In this work, we propose to model the interaction between visual and textual features for multi-modal neural machine translation (MMT) through a latent variable model. This latent variable can be seen as a multi-modal stochastic embedding of an image and its description in a foreign language. It is used in a target-language decoder and also to predict image features. Importantly, our model formulation utilises visual and textual inputs during training but does not require that images be available at test time. We show that our latent variable MMT formulation improves considerably over strong baselines, including a multi-task learning approach (Elliott and Kádár, 2017) and a conditional variational auto-encoder approach (Toyama et al., 2016). Finally, we show improvements due to (i) predicting image features in addition to only conditioning on them, (ii) imposing a constraint on the minimum amount of information encoded in the latent variable, and (iii) by training on additional target-language image descriptions (i.e. synthetic data).1 Code and pre-trained models will be released soon.

No abstract

This paper describes a study of the impact of coreference resolution on NLP applications. Further to our previous study [1], in which we investigated whether anaphora resolution could be beneficial to NLP applications, we now seek to establish whether a different, but related task -that of coreference resolution, could improve the performance of three NLP applications: text summarisation, recognising textual entailment and text classification. The study discusses experiments in which the aforementioned applications were implemented in two versions, one in which the BART coreference resolution system was integrated and one in which it was not, and then tested in processing input text. The paper discusses the results obtained.

Abstract-We present a Recognizing Textual Entailment (RTE) system based on different similarity metrics. The metrics used are string-based metrics and the Semantic Edit Distance Metric, which is proposed in this paper to address limitations of known semantic-based metrics and to support the decisions made by a simple method based on lexical similarity metrics. We add the scores of the metrics as features for a machine learning algorithm. The performance of our system is comparable with the average performance of the Recognizing Textual Entailment Challenges, though lower than that of the state-ofthe-art methods.

This paper describes the joint submission of the QT21 projects for the English→Latvian translation task of the EMNLP 2017 Second Conference on Machine Translation (WMT 2017). The submission is a system combination which combines seven different statistical machine translation systems provided by the different groups.The systems are combined using either RWTH's system combination approach, or USFD's consensus-based systemselection approach. The final submission shows an improvement of 0.5 BLEU compared to the best single system on newstest2017.

We report results obtained by the UoW method in SemEval-2014's Task 10 -Multilingual Semantic Textual Similarity. We propose to model Semantic Textual Similarity in the context of Multi-task Learning in order to deal with inherent challenges of the task such as unbalanced performance across domains and the lack of training data for some domains (i.e. unknown domains). We show that the Multi-task Learning approach outperforms previous work on the 2012 dataset, achieves a robust performance on the 2013 dataset and competitive results on the 2014 dataset. We highlight the importance of the challenge of unknown domains, as it affects overall performance substantially.

This work exploits translation data as a source of semantically relevant learning signal for models of word representation. In particular, we exploit equivalence through translation as a form of distributional context and jointly learn how to embed and align with a deep generative model. Our EMBEDALIGN model embeds words in their complete observed context and learns by marginalisation of latent lexical alignments. Besides, it embeds words as posterior probability densities, rather than point estimates, which allows us to compare words in context using a measure of overlap between distributions (e.g. KL divergence). We investigate our model's performance on a range of lexical semantics tasks achieving competitive results on several standard benchmarks including natural language inference, paraphrasing, and text similarity.

Abstract. We propose a novel approach to recognise textual entailment (RTE) following a two-stage architecture -alignment and decision -where both stages are based on semantic representations. In the alignment stage the entailment candidate pairs are represented and aligned using predicate-argument structures. In the decision stage, a Markov Logic Network (MLN) is learnt using rich relational information from the alignment stage to predict an entailment decision. We evaluate this approach using the RTE Challenge datasets. It achieves the best results for the RTE-3 dataset and shows comparable performance against the state of the art approaches for other datasets.