To Softmax, or not to Softmax: that is the question when applying Active Learning for Transformer Models

Despite achieving state-of-the-art results in nearly all Natural Language Processing applications, fine-tuning Transformer-based language models still requires a significant amount of labeled data to work. A well known technique to reduce the amount of human effort in acquiring a labeled dataset is \textit{Active Learning} (AL): an iterative process in which only the minimal amount of samples is labeled. AL strategies require access to a quantified confidence measure of the model predictions. A common choice is the softmax activation function for the final layer. As the softmax function provides misleading probabilities, this paper compares eight alternatives on seven datasets. Our almost paradoxical finding is that most of the methods are too good at identifying the true most uncertain samples (outliers), and that labeling therefore exclusively outliers results in worse performance. As a heuristic we propose to systematically ignore samples, which results in improvements of various methods compared to the softmax function.

ImitAL: Learned Active Learning Strategy on Synthetic Data

Active Learning (AL) is a well-known standard method for efficiently obtaining annotated data by first labeling the samples that contain the most information based on a query strategy. In the past, a large variety of such query strategies has been proposed, with each generation of new strategies increasing the runtime and adding more complexity. However, to the best of our our knowledge, none of these strategies excels consistently over a large number of datasets from different application domains. Basically, most of the the existing AL strategies are a combination of the two simple heuristics informativeness and representativeness, and the big differences lie in the combination of the often conflicting heuristics. Within this paper, we propose ImitAL, a domain-independent novel query strategy, which encodes AL as a learning-to-rank problem and learns an optimal combination between both heuristics. We train ImitAL on large-scale simulated AL runs on purely synthetic datasets. To show that ImitAL was successfully trained, we perform an extensive evaluation comparing our strategy on 13 different datasets, from a wide range of domains, with 7 other query strategies.

ImitAL: Learning Active Learning Strategies from Synthetic Data

One of the biggest challenges that complicates applied supervised machine learning is the need for huge amounts of labeled data. Active Learning (AL) is a well-known standard method for efficiently obtaining labeled data by first labeling the samples that contain the most information based on a query strategy. Although many methods for query strategies have been proposed in the past, no clear superior method that works well in general for all domains has been found yet. Additionally, many strategies are computationally expensive which further hinders the widespread use of AL for large-scale annotation projects. We, therefore, propose ImitAL, a novel query strategy, which encodes AL as a learning-to-rank problem. For training the underlying neural network we chose Imitation Learning. The required demonstrative expert experience for training is generated from purely synthetic data. To show the general and superior applicability of \ImitAL{}, we perform an extensive evaluation comparing our strategy on 15 different datasets, from a wide range of domains, with 10 different state-of-the-art query strategies. We also show that our approach is more runtime performant than most other strategies, especially on very large datasets.

RETRO: Relation Retrofitting For In-Database Machine Learning on Textual Data

There are massive amounts of textual data residing in databases, valuable for many machine learning (ML) tasks. Since ML techniques depend on numerical input representations, word embeddings are increasingly utilized to convert symbolic representations such as text into meaningful numbers. However, a naive one-to-one mapping of each word in a database to a word embedding vector is not sufficient and would lead to poor accuracies in ML tasks. Thus, we argue to additionally incorporate the information given by the database schema into the embedding, e.g. which words appear in the same column or are related to each other. In this paper, we propose RETRO (RElational reTROfitting), a novel approach to learn numerical representations of text values in databases, capturing the best of both worlds, the rich information encoded by word embeddings and the relational information encoded by database tables. We formulate relation retrofitting as a learning problem and present an efficient algorithm solving it. We investigate the impact of various hyperparameters on the learning problem and derive good settings for all of them. Our evaluation shows that the proposed embeddings are ready-to-use for many ML tasks such as classification and regression and even outperform state-of-the-art techniques in integration tasks such as null value imputation and link prediction.