The components underpinning PLMs—large weight matrices—were shown to bear considerable redundancy. Matrix factorization, a well-established technique from matrix theory, has been utilized to reduce the number of parameters in PLM. However, it fails to retain satisfactory performance under moderate to high compression rates. In this paper, we identify the full-rankness of fine-tuned PLM as the fundamental bottleneck for the failure of matrix factorization and explore the use of network pruning to extract low-rank sparsity pattern desirable to matrix factorization. We find such a low-rank sparsity pattern exclusively exists in models generated by first-order pruning, which motivates us to unite the two approaches and achieve more effective model compression. We further propose two techniques: sparsity-aware SVD and mixed-rank fine-tuning, which improve the initialization and training of the compression procedure, respectively. Experiments on GLUE and question-answering tasks show that the proposed method has a superior compression-performance trade-off compared to existing approaches.
Although Large Language Models (LLMs) demonstrate remarkable ability in processing and generating human-like text, they do have limitations when it comes to comprehending and expressing world knowledge that extends beyond the boundaries of natural language(e.g., chemical molecular formula). Injecting a collection of symbolic data directly into the training of LLMs can be problematic, as it disregards the synergies among different symbolic families and overlooks the need for a balanced mixture of natural and symbolic data. In this work, we tackle these challenges from both a data and framework perspective and introduce Symbol-LLM series models. First, we curated a data collection consisting of 34 tasks and incorporating 20 distinct symbolic families, intending to capture the interrelations and foster synergies between symbols. Then, a two-stage tuning framework succeeds in injecting symbolic knowledge without loss of the generality ability. Extensive experiments on both symbol- and NL-centric tasks demonstrate the balanced and superior performances of Symbol-LLM series models.
Iterative pruning is one of the most effective compression methods for pre-trained language models. We discovered that finding the optimal pruning decision is an equality-constrained 0-1 Integer Linear Programming problem. The solution to this optimization problem leads to a principled importance criterion which we use to rank parameters during iterative model pruning. To mitigate the poor generalization at high sparsity levels, we propose a self-regularization scheme where model prediction is regularized by the latest checkpoint with increasing sparsity throughout pruning. Our experiments on natural language understanding, question answering, named entity recognition, and data-to-text generation with various Transformer-based PLMs show the effectiveness of the approach at various sparsity levels.
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.
Pretrained masked language models (PLMs) were shown to be inheriting a considerable amount of relational knowledge from the source corpora. In this paper, we present an in-depth and comprehensive study concerning specializing PLMs into relational models from the perspective of network pruning. We show that it is possible to find subnetworks capable of representing grounded commonsense relations at non-trivial sparsity while being more generalizable than original PLMs in scenarios requiring knowledge of single or multiple commonsense relations.
Current text-image approaches (e.g., CLIP) typically adopt dual-encoder architecture using pre-trained vision-language representation. However, these models still pose non-trivial memory requirements and substantial incremental indexing time, which makes them less practical on mobile devices. In this paper, we present an effective two-stage framework to compress large pre-trained dual-encoder for lightweight text-image retrieval. The resulting model is smaller (39% of the original), faster (1.6x/2.9x for processing image/text respectively), yet performs on par with or better than the original full model on Flickr30K and MSCOCO benchmarks. We also open-source an accompanying realistic mobile image search application.
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.