Despite tremendous improvements in natural language generation, summarization models still suffer from the unfaithfulness issue. Previous work evaluates faithfulness either using models trained on the other tasks or in-domain synthetic data, or prompting a large model such as ChatGPT. This paper proposes to do zero-shot faithfulness evaluation simply with a moderately-sized foundation language model. We introduce a new metric FFLM, which is a combination of probability changes based on the intuition that prefixing a piece of text that is consistent with the output will increase the probability of predicting the output. Experiments show that FFLM performs competitively with or even outperforms ChatGPT on both inconsistency detection and faithfulness rating with 24x fewer parameters. FFLM also achieves improvements over other strong baselines.

The quadratic complexity of the attention module makes it gradually become the bulk of compute in Transformer-based LLMs during generation. Moreover, the excessive key-value cache that arises when dealing with long inputs also brings severe issues on memory footprint and inference latency. In this work, we propose a plug-and-play approach that is able to incrementally compress the intermediate activation of a specified span of tokens into compact ones, thereby reducing both memory and computational cost when processing subsequent context. Experiments on both in-domain language modeling and zero-shot open-ended document generation demonstrate the advantage of our approach over sparse attention baselines in terms of fluency, n-gram matching, and semantic similarity. At last, we comprehensively profile the benefit of context compression on improving the system throughout. Code is available at https://github.com/DRSY/KV_Compression.

Curriculum learning has shown promising improvements in multiple domains by training machine learning models from easy samples to hard ones. Previous works which either design rules or train models for scoring the difficulty highly rely on task-specific expertise, and cannot generalize. Inspired by the “easy-to-hard” intuition, we propose to do in-sample curriculum learning for natural language generation tasks. Our learning strategy starts training the model to generate the last few words, i.e., do sequence completion, and gradually extends to generate the whole output sequence. Comprehensive experiments show that it generalizes well to different tasks and achieves significant improvements over strong baselines.

Changing speaker names consistently throughout a dialogue should not affect its meaning and corresponding outputs for text generation from dialogues. However, pre-trained language models, serving as the backbone for dialogue-processing tasks, have shown to be sensitive to nuances. This may result in unfairness in real-world applications. No comprehensive analysis of this problem has been done in the past. In this work, we propose to quantitatively measure a model’s sensitivity on speaker names, and comprehensively evaluate a number of known methods for reducing speaker name sensitivity, including a novel approach of our own. Extensive experiments on multiple datasets provide a benchmark for this problem and show the favorable performance of our approach in sensitivity reduction and quality of generation.

Opinion summarization of multiple reviews suffers from the lack of reference summaries for training. Most previous approaches construct multiple reviews and their summary based on textual similarities between reviews,resulting in information mismatch between the review input and the summary. In this paper, we convert each review into a mixof structured and unstructured data, which we call opinion-aspect pairs (OAs) and implicit sentences (ISs).We propose a new method to synthesize training pairs of such mix-structured data as input and the textual summary as output,and design a summarization model with OA encoder and IS encoder. Experiments show that our approach outperforms previous methods on Yelp, Amazon and RottenTomatos datasets.

Previous length-controllable summarization models mostly control lengths at the decoding stage, whereas the encoding or the selection of information from the source document is not sensitive to the designed length. They also tend to generate summaries as long as those in the training data. In this paper, we propose a length-aware attention mechanism (LAAM) to adapt the encoding of the source based on the desired length. Our approach works by training LAAM on a summary length balanced dataset built from the original training data, and then fine-tuning as usual. Results show that this approach is effective in generating high-quality summaries with desired lengths and even those short lengths never seen in the original training set.

Previous dialogue summarization techniques adapt large language models pretrained on the narrative text by injecting dialogue-specific features into the models. These features either require additional knowledge to recognize or make the resulting models harder to tune. To bridge the format gap between dialogues and narrative summaries in dialogue summarization tasks, we propose to post-train pretrained language models (PLMs) to rephrase from dialogue to narratives. After that, the model is fine-tuned for dialogue summarization as usual. Comprehensive experiments show that our approach significantly improves vanilla PLMs on dialogue summarization and outperforms other SOTA models by the summary quality and implementation costs.

A document can be summarized in a number of ways. Reference-based evaluation of summarization has been criticized for its inflexibility. The more sufficient the number of abstracts, the more accurate the evaluation results. However, it is difficult to collect sufficient reference summaries. In this paper, we propose a new automatic reference-free evaluation metric that compares semantic distribution between source document and summary by pretrained language models and considers summary compression ratio. The experiments show that this metric is more consistent with human evaluation in terms of coherence, consistency, relevance and fluency.

Multi-turn response selection is a task designed for developing dialogue agents. The performance on this task has a remarkable improvement with pre-trained language models. However, these models simply concatenate the turns in dialogue history as the input and largely ignore the dependencies between the turns. In this paper, we propose a dialogue extraction algorithm to transform a dialogue history into threads based on their dependency relations. Each thread can be regarded as a self-contained sub-dialogue. We also propose Thread-Encoder model to encode threads and candidates into compact representations by pre-trained Transformers and finally get the matching score through an attention layer. The experiments show that dependency relations are helpful for dialogue context understanding, and our model outperforms the state-of-the-art baselines on both DSTC7 and DSTC8*, with competitive results on UbuntuV2.