Recent surveys consider linguistic typology as a potential source of knowledge to support multilingual natural language processing (NLP) tasksohoran-etal-2016-survey; ponti-etal-2019-modeling. Linguistic typology studies language variation in terms of their functional processes comrie1989language. Several typological knowledge bases (KB) have been crafted, from where we can extract categorical language features littell-etal-2017-uriel. Nevertheless, their sparsity and reduced coverage present a challenge for an end-to-end integration into NLP algorithms. For example, the World Atlas of Language Structure (WALS; wals) encodes 143 features for 2,679 languages, but their mean coverage per language is barely around 14%.
Dense and data-driven language representations have emerged in response. They are computed from multilingual settings of language modelling ostling-tiedemann-2017-continuous
and neural machine translation (NMT)malaviya-etal-2017-learning. However, the language diversity in the corpus-based representations is limited. The language coverage could be broadened with other knowledge, such as that encoded in WALS, to distinguish even more language properties. Therefore, to obtain the best of both views (KB and task-learned) with minimal information loss, we project a shared space of discrete and continuous features using a variant of canonical correlation analysis svcca-NIPS2017_7188.
For our study, we fuse language-level embeddings from multilingual machine translation with syntactic features of WALS. We inspect how much typological knowledge is present by predicting features for new languages. Then, we infer language phylogenies and inspect whether specific relationships are induced from the task-learned vectors.
Furthermore, to demonstrate that our approach has practical benefits in NLP, we apply our language vectors in multilingual NMT with language clustering tan-etal-2019-multilingual and adapt the ranking of related languages for multilingual transfer lin-etal-2019-choosing. As a side outcome, we identify that there is an ideal setting to encode language relationships in language embeddings from NMT. Finally, we present a simple tool to allow everyone to fuse, extend and compare their own representations111Once the paper is published, we will release the code at: https://github.com/aoncevay/multiview-langrep.
2 Multi-view language representations
Our primary goal is to fuse parallel representations of the same language in one shared space, and canonical correlation analysis (CCA) allows us to find a projection of two views for a given set of data. With CCA, we look for linear combinations that maximise the correlation of the two sources in each coordinate iteratively hardoon2004canonical. After training, we can apply the transformation learned on a new sample from any view to obtain a CCA-based language representation222With language representations, we refer to an annotated or unsupervised characterisation of a language itself (e.g. Spanish or English), and not to word or sentence-level representations, as it is used in the recent NLP literature..
CCA considers all dimensions of the two views as equally important. However, our sources are potentially redundant: KB features are mostly one-hot-encoded, whereas task-learned ones inherit the high dimensionality of the embedding layer. Moreover, few samples and sparsity could make the convergence harder. For the redundancy issue,singular value decomposition
(SVD) is an appealing alternative. With SVD, we factorise the source data matrix to compute the principal components and singular values. Furthermore, to deal with sparsity, we adopt a truncated SVD approximation, which is also known as latent semantic analysis in the context of linear dimensionality reduction for term-count matricesdumais2004latent.
The two-step transformation of SVD followed by CCA is called singular vector canonical correlation analysis (SVCCA; svcca-NIPS2017_7188)
in the context of understanding the representation learning throughout neural network layers. That being said, we use SVCCA to get language representations and not to inspect a neural architecture333
As the SVD step performs a dimensionality reduction while preserving the most explained variance as possible, we can consider two additional parameters: a threshold value in the [0.5,1.0] range with 0.05 incremental steps, for the explained variance ratio of each view. With a value equal to 1, we bypass SVD and compute CCA only. We then tuned all our following experiments (see AppendixC for details)..
3 Methodology and research questions
To embed linguistic typology knowledge in dense representations for a broad set of languages, we employ SVCCA (§2) with the following sources:
We employ the language vectors from the URIEL and database littell-etal-2017-uriel. Precisely, we work with the -NN vectors of the Syntax feature class (; 103 feats.), that are composed of binary features encoded from WALS wals.
(NMT) Learned view.
Firstly, we exploit the NMT-learned embeddings from the Bible (; 512 dim.) malaviya-etal-2017-learning. Up to 731 entries are available in that intersects with . They were trained in a many-to-English NMT model with a pseudo-token identifying the source language at the beginning of every input sentence.
Secondly, we take the many-to-English language embeddings learned for the language clustering task on multilingual NMT (; 256 dim.) tan-etal-2019-multilingual, where they use 23 languages of the WIT corpus cettoloEtAl:EAMT2012.
One main difference for the latter is the use of factors in the architecture, meaning that the embedding of every input token was concatenated with the embedded pseudo-token that identifies the source language. The second difference is the neural architecture used to extract the embeddings: the former use a recurrent neural network, whereas the latter a small transformer modelNIPS2017_7181.
Finally, we train a new set of embeddings () that we extracted from the 53 languages of the TED corpus (many-to-English) processed by qi-etal-2018-pre, using the approach of tan-etal-2019-multilingual444We prefer to use factored embeddings over initial pseudo-tokens as we identified that there is a difference for encoding information about language similarity (see §7)..
What knowledge do we represent?
Each source embeds specialised knowledge to assess language relatedness. The KB vectors can measure typological similarity, whereas task-learned embeddings correlates with other kinds of language relationships (e.g. genetic) Bjerva2019-bl. To analyse whether each kind of knowledge is induced with SVCCA, we assess the tasks of typological feature prediction (§4) and reconstruction of a language phylogeny (§5).
What is the benefit for multilingual NMT (and NLP)?
Language-level representations can evaluate the distance between languages in a vector space. We then can assess their applicability on multilingual NMT tasks that require guidance from language relationships. Therefore, language clustering and ranking related partner languages for (multilingual) transfer are our study cases (§6).
4 Prediction of typological features
An example of a typological feature is a word order specification, like whether the adjective is predominately placed before or after the noun (features #24 and #25 of ). Our task consists in predicting syntactic features () leaving one-language and one-language-family out to control phylogenetic relationships bjerva-etal-2019-probabilistic. Previous work has shown that task-learned embeddings are potential candidates to predict features of a linguistic typology KB malaviya-etal-2017-learning, and our goal is to evaluate whether SVCCA can enhance the NMT-learned language embeddings with typological knowledge from their KB parallel view.
In Table 1, we observe that SVCCA outperformed their NMT-learned counterparts for and , where the performance is significantly better for the one-language-out setting. In the case of (with 731 entries), we notice that the overall performance drops, and the SVCCA transformation cannot improve it. We argue that a potential reason for the accuracy dropping is the method used to extract the NMT-learned embeddings (initial pseudo-token instead of factors: §7), which could diminishes the information embedded about each language, and consequently, impacts the SVCCA projection. In conclusion, we notice that specific typological knowledge is usually hard to learn in an unsupervised way, and fusing them with KB vectors using SVCCA is feasible for inducing information of linguistics typology in some scenarios.
5 Language phylogeny analysis
According to Bjerva2019-bl, there is a positive correlation between the language distances in a phylogenetic tree and a pairwise distance-matrix of task-learned representations. Our goal therefore is to investigate whether fusing linguistic typology with SVCCA can preserve or enhance the embedded relationship information. For that reason, we first examine how well a language phylogeny can be reconstructed from language representations (§5.1), and study the correlation afterwards (§5.2).
5.1 Inference of a phylogenetic tree
Based on previous work rabinovich-etal-2017-found, we take a tree of 17 Indo-European languages Serva2008-fb as a Gold Standard (GS), which is shown in Figure 1a. We also use agglomerative clustering with variance minimisation ward1963hierarchical
as linkage, but we employ cosine similarity asBjerva2019-bl. We also consider a concatenation () of the KB and NMT-learned views as a baseline.
It is essential to highlight that none of the NMT-learned and vectors have all the 17 language entries of the GS. Therefore, we can quickly preview one of the significant advantages of the SVCCA vectors, as we are able to represent “unknown” languages using one of the views. The NMT-learned views lack English, since they were extracted from the source side of a many-to-English system, but we were able to project the KB English vectors into the shared space. In addition, we project other four languages (Swedish, Danish, Latvian, Lithuanian) to complete the embeddings of tan-etal-2019-multilingual and Latvian to complete our own set.
We differ from previous studies and use a tree edit distance metric, which is defined as the minimum cost of transforming one tree into another by inserting, deleting or modifying (the label of) a node. Specifically, we used the All Path Tree Edit Distance algorithm (APTED; pawlik-augsten-2015-efficient; pawlik-augsten-2016-tree), a novel one for the task. We chose an edit-distance method as it is more transparent for assessing what is the degree of impact for a single change of linkage in the GS.
As we need to compare inferred pruned trees with different number of nodes, we propose a normalised version given by: , where is the inferred tree, and indicates the number of nodes. The denominator then is the maximum cost possible of deleting all nodes of and inserting each GS node.
|(Syntax)||30 / 0.45|
|(Bible)||35 / 0.54||27 / 0.42||23 / 0.34|
|(WIT-23)||35 / 0.62||23 / 0.41||27 / 0.48|
|(TED-53)||15 / 0.26||18 / 0.29||10 / 0.15|
Table 2 shows the results for all settings, where the single-view scores are meagre in most of the cases. For instance, the inferred tree (Fig.1c) requires 30 editions to resemble the GS. The exception is (Fig.1d), which requires half the editions, although it is incomplete.
We observe that the best absolute and normalised scores are obtained by fusing and with SVCCA (Fig.1b). English is projected in the Germanic branch, although Latvian is separated from the Balto-Slavic group. The latter case is similar for Bulgarian, which is misplaced in the original tree as well. Nevertheless, we only require ten editions to equate the GS (where 66 is the maximum cost possible), confirming that our approach is a robust alternative for completing language entries and inferring a language phylogeny. We then proceed to discuss what kind of relationship we are representing.
5.2 Correlation with lexical similarity
Bjerva2019-bl argued that raw language embeddings from language modelling correlates with genetic and structural similarity555We note that Bjerva2019-bl used monolingual texts translated from different languages to investigate what kind of genetic information is preserved. Concerning structural similarity, they computed a distance matrix using syntax-dependency-tags counts per language from annotated treebanks. We leave this analysis for further work.. For the former, they correlated a distance matrix with pairwise-leaf-distances of the GS. However, Serva2008-fb originally inferred the phylogeny by comparing the translated Swadesh list of 200-words dyen1992indoeuropean with Levenshtein (edit) distance. The list is a crafted set of concepts for comparative linguistics (e.g. I, eye, sleep), and it is usually processed by lexicostatistics methods to study language relationship through time. Therefore, we prefer to argue that corpus-based embeddings could partially encode lexical similarity of languages.
We perform an Spearman correlation between the cophenetic matrix666 Pairwise-distances of the hierarchy’s leaves (languages). of the GS and the pairwise cosine-distance matrices of , and SVCCA(,), where we obtain correlation coefficients of 0.48, 0.68 and 0.80, respectively (p-values0.001). Our conclusion is that typological knowledge strengthen the representation of lexical similarity within NMT-learned embeddings.
6 Application in multilingual NMT
With multilingual NMT, we can translate several language-pairs using a single model. Low-resource languages are usually benefited through multilingual transfer, which resembles a simultaneous training of the parent(s) and child models. Therefore, we want to take advantage of a language-level vector space for relating similar languages and enhancing multilingual transfer within multilingual NMT. For that reason, we first address the language clustering task proposed by tan-etal-2019-multilingual, and afterwards, the language ranking model of lin-etal-2019-choosing.
The main idea is to obtain smaller multilingual NMT models as an intermediate point between maintaining many pairwise systems and a single massive multilingual model. With limited resources, it is challenging to support the first scenario, whereas the advantages for the massive setting are also very appealing (e.g. simplified training process, translation improvement for low-resource languages or zero-shot translation johnson-etal-2017-googles). Therefore, to address the task, tan-etal-2019-multilingual trained a factored multilingual NMT model of 23 languages from cettoloEtAl:EAMT2012
, where the language embedding is concatenated in every input token. Then, they performed hierarchical clustering with the representations, and selected a number of clusters guided by the Elbow method. Finally, they compared the systems against individual, massive and language family-based cluster models.
Rather than only using our multi-view representations to compute a set of clusters, we also address the question: do we need to train the massive model again if we want to add one or more new languages to our setting? If one of the goals of working with clustered NMT models is to avoid training and maintaining massive systems, it is quite a significant problem for the new language scenario.
The original goal of LangRank
is to choose a parent language to perform transfer learning in different tasks, NMT included. To achieve this,lin-etal-2019-choosing trained a model based on the performance of several hundred pairwise MT systems using the dataset of qi-etal-2018-pre. For the input features, they considered linguistically-informed vectors from littell-etal-2017-uriel and corpus-based statistics, such as word/sub-word overlapping and the ratio of the token-types or the data size between the target child and potential candidates, where the latter features were one of the most relevant.
Considering the transfer capabilities within multilingual NMT and the possibility to obtain a ranked list of candidates from LangRank, we propose an adapted task of choosing -related languages for multilingual transfer. We then use our multi-view representations to rank related languages from the vector space, as they embed information about typological and lexical relationships. This is similar to the features that lin-etal-2019-choosing considers, but without training a ranking model fed with scores from pairwise MT systems.
6.1 Experimental setup
We focus on the many-to-one (English) multilingual NMT setting to simplify the findings in both tasks. However, similar experiments could be performed in a one-to-many direction. Details about models, training and inference are described in Appendix B.
We use the dataset processed and tokenised by qi-etal-2018-pre of 53 languages (TED-53), from where we learned our embeddings. We opted for TED-53 to better evaluate the extensibility of clusters and because it is also used to train the LangRank model. The list of languages, set sizes and other details are included in Appendix A. Before preprocessing the text, we drop any sentences from the training sets which overlap with any of the test sets. Since we are building many-to-English multilingual systems, this is important, as any such overlap will bias the results.
We first list the baselines and our approaches, with the number of clusters/models between brackets:
Individual : Pairwise model per language.
Massive : A single model for all languages.
Language families : Based on historical linguistics. We divide the 33 Indo-European languages into 7 branches. Moreover, 11 groups only have one language.
KB : (Syntax) tends to agglomerate large clusters (with 4-13-33 languages) and behaves similar to the massive model.
Learned : We train a set of 53 factored embeddings () similar to tan-etal-2019-multilingual.
SVCCA-53 : Multi-view representation with SVCCA composing both and vectors. Figure 2 shows the inferred hierarchy.
SVCCA-23 : Similar to the previous setting, but we use the set of 23 language embeddings instead tan-etal-2019-multilingual, and project the 30 complementary languages with SVCCA(,).
With the last setting, we are interrogating whether SVCCA is a useful method for rapidly increasing the number of languages without retraining massive models given new entries that require their NMT-learned embeddings for clustering.
Similar to tan-etal-2019-multilingual, we use hierarchical agglomeration with average linkage and cosine similarity. However, we choose a different criterion for choosing the optimal number of clusters.
Selection of number of clusters.
The Elbow criteria has been suggested for this purpose tan-etal-2019-multilingual; however, as we can see in Figure 2
, it might be ambiguous. Thus, we propose using a heuristic called Silhouetterousseeuw1987silhouettes, which returns a score in the [-1,1] range. A sample cluster with a silhouette close to 1 indicates that it is cohesive and well-separated. With the average silhouette of all samples, we vary the number of cluster partitions, and look for the peak value.
We focus on five low-resource languages from TED-53: Bosnian (bos, Indo-European/Balto-Slavic), Galician (glg, Indo-European/Italic), Malay (zlm, Austronesian), Estonian (est, Uralic) and Georgian (kat, Kartvelian). They have between 5k and 13k translated sentences with English, and we chose them as they achieved the most significant improvement from the individual to the massive setting. We then identified the top-3 related languages using LangRank, which give us a multilingual training set of around 500 thousand sentences for each case. Given that LangRank usually prefers to choose candidates with larger data size lin-etal-2019-choosing, for a fair comparison, we use SVCCA and cosine similarity to choose the closest languages that can agglomerate a similar amount of parallel sentences.
6.2 Language clustering results
We first briefly discuss the composition of clusters obtained by SVCCA. Then, we analyse the results grouped by training size bins. We complement the analysis by family groups in Appendix D.
In Figure 2, we observe that SVCCA-53 has adopted ten clusters with a proportionally distributed number of languages (the smallest one is Greek-Arabic-Hebrew, and the largest one has seven entries). Moreover, the languages are usually grouped by phylogenetic or geographical criteria.
From a more detailed inspection, there are entries that do not correspond to their respective family branches, although the single-view sources might induce the bias. For instance, the tree (Fig1d) “misplaced” Bulgarian within Italic languages. Nevertheless, the unexpected agglomerations rely on the features encoded in the KB or the NMT learning process, and we expect they can uncover surprising clusters to avoid isolating languages without close relatives (e.g. Basque, or even Japanese as the only Japonic member in the set).
Training size bins:
We manually define the upper bounds of the bins as [10,75,175,215] thousands of training sentences, which results in groups composed by [14,14,13,12] languages. Figure 3 shows the box plots of BLEU from where we can analyse each distribution (mean, variance).
Throughout all the bins, we observe that both SVCCA-53 and SVCCA-23 accomplish a comparable accuracy with the best setting in each group. In other words, their clusters provide stable performance for both low or high-resource languages.
In the first bin of the smallest corpora, the Massive baseline and the large clusters of barely surpass the SVCCA schemes. Nevertheless, SVCCA contributes a notable advantage if we want to train a multilingual NMT model for a specific low-resource language, and we do not have the resources for training a massive system. We further analyse this scenario in §6.3.
In the rightmost bin, for the highest resource languages, the Massive and performed worse than SVCCA. Furthermore, we show a competitive accuracy for the Individual and Family approaches. The former’s clusters have steady performance across most of the bins as well. Nevertheless, they double the number of clusters that we have in both SVCCA settings, and with more than half of the “clusters” having only one language.
Other approaches, like using the NMT-learned embeddings () as tan-etal-2019-multilingual or the concatenation baseline, obtain similar translation results in the last three bins. However, we need to obtain the NMT-learned embeddings first in order to fulfil those methods (from a 53-languages massive model). Using SVCCA and a pre-trained smaller set of language embeddings is enough for projecting new representations, as we present with our SVCCA-23 approach.
6.3 Language ranking results
After discussing overall translation accuracy for all the languages, we now focus on five specific low-resource cases and how multilingual transfer enhance their performance. Table 3 shows the BLEU scores of the translation into English for the smaller multilingual models that group each child language with their candidates ranked by LangRank and our SVCCA-53 representations.
We also include the results of the individual and massive MT systems. Even when the latter baseline provides a significant improvement over the former, we observe that many of the smaller multilingual models outperform the translation accuracy of the massive system. The result suggests that the amount of data is not the most important confound for supporting multilingual transfer in a low-resource language.
Comparing the two ranking approaches, we observe that SVCCA achieves a comparable performance in most of the cases. We note that LangRank prefers related languages with large datasets, as it only requires three candidates to group around half a million training samples, whereas SVCCA suggests to include from three to ten languages to reach a similar amount of parallel sentences. However, increasing the number of languages could impact the multilingual transfer negatively (see the case of Georgian or kat), and it is analogous to adding different “out-of-domain” samples. To alleviate this, we could bypass candidate languages that do not possess a specific amount of training samples.
We argue that our method provides a robust alternative to determine which languages are suitable for multilingual transfer learning. A notable advantage is that we do not need to pre-train MT systems from a specific dataset, and we can easily extend the coverage of languages without re-training the ranking model to consider new entries777However, we do not answer what multilingual NMT really transfers to the low-resource languages. We left that question for further research, together with optimising the number of languages or the amount of data per each language..
|bos||4.2||26.6||28.8 (434)||28.2 (L=5)|
|glg||8.4||24.9||27.7 (443)||28.4 (L=3)|
|zlm||4.1||20.1||21.2 (463)||21.0 (L=4)|
|est||5.8||13.5||13.5 (533)||12.1 (L=6)|
|kat||5.8||14.3||13.3 (499)||10.5 (L=10)|
7 Factors over initial pseudo-tokens
We additionally argue that the configuration used to compute the language embeddings impacts what relationship they can learn. For the analysis, we extract an alternative set of 53 language embeddings () but using the initial pseudo-token setting instead of factors. Then, we perform a silhouette analysis to identify whether we can build cohesive and well-separated clusters of languages.
Figure 4 shows the silhouette analysis for the aforementioned embeddings () together with the Bible embeddings () that were trained with the same configuration. We observe that the silhouette score never exceeds 0.2, and the curve keeps degrading when we examine a higher number of clusters, which contrast the trend shown in Figure 2. The pattern proves that the vectors are not suitable for clustering (the hierarchies are shown in Figure 6 in the Appendix), and they might only encode enough information to perform a classification task in the multilingual NMT training and inference. For that reason, we consider it essential to use language embeddings from factors for extracting language relationships.
8 Related work
For language-level representations, URIEL and littell-etal-2017-uriel allow a straightforward extraction of typological binary features from different KBs. murawaki-2015-continuous; murawaki-2017-diachrony; murawaki-2018-analyzing
exploits them to build latent language representations with independent binary variables. Language features are encoded from data-driven tasks as well, such as NMTmalaviya-etal-2017-learning or language modelling tsvetkov-etal-2016-polyglot; ostling-tiedemann-2017-continuous; bjerva-augenstein-2018-tracking with complementary linguistic-related target tasks bjerva-augenstein-2018-phonology.
Our approach is most similar to bjerva-etal-2019-probabilistic, as they build a generative model from typological features and use language embeddings, extracted from factored language modelling at character-level, as a prior of the model to extend the language coverage. However, our method primarily differs as it is mainly based in linear algebra, encodes information from both sources since the beginning, and can deal with a small number of shared entries (e.g. 23 from ) to compute robust representations.
There has been very little work on adopting typology knowledge for NMT. There is not a deep integration of the topics ponti-etal-2019-modeling, but one shallow and prominent case is the ranking method lin-etal-2019-choosing that we analysed in §6.
Finally, CCA and its variants have been previously used to derive embeddings at word-level faruqui-dyer-2014-improving; dhillon2015eigenwords; osborne-etal-2016-encoding. kudugunta-etal-2019-investigating also used SVCCA but to inspect sentence-level representations, where they uncover relevant insights about language similarity that are aligned with our results in §5. However, as far as we know, this is the first time a CCA-based method has been used to compute language-level representations.
9 Takeaways and practical tool
We summarise our key findings as follows:
SVCCA can fuse linguistic typology KB entries with NMT-learned embeddings without diminishing the originally encoded typological and lexical similarity of languages.
Our method is a robust alternative to identify clusters and choose related languages for an small-scale multilingual transfer in NMT. The advantage is notable when it is not feasible to pre-train a ranking model or learn embeddings from a massive multilingual system.
Factored language embeddings encodes more information to agglomerate related languages than the initial pseudo-token setting.
Furthermore, we release a tool to compute language representations using SVCCA, together with our vectors. It is possible to use language vectors from KBs (e.g. contains features from Phonology or Phonetic Inventory) or task-learned embeddings from different settings, such as one-to-many or many-to-many NMT and multilingual language modelling. Besides, we could rapidly project new language representations to assess tasks like clustering or ranking candidates for multilingual NMT (and NLP) that involves massive datasets of hundreds of languages.
We compute multi-view language representations with SVCCA using two sources: KB and NMT-learned vectors. We investigated that the knowledge contained in each source (typological and lexical similarity) is preserved in the combined representation. Moreover, our approach offers important advantages because we can evaluate projected languages with entries in only one of the views. The benefits are noticeable in multilingual NMT tasks, like language clustering and ranking related languages for multilingual transfer. We plan to study how to deeply incorporate our typologically-enriched embeddings in multilingual NMT, where there are promising avenues in parameter selection sachan-neubig-2018-parameter and generation platanios-etal-2018-contextual.
[image=true, lines=2, findent=1ex, nindent=0ex, loversize=.15]figs/eu-logo.pngThis work was supported by funding from the European Union’s Horizon 2020 research and innovation programme under grant agreements No 825299 (GoURMET) and the EPSRC fellowship grant EP/S001271/1 (MTStretch). Also, it was performed using resources provided by the Cambridge Service for Data Driven Discovery (CSD3) operated by the University of Cambridge Research Computing Service (http://www.csd3.cam.ac.uk/), provided by Dell EMC and Intel using Tier-2 funding from the Engineering and Physical Sciences Research Council (capital grant EP/P020259/1), and DiRAC funding from the Science and Technology Facilities Council (www.dirac.ac.uk). We express our thanks to Kenneth Heafield, who provided us with access to the computing resources.
Appendix A Languages and individual BLEU scores
We work with 53 languages pre-processed by qi-etal-2018-pre, from where we mapped the ISO 639-1 codes to the ISO 693-2 standard. However, we need to manually correct the mapping of some codes to identify the correct language vector in the URIEL littell-etal-2017-uriel library:
zh (zho , Chinese macro-language) mapped to cmn (Mandarin Chinese).
fa (fas , Persian inclusive code for 11 dialects) mapped to pes (Western/Iranian Persian).
ar (ara , Arabic) mapped to arb (Standard Arabic).
We disregard working with artificial languages like Esperanto (eo) or variants like Brazilian Portuguese (pt-br) and Canadian French (fr-ca).
Table 4 presents the list of all the languages with the following details: ISO 693-2 code, language family, size of the training set in thousands of sentences (with their respective training size bin) and the individual BLEU score obtained per clustering approach and other baselines.
|BLEU score per approach|
|ISO||Language||Lang. family||Size (k)||Bin||Individual||Massive||Family||SVCCA-53||SVCCA-23|
Appendix B Model and training details
Similar to tan-etal-2019-multilingual, we train small transformer models NIPS2017_7181. We jointly learn 90k shared sub-words with the byte pair encoding sennrich-etal-2016-neural algorithm built in SentencePiece kudo-richardson-2018-sentencepiece. We also oversample all the training data of the less-resourced languages in each cluster, and shuffle them proportionally in all batches.
We use Nematus sennrich-EtAl:2017:EACLDemo only to extract the factored language embeddings from the TED-53 corpus (). Given the large number of experiments, we choose the efficient Marian NMT junczys-dowmunt-etal-2018-marian toolkit for training the rest of systems. With Marian NMT, we only use the basic pseudo-token setting for identifying the source language, as we did not need to retrieve new language embeddings after training. Besides, we allow the Marian NMT framework to automatically determine the mini-batch size given the sentence-length and available memory (mini-batch-fit parameter)
We train our models with up to four NVIDIA P100 GPUs using Adam optimiser kingma2014adam with default parameters () and early stopping at 5 validation steps for the cross-entropy metric. Finally, the sacreBLEU version string post-2018-call is as follows: BLEU+case.mixed+numrefs.1+smooth.exp +tok.13a+version.1.3.7.
|Lang. families||# L||Size (k)||Individual||Massive||Family||SVCCA-53||SVCCA-23|
|Number of clusters/models||53||1||20||3||11||18||10||10|
|(Syntax)||30 / 0.45 (0.5)|
|(Bible)||35 / 0.54 (0.9)||27 / 0.42 (0.70,0.55)||23 / 0.34 (0.70,0.75)|
|(WIT-23)||35 / 0.62 (0.8)||23 / 0.41 (0.75,0.95)||27 / 0.48 (0.50,0.95)|
|(TED-53)||15 / 0.26 (0.6)||18 / 0.29 (0.70,0.55)||10 / 0.15 (1.00,0.55)|
Appendix C SVD explained variance selection
To compute SVCCA, we transform each source space using SVD, where we can choose to preserve a number of dimensions that represents an accumulated explained variance of the original dataset. For that reason, we perform a parameter sweep between 0.5 and 1.0 using 0.05 incremental steps. For a fair comparison, we also transform the single spaces (KB or Learned) with SVD and look for the optimal threshold.
Prediction of typological features.
We selected a 0.5 threshold for the NMT-learned vectors of and , and 0.7 for . In case of the SVCCA representation, uses [0.75,0.70], whereas and employ [0.95,0.50] values. The parameter values are for both one-language-out and one-family-out settings. We can argue that there is redundancy in the NMT-learned embeddings, as the prediction of typological features with Logistic Regression always prefers a dimensionality-reduced version instead of the original data (threshold at 1.0).
Language phylogeny inference.
In Table 6, we report the optimal value for the SVD explained variance ratio in each single and multi-view (concatenation and SVCCA) setting.
Language clustering (and ranking).
We cannot perform an exhaustive analysis for the threshold of the explained variance ratio per view. As our main goal is to increase the coverage of languages steadily, we must determine what configuration allows a stable growth of the hierarchy.
We thereupon take inspiration from bootstrap clustering nerbonne2008projecting, and increase the number of language entries from few entries (e.g. 10) to 53 by resample bootstrapping using each of the source vectors: , and . Afterwards, we search for the threshold value that preserves a stable number of clusters given the peak silhouette value. Our heuristic looks for the least variability throughout the incremental bootstrapping (Fig. 5).
We found that 0.65 is the most stable value for , whereas 0.60 is the best one for both and , so we thereupon fix SVCCA-53 and SVCCA-23 to [0.65,0.6]. We also apply the chosen thresholds on the concatenation baseline for a fair comparison. In the single-view cases, the transformations with the tuned variance ratio do not overcome any non-optimised counterparts.
Analysis of the number of clusters (blue) and the ratio of number of clusters per total languages (red) given the chosen thresholds of explained variance ratio. We show the confidence interval computed from the bootstrapping, and we observe that the number of clusters is stable since 42 and 38 languages forand vectors, respectively.
Appendix D Language clustering results by language families
Following a guide for evaluating multilingual benchmarks anastasopoulos-2019-multilingual, we also group the scores by language families. Table 5 includes the overall weighted average per number of languages in each family branch. We observe that most of the approaches have obtained clusters with similar overall translation accuracy. The individual models are the only ones that significantly underperform. The poor performance is transferred to the Family baseline, as most of the groups contains only one language given the low language diversity of the dataset.
The vectors obtain the highest overall accuracy, mostly from their few large clusters (see Fig. 5(b)). Meanwhile, SVCCA-53 achieves the second-best overall result, by a minimal margin, and with 3 to 7 languages per cluster, which are usually faster to converge. Besides, the massive model, the embeddings and the concatenation baseline present a competitive achievement as well. However, the first requires more resources to train until convergence, whereas the last two need the 53 pre-trained embeddings from a previous massive system.
In contrast, SVCCA-23 is a faster alternative if we want to target specific new languages (see Fig. 5(a)). We only require a small group of language embeddings (e.g. of 23 entries) and project the rest with SVCCA and a set KB-vectors as a side view. For instance, if we need to deploy a translation model for Basque or Thai, we could reach a comparable or better accuracy to a massive model with the SVCCA-23 chosen clusters of only 3 (Arabic, Hebrew) or 5 (Chinese, Indonesian, Vietnamese, Malay) languages, respectively.