Text Alignment Is An Efficient Unified Model for Massive NLP Tasks

Large language models (LLMs), typically designed as a function of next-word prediction, have excelled across extensive NLP tasks. Despite the generality, next-word prediction is often not an efficient formulation for many of the tasks, demanding an extreme scale of model parameters (10s or 100s of billions) and sometimes yielding suboptimal performance. In practice, it is often desirable to build more efficient models -- despite being less versatile, they still apply to a substantial subset of problems, delivering on par or even superior performance with much smaller model sizes. In this paper, we propose text alignment as an efficient unified model for a wide range of crucial tasks involving text entailment, similarity, question answering (and answerability), factual consistency, and so forth. Given a pair of texts, the model measures the degree of alignment between their information. We instantiate an alignment model (Align) through lightweight finetuning of RoBERTa (355M parameters) using 5.9M examples from 28 datasets. Despite its compact size, extensive experiments show the model's efficiency and strong performance: (1) On over 20 datasets of aforementioned diverse tasks, the model matches or surpasses FLAN-T5 models that have around 2x or 10x more parameters; the single unified model also outperforms task-specific models finetuned on individual datasets; (2) When applied to evaluate factual consistency of language generation on 23 datasets, our model improves over various baselines, including the much larger GPT-3.5 (ChatGPT) and sometimes even GPT-4; (3) The lightweight model can also serve as an add-on component for LLMs such as GPT-3.5 in question answering tasks, improving the average exact match (EM) score by 17.94 and F1 score by 15.05 through identifying unanswerable questions.

Paper

Similar papers

Peer review

Reviewer Bk1X7/10 · confidence 4/52023-07-05

Summary

While next word prediction produces task-general models that can do a wide variety of tasks when prompted, it is not an efficient formulation in that it requires very large models. On the other hand, fine-tuned models achieve higher performance at smaller sizes but are specific to a few tasks. This paper proposes text alignment as a middle ground that encompasses a wide range of tasks while allowing for smaller models than next word prediction. Concretely, they convert 28 datasets (encompassing tasks like entailment, IR, QA, coref, and consistency) into the text alignment format and fine-tune RoBERTa on them. The resulting model outperforms much larger models (that are instruction-finetuned), as well as RoBERTa with task-specific fine-tuning.

Strengths

(1) The idea of using text alignment as a task-general interface is interesting and seems useful for producing useful task-general models at smaller sizes. (2) The experiments are extremely thorough, and the method performs well across the board. (3) The paper is well-written and clear.

Weaknesses

While this section contains some suggested experiments, I support the acceptance of this paper regardless of whether or not they are run during the rebuttal period. (1) While the paper frames alignment as being more general than multi-task finetuning, they only evaluate the model on tasks seen during alignment finetuning. Indeed, as shown in the task ablations in Table 5, it seems that the model can only do tasks that are included during training. Therefore, I wonder how different the model is from simply doing multi-task fine-tuning, and what clear advantages are provided by using a unified interface. While already strong, I think the results would be even stronger if there were examples of the model doing unseen tasks. (2) Related to questions of how general alignment is, the tasks included in the training and evaluation feel very close to entailment, which might be a key factor in enabling them all to use the same interface without task-specific heads. Therefore, I wonder how performance would be affected if tasks very far from entailment were included. (For example, you could include POS tagging, where x_1 is the original sentence and x_2 is the sentence with some or all of the words replaced by their part of speech.) (3) Related to the above, while the paper claims that changing the interface from next word prediction to alignment allows for smaller models, I wonder if the smaller model sizes are simply a result of considering a narrower set of tasks than the instruction-finetuned models. I suspect that FLAN-T5 needs larger model sizes simply because it needs more capacity to do more tasks. One relevant ablation testing this question would be to take the same RoBERTa model and instruction-finetune it on the same 28 datasets as the alignment model.

Questions

(1) For handling longer contexts, I wonder if it would make sense to take min_j max_i f(x_1^i, x_2^j) instead of mean_j max_i f(x_1^i, x_2^j) as in the paper, with the interpretation that all of the facts in x_2 should be supported. The mean could fail if x_2 contains mostly supported statements, except for an egregiously unsupported statement in a single chunk.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

4 excellent

Presentation

4 excellent

Contribution

4 excellent

Limitations

Addressed

Reviewer zhhb4/10 · confidence 4/52023-07-05

Summary

This work proposed a text alignment model for a wide range of tasks that aims to measure the degree of alignment between their information. To be more specific, 5.9M examples from 28 datasets are used to fine-tune RoBERTa model. Experimental results show that the text alignment-enhanced model delivers comparable or superior performance compared to larger LMs, validating the effectiveness of the proposed method.

Strengths

1. 5.9M examples from 28 datasets are extracted for LMs fine-tuning, and the experimental result on in-domain datasets (Table 1) and zero-shot setup (Table 2) demonstrate the effectiveness of Alignment-RoBERTa. 2. Experimental code has been submitted, benefitting future pre-fine-tuning research.

Weaknesses

1. The previous pre-fine-tuning work proposed combining multiple losses with different weights [1], which diminishes the technical contribution of this work in terms of pre-fine-tuning. 2. Additional synthetic data is needed, increasing the complexity of the proposed pre-fine-tuning method. Also, the ablation study and corresponding are insufficient in terms of how much contribution synthetic data makes to the overall performance. 3. Experimental setup is not convincing because of the contamination of the training set (seen task) and test set (unseen task). Task clustering is needed to remove the concern. [1] https://aclanthology.org/2021.emnlp-main.468.pdf

Questions

As for Flan models, there is a potential concern regarding contamination of the training set and test set, especially when it comes to the in-domain test, as shown in Table 1 and Figure 2. I understand you have seen and unseen datasets. However, the training and test sets may not have the exact same data points, but they share similar patterns or characteristics. In this case, things become very trivial: the improvement is just because your 5.9M contains similar data instances as in the test set. Given this context, I have two questions: 1. Did you experiment with an out-of-domain test setting for the alignment-based RoBERTa model? I'd like to know if you cluster the tasks to determine the task boundary, as many tasks have a high overlap. 2. Line 634 in the appendix: you mentioned for each dataset, we only use the first 500k samples. How did you determine the number "500k," and how will performance change if you vary this number?

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

The authors state the limitation regarding biases in the constructed dataset, and it is an inevitable issue in the training data.

Reviewer 6RaC7/10 · confidence 4/52023-07-06

Summary

This paper leverages the fact that a lot of popular comparison based NLP tasks like entailment, paraphrase detection, semantic similarity judgement, multiple choice passage based QA etc. amount to learning a specific similarity function between two sets of sequences that is a proxy for “information alignment” between the two sets. Hence, this paper gathers together public datasets for many such tasks and with some light rule-based data augmentation results in 5.9 training examples from 28 datasets. Then a moderately sized RoBERTa model is finetuned on this big comparison-based dataset which is then compared to task specific models and larger models like FLAN T5 on many such datasets. This model has also been used as a metric for measuring factual consistency of NLG models like summarization models. The authors also use this model to detect questions that are unanswerable from the accompanying context to boost performance of systems on some QA datasets.

Strengths

– The paper is well-motivated and the large aggregated dataset and the model trained on it will be useful to the community for further study. – This model outperforms larger general models like FLAN T5 and is competitive with task specific finetuned RoBERTa models on various semantic comparison tasks (some of them unseen during training) which shows the effectiveness of similarity between various such tasks. – The results on summarization evaluation are promising and the usage of this model for identifying unanswerable questions is interesting. – The ablation study shows an interesting trend indicating that various comparison-based datasets and tasks are very similar and compatible with each other. This raises interesting questions related to the nature of these tasks and datasets.

Weaknesses

– No comparison is made against specialized task-specific models. While fine-tuning RoBERTa on task-specific datasets is informative, a deeper insight into how the proposed model fares in comparison to more focused task-specific models will strengthen the comparison. – While this paper focuses on comparison and understanding tasks, it is compared to larger generative models that are specially designed for natural language generation. While these models show impressive performance on these understanding/comparison tasks, a more informative comparison would be against larger encoder-based models that are specifically trained to do well on these datasets and benchmarks like SuperGLUE.

Questions

See above

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

See above

Reviewer ayzf5/10 · confidence 4/52023-07-07

Summary

This paper proposes a way to cast a variety of classification tasks into a single text alignment task. The authors found out that using the text alignment task could generate better results on certain downstream tasks, compared to Flan-T5 and GPT-3.5.

Strengths

The paper presents a novel approach by framing everything as an alignment task. The results are good, and I think the verifier results are interesting.

Weaknesses

I understand that the text-alignment model is not well-suited for generative tasks. Therefore, I believe that the conclusion of the paper is a bit unfair to Flan T5 and Gpt-3, which can be used for generative tasks. It is well-known that sometimes specialized small models can be performed better than by much larger general models. Somehow, the proposed method sounds like another trade-off. I think the proposed method is reasonable, but I believe that further research into different trade-offs would be beneficial.

Questions

Also, more baseline should be compared to. If the model just train the the same subset of the task without casting them into alignment model, will the model still perform well?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The proposed model cannot handle generative tasks.

Reviewer Bk1X2023-08-15

Thanks for the very thorough response. All of my questions were answered.

Reviewer 6RaC2023-08-16

thanks for the rebuttal

The author response was informative and i am keeping my current score.

Reviewer zhhb2023-08-21

Thank you for your answers. The term "Pre-fine-tuning" is used in the title of [1]. It herein refers to using 5.9M data instances (in your case) to further tune the model and test in the downstream tasks without additional task-specific fine-tuning. So I think I get this right. Your answer partially addresses my concerns. However, you have not addressed the concern over test set contamination. In the previous FLAN paper [2], section 2 stated that "we group datasets into clusters by task type and hold out each task cluster for evaluation while instruction tuning on all remaining clusters." The reason for doing this is straightforward: training on one QA dataset, for example, might help the model do better on another QA dataset. Therefore, they group all datasets into clusters by type of task and hold out not just the single QA dataset but the entire QA task cluster to which the dataset belongs. In contrast, in your case, the QA datasets appear in both your training data (Table 1) and test data (Table 2). That's why I have this concern regarding training/test set contamination. The model may learn very simple superficial cues/patterns in the training set without literally understanding them. In short, seen/unseen datasets SHOULD be determined by task cluster instead of a single task. [1] https://aclanthology.org/2021.emnlp-main.468.pdf [2] https://arxiv.org/pdf/2109.01652.pdf

Authorsrebuttal2023-08-21

Thanks for your feedback. **Pre-fine-tuning** We’d like to clarify that the term “pre-fine-tuning” as used in [1] does not refer to what you described. In particular, **“pre-fine-tuning” DOES require additional task-specific fine-tuning** before applying the “pre-fine-tuned” model to a downstream task. In contrast, as you pointed out, **our method** applies to downstream tasks **WITHOUT** additional task-specific fine-tuning. Our work is thus fundamentally different from “pre-fine-tuning” methods as in [1]. More specifically, the work [1] explicitly uses task-specific finetuning after “pre-fine-tuning”: 1) The MUPPET authors clearly state in the abstract they *“propose pre-finetuning, an additional large scale learning stage **between** language model **pre-training** and **fine-tuning**”*. Furthermore, in the experimental setup, they mention that *“We first show that pre-finetuning improves the representations of pre-training models. To do so, we **fine-tune** our pre-finetuned models on a large set of tasks”* and *“**Finetuning** Outside of Pre-Finetuning Domain”*. These statements highlight the need for finetuning in the pre-finetuning method, which our alignment model does not require. 2) In 3.5 Experimental Setup of the MUPPET paper, the authors mention that *“Every Sentence Prediction dataset gets a **separate** classification head, for Commonsense and MRC we utilize a separate unified head for each task.”* However, our alignment model does not necessitate assigning a specific head for each downstream task. **Test Set Contamination** **There could be different ways of “grouping” and distinguishing between seen and unseen tasks/datasets.** For example, the FLAN paper groups tasks/datasets based on the *task type* as you mentioned. Under this perspective, the task of *factual consistency evaluation* (i.e., predicting a factual consistency score, Section 4.2) is of a different type than any training tasks (e.g., QA), and our model shows substantial improvement over diverse baselines on as many as 23 datasets. On the other hand, as we mentioned in the initial response, one could also define unseen or out-of-domain tasks/datasets as those of different textual patterns/characteristics than training tasks/datasets. For example (as mentioned in the initial response), the dialogue datasets in *factual consistency evaluation* (Section 4.2) are of very different patterns/characteristics than training data (and they are also of different task types); the DREAM QA dataset in Section 4.1 also involves very different forms of context (i.e., dialogue) compared to the QA tasks in training. **In either way, our rich experimental results have shown that our approach does generalize to unseen tasks and show strong improvements over diverse baselines.** Reviewer Bk1X has commended our rich experiments, while both reviewer ayzf and 6RaC concur on our strong results. We will make this clearer. **Reference** [1] Aghajanyan, Armen, et al. "Muppet: Massive Multi-task Representations with Pre-Finetuning." Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

Reviewer zhhb2023-08-21

Thank you for your answers. Regarding the term "pre-fine-tuning," I was referring to your case rather than [1]. For test set contamination, you mentioned, "one could also define unseen or out-of-domain tasks/datasets as those of different textual patterns/characteristics than training tasks/datasets..." Then, it would be necessary to justify why you think your textual patterns/characteristics-based method is better than the task category-based method (used in [2]) regarding seen/unseen split. Alternatively, it is better to have a quantitative analysis of your division between seen and unseen tasks to endorse your strong performance improvement, such as high-order n-gram overlap checking in [3] (see section 4) or manual checking in the PaLM model [4] (see section 8). Therefore, I stick to my previous score. [1] https://aclanthology.org/2021.emnlp-main.468.pdf [2] https://arxiv.org/pdf/2109.01652.pdf [3] Language Models are Few-Shot Learners [4] PaLM: Scaling Language Modeling with Pathways

Authorsrebuttal2023-08-21

Thank you for your feedback again. **(1) "pre-fine-tuning"** It looks like we've already addressed your original concerns regarding the novelty of our work (i.e., the weakness-1 in your original comment). It seems we both agree our work is different from [1] as we don't require additional task-specific finetuning. Please let us know if you have any further questions in this regard. **(2) contamination** We'd like to clarify again that we have evaluated on "unseen" tasks/datasets of both **unseen task types** and **unseen patterns/characteristics**. We've given concrete examples of **both** cases in both our initial response and our second response. We were not arguing one way of seen/unseen splitting is better than the other -- We have evaluated and shown strong improvements in **both** ways.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC