# Math23K

Math23K is a Chinese dataset of 23,162 math word problems used to train and evaluate AI models for automatic problem solving and equation generation.

Math23K is a widely used benchmark dataset in the field of [machine-learning](https://www.wikiprompt.org/wiki/machine-learning) and [artificial-intelligence](https://www.wikiprompt.org/wiki/artificial-intelligence), specifically designed for the task of solving math word problems. The dataset consists of 23,162 Chinese-language elementary school math problems, each paired with its corresponding equation or answer. It was introduced to address the scarcity of large-scale, non-English datasets for mathematical reasoning, which had previously been dominated by English resources. Researchers use Math23K to develop and test models that can parse natural language, extract numerical relationships, and generate correct mathematical expressions.

The dataset was created by a team from the Chinese Academy of Sciences and was first presented in a 2017 paper titled "A Goal-Driven Tree-Structured Neural Model for Math Word Problems." The problems cover a range of arithmetic operations, including addition, subtraction, multiplication, and division, often involving real-world scenarios such as shopping, travel, and age calculations. Each problem is annotated with a target equation template, allowing models to be trained in a supervised manner. Math23K has become a standard evaluation set for Chinese math word problem solvers, complementing English benchmarks like the MAWPS and ASDiv datasets.

## Dataset Structure and Annotation

Math23K contains 23,162 problems, split into training, validation, and test sets. The standard split, as used in most published research, allocates 21,162 problems for training, 1,000 for validation, and 1,000 for testing. Each problem is a short Chinese sentence or paragraph, typically 20 to 50 characters in length, followed by a gold-standard equation. For example, a problem might state: "小明有5个苹果，小红给了他3个，他一共有多少个？" (Xiao Ming has 5 apples, Xiao Hong gives him 3, how many does he have in total?), with the equation "5+3=8".

The annotations are normalized to a canonical form, where numbers are replaced with placeholders (e.g., "n1", "n2") to create equation templates. This design enables models to learn structural patterns independent of specific numeric values, improving generalization. The dataset also includes multiple-choice answers in some versions, but the primary format is open-ended equation generation.

## Model Architectures and Approaches

Since its release, Math23K has been used to benchmark a variety of [neural-network](https://www.wikiprompt.org/wiki/neural-network) architectures. Early approaches relied on [sequence-to-sequence](https://www.wikiprompt.org/wiki/sequence-to-sequence) models with [encoder-decoder](https://www.wikiprompt.org/wiki/encoder-decoder) structures, which treated the problem as a translation task from text to equations. These were later enhanced with attention mechanisms and [multi-head-attention](https://www.wikiprompt.org/wiki/multi-head-attention) to better align words with numbers. A notable breakthrough came with tree-structured models that explicitly represent the hierarchical nature of arithmetic expressions, such as the goal-driven tree-structured neural model from the original paper.

More recent work has applied [transformer](https://www.wikiprompt.org/wiki/transformer)-based models, including [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)s fine-tuned on Math23K. These models leverage pretrained representations from Chinese corpora and achieve state-of-the-art accuracy, often exceeding 80% on the test set. However, the dataset remains challenging due to its diverse problem types and the need for multi-step reasoning. Researchers have also explored [curriculum-learning](https://www.wikiprompt.org/wiki/curriculum-learning) and [data-augmentation](https://www.wikiprompt.org/wiki/data-augmentation) techniques to improve robustness.

## Evaluation Metrics and Challenges

The primary metric for Math23K is equation accuracy, which measures the percentage of problems where the generated equation exactly matches the gold-standard equation. Some studies also report answer accuracy, which considers the final numeric result. The dataset's difficulty stems from several factors: ambiguous language, implicit operations, and the need to infer quantities from context. For instance, problems may use words like "多" (more) or "少" (less) to indicate addition or subtraction, but the exact operation depends on the sentence structure.

Another challenge is the presence of irrelevant or redundant information, which can mislead models. Additionally, Math23K includes problems with multiple valid equations, but the gold standard only provides one, making exact-match evaluation strict. As a result, models that produce correct answers but different equation forms are penalized, prompting researchers to develop more flexible evaluation methods.

## Impact and Applications

Math23K has significantly influenced research in mathematical reasoning within [deep-learning](https://www.wikiprompt.org/wiki/deep-learning). It has been used to study the generalization capabilities of models across different problem types and to compare the effectiveness of various training strategies. The dataset has also been integrated into larger benchmarks, such as the Chinese Math Word Problem Corpus, and has inspired similar datasets in other languages, including English and Japanese.

In practical applications, models trained on Math23K are used in educational technology, such as intelligent tutoring systems and automated homework grading. Companies like [alibaba-cloud](https://www.wikiprompt.org/wiki/alibaba-cloud) and research institutions have explored deploying these models in real-world classrooms. However, the dataset's focus on elementary-level arithmetic limits its applicability to advanced mathematics, and efforts are underway to create more complex datasets.

## Limitations and Future Directions

Despite its popularity, Math23K has known limitations. The problems are relatively simple, involving at most two or three operations, and lack the complexity of real-world mathematical reasoning. The Chinese language also introduces specific linguistic challenges, such as the use of measure words and flexible word order, which may not transfer to other languages. Furthermore, the dataset does not include visual or multi-modal information, restricting its use in tasks that combine text and diagrams.

Future research aims to expand Math23K with more diverse problem types, richer annotations, and multi-step reasoning chains. Some efforts are integrating it with [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)s to explore few-shot and zero-shot learning, where models are prompted with a few examples instead of being fine-tuned. As of 2025, Math23K remains a standard baseline, but newer datasets like Math23K-v2 and Chinese MathQA are emerging to address its shortcomings.

## See Also

- [machine-learning](https://www.wikiprompt.org/wiki/machine-learning)
- [neural-network](https://www.wikiprompt.org/wiki/neural-network)
- [transformer](https://www.wikiprompt.org/wiki/transformer)
- [large-language-model](https://www.wikiprompt.org/wiki/large-language-model)
- [sequence-to-sequence](https://www.wikiprompt.org/wiki/sequence-to-sequence)

---
Source: https://www.wikiprompt.org/wiki/math23k
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T03:53:51.285001+00:00