Online translation systems aim to incorporate human post-editing feedback (or the corrected version of the MT output) into their models in realtime, as soon as it becomes available. This feedback helps the system to learn from the mistakes made in the past translations and avoid to repeat them in future translations. This continuous learning capability will eventually improve the quality of the translations and consequently increase the productivity of the translators/post-editors (Tatsumi, 2009) working with MT suggestions in a CAT environment. The basic workflow of an online translation system goes through the following steps repeatedly: i) the system receives an input segment; ii) the input segment is translated and provided to the post-editor to fix any errors 1A domain is made of segments belonging to the same text genre and the MT outputs are generated by the same MT system. in it; and iii) the human post-edited version of the translation is incorporated back into the system, by stepwise updating the underlying models and parameters. In the APE context, the input is a machine-translated segment (optionally with its corresponding source segment), which is processed by the online APE system to fix errors, and then verified by the post-editors. Several online translation systems have been proposed over the years (Hardt and Elming, 2010; Bertoldi et al., 2013; Mathur et al., 2013; Simard and Foster, 2013; Ortiz-Mart¨ınez and Casacuberta, 2014; Denkowski et al., 2014; Wuebker et al., 2015) .

The state-of-the-art online APE system is the Thot toolkit (Ortiz-Mart¨ınez and Casacuberta, 2014) that has been previously developed to support fully automatic and interactive statistical machine translation and then used in the APE task (Lagarda et al., 2015) . To update the inner models with the user feedback, a set of sufficient statistics was maintained and incrementally updated. In the case of language model, only the n-gram counts are required to maintain sufficient statistics. To update the translation model, an incremental version of EM algorithm is used to first obtain word alignment and then phrase pairs counts were extracted to update the sufficient statistics. Other features like source/target phrase-length models or distortion model are implemented by means of geometric distributions with fixed parameters. However, Thot differs from our approach because it does not embed any techniques for selecting the most relevant training data. In the long-run, when data points from different domain are continuously analysed, this system tends to become more and more generic, which may not be useful and even harmful for automatically post-editing domain-specific segments.

To preserve all the knowledge gained in the online learning process and at the same time being able to apply specific post-editing rules when needed, we propose an instance selection technique for online APE that has the ability to retrieve specific data points whose context is similar to the segment to be post-edited. These data points are then used to build reliable APE models. When there are no reliable data points in the knowledge base, the MT output is kept untouched, as opposed to the existing APE systems, which tends to always translate the given input segment independently from the reliability of the applicable correction rules.

Our proposed algorithm emulates an online APE system and assumes to have the following data to run the online experiments: i) source (src); ii) MT output (mt); and iii) human post-edits (pe) of the MT output. At the beginning the knowledge base of our online APE system is empty and it will be updated whenever an instance (a tuple containing parallel segments from all the above mentioned documents) is processed. When the system receives an input (src, mt), the most relevant training instances from a pool of multi-domain data stored in our knowledge base are retrieved. The similarity between the training instances and the input segment is measured by a score based on the term frequency inverse document frequency (tfidf ), generally used in information retrieval. The larger the number of words in common between the training and the input sentences, the higher is the score. In our system, these scores are computed using the Lucene library.2 Only those training instances that have similarity score above a certain threshold (decided over a held-out development set) are used to build: i) a tri-gram local language model over the target side of the training corpus with the IRSTLM toolkit (Federico et al., 2008) ; ii) the translation and reordering models using the Moses toolkit (Koehn et al., 2007) and the word alignment of each sentence pair is computed using the incremental GIZA++ software.3 The log-linear model parameters are optimized over a part of the selected instances. To obtain reliably-tuned weights and a fast optimization process, multiple instances of MIRA (Chiang, 2012) are run in parallel on three small development sets randomly selected from the retrieved sentences. The obtained weights are then averaged. If a minimum value of retrieved sentences is not reached, the optimization step is skipped because having few sentences might not yield reliable weights. In this case, the weights computed on the previous input segment are used. The tuned weights and the models built on all the data are then used to post-edit the input sentences.

In a real translation workflow, the APE segment is then passed to the human translator that creates the post-edited segment. Once the post-edit is 2https://lucene.apache.org/ 3https://code.google.com/archive/p/ inc-giza-pp/ available it is added to the knowledge base along with the source and the mt sentences. In our experiments we emulate the post-edited sentence of the APE segment with the post-edit of the mt output. 4

Data To examine the performance of the online APE systems in a multi-domain translation environment, we select two data sets for the EnglishGerman language pair belonging to information technology (IT). Although they come from the same category (IT), they feature variability in terms of vocabulary coverage, MT errors, and post-editing style. The two data sets are respectively a subset of the Autodesk Post-Editing Data corpus and the resources used at the second round of the APE shared task at the first conference on machine translation (WMT2016).4 The data sets are pre-processed to obtain a joint-representation that links each source word with a MT word (mt#src). This representation has been proposed in the context-aware APE approach by (Be´chara et al., 2011) and leverages the source information to disambiguate post-editing rules. Recently, (Chatterjee et al., 2015b) also confirmed this approach to work better than translating from raw MT segments over multiple language pairs. The jointrepresentation is used as a source corpus to train all the APE systems reported in this paper and it is obtained by first aligning the words of source (src) and MT (mt) segments using MGIZA++ (Gao and Vogel, 2008) , and then each mt word is concatenated with its corresponding src words.

The Autodesk training, and development sets consist of 12,238, and 1,948 segments respectively, while the WMT2016 data contains 12,000, and 1,000 segments. To measure the diversity of the two data sets we compute the vocabulary overlap between the two joint-representations. This is performed internally to each data set (splitting the training data in two halves) and across them. As expected, in the first case the vocabulary overlap is much larger (> 40%) than in the second one ( 15%); this indicates that the two data sets are quite different and few information can be shared.

To emulate the multi-domain scenario, the two training data sets are first merged together and then shuffled. The same strategy is also used for the development sets. This represents the situation in 4http://www.statmt.org/wmt16/ ape-task.html which an APE system serves two CAT tools that process documents from two domains and the sequence of points is random. Our approach and the competitors are run on all the shuffled training data and evaluated on the second half (12,100 points). Evaluation metrics The performance of the different APE systems is evaluated using the Translation Error rate (TER) (Snover et al., 2006) , BLEU (Papineni et al., 2002) and the precision (Chatterjee et al., 2015a) . TER and BLEU measures the similarity between the MT outputs and their references by looking at the word/n-gram overlaps, while precision is the ratio of number of sentences an APE system improves (with respect to the MT output) over all the sentences it modifies.5 Larger values indicate that the APE system is able to improve the quality of most of the sentences it changes. The statistical significance test for BLEU is computed using the paired bootstrap resampling technique (Koehn, 2004) , and for TER using the stratified approximate randomization technique (Clark et al., 2011) .

Terms of comparison We evaluate our online learning approach against the output produced by the MT system, the batch APE system that follows the approach proposed in (Chatterjee et al., 2015b) , and the Thot toolkit. 5

The main goal of this research is to examine the performance of online APE methods in a multidomain scenario, where the APE system receives a stream of data coming from different domains. The parameters of our approach (i.e. similarity score threshold and minimum number of selected sentence) are optimised following the grid search strategy. We set the threshold values to 1 and the minimum number of selected sentences to 20. The results of all the systems are reported in Table 1.

The batch APE system that is trained only on the first half of the data is able to slightly improve the performance of the MT system, but it damages most of the sentence it changes (precision smaller than 45%). Although Thot can learn from all the data, it is interesting to note that it does not significantly improve over the MT system and the batch APE system. This suggests that using all the data 5For each sentence in the test set, if the TER score of APE system is different than the baseline then it is considered as a modified sentence MT Batch APE Thot Our approach without considering the peculiarities of each domain does not allow an APE system to efficiently learn reliable correction rules and to improve the machine translation quality. Moreover, these results also show that few information can be shared between the two data sets. This is expected considering the limited overlap between the two corpora.

Our approach provides significant improvements in BLEU, TER and precision over all the competitors. In particular, it can obtain more than one TER and BLEU point improvement, and more than 20% precision points increment over the best APE system (the Thot toolkit). Such gains confirm that the instance selection mechanism allows our APE system to identify domain-specific data and to leverage it for extracting reliable correction rules. Further analysis of the performance of the online systems revealed that our approach modifies less segments compared with Thot, because it builds a model only if it finds relevant data, leaving the MT segment untouched otherwise. These untouched MT segments, when modified by Thot, often lead to deterioration. This suggests that, the output obtained with our solution has a higher potential for being useful to human translators. Such usefulness comes not only in terms of a more pleasant post-editing activity, but also in terms of time savings yield by overall better suggestions. 6

We addressed the problem of building a robust online APE system that is able to efficiently work on a stream of data points belonging to different domains. In this condition, our APE has shown its capability to continuously adapt to the dynamics of diverse data processed in real-time. In particular, the instance selection mechanism allows our APE method to reduce the number of wrong modifications, which result in significant improvements in precision over the state-of-the-art online APE system, and thus making it a viable solution to be deployed in a real-word CAT framework.

This work has been partially supported by the ECfunded H2020 project QT21 (grant agreement no. 645452).