Sequential fine-tuning and multi-task learning are methods aiming to incorporate knowledge from multiple tasks; however, they suffer from catastrophic forgetting and difficulties in dataset balancing. To address these shortcomings, we propose AdapterFusion, a new two stage learning algorithm that leverages knowledge from multiple tasks. First, in the knowledge extraction stage we learn task specific parameters called adapters, that encapsulate the task-specific information. We then combine the adapters in a separate knowledge composition step. We show that by separating the two stages, i.e., knowledge extraction and knowledge composition, the classifier can effectively exploit the representations learned from multiple tasks in a non-destructive manner. We empirically evaluate AdapterFusion on 16 diverse NLU tasks, and find that it effectively combines various types of knowledge at different layers of the model. We show that our approach outperforms traditional strategies such as full fine-tuning as well as multi-task learning. Our code and adapters are available at AdapterHub.ml.
The current modus operandi in NLP involves downloading and fine-tuning pre-trained models consisting of millions or billions of parameters. Storing and sharing such large trained models is expensive, slow, and time-consuming, which impedes progress towards more general and versatile NLP methods that learn from and for many tasks. Adapters—small learnt bottleneck layers inserted within each layer of a pre-trained model— ameliorate this issue by avoiding full fine-tuning of the entire model. However, sharing and integrating adapter layers is not straightforward. We propose AdapterHub, a framework that allows dynamic “stiching-in” of pre-trained adapters for different tasks and languages. The framework, built on top of the popular HuggingFace Transformers library, enables extremely easy and quick adaptations of state-of-the-art pre-trained models (e.g., BERT, RoBERTa, XLM-R) across tasks and languages. Downloading, sharing, and training adapters is as seamless as possible using minimal changes to the training scripts and a specialized infrastructure. Our framework enables scalable and easy access to sharing of task-specific models, particularly in low-resource scenarios. AdapterHub includes all recent adapter architectures and can be found at AdapterHub.ml
Semantic specialization methods fine-tune distributional word vectors using lexical knowledge from external resources (e.g. WordNet) to accentuate a particular relation between words. However, such post-processing methods suffer from limited coverage as they affect only vectors of words seen in the external resources. We present the first post-processing method that specializes vectors of all vocabulary words – including those unseen in the resources – for the asymmetric relation of lexical entailment (LE) (i.e., hyponymy-hypernymy relation). Leveraging a partially LE-specialized distributional space, our POSTLE (i.e., post-specialization for LE) model learns an explicit global specialization function, allowing for specialization of vectors of unseen words, as well as word vectors from other languages via cross-lingual transfer. We capture the function as a deep feed-forward neural network: its objective re-scales vector norms to reflect the concept hierarchy while simultaneously attracting hyponymy-hypernymy pairs to better reflect semantic similarity. An extended model variant augments the basic architecture with an adversarial discriminator. We demonstrate the usefulness and versatility of POSTLE models with different input distributional spaces in different scenarios (monolingual LE and zero-shot cross-lingual LE transfer) and tasks (binary and graded LE). We report consistent gains over state-of-the-art LE-specialization methods, and successfully LE-specialize word vectors for languages without any external lexical knowledge.
This paper introduces rank-based training of structured prediction energy networks (SPENs). Our method samples from output structures using gradient descent and minimizes the ranking violation of the sampled structures with respect to a scalar scoring function defined with domain knowledge. We have successfully trained SPEN for citation field extraction without any labeled data instances, where the only source of supervision is a simple human-written scoring function. Such scoring functions are often easy to provide; the SPEN then furnishes an efficient structured prediction inference procedure.