In this paper, we study the utilization of pre-trained language models to enable few-shotNatural Language Generation (NLG) in task-oriented dialog systems. We introduce a system consisting of iterative self-training and an extensible mini-template framework that textualizes the structured input data into semi-natural text to fully take advantage of pre-trained language models. We compare var-ious representations of NLG models’ input and output and show that transforming the input and output to be similar to what the language model has seen before during pre-training improves the model’s few-shot performance substantially. We show that neural mod-els can be trained with as few as 300 annotated examples while providing high fidelity, considerably lowering the resource requirements for standing up a new domain or language. This level of data efficiency removes the need for crowd-sourced data collection resulting in higher quality data annotated by expert linguists. In addition, model maintenance and debugging processes will improve in this few-shot setting. Finally, we explore distillation and using a caching system to satisfy latency requirements of real-world systems.

We propose a novel framework to train models to classify acceptability of responses generated by natural language generation (NLG) models, improving upon existing sentence transformation and model-based approaches. An NLG response is considered acceptable if it is both semantically correct and grammatical. We don’t make use of any human references making the classifiers suitable for runtime deployment. Training data for the classifiers is obtained using a 2-stage approach of first generating synthetic data using a combination of existing and new model-based approaches followed by a novel validation framework to filter and sort the synthetic data into acceptable and unacceptable classes. Our 2-stage approach adapts to a wide range of data representations and does not require additional data beyond what the NLG models are trained on. It is also independent of the underlying NLG model architecture, and is able to generate more realistic samples close to the distribution of the NLG model-generated responses. We present results on 5 datasets (WebNLG, Cleaned E2E, ViGGO, Alarm, and Weather) with varying data representations. We compare our framework with existing techniques that involve synthetic data generation using simple sentence transformations and/or model-based techniques, and show that building acceptability classifiers using data that resembles the generation model outputs followed by a validation framework outperforms the existing techniques, achieving state-of-the-art results. We also show that our techniques can be used in few-shot settings using self-training.

Natural language generation (NLG) is a critical component in conversational systems, owing to its role of formulating a correct and natural text response. Traditionally, NLG components have been deployed using template-based solutions. Although neural network solutions recently developed in the research community have been shown to provide several benefits, deployment of such model-based solutions has been challenging due to high latency, correctness issues, and high data needs. In this paper, we present approaches that have helped us deploy data-efficient neural solutions for NLG in conversational systems to production. We describe a family of sampling and modeling techniques to attain production quality with light-weight neural network models using only a fraction of the data that would be necessary otherwise, and show a thorough comparison between each. Our results show that domain complexity dictates the appropriate approach to achieve high data efficiency. Finally, we distill the lessons from our experimental findings into a list of best practices for production-level NLG model development, and present them in a brief runbook. Importantly, the end products of all of the techniques are small sequence-to-sequence models (~2Mb) that we can reliably deploy in production. These models achieve the same quality as large pretrained models (~1Gb) as judged by human raters.

Multilingual contextual embeddings have demonstrated state-of-the-art performance in zero-shot cross-lingual transfer learning, where multilingual BERT is fine-tuned on one source language and evaluated on a different target language. However, published results for mBERT zero-shot accuracy vary as much as 17 points on the MLDoc classification task across four papers. We show that the standard practice of using English dev accuracy for model selection in the zero-shot setting makes it difficult to obtain reproducible results on the MLDoc and XNLI tasks. English dev accuracy is often uncorrelated (or even anti-correlated) with target language accuracy, and zero-shot performance varies greatly at different points in the same fine-tuning run and between different fine-tuning runs. These reproducibility issues are also present for other tasks with different pre-trained embeddings (e.g., MLQA with XLM-R). We recommend providing oracle scores alongside zero-shot results: still fine-tune using English data, but choose a checkpoint with the target dev set. Reporting this upper bound makes results more consistent by avoiding arbitrarily bad checkpoints.

End-to-end neural models for goal-oriented conversational systems have become an increasingly active area of research, though results in real-world settings are few. We present real-world results for two issue types in the customer service domain. We train models on historical chat transcripts and test on live contacts using a human-in-the-loop research platform. Additionally, we incorporate customer profile features to assess their impact on model performance. We experiment with two approaches for response generation: (1) sequence-to-sequence generation and (2) template ranking. To test our models, a customer service agent handles live contacts and at each turn we present the top four model responses and allow the agent to select (and optionally edit) one of the suggestions or to type their own. We present results for turn acceptance rate, response coverage, and edit rate based on approximately 600 contacts, as well as qualitative analysis on patterns of turn rejection and edit behavior. Top-4 turn acceptance rate across all models ranges from 63%-80%. Our results suggest that these models are promising for an agent-support application.

Contextual word embeddings (e.g. GPT, BERT, ELMo, etc.) have demonstrated state-of-the-art performance on various NLP tasks. Recent work with the multilingual version of BERT has shown that the model performs surprisingly well in cross-lingual settings, even when only labeled English data is used to finetune the model. We improve upon multilingual BERT’s zero-resource cross-lingual performance via adversarial learning. We report the magnitude of the improvement on the multilingual MLDoc text classification and CoNLL 2002/2003 named entity recognition tasks. Furthermore, we show that language-adversarial training encourages BERT to align the embeddings of English documents and their translations, which may be the cause of the observed performance gains.

We incorporate an explicit neural interlingua into a multilingual encoder-decoder neural machine translation (NMT) architecture. We demonstrate that our model learns a language-independent representation by performing direct zero-shot translation (without using pivot translation), and by using the source sentence embeddings to create an English Yelp review classifier that, through the mediation of the neural interlingua, can also classify French and German reviews. Furthermore, we show that, despite using a smaller number of parameters than a pairwise collection of bilingual NMT models, our approach produces comparable BLEU scores for each language pair in WMT15.

We describe results of a word sense annotation task using WordNet, involving half a dozen well-trained annotators on ten polysemous words for three parts of speech. One hundred sentences for each word were annotated. Annotators had the same level of training and experience, but interannotator agreement (IA) varied across words. There was some effect of part of speech, with higher agreement on nouns and adjectives, but within the words for each part of speech there was wide variation. This variation in IA does not correlate with number of senses in the inventory, or the number of senses actually selected by annotators. In fact, IA was sometimes quite high for words with many senses. We claim that the IA variation is due to the word meanings, contexts of use, and individual differences among annotators. We find some correlation of IA with sense confusability as measured by a sense confusion threshhold (CT). Data mining for association rules on a flattened data representation indicating each annotator's sense choices identifies outliers for some words, and systematic differences among pairs of annotators on others.