Uni-Parser 1.4：化学文献与专利解析能力全面跃升

今天，我们正式发布 Uni-Parser 1.4。

这是 Uni-Parser 面向科学文献与专利解析的一次重要升级。在化学文献与专利解析的精度与吞吐能力上，我们都实现了大幅提升，同时进一步增强了对复杂科学文档结构、多模态内容和化学信息的理解与提取能力。

Uni-Parser 不只是把 PDF 转成文本，更重要的是尽可能恢复文档原有的结构与关系：图表对应哪段图注，公式如何与编号配对，分子结构对应哪个标识符，反应路线又如何与上下文关联。在 Uni-Parser 1.4 中，我们继续围绕三件事推进：更准确地理解复杂科学版面和多模态内容，更完整地提取化学信息，更高效地将解析能力接入真实的数据与 Agent 工作流。

精度提升：全面优化复杂版面、多模态内容与化学信息解析，显著提升文本、公式、表格、图表及化学结构的识别与关联准确率。
吞吐提升：针对大规模科学文献与专利解析进行系统优化，在更低资源开销下实现更高吞吐，更适合百万乃至千万级文档处理。
结构理解：统一解析文本、公式、表格、图表、图片、化学反应式和分子结构，并通过 group / pages_tree 保留跨模态关联。
化学解析：从分子区域定位、分子—编号关联，到 SMILES / E-SMILES 与 Markush 标记，让化学信息不再停留在图片中。
开发者接入：在线 Demo、API、Python SDK（Uni-Parser Tools）、CLI、Agent Skill 与 MCP Server 均可按不同工作方式接入。

Uni-Parser 1.4 已上线在线体验，并可搭配开源工具箱 UniParser-Tools 进行Python调用使用。

文档解析能力全面跃升

Uni-Parser 1.4 文档解析能力评测

Uni-Parser 1.4 在通用文档与化学专业解析上实现同步提升：

通用的文档解析能力评测，Uni-Parser-G 1.4（通用模式）在 OmniDocBench 1.5 上获得 95.24 的综合得分，位于第一梯队水平；
化学文档解析能力的评测，Uni-Parser 1.4 在 BioVista 上，化学结构定位mAP@50达到 0.981 ，结构识别综合准确率达到 0.845，均达到SOTA。

从多语言科学文献到复杂化学专利，Uni-Parser 正在把“页面上看得见的内容”转化为可检索、可计算、可供 RAG、Agent 与 AI4Science 直接使用的结构化数据。

此外，PDF 解析能力现已更充分的覆盖 100+ 种语言，对于小语种也能够完整保留并识别页面中的文字、公式、化学结构、图片与复杂版式。

Uni-Parser 多语言与多模态解析示例

在 Uni-Parser 1.4 中，我们也全面开放了化学反应、Chart、谱图及图像解析能力，进一步将科学文档中的视觉信息转化为可直接使用的结构化数据：

化学反应图像：自动解析并拆解为 反应物–反应条件–产物三元组

Chart 与谱图：从图表中恢复底层数据，转换为结构化数据表

段落中的分子结构：识别分子结构，并保留其与上下文的对应关系

表格中的分子结构：恢复表格的行列结构及其中的分子信息

让科学文献中的图、表、谱图和化学信息，不再只是“图片”，而是能够被理解、检索、计算和进一步处理的数据。

面向规模化任务的吞吐提升

解析精度决定单份结果是否可用，吞吐能力则决定解析系统能否真正进入规模化数据生产。面对数十万乃至更大规模的论文与专利集合，仅提升单个模型的推理速度并不够，还需要让文档预处理、语义识别、结果聚合与服务调度形成高效协同的流水线。

Uni-Parser 1.4 service throughput benchmark

本次吞吐测试在配备 24 GB GPU 显存的环境下以单服务、多进程方式进行，使用 256 张不同的真实图片，并分别采用 4 路和 8 路并发请求。测试结果显示，Uni-Parser 1.4 已能够在文本、表格、公式、图表、图片和分子等不同解析模块之间建立稳定的并行处理能力：

通用内容解析：Fast 模式下，文本解析达到 93.72 img/s，表格解析达到 20.67 img/s；High Quality 模式下，公式、图表和图片解析分别达到 30.62 img/s、3.88 img/s 和 2.55 img/s。
化学信息解析：分子 Fast 模式达到 62.53 img/s，分子与表达式模块在参考优化对比中均获得约 1.63 倍吞吐提升，分子区域检测模块提升约 1.27 倍。
端到端并行扩展：在检测配置下，整体吞吐从单进程的 2.461 file/s 提升至双进程的 4.738 file/s、三进程的 7.054 file/s，并在四进程时达到 9.024 file/s，约为单进程的 3.67 倍。

这组结果不仅来自单点模型加速，也来自异步调度、多进程并行和不同语义服务之间的协同。不同解析模块可以根据实际负载独立扩容，使 CPU 预处理、GPU 推理、服务传输和结果聚合尽可能重叠执行，从而减少硬件等待，让有限计算资源持续处理有效任务。

吞吐数据用于展示特定测试配置下的处理能力，并不构成固定性能承诺。实际速度仍会受到页面尺寸、内容密度、解析模式、并发规模、硬件环境和服务部署方式影响；在生产场景中，应使用具有代表性的业务数据完成容量测试与配置评估。

化学信息提取：让分子真正进入数据链路

MolParser 2.0 preview 化学结构识别评测

Uni-Parser 1.4 中，我们带来了全新的 OCSR（化学结构识别）模型： MolParser-2.0-preview ；

MolParser-2.0-preview 在多个核心基准上均取得当前最佳成绩：在 Uni-Parser Bench 上达到 90.3%，在 WildMol-10k 上达到 80.2%，在 BioVista 上达到 84.5%，在 USPTO 上达到 94.0%，全面超过此前版本及其他主流 OCSR 方法。

从规则版面到复杂科学文献，从分子结构到生物医学图像，Uni-Parser 1.4 正在进一步提升科学文档解析的准确性与可靠性。

此外，在化学文献与专利中，最有价值的信息往往不是文本本身，而是嵌在图片里的分子结构、Markush 表达和反应路线。仅仅裁出这些区域还不够；要让它们进入分子库、结构检索和数据分析，还需要把图像继续转换为机器可处理的结构表达，并保留其与原文编号和上下文的联系。

这意味着，一份专利中的化学信息可以从“页面上的一张图”继续流向：

专利分子库与反应库构建；
子结构、相似性与现有技术检索；
化合物—化合物编号—上下文联合索引；
面向 AI4Science 模型的数据生产与质检。

而 Uni-Parser 正是为结构化数据抽取设计，可用于大规模化学结构、化学反应数据库的抽取和构建。

不止于 Markdown：同一份文档，面向不同下游任务输出

解析的终点不是一份“看起来正确”的文本，而是让数据能够进入下一步工作。

Uni-Parser 提供四个层次的结果：

content：格式化全文，适合阅读、搜索、RAG 与 LLM 输入；
objects：包含类型、置信度、坐标与页码的扁平语义块；
pages_dict：按页面组织的版面结果；
pages_tree：保留父子层级与语义分组的树形结构。

不同元素还可以选择不同输出格式：正文使用 Markdown，表格使用 HTML，公式使用 LaTeX，图表或分子区域保留原始图像。对下游系统而言，同一次解析既可以生成便于阅读的全文，也可以提供用于切图、配对、抽取、质检和数据库构建的精细结构。

从一篇论文，到规模化文献数据工程

真实的科学数据工程面对的通常不是一篇论文，而是数万、数百万甚至更大规模的文献集合。解析能力要进入生产环境，除了准确率，还要考虑吞吐、资源利用率、稳定性、失败恢复与扩展成本。

Uni-Parser 将版面分析后的任务送入不同语义队列，并通过异步调度、动态负载均衡与流水线并行，让 CPU 预处理、GPU 推理、服务间传输和结果聚合尽可能重叠执行。不同解析模块可以独立扩容或更新，以适应真实业务中不断变化的文档类型与流量。

对于应用集成，API 支持异步回调；CLI 会保存任务 token 与结构化结果，任务中断后可以继续获取；开发者也可以通过 SDK、Agent Skill 或 MCP Server，把文档解析嵌入现有科研与 Agent 工作流。

开始使用 Uni-Parser 1.4

Uni-Parser 1.4 已通过在线产品与 API 提供服务。开发者也可以使用 UniParser-Tools （推荐），把解析能力接入脚本、批处理、notebook 与 Agent 应用。

在线体验

访问 Uni-Parser 在线 Demo，上传 PDF、图片或提交公网可访问的 PDF 链接，即可查看格式化全文与页面级解析结果。

首次登录会自动获得试用额度。API 按解析成功的页数固定扣减试用额度，每页扣减 0.05 试用额度，不区分解析内容，不进行其他附加计算；解析失败或超时不扣减。

CLI：一条命令解析论文

pip install "git+https://github.com/dptech-corp/UniParser-Tools.git"
uniparser auth
uniparser parse paper.pdf

解析完成后，可以获得完整 Markdown、pages_tree.json 与任务元数据。任务中断或需要重新获取结果时，可使用保存的 token：

uniparser fetch --token "TASK_TOKEN"

Python SDK：接入数据工作流

import os

from uniparser_tools.api.clients import UniParserClient
from uniparser_tools.common.constant import FormatFlag, ParseMode, ParseModeTextual

parser = UniParserClient(
    host="https://uniparser.dp.tech/",
    api_key=os.environ["UNIPARSER_API_KEY"],
)

task = parser.trigger_file(
    file_path="paper.pdf",
    textual=ParseModeTextual.OCRHighQuality,
    equation=ParseMode.OCRHighQuality,
    table=ParseMode.OCRHighQuality,
    molecule=ParseMode.OCRFast,
)

result = parser.get_formatted(
    task["token"],
    content=True,
    textual=FormatFlag.Markdown,
    table=FormatFlag.Markdown,
    equation=FormatFlag.Markdown,
)

print(result["content"])

Agent Skill 与 MCP

UniParser-Tools 提供可安装的 Agent Skill：

npx skills add dptech-corp/UniParser-Tools

安装并配置 API Key 后，可以直接向支持 Skill 的编程助手提出“解析这个 PDF”“提取论文表格”或“将 PDF 转为 Markdown”等请求。

对于采用 Model Context Protocol 的应用，独立的 MCP Server 可以把工具调用转发到 Uni-Parser HTTP API，使科学文档解析成为 Agent 工作流中的标准能力。

AI 与生产资料的结合，才是 AI 真正转化为生产力的开始。

今天，大量科学知识和科学家的探索积累仍沉睡在 PDF、扫描件和图片中。估算，超过 90% 的生产资料尚以非结构化形态存在，构成了一个庞大的“暗数据”宇宙。对 AI 而言，无法被准确读取、关联和计算的数据，几乎等同于不存在的数据。

Uni-Parser 正是为此而生：跨越论文、专利、报告与技术资料等不同场景，将 PDF、图片等非结构化内容转化为统一的 Markdown、JSON 等标准结构化数据。我们的目标，面向大批量科学文献解析任务的高吞吐能力，让数据在进入大模型训练、RAG、知识抽取和 Agent 之前，先成为准确、完整、可追溯的机器可用资产。围绕 "快速、准确、安全、可靠" 四项核心要求，Uni-Parser 支持大批量文档解析，并可根据业务需求提供 SaaS 在线服务、API 调用与私有化部署等交付方式，

对于 AI4Science，这不仅是一次文档解析能力的升级，更是在打通一条从科学资料到科学数据、再到模型与智能体的生产链路。当论文、专利和实验资料中的结构、关系与专业语义被保留下来，它们才能真正进入检索、数据库、模型训练与科研 Agent，成为可以持续流动和复用的生产资料。

我们期待与科研机构、开发者和产业伙伴一起，让更多科学知识走出“暗数据”宇宙——从可阅读走向可检索、可计算，最终成为推动科学发现与产业创新的新生产力。

在线体验：https://uniparser.dp.tech/
API 文档：https://uniparser.dp.tech/api
UniParser-Tools：https://github.com/dptech-corp/UniParser-Tools
技术报告：https://arxiv.org/abs/2512.15098
商业咨询：uniparser@dp.tech / bd@dp.tech

从在线服务到私有化部署，让解析能力真正进入企业生产环境

本服务支持私有化部署，私有化部署请联系 uniparser@dp.tech。

针对 PDF 高并发抽取、中英文解析和表格解析等生产场景，可参考以下服务器配置：

Uni-Parser 1.4: A Major Leap in Parsing Chemical Literature and Patents

Today, we are officially releasing Uni-Parser 1.4.

This is a major upgrade for scientific-literature and patent parsing. We have substantially improved both accuracy and throughput for chemical literature and patents, while further strengthening the system's ability to understand and extract complex scientific-document structures, multimodal content, and chemical information.

Uni-Parser does more than convert PDFs into text. More importantly, it reconstructs as much of the document's original structure and relationships as possible: which caption belongs to a chart, how an equation pairs with its number, which identifier belongs to a molecular structure, and how a reaction scheme connects to its surrounding context. In Uni-Parser 1.4, we continue to focus on three goals: understanding complex scientific layouts and multimodal content more accurately, extracting chemical information more completely, and integrating parsing more efficiently into real-world data and Agent workflows.

Higher accuracy: Comprehensive improvements to complex layouts, multimodal content, and chemical-information parsing significantly improve recognition and association accuracy for text, equations, tables, charts, and chemical structures.
Higher throughput: System-level optimizations for large-scale scientific-literature and patent parsing deliver greater throughput with lower resource overhead, making the service better suited to collections containing millions or tens of millions of documents.
Structural understanding: Text, equations, tables, charts, figures, chemical reactions, and molecular structures are parsed in a unified workflow, while group / pages_tree preserves cross-modal relationships.
Chemical parsing: From molecule-region detection and molecule-to-identifier association to SMILES, E-SMILES, and Markush notation, chemical information no longer remains trapped in images.
Developer access: The online Demo, API, Python SDK (UniParser-Tools), CLI, Agent Skill, and MCP Server support different ways of working.

Uni-Parser 1.4 is now available through the online experience and can be used from Python with the open-source UniParser-Tools toolkit.

A Major Leap in Document Parsing

Uni-Parser 1.4 document-parsing benchmarks

Uni-Parser 1.4 improves both general document parsing and specialized chemical-document parsing:

In general document-parsing evaluations, Uni-Parser-G 1.4 (General mode) achieves an overall score of 95.24 on OmniDocBench 1.5, placing it in the leading tier.
In chemical document-parsing evaluations on BioVista, Uni-Parser 1.4 reaches a chemical-structure localization mAP@50 of 0.981 and an overall structure-recognition accuracy of 0.845, both at state-of-the-art levels.

From multilingual scientific papers to complex chemical patents, Uni-Parser turns content visible on the page into structured data that can be searched, computed, and used directly by RAG systems, Agents, and AI4Science workflows.

PDF parsing now provides broader coverage for more than 100 languages, including low-resource languages, while preserving and recognizing text, equations, chemical structures, figures, and complex layouts.

Uni-Parser multilingual and multimodal parsing examples

In Uni-Parser 1.4, we have also made chemical-reaction, chart, spectrum, and image parsing broadly available, further converting visual information in scientific documents into directly usable structured data:

Chemical-reaction images: Automatically parse and decompose reactions into reactant–condition–product triplets

Charts and spectra: Recover underlying data and convert it into structured data tables

Molecular structures in paragraphs: Recognize molecular structures while preserving their relationship with the surrounding context

Paragraph containing molecular structures

Parsed paragraph containing molecular structures

Molecular structures in tables: Recover the table layout and the molecular information it contains

Parsed table containing molecular structures

Figures, tables, spectra, and chemical information in scientific literature are no longer merely images; they become data that can be understood, retrieved, computed, and processed further.

Higher Throughput for Large-Scale Workloads

Parsing accuracy determines whether an individual result is usable; throughput determines whether a parsing system can enter large-scale data production. For collections containing hundreds of thousands of papers and patents—or more—speeding up a single model is not enough. Document preprocessing, semantic recognition, result aggregation, and service scheduling must work together as an efficient pipeline.

Uni-Parser 1.4 service throughput benchmark

This throughput benchmark used a single service with multiple processes in an environment equipped with 24 GB of GPU memory. It processed 256 distinct real-world images with 4 and 8 concurrent requests. The results show that Uni-Parser 1.4 can maintain stable parallel processing across text, table, equation, chart, figure, and molecule modules:

General-content parsing: In Fast mode, text parsing reaches 93.72 img/s and table parsing reaches 20.67 img/s. In High Quality mode, equation, chart, and figure parsing reach 30.62 img/s, 3.88 img/s, and 2.55 img/s, respectively.
Chemical-information parsing: Molecule parsing reaches 62.53 img/s in Fast mode. In the reference optimization comparison, the molecule and expression modules each gain approximately 1.63x in throughput, while molecule-region detection improves by approximately 1.27x.
End-to-end parallel scaling: Under the detection configuration, overall throughput increases from 2.461 file/s with one process to 4.738 file/s with two, 7.054 file/s with three, and 9.024 file/s with four—approximately 3.67x the single-process throughput.

These results come not only from accelerating individual models but also from asynchronous scheduling, multi-process parallelism, and coordination among semantic services. Individual parsing modules can scale independently according to actual load, allowing CPU preprocessing, GPU inference, service transfer, and result aggregation to overlap as much as possible. This reduces hardware waits and keeps limited computational resources focused on useful work.

The throughput figures demonstrate processing capacity under specific test settings and do not constitute a fixed performance commitment. Actual speed varies with page dimensions, content density, parsing mode, concurrency, hardware environment, and service topology. Production deployments should be sized and evaluated with representative business data.

Chemical Information Extraction: Bringing Molecules into the Data Pipeline

MolParser 2.0 preview chemical-structure recognition benchmarks

Uni-Parser 1.4 introduces a new OCSR (optical chemical structure recognition) model: MolParser-2.0-preview.

MolParser-2.0-preview achieves state-of-the-art results across several core benchmarks: 90.3% on Uni-Parser Bench, 80.2% on WildMol-10k, 84.5% on BioVista, and 94.0% on USPTO, outperforming the previous version and other mainstream OCSR methods.

From regular layouts to complex scientific literature, and from molecular structures to biomedical images, Uni-Parser 1.4 further improves the accuracy and reliability of scientific-document parsing.

In chemical literature and patents, some of the most valuable information is not text but molecular structures, Markush representations, and reaction schemes embedded in images. Cropping these regions is not enough. To bring them into molecule libraries, structure search, and data analysis, the images must be converted into machine-processable structural representations while retaining their relationships with identifiers and surrounding context.

This allows chemical information in a patent to move from an image on the page into:

Patent molecule and reaction database construction;
Substructure, similarity, and prior-art search;
Joint indexing of compounds, compound identifiers, and context;
Data production and quality control for AI4Science models.

Uni-Parser is designed for structured data extraction and can support the large-scale construction of chemical-structure and chemical-reaction databases.

Beyond Markdown: One Document, Multiple Downstream Outputs

The goal of parsing is not a document that merely looks correct; it is data that can move directly into the next stage of work.

Uni-Parser provides four levels of results:

content: Formatted full text for reading, search, RAG, and LLM input;
objects: Flat semantic blocks containing type, confidence, coordinates, and page number;
pages_dict: Layout results organized by page;
pages_tree: A tree structure preserving parent-child hierarchy and semantic grouping.

Different elements can also use different output formats: Markdown for body text, HTML for tables, LaTeX for equations, and original images for chart or molecule regions. A single parse can therefore produce readable full text alongside detailed structures for cropping, association, extraction, quality control, and database construction.

From One Paper to Large-Scale Literature Data Engineering

Real scientific-data engineering rarely involves only one paper. It often involves tens of thousands, millions, or even larger collections of documents. To operate in production, parsing must account not only for accuracy but also for throughput, resource utilization, stability, failure recovery, and scaling cost.

Uni-Parser sends regions identified during layout analysis to different semantic queues. Through asynchronous scheduling, dynamic load balancing, and pipeline parallelism, CPU preprocessing, GPU inference, inter-service transfer, and result aggregation overlap as much as possible. Individual parsing modules can be scaled or updated independently to accommodate changing document types and traffic patterns.

For application integration, the API supports asynchronous callbacks. The CLI saves task tokens and structured results, allowing interrupted tasks to resume result retrieval. Developers can also use the SDK, Agent Skill, or MCP Server to embed document parsing into existing research and Agent workflows.

Get Started with Uni-Parser 1.4

Uni-Parser 1.4 is available through the online product and API. Developers can also use UniParser-Tools (recommended) to integrate parsing into scripts, batch processing, notebooks, and Agent applications.

Online Experience

Visit the Uni-Parser online Demo to upload a PDF or image, or submit a publicly accessible PDF URL, and inspect formatted full text together with page-level parsing results.

Trial credits are granted automatically on first sign-in. The API deducts a fixed 0.05 trial credit for each successfully parsed page, regardless of parsed content and with no additional calculation. Failed or timed-out parsing tasks do not consume credits.

CLI: Parse a Paper with One Command

pip install "git+https://github.com/dptech-corp/UniParser-Tools.git"
uniparser auth
uniparser parse paper.pdf

After parsing, you receive complete Markdown, pages_tree.json, and task metadata. If a task is interrupted or its results need to be fetched again, use the saved token:

uniparser fetch --token "TASK_TOKEN"

Python SDK: Integrate with Data Workflows

import os

from uniparser_tools.api.clients import UniParserClient
from uniparser_tools.common.constant import FormatFlag, ParseMode, ParseModeTextual

parser = UniParserClient(
    host="https://uniparser.dp.tech/",
    api_key=os.environ["UNIPARSER_API_KEY"],
)

task = parser.trigger_file(
    file_path="paper.pdf",
    textual=ParseModeTextual.OCRHighQuality,
    equation=ParseMode.OCRHighQuality,
    table=ParseMode.OCRHighQuality,
    molecule=ParseMode.OCRFast,
)

result = parser.get_formatted(
    task["token"],
    content=True,
    textual=FormatFlag.Markdown,
    table=FormatFlag.Markdown,
    equation=FormatFlag.Markdown,
)

print(result["content"])

Agent Skill and MCP

UniParser-Tools provides an installable Agent Skill:

npx skills add dptech-corp/UniParser-Tools

After installation and API Key configuration, users can ask a Skill-enabled coding assistant to "parse this PDF," "extract the tables from this paper," or "convert this PDF to Markdown."

For applications that use the Model Context Protocol, a standalone MCP Server forwards tool calls to the Uni-Parser HTTP API, making scientific-document parsing a standard capability in Agent workflows.

AI Becomes Truly Productive When It Connects with the Raw Materials of Production

Today, a vast amount of scientific knowledge and accumulated research remains dormant in PDFs, scans, and images. It is estimated that more than 90% of production materials still exist in unstructured form, creating a vast universe of "dark data." For AI, data that cannot be accurately read, connected, and computed is almost equivalent to data that does not exist.

Uni-Parser was created to address this problem. Across papers, patents, reports, and technical materials, it converts unstructured PDFs and images into standardized structured data such as Markdown and JSON. Our goal is to deliver the throughput required for large-scale scientific-document parsing, turning data into accurate, complete, and traceable machine-usable assets before it enters foundation-model training, RAG, knowledge extraction, or Agent workflows. Guided by four core requirements—fast, accurate, secure, and reliable—Uni-Parser supports large document collections and can be delivered as a SaaS online service, API, or private deployment, depending on business needs.

For AI4Science, this is more than an upgrade to document parsing. It connects a production pipeline from scientific materials to scientific data, then onward to models and intelligent Agents. Only when the structures, relationships, and specialized semantics in papers, patents, and experimental materials are preserved can they enter retrieval systems, databases, model training, and research Agents as reusable production assets.

We look forward to working with research institutions, developers, and industry partners to bring more scientific knowledge out of the "dark data" universe—from readable content to searchable and computable data, and ultimately to new productivity that advances scientific discovery and industrial innovation.

Online experience: https://uniparser.dp.tech/
API documentation: https://uniparser.dp.tech/api
UniParser-Tools: https://github.com/dptech-corp/UniParser-Tools
Technical report: https://arxiv.org/abs/2512.15098
Business inquiries: uniparser@dp.tech / bd@dp.tech

From Online Services to Private Deployment: Bringing Document Parsing into Enterprise Production Environments

This service supports private deployment. For private deployment inquiries, contact uniparser@dp.tech.

For production workloads such as high-concurrency PDF extraction, Chinese and English document parsing, and table parsing, the following server configuration can be used as a reference.

Recommended Configuration

CPU: At least 96 logical cores;
GPU: At least 8 GPU accelerators, with at least 24 GB of memory per accelerator;
Memory: At least 1 TB (16 × 64 GB);
Storage: At least 2 × 960 GB SSD + 2 × 3.84 TB SSD;
Network: At least 2 × 25 Gbps.

让科学文献成为可计算的数据：Uni-Parser 1.4 正式发布

Turning Scientific Documents into Computable Data: Uni-Parser 1.4