Meaning Representations from Trajectories in Autoregressive Models
Tian Yu Liu, Matthew Trager, Alessandro Achille, Pramuditha Perera,, Luca Zancato, Stefano Soatto
TL;DR
This paper introduces a novel method to derive meaning representations from autoregressive language models by analyzing the distribution of all possible output trajectories, enabling better semantic understanding and relation modeling without fine-tuning.
Contribution
It presents a prompt-free, distribution-based approach to extract semantic representations from pre-trained autoregressive models, capable of modeling asymmetric relations and multimodal data.
Findings
Distribution-based representations align with human annotations.
Outperform zero-shot methods on semantic similarity tasks.
Enable complex entailment and containment reasoning.
Abstract
We propose to extract meaning representations from autoregressive language models by considering the distribution of all possible trajectories extending an input text. This strategy is prompt-free, does not require fine-tuning, and is applicable to any pre-trained autoregressive model. Moreover, unlike vector-based representations, distribution-based representations can also model asymmetric relations (e.g., direction of logical entailment, hypernym/hyponym relations) by using algebraic operations between likelihood functions. These ideas are grounded in distributional perspectives on semantics and are connected to standard constructions in automata theory, but to our knowledge they have not been applied to modern language models. We empirically show that the representations obtained from large models align well with human annotations, outperform other zero-shot and prompt-free methods…
Peer Reviews
Decision·ICLR 2024 poster
1. The paper's trajectory-based method for understanding LLMs is original, diverging from typical vector space models and prompt-based approaches, and introducing a new angle to distributional semantics. 2. The approach has been empirically tested against established benchmarks, indicating robust methodology and results that surpass existing techniques like CLIP embeddings on image-image similarity tasks. 3. The paper is clearly articulated, systematically presenting the new method and its imp
My main criticism of the paper is in the results for semantic similarity, they are far below those of contrastive methods (like 10 pts or so). I also the Sentence-T5 results are a bit misleading, that is the case without any fine-tuning. This isn't explicitly stated. There are other approaches that achieve far higher results on these tasks that do not use the training data for these tasks. I think the Sentence-T5 results in this paper are actually worse than a random encoder where random word em
- The paper is overall well-written. I had no difficulty in reading and understanding this paper. - This paper presents an interesting and unique usage of decoder-only language models for measuring sentence similarity. - The proposed method shows better performance on sentence similarity tasks over encoder-based baselines.
- For the baseline encoder-only models, it is better to include larger models like BERT-large and RoBERTa-large using their CLS tokens and token averages. - The discussion about partial ordering between sentences is a bit puzzling. Since Tu and Tv are samples from u and v respectively, the former set of trajectories usually gives high Mu values for u and vise versa, so it hardly happens that Mu < Mv or Mu > Mv for all t in Tu U Tv. Besides, the discussion of entailment suddenly shifts from the c
- The work seems to outperform other methods for autoregressive representation of sentences on the sentence similarity tasks studied, indicating its potential for continued relevance and utility. - The ability to modify similarity using prompting is very clever. - On a theoretical level, thinking about sentence meaning in terms of this theoretical notion of a trajectory of meanings is a great framing.
- It has a relatively rigid and narrow use case, since this method can only be used for pairwise comparison and since it's not obvious how to fine-tune it. - The work frames it as producing "interpretable" vectors, but the work was somewhat lacking in an actual exploration of that interpretability. - I liked the idea of the entailment and hypernymy work, but it felt a bit convoluted: the way they approached both tasks seems to have lead to them comparing to weak baselines, despite NLI and wor
Code & Models
Videos
Taxonomy
TopicsNatural Language Processing Techniques · Topic Modeling · Text Readability and Simplification
MethodsALIGN