Automatic document summarization aims to produce a concise summary covering the input document’s salient information. Within a report document, the salient information can be scattered in the textual and non-textual content. However, existing document summarization datasets and methods usually focus on the text and filter out the non-textual content. Missing tabular data can limit produced summaries’ informativeness, especially when summaries require covering quantitative descriptions of critical metrics in tables. Existing datasets and methods cannot meet the requirements of summarizing long text and multiple tables in each report. To deal with the scarcity of available data, we propose FINDSum, the first large-scale dataset for long text and multi-table summarization. Built on 21,125 annual reports from 3,794 companies, it has two subsets for summarizing each company’s results of operations and liquidity. To summarize the long text and dozens of tables in each report, we present three types of summarization methods. Besides, we propose a set of evaluation metrics to assess the usage of numerical information in produced summaries. Dataset analyses and experimental results indicate the importance of jointly considering input textual and tabular data when summarizing report documents.

Aspect term extraction aims to extract aspect terms from a review sentence that users have expressed opinions on. One of the remaining challenges for aspect term extraction resides in the lack of sufficient annotated data. While self-training is potentially an effective method to address this issue, the pseudo-labels it yields on unlabeled data could induce noise. In this paper, we use two means to alleviate the noise in the pseudo-labels. One is that inspired by the curriculum learning, we refine the conventional self-training to progressive self-training. Specifically, the base model infers pseudo-labels on a progressive subset at each iteration, where samples in the subset become harder and more numerous as the iteration proceeds. The other is that we use a discriminator to filter the noisy pseudo-labels. Experimental results on four SemEval datasets show that our model significantly outperforms the previous baselines and achieves state-of-the-art performance.

Sentiment transfer aims to change the underlying sentiment of input sentences. The two major challenges in existing works lie in (1) effectively disentangling the original sentiment from input sentences; and (2) preserving the semantic content while transferring the sentiment. We find that identifying the sentiment-irrelevant content from input sentences to facilitate generating output sentences could address the above challenges and then propose the Decode with Template model in this paper. We first mask the explicit sentiment words in input sentences and use the rest parts as templates to eliminate the original sentiment. Then, we input the templates and the target sentiments into our bidirectionally guided variational auto-encoder (VAE) model to generate output. In our method, the template preserves most of the semantics in input sentences, and the bidirectionally guided decoding captures both forward and backward contextual information to generate output. Both two parts contribute to better content preservation. We evaluate our method on two review datasets, Amazon and Yelp, with automatic evaluation methods and human rating. The experimental results show that our method significantly outperforms state-of-the-art models, especially in content preservation.

Word embedding learning is the task to map each word into a low-dimensional and continuous vector based on a large corpus. To enhance corpus based word embedding models, researchers utilize domain knowledge to learn more distinguishable representations via joint optimization and post-processing based models. However, joint optimization based models require much training time. Existing post-processing models mostly consider semantic knowledge while learned embedding models show less functional information. Glossary is a comprehensive linguistic resource. And in previous works, the glossary is usually used to enhance the word representations via joint optimization based methods. In this paper, we post-process pre-trained word embedding models with incorporating the glossary and capture more topical and functional information. We propose GGP (Glossary Guided Post-processing word embedding) model which consists of a global post-processing function to fine-tune each word vector, and an auto-encoding model to learn sense representations, furthermore, constrains each post-processed word representation and the composition of its sense representations to be similar. We evaluate our model by comparing it with two state-of-the-art models on six word topical/functional similarity datasets, and the results show that it outperforms competitors by an average of 4.1% across all datasets. And our model outperforms GloVe by more than 7%.

Automated Essay Scoring (AES) is a critical text regression task that automatically assigns scores to essays based on their writing quality. Recently, the performance of sentence prediction tasks has been largely improved by using Pre-trained Language Models via fusing representations from different layers, constructing an auxiliary sentence, using multi-task learning, etc. However, to solve the AES task, previous works utilize shallow neural networks to learn essay representations and constrain calculated scores with regression loss or ranking loss, respectively. Since shallow neural networks trained on limited samples show poor performance to capture deep semantic of texts. And without an accurate scoring function, ranking loss and regression loss measures two different aspects of the calculated scores. To improve AES’s performance, we find a new way to fine-tune pre-trained language models with multiple losses of the same task. In this paper, we propose to utilize a pre-trained language model to learn text representations first. With scores calculated from the representations, mean square error loss and the batch-wise ListNet loss with dynamic weights constrain the scores simultaneously. We utilize Quadratic Weighted Kappa to evaluate our model on the Automated Student Assessment Prize dataset. Our model outperforms not only state-of-the-art neural models near 3 percent but also the latest statistic model. Especially on the two narrative prompts, our model performs much better than all other state-of-the-art models.