Transfer Learning has been shown to be a powerful tool for Natural Language Processing (NLP) and has outperformed the standard supervised learning paradigm, as it takes benefit from the pre-learned knowledge. Nevertheless, when transfer is performed between less related domains, it brings a negative transfer, i.e. hurts the transfer performance. In this research, we shed light on the hidden negative transfer occurring when transferring from the News domain to the Tweets domain, through quantitative and qualitative analysis. Our experiments on three NLP taks: Part-Of-Speech tagging, Chunking and Named Entity recognition, reveal interesting insights.

Two prevalent transfer learning approaches are used in recent works to improve neural networks performance for domains with small amounts of annotated data: Multi-task learning which involves training the task of interest with related auxiliary tasks to exploit their underlying similarities, and Mono-task fine-tuning, where the weights of the model are initialized with the pretrained weights of a large-scale labeled source domain and then fine-tuned with labeled data of the target domain (domain of interest). In this paper, we propose a new approach which takes advantage from both approaches by learning a hierarchical model trained across multiple tasks from a source domain, and is then fine-tuned on multiple tasks of the target domain. Our experiments on four tasks applied to the social media domain show that our proposed approach leads to significant improvements on all tasks compared to both approaches.

L’apprentissage par transfert représente la capacité qu’un modèle neuronal entraîné sur une tâche à généraliser suffisamment et correctement pour produire des résultats pertinents sur une autre tâche proche mais différente. Nous présentons dans cet article une approche fondée sur l’apprentissage par transfert pour construire automatiquement des outils d’analyse de textes des réseaux sociaux en exploitant les similarités entre les textes d’une langue bien dotée (forme standard d’une langue) et les textes d’une langue peu dotée (langue utilisée en réseaux sociaux). Nous avons expérimenté notre approche sur plusieurs langues ainsi que sur trois tâches d’annotation linguistique (étiquetage morpho-syntaxique, annotation en parties du discours et reconnaissance d’entités nommées). Les résultats obtenus sont très satisfaisants et montrent l’intérêt de l’apprentissage par transfert pour tirer profit des modèles neuronaux profonds sans la contrainte d’avoir à disposition une quantité de données importante nécessaire pour avoir une performance acceptable.

Fine-tuning neural networks is widely used to transfer valuable knowledge from high-resource to low-resource domains. In a standard fine-tuning scheme, source and target problems are trained using the same architecture. Although capable of adapting to new domains, pre-trained units struggle with learning uncommon target-specific patterns. In this paper, we propose to augment the target-network with normalised, weighted and randomly initialised units that beget a better adaptation while maintaining the valuable source knowledge. Our experiments on POS tagging of social media texts (Tweets domain) demonstrate that our method achieves state-of-the-art performances on 3 commonly used datasets.

In this paper, we describe a morpho-syntactic tagger of tweets, an important component of the CEA List DeepLIMA tool which is a multilingual text analysis platform based on deep learning. This tagger is built for the Morpho-syntactic Tagging of Tweets (MTT) Shared task of the 2018 VarDial Evaluation Campaign. The MTT task focuses on morpho-syntactic annotation of non-canonical Twitter varieties of three South-Slavic languages: Slovene, Croatian and Serbian. We propose to use a neural network model trained in an end-to-end manner for the three languages without any need for task or domain specific features engineering. The proposed approach combines both character and word level representations. Considering the lack of annotated data in the social media domain for South-Slavic languages, we have also implemented a cross-domain Transfer Learning (TL) approach to exploit any available related out-of-domain annotated data.