Leveraging representations from pre-trained transformer-based encoders achieves state-of-the-art performance on numerous NLP tasks. Larger encoders can improve accuracy for spoken language understanding (SLU) but are challenging to use given the inference latency constraints of online systems (especially on CPU machines).We evaluate using a larger 170M parameter BERT encoder that shares representations across languages, domains and tasks for SLU compared to using smaller 17M parameter BERT encoders with language-, domain- and task-decoupled finetuning.Running inference with a larger shared encoder on GPU is latency neutral and reduces infrastructure cost compared to running inference for decoupled smaller encoders on CPU machines. The larger shared encoder reduces semantic error rates by 4.62% for test sets representing user requests to voice-controlled devices and 5.79% on the tail of the test sets on average across four languages.

In production SLU systems, new training data becomes available with time so that ML models need to be updated on a regular basis. Specifically, releasing new features adds new classes of data while the old data remains constant. However, retraining the full model each time from scratch is computationally expensive. To address this problem, we propose to consider production releases from the curriculum learning perspective and to adapt the local-to-global learning (LGL) schedule (Cheng et. al, 2019) for a statistical model that starts with fewer output classes and adds more classes with each iteration. We report experiments for the tasks of intent classification and slot filling in the context of a production voice-assistant. First, we apply the original LGL schedule on our data and then adapt LGL to the production setting where the full data is not available at initial training iterations. We demonstrate that our method improves model error rates by 7.3% and saves up to 25% training time for individual iterations.

In productionized machine learning systems, online model performance is known to deteriorate over time when there is a distributional drift between offline training and online application data. As a remedy, models are typically retrained at fixed time intervals, implying high computational and manual costs. This work aims at decreasing such costs in productionized, large-scale Spoken Language Understanding systems. In particular, we develop a need-based re-training strategy guided by an efficient drift detector and discuss the arising challenges including system complexity, overlapping model releases, observation limitation and the absence of annotated resources at runtime. We present empirical results on historical data and confirm the utility of our design decisions via an online A/B experiment.

In this paper, we explore class-incremental learning for intent classification (IC) in a setting with limited old data available. IC is the task of mapping user utterances to their corresponding intents. Even though class-incremental learning without storing the old data yields high potential of reducing human and computational resources in industry NLP model releases, to the best of our knowledge, it hasn’t been studied for NLP classification tasks in the literature before. In this work, we compare several contemporary class-incremental learning methods, i.e., BERT warm start, L2, Elastic Weight Consolidation, RecAdam and Knowledge Distillation within two realistic class-incremental learning scenarios: one where only the previous model is assumed to be available, but no data corresponding to old classes, and one in which limited unlabeled data for old classes is assumed to be available. Our results indicate that among the investigated continual learning methods, Knowledge Distillation worked best for our class-incremental learning tasks, and adding limited unlabeled data helps the model in both adaptability and stability.

Recent progress through advanced neural models pushed the performance of task-oriented dialog systems to almost perfect accuracy on existing benchmark datasets for intent classification and slot labeling. However, in evolving real-world dialog systems, where new functionality is regularly added, a major additional challenge is the lack of annotated training data for such new functionality, as the necessary data collection efforts are laborious and time-consuming. A potential solution to reduce the effort is to augment initial seed data by paraphrasing existing utterances automatically. In this paper, we propose a new, data-efficient approach following this idea. Using an interpretation-to-text model for paraphrase generation, we are able to rely on existing dialog system training data, and, in combination with shuffling-based sampling techniques, we can obtain diverse and novel paraphrases from small amounts of seed data. In experiments on a public dataset and with a real-world dialog system, we observe improvements for both intent classification and slot labeling, demonstrating the usefulness of our approach.

In large-scale commercial dialog systems, users express the same request in a wide variety of alternative ways with a long tail of less frequent alternatives. Handling the full range of this distribution is challenging, in particular when relying on manual annotations. However, the same users also provide useful implicit feedback as they often paraphrase an utterance if the dialog system failed to understand it. We propose MARUPA, a method to leverage this type of feedback by creating annotated training examples from it. MARUPA creates new data in a fully automatic way, without manual intervention or effort from annotators, and specifically for currently failing utterances. By re-training the dialog system on this new data, accuracy and coverage for long-tail utterances can be improved. In experiments, we study the effectiveness of this approach in a commercial dialog system across various domains and three languages.

The most approaches to Knowledge Base Question Answering are based on semantic parsing. In this paper, we address the problem of learning vector representations for complex semantic parses that consist of multiple entities and relations. Previous work largely focused on selecting the correct semantic relations for a question and disregarded the structure of the semantic parse: the connections between entities and the directions of the relations. We propose to use Gated Graph Neural Networks to encode the graph structure of the semantic parse. We show on two data sets that the graph networks outperform all baseline models that do not explicitly model the structure. The error analysis confirms that our approach can successfully process complex semantic parses.

Common-sense argumentative reasoning is a challenging task that requires holistic understanding of the argumentation where external knowledge about the world is hypothesized to play a key role. We explore the idea of using event knowledge about prototypical situations from FrameNet and fact knowledge about concrete entities from Wikidata to solve the task. We find that both resources can contribute to an improvement over the non-enriched approach and point out two persisting challenges: first, integration of many annotations of the same type, and second, fusion of complementary annotations. After our explorations, we question the key role of external world knowledge with respect to the argumentative reasoning task and rather point towards a logic-based analysis of the chain of reasoning.

The Fact Extraction and VERification (FEVER) shared task was launched to support the development of systems able to verify claims by extracting supporting or refuting facts from raw text. The shared task organizers provide a large-scale dataset for the consecutive steps involved in claim verification, in particular, document retrieval, fact extraction, and claim classification. In this paper, we present our claim verification pipeline approach, which, according to the preliminary results, scored third in the shared task, out of 23 competing systems. For the document retrieval, we implemented a new entity linking approach. In order to be able to rank candidate facts and classify a claim on the basis of several selected facts, we introduce two extensions to the Enhanced LSTM (ESIM).

Most approaches to Knowledge Base Question Answering are based on semantic parsing. In this paper, we present a tool that aids in debugging of question answering systems that construct a structured semantic representation for the input question. Previous work has largely focused on building question answering interfaces or evaluation frameworks that unify multiple data sets. The primary objective of our system is to enable interactive debugging of model predictions on individual instances (questions) and to simplify manual error analysis. Our interactive interface helps researchers to understand the shortcomings of a particular model, qualitatively analyze the complete pipeline and compare different models. A set of sit-by sessions was used to validate our interface design.

The first stage of every knowledge base question answering approach is to link entities in the input question. We investigate entity linking in the context of question answering task and present a jointly optimized neural architecture for entity mention detection and entity disambiguation that models the surrounding context on different levels of granularity. We use the Wikidata knowledge base and available question answering datasets to create benchmarks for entity linking on question answering data. Our approach outperforms the previous state-of-the-art system on this data, resulting in an average 8% improvement of the final score. We further demonstrate that our model delivers a strong performance across different entity categories.

We demonstrate that for sentence-level relation extraction it is beneficial to consider other relations in the sentential context while predicting the target relation. Our architecture uses an LSTM-based encoder to jointly learn representations for all relations in a single sentence. We combine the context representations with an attention mechanism to make the final prediction. We use the Wikidata knowledge base to construct a dataset of multiple relations per sentence and to evaluate our approach. Compared to a baseline system, our method results in an average error reduction of 24 on a held-out set of relations. The code and the dataset to replicate the experiments are made available at https://github.com/ukplab/.

The Story Cloze test is a recent effort in providing a common test scenario for text understanding systems. As part of the LSDSem 2017 shared task, we present a system based on a deep learning architecture combined with a rich set of manually-crafted linguistic features. The system outperforms all known baselines for the task, suggesting that the chosen approach is promising. We additionally present two methods for generating further training data based on stories from the ROCStories corpus.