A key challenge in the creation and refinement of virtual assistants is the ability to mine unlabeled utterance data to discover common intents. We develop an approach to this problem that combines large-scale pre-training and multi-task learning to derive a semantic embedding that can be leveraged to identify clusters of utterances that correspond to unhandled intents. An utterance encoder is first trained with a language modeling objective and subsequently adapted to predict intent labels from a large collection of cross-domain enterprise virtual assistant data using a multi-task cosine softmax loss. Experimental evaluation shows significant advantages for this multi-step pre-training approach, with large gains in downstream clustering accuracy on new applications compared to standard sentence embedding approaches. The approach has been incorporated into an interactive discovery tool that enables visualization and exploration of intents by system analysts and builders.

To understand how training on conversational language impacts performance of pre-trained models on downstream dialogue tasks, we build compact Transformer-based Language Models from scratch on several large corpora of conversational data. We compare the performance and characteristics of these models against BERT and other strong baselines on dialogue probing tasks. Commercial dialogue systems typically require a small footprint and fast execution time, but recent trends are in the other direction, with an ever-increasing number of parameters, resulting in difficulties in model deployment. We focus instead on training fast, lightweight models that excel at natural language understanding (NLU) and can replace existing lower-capacity conversational AI models with similar size and speed. In the process, we develop a simple but unique curriculum-based approach that moves from general-purpose to dialogue-targeted both in terms of data and objective. Our resultant models have around 1/3 the number of parameters of BERT-base and produce better representations for a wide array of intent detection datasets using linear and Mutual-Information probing techniques. Additionally, the models can be easily fine-tuned on a single consumer GPU card and deployed in near real-time production environments.

Current state-of-the-art models for named entity recognition (NER) are neural models with a conditional random field (CRF) as the final layer. Entities are represented as per-token labels with a special structure in order to decode them into spans. Current work eschews prior knowledge of how the span encoding scheme works and relies on the CRF learning which transitions are illegal and which are not to facilitate global coherence. We find that by constraining the output to suppress illegal transitions we can train a tagger with a cross-entropy loss twice as fast as a CRF with differences in F1 that are statistically insignificant, effectively eliminating the need for a CRF. We analyze the dynamics of tag co-occurrence to explain when these constraints are most effective and provide open source implementations of our tagger in both PyTorch and TensorFlow.

Because large, human-annotated datasets suffer from labeling errors, it is crucial to be able to train deep neural networks in the presence of label noise. While training image classification models with label noise have received much attention, training text classification models have not. In this paper, we propose an approach to training deep networks that is robust to label noise. This approach introduces a non-linear processing layer (noise model) that models the statistics of the label noise into a convolutional neural network (CNN) architecture. The noise model and the CNN weights are learned jointly from noisy training data, which prevents the model from overfitting to erroneous labels. Through extensive experiments on several text classification datasets, we show that this approach enables the CNN to learn better sentence representations and is robust even to extreme label noise. We find that proper initialization and regularization of this noise model is critical. Further, by contrast to results focusing on large batch sizes for mitigating label noise for image classification, we find that altering the batch size does not have much effect on classification performance.

We introduce Baseline: a library for reproducible deep learning research and fast model development for NLP. The library provides easily extensible abstractions and implementations for data loading, model development, training and export of deep learning architectures. It also provides implementations for simple, high-performance, deep learning models for various NLP tasks, against which newly developed models can be compared. Deep learning experiments are hard to reproduce, Baseline provides functionalities to track them. The goal is to allow a researcher to focus on model development, delegating the repetitive tasks to the library.