Zero-shot transfer for implicit discourse relation classification

TL;DR

This paper introduces a zero-shot transfer learning approach for classifying implicit discourse relations across multiple languages using only English training data and unannotated parallel text, demonstrating effective results on the TED-MDB corpus.

Contribution

It proposes a novel zero-shot transfer method for implicit discourse relation classification that requires no annotated data in target languages.

Findings

Achieved good classification results on seven languages using only English data.

Demonstrated the effectiveness of unannotated parallel text for cross-lingual transfer.

Showed that zero-shot transfer can be viable for discourse relation classification.

Abstract

Automatically classifying the relation between sentences in a discourse is a challenging task, in particular when there is no overt expression of the relation. It becomes even more challenging by the fact that annotated training data exists only for a small number of languages, such as English and Chinese. We present a new system using zero-shot transfer learning for implicit discourse relation classification, where the only resource used for the target language is unannotated parallel text. This system is evaluated on the discourse-annotated TED-MDB parallel corpus, where it obtains good results for all seven languages using only English training data.

Tables

Table 1. Table 1: Distribution of top level senses of the implicit discourse relations in PDTB 2.0 and PDTB 3.0 training, development and test sets: comp(arison), cont(ingency), exp(ansion), temp(oral).

	PDTB2			PDTB3
Sense	Train	Dev	Test	Train	Dev	Test
Comp.	1894	401	146	1828	404	153
Cont.	3281	628	276	5872	1159	527
Exp.	6792	1253	556	7939	1466	643
Temp.	665	93	68	1413	230	148

Table 2. Table 2: Distribution of top level senses of the implicit discourse relations in the TED-MDB and TDB corpora. The numbers within the parenthesis indicate the ratio. Since there is no official training/dev split for TDB, we arbitrarily chose two sections with different genres for the development set.

Language	Comparison	Contingency	Expansion	Temporal	Total
English	20 (10.31%)	52 (26.80%)	107 (55.15%)	15 (7.73%)	194 (100%)
German	13 (6.07%)	41 (19.16%)	148 (69.16%)	12 (5.61%)	214 (100%)
Lithuanian	26 (10.57%)	53 (21.54%)	154 (62.60%)	13 (5.28%)	246 (100%)
Polish	19 (9.74%)	28 (14.36%)	130 (66.67%)	18 (9.23%)	195 (100%)
Portuguese	23 (9.06%)	47 (18.50%)	169 (66.54%)	15 (5.91%)	254 (100%)
Russian	16 (7.24%)	31 (14.03%)	169 (76.47%)	5 (2.26%)	221 (100%)
Turkish	20 (9.90%)	29 (14.36%)	140 (69.31%)	13 (6.44%)	202 (100%)
TDB (training)	71 (10.94%)	142 (21.88%)	363 (55.93%)	73 (11.25%)	649 (100%)
TDB (dev)	11 (9.82%)	31 (27.68%)	49 (43.75%)	21 (18.75%)	112 (100%)

Table 3. Table 3: Comparison of the F scores (%) of binary classifiers on PDTB 2.0 test set. Left out scores refer to the results where EntRel relations are also considered to be Expansion.

	Comparison	Contingency	Expansion	Temporal
Pitler et al. (2009)	21.96	47.13	-	16.76
Zhou et al. (2010)	31.79	47.16	70.11	20.30
Park and Cardie (2012)	31.32	49.82	-	26.57
Rutherford and Xue (2014)	39.70	54.42	70.23	28.69
Zhang et al. (2015)	33.22	52.04	69.59	30.54
Ji and Eisenstein (2015)	35.93	52.78	-	27.63
Lan et al. (2017)	40.73	58.96	72.47	38.50
Bai and Zhao (2018)	47.85	54.47	70.60	36.87
Dai and Huang (2018)	46.79	57.09	70.41	45.61
Baseline	24.49	41.75	69.41	12.20
Our system	28.19 ($\pm$0.83)	50.63 ($\pm$1.00)	64.07 ($\pm$1.90)	29.22 ($\pm$2.53)

Table 4. Table 4: F scores (%) when the model is trained only on PDTB 3.0. In the table, PDTB 3.0 refers to the test set of the PDTB 3.0 corpus. The remaining rows refer to evaluations using TED-MDB.

Language	Comparison	Contingency	Expansion	Temporal	Average
Baseline (PDTB 3.0)	18.84	52.75	60.83	18.28	37.67
PDTB 3.0	24.90 ($\pm$0.87)	59.18 ($\pm$0.72)	60.10 ($\pm$1.32)	36.73 ($\pm$1.45)	45.23
German	8.62 ($\pm$1.61)	37.34 ($\pm$1.43)	70.81 ($\pm$3.16)	40.11 ($\pm$4.32)	39.22
English	10.18 ($\pm$3.31)	40.92 ($\pm$1.80)	62.28 ($\pm$2.16)	50.45 ($\pm$5.26)	40.96
Lithuanian	23.50 ($\pm$2.33)	34.64 ($\pm$1.43)	62.35 ($\pm$2.65)	36.78 ($\pm$3.28)	39.32
Polish	16.50 ($\pm$3.51)	29.19 ($\pm$1.36)	60.32 ($\pm$2.84)	44.17 ($\pm$3.37)	37.54
Portuguese	19.59 ($\pm$1.99)	33.85 ($\pm$1.27)	66.83 ($\pm$2.57)	37.04 ($\pm$3.43)	39.33
Russian	14.90 ($\pm$2.07)	26.76 ($\pm$1.08)	70.06 ($\pm$3.97)	28.28 ($\pm$4.41)	35.00
Turkish	10.99 ($\pm$3.16)	25.28 ($\pm$1.23)	64.14 ($\pm$2.96)	33.66 ($\pm$4.31)	33.52

Table 5. Table 5: Comparison of average F-scores (%) when the model is trained on different training sets. Bold means significantly higher F-score than the second highest column ( p < 0.001 𝑝 0.001 p<0.001 , Mann-Whitney U test).

Language	TDB	PDTB3	PTDB3+TDB
PDTB3 Test	35.35	45.23	45.62
German	36.93	39.22	41.44
English	38.06	40.96	42.22
Lithuanian	36.92	39.32	41.94
Polish	35.48	37.54	39.65
Portuguese	37.58	39.33	41.04
Russian	30.92	35.00	38.23
Turkish	39.58	33.52	37.14

Equations

$$V_{avg}=\frac{1}{2}(V_{arg1}+V_{arg2})$$

$$V_{sub}=V_{arg1}-V_{arg2}$$

$$V_{mul}=V_{arg1}\ast V_{arg2}$$

$$V_{dr}=[V_{arg1},V_{arg2},V_{avg},V_{sub},V_{mul}]$$

Full text

1 Introduction

The difference between a set of randomly selected sentences and a discourse lies in coherence. Among other attempts at defining the elusive nature of coherence, one way is to look at the meaning conveyed between the adjacent pair of sentences. In the current study, we follow the Penn Discourse Treebank (PDTB) framework which regards abstract objects (Asher, 2012) as the units of discourse and views the text as a collection of discourse level predicates, each taking two arguments. Such predicates, called discourse connectives, may (Ex. 1) or may not (Ex. 2) be represented in the surface form:

Because the drought reduced U.S. stockpiles, they have more than enough storage space for their new crop, and that permits them to wait for prices to rise. 2. 2.

But a few funds have taken other defensive steps. Some have raised their cash positions to record levels. Implicit = BECAUSE High cash positions help buffer a fund when the market falls.

where italics represents the first and boldface the second argument to the underlined discourse connective. The discourse relations which lack an overt discourse connective (Ex. 2) are referred as implicit discourse relations and are shown to be the most challenging part of the discourse parsing (e.g. Pitler et al., 2009).

In this paper, we perform implicit discourse relation classification using three recent data sets annotated according to the same guidelines: Penn Discourse Treebank (PDTB) 3.0, the Turkish Discourse Bank (TDB), and the multilingual TED-MDB. To the best of our knowledge, multilingual training and zero-shot transfer has not previously been investigated for this problem. The results suggest that an implicit discourse relation classifier can transfer well across dissimilar languages, and that pooling training data from unrelated languages (English and Turkish) leads to significantly better performance for all languages.

2 Related Work

Implicit discourse relation recognition is often handled as a classification task, where earlier studies focused on using linguistically rich features Pitler et al. (2009); Zhou et al. (2010); Park and Cardie (2012); Rutherford and Xue (2014).

Recently, neural network approaches have become popular. Ji and Eisenstein (2015) use two RNNs on the syntactic trees of the arguments whereas Zhang et al. (2015) use a CNN to perform discourse parsing in a multi-task setting where they consider both explicit and implicit discourse relations.

Rutherford and Xue (2016) use a simple yet robust feedforward network and achieves the highest performance on the out-of-domain blind test in the CoNLL 2016 shared task (Xue et al., 2016).

Lan et al. (2017) apply a multi-task attention-based neural network model whereas Bai and Zhao (2018) focus on the representation of the sentence pair and take different levels of text, from character to sentence pair, into account to achieve a richer representation.

Dai and Huang (2018) adopt a similiar approach and represent discourse units by considering a wider paragraph-level context. The discourse unit representations are created by a Bi-LSTM which takes a sequence of discourse relations in a paragraph which enables capturing the inter-dependencies between discourse relations as well.

3 Data

We use four different data sets: the Penn Discourse Treebank (PDTB) version 2.0 Prasad et al. (2008) and version 3.0 Prasad et al. (2018), as well as the TED Multilingual Discourse Bank (TED-MDB, Zeyrek et al. 2018) and the Turkish Discourse Bank (TDB, Zeyrek and Kurfalı 2017).

The PDTB is built upon the 1 million word Wall Street Journal corpus and is the largest available resource for discourse relations. Most related work uses PDTB 2.0, so we include this for comparing our baseline to previous work.

The recently released PDTB 3.0 adopts a new annotation schema as well as an updated sense hierarchy. PDTB 3.0 includes the annotations of PDTB 2.0 updated according to the new annotation schema, as well as about 13 thousand new annotations, of which about 5K are implicit relations (Prasad et al., 2018). The distribution of the top level senses of the implicit discourse relations in both PDTB versions is provided in \Freftab:pdtb-stats.

TED-MDB (Zeyrek et al., 2018) is the first parallel corpus annotated for discourse relations. It closely follows the PDTB 3.0 framework and includes the manual annotations of six TED talks in seven languages (English, Turkish, European Portuguese, Polish, German, Russian) aiming to allow crosslingual comparison of discourse relations111The TED-MDB annotations are available at: https://github.com/MurathanKurfali/Ted-MDB-Annotations. It has recently also been updated with Lithuanian (Oleskeviciene et al., 2019).

Despite the high number of languages covered by TED-MDB, the amount of annotated text per language is limited (see Table 2). Therefore, in the current study, we limit ourselves with the top level senses, namely Expansion, Contingency, Comparison and Temporal. We only use TED-MDB for evaluation.

Among the TED-MDB languages other than English, only Turkish has another corpus annotated with PDTB 3.0 discourse annotations, namely the Turkish Discourse Bank (TDB). TDB is a multi-genre corpus of 40 000 words, considerably less than the PDTB (see Table 2), but it provides the only directly comparable baseline to assess the performance of zero-shot learning.

4 Model

The main purpose of this study is to assess the performance of transfer learning on the implicit discourse relation classification task. To this end, we use a simple feedforward network fed with multilingual sentence embeddings following the finding of Rutherford et al. (2017) which shows that simple discourse models with feedforward layers perform on par or better than those of with surface features or recurrent and convolutional architectures.

We follow the model of Rutherford and Xue (2016) due to its simplicity and robust nature even in the multilingual setting with different argument and discourse relation representations. We represent the arguments of the discourse relation via pre-trained LASER model Artetxe and Schwenk (2018). LASER is chosen as it is the current state-of-the-art model on several Natural Language Inference (NLI) transfer learning tasks, a sentence relation classification problem similar to discourse relation classification.

Given the argument vectors, $V_{arg1}$ and $V_{arg2}$, the next step is to represent the discourse relation in a way that the interactions between them are captured. To this end, we model the discourse relation vector, $V_{dr}$, by performing the following pair-wise vector operations following the DisSent model of Nie et al. (2017):

[TABLE]

[TABLE]

[TABLE]

[TABLE]

The resulting vector is further fed into a hidden layer $h_{t}$ with $d$ hidden units222We use d=100 in the experiments to achieve a more abstract representation of the relation and finally the output $o$ is calculated using the sigmoid function. This model is also essentially the same as was used by Artetxe and Schwenk (2018) for NLI transfer learning.

5 Experiments

We formulate the implicit relation classification as four ”one vs other” binary classification task. We follow the conventional setting of the first study Pitler et al. (2009) and split the PDTB 2.0 into training (sections 2-20), development (sections 0-1 and 23-24) and test sets (sections 21-22) to have directly comparable results with the previous work. However, following the PDTB’s original distinction but unlike some previous work, we distinguish Entity-based relations from implicit relations. Each classifier is trained on an equal number of positive and negative instances where the negative instances are randomly selected in each epoch to have a better representation of the data during the training. This model is evaluated on the PDTB 2.0 test set to confirm whether our model performs adequatly on same-language, same-domain data. These results are directly comparable to previous work.

As TED-MDB is annotated according to the PDTB 3.0 framework, we train separate classifiers on PDTB 3.0 following the same convention as above. We test the trained models on all the implicit discourse relations in the TED-MDB corpus.

The PDTB framework allows annotations to be labelled with more than one label. In such cases we only keep the first label, in line with previous studies (among others Ji and Eisenstein, 2015; Rutherford et al., 2017).

The argument vectors are kept fixed during the training, and we do not update the parameters of the LASER model. We use cross-entropy loss, and AdaGrad as the optimizer. We evaluate using the model which achieved the highest F-score on the development set. As for the regularization, we use a dropout layer between the input and the hidden layer with a dropout probability of 0.3. All models are run 100 times to estimate the variance due to random initialization and stochastic training. All the models are implemented in PyTorch333https://pytorch.org/.

6 Results and Discussion

\Fref

tab:pdtb2 shows the same-language, same-domain performance of our system, in comparison to previous work. All figures refer to PDTB 2.0 test set F-score, when trained on the PDBT 2.0 training set, and are directly comparable. While our model does not achieve state-of-the-art performance in this setting, this experiment shows that it performs adequately for English, and provides a reasonable baseline for the zero-shot experiments presented in Tables 4 and 5. We also include a naive baseline system which always predicts true and is evaluated on the respective (PDTB 2.0 or PDTB 3.0) test set in our comparisons.

In all zero-shot experiments, evaluation is performed on the available test data with PDTB 3.0 annotations: TED-MDB, and the PDTB 3.0 test set itself. Results in \Freftab:pdtb3_res use PDTB 3.0 only for training, whereas \Freftab:tdb-effect presents the effect of having additional training data from Turkish (a language unrelated to English). Pooling training data from different languages is possible since our model is language-agnostic.

In all zero-shot experiments, we see similar levels of performance across all the evaluated languages in TED-MDB. While not completely comparable numerically since annotations differ slightly between languages, this evaluation set consists of parallel sentences annotated according to the same guidelines. The similarity in scores between the training language(s)—English and/or Turkish—and the remaining languages indicates that little accuracy is lost during transfer.

Comparing the performances with and without additional Turkish data, TDB, reveals that adding a small amount (relative to the size of PDTB 3.0) of Turkish training data improves the F-scores by a statistically significant amount444On a sense-wise analysis, we observe that the main increase is in the Expansion relations; however, there is no decrease in any of the other senses. for not only Turkish, but for all the languages in TED-MDB \Freftab:tdb-effect.

7 Conclusion

In the current paper we have presented the (to the best of our knowledge) first study of zero-shot learning in the implicit discourse relation classification task. Our method does not require any discourse level annotation for the target languages, yet still achieves good performance even for those languages where no training data is available. The performance is further increased by pooling training data from multiple languages. Using our published code555https://github.com/MurathanKurfali/multilingual_IDRC and publicly available resources it can used for implicit discourse classification in nearly a hundred languages.

Acknowledgments

We would like to thank Bonnie Webber for her help in obtaining PDTB 3.0 and Mats Wirén for his useful comments.