the number of times translation t is reached, 2 (0, 1) is a penalisation constant and l is the length of the path leading to wt. l penalises the translation candidates in a way that paths of lower length and higher frequency get a lower weight. On the other hand, a longer translation path results in a lower weight factor. For instance, in the translation graph in Figure 3, the frequency of the words primavera, font and origem is respectively 2, 1 and 1 and the length of their translation path is 7. For the current task, we set = 0:5 and have included the part-of-speech tags in the inference.
Finally, the weights are normalised such that Pt2T wt = 1.
iturri fuente source fonto origen font brollador origem fonte primavero primavera
primavera printempo spring udaberri primavera
printemps malguki
muelle English (en) Basque (eu)
Spanish (es)
French (fr) Esperanto (eo)
Catalan (ca) Portuguese (pt)
To perform experiments on NMT models with a minimal set of parallel data, i.e.
for less-resourced languages, we trained a multi-source and multi-target NMT
model [
For the continuous training of the multi-way model, we experimented with two di erent datasets: each one of the dictionaries generated following the cycle (Section 2.1) and path (Section 2.2) strategies.
Filtering The NMT models were trained without POS information; hence, the NMT model is unable to assign a POS tag to each predicted translation. As the 3 Trained on English-Spanish, French-Romanian and Italian-Portuguese parallel data 25;594
TIAD 2019 shared task required POS information, we used a further ltering strategy. We used monolingual Portuguese, French and English dictionaries extracted from Wiktionary, and, for each translation prediction, we examined if the word exists in both dictionaries with the same POS tag.4 One entry is generated for each shared POS tag.
For example, targeting English to Portuguese, when we request the translation for snake the system generates serpente; snake appears both as a noun and a verb in the English Wiktionary, but serpente only appears as a noun in the Portuguese Wiktionary. Hence, we generate an entry in the English-Portuguese dictionary where the noun snake has serpente as a possible translation. As a side e ect this approach removes those words that are incorrectly generated due to the use of subword units, and in ected words (i.e. words not in a canonical form) that are seldom included in dictionaries. 2.4
Datasets For the cycle (Section 2.1) and path (Section 2.2) strategies, we used the Apertium dictionaries5 shown in Figure 1.
In order to train the multi-way model described in Section 2.3, we used the
Directorate-General for Translation (DGT) corpus [
Neural Machine Translation Toolkit
We used OpenNMT [
framework. Due to computational complexity, the vocabulary in NMT models
had to be limited. In order to overcome this limitation, we used byte pair encoding
(BPE) to generate subword units [
The Translation Inference Across Dictionaries 2019 Shared Task focuses on the automatic generation of dictionary entries in order to enrich knowledge graphs with multilingual knowledge. The task focuses on the generation of French-English, Portuguese-English and French-Portuguese dictionaries. The organisers recommended the use of the Apertium RDF dataset6 (Figure 1 shows the languages in Apertium RFT used by our submission), although the usage of other resources, other than parallel data between the targeted languages, was permitted.
Evaluation was carried out by the organisers of the task, using manually compiled pairs of K Dictionaries as a gold standard. Precision, recall, F-measure and coverage were reported and participants submitted 11 di erent submissions to the task. The averaged results for all the di erent models can be seen in Table 2 and the in-depth results for our submitted models can be seen in Table 3. 4
Despite the fact that we are using a straightforward approach to extract candidates from Apertium RDF, the precision of our models rank among the highest for the
task. We demonstrate the multi-way NMT approach, which generates translations without requiring any parallel data between the targeted languages. Additionally, when using the path or NMT approaches to generate entries, a threshold can be modi ed to obtain dictionaries with higher recall and lower precision or vice-versa, what can be used to adapt the method for di erent use cases.
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 Union's Horizon 2020 research and innovation programme under grant agreement No 731015, ELEXIS - European Lexical Infrastructure.
U 4 C E N U :8 :4 :2 :1 :1 :1 :1 :1 :1 :1 :1 P
P :1 :2 :2 :2 :2 :2 :2 :2 :2 :2 R
N
L 0 0 0 0 0 0 0 0 0 0 0 r o ie :1 :1 :1 :1 :1 :1 :1 :1 :1 :1 :1 C ld d 3 3 3 3 3 3 3 3 3 3 3 c T ov ab e ra le g e 3
. ) e
s f u ro lt th th e e .) b y l o :7 :7 :7 :7 :7 :7 :7 :7 :7 :7 :7 P l f 5 5 5 5 5 5 5 5 5 5 5
N 0 0 0 0 0 0 0 0 0 0 0 d e i v :0 :0 :0 :0 :0 :0 :0 :0 :0 :0 :0 R LP
7 7 7 7 7 7 7 7 7 7 7 t i 0 0 0 0 0 0 0 0 0 0 0 F Y
o :11 :11 :11 :11 :11 :11 :11 :11 :11 :11 :11 1 C th n
L :0 :2 :1 :1 :1 :1 :1 :1 :1 :1 :1 1 3
P
0 0 0 0 0 0 0 0 0 0 0 ,) :8 :3 :1 :1 :1 :1 :0 :0 :0 :0 :0 P U
9 9 9 9 9 N :1 :2 :2 :2 :2 :2 :2 :2 :2 :2 R -N 6 1 3 4 4 4 4 4 4 4 M :2 :1 :1 :1 :1 :1 :1 :1 :1 :1 1 C
Y 1 6 4 4 3 3 3 3 3 3 e a s u 0 0 0 0 0 0 0 0 0 0 0 r e 3 4 1 1 1 1 1 1 1