As a long-standing central problem in natural language processing (NLP), dependency parsing has been dominated by data-driven approaches with supervised learning for decades. The foundation of data-driven parsing is the availability and scale of annotated training data (i.e.,treebanks). Numerous efforts have been made towards the construction of treebanks which established the benchmark research on dependency parsing, such as the widely-used Penn Treebank [Marcus et al.1993]. However, the heavy cost of treebanking typically limits the existing treebanks in both scale and coverage of languages.
To address the problem, a variety of authors have proposed to exploit existing heterogeneous treebanks with different annotation schemes via grammar conversion [Niu et al.2009], quasi-synchronous grammar features [Li et al.2012] or shared feature representations [Johansson2013] for the enhancement of parsing models. Despite their effectiveness in specific datasets, these methods typically require manually designed rules or features, and in most cases, they are limited to the data resources that can be used. Furthermore, for the majority of world languages, such heterogeneous treebanks are not even available. In these cases, cross-lingual treebanks may lend a helping hand.
In this paper, we aim at developing an universal framework that can exploit multi-typed source treebanks to improve parsing of a target treebank. Specifically, we will consider two kinds of source treebanks, that are multilingual universal treebanks and monolingual heterogeneous treebanks.
Cross-lingual supervision has proven highly beneficial for low-resource language parsing [Hwa et al.2005, McDonald et al.2011], implying that different languages have a great deal of common ground in grammars. But unfortunately, linguistic inconsistencies also exist in both typologies and lexical representations across languages. Figure 1(a) illustrates two sentences in German and English with universal dependency annotations. The typological differences (subject-verb-object order) results in the opposite directions of the dobj arcs, while the rest arcs remain consistent.
Similar problems also come with monolingual heterogeneous treebanks. Figure 1(b) shows an English sentence annotated with respectively the universal dependencies which are content-head and the CoNLL dependencies which instead take the functional heads. Despite the structural divergences, these treebanks express the syntax of the same language, thereby sharing a large amount of common knowledge that can be effectively transferred.
The present paper proposes a simple and effective framework that aims at making full use of the consistencies while avoids suffering from the inconsistencies across treebanks. Our framework effectively ties together the deep neural parsing models with multi-task learning, using multi-level parameter sharing to control the information flow across tasks. More specifically, learning with each treebank is maintained as an individual task, and their interactions are achieved through parameter sharing in different abstraction levels on the deep neural network, thus referred to asdeep multi-task learning. We find that different parameter sharing strategies should be applied for different typed source treebanks adaptively, due to the different types of consistencies and inconsistencies (Figure 1).
We investigate the effect of multilingual treebanks as source using the Universal Dependency Treebanks (UDT) [McDonald et al.2013]. We show that our approach improves significantly over strong supervised baseline systems in six languages. We further study the effect of monolingual heterogeneous treebanks as source using UDT and the CoNLL-X shared task dataset [Buchholz and Marsi2006]. We consider using UDT and CoNLL-X as source treebanks respectively, to investigate their mutual benefits. Experiment results show significant improvements under both settings. Moreover, indirect comparisons on the Chinese Penn Treebank 5.1 (CTB5) using the Chinese Dependency Treebank (CDT)111catalog.ldc.upenn.edu/LDC2012T05 as source treebank show the merits of our approach over previous work.
2 Related Work
The present work is related to several strands of previous studies.
Monolingual resources for parsing. Exploiting heterogeneous treebanks for parsing has been explored in various ways. niu-wang-wu:2009:ACLIJCNLP automatically convert the dependency-structure CDT into the phrase-structure style of CTB5 using a trained constituency parser on CTB5, and then combined the converted treebanks for constituency parsing. li-liu-che:2012:ACL2012 capture the annotation inconsistencies among different treebanks by designing several types of transformation patterns, based on which they introduce quasi-synchronous grammar features [Smith and Eisner2009] to augment the baseline parsing models. johansson:2013:NAACL-HLT also adopts the idea of parameter sharing to incorporate multiple treebanks. They focused on parameter sharing at feature-level with discrete representations, which limits its scalability to multilingual treebanks where feature surfaces might be totally different. On the contrary, our approach are capable of utilizing representation-level parameter sharing, making full use of the multi-level abstractive representations generated by deep neural network. This is the key that makes our framework scalable to multi-typed treebanks and thus more practically useful.
Aside from resource utilization, attempts have also been made to integrate different parsing models through stacking [Torres Martins et al.2008, Nivre and McDonald2008] or joint inference [Zhang and Clark2008, Zhang et al.2014].
Multilingual resources for parsing. Cross-lingual transfer has proven to be a promising way of inducing parsers for low-resource languages, either through data transfer [Hwa et al.2005, Tiedemann2014, Rasooli and Collins2015] or model transfer [McDonald et al.2011, Täckström et al.2012, Guo et al.2015, Zhang and Barzilay2015].
duong-EtAl:2015:EMNLP and waleed2016unidep both adopt parameter sharing to exploit multilingual treebanks in parsing, but with a few important differences to our work. In both of their models, most of the neural network parameters are shared in two (or multiple) parsers except the feature embeddings,222duong-EtAl:2015:EMNLP used regularizers to tie the lexical embeddings with a bilingual dictionary. which ignores the important syntactical inconsistencies of different languages and is also inapplicable for heterogeneous treebanks that have different transition actions. Besides, duong-EtAl:2015:EMNLP focus on low resource parsing where the target language has a small treebank of 3K tokens. Their models may sacrifice accuracy on target languages with a large treebank. waleed2016unidep instead train a single parser on a multilingual set of rich-resource treebanks, which is a more similar setting to ours. We refer to their approach as shallow multi-task learning (SMTL) and will include as one of our baseline systems (Section 4.2). Note that SMTL is a special case of our approach in which all tasks use the same set of parameters.
, demonstrating the potential of cross-lingual transfer learning.
Multi-task learning for NLP. There has been a line of research on joint modeling pipelined NLP tasks, such as word segmentation, POS tagging and parsing [Hatori et al.2012, Li et al.2011, Bohnet and Nivre2012]. Most multi-task learning or joint training frameworks can be summarized as parameter sharing approaches proposed by Ando:2005:FLP:1046920.1194905. In the context of neural models for NLP, the most notable work was proposed by collobert2008unified, which aims at solving multiple NLP tasks within one framework by sharing common word embeddings. henderson2013multilingual present a joint dependency parsing and semantic role labeling model with the Incremental Sigmoid Belief Networks (ISBN) [Henderson and Titov2010]
. More recently, the idea of neural multi-task learning was applied to sequence-to-sequence problems with recurrent neural networks. dong-EtAl:2015:ACL-IJCNLP2 use multiple decoders in neural machine translation systems that allows translating one source language to many target languages. luong2015multitask study the ensemble of a wide range of tasks (e.g., syntactic parsing, machine translation, image caption, etc.) with multi-task sequence-to-sequence models.
To the best of our knowledge, we present the first work that successfully integrate both monolingual and multilingual treebanks for parsing, with or without consistent annotation schemes.
This section describes the deep multi-task learning architecture, using a formalism that extends on the transition-based dependency parsing model with LSTM networks [Dyer et al.2015] which is further enhanced by modeling characters [Ballesteros et al.2015]. We first revisit the parsing approach of ballesteros-dyer-smith:2015:EMNLP, then present our framework for learning with multi-typed source treebanks.
3.1 Transition-based Neural Parsing
Neural models for parsing have gained a lot of interests in recent years, particularly boosted by chen2014fast. The heart of transition-based parsing is the challenge of representing the state (configuration) of a transition system, based on which the most likely transition action is determined. Typically, a state includes three primary components, a stack, a buffer and a set of dependency arcs
. Traditional parsing models deal with features extracted from manually defined feature templates in a discrete feature space, which suffers from the problems ofSparsity, Incompleteness and Expensive feature computation
. The neural network model proposed by chen2014fast instead represents features as continuous, low-dimensional vectors and use acubeactivation function for implicit feature composition. More recently, this architecture has been improved in several different ways [Dyer et al.2015, Weiss et al.2015, Zhou et al.2015, Andor et al.2016]. Here, we employ the LSTM-based architecture enhanced with character bidirectional LSTMs [Ballesteros et al.2015] for the following major reasons:
Compared with Chen & Manning’s architecture, it makes full use of the non-local features by modeling the full history information of a state with stack LSTMs.
By modeling words, stack, buffer and action sequence separately which indicate hierarchical abstractions of representations, we can control the information flow across tasks via parameter sharing with more flexibility (Section 3.2).
Besides, we did not use the earlier ISBN parsing model [Titov and Henderson2007] due to its lack of scalability to large vocabulary. Figure 2 illustrates the transition-based parsing architecture using LSTMs. Bidirectional LSTMs are used for modeling the word representations (Figure 3), which we refer to as Char-BiLSTMs henceforth. Char-BiLSTMs learn features for each word, and then the representation of each token can be calculated as:
where is the POS tag embedding. The token embeddings are then fed into subsequent LSTM layers to obtain representations of the stack, buffer and action sequence respectively referred to as and (The subscript represents the time step). Note that the subtrees within the stack and buffer are modeled with a recursive neural network (RecNN) as described in dyer-EtAl:2015:ACL-IJCNLP. Next, a linear mapping () is applied to the concatenation of and , and passed through a component-wise :
Finally, the probability of next action
is estimated using afunction:
where represents the set of valid actions given the current content in the stack and buffer.
We apply the non-projective transition system originally introduced by nivre2009non since most of the treebanks we consider in this study has a noticeable proportion of non-projective trees. In the Swap-based system, both the stack and buffer may contain tree fragments, so RecNN is applied both in S and B to obtain representations of each position.
3.2 Deep Multi-task Learning
Multi-task learning (MTL) is the procedure of inductive transfer that improves learning for one task by using the information contained in the training signals of other related tasks. It does this by learning tasks in parallel while using a shared representation. A good overview, especially focusing on neural networks, can be found in caruana1997multitask.
We illustrate our multi-task learning architecture in Figure 4. As discussed in previous sections, multiple treebanks, either multilingual or monolingual heterogeneous, contain knowledge that can be mutually beneficial. We consider the target treebank processing as the primary task, and the source treebank as a related task. The two tasks are interacted through multi-level parameter sharing (Section 3.2.1). Inspired by waleed2016unidep, we introduce a task-specific vector (task embedding) which is first combined with to compute , and then further concatenated with
to compute the probability distribution of transition actions. Therefore, Eqn2, 3 become:
Each task uses the same formalism for optimization, and the joint cross-entropy is used as the objective function. The key of multi-task learning is parameter sharing, without which the correlation between tasks will not be exploited. Conventional multi-task learning models typically share a small proportion of parameters across tasks. For example, collobert2008unified only shares word embeddings, and dong-EtAl:2015:ACL-IJCNLP2 shares the encoder of sequence-to-sequence models. In this work, we use more sophisticated parameter sharing strategies according to the linguistic similarities and differences between the tasks.
3.2.1 Parameter Sharing
Deep neural networks automatically learn features for a specific task with hierarchical abstractions, which gives us the flexibility to control parameter sharing in different levels accordingly.
In this study, different parameter sharing strategies are applied according to the source and target treebanks being used. We consider two different scenarios: MTL with multilingual universal treebanks as source (Multi-Univ) and MTL with monolingual heterogeneous treebanks as source (Mono-Hetero). Table 1 presents our parameter sharing strategies for each setting.
Multi-Univ. Multilingual universal treebanks are annotated with the same set of POS tags [Petrov et al.2012], dependency relations, and thus share the same set of transition actions. However, the vocabularies (word, characters) are language-specific. Additionally, linguistic typologies such as the order of subject-verb-object and adjective-noun (Figure 1(a)) also varies across languages, which result in the divergence of inherent grammars of transition actions. Therefore, it makes sense to share the lookup tables (embeddings) of POS tags (), relations () and actions (), but separate the character embeddings () as well as the Char-BiLSTM (BiLSTM(chars)), and also the LSTM modeling action sequence (LSTM(A))
Mono-Hetero. Monolingual heterogeneous treebanks instead share the same lexical representations, but have different POS tags, structures and relations (Figure 1(b)) due to the different annotation schemes. Hence the transition actions set varies across treebanks. For simplicity reasons, we convert the language-specific POS tags in the heterogeneous treebanks into universal POS tags [Petrov et al.2012]. Consequently, and BiLSTM(chars), are shared across tasks, but , , LSTM(A) are separated.
Besides, the LSTM parameters for modeling the stack and buffer (LSTM(S), LSTM(B)), the RecNN for modeling tree compositions, and the weights from S, B, A to the state () are shared for both Multi-Univ and Mono-Hetero. As standard in multi-task learning, the weights at the output layer () are task-specific in both settings.
Training is achieved in a stochastic manner by looping over the tasks:
Randomly select a task.
Select a sentence from the task, and generate instances for classification.
Update the corresponding parameters by back-propagation w.r.t. the instances.
Go to 1.
We adopt the development data of the target treebank (primary task) for early-stopping.
We first describe the data and settings in our experiments, then the results and analysis.
4.1 Data and Settings
We conduct experiments on UDT v2.0333github.com/ryanmcd/uni-dep-tb and the CoNLL-X shared task data. For monolingual heterogeneous source, we also experiment on CTB5 using CDT as the source treebank, to compare with the previous work of li-liu-che:2012:ACL2012. Statistics of the datasets are summarized in Table 2. We investigate the following experiment settings:
Multilingual (UnivUniv). In this setting, we study the integration of multilingual universal treebanks. Experiments are conducted using the UDT dataset. Specifically, we consider DE, ES, FR, PT, IT and SV treebanks as target treebanks, and the EN treebank as the common source treebank.
Monolingual (ConllUniv). Here we study the integration of monolingual heterogeneous treebanks. The CoNLL-X corporas (DE, ES, PT, SV) and the UDT treebank of corresponding languages are used as source and target treebanks mutually.
Monolingual (CDTCTB5). We follow the same settings of li-liu-che:2012:ACL2012, and consider two scenarios using automatic POS tags and gold-standard POS tags respectively.
We use the widely-adopted unlabeled attachment score (UAS) and labeled attachment score (LAS) for evaluation.
4.2 Baseline Systems
We compare our approach with the following baseline systems.
Monolingual supervised training (Sup). Models are trained only on the target treebank, with the LSTM-based parser.
Cascaded training (Cas). This system has two stages. First, models are trained using the source treebank. Then the parameters are used to initialize the neural network for training target parsers. Similar approach was studied in duong-EtAl:2015:ACL-IJCNLP and guo2016multisource for low-resource parsing.
For Multilingual (UnivUniv), we also compare with the shallow multi-task learning (SMTL) system, as described in Section 2, which is representative of the approach of duong-EtAl:2015:EMNLP and waleed2016unidep. In SMTL all the parameters are shared except the character embeddings (), and task embeddings () are not used. Unlike duong-EtAl:2015:EMNLP and waleed2016unidep, we don’t use external resources such as cross-lingual word clusters, embeddings and dictionaries which is beyond the scope of this work.
|Multilingual (Univ Univ)|
In this section, we present empirical evaluations under different settings.
4.3.1 Multilingual Universal Source Treebanks
Table 3 shows the results under the Multilingual (UnivUniv) setting. Cas yields slightly better performance than Sup, especially for SV (+1.52% UAS and +2.04% LAS), indicating that pre-training with EN training data indeed provides a better initialization of the parameters for cascaded training. SMTL in turn outperforms Cas overall (comparable for IT), which implies that training two treebanks jointly helps even with an unique model.
Furthermore, with appropriate parameter sharing, our deep multi-task learning approach (MTL) outperforms Sup overall and achieves the best performances in five out of six languages. An exception is Swedish. As we can see, both Cas and SMTL outperforms MTL by a significant margin for SV. The underlying reasons we suggest are two-fold.
SV morphology is similar to EN with less inflections, encouraging the morphology-related parameters like BiLSTM(chars) to be shared.
SV has a much smaller treebank compared with EN (1:9). Intuitively, SMTL and Cas work better in low resource setting.
To verify the first issue, we conduct tests on SMTL without sharing Char-BiLSTMs. As shown in Table 4, the performance of SMTL decreases significantly (-0.73 in UAS). This observation also indicates that MTL has the potential to reach higher performances through language-specific tuning of parameter sharing strategies.
To verify the second issue, we consider a low resource setup following duong-EtAl:2015:EMNLP, where the target language has a small treebank (3K tokens). We train our models on identical sampled dataset shared by duong-EtAl:2015:EMNLP on DE, ES and FR. As we can find in Table 5, while all the models outperform Sup, both Cas and SMTL work better than MTL, which confirms our assumption. Although not the primary focus of this work, we find that SMTL and MTL can be significantly improved in low resource setting through weighted sampling of tasks during training. Specifically, in the training procedure (Section 3.2.2), we sample from the source language (EN) which has a much richer treebank with larger probability of 0.9, while sample from the target language with probability of 0.1. In this way, the two tasks are encouraged to converge at a similar rate. As shown in Table 5, both SMTL and MTL benefit from weighted task sampling.
|+ weighted sampling||63.50||70.17||68.52|
|+ weighted sampling||64.22||68.42||66.67|
|Duong et al.||61.2||69.1||65.3|
|Duong et al. + Dict||61.8||70.5||67.2|
4.3.2 Monolingual Hetero. Source Treebanks
|Ours||UAS||79.34||80.25 (+0.91)||81.13 (+1.79)||85.25||86.29 (+1.04)||86.69 (+1.44)|
|LAS||76.23||77.26 (+1.03)||78.24 (+2.01)||83.59||84.72 (+1.13)||85.18 (+1.59)|
|Sup||with QG||Sup||with QG|
|Li12-O2||UAS||79.67||81.04 (+1.37)||86.13||86.44 (+0.31)|
|Li12-O2sib||79.25||80.45 (+1.20)||85.63||86.17 (+0.54)|
Table 6 shows the results of Monolingual (ConllUniv). Overall MTL systems outperforms the supervised baselines by significant margins in both conditions, showing the mutual benefits of UDT and CoNLL-X treebanks.444An exception is PT in Monolingual (UnivConll), in which both Cas and MTL get slightly degradation in performance. This may be due to the low quality of the PT universal treebank caused by the automatic construction process. We discussed and verified this with the author of UDT v2.0.
In addition, among the four languages here, the SV universal treebank is mainly converted from the Talbanken part of the Swedish bank [Nivre and Megyesi2007], thus has a large overlap with the CoNLL-X Swedish treebank. In fact, we find a large proportion of the SV test data in UDT/CoNLL-X appears in CoNLL-X/UDT SV training data. Typically we expect fully unseen data for testing, so we further separate the SV testing data into two parts: In-Src and Out-Src including sentences that appear in the source treebank or not, respectively. Statistics are shown below.
The SV row in Table 6 presents the Out-Src results of SV, which shows consistent improvements.
To show the merit of our approach against previous approaches, we further conduct experiments on CTB5 using CDT as heterogeneous source treebank (Table 2). For CTB5, we follow [Li et al.2012] and consider two scenarios which use automatic POS tags and gold-standard POS tags respectively. To compare with their results, we run Sup, Cas and MTL on CTB5. Table 7 presents the results. The indirect comparison indicates that our approach can achieve larger improvement than their method in both scenarios. Beside the empirical comparison, our method has the additional advantages in its scalability to multi-typed source treebanks without the painful human efforts of feature design.
Overall, our approach obtains substantial gains over supervised baselines with either multilingual universal treebanks or monolingual heterogeneous treebanks as source. With multilingual source treebanks, our model has the potential to improve even further via language-specific tuning. While not the primary focus of this study, in low resource setting, we show that more emphasize may be put on the source treebanks through weighted task sampling.
This paper propose an universal framework based on deep multi-task learning that can integrate arbitrary-typed source treebanks to enhance the parsing models on target treebanks. We study two scenarios, respectively using multilingual universal source treebanks and monolingual heterogeneous source treebanks, and design effective parameter sharing strategies for each scenario.
We conduct extensive experiments on several benchmark treebanks in various languages. Results demonstrate that our approach significantly improves over baseline systems under various experiment setting. Furthermore, our framework can flexibly incorporate richer treebanks and more related tasks, which we leave to future exploration.
We thank Ryan McDonald for fruitful discussions, and thank Dr. Zhenghua Li for sharing the processed CTB and CDT dataset. This work was supported by the National Key Basic Research Program of China via grant 2014CB340503 and the National Natural Science Foundation of China (NSFC) via grant 61133012 and 61370164.
- [Ammar et al.2016] Waleed Ammar, George Mulcaire, Miguel Ballesteros, Chris Dyer, and Noah A Smith. 2016. One parser, many languages. arXiv preprint arXiv:1602.01595.
- [Ando and Zhang2005] Rie Kubota Ando and Tong Zhang. 2005. A framework for learning predictive structures from multiple tasks and unlabeled data. JMLR, 6:1817–1853, December.
- [Andor et al.2016] Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav Petrov, and Michael Collins. 2016. Globally normalized transition-based neural networks. arXiv preprint arXiv:1603.06042.
- [Ballesteros et al.2015] Miguel Ballesteros, Chris Dyer, and Noah A. Smith. 2015. Improved transition-based parsing by modeling characters instead of words with lstms. In Proc. of the 2015 Conference on EMNLP, pages 349–359, September.
- [Bohnet and Nivre2012] Bernd Bohnet and Joakim Nivre. 2012. A transition-based system for joint part-of-speech tagging and labeled non-projective dependency parsing. In Proc. of the 2012 Joint Conference on EMNLP and CoNLL, pages 1455–1465, July.
- [Buchholz and Marsi2006] Sabine Buchholz and Erwin Marsi. 2006. Conll-x shared task on multilingual dependency parsing. In Proc. of the Tenth Conference on Computational Natural Language Learning (CoNLL-X), pages 149–164, June.
- [Burkett and Klein2008] David Burkett and Dan Klein. 2008. Two languages are better than one (for syntactic parsing). In Proc. of the 2008 Conference on EMNLP, pages 877–886, October.
- [Caruana1997] Rich Caruana. 1997. Multitask learning. Machine learning, 28(1):41–75.
- [Chen and Manning2014] Danqi Chen and Christopher Manning. 2014. A fast and accurate dependency parser using neural networks. In Proc. of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 740–750, October.
- [Chen et al.2010] Wenliang Chen, Jun’ichi Kazama, and Kentaro Torisawa. 2010. Bitext dependency parsing with bilingual subtree constraints. In Proc. of the 48th ACL, pages 21–29, July.
- [Collobert and Weston2008] Ronan Collobert and Jason Weston. 2008. A unified architecture for natural language processing: Deep neural networks with multitask learning. In Proc. of the 25th ICML, pages 160–167.
- [Dong et al.2015] Daxiang Dong, Hua Wu, Wei He, Dianhai Yu, and Haifeng Wang. 2015. Multi-task learning for multiple language translation. In Proc. of the 53rd ACL and the 7th IJCNLP (Volume 1: Long Papers), pages 1723–1732, July.
- [Duong et al.2015a] Long Duong, Trevor Cohn, Steven Bird, and Paul Cook. 2015a. Low resource dependency parsing: Cross-lingual parameter sharing in a neural network parser. In Proc. of the 53rd ACL and the 7th IJCNLP (Volume 2: Short Papers), pages 845–850, July.
- [Duong et al.2015b] Long Duong, Trevor Cohn, Steven Bird, and Paul Cook. 2015b. A neural network model for low-resource universal dependency parsing. In Proc. of the 2015 Conference on EMNLP, pages 339–348, September.
[Dyer et al.2015]
Chris Dyer, Miguel Ballesteros, Wang Ling, Austin Matthews, and Noah A. Smith.
Transition-based dependency parsing with stack long short-term memory.In Proc. of the 53rd ACL and the 7th IJCNLP (Volume 1: Long Papers), pages 334–343, July.
[Guo et al.2015]
Jiang Guo, Wanxiang Che, David Yarowsky, Haifeng Wang, and Ting Liu.
Cross-lingual dependency parsing based on distributed representations.In Proc. of the 53rd ACL and the 7th IJCNLP (Volume 1: Long Papers), pages 1234–1244, July.
[Guo et al.2016]
Jiang Guo, Wanxiang Che, David Yarowsky, Haifeng Wang, and Ting Liu.
A representation learning framework for multi-source transfer
Proc. of the Thirtieth AAAI Conference on Artificial Intelligence (AAAI), February.
- [Hatori et al.2012] Jun Hatori, Takuya Matsuzaki, Yusuke Miyao, and Jun’ichi Tsujii. 2012. Incremental joint approach to word segmentation, pos tagging, and dependency parsing in chinese. In Proc. of the 50th ACL (Volume 1: Long Papers), pages 1045–1053, July.
- [Henderson and Titov2010] James Henderson and Ivan Titov. 2010. Incremental sigmoid belief networks for grammar learning. J. Mach. Learn. Res., 11:3541–3570, December.
- [Henderson et al.2013] James Henderson, Paola Merlo, Ivan Titov, and Gabriele Musillo. 2013. Multilingual joint parsing of syntactic and semantic dependencies with a latent variable model. Computational Linguistics, 39(4):949–998.
- [Huang et al.2009] Liang Huang, Wenbin Jiang, and Qun Liu. 2009. Bilingually-constrained (monolingual) shift-reduce parsing. In Proc. of the 2009 Conference on EMNLP, pages 1222–1231, August.
- [Hwa et al.2005] Rebecca Hwa, Philip Resnik, Amy Weinberg, Clara Cabezas, and Okan Kolak. 2005. Bootstrapping parsers via syntactic projection across parallel texts. Natural language engineering, 11(03):311–325.
- [Johansson2013] Richard Johansson. 2013. Training parsers on incompatible treebanks. In Proc. of NAACL: HLT, pages 127–137, June.
- [Li et al.2011] Zhenghua Li, Min Zhang, Wanxiang Che, Ting Liu, Wenliang Chen, and Haizhou Li. 2011. Joint models for chinese pos tagging and dependency parsing. In Proc. of the 2011 Conference on EMNLP, pages 1180–1191, July.
- [Li et al.2012] Zhenghua Li, Ting Liu, and Wanxiang Che. 2012. Exploiting multiple treebanks for parsing with quasi-synchronous grammars. In Proc. of the 50th ACL (Volume 1: Long Papers), pages 675–684, July.
- [Luong et al.2015] Minh-Thang Luong, Quoc V. Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. 2015. Multi-task sequence to sequence learning. CoRR, abs/1511.06114.
- [Marcus et al.1993] Mitchell P Marcus, Mary Ann Marcinkiewicz, and Beatrice Santorini. 1993. Building a large annotated corpus of english: The penn treebank. Computational linguistics, 19(2):313–330.
- [McDonald et al.2011] Ryan McDonald, Slav Petrov, and Keith Hall. 2011. Multi-source transfer of delexicalized dependency parsers. In Proc. of the 2011 Conference on EMNLP, pages 62–72, July.
- [McDonald et al.2013] Ryan McDonald, Joakim Nivre, Yvonne Quirmbach-Brundage, Yoav Goldberg, Dipanjan Das, Kuzman Ganchev, Keith Hall, Slav Petrov, Hao Zhang, Oscar Täckström, Claudia Bedini, Núria Bertomeu Castelló, and Jungmee Lee. 2013. Universal dependency annotation for multilingual parsing. In Proc. of the 51st ACL (Volume 2: Short Papers), pages 92–97, August.
- [Niu et al.2009] Zheng-Yu Niu, Haifeng Wang, and Hua Wu. 2009. Exploiting heterogeneous treebanks for parsing. In Proc. of the Joint Conference of the 47th ACL and the 4th IJCNLP of the AFNLP, pages 46–54, August.
- [Nivre and McDonald2008] Joakim Nivre and Ryan McDonald. 2008. Integrating graph-based and transition-based dependency parsers. In Proc. of ACL-08: HLT, pages 950–958, June.
- [Nivre and Megyesi2007] Joakim Nivre and Beata Megyesi. 2007. Bootstrapping a swedish treebank using cross-corpus harmonization and annotation projection. In Proc. of the 6th International Workshop on Treebanks and Linguistic Theories, pages 97–102.
- [Nivre2009] Joakim Nivre. 2009. Non-projective dependency parsing in expected linear time. In Proc. of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 351–359, August.
- [Petrov et al.2012] Slav Petrov, Dipanjan Das, and Ryan McDonald. 2012. A universal part-of-speech tagset. In Proc. of the Eighth International Conference on Language Resources and Evaluation (LREC-2012), pages 2089–2096, May.
- [Rasooli and Collins2015] Mohammad Sadegh Rasooli and Michael Collins. 2015. Density-driven cross-lingual transfer of dependency parsers. In Proc. of the 2015 Conference on EMNLP, pages 328–338, September.
- [Smith and Eisner2009] David A. Smith and Jason Eisner. 2009. Parser adaptation and projection with quasi-synchronous grammar features. In Proc. of the 2009 Conference on EMNLP, pages 822–831, August.
- [Täckström et al.2012] Oscar Täckström, Ryan McDonald, and Jakob Uszkoreit. 2012. Cross-lingual word clusters for direct transfer of linguistic structure. In Proc. of NAACL: HLT, pages 477–487, June.
- [Tiedemann2014] Jörg Tiedemann. 2014. Rediscovering annotation projection for cross-lingual parser induction. In Proc. of COLING 2014, pages 1854–1864, August.
- [Titov and Henderson2007] Ivan Titov and James Henderson. 2007. Fast and robust multilingual dependency parsing with a generative latent variable model. In Proc. of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pages 947–951, June.
- [Torres Martins et al.2008] André Filipe Torres Martins, Dipanjan Das, Noah A. Smith, and Eric P. Xing. 2008. Stacking dependency parsers. In Proc. of the 2008 Conference on EMNLP, pages 157–166, October.
- [Weiss et al.2015] David Weiss, Chris Alberti, Michael Collins, and Slav Petrov. 2015. Structured training for neural network transition-based parsing. In Proc. of the 53rd ACL and the 7th IJCNLP (Volume 1: Long Papers), pages 323–333, July.
[Zhang and Barzilay2015]
Yuan Zhang and Regina Barzilay.
Hierarchical low-rank tensors for multilingual transfer parsing.In Proc. of the 2015 Conference on EMNLP, pages 1857–1867, September.
- [Zhang and Clark2008] Yue Zhang and Stephen Clark. 2008. A tale of two parsers: Investigating and combining graph-based and transition-based dependency parsing. In Proc. of the 2008 Conference on EMNLP, pages 562–571, October.
- [Zhang et al.2014] Meishan Zhang, Wanxiang Che, Yanqiu Shao, and Ting Liu. 2014. Jointly or separately: Which is better for parsing heterogeneous dependencies? In Proc. of COLING 2014, pages 530–540, August.
- [Zhou et al.2015] Hao Zhou, Yue Zhang, Shujian Huang, and Jiajun Chen. 2015. A neural probabilistic structured-prediction model for transition-based dependency parsing. In Proc. of the 53rd ACL and the 7th IJCNLP (Volume 1: Long Papers), pages 1213–1222, July.