Existing literature does not give much guidance on how to build the best possible multi-domain summarization model from existing components. We present an extensive evaluation of popular pre-trained models on a wide range of datasets to inform the selection of both the model and the training data for robust summarization across several domains. We find that fine-tuned BART performs better than T5 and PEGASUS, both on in-domain and out-of-domain data, regardless of the dataset used for fine-tuning. While BART has the best performance, it does vary considerably across domains. A multi-domain summarizer that works well for all domains can be built by simply fine-tuning on diverse domains. It even performs better than an in-domain summarizer, even when using fewer total training examples. While the success of such a multi-domain summarization model is clear through automatic evaluation, by conducting a human evaluation, we find that there are variations that can not be captured by any of the automatic evaluation metrics and thus not reflected in standard leaderboards. Furthermore, we find that conducting reliable human evaluation can be complex as well. Even experienced summarization researchers can be inconsistent with one another in their assessment of the quality of a summary, and also with themselves when re-annotating the same summary. The findings of our study are two-fold. First, BART fine-tuned on heterogeneous domains is a great multi-domain summarizer for practical purposes. At the same time, we need to re-examine not just automatic evaluation metrics but also human evaluation methods to responsibly measure progress in summarization.

How to usefully encode compositional task structure has long been a core challenge in AI. Recent work in chain of thought prompting has shown that for very large neural language models (LMs), explicitly demonstrating the inferential steps involved in a target task may improve performance over end-to-end learning that focuses on the target task alone. However, chain of thought prompting has significant limitations due to its dependency on huge pretrained LMs. In this work, we present compositional fine-tuning (CFT): an approach based on explicitly decomposing a target task into component tasks, and then fine-tuning smaller LMs on a curriculum of such component tasks. We apply CFT to recommendation tasks in two domains, world travel and local dining, as well as a previously studied inferential task (sports understanding). We show that CFT outperforms end-to-end learning even with equal amounts of data, and gets consistently better as more component tasks are modeled via fine-tuning. Compared with chain of thought prompting, CFT performs at least as well using LMs only 7.4% of the size, and is moreover applicable to task domains for which data are not available during pretraining.

The capabilities of today’s natural language processing systems are typically evaluated using large datasets of curated questions and answers. While these are critical benchmarks of progress, they also suffer from weakness due to artificial distributions and incomplete knowledge. Artifacts arising from artificial distributions can overstate language model performance, while incomplete knowledge limits fine-grained analysis. In this work, we introduce a complementary benchmarking approach based on SimPlified Language Activity Traces (SPLAT). SPLATs are corpora of language encodings of activity in some closed domain (we study traces from chess and baseball games in this work). SPLAT datasets use naturally-arising distributions, allow the generation of question-answer pairs at scale, and afford complete knowledge in their closed domains. We show that language models of three different architectures can answer questions about world states using only verb-like encodings of activity. Our approach is extensible to new language models and additional question-answering tasks.

Neural Network Language Models (NNLMs) generate probability distributions by applying a softmax function to a distance metric formed by taking the dot product of a prediction vector with all word vectors in a high-dimensional embedding space. The dot-product distance metric forms part of the inductive bias of NNLMs. Although NNLMs optimize well with this inductive bias, we show that this results in a sub-optimal ordering of the embedding space that structurally impoverishes some words at the expense of others when assigning probability. We present numerical, theoretical and empirical analyses which show that words on the interior of the convex hull in the embedding space have their probability bounded by the probabilities of the words on the hull.