| { | |
| "paper_id": "2021", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T12:14:08.314783Z" | |
| }, | |
| "title": "Hierarchical Transformer for Multilingual Machine Translation", | |
| "authors": [ | |
| { | |
| "first": "Albina", | |
| "middle": [], | |
| "last": "Khusainova", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Innopolis University", | |
| "location": { | |
| "settlement": "Innopolis", | |
| "country": "Russia" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Adil", | |
| "middle": [], | |
| "last": "Khan", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Innopolis University", | |
| "location": { | |
| "settlement": "Innopolis", | |
| "country": "Russia" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Ad\u00edn", | |
| "middle": [ | |
| "Ram\u00edrez" | |
| ], | |
| "last": "Rivera", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "University of Campinas Campinas", | |
| "location": { | |
| "country": "Brazil" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Vitaly", | |
| "middle": [], | |
| "last": "Romanov", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Innopolis University", | |
| "location": { | |
| "settlement": "Innopolis", | |
| "country": "Russia" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "The choice of parameter sharing strategy in multilingual machine translation models determines how optimally parameter space is used and hence, directly influences ultimate translation quality. Inspired by linguistic trees that show the degree of relatedness between different languages, the new general approach to parameter sharing in multilingual machine translation was suggested recently. The main idea is to use these expert language hierarchies as a basis for multilingual architecture: the closer two languages are, the more parameters they share. In this work, we test this idea using the Transformer architecture and show that despite the success in previous work there are problems inherent to training such hierarchical models. We demonstrate that in case of carefully chosen training strategy the hierarchical architecture can outperform bilingual models and multilingual models with full parameter sharing.", | |
| "pdf_parse": { | |
| "paper_id": "2021", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "The choice of parameter sharing strategy in multilingual machine translation models determines how optimally parameter space is used and hence, directly influences ultimate translation quality. Inspired by linguistic trees that show the degree of relatedness between different languages, the new general approach to parameter sharing in multilingual machine translation was suggested recently. The main idea is to use these expert language hierarchies as a basis for multilingual architecture: the closer two languages are, the more parameters they share. In this work, we test this idea using the Transformer architecture and show that despite the success in previous work there are problems inherent to training such hierarchical models. We demonstrate that in case of carefully chosen training strategy the hierarchical architecture can outperform bilingual models and multilingual models with full parameter sharing.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Machine translation (MT) today is gradually approaching near-human quality; however, this holds true only when massive parallel corpus is available. As for the low-resource machine translation, the main way to improve it is to use additional data such as monolingual texts or parallel data in other languages. One of the ways to use the latter is to build a multilingual model: instead of training separate models for each translation direction, multiple parallel corpora can be combined to train a single model where languages can share some parameters to help each other learn. Given that languages have a lot in common, a properly organized parameter sharing strategy could compensate for the lack of training examples in low-resource pairs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Exploiting language relatedness can substantially improve translation quality (Tan et al., 2019) . The question is which architectural setup allows getting the most benefit from between-language commonalities. While there are different approaches to this problem, such as full parameter sharing (Johnson et al., 2017) or shared encoder side with language-specific decoders (Dong et al., 2015) , we find the recent approach by Azpiazu and Pera (2020) promising because it accounts for the degree of relatedness between languages in a multilingual model in a systematic way.", | |
| "cite_spans": [ | |
| { | |
| "start": 78, | |
| "end": 96, | |
| "text": "(Tan et al., 2019)", | |
| "ref_id": "BIBREF16" | |
| }, | |
| { | |
| "start": 295, | |
| "end": 317, | |
| "text": "(Johnson et al., 2017)", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 373, | |
| "end": 392, | |
| "text": "(Dong et al., 2015)", | |
| "ref_id": "BIBREF4" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Linguistic trees organize languages in hierarchies by the degree of kinship, and the same approach can be applied to multilingual machine translation models. The idea is to organize both encoder and decoder in a hierarchical fashion, reflecting the degree of relatedness between languages, such that the most related languages share the largest number of parameters. Figure 1 shows the outline of such model (will be explained in more detail in section 3.1).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 367, | |
| "end": 375, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Unlike Azpiazu and Pera (2020), who used Long Short Term Memory (LSTM) recurrent neural network as a basic architecture, we implemented this idea using the state-of-the-art Transformer (Vaswani et al., 2017) architecture. Our experiments demonstrated some persistent training problems associated with the hierarchical architecture-namely, the model is prone to early overfitting in low-resource directions leading to low translation scores. Given that one of the main goals of introducing multilingual models in general and the hierarchical model in particular is improving Figure 1 : A high-level view of a sample multilingual hierarchical model. Horizontal lines denote encoder/decoder blocks, vertical lines and arrows show points of merging/splitting, and subscripts clarify which languages share a particular encoder/decoder. BCP-47 language codes are given here and below (Phillips and Davis, 2009) . translation accuracy for low-resource directions, this problem is critical and needs a solution.", | |
| "cite_spans": [ | |
| { | |
| "start": 185, | |
| "end": 207, | |
| "text": "(Vaswani et al., 2017)", | |
| "ref_id": "BIBREF18" | |
| }, | |
| { | |
| "start": 878, | |
| "end": 904, | |
| "text": "(Phillips and Davis, 2009)", | |
| "ref_id": "BIBREF9" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 574, | |
| "end": 582, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "These are the main contributions of this work: 1. Testing the hierarchical multilingual translation model using the Transformer architecture;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "2. Identifying and analyzing problems related to the hierarchical nature of the model-namely, overfitting in low-resource directions;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "3. Suggesting and testing a number of solutions, which could be summarized as various forms of regularization.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "In this section, we review the existing approaches to parameter sharing in multilingual MT models to contrast them with the hierarchical approach.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Probably, the simplest form of parameter sharing was introduced by Johnson et al. (2017) , where all parameters are shared and identifier tokens are used to distinguish between languages. While being easy to implement, such model may require more resources to capture relationships between languages, as no architectural hints are given. Dong et al. (2015) views the translation problem as a multi-task one, with a shared encoder and separate decoders for each target language. Here the potential knowledge sharing between target languages is not considered. Firat et al. (2016) and Wang et al. (2019) introduce shared parameters between the encoder and decoder. This is an interesting idea that can be considered in the future. Sachan and Neubig (2018) propose a one-tomany model, where parameters are partially shared between the multiple decoders. This is similar in spirit to the hierarchical approach; however, instead of having shared parts between individual decoders, the idea is to build a hierarchy of decoders.", | |
| "cite_spans": [ | |
| { | |
| "start": 67, | |
| "end": 88, | |
| "text": "Johnson et al. (2017)", | |
| "ref_id": "BIBREF7" | |
| }, | |
| { | |
| "start": 338, | |
| "end": 356, | |
| "text": "Dong et al. (2015)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 559, | |
| "end": 578, | |
| "text": "Firat et al. (2016)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 583, | |
| "end": 601, | |
| "text": "Wang et al. (2019)", | |
| "ref_id": "BIBREF19" | |
| }, | |
| { | |
| "start": 729, | |
| "end": 753, | |
| "text": "Sachan and Neubig (2018)", | |
| "ref_id": "BIBREF10" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Many other recent works, such as and Fan et al. (2020) add language-specific components/layers to the decoder side of a model.", | |
| "cite_spans": [ | |
| { | |
| "start": 37, | |
| "end": 54, | |
| "text": "Fan et al. (2020)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "We observe that in these works languages on the encoder/decoder side of multilingual models are treated the same, regardless of the degree of their relatedness. In fact, Fan et al. (2020) do group languages by vocabulary overlap, nevertheless, in their model close languages may end up being in different groups. In the hierarchical model, however, the number of shared parameters directly depends on the degree of kinship between languages.", | |
| "cite_spans": [ | |
| { | |
| "start": 170, | |
| "end": 187, | |
| "text": "Fan et al. (2020)", | |
| "ref_id": "BIBREF5" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "Using trees to model genealogical relationships between different languages has historically been a common approach in linguistics (Schleicher and Schleicher, 1853) . Although this is not the only model, and there exist alternatives such as wave model (Schmidt, 1872), tree structure naturally suits for depicting origination of languages one from another over time.", | |
| "cite_spans": [ | |
| { | |
| "start": 131, | |
| "end": 164, | |
| "text": "(Schleicher and Schleicher, 1853)", | |
| "ref_id": "BIBREF12" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Approach", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "This expert knowledge of relationships between languages can be utilized when building a multilingual MT model. The current trend in natural language processing is to make models learn linguistic rules and patterns on their own, without explicitly guiding them. However, in this case, when supervision does not really cost anything and language relatedness information is readily available, this expert knowledge can let models train faster and manage parameter space better. Belinkov et al. (2017) found out that different encoder levels specialize in different language as-pects: lower-level representations better capture part-of-speech tags, and higher-level ones are better at grasping semantics.", | |
| "cite_spans": [ | |
| { | |
| "start": 476, | |
| "end": 498, | |
| "text": "Belinkov et al. (2017)", | |
| "ref_id": "BIBREF3" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Approach", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "Another recent observation by Kudugunta et al. (2019) showed that in multilingual MT models with a single shared encoder, representations of different languages cluster by language families, and as they move up the encoder, similarity between source languages increases.", | |
| "cite_spans": [ | |
| { | |
| "start": 30, | |
| "end": 53, | |
| "text": "Kudugunta et al. (2019)", | |
| "ref_id": "BIBREF8" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Approach", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "These facts taken together suggest that the encoder tries to find a common representation of different languages that initially cluster together by language families. As they move up the encoder, a model finds intermediate representations that smooth out the dissimilarities on different levels -morphological, syntactic, semantic.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Approach", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "This closely resembles the structure of phylogenetic trees, where connections on the bottom level mean the greatest linguistic similarity between languages, and connections on the highest level of a hierarchy indicate that languages are far from each other. See the example of such tree in Figure 2 . If we organize the architecture of a multilingual MT model according to language relationships in a phylogenetic tree, this will allow sharing parameters between languages on appropriate levels.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 290, | |
| "end": 298, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Approach", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "For example, we can take two very close languages, whose vocabularies overlap significantly. Since these languages are highly related, we assume that it will not take long to reduce them to the same representation. So, we \"combine\" them early on, i.e., introduce shared parameters for these languages in the first layers of the encoder. The third language, let us suppose, comes from the same family, but many base words are different. Still, sentence structure remains the same as in the first two languages. Therefore, we combine this language with the first two in the later stages (parameter sharing starts deeper in encoder layers), and so on. This parameter sharing strategy allows the economic utilization of parameter space and has the potential to lead to better, i.e., source-languageindependent representations of different languages on the output level of the encoder.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Approach", | |
| "sec_num": "3" | |
| }, | |
| { | |
| "text": "The principal idea is to organize languages on the encoder and decoder sides according to their linguistic similarity. Namely, each side of the model is built as a hierarchy that corresponds to how languages connect in phylogenetic trees: the closer Figure 2 : A fragment of the phylogenetic tree for Turkic languages (Savelyev and Robbeets, 2020) . two languages are, the more parameters they share.", | |
| "cite_spans": [ | |
| { | |
| "start": 318, | |
| "end": 347, | |
| "text": "(Savelyev and Robbeets, 2020)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 250, | |
| "end": 258, | |
| "text": "Figure 2", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Model Description", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "The example of the hierarchical model is given in Figure 1 . There are four source and five target languages connected through the chain of hierarchically organized encoders and decoders. Let us consider the encoder side. All source languages have their own respective encoders e ru , e en , e de , e az , and parameters of these encoders are not shared with any other language since they are intended to learn language-specific features. Deeper in the model, the outputs of some encoders are stacked together and passed to shared encoders (connections are denoted with arrows). These encoders (e en\u2212de , e ru\u2212en\u2212de ) are shared among two or more languages so that they can capture knowledge common to these languages. Finally, there is the last encoder e ru\u2212en\u2212de\u2212az that is shared by all source languages and that combines the outputs of all remaining encoders together.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 50, | |
| "end": 58, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Model Description", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "This architecture is designed to enable knowledge sharing between different languages on different levels. We see that English and German are connected first, as they both come from the Germanic branch of the Indo-European language family. Further in the model, the Russian language mixes in, since it belongs to a different branch of the same language family. Later on, we join these Indo-European languages with Azerbaijani, which comes from an entirely different Turkic language family. Logic on the decoder side is analogousmost dissimilar languages split first.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model Description", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "We suggest having the same number of parameters along any path from source to target. For example, if we compare ru-tt and az-en translation paths in Figure 1 , although the latter has considerably less shared parameters, their total number should be the same for both paths. That is why some encoder blocks are longer than others, pointing that, for instance, the number of parameters in e az should be the same as there are cumulatively in e ru and e ru\u2212en\u2212de .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 150, | |
| "end": 158, | |
| "text": "Figure 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Model Description", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "Overall, we can say that the model adheres to the framework by Azpiazu and Pera (2020), but there are important differences. We use a different underlying model-Transformer instead of LSTM, we do not limit the number of layers and do not prune language families. There are also differences in training procedure which will be explained in section 4.2. And, most importantly, we identify the problem specific to training hierarchical models and suggest the improved training techniques for such models, as will be described in section 4.3.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model Description", | |
| "sec_num": "3.1" | |
| }, | |
| { | |
| "text": "To test the hierarchical approach to parameter sharing we experimented with several hierarchical setups in multilingual models. Unlike Azpiazu and Pera (2020) who only test complex multi-source multi-target models, we decided to start with simple setups when the hierarchy is just on the one side of the model (encoder or decoder) and then increase the complexity to the general case.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We tested the hierarchical model's performance in three different cases:", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "1. Simple case with two related source languages and one target language, Figure 3 ;", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 74, | |
| "end": 82, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "2. Simple case with one source language and two related target languages;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "3. General case with several languages of different degree of relatedness both on the source and target side, Figure 4 .", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 110, | |
| "end": 118, | |
| "text": "Figure 4", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "For each of these cases we trained two hierarchical models using different language sets. Further, when we identified specific training problems of hierarchical models (described in section 4.3.2), we suggested and applied two variations to each hierarchical model. Each hierarchical model is compared to two baselines: bilingual models and multilingual model with full parameter sharing (will be also called full sharing model later). Comparing to bilingual models trained on the same parallel corpora helps to understand whether languages can learn from each other in the hierarchical model. Whereas contrasting hierarchical models to full sharing models, with single shared encoder and decoder for all languages, as in Johnson et al. (2017) , answers the question about the use of explicitly introducing the hierarchy into the model. In every multilingual model we trained there are low-resource and high-resource directions, and we explored whether low-resource ones can learn from high-resource ones, and which model structure better fits for this purpose.", | |
| "cite_spans": [ | |
| { | |
| "start": 722, | |
| "end": 743, | |
| "text": "Johnson et al. (2017)", | |
| "ref_id": "BIBREF7" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "We used four parallel corpora, all coming from the same JW300 dataset (Agi\u0107 and Vuli\u0107, 2019) to enable fair comparison within the same domain:", | |
| "cite_spans": [ | |
| { | |
| "start": 70, | |
| "end": 92, | |
| "text": "(Agi\u0107 and Vuli\u0107, 2019)", | |
| "ref_id": "BIBREF0" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "\u2022 Turkish-German (tr-de), 500k sentences;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "\u2022 Azerbaijani-German (az-de), 110k;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "\u2022 English-Polish (en-pl), 500k;", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "\u2022 German-Polish (de-pl), 110k.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Turkish and Azerbaijani languages are related and come from the Oghuz sub-branch of the Turkic language family. English, German, and Polish all belong to the Indo-European language family, however, English and German are more related as they both come from the Germanic branch, while Polish is under the Balto-Slavic branch.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "For training simple models with a hierarchy on one side (case 1 and case 2) we used tr-de and az-de corpora together (resulting models can be denoted as az-tr\u2192de ( Figure 3 ) and de\u2192az-tr). Similarly, en-pl and de-pl corpora resulted in en-de\u2192pl and pl\u2192en-de models.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 164, | |
| "end": 172, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "For general hierarchical models (case 3) all four corpora were used together and resulted in 2 models: az-tr-pl\u2192en-de ( Figure 4 ) and en-de\u2192az-tr-pl.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 120, | |
| "end": 128, | |
| "text": "Figure 4", | |
| "ref_id": "FIGREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "The sizes of corpora are different: tr-de and enpl are considered as high-resource language pairs, and az-de and de-pl are low-resource pairs.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "All corpora were filtered by maximum sentence length of 40 BPE tokens (Sennrich et al., 2016) . ", | |
| "cite_spans": [ | |
| { | |
| "start": 70, | |
| "end": 93, | |
| "text": "(Sennrich et al., 2016)", | |
| "ref_id": "BIBREF15" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Data", | |
| "sec_num": "4.1" | |
| }, | |
| { | |
| "text": "Our multilingual machine translation system implementation is based on the Transformer architecture (Vaswani et al., 2017) with the reduced number of parameters:", | |
| "cite_spans": [ | |
| { | |
| "start": 100, | |
| "end": 122, | |
| "text": "(Vaswani et al., 2017)", | |
| "ref_id": "BIBREF18" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "\u2022 d model = 128", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "\u2022 dff = 512", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "\u2022 num heads = 8", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "\u2022 dropout rate = 0.1 For all hierarchical and bilingual models there are 6 layers on the path from any source to any target language. Thus, any bilingual model consists of 3 encoder and 3 decoder layers. Case 1 model (with hierarchical encoder) consists of 1 layer in each of the individual encoders, 2 layers in the shared encoder, and 3 layers in the decoder. Adding more languages in the general hierarchical model results in more overall layers, but there are still 6 layers between any pair of source and target languages.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "To make the models with full parameter sharing comparable to hierarchical ones, we keep the overall number of layers in the encoder and decoder the same between the models. For example, the full parameter sharing model trained for comparison with case 1 hierarchical model has 4 encoder and 3 decoder layers (corresponds to 1+1+2 and 3). However, when we came to the general case with more layers, the full sharing model could not learn anything. So, for en-de\u2192az-tr-pl doubling batch size solved the problem, and for az-tr-pl\u2192en-de we had to decrease the number of layers in the encoder from 6 to 4.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Azpiazu and Pera (2020) did not mention whether they kept the sizes of the hierarchical model and their baselines comparable, which makes it difficult to analyze the results.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "Since training corpora are of different sizes, we oversampled low-resource corpora when training hierarchical models. For full parameter sharing models, however, it turned out that oversampling hurts their performance, so we did not apply it.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "As for the training procedure, one way to train the hierarchical model is to alternate between all translation directions in a system (as Azpiazu and Pera (2020)). In this case, the possible concern is that the model parameters may start to oscillate between these directions. Therefore, we decided to simultaneously feed data from all source languages to the respective encoders, stack representations at points where specific encoders merge into shared ones, and pass them through the chain of decoders down to individual decoders. We trained all models with batch size 128 until convergence (for 50 epochs maximum) and report best BLEU scores reached over epochs. When training hierarchical models, we stack full-size (128) batches of different language pairs in the encoder and split them in the decoder.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "To facilitate knowledge sharing in multilingual models we use shared vocabularies.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Model and Training", | |
| "sec_num": "4.2" | |
| }, | |
| { | |
| "text": "In this section, we present and discuss the results of the experiments.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Results and Analysis", | |
| "sec_num": "4.3" | |
| }, | |
| { | |
| "text": "To evaluate the hierarchical approach, we used two baselines: bilingual and full sharing models. We looked at several metrics. First, we calculated the average difference in BLEU scores between multilingual and bilingual models. To do so, we averaged across total 16 translation paths in all trained full sharing models, and, separately, in all hierarchical models. Second, we divided these translation paths to low-resource (8) and high-resource ones (8), and computed average BLEU scores difference across high-resource directions only and, third, across low-resource directions only. This information is visualized in Figure 5 . The height of the bars reflects the magnitude of the difference, and their direction indicates whether scores improve or degrade, compared to a bilingual baseline. For now, we are only interested in full sharing (Full) and basic hierarchical (Hie) models.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 621, | |
| "end": 629, | |
| "text": "Figure 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "On the left of the Figure 5 there are average differences in scores, and we see that full sharing models perform on average almost the same as bilingual ones, and hierarchical are even slightly worse (-0.29) . However, the comparison becomes more insightful if we look at high-resource and low-resource directions separately. Although on average full sharing models may seem to behave the same as bilingual, now it is clear that this actually happens because low-resource directions improve, learning from related parallel data (+1.28), and high-resource directions degrade (-1.26).", | |
| "cite_spans": [ | |
| { | |
| "start": 200, | |
| "end": 207, | |
| "text": "(-0.29)", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 19, | |
| "end": 27, | |
| "text": "Figure 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "Exactly the same pattern is observed for hierarchical models: low-resource directions do benefit from multilinguality (+1.31), and high-resource directions are hurt by it (-1.88), even more than in full sharing models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "So, we see two problems here: degradation of high-resource directions in multilingual models in general; and overall low performance of hierarchical models, compared to full sharing models. The first problem is a broad one, it was observed in earlier works on multilingual MT models (Firat et al., 2016) and thus is outside of the scope of this paper. However, the second problem is surprising, since this violates our assumption about the usefulness of the hierarchical organization of a multilingual model, and hence, needs an investigation.", | |
| "cite_spans": [ | |
| { | |
| "start": 283, | |
| "end": 303, | |
| "text": "(Firat et al., 2016)", | |
| "ref_id": "BIBREF6" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "It would be interesting to compare these results with Azpiazu and Pera (2020), however, the way they present their findings does not let us do it. They average BLEU scores across unique source languages, and there is no strict division to highresource and low-resource pairs. According to their results, the hierarchical model performs better than all other baselines, including bilingual and full sharing ones. However, in the GlobalVoices dataset (Tiedemann, 2012 ) they used the absolute majority of pairs are very low-resource. Thus, even if high-resource pairs' scores degrade in multilingual models, it would not be seen if for the majority of low-resource pairs scores do improve. The \"bad\" high-resource score will simply be lost among lowresource improvements. We cannot assert this is the case, but it certainly can be the case that high-resource pairs' scores in their hierarchical models also degrade. This question could have been clarified in Figure 6 of their paper, where they group language pairs by corpus size, but, unfortunately, they decided to omit results for language pairs with relatively big corpus size (more than 100k).", | |
| "cite_spans": [ | |
| { | |
| "start": 449, | |
| "end": 465, | |
| "text": "(Tiedemann, 2012", | |
| "ref_id": "BIBREF17" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 957, | |
| "end": 965, | |
| "text": "Figure 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "Similarly, it also can be the case that highresource pairs in their full sharing models (called \"one-to-one\" there) have higher scores than in hierarchical models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "Overall, it is hard to compare the results in this high/low-resource aspect because settings are very different. Namely, what is low-resource here (100k) could be considered high-resource there. Also, the distribution of high-resource and lowresource directions is different: in our case, half of the directions in multilingual models are highresource, in their case, they constitute a minority.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "So, to sum it up, the identified problems can be the case in Azpiazu and Pera (2020) too, so, our following findings may benefit the overall hierarchical framework they suggested.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Evaluating Hierarchical Model", | |
| "sec_num": "4.3.1" | |
| }, | |
| { | |
| "text": "The problem we identified and decided to explore here is that the hierarchical model does not outperform the full sharing one as it should have been according to our assumption. The hierarchical idea does not work as expected-it shows almost the same improvements for low-resource pairs and greater degradation for high-resource ones.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "In an attempt to solve this problem we investigated training dynamics (train and validation losses) for both model types. It turns out there is a persistent problem with training hierarchical models: for all trained models there is a clear overfitting for low-resource pairs, see Figure 6 , whereas highresource pairs train normally. This is interesting given that low-resource scores improved compared to bilingual baselines. Probably, the hierarchical model allows languages to learn from each other, but this is hindered by overfitting, and if we fix it, scores should grow even higher.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 280, | |
| "end": 288, | |
| "text": "Figure 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "At the same time, this problem never arises in full sharing models (even when low-resource pairs are oversampled). There might be several reasons why this is happening. In hierarchical models, different translation paths have the same number of parameters regardless of the corpus size. This means that low-resource pairs may have too many parameters for the amount of data they have, hence, overfitting. And high-resource pairs may need more Figure 5 : Difference in BLEU scores between multilingual models and bilingual models. Zero level represents average bilingual models' score. Positive values mean improvement over bilingual baseline, negative values, respectively, signify decrease. Full denotes models with full parameter sharing, Hie stands for hierarchical models, Hie+Downweight is for hierarchical models trained by down-weighting low-resource samples, and Hie+Regularize is for hierarchical models regularized using high-resource samples.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 443, | |
| "end": 451, | |
| "text": "Figure 5", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "parameters than low-resource pairs. In this sense, full sharing models may be more effective, since the parameter space is distributed automatically. Another possible explanation is that in full sharing models high-resource data acts as an implicit regularizer, preventing model parameters from being overspecialized in a low-resource pair. This is not exactly so in hierarchical models-there are individual layers, and there can be layers shared by several close low-resource languages.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "Yet another concern is oversampling lowresource data when training hierarchical models, which we applied and which itself can be the reason for overfitting. Indeed, we did not observe overfitting in our initial experiments without oversampling; However, in this case, low-resource scores are considerably lower than the ones we reported. Hence, we decided to fight overfitting using a different approach rather than removing oversampling.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "Based on this reasoning we suggest two ways to solve the problem of overfitting: 1) Downweight low-resource samples, i.e. decrease their weight in the loss function proportionally to data disbalance. Disbalance here is taken to be the ratio between the low-resource corpus size and the closest high-resource corpus size. For example, if the ratio is 1:5, then low-resource samples have five times less weight in the loss function. The idea is that this will act as a regularization for low-resource directions and at the same time high-resource samples will have more influence in shared parts which will allow improving highresource scores too.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "2) Regularize low-resource paths with highresource data. This idea is based on the above hypothesis about an implicit regularization happening in full sharing models. Here we explicitly apply this regularization by feeding the closest high-resource data instead of the low-resource data intended for the path. I.e., during the epoch the low-resource path is fed all available low-resource samples (once, we do not oversample here) plus related high-resource samples. We downweight high-resource samples used for regularization proportionally to data disbalance to limit the possible negative effect.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "We applied both these ideas. First of all, both of them solved the problem of overfitting. Now let us refer again to Figure 5 . As demonstrated there, downweighting low-resource samples turned out to bring a lot of improvement. First, the model now performs on average better than both previous mod- Figure 6 : The example of overfitting in hierarchical models. Train (left) and validation (right) losses for lowresource pl-de pair when training pl\u2192en-de hierarchical model. els (+0.42). Second, the issue with high-resource pairs is greatly alleviated, from -1.88 to -0.92. And third, low-resource scores have also noticeably increased, from +1.31 to +1.76. As for the second approach, it also improved high-resource scores (-1.09) but severely hit low-resource pairs. Probably, the reason is that the regularization was too strong, which prevented possible learning. So, to summarize, both approaches solved the problem of overfitting, and while regularization with high-resource data turned out to be too strict, downweighting low-resource samples was able to greatly improve the scores for both high-resource and low-resource directions. Now, taking the improved version of the hierarchical model as a primary one, we will try to answer the questions we put at the beginning of section 4. We were interested to know whether there is a difference when the hierarchy is on the encoder or decoder side. From our data we can say that there is no substantial difference: the hierarchical model performs at the same level regardless of where the hierarchy is. What is interesting though, simple hierarchical models perform considerably better than bigger ones: high-resource directions degrade less (-0.52 on average versus -1.33), and low-resource improve more (+2.02 versus +1.49). This may happen because adding more languages to a multilingual model inevitably leads to more distant pairs being mixed together, and they may interfere with each other instead of helping.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 117, | |
| "end": 125, | |
| "text": "Figure 5", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 300, | |
| "end": 308, | |
| "text": "Figure 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "Comparing hierarchical models to bilingual ones confirmed our supposition that hierarchical architecture allows languages to learn from each other, however, this is true mostly for low-resource pairs. If we look at the exact scores, high-resource directions do not always degrade: in 2 out of 8 cases they improve, although, not by a large margin. Low-resource directions improve in all 8 cases.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "As for the comparison of hierarchical models with full sharing ones, the improved hierarchical model surpasses the full sharing model in both highresource (+0.34) and low-resource (+0.48) pairs. At the same time, the full sharing model is beneficial in the sense that it comes with an intrinsic regularization mechanism, and perhaps this aspect could be further improved in hierarchical models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Improving Hierarchical Model", | |
| "sec_num": "4.3.2" | |
| }, | |
| { | |
| "text": "This work tested a new hierarchical approach in multilingual MT. We implemented the hierarchical model based on the Transformer architecture and contrasted it to bilingual and full parameter sharing baselines. The straightforward approach to training turned out to be hampering the hierarchical model's performance. We found out that there is a problem of overfitting which is specific to training hierarchical models. Regularizing low-resource directions solved the problem, substantially improving the model's performance.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "We showed that the hierarchical model greatly improves low-resource pairs' scores, however, at the expense of high-resource pairs. The comparison with the full sharing model provided positive evidence supporting the assumption about the usefulness of explicitly defining parameter sharing strategy in a multilingual model. Overall, the hierarchical approach looks promising and hence should be further explored and developed. In the future, we would like to analyze how exactly learning happens in hierarchical models and whether hierarchical architecture is indeed capable of capturing different language aspects (morphological, syntactic, semantic) on different levels in a hierarchical manner.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "5" | |
| } | |
| ], | |
| "back_matter": [], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "JW300: A widecoverage parallel corpus for low-resource languages", | |
| "authors": [ | |
| { | |
| "first": "Zeljko", | |
| "middle": [], | |
| "last": "Agi\u0107", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ivan", | |
| "middle": [], | |
| "last": "Vuli\u0107", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "3204--3210", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.18653/v1/P19-1310" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Zeljko Agi\u0107 and Ivan Vuli\u0107. 2019. JW300: A wide- coverage parallel corpus for low-resource languages. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3204-3210, Florence, Italy. Association for Compu- tational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Ion Madrazo Azpiazu and Maria Soledad Pera. 2020. A framework for hierarchical multilingual machine translation", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ion Madrazo Azpiazu and Maria Soledad Pera. 2020. A framework for hierarchical multilingual machine translation.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "Simple, scalable adaptation for neural machine translation", | |
| "authors": [ | |
| { | |
| "first": "Ankur", | |
| "middle": [], | |
| "last": "Bapna", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "N", | |
| "middle": [], | |
| "last": "Arivazhagan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Orhan", | |
| "middle": [], | |
| "last": "Firat", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "EMNLP/IJCNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ankur Bapna, N. Arivazhagan, and Orhan Firat. 2019. Simple, scalable adaptation for neural machine trans- lation. In EMNLP/IJCNLP.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Evaluating layers of representation in neural machine translation on part-of-speech and semantic tagging tasks", | |
| "authors": [ | |
| { | |
| "first": "Yonatan", | |
| "middle": [], | |
| "last": "Belinkov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Llu\u00eds", | |
| "middle": [], | |
| "last": "M\u00e0rquez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hassan", | |
| "middle": [], | |
| "last": "Sajjad", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nadir", | |
| "middle": [], | |
| "last": "Durrani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Fahim", | |
| "middle": [], | |
| "last": "Dalvi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "James", | |
| "middle": [], | |
| "last": "Glass", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Proceedings of the Eighth International Joint Conference on Natural Language Processing", | |
| "volume": "1", | |
| "issue": "", | |
| "pages": "1--10", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yonatan Belinkov, Llu\u00eds M\u00e0rquez, Hassan Sajjad, Nadir Durrani, Fahim Dalvi, and James Glass. 2017. Evaluating layers of representation in neural ma- chine translation on part-of-speech and semantic tagging tasks. In Proceedings of the Eighth In- ternational Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1-10, Taipei, Taiwan. Asian Federation of Natural Lan- guage Processing.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "Multi-task learning for multiple language translation", | |
| "authors": [ | |
| { | |
| "first": "Daxiang", | |
| "middle": [], | |
| "last": "Dong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hua", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Wei", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dianhai", | |
| "middle": [], | |
| "last": "Yu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Haifeng", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2015, | |
| "venue": "Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1723--1732", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.3115/v1/P15-1166" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Daxiang Dong, Hua Wu, Wei He, Dianhai Yu, and Haifeng Wang. 2015. Multi-task learning for mul- tiple language translation. In Proceedings of the 53rd Annual Meeting of the Association for Compu- tational Linguistics and the 7th International Joint Conference on Natural Language Processing (Vol- ume 1: Long Papers), pages 1723-1732, Beijing, China. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Beyond english-centric multilingual machine translation", | |
| "authors": [ | |
| { | |
| "first": "A", | |
| "middle": [], | |
| "last": "Fan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shruti", | |
| "middle": [], | |
| "last": "Bhosale", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Holger", | |
| "middle": [], | |
| "last": "Schwenk", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Z", | |
| "middle": [], | |
| "last": "Ma", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ahmed", | |
| "middle": [], | |
| "last": "El-Kishky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Siddharth", | |
| "middle": [], | |
| "last": "Goyal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mandeep", | |
| "middle": [], | |
| "last": "Baines", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "\u00c7", | |
| "middle": [], | |
| "last": "Onur", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Guillaume", | |
| "middle": [], | |
| "last": "Elebi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Vishrav", | |
| "middle": [], | |
| "last": "Wenzek", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Naman", | |
| "middle": [], | |
| "last": "Chaudhary", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tom", | |
| "middle": [], | |
| "last": "Goyal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Vitaliy", | |
| "middle": [], | |
| "last": "Birch", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Liptchinsky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "E", | |
| "middle": [], | |
| "last": "Sergey Edunov", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Grave", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Armand", | |
| "middle": [], | |
| "last": "Auli", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Joulin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "ArXiv", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "A. Fan, Shruti Bhosale, Holger Schwenk, Z. Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur \u00c7 elebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, E. Grave, M. Auli, and Armand Joulin. 2020. Beyond english-centric multi- lingual machine translation. ArXiv, abs/2010.11125.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Multi-way, multilingual neural machine translation with a shared attention mechanism", | |
| "authors": [ | |
| { | |
| "first": "Orhan", | |
| "middle": [], | |
| "last": "Firat", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kyunghyun", | |
| "middle": [], | |
| "last": "Cho", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yoshua", | |
| "middle": [], | |
| "last": "Bengio", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "866--875", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.18653/v1/N16-1101" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Orhan Firat, Kyunghyun Cho, and Yoshua Bengio. 2016. Multi-way, multilingual neural machine trans- lation with a shared attention mechanism. In Pro- ceedings of the 2016 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 866-875, San Diego, California. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "Google's multilingual neural machine translation system: Enabling zero-shot translation", | |
| "authors": [ | |
| { | |
| "first": "Melvin", | |
| "middle": [], | |
| "last": "Johnson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mike", | |
| "middle": [], | |
| "last": "Schuster", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Quoc", | |
| "middle": [ | |
| "V" | |
| ], | |
| "last": "Le", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Maxim", | |
| "middle": [], | |
| "last": "Krikun", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yonghui", | |
| "middle": [], | |
| "last": "Wu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Zhifeng", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nikhil", | |
| "middle": [], | |
| "last": "Thorat", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Fernanda", | |
| "middle": [], | |
| "last": "Vi\u00e9gas", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Martin", | |
| "middle": [], | |
| "last": "Wattenberg", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Greg", | |
| "middle": [], | |
| "last": "Corrado", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Macduff", | |
| "middle": [], | |
| "last": "Hughes", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeffrey", | |
| "middle": [], | |
| "last": "Dean", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Transactions of the Association for Computational Linguistics", | |
| "volume": "5", | |
| "issue": "", | |
| "pages": "339--351", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1162/tacl_a_00065" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Melvin Johnson, Mike Schuster, Quoc V. Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Thorat, Fernanda Vi\u00e9gas, Martin Wattenberg, Greg Corrado, Macduff Hughes, and Jeffrey Dean. 2017. Google's multilingual neural machine translation system: En- abling zero-shot translation. Transactions of the As- sociation for Computational Linguistics, 5:339-351.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Investigating multilingual NMT representations at scale", | |
| "authors": [ | |
| { | |
| "first": "Sneha", | |
| "middle": [], | |
| "last": "Kudugunta", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ankur", | |
| "middle": [], | |
| "last": "Bapna", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Isaac", | |
| "middle": [], | |
| "last": "Caswell", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Orhan", | |
| "middle": [], | |
| "last": "Firat", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1565--1575", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.18653/v1/D19-1167" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sneha Kudugunta, Ankur Bapna, Isaac Caswell, and Orhan Firat. 2019. Investigating multilingual NMT representations at scale. In Proceedings of the 2019 Conference on Empirical Methods in Natu- ral Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 1565-1575, Hong Kong, China. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Tags for Identifying Languages", | |
| "authors": [ | |
| { | |
| "first": "Addison", | |
| "middle": [], | |
| "last": "Phillips", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mark", | |
| "middle": [], | |
| "last": "Davis", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2009, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.17487/RFC5646" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Addison Phillips and Mark Davis. 2009. Tags for Iden- tifying Languages. RFC 5646.", | |
| "links": null | |
| }, | |
| "BIBREF10": { | |
| "ref_id": "b10", | |
| "title": "Parameter sharing methods for multilingual self-attentional translation models", | |
| "authors": [ | |
| { | |
| "first": "Devendra", | |
| "middle": [], | |
| "last": "Sachan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Graham", | |
| "middle": [], | |
| "last": "Neubig", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the Third Conference on Machine Translation: Research Papers", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "261--271", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.18653/v1/W18-6327" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Devendra Sachan and Graham Neubig. 2018. Parame- ter sharing methods for multilingual self-attentional translation models. In Proceedings of the Third Con- ference on Machine Translation: Research Papers, pages 261-271, Brussels, Belgium. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Bayesian phylolinguistics infers the internal structure and the time-depth of the Turkic language family", | |
| "authors": [ | |
| { | |
| "first": "Alexander", | |
| "middle": [], | |
| "last": "Savelyev", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Martine", | |
| "middle": [], | |
| "last": "Robbeets", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2020, | |
| "venue": "Journal of Language Evolution", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.1093/jole/lzz010" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Alexander Savelyev and Martine Robbeets. 2020. Bayesian phylolinguistics infers the internal struc- ture and the time-depth of the Turkic language fam- ily. Journal of Language Evolution. Lzz010.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "Die ersten spaltungen des indogermanischen urvolkes", | |
| "authors": [ | |
| { | |
| "first": "August", | |
| "middle": [], | |
| "last": "Schleicher", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Schleicher", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "August Schleicher and S Schleicher. 1853. Die ersten spaltungen des indogermanischen urvolkes [the first splits of the proto-indo-european people].", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Die verwantschaftsverh\u00e4ltnisse der indogermanischen sprachen", | |
| "authors": [ | |
| { | |
| "first": "Johannes", | |
| "middle": [], | |
| "last": "", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Johannes Schmidt. 1872. Die ver- wantschaftsverh\u00e4ltnisse der indogermanischen sprachen [The Relationship of the Indo-European Languages].", | |
| "links": null | |
| }, | |
| "BIBREF15": { | |
| "ref_id": "b15", | |
| "title": "Neural machine translation of rare words with subword units", | |
| "authors": [ | |
| { | |
| "first": "Rico", | |
| "middle": [], | |
| "last": "Sennrich", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Barry", | |
| "middle": [], | |
| "last": "Haddow", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Alexandra", | |
| "middle": [], | |
| "last": "Birch", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, ACL 2016", | |
| "volume": "1", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Lin- guistics, ACL 2016, August 7-12, 2016, Berlin, Ger- many, Volume 1: Long Papers. The Association for Computer Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "Multilingual neural machine translation with language clustering", | |
| "authors": [ | |
| { | |
| "first": "Xu", | |
| "middle": [], | |
| "last": "Tan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jianxian", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Di", | |
| "middle": [], | |
| "last": "He", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Yingce", | |
| "middle": [], | |
| "last": "Xia", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tao", | |
| "middle": [], | |
| "last": "Qin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tie-Yan", | |
| "middle": [], | |
| "last": "Liu", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "EMNLP/IJCNLP", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Xu Tan, Jianxian Chen, Di He, Yingce Xia, Tao Qin, and Tie-Yan Liu. 2019. Multilingual neural machine translation with language clustering. In EMNLP/IJCNLP.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Parallel data, tools and interfaces in opus", | |
| "authors": [ | |
| { | |
| "first": "J\u00f6rg", | |
| "middle": [], | |
| "last": "Tiedemann", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Proceedings of the Eight International Conference on Language Resources and Evaluation (LREC'12)", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "J\u00f6rg Tiedemann. 2012. Parallel data, tools and inter- faces in opus. In Proceedings of the Eight Interna- tional Conference on Language Resources and Eval- uation (LREC'12), Istanbul, Turkey. European Lan- guage Resources Association (ELRA).", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Attention is all you need", | |
| "authors": [ | |
| { | |
| "first": "Ashish", | |
| "middle": [], | |
| "last": "Vaswani", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Noam", | |
| "middle": [], | |
| "last": "Shazeer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Niki", | |
| "middle": [], | |
| "last": "Parmar", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jakob", | |
| "middle": [], | |
| "last": "Uszkoreit", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Llion", | |
| "middle": [], | |
| "last": "Jones", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Aidan", | |
| "middle": [ | |
| "N" | |
| ], | |
| "last": "Gomez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Illia", | |
| "middle": [], | |
| "last": "Kaiser", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Polosukhin", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "Advances in Neural Information Processing Systems", | |
| "volume": "30", | |
| "issue": "", | |
| "pages": "5998--6008", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, \u0141 ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Gar- nett, editors, Advances in Neural Information Pro- cessing Systems 30, pages 5998-6008. Curran Asso- ciates, Inc.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "A compact and language-sensitive multilingual translation method", | |
| "authors": [ | |
| { | |
| "first": "Yining", | |
| "middle": [], | |
| "last": "Wang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Long", | |
| "middle": [], | |
| "last": "Zhou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jiajun", | |
| "middle": [], | |
| "last": "Zhang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Feifei", | |
| "middle": [], | |
| "last": "Zhai", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jingfang", | |
| "middle": [], | |
| "last": "Xu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Chengqing", | |
| "middle": [], | |
| "last": "Zong", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "1213--1223", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.18653/v1/P19-1117" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Yining Wang, Long Zhou, Jiajun Zhang, Feifei Zhai, Jingfang Xu, and Chengqing Zong. 2019. A com- pact and language-sensitive multilingual translation method. In Proceedings of the 57th Annual Meet- ing of the Association for Computational Linguis- tics, pages 1213-1223, Florence, Italy. Association for Computational Linguistics.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "type_str": "figure", | |
| "num": null, | |
| "text": "The example of a simple hierarchical model with two source and one target language. Rounded rectangles denote encoder/decoder blocks.", | |
| "uris": null | |
| }, | |
| "FIGREF1": { | |
| "type_str": "figure", | |
| "num": null, | |
| "text": "The example of a general multi-source multi-target hierarchical model.", | |
| "uris": null | |
| } | |
| } | |
| } | |
| } |