Generating the English transliteration of a name written in a foreign script is an important and challenging step in multilingual knowledge acquisition and information extraction. Existing approaches to transliteration generation require a large (>5000) number of training examples. This difficulty contrasts with transliteration discovery, a somewhat easier task that involves picking a plausible transliteration from a given list. In this work, we present a bootstrapping algorithm that uses constrained discovery to improve generation, and can be used with as few as 500 training examples, which we show can be sourced from annotators in a matter of hours. This opens the task to languages for which large number of training examples are unavailable. We evaluate transliteration generation performance itself, as well the improvement it brings to crosslingual candidate generation for entity linking, a typical downstream task. We present a comprehensive evaluation of our approach on nine languages, each written in a unique script. 1

With the advent of powerful neural language models over the last few years, research attention has increasingly focused on what aspects of language they represent that make them so successful. Several testing methodologies have been developed to probe models' syntactic representations. One popular method for determining a model's ability to induce syntactic structure trains a model on strings generated according to a template then tests the model's ability to distinguish such strings from superficially similar ones with different syntax. We illustrate a fundamental problem with this approach by reproducing positive results from a recent paper with two non-syntactic baseline language models: an n-gram model and an LSTM model trained on scrambled inputs.

In ‘Against Stored Abstractions,’ Ambridge uses neural and computational evidence to make his case against abstract representations. He argues that storing only exemplars is more parsimonious – why bother with abstraction when exemplar models with on-the-fly calculation can do everything abstracting models can and more – and implies that his view is well supported by neuroscience and computer science. We argue that there is substantial neural, experimental, and computational evidence to the contrary: while both brains and machines can store exemplars, forming categories and storing abstractions is a fundamental part of what they do.

Child language acquisition is often identified as one of the primary drivers of language change, but the lack of historical child data presents a challenge for empirically investigating its effect. In this work, I observe the relationship between lexicons extracted from modern child-directed speech and those drawn from modern and historical literary corpora in order to better understand when language acquisition can be modeled over historical and non-child corpora as it is over child corpora. The type frequencies of morphophonological and syntactic-semantic patterns occur at similar type frequencies in these corpora among high token frequency items, and furthermore, when a learning algorithm is applied to lexicons sampled from these sources, it consistently achieves the same learning outcomes in each. With appropriate care and pre-processing, modern and historical text corpora are effectively interchangeable with child-directed speech corpora for the purpose of estimating child lexical experience, opening a path for modeling language acquisition where child-directed corpora are not available.

This year's iteration of the SIGMORPHON-UniMorph shared task on "human-like" morphological inflection generation focuses on generalization and errors in language acquisition. Systems are trained on data sets extracted from corpora of child-directed speech in order to simulate a natural learning setting, and their predictions are evaluated against what is known about children's developmental trajectories for three well-studied patterns: English past tense, German noun plurals, and Arabic noun plurals. Three submitted neural systems were evaluated together with two baselines. Performance was generally good, and all systems were prone to human-like over-regularization. However, all systems were also prone to non-human-like over-irregularization and nonsense productions to varying degrees. We situate this behavior in a discussion of the Past Tense Debate. 1