What to talk about and how? Selective Generation using LSTMs with Coarse-to-Fine Alignment

TL;DR

This paper introduces a neural encoder-aligner-decoder model for selective generation that effectively chooses relevant data and generates descriptive text, achieving state-of-the-art results on weather and RoboCup datasets.

Contribution

The paper presents a novel end-to-end neural model with a coarse-to-fine aligner for content selection and surface realization, outperforming previous methods without relying on specialized features.

Findings

59% relative improvement in generation on WeatherGov dataset

Effective generalization to RoboCup dataset

Model achieves state-of-the-art results without linguistic resources

Abstract

We propose an end-to-end, domain-independent neural encoder-aligner-decoder model for selective generation, i.e., the joint task of content selection and surface realization. Our model first encodes a full set of over-determined database event records via an LSTM-based recurrent neural network, then utilizes a novel coarse-to-fine aligner to identify the small subset of salient records to talk about, and finally employs a decoder to generate free-form descriptions of the aligned, selected records. Our model achieves the best selection and generation results reported to-date (with 59% relative improvement in generation) on the benchmark WeatherGov dataset, despite using no specialized features or linguistic resources. Using an improved k-nearest neighbor beam filter helps further. We also perform a series of ablations and visualizations to elucidate the contributions of our key model components. Lastly, we evaluate the generalizability of our model on the RoboCup dataset, and get results that are competitive with or better than the state-of-the-art, despite being severely data-starved.