企业智能知识库解决方案
智能客服解决方案
对话式 BI 解决方案澜舟论文助写，由 NLP 领域扩展至面向 AI 全领域
2023-01-11
新年伊始，很开心向大家宣布，澜舟论文助写完成一次产品升级！本次发布的版本由仅支持 NLP 领域的模型扩展到可面向 AI 全领域，涵盖自然语言处理、图像处理、语音处理、机器人、机器学习、深度学习等众多 AI 相关的写作领域。
接下来将为大家展示本次升级在各领域的具体应用，以及与最近火热的 ChatGPT 模型简单做一下效果对比。
澜舟论文助写简介#
澜舟论文助写（Langboat Paper Assistant，简称 LPA）是一款帮助中国学生、老师、研究者撰写专业英文论文的神器。
澜舟论文助写提供续写和组句两种能力生成推荐例句：
- 续写：输入一句话，它会推荐你接下来怎么写，通过生成的句子推荐用词、句型和调子；
- 组句：输入若干关键词，它会帮你组织成一个合理的句子，供参考。
通过这两种能力，用户可以得到更加专业的写作风格推荐，同时在参考推荐句式的基础上，可以根据自身需求对内容进行修改，来获得自己想要表达的内容。基于修改后的内容可以继续进行生成，通过交互式写作，可以更好把握用户想要表达的内容，从而在下文推荐更好的符合语境的例句。
在技术实现上，澜舟论文助写基于自研的孟子轻量化模型技术:
- 通过可控学习，用户可以提供关键词引导模型推荐符合用户期望的例句，使得模型推荐内容更加贴近用户需求；
- 通过多任务联合学习，最终可以通过一个模型实现可控推荐和续写推荐的能力，从而有效降低模型部署成本；
- 基于多实例对比的强化学习，首先训练针对论文写作质量的评分模型，然后通过标准数据和采样生成的数据构建多组对比实例，通过强化学习优化模型，提高模型对得分高的例句的生成概率，降低模型对得分低的例句的生成概率，优化时同时考虑多个例句，从而可以有效提高模型生成高质量例句的能力。
澜舟论文助写示例#
下面展示了澜舟论文助写在多个 AI 领域的写作效果，首先我们借助不同领域对比两者的续写能力。
| NLP领域 | |
|---|---|
| 输入上文 | Making language models bigger does not inherently make them better at following a user’s intent. For example, large language models can generate outputs that are untruthful, |
| 澜舟 LPA 生成结果 | inconsistent, and often lack a coherent personality. A growing body of work has shown that these weaknesses can be remedied by using human examples as prompts to guide the generation. One could use this line of work to guide the rest of our models towards more understandable texts, and perhaps better coherence. However, we can be somewhat assured that our language models are following the desired intents to the best of our abilities. In this paper, we explore two mechanisms for using human-written examples to guide generation: using a large language model as a decoder, and adding a classifier to the language models. We conduct a thorough analysis of these methods on the tasks of open domain question answering and dialog modeling, comparing with standard fine-tuning, as well as other baselines. In both cases, our methods show consistent improvements over strong fine tuning baselines, and outperforms existing state of the art approaches. |
| ChatGPT 生成结果 | biased, or simply nonsensical. While larger models may have a greater capacity to memorize patterns in data and generate more diverse outputs, they may also be more prone to overfitting and less able to generalize to new situations. Additionally, simply increasing the size of a language model does not address the underlying issue of providing a model with a clear understanding of the user's intent or context. To truly improve a model's ability to follow a user's intent, it is important to carefully consider not just the size of the model, but also the quality and diversity of the training data, the design of the model architecture, and the use of appropriate task-specific objectives and evaluation metrics. |
| CV领域 | |
|---|---|
| 输入上文 | Image synthesis is one of the computer vision fields with the most spectacular recent development, but also among those with the greatest computational demands. |
| 澜舟 LPA 生成结果 | The recent generative adversarial network models allow for high-quality image synthesis, but their training is often time consuming and prone to failure. One of the main challenges is how to use multiple generators in a way that maintains consistency between generators, while maintaining image quality. In this work, we tackle the consistency problem of GANs by introducing a novel differentiable image synthesis method and its application in 3D-aware image generation. The novel differentiability comes from a new formulation called gradient-matching that naturally extends the gradient optimization of GANs to the case when the image is a 3D array. |
| ChatGPT 生成结果 | Image synthesis involves the generation of new images or videos from a set of input images or videos, using computer algorithms. One of the main goals of image synthesis is to create realistic images or videos that are indistinguishable from real ones. Some of the applications of image synthesis include creating virtual reality experiences, augmenting or enhancing images or videos, and creating new content for movies or video games. Despite the impressive results that have been achieved in image synthesis, it remains a challenging task due to the large amount of data and computational resources required. This has led to the development of specialized hardware and algorithms that are optimized for image synthesis tasks. |
| 语音领域 | |
|---|---|
| 输入上文 | These pre-trained audio encoders learn high-quality representations of speech, but because they are purely unsupervised they lack an equivalently performant decoder mapping those representations to usable outputs. |
| 澜舟 LPA 生成结果 | We find that this limitation can be removed with a small amount of additional task-specific supervised fine-tuning. We observe consistent and significant improvements across a wide array of tasks, including the recent challenging LRS3 speech translation benchmark, and our fine-tuned models are able to outperform comparable prior work. Furthermore, the benefits of this approach do not require as many models or compute for training. In summary: • Audio pre-training learns general purpose speech representations useful for downstream tasks. • Audio pre-trained models are able to take advantage of task-specific training data e.g., surpassing prior unsupervised methods on a wide range of tasks, including the challenging LRS3 benchmark. |
| ChatGPT 生成结果 | As a result, while they may be effective at recognizing and transcribing spoken language, they are limited in their ability to generate realistic synthesized speech or perform other tasks that require a decoder. To overcome this limitation, researchers have explored techniques such as fine-tuning the encoder-decoder architecture on supervised tasks or using supervised learning to train the decoder independently. However, these approaches can be time-consuming and resource-intensive, and may not always achieve the desired level of performance. As such, there is ongoing research into more effective methods for training both encoders and decoders for speech processing tasks. |
除了续写能力，澜舟 LPA 还提供了组句能力，可以根据关键词控制后续推荐例句的内容，同样对比看一下：
| NLP领域 | |
|---|---|
| 输入上文 | Making language models bigger does not inherently make them better at following a user’s intent. In this work, we proposed ChatGPT, |
| 输入关键词 | Transformer, reinforcement learning, human feedback |
| 澜舟 LPA 生成结果 | a new Transformer LM for dialogue, trained using reinforcement learning from human feedback. We demonstrate via human and automatic metrics that ChatGPT outperforms a strong conversational GPT2 baseline. The key factor behind this is the reward function which helps the model to learn better dialogues without relying on task-specific labels or train-inference mismatch. We show that the reward from even a small number of feedback iterations can be used in a semi-supervised framework to scale a dialogue model for production. |
| ChatGPT 生成结果 | a variant of the Transformer architecture that uses reinforcement learning and human feedback to improve the model’s ability to follow a user’s intent. Our experimental results showed that ChatGPT was able to significantly outperform a baseline Transformer model on a range of conversational tasks, demonstrating the effectiveness of our approach. |
相比 ChatGPT 模型，通过组句和续写这两种能力生成的写作效果来看，澜舟论文助写生成的结果在句式上和前后文结构上更具学术范儿，生成论文的风格更偏专业性，前后文的连贯性和句式表达上非常切合 ACL、CVPR、AAAI、ICML 等 AI 专业领域论文的风格；而 ChatGPT 生成的结果更偏向通用的解释和表达。由此可见，通过对特定领域的持续优化和改进，孟子模型在严谨的论文助写方面也是表现非凡的。
LPA 体验网址：https://acl.langboat.com/ （目前限于PC浏览器）
展望#
一篇好的英语论文，不仅需要需要好的内容创新也需要更好地组织和表达内容，同时提高作者的写作效率也至关重要，围绕论文写作领域，我们即将发布论文润色、改写和语法检查等能力，期待能为英语论文写作者提供更多的辅助。
声明：LPA 提供的两个功能关键词组句和续写推荐均是利用孟子模型动态生成句子。注意它不是到一个论文库里面检索已有的发表文章的句子，因此不会存在抄袭问题。同时，不要指望 AI 帮你把一篇论文从头到尾臆造出来。LPA 系统生成的句子的目的仅仅是给作者推荐用词和句型，作者必须根据自己的思路对句子进行编辑修正以最佳地体现自己要表达的内容。
