We consider the problem of estimating the causal effects of linguistic properties on downstream outcomes. For example, does writing a complaint politely lead to a faster response time? How much will a positive product review increase sales? This paper focuses on two challenges related to the problem. First, we formalize the causal quantity of interest as the effect of a writer's intent, and establish the assumptions necessary to identify this from observational data. Second, in practice we only have access to noisy proxies for these linguistic properties---e.g., predictions from classifiers and lexicons. We propose an estimator for this setting and prove that its bias is bounded when we perform an adjustment for the text. The method leverages (1) a pre-trained language model (BERT) to adjust for the text, and (2) distant supervision to improve the quality of noisy proxies. We show that our algorithm produces better causal estimates than related methods on two datasets: predicting the effect of music review sentiment on sales, and complaint politeness on response time.
Neural image-to-text radiology report generation systems offer the potential to accelerate clinical processes by saving radiologists from the repetitive labor of drafting radiology reports and preventing medical errors. However, existing report generation systems, despite achieving high performances on natural language generation metrics such as CIDEr or BLEU, still suffer from incomplete and inconsistent generations, rendering these systems unusable in practice. In this work, we aim to overcome this problem by proposing two new metrics that encourage the factual completeness and consistency of generated radiology reports. The first metric, the Exact Entity Match score, evaluates a generation by its coverage of radiology domain entities against the references. The second metric, the Entailing Entity Match score, augments the first metric by introducing a natural language inference model into the entity match process to encourage consistent generations that can be entailed from the references. To achieve this, we also developed an in-domain NLI model via weak supervision to improve its performance on radiology text. We further propose a report generation system that optimizes these two new metrics via reinforcement learning. On two open radiology report datasets, our system not only achieves the best performance on these two metrics compared to baselines, but also leads to as much as +2.0 improvement on the F1 score of a clinical finding metric. We show via analysis and examples that our system leads to generations that are more complete and consistent compared to the baselines.
Benchmarks such as GLUE have helped drive advances in NLP by incentivizing the creation of more accurate models. While this leaderboard paradigm has been remarkably successful, a historical focus on performance-based evaluation has been at the expense of other qualities that the NLP community values in models, such as compactness, fairness, and energy efficiency. In this opinion paper, we study the divergence between what is incentivized by leaderboards and what is useful in practice through the lens of microeconomic theory. We frame both the leaderboard and NLP practitioners as consumers and the benefit they get from a model as its utility to them. With this framing, we formalize how leaderboards -- in their current form -- can be poor proxies for the NLP community at large. For example, a highly inefficient model would provide less utility to practitioners but not to a leaderboard, since it is a cost that only the former must bear. To allow practitioners to better estimate a model's utility to them, we advocate for more transparency on leaderboards, such as the reporting of statistics that are of practical concern (e.g., model size, energy efficiency, and inference latency).
Despite its importance to experimental design, statistical power (the probability that, given a real effect, an experiment will reject the null hypothesis) has largely been ignored by the NLP community. Underpowered experiments make it more difficult to discern the difference between statistical noise and meaningful model improvements, and increase the chances of exaggerated findings. By meta-analyzing a set of existing NLP papers and datasets, we characterize typical power for a variety of settings and conclude that underpowered experiments are common in the NLP literature. In particular, for several tasks in the popular GLUE benchmark, small test sets mean that most attempted comparisons to state of the art models will not be adequately powered. Similarly, based on reasonable assumptions, we find that the most typical experimental design for human rating studies will be underpowered to detect small model differences, of the sort that are frequently studied. For machine translation, we find that typical test sets of 2000 sentences have approximately 75% power to detect differences of 1 BLEU point. To improve the situation going forward, we give an overview of best practices for power analysis in NLP and release a series of notebooks to assist with future power analyses.
We introduce $k$-nearest-neighbor machine translation ($k$NN-MT), which predicts tokens with a nearest neighbor classifier over a large datastore of cached examples, using representations from a neural translation model for similarity search. This approach requires no additional training and scales to give the decoder direct access to billions of examples at test time, resulting in a highly expressive model that consistently improves performance across many settings. Simply adding nearest neighbor search improves a state-of-the-art German-English translation model by 1.5 BLEU. $k$NN-MT allows a single model to be adapted to diverse domains by using a domain-specific datastore, improving results by an average of 9.2 BLEU over zero-shot transfer, and achieving new state-of-the-art results---without training on these domains. A massively multilingual model can also be specialized for particular language pairs, with improvements of 3 BLEU for translating from English into German and Chinese. Qualitatively, $k$NN-MT is easily interpretable; it combines source and target context to retrieve highly relevant examples.
Hungarian is often referred to as a discourse-configurational language, since the structural position of constituents is determined by their logical function (topic or comment) rather than their grammatical function (e.g., subject or object). We build on work by Koml\'osy (1989) and argue that in addition to discourse context, the lexical semantics of the verb also plays a significant role in determining Hungarian word order. In order to investigate the role of lexical semantics in determining Hungarian word order, we conduct a large-scale, data-driven analysis on the ordering of 380 transitive verbs and their objects, as observed in hundreds of thousands of examples extracted from the Hungarian Gigaword Corpus. We test the effect of lexical semantics on the ordering of verbs and their objects by grouping verbs into 11 semantic classes. In addition to the semantic class of the verb, we also include two control features related to information structure, object definiteness and object NP weight, chosen to allow a comparison of their effect size to that of verb semantics. Our results suggest that all three features have a significant effect on verb-object ordering in Hungarian and among these features, the semantic class of the verb has the largest effect. Specifically, we find that stative verbs, such as fed "cover", jelent "mean" and \"ovez "surround", tend to be OV-preferring (with the exception of psych verbs which are strongly VO-preferring) and non-stative verbs, such as b\'ir\'al "judge", cs\"okkent "reduce" and cs\'okol "kiss", verbs tend to be VO-preferring. These findings support our hypothesis that lexical semantic factors influence word order in Hungarian.
Recent models for unsupervised representation learning of text have employed a number of techniques to improve contextual word representations but have put little focus on discourse-level representations. We propose CONPONO, an inter-sentence objective for pretraining language models that models discourse coherence and the distance between sentences. Given an anchor sentence, our model is trained to predict the text k sentences away using a sampled-softmax objective where the candidates consist of neighboring sentences and sentences randomly sampled from the corpus. On the discourse representation benchmark DiscoEval, our model improves over the previous state-of-the-art by up to 13% and on average 4% absolute across 7 tasks. Our model is the same size as BERT-Base, but outperforms the much larger BERT- Large model and other more recent approaches that incorporate discourse. We also show that CONPONO yields gains of 2%-6% absolute even for tasks that do not explicitly evaluate discourse: textual entailment (RTE), common sense reasoning (COPA) and reading comprehension (ReCoRD).
We propose a novel methodology for analyzing the encoding of grammatical structure in neural language models through transfer learning. We test how a language model can leverage its internal representations to transfer knowledge across languages and symbol systems. We train LSTMs on non-linguistic, structured data and test their performance on human language to assess which kinds of data induce generalizable encodings that LSTMs can use for natural language. We find that models trained on structured data such as music and Java code have internal representations that help in modelling human language, and that, surprisingly, adding minimal amounts of structure to the training data makes a large difference in transfer to natural language. Further experiments on transfer between human languages show that zero-shot performance on a test language is highly correlated with syntactic similarity to the training language, even after removing any vocabulary overlap. This suggests that the internal representations induced from natural languages are typologically coherent: they encode the features and differences outlined in typological studies. Our results provide insights into how neural networks represent linguistic structure, and also about the kinds of structural biases that give learners the ability to model language.