Thailand

In the era of large language models, Mixture-of-Experts (MoE) is a promising architecture for managing computational costs when scaling up model parameters. However, conventional MoE architectures like GShard, which activate the top-$K$ out of $N$ experts, face challenges in ensuring expert specialization, i.e. each expert acquires non-overlapping and focused knowledge. In response, we propose the DeepSeekMoE architecture towards ultimate expert specialization. It involves two principal strategies: (1) finely segmenting the experts into $mN$ ones and activating $mK$ from them, allowing for a more flexible combination of activated experts; (2) isolating $K_s$ experts as shared ones, aiming at capturing common knowledge and mitigating redundancy in routed experts. Starting from a modest scale with 2B parameters, we demonstrate that DeepSeekMoE 2B achieves comparable performance with GShard 2.9B, which has 1.5 $\times$ expert parameters and computation. In addition, DeepSeekMoE 2B nearly approaches the performance of its dense counterpart with the same number of total parameters, which sets the upper bound of MoE models. Subsequently, we scale up DeepSeekMoE to 16B parameters and show that it achieves comparable performance with DeepSeek 7B and LLaMA2 7B, with only about 40% of computations.

ACL 2024

DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models

large language model

mixture-of-experts

poster

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

In-context learning (ICL), which promotes inference with several demonstrations, has become a widespread paradigm to stimulate LLM capabilities for downstream tasks. 
Due to context length constraints, it cannot be further improved in spite of more training data, and general features directly from LLMs in ICL are not adaptive to the specific downstream task. 
In this paper, we propose a feature-adaptive and data-scalable in-context learning framework (FADS-ICL), which can leverage task-adaptive features to promote inference on the downstream task, with the supervision of beyond-context samples.
Specifically, it first extracts general features of beyond-context samples via the LLM with ICL input form one by one, and introduces a task-specific modulator to perform feature refinement and prediction after fitting a specific downstream task. 
We conduct extensive experiments on FADS-ICL under varying data settings (4$\sim$128 shots) and LLM scale (0.8$\sim$70B) settings. Experimental results show that FADS-ICL consistently outperforms previous state-of-the-art methods by a significant margin under all settings, verifying the effectiveness and superiority of FADS-ICL. For example, under the 1.5B and 32 shots setting, FADS-ICL can achieve \textbf{+14.3} average accuracy from feature adaptation over vanilla ICL on 10 datasets, with \textbf{+6.2} average accuracy over the previous state-of-the-art method, and the performance can further improve with increasing training data.

Feature-Adaptive and Data-Scalable In-Context Learning

Syntactic Transformer language models aim to achieve better generalization through simultaneously modeling syntax trees and sentences. While prior work has been focusing on adding constituency-based structures to Transformers, we introduce Dependency Transformer Grammars (DTGs), a new class of Transformer language model with explicit dependency-based inductive bias. DTGs simulate dependency transition systems with constrained attention patterns by modifying attention masks, incorporate the stack information through relative positional encoding, and augment dependency arc representation with a combination of token embeddings and operation embeddings. When trained on a dataset of sentences annotated with dependency trees, DTGs achieve better generalization while maintaining comparable perplexity with Transformer language model baselines. DTGs also outperform recent constituency-based models, showing that dependency can better guide Transformer language models. Our code is released at https://github.com/zhaoyd1/Dep_Transformer_Grammars.

Dependency Transformer Grammars: Integrating Dependency Structures into Transformer Language Models

Text embedding is pivotal in semantic textual similarity (STS) tasks, which are crucial components in Large Language Model (LLM) applications. STS learning largely relies on the cosine function as the optimization objective to reflect semantic similarity. However, the cosine has saturation zones rendering vanishing gradients and hindering learning subtle semantic differences in text embeddings. To address this issue, we propose a novel Angle-optimized Embedding model, AoE. It optimizes angle differences in complex space to explore similarity in saturation zones better. To set up a comprehensive evaluation, we experimented with existing short-text STS, our newly collected long-text STS, and downstream task datasets. Extensive experimental results on STS and MTEB benchmarks show that AoE significantly outperforms popular text embedding models neglecting cosine saturation zones. It highlights that AoE can produce high-quality text embeddings and broadly benefit downstream tasks.

AoE: Angle-optimized Embeddings for Semantic Textual Similarity

Unsupervised bilingual lexicon induction (BLI) task aims to find word translations between languages and has achieved great success in similar language pairs. However, related works mostly rely on a single linear mapping for language alignment and fail on distant or low-resource language pairs, achieving less than half the performance observed in rich-resource language pairs. In this paper, we introduce DM-BLI, a Dynamic Multiple subspaces alignment framework for unsupervised BLI. DM-BLI improves language alignment by utilizing multiple subspace alignments instead of a single mapping. We begin via unsupervised clustering to discover these subspaces in source embedding space. Then we identify and align corresponding subspaces in the target space using a rough global alignment. DM-BLI further employs intra-cluster and inter-cluster contrastive learning to refine precise alignment for each subspace pair. Experiments conducted on standard BLI datasets for 12 language pairs (6 rich-resource and 6 low-resource) demonstrate substantial gains achieved by our framework. We release our code at \href{https://github.com/huling-2/DM-BLI/tree/master}{https://github.com/huling-2/DM-BLI.git.}

DM-BLI: Dynamic Multiple Subspaces Alignment for Unsupervised Bilingual Lexicon Induction

Open domain question answering (ODQA) aims to answer questions with knowledge from an external corpus. Fusion-in-Decoder (FiD) is an effective retrieval-augmented reader model to address this task. Given that FiD independently encodes passages, which overlooks the semantic relationships between passages, some studies use knowledge graphs (KGs) to establish dependencies among passages. However, they only leverage knowledge triples from existing KGs, which suffer from incompleteness and may lack certain information critical for answering given questions. To this end, in order to capture the dependencies between passages while tacking the issue of incompleteness in existing KGs, we propose to enhance the retrieval-augmented reader model with a knowledge graph generation module (REANO). Specifically, REANO consists of a KG generator and an answer predictor. The KG generator aims to generate KGs from the passages and the answer predictor then generates answers based on the passages and the generated KGs. Experimental results on five ODQA datasets indicate that compared with baselines, REANO can improve the exact match score by up to 2.7% on the EntityQuestion dataset, with an average improvement of 1.8% across all the datasets.

REANO: Optimising Retrieval-Augmented Reader Models through Knowledge Graph Generation

Disentangled latent spaces usually have better semantic separability and geometrical properties, which leads to better interpretability and more controllable data generation. While this has been well investigated in Computer Vision, in tasks such as image disentanglement, in the NLP domain, sentence disentanglement is still comparatively under-investigated. Most previous work have concentrated on disentangling task-specific generative factors, such as sentiment, within the context of style transfer. In this work, we focus on a more general form of sentence disentanglement, targeting the localised modification and control of more general sentence semantic features. To achieve this, we contribute to a novel notion of sentence semantic disentanglement and introduce a flow-based invertible neural network (INN) mechanism integrated with a transformer-based language Autoencoder (AE) in order to deliver latent spaces with better separability properties. Experimental results demonstrate that the model can conform the distributed latent space into a better semantically disentangled sentence space, leading to improved language interpretability and controlled generation when compared to the recent state-of-the-art language VAE models.

Learning Disentangled Semantic Spaces of Explanations via Invertible Neural Networks

Natural Language Inference (NLI) remains an important benchmark task for LLMs. NLI datasets are a springboard for transfer learning to other semantic tasks, and NLI models are standard tools for identifying the faithfulness of model-generated text. There are several large scale NLI datasets today, and models have improved greatly by hill-climbing on these collections. Yet their realistic performance on out-of-distribution/domain data is less well-understood. We explore the opportunity for synthetic high-quality datasets to adapt NLI models for zero-shot use in downstream applications across new and unseen text domains. We demonstrate a new approach for generating NLI data in diverse domains and lengths, so far not covered by existing training sets. The resulting examples have meaningful premises, 
the hypotheses are formed in creative ways rather than simple edits to a few premise tokens, and the labels have high accuracy. 
We show that models trained on this data ($685$K synthetic examples) have the best generalization to completely new downstream test settings. On the TRUE benchmark, a T5-small model trained with our data improves around $7\%$ on average compared to training on the best alternative dataset. The improvements are more pronounced for smaller models, while still meaningful on a T5 XXL model. We also demonstrate gains on test sets when in-domain training data is augmented with our domain-general synthetic data.

A synthetic data approach for domain generalization of NLI models

Sentence Representation Learning (SRL) is a crucial task in Natural Language Processing (NLP), where contrastive Self-Supervised Learning (SSL) is currently a mainstream approach. However, the reasons behind its remarkable effectiveness remain unclear. Specifically, many studies have investigated the similarities between contrastive and non-contrastive SSL from a theoretical perspective. Such similarities can be verified in classification tasks, where the two approaches achieve comparable performance. But in ranking tasks (i.e., Semantic Textual Similarity (STS) in SRL), contrastive SSL significantly outperforms non-contrastive SSL. Therefore, two questions arise: First, *what commonalities enable various contrastive losses to achieve superior performance in STS?* Second, *how can we make non-contrastive SSL also effective in STS?* To address these questions, we start from the perspective of gradients and discover that four effective contrastive losses can be integrated into a unified paradigm, which depends on three components: the **Gradient Dissipation**, the **Weight**, and the **Ratio**. Then, we conduct an in-depth analysis of the roles these components play in optimization and experimentally demonstrate their significance for model performance. Finally, by adjusting these components, we enable non-contrastive SSL to achieve outstanding performance in STS.

Towards Better Understanding of Contrastive Sentence Representation Learning: A Unified Paradigm for Gradient

In this paper, we explore the capabilities of LLMs in capturing lexical-semantic knowledge from WordNet on the example of the LLaMA-2-7b model and test it on multiple lexical semantic tasks. As the outcome of our experiments, we present TaxoLLaMA, the "all-in-one" model for taxonomy-related tasks, lightweight due to 4-bit quantization and LoRA. TaxoLLaMA achieves 11 SOTA results, and 4 top-2 results out of 16 tasks on the Taxonomy Enrichment, Hypernym Discovery, Taxonomy Construction, and Lexical Entailment tasks. Moreover, it demonstrates a very strong zero-shot performance on Lexical Entailment and Taxonomy Construction with no fine-tuning. We also explore its hidden multilingual and domain adaptation capabilities with a little tuning or few-shot learning. All datasets, code, and pre-trained models are available online (code: https://github.com/VityaVitalich/TaxoLLaMA)

TaxoLLaMA: WordNet-based Model for Solving Multiple Lexical Semantic Tasks

The sequential process of conceptualization and instantiation is essential to generalizable commonsense reasoning as it allows the application of existing knowledge to unfamiliar scenarios. However, existing works tend to undervalue the step of instantiation and heavily
rely on pre-built concept taxonomies and human annotations to collect both types of knowledge, resulting in a lack of instantiated knowledge to complete reasoning, high cost, and limited scalability. To tackle these challenges, we introduce CANDLE (ConceptuAlization
and INstantiation Distillation from Large Language ModEls), a distillation framework that iteratively performs contextualized conceptualization and instantiation over commonsense knowledge bases by instructing large language models to generate both types of knowledge with critic filtering. By applying CANDLE to ATOMIC (Sap et al., 2019a), we construct a comprehensive knowledge base comprising six million conceptualizations and instantiated commonsense knowledge triples. Both types of knowledge are firmly rooted in the original ATOMIC dataset, and intrinsic evaluations demonstrate their exceptional quality and diversity. Empirical results indicate that distilling CANDLE on student models provides benefits across three downstream tasks. Our data and models are publicly available at https://github.com/HKUST-KnowComp/CANDLE.

Premium content

DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models

Downloads

Next from ACL 2024

Feature-Adaptive and Data-Scalable In-Context Learning

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES