Research Seminar

We are pleased to announce we are restarting our research seminars  and aiming for this to be a monthly series on the first Thursday of every month. Please find details of the first seminar below.

Date:                     1st October 2020

Time:                     11:00

Platform:             Teams, please email A[dot]harper2[at]wlv.ac.uk for a link

TransQuest: Translation Quality Estimation with Cross-lingual Transformers

Tharindu Ranasinghe

High-accuracy translation quality estimation (QE) that can be easily deployed for a number of language pairs is the missing piece in many commercial translation workflows. Even though there are many systems that can do QE, majority of these methods work only on the language pair they are trained on and need retraining for new language pairs which can be usually computationally expensive and difficult especially for low-resource language pairs. As a solution, in this presentation, we introduce TransQuest – a simple QE framework based on cross-lingual transformers. TransQuest outperforms the current state of the art quality estimation methods like DeepQuest and OpenKiwi. This is also the winning solution in recently concluded WMT 2020 sentence-level Direct Assessment shared task, winning all the language pairs with the multilingual track too. 

Wolverhampton researchers collaborate on BBC’s “Novels That Shaped Our World”

We are working with BBC Arts and Faculty of Arts on a new engagement project to mark the 300th anniversary of the English Language Novel.

The interdisciplinary project unites research by Wolverhampton’s Research Group for Computational Linguistics, including Dr Sara Moze, Richard Evans and Dr Emad Mohamed, and from English and Creative Writing staff, led by Dr Aidan Byrne. The project is seeking support from the Arts and Humanities Research Council.

There is more information on the University website if you would like to find out more about the project: https://www.wlv.ac.uk/about-us/news-and-events/latest-news/2019/august-2019/wolverhampton-researchers-collaborate-on-bbcs-novels-that-shaped-our-world.php

Paper accepted at NAACL 2019

Congratulations to Marcos Zampieri, whose paper has been accepted at NAACL 2019.

Reference:  Marcos  Zampieri,  Shervin  Malmasi,  Preslav  Nakov, Sara  Rosenthal,  Noura  Farra,  and  Ritesh  Kumar (2019)  Predicting the Type and Target of Offensive Posts in Social Media. 

You may access the NAACL paper here: https://arxiv.org/abs/1902.09666

Paper accepted at NAACL 2019

We are pleased to announce that the paper titled “Bridging the Gap: Attending to Discontinuity in Identification of Multiword Expressions” from researchers in RGCL has been accepted into the main track of the Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL 2019). This is joint work by Omid Rohanian, Shiva Taslimipoor, Le An Ha, Samaneh Kouchaki, and Prof. Ruslan Mitkov.

A preprint of this paper will soon be available on ArXiv.

Datasets annotated with signs of syntactic complexity

In our paper:

Evans, R., & Orasan, C. (2013). Annotating signs of syntactic complexity to support sentence simplification. In I. Habernal & V. Matousek (Eds.), Text, Speech and Dialogue. Proceedings of the 16th International Conference TSD 2013. Plzen, Czech Republic: Springer. pp. 92 – 104

we present the annotation of a dataset that is used by our syntactic simplification method to identify places where rewriting rules have to be applied in order to produce simpler sentences. 

The datasets are available in XML format as three independent files, each representing a different genre

Each file contains a list of sentences annotated using the following format:

<S ID="2"><SIGN ID="2" CLASS="SSEV">That</SIGN> is 
<SIGN ID="3" CLASS="HELP">,</SIGN> a high-fibre diet, 
fluid <SIGN ID="4" CLASS="CLN">,</SIGN> etc.</S>

The sentences are marked using the S tag, whilst the signs by the tag SIGN. The type of sign is encoded by the attribute CLASS. The sentences were annotated in isolation, so the files above do not contain coherent texts, but sequences of sentences extracted from different files. 

To understand the difference between different classes and how the annotation process was carried out please consult the annotation guidelines. Specific questions about the annotation should be sent to Richard Evans. A demo of the sign tagger is available at http://rgcl.wlv.ac.uk/demos/SignTaggerWebDemo/  

You can find out more about our approach for syntactic simplification in our recent paper

Evans, Richard, and Constantin Orǎsan. 2018. “Identifying Signs of Syntactic Complexity for Rule-Based Sentence Simplification.” Natural Language Engineering. https://doi.org/10.1017/S1351324918000384.

Sheila Castilho and Natalia Resende spend a week at RGCL

This week we have had the pleasure of welcoming Dr Sheila Castilho and Dr Natalia Resende for a one week research stay at the Research Group in Computational Linguistics. Sheila and Natalia both come from the ADAPT Centre, Dublin and have come to discuss collaborations with members of our research group. During their stay, both Natalia and Sheila gave the group a talk about their research. The details of which can be found below:- 

 

Speaker: Dr Sheila Castilho

Date of talk: 19th November 2018

Title: Attaining the Unattainable? Reassessing Claims of Human Parity in Neural Machine Translation

Abstract: We reassess a recent study (Hassan et al., 2018) that claimed that machine translation (MT) has reached human parity for the translation of news from Chinese into English, using pairwise ranking and considering three variables that were not taken into account in that previous study: the language in which the source side of the test set was originally written, the translation proficiency of the evaluators, and the provision of inter-sentential context. If we consider only original source text (i.e. not translated from another language, or translationese), then we find evidence showing that human parity has not been achieved. We compare the judgments of professional translators against those of non-experts and discover that those of the experts result in higher inter-annotator agreement and better discrimination between human and machine translations. In addition, we analyse the human translations of the test set and identify important translation issues. Finally, based on these findings, we provide a set of recommendations for future human evaluations of MT.

Speaker: Dr Natalia Resende

Date of talk: 21st November 2018

Title: Classifying nouns in Portuguese into gender categories: a deep learning approach

Abstract: In Portuguese, all nouns are distributed into two gender categories: feminine and masculine. On one hand, gender can be predicted from the phonological cues present in the endings of the nouns. For example, nouns ending in -a  tend to be feminine and nouns ending in -o   tend to be masculine. On the other hand, the relationship between word ending and gender is far from being a consistent rule, since nouns ending in other phonemes may be of either gender. In the present study, a connectionist network was trained to classify Portuguese nouns into gender categories considering their phonological structure as whole. The performance of the network was analysed in detail to check whether the network considers only the endings of the nouns or their whole phonological structure for gender decisions. In addition, it was analysed what type of information the network takes into account to decide the gender of nouns whose endings are not predictive of gender. Results show an error-free performance when the network takes into account the phonological information present in the endings of the nouns and frequency effects for nonpredictive endings. The present study has implications to the training of NLP systems when classifying nouns into gender categories.