Low- and high-resource opinion summarization

Abstract

Customer reviews play a vital role in the online purchasing decisions we make. The reviews express user opinions that are useful for setting realistic expectations and uncovering important details about products. However, some products receive hundreds or even thousands of reviews, making them time-consuming to read. Moreover, many reviews contain uninformative content, such as irrelevant personal experiences. Automatic summarization offers an alternative – short text summaries capturing the essential information expressed in reviews. Automatically produced summaries can reflect overall or particular opinions and be tailored to user preferences. Besides being presented on major e-commerce platforms, home assistants can also vocalize them. This approach can improve user satisfaction by assisting in making faster and better decisions.

Modern summarization approaches are based on neural networks, often requiring thousands of annotated samples for training. However, human-written summaries for products are expensive to produce because annotators need to read many reviews. This has led to annotated data scarcity where only a few datasets are available. Data scarcity is the central theme of our works, and we propose a number of approaches to alleviate the problem. The thesis consists of two parts where we discuss low- and high-resource data settings.

In the first part, we propose self-supervised learning methods applied to customer reviews and few-shot methods for learning from small annotated datasets. Customer reviews without summaries are available in large quantities, contain a breadth of in-domain specifics, and provide a powerful training signal. We show that reviews can be used for learning summarizers via a self-supervised objective. Further, we address two main challenges associated with learning from small annotated datasets. First, large models rapidly overfit on small datasets leading to poor generalization. Second, it is not possible to learn a wide range of in-domain specifics (e.g., product aspects and usage) from a handful of gold samples. This leads to subtle semantic mistakes in generated summaries, such as ‘great dead on arrival battery.’ We address the first challenge by explicitly modeling summary properties (e.g., content coverage and sentiment alignment). Furthermore, we leverage small modules – adapters – that are more robust to overfitting. As we show, despite their size, these modules can be used to store in-domain knowledge to reduce semantic mistakes. Lastly, we propose a simple method for learning personalized summarizers based on aspects, such as ‘price,’ ‘battery life,’ and ‘resolution.’ This task is harder to learn, and we present a few-shot method for training a query-based summarizer on small annotated datasets.

In the second part, we focus on the high-resource setting and present a large dataset with summaries collected from various online resources. The dataset has more than 33,000 humanwritten summaries, where each is linked up to thousands of reviews. This, however, makes it challenging to apply an ‘expensive’ deep encoder due to memory and computational costs. To address this problem, we propose selecting small subsets of informative reviews. Only these subsets are encoded by the deep encoder and subsequently summarized. We show that the selector and summarizer can be trained end-to-end via amortized inference and policy gradient methods.