--- title: "Construction of English Resume Corpus and Test with" description: "The authors want to extract information from English resumes without losing sentences that NER cannot capture as spans… 8 references." authors: ["Chengguang Gan", "Tatsunori Mori"] year: 2022 venue: "arXiv:2208.03219" source: "https://arxiv.org/abs/2208.03219" doi: "10.48550/arxiv.2208.03219" arxiv: "2208.03219" licence: "arXiv — non-exclusive license to distribute" licence_url: "http://arxiv.org/licenses/nonexclusive-distrib/1.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/gan-2022-construction-english-resume-corpus-test/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/gan-2022-construction-english-resume-corpus-test/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2022](https://md.research.elamur.ai/free/en/papers/2022/page/9/) / Construction of English Resume Corpus and Test with Pre-trained Language Models # Construction of English Resume Corpus and Test with Pre-trained Language Models **Authors:** Chengguang Gan and Tatsunori Mori · **Year:** 2022 · **Venue:** arXiv:2208.03219 · **Licence:** [arXiv — non-exclusive license to distribute](http://arxiv.org/licenses/nonexclusive-distrib/1.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2208.03219) · [PDF](https://arxiv.org/pdf/2208.03219) · [DOI 10.48550/arxiv.2208.03219](https://doi.org/10.48550/arxiv.2208.03219) · [arXiv 2208.03219](https://arxiv.org/abs/2208.03219) · [OpenAlex](https://openalex.org/W4290772593) **Citations:** 1 citations (OpenAlex, 2026-10-14) ## Summary The authors want to extract information from English resumes without losing sentences that NER cannot capture as spans, such as a candidate's statement of career goals. Following earlier work, they treat each resume sentence as a unit and assign it a section label. The earlier corpus had 5 loosely defined categories, including a catch-all "others" class, and covered 500 resumes. The authors redesigned the scheme into 7 labels so that each sentence belongs clearly to one section. They then used a small custom tool to annotate 1,000 resumes, which gave about 78,000 sentences. They fine-tuned four large PLMs on this data. T5 large scored highest (87.35 F1-micro), and every model beat a BERT baseline trained on the older corpus, though the two corpora use different label sets. A learning-curve experiment indicates that performance levels off at roughly 40,000 training sentences. The resulting resource is a fine-grained, highly imbalanced benchmark for resume sentence classification, which could support automated HR screening. ## Key points - **Corpus:** 1,000 resumes drawn from a Kaggle collection of 15,000, split into about 78,000 sentences. The predecessor used 500 resumes and about 40,000 sentences. - **Label set (7 classes):** Experience, Personal Information, Summary, Education, Qualifications, Skill, Object. These replace experience/knowledge/education/project/others. - **Results (F1-micro, 7:1.5:1.5 random split, mean of 3 runs):** | Model | F1-micro | |---|---| | BERT large, old corpus (baseline) | 85.97 | | BERT large | 86.67 | | ALBERT | 86.40 | | RoBERTa large | 87.00 | | T5 large | 87.35 | The model ranking matches the usual standing of these PLMs on general NLP benchmarks. - **Learning curve:** The setup used 58k train, 10k validation and 10k test sentences, with the training set grown in steps of 5,000. - Validation F1 peaked at 86.6 with 45k training sentences. - Test F1 peaked at 85.9 with both 40k and 45k. - **Class imbalance:** - Experience accounts for about half of all sentences. - Skill, Object and Qualifications make up only 7%, 3% and 1%. - Confusion matrices show that RoBERTa handles the rare Qualifications class better, while T5 is slightly stronger across all classes. ## Interesting ideas - When the useful content is a whole sentence rather than a span, labelling sentences by function keeps the full wording that NER-style extraction throws away. - The authors rebuilt the taxonomy so that no sentence falls into a vague leftover bucket, accepting more classes in exchange for clearer boundaries. - They held validation and test sets fixed and grew the training set in equal steps to find where accuracy stops improving, which tells you when more annotation stops paying off. - Their annotation tool pre-segments each document by rule, gives one button per label, and auto-exports the file before loading the next document, so labelling becomes rapid clicking. - With one class at only 1% of the data, they inspect confusion matrices per class instead of relying on micro-F1 alone, since the majority class dominates the aggregate score. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (frames): A coarse sentence-role layer could run before frame extraction, tagging sentences as goal statements, background or narrative events, so that each sentence's frames inherit a discourse context. - **I2** (learn): When designing labels for silver data or LLM-teacher output, avoid catch-all classes, so that disagreements between witnesses point to real ambiguity rather than to an undefined residue. - **I3** (learn): The same saturation curve could show how much agreed silver treebank data gparse actually needs before extra sentences are wasted effort. - **I4** (learn): A similar minimal interface could let a human quickly resolve the sentences where witness parsers or teacher tiers disagree. - **I5** (tree): Parser evaluation should report per-deprel scores and confusions for rare UD relations, not just overall LAS, so that weak spots on infrequent constructions stay visible. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions Also on the byline: [Yokohama National University](https://ror.org/03zyp6p76). 📚 Shelves (7): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works · [👁️ Computer Vision and Pattern Recognition](https://md.research.elamur.ai/free/en/shelves/computer-science/computer-vision-and-pattern-recognition/) · 3,486 works · [💼 Business, Management and Accounting](https://md.research.elamur.ai/free/en/shelves/business-management-and-accounting/) · 40 works · [💼 Organizational Behavior and Human Resource Management](https://md.research.elamur.ai/free/en/shelves/business-management-and-accounting/organizational-behavior-and-human-resource-management/) · 15 works 🧭 Topics (1): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,379 works ## Cite Cite as: Chengguang Gan and Tatsunori Mori. 2022. Construction of English Resume Corpus and Test with Pre-trained Language Models. arXiv:2208.03219. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{gan2022construction, title = {{Construction of English Resume Corpus and Test with Pre-trained Language Models}}, author = {Chengguang Gan and Tatsunori Mori}, year = {2022}, doi = {10.48550/arxiv.2208.03219}, eprint = {2208.03219}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2208.03219} } ``` - [← previous CompLx@SMM4H'22: In-domain pretrained language models for detection of adverse drug reaction mentions in English tweets](https://md.research.elamur.ai/free/en/papers/xherija-2022-complx-smm4h-22-domain-pretrained/) - [next → Cross-Dialect Social Media Dependency Parsing for Social Scientific Entity Attribute Analysis](https://md.research.elamur.ai/free/en/papers/eggleston-2022-cross-dialect-social-media-dependency/) - [↑ back to papers of 2022](https://md.research.elamur.ai/free/en/papers/2022/page/9/) ## References 8 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. [Bert: Pre-training of deep bidirectional transformers for language understanding](https://md.research.elamur.ai/free/en/papers/devlin-2019-bert-pre-training-deep-bidirectional/). arXiv preprint arXiv:1810.04805. [↗](https://arxiv.org/abs/1810.04805) 2. Chengguang Gan and Ryouei Takahashi. 2021. Verification of the applicability of bert in the english resume data extraction. In Information Processing Society of Japan Kansai Branch IPSJ Kansai-Branch Convention 2021, volume 2021. [find](https://scholar.google.com/scholar?q=Chengguang+Gan+and+Ryouei+Takahashi%2E+2021%2E+Verification+of+the+applicability+of+bert+in+the+english+resume+data+extraction%2E+In+Information+Processing+Society+of+Japan+Kansai+Branch+IPSJ+Kansai%2DBranch+) 3. S. Huang, L. I. Wei, and J. Zhang. 2018. Entity extraction method of resume information based on deep learning. Computer Engineering and Design. Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Figure 5: Confuse Matrix of RoBERTa large and T5 large model in test set. 2019. [Albert: A lite bert for self-supervised learning of language representations](https://md.research.elamur.ai/free/en/papers/lan-2019-albert-lite-bert-self-supervised/). arXiv preprint arXiv:1909.11942. [↗](https://arxiv.org/abs/1909.11942) 4. Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. [Roberta: A robustly optimized bert pretraining approach](https://md.research.elamur.ai/free/en/papers/liu-2019-roberta-robustly-optimized-bert-pretraining/). arXiv preprint arXiv:1907.11692. [↗](https://arxiv.org/abs/1907.11692) 5. R Mooney. 1999. Relational learning of pattern-match rules for information extraction. In Proceedings of the sixteenth national conference on artificial intelligence, volume 328, page 334. [find](https://scholar.google.com/scholar?q=R+Mooney%2E+1999%2E+Relational+learning+of+pattern%2Dmatch+rules+for+information+extraction%2E+In+Proceedings+of+the+sixteenth+national+conference+on+artificial+intelligence%2C+volume+328%2C+page+334%2E) 6. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. 2020. [Exploring the limits of transfer learning with a unified text-to-text transformer](https://md.research.elamur.ai/free/en/papers/raffel-2019-exploring-limits-transfer-learning-unified/). J. Mach. Learn. Res., 21(140):1–67. 7. Yanyuan Su, Jian Zhang, and Jianhao Lu. 2019. The resume corpus: A large dataset for research in information extraction systems. In 2019 15th International Conference on Computational Intelligence and Security (CIS), pages 375–378. IEEE. [find](https://scholar.google.com/scholar?q=Yanyuan+Su%2C+Jian+Zhang%2C+and+Jianhao+Lu%2E+2019%2E+The+resume+corpus%3A+A+large+dataset+for+research+in+information+extraction+systems%2E+In+2019+15th+International+Conference+on+Computational+Intelligence+and) 8. Kun Yu, Gang Guan, and Ming Zhou. 2005. Resume information extraction with cascaded hybrid model. In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL’05), pages 499–506. [find](https://scholar.google.com/scholar?q=Kun+Yu%2C+Gang+Guan%2C+and+Ming+Zhou%2E+2005%2E+Resume+information+extraction+with+cascaded+hybrid+model%2E+In+Proceedings+of+the+43rd+Annual+Meeting+of+the+Association+for+Computational+Linguistics+%28ACL%E2%80%9905%29%2C+p) *Updated 2 November 2026*