In this paper, we present our submission to Shared Metrics Task: RoBLEURT (Robustly Optimizing the training of BLEURT). After investigating the recent advances of trainable metrics, we conclude several aspects of vital importance to obtain a well-per
formed metric model by: 1) jointly leveraging the advantages of source-included model and reference-only model, 2) continuously pre-training the model with massive synthetic data pairs, and 3) fine-tuning the model with data denoising strategy. Experimental results show that our model reaching state-of-the-art correlations with the WMT2020 human annotations upon 8 out of 10 to-English language pairs.
We propose pre-finetuning, an additional large-scale learning stage between language model pre-training and fine-tuning. Pre-finetuning is massively multi-task learning (around 50 datasets, over 4.8 million total labeled examples), and is designed to
encourage learning of representations that generalize better to many different tasks. We show that pre-finetuning consistently improves performance for pretrained discriminators (e.g. RoBERTa) and generation models (e.g. BART) on a wide range of tasks (sentence prediction, commonsense reasoning, MRC, etc.), while also significantly improving sample efficiency during fine-tuning. We also show that large-scale multi-tasking is crucial; pre-finetuning can hurt performance when few tasks are used up until a critical point (usually above 15) after which performance improves linearly in the number of tasks.
The transformer-based pre-trained language models have been tremendously successful in most of the conventional NLP tasks. But they often struggle in those tasks where numerical understanding is required. Some possible reasons can be the tokenizers a
nd pre-training objectives which are not specifically designed to learn and preserve numeracy. Here we investigate the ability of text-to-text transfer learning model (T5), which has outperformed its predecessors in the conventional NLP tasks, to learn numeracy. We consider four numeracy tasks: numeration, magnitude order prediction, finding minimum and maximum in a series, and sorting. We find that, although T5 models perform reasonably well in the interpolation setting, they struggle considerably in the extrapolation setting across all four tasks.
Supplementary Training on Intermediate Labeled-data Tasks (STILT) is a widely applied technique, which first fine-tunes the pretrained language models on an intermediate task before on the target task of interest. While STILT is able to further impro
ve the performance of pretrained language models, it is still unclear why and when it works. Previous research shows that those intermediate tasks involving complex inference, such as commonsense reasoning, work especially well for RoBERTa-large. In this paper, we discover that the improvement from an intermediate task could be orthogonal to it containing reasoning or other complex skills --- a simple real-fake discrimination task synthesized by GPT2 can benefit diverse target tasks. We conduct extensive experiments to study the impact of different factors on STILT. These findings suggest rethinking the role of intermediate fine-tuning in the STILT pipeline.
This paper describes the Fujitsu DMATH systems used for WMT 2021 News Translation and Biomedical Translation tasks. We focused on low-resource pairs, using a simple system. We conducted experiments on English-Hausa, Xhosa-Zulu and English-Basque, and
submitted the results for Xhosa→Zulu in the News Translation Task, and English→Basque in the Biomedical Translation Task, abstract and terminology translation subtasks. Our system combines BPE dropout, sub-subword features and back-translation with a Transformer (base) model, achieving good results on the evaluation sets.
Recent impressive improvements in NLP, largely based on the success of contextual neural language models, have been mostly demonstrated on at most a couple dozen high- resource languages. Building language mod- els and, more generally, NLP systems fo
r non- standardized and low-resource languages remains a challenging task. In this work, we fo- cus on North-African colloquial dialectal Arabic written using an extension of the Latin script, called NArabizi, found mostly on social media and messaging communication. In this low-resource scenario with data display- ing a high level of variability, we compare the downstream performance of a character-based language model on part-of-speech tagging and dependency parsing to that of monolingual and multilingual models. We show that a character-based model trained on only 99k sentences of NArabizi and fined-tuned on a small treebank of this language leads to performance close to those obtained with the same architecture pre- trained on large multilingual and monolingual models. Confirming these results a on much larger data set of noisy French user-generated content, we argue that such character-based language models can be an asset for NLP in low-resource and high language variability set- tings.
The main idea of this solution has been to focus on corpus cleaning and preparation and after that, use an out of box solution (OpenNMT) with its default published transformer model. To prepare the corpus, we have used set of standard tools (as Moses
scripts or python packages), but also, among other python scripts, a python custom tokenizer with the ability to replace numbers for variables, solve the upper/lower case issue of the vocabulary and provide good segmentation for most of the punctuation. We also have started a line to clean corpus based on statistical probability estimation of source-target corpus, with unclear results. Also, we have run some tests with syllabical word segmentation, again with unclear results, so at the end, after word sentence tokenization we have used BPE SentencePiece for subword units to feed OpenNMT.
Incorporating lexical knowledge into deep learning models has been proved to be very effective for sequence labeling tasks. However, previous works commonly have difficulty dealing with large-scale dynamic lexicons which often cause excessive matchin
g noise and problems of frequent updates. In this paper, we propose DyLex, a plug-in lexicon incorporation approach for BERT based sequence labeling tasks. Instead of leveraging embeddings of words in the lexicon as in conventional methods, we adopt word-agnostic tag embeddings to avoid re-training the representation while updating the lexicon. Moreover, we employ an effective supervised lexical knowledge denoising method to smooth out matching noise. Finally, we introduce a col-wise attention based knowledge fusion mechanism to guarantee the pluggability of the proposed framework. Experiments on ten datasets of three tasks show that the proposed framework achieves new SOTA, even with very large scale lexicons.
The task of converting a nonstandard text to a standard and readable text is known as lexical normalization. Almost all the Natural Language Processing (NLP) applications require the text data in normalized form to build quality task-specific models.
Hence, lexical normalization has been proven to improve the performance of numerous natural language processing tasks on social media. This study aims to solve the problem of Lexical Normalization by formulating the Lexical Normalization task as a Sequence Labeling problem. This paper proposes a sequence labeling approach to solve the problem of Lexical Normalization in combination with the word-alignment technique. The goal is to use a single model to normalize text in various languages namely Croatian, Danish, Dutch, English, Indonesian-English, German, Italian, Serbian, Slovenian, Spanish, Turkish, and Turkish-German. This is a shared task in 2021 The 7th Workshop on Noisy User-generated Text (W-NUT)'' in which the participants are expected to create a system/model that performs lexical normalization, which is the translation of non-canonical texts into their canonical equivalents, comprising data from over 12 languages. The proposed single multilingual model achieves an overall ERR score of 43.75 on intrinsic evaluation and an overall Labeled Attachment Score (LAS) score of 63.12 on extrinsic evaluation. Further, the proposed method achieves the highest Error Reduction Rate (ERR) score of 61.33 among the participants in the shared task. This study highlights the effects of using additional training data to get better results as well as using a pre-trained Language model trained on multiple languages rather than only on one language.
Large-scale pretrained transformer models have demonstrated state-of-the-art (SOTA) performance in a variety of NLP tasks. Nowadays, numerous pretrained models are available in different model flavors and different languages, and can be easily adapte
d to one's downstream task. However, only a limited number of models are available for dialogue tasks, and in particular, goal-oriented dialogue tasks. In addition, the available pretrained models are trained on general domain language, creating a mismatch between the pretraining language and the downstream domain launguage. In this contribution, we present CS-BERT, a BERT model pretrained on millions of dialogues in the customer service domain. We evaluate CS-BERT on several downstream customer service dialogue tasks, and demonstrate that our in-domain pretraining is advantageous compared to other pretrained models in both zero-shot experiments as well as in finetuning experiments, especially in a low-resource data setting.
