Entity linking in dialogue is the task of mapping entity mentions in utterances to a target knowledge base. Prior work on entity linking has mainly focused on well-written articles such as Wikipedia, annotated newswire, or domain-specific datasets. We extend the study of entity linking to open domain dialogue by presenting the OpenEL corpus: an annotated multi-domain corpus for linking entities in natural conversation to Wikidata. Each dialogic utterance in 179 dialogues over 12 topics from the EDINA dataset has been annotated for entities realized by definite referring expressions as well as anaphoric forms such as he, she, it and they. This dataset supports training and evaluation of entity linking in open-domain dialogue, as well as analysis of the effect of using dialogue context and anaphora resolution in model training. It could also be used for fine-tuning a coreference resolution algorithm. To the best of our knowledge, this is the first substantial entity linking corpus publicly available for open-domain dialogue. We also establish baselines for this task using several existing entity linking systems. We found that the Transformer-based system Flair + BLINK has the best performance with a 0.65 F1 score. Our results show that dialogue context is extremely beneficial for entity linking in conversations, with Flair + Blink achieving an F1 of 0.61 without discourse context. These results also demonstrate the remaining performance gap between the baselines and human performance, highlighting the challenges of entity linking in open-domain dialogue, and suggesting many avenues for future research using OpenEL.

We demonstrate that knowledge distillation can be used not only to reduce model size, but to simultaneously adapt a contextual language model to a specific domain. We use Multilingual BERT (mBERT; Devlin et al., 2019) as a starting point and follow the knowledge distillation approach of (Sahn et al., 2019) to train a smaller multilingual BERT model that is adapted to the domain at hand. We show that for in-domain tasks, the domain-specific model shows on average 2.3% improvement in F1 score, relative to a model distilled on domain-general data. Whereas much previous work with BERT has fine-tuned the encoder weights during task training, we show that the model improvements from distillation on in-domain data persist even when the encoder weights are frozen during task training, allowing a single encoder to support classifiers for multiple tasks and languages.

We describe recent work on MedSLT, a medium-vocabulary interlingua-based medical speech translation system, focussing on issues that arise when handling languages of which the grammar engineer has little or no knowledge. We show how we can systematically create and maintain multiple forms of grammars, lexica and interlingual representations, with some versions being used by language informants, and some by grammar engineers. In particular, we describe the advantages of structuring the interlingua definition as a simple semantic grammar, which includes a human-readable surface form. We show how this allows us to rationalise the process of evaluating translations between languages lacking common speakers, and also makes it possible to create a simple generic tool for debugging to-interlingua translation rules. Examples presented focus on the concrete case of translation between Japanese and Arabic in both directions.

We present an overview of Regulus, an Open Source platform that supports corpus-based derivation of efficient domain-specific speech recognisers from general linguistically motivated unification grammars. We list available Open Source resources, which include compilers, resource grammars for various languages, documentation and a development environment. The greater part of the paper presents a series of experiments carried out using a medium-vocabulary medical speech translation application and a corpus of 801 recorded domain utterances, designed to investigate the impact on speech understanding performance of vocabulary size, grammatical coverage, presence or absence of various linguistic features, degree of generality of thegrammar and use or otherwise of probabilistic weighting in the CFGlanguage model. In terms of task accuracy, the most significant factors were the use of probabilistic weighting, the degree of generality of the grammar and the inclusion of features which model sortal restrictions.

In this paper, we present evidence that providing users of a speech to speech translation system for emergency diagnosis (MedSLT) with a tool that helps them to learn the coverage greatly improves their success in using the system. In MedSLT, the system uses a grammar-based recogniser that provides more predictable results to the translation component. The help module aims at addressing the lack of robustness inherent in this type of approach. It takes as input the result of a robust statistical recogniser that performs better for out-of-coverage data and produces a list of in-coverage example sentences. These examples are selected from a defined list using a heuristic that prioritises sentences maximising the number of N-grams shared with those extracted from the recognition result.