Training and serving long-context large language models (LLMs) incurs substantial overhead. To address this, two critical steps are often required: a pretrained LLM typically undergoes a separate stage for context length extension by training on long-context data, followed by architectural modifications to reduce the overhead of KV cache during serving. This paper argues that integrating length extension with a GPU-friendly KV cache reduction architecture not only reduces training overhead during length extension, but also achieves better long-context performance. This leads to our proposed LongGen, which finetunes a pretrained LLM into an efficient architecture during length extension. LongGen builds on three key insights: (1) Sparse attention patterns, such as window attention (attending to recent tokens), attention sink (initial ones), and blockwise sparse attention (strided token blocks) are well-suited for building efficient long-context models, primarily due to their GPU-friendly memory access patterns, enabling efficiency gains not just theoretically but in practice as well. (2) It is essential for the model to have direct access to all tokens. A hybrid architecture with 1/3 full attention layers and 2/3 efficient ones achieves a balanced trade-off between efficiency and long-context performance. (3) Lightweight training on 5B long-context data is sufficient to extend the hybrid model's context length from 4K to 128K. We evaluate LongGen on both Llama-2 7B and Llama-2 70B, demonstrating its effectiveness across different scales. During training with 128K-long contexts, LongGen achieves 1.55x training speedup and reduces wall-clock time by 36%, compared to a full-attention baseline. During inference, LongGen reduces KV cache memory by 62%, achieving 1.67x prefilling speedup and 1.41x decoding speedup. Compared to baselines that apply KV-cache reduction techniques to full-attention long-context LLMs, LongGen achieves substantially stronger performance not only on the Needle-in-a-Haystack retrieval task, but also on more challenging long-context reasoning tasks, including BABILong and RULER.

In this study, we introduce adaptive KV cache compression, a plug-and-play method that reduces the memory footprint of generative inference for Large Language Models (LLMs). Different from the conventional KV cache that retains key and value vectors for all context tokens, we conduct targeted profiling to discern the intrinsic structure of attention modules. Based on the recognized structure, we then construct the KV cache in an adaptive manner: evicting long-range contexts on attention heads emphasizing local contexts, discarding non-special tokens on attention heads centered on special tokens, and only employing the standard KV cache for attention heads that broadly attend to all tokens. Moreover, with the lightweight attention profiling used to guide the construction of the adaptive KV cache, FastGen can be deployed without resource-intensive fine-tuning or re-training. In our experiments across various asks, FastGen demonstrates substantial reduction on GPU memory consumption with negligible generation quality loss. We will release our code and the compatible CUDA kernel for reproducibility.

Suyu Ge, Chunting Zhou, Rui Hou, Madian Khabsa, Yi-Chia Wang, Qifan Wang, Jiawei Han, Yuning Mao. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

In this paper we improve the zero-shot generalization ability of language models via Mixture-Of-Memory Augmentation (MoMA), a mechanism that retrieves augmentation documents from multiple information corpora ("external memories"), with the option to "plug in" unseen memory at inference time.We develop a joint learning mechanism that trains the augmentation component with latent labels derived from the end retrieval task, paired with hard negatives from the memory mixture.We instantiate the model in a zero-shot dense retrieval setting by augmenting strong T5-based retrievers with MoMA.With only T5-base, our model obtains strong zero-shot retrieval accuracy on the eighteen tasks included in the standard BEIR benchmark, outperforming some systems with larger model sizes.As a plug-inplay model, our model can efficiently generalize to any unseen corpus, meanwhile achieving comparable or even better performance than methods relying on target-specific pretraining.Our analysis further illustrates the necessity of augmenting with mixture-of-memory for robust generalization, the benefits of augmentation learning, and how MoMA utilizes the plugin memory at inference time without changing its parameters.

Target-oriented opinion summarization is to profile a target by extracting user opinions from multiple related documents. Instead of simply mining opinion ratings on a target (e.g., a restaurant) or on multiple aspects (e.g., food, service) of a target, it is desirable to go deeper, to mine opinion on fine-grained sub-aspects (e.g., fish). However, it is expensive to obtain high-quality annotations at such fine-grained scale. This leads to our proposal of a new framework, FineSum, which advances the frontier of opinion analysis in three aspects: (1) minimal supervision, where no document-summary pairs are provided, only aspect names and a few aspect/sentiment keywords are available; (2) fine-grained opinion analysis, where sentiment analysis drills down to a specific subject or characteristic within each general aspect; and (3) phrase-based summarization, where short phrases are taken as basic units for summarization, and semantically coherent phrases are gathered to improve the consistency and comprehensiveness of summary. Given a large corpus with no annotation, FineSum first automatically identifies potential spans of opinion phrases, and further reduces the noise in identification results using aspect and sentiment classifiers. It then constructs multiple fine-grained opinion clusters under each aspect and sentiment. Each cluster expresses uniform opinions towards certain sub-aspects (e.g., "fish" in "food" aspect) or characteristics (e.g., "Mexican" in "food" aspect). To accomplish this, we train a spherical word embedding space to explicitly represent different aspects and sentiments. We then distill the knowledge from embedding to a contextualized phrase classifier, and perform clustering using the contextualized opinion-aware phrase embedding. Both automatic evaluations on the benchmark and quantitative human evaluation validate the effectiveness of our approach.

Cyberbullying, identified as intended and repeated online bullying behavior, has become increasingly prevalent in the past few decades. Despite the significant progress made thus far, the focus of most existing work on cyberbullying detection lies in the independent content analysis of different comments within a social media session. We argue that such leading notions of analysis suffer from three key limitations: they overlook the temporal correlations among different comments; they only consider the content within a single comment rather than the topic coherence across comments; they remain generic and exploit limited interactions between social media users. In this work, we observe that user comments in the same session may be inherently related, e.g., discussing similar topics, and their interaction may evolve over time. We also show that modeling such topic coherence and temporal interaction are critical to capture the repetitive characteristics of bullying behavior, thus leading to better predicting performance. To achieve the goal, we first construct a unified temporal graph for each social media session. Drawing on recent advances in graph neural network, we then propose a principled graph-based approach for modeling the temporal dynamics and topic coherence throughout user interactions. We empirically evaluate the effectiveness of our approach with the tasks of session-level bullying detection and comment-level case study. Our code is released to public. 1

OBJECTIVE: We address a first step toward using social media data to supplement current efforts in monitoring population-level medication nonadherence: detecting changes to medication treatment. Medication treatment changes, like changes to dosage or to frequency of intake, that are not overseen by physicians are, by that, nonadherence to medication. Despite the consequences, including worsening health conditions or death, 50% of patients are estimated to not take medications as indicated. Current methods to identify nonadherence have major limitations. Direct observation may be intrusive or expensive, and indirect observation through patient surveys relies heavily on patients' memory and candor. Using social media data in these studies may address these limitations. METHODS: We annotated 9830 tweets mentioning medications and trained a convolutional neural network (CNN) to find mentions of medication treatment changes, regardless of whether the change was recommended by a physician. We used active and transfer learning from 12 972 reviews we annotated from WebMD to address the class imbalance of our Twitter corpus. To validate our CNN and explore future directions, we annotated 1956 positive tweets as to whether they reflect nonadherence and categorized the reasons given. RESULTS: Our CNN achieved 0.50 F1-score on this new corpus. The manual analysis of positive tweets revealed that nonadherence is evident in a subset with 9 categories of reasons for nonadherence. CONCLUSION: We showed that social media users publicly discuss medication treatment changes and may explain their reasons including when it constitutes nonadherence. This approach may be useful to supplement current efforts in adherence monitoring.

With the explosion of online news, personalized news recommendation becomes increasingly important for online news platforms to help their users find interesting information. Existing news recommendation methods achieve personalization by building accurate news representations from news content and user representations from their direct interactions with news (e.g., click), while ignoring the high-order relatedness between users and news. Here we propose a news recommendation method which can enhance the representation learning of users and news by modeling their relatedness in a graph setting. In our method, users and news are both viewed as nodes in a bipartite graph constructed from historical user click behaviors. For news representations, a transformer architecture is first exploited to build news semantic representations. Then we combine it with the information from neighbor news in the graph via a graph attention network. For user representations, we not only represent users from their historically clicked news, but also attentively incorporate the representations of their neighbor users in the graph. Improved performances on a large-scale real-world dataset validate the effectiveness of our proposed method.

Chuhan Wu, Fangzhao Wu, Suyu Ge, Tao Qi, Yongfeng Huang, Xing Xie. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Chuhan Wu, Fangzhao Wu, Tao Qi, Suyu Ge, Yongfeng Huang, Xing Xie. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.