Rephrased Labels Improve Zero-Shot Text Classification by 30%

Consider this: given an embedding model $f$, a document $x$, and a candidate label set ${y_1, y_2, y_3}$, you need to determine which label $y$ should be assigned to $x$. Essentially, you are asked to use an embedding model for classification tasks. Formally, this is represented as $y:=\arg\max_{y \in \{y_1, y_2, y_3\}}(c(f(x), y))$, where $c$ is the classifier. This is a classic classification problem that you might learn about in your ML101 course.

However, in this post, we focus on the zero-shot setting, which comes with some constraints:

1. You cannot alter the embedding model $f$—it remains a black box that you cannot fine-tune.
2. You cannot train $c$ since there are no labeled data points available for training.

So, how would you approach this?

Before answering, let's look at whether this is a meaningful question. First, is this a real problem? Second, is it feasible?

• Real-World Relevance: Yes, this is a real scenario. Consider a classification API where users submit data and candidate labels, asking the server to classify the data. In this case, there is no training data, and the server does not fine-tune the embedding model for each user or problem setting. Thus, this is precisely the situation we're dealing with.
• Feasibility: It is indeed feasible. Embedding models, often encoder-only transformers, are trained on vast amounts of data for general purposes. They are evaluated on multiple tasks in the MTEB (Massive Text Embedding Benchmark), including clustering, sentence similarity, retrieval, and classification. Recently, using LLMs or decoder-only transformers as embedding models, such as NV-embed, has become popular. These models can have parameter sizes up to 7 billion! With such large parameter sizes and extensive training data, these embedding models should possess some zero-shot classification capabilities. It is not surprising that they can handle zero-shot, out-of-domain settings, though not perfectly.

tagZero-Shot Classification Baseline

So, how would you solve this problem? Since you cannot train any classifier with zero labeled data, $c$ cannot be parameterized and is mostly just a metric function, like cosine similarity. Therefore, a naive way of performing classification is to embed both the data and all candidate labels using the embedding model, compare the cosine similarity between the data and label embeddings, and select the label with the maximum cosine similarity, i.e. $$y:=\arg\max_{y \in \{y_1, y_2, y_3\}}(\mathrm{cos}(f(x), f(y)))$$ Simple but effective. In fact, this approach is commonly used by researchers when evaluating MTEB classification tasks.

tagCan We Do Better?

Is it possible to improve upon this method? The information we can leverage is limited. We have labels like [sports, politics, finance] or [positive, negative, neutral]. These labels have some inherent semantics (and ambiguities), especially when considered in context with $(x, y)$, but it’s still not a lot to work with. Some approaches rely on external knowledge, such as data domain assumptions or label descriptions, to enhance the label semantics—often through data generation using LLMs. While effective, these methods also introduce additional resource consumption and latency, which we will not consider here.

Liberating Seen Classes: Boosting Few-Shot and Zero-Shot Text Classification via Anchor Generation and Classification Reframing

Few-shot and zero-shot text classification aim to recognize samples from novel classes with limited labeled samples or no labeled samples at all. While prevailing methods have shown promising performance via transferring knowledge from seen classes to unseen classes, they are still limited by (1) Inherent dissimilarities among classes make the transformation of features learned from seen classes to unseen classes both difficult and inefficient. (2) Rare labeled novel samples usually cannot provide enough supervision signals to enable the model to adjust from the source distribution to the target distribution, especially for complicated scenarios. To alleviate the above issues, we propose a simple and effective strategy for few-shot and zero-shot text classification. We aim to liberate the model from the confines of seen classes, thereby enabling it to predict unseen categories without the necessity of training on seen classes. Specifically, for mining more related unseen category knowledge, we utilize a large pre-trained language model to generate pseudo novel samples, and select the most representative ones as category anchors. After that, we convert the multi-class classification task into a binary classification task and use the similarities of query-anchor pairs for prediction to fully leverage the limited supervision signals. Extensive experiments on six widely used public datasets show that our proposed method can outperform other strong baselines significantly in few-shot and zero-shot tasks, even without using any seen class samples.

arXiv.orgHan Liu

Zero-Shot Robustification of Zero-Shot Models

Zero-shot inference is a powerful paradigm that enables the use of large pretrained models for downstream classification tasks without further training. However, these models are vulnerable to inherited biases that can impact their performance. The traditional solution is fine-tuning, but this undermines the key advantage of pretrained models, which is their ability to be used out-of-the-box. We propose RoboShot, a method that improves the robustness of pretrained model embeddings in a fully zero-shot fashion. First, we use language models (LMs) to obtain useful insights from task descriptions. These insights are embedded and used to remove harmful and boost useful components in embeddings -- without any supervision. Theoretically, we provide a simple and tractable model for biases in zero-shot embeddings and give a result characterizing under what conditions our approach can boost performance. Empirically, we evaluate RoboShot on nine image and NLP classification tasks and show an average improvement of 15.98% on worst group accuracy, with trivial decrease in overall accuracy over several zero-shot baselines. Additionally, we demonstrate that RoboShot is compatible with a variety of pretrained and language models and propose a way to further boost performance with a zero-shot adaptation variant.

arXiv.orgDyah Adila

One key observation is that most general-purpose embeddings are better at sentence similarity tasks than on classification tasks. So, if we can somehow rephrase the label into a sentence and compare the data embedding to this rephrased label sentence, could we achieve better performance? If so, what’s the best way to rephrase the label? How much improvement can we expect? And is this method generalizable across different datasets and embedding models? In this post, I conduct some experiments to answer these questions.

tagExperimental Setup

Instead of directly embedding the LABEL and computing the cosine similarity with the embedded document, I first rephrase the LABEL into a more descriptive sentence without looking at the data first. For example, I use constructions like "This is something about $LABEL". Note, this is a pretty generic construction as it does not assume anything of the data: it could be a code snippet, a tweet, a long document, a news article, a soundbite, a docstring, or anything . Specifically, I tested different rephrasing constructions as follows:

label_constructions = [
    lambda LABEL: f'{LABEL}',
    lambda LABEL: f'{LABEL} ' * 4,
    lambda LABEL: f'Label: "{LABEL}".',
    lambda LABEL: f'This is something about "{LABEL}".',
    lambda LABEL: f'This is seriously about "{LABEL}".',
    lambda LABEL: f'This article is about "{LABEL}".',
    lambda LABEL: f'This news is about "{LABEL}".',
    lambda LABEL: f'This is categorized as "{LABEL}".',

    # Descriptive Context Embedding
    lambda LABEL: f'This text discusses the topic of "{LABEL}" in detail.',
    lambda LABEL: f'A comprehensive overview of "{LABEL}".',
    lambda LABEL: f'Exploring various aspects of "{LABEL}".',

    # Question-Based Construction
    lambda LABEL: f'What are the key features of "{LABEL}"?',
    lambda LABEL: f'How does "{LABEL}" impact our daily lives?',
    lambda LABEL: f'Why is "{LABEL}" important in modern society?',

    # Comparative Context Construction
    lambda LABEL: f'"{LABEL}" compared to other related concepts.',
    lambda LABEL: f'The differences between "{LABEL}" and its alternatives.',
    lambda LABEL: f'Understanding "{LABEL}" and why it stands out.',
]

The datasets I’m using for this experiment are TREC, AG News, GoEmotions, and Twitter Eval text classification data. I've also confirmed with our team that those dataset were not used for training our embedding models. Specifically, the labels we try to predict in these datasets are as follows:

• TREC: ['ABBR', 'ENTY', 'DESC', 'HUM', 'LOC', 'NUM']
• AG News: ['World', 'Sports', 'Business', 'Sci/Tech']
• GoEmotions: ['admiration','amusement','anger','annoyance','approval','caring','confusion','curiosity','desire','disappointment','disapproval','disgust','embarrassment','excitement','fear','gratitude','grief','joy','love','nervousness','optimism','pride','realization','relief','remorse','sadness','surprise','neutral']
• Twitter Eval: ['negative', 'neutral', 'positive']

Unlike classic classification tasks, we should not encode the labels into integers (e.g., 0, 1, 2) because we would lose the semantic information inherent in the labels. Moreover, I do not expand the abbreviated labels into their full forms. For example, the full names of the TREC labels are [Description, Entity, Abbreviation, Human, Location, Numeric], but I keep using the original abbreviations in the experiment as:

• It more closely mirrors real-world settings where abbreviated labels are commonly used.
• It allows us to see how well our embedding model can handle labels with extremely weak semantic information. (As I mentioned above, the other extreme is to leverage label descriptions, enhanced or generated by LLMs, which is out of scope.)

The embedding models I'm testing are:

• jina-embeddings-v2-base-en (released in October 2023)
• jina-clip-v1 (released in June 2024)

It might be surprising for some readers to see that I included jina-clip-v1, which is a multimodal embedding model for both image and text. However, our innovation with jina-clip-v1 exactly focused on improving its text retrieval/similarity ability vs. the original CLIP models. So a good chance to test it.

You can try out both models via our Embedding API:

Embedding API

Multimodal, bilingual long-context embeddings for your search and RAG.

The full implementation can be found in the notebook below:

Google Colab

tagExperimental Results

We measure the Precision, Recall, and (weighted average) F1 score for our experiments. To evaluate the effectiveness of our method, we compute the relative improvement over the zero-shot classification baseline where the LABEL is used as it is. Note jina-embeddings-v2-base-en and jina-clip-v1 have different baseline performance. So we only compare the results within the model not between the two models. The results are then sorted in descending order by their (weighted average) F1 score.

For all metrics, the larger the better. The delta percentage represents the relative improvement over the baseline, i.e. the first row.

tagjina-embeddings-v2-base-en on AG News

LABEL = ['World', 'Sports', 'Business', 'Sci/Tech']

tagjina-clip-v1 on AG News

LABEL = ['World', 'Sports', 'Business', 'Sci/Tech']

tagjina-embeddings-v2-base-en on TREC

LABEL = ['ABBR', 'ENTY', 'DESC', 'HUM', 'LOC', 'NUM']

tagjina-clip-v1 on TREC

LABEL = ['ABBR', 'ENTY', 'DESC', 'HUM', 'LOC', 'NUM']

tagjina-embeddings-v2-base-en on GoEmotions

LABEL = ['admiration','amusement','anger','annoyance','approval','caring','confusion','curiosity','desire','disappointment','disapproval','disgust','embarrassment','excitement','fear','gratitude','grief','joy','love','nervousness','optimism','pride','realization','relief','remorse','sadness','surprise','neutral']

tagjina-clip-v1 on GoEmotions

LABEL = ['admiration','amusement','anger','annoyance','approval','caring','confusion','curiosity','desire','disappointment','disapproval','disgust','embarrassment','excitement','fear','gratitude','grief','joy','love','nervousness','optimism','pride','realization','relief','remorse','sadness','surprise','neutral']

tagjina-embeddings-v2-base-en on Twitter Eval

LABEL = ['negative', 'neutral', 'positive']

tagjina-clip-v1 on Twitter Eval

LABEL = ['negative', 'neutral', 'positive']

tagConclusion

Depending on the model and dataset, rephrasing labels can be a highly effective strategy, significantly improving zero-shot classification performance by 30%. Some rephrasing constructions work better when they "hit" exactly the data domain. For example, "This news is about 'LABEL'" performs well on AG News, as observed with both models. Interestingly, generic constructions such as "This is something about 'LABEL'" also work quite well.

Among all the constructions tested, my favorite is "This is seriously about 'LABEL'" as it resonates the "my grandma is dying" vibe in LLM prompting. Either way, this confirms our intuition that converting a classification task into a sentence similarity task can boost zero-shot classification performance.

However, we must acknowledge the limitations of this label rephrasing strategy. When applied to the TREC dataset, where labels are abbreviations with very weak semantics (e.g., ['ABBR', 'ENTY', 'DESC', 'HUM', 'LOC', 'NUM']), rephrasing does not work well. In fact, any kind of label rephrasing underperforms the baseline. This is likely because the very limited semantics in these abbreviations get "whitened out" by other words during the rephrasing process. This also explains why repeating the LABEL four times does not drop the performance too much compared to other constructions.

This leaves us with one final question: what is the best and most generic label construction? Is there a specific trick like "This is seriously about 'LABEL'" or can we use DSPy to figure it out via discrete optimization?

In fact, this line of thinking is at the heart of jina-embeddings-v3. The idea is to introduce a learnable special token during the training time, such as [ZEROCLS], to indicate the task; and prepend this token during the inference time with [ZEROCLS] LABEL to signal the embedding model to "switch" to zero-shot classification mode. In our experiment above, "This is seriously about" essentially plays the role of [ZEROCLS]. There are just a few weeks left until our v3 release, so stay tuned!