Neural natural language generation (NNLG) systems are known for their pathological outputs, i.e. generating text which is unrelated to the input specification. In this paper, we show the impact of semantic noise on state-of-theart NNLG models which implement different semantic control mechanisms. We find that cleaned data can improve semantic correctness by up to 97%, while maintaining fluency. We also find that the most common error is omitting information, rather than hallucination.

Mixed initiative in open-domain dialogue requires a system to pro-actively introduce new topics. The one-turn topic transition task explores how a system connects two topics in a cooperative and coherent manner. The goal of the task is to generate a "bridging" utterance connecting the new topic to the topic of the previous conversation turn. We are especially interested in commonsense explanations of how a new topic relates to what has been mentioned before. We first collect a new dataset of human one-turn topic transitions, which we call OTTers 1 . We then explore different strategies used by humans when asked to complete such a task, and notice that the use of a bridging utterance to connect the two topics is the approach used the most. We finally show how existing state-of-the-art text generation models can be adapted to this task and examine the performance of these baselines on different splits of the OTTers data.

A controversial issue in psycholinguistics is the degree to which speakers employ audience design during language production. Hypothesising that a consideration of the listener's needs is particularly relevant when the listener is under cognitive load, we had speakers describe objects for a listener performing an easy or a difficult simulated driving task. We predicted that speakers would introduce more redundancy in their descriptions in the difficult driving task, thereby accommodating the listener's reduced cognitive capacity. The results showed that speakers did not adapt their descriptions to a change in the listener's cognitive load. However, speakers who had experienced the driving task themselves before and who were presented with the difficult driving task first were more redundant than other speakers. These findings may suggest that speakers only consider the listener's needs in the presence of strong enough cues, and do not update their beliefs about these needs during the task.

Natural language generation (NLG) systems rely on corpora for both hand-crafted approaches in a traditional NLG architecture and for statistical end-to-end (learned) generation systems. Limitations in existing resources, however, make it difficult to develop systems which can vary the linguistic properties of an utterance as needed. For example, when users' attention is split between a linguistic and a secondary task such as driving, a generation system may need to reduce the information density of an utterance to compensate for the reduction in user attention. We introduce a new corpus in the restaurant recommendation and comparison domain, collected in a paraphrasing paradigm, where subjects wrote texts targeting either a general audience or an elderly family member. This design resulted in a corpus of more than 5000 texts which exhibit a variety of lexical and syntactic choices and differ with respect to average word & sentence length and surprisal. The corpus includes two levels of meaning representation: flat 'semantic stacks' for propositional content and Rhetorical Structure Theory (RST) relations between these propositions.

While previous research on readability has typically focused on document-level measures, recent work in areas such as natural language generation has pointed out the need of sentence-level readability measures. Much of psycholinguistics has focused for many years on processing measures that provide difficulty estimates on a word-by-word basis. However, these psycholinguistic measures have not yet been tested on sentence readability ranking tasks. In this paper, we use four psycholinguistic measures: idea density, surprisal, integration cost, and embedding depth to test whether these features are predictive of readability levels. We find that psycholinguistic features significantly improve performance by up to 3 percentage points over a standard document-level readability metric baseline.