Extracting information from résumés is typically formulated as a two-stage problem, where the document is first segmented into sections and then each section is processed individually to extract the target entities. Instead, we cast the whole problem as sequence labeling in two levels -lines and tokens- and study model architectures for solving both tasks simultaneously. We build high-quality résumé parsing corpora in English, French, Chinese, Spanish, German, Portuguese, and Swedish. Based on these corpora, we present experimental results that demonstrate the efectiveness of the proposed models for the information extraction task, outperforming approaches introduced in previous work. We conduct an ablation study of the proposed architectures. We also analyze both model performance and resource eficiency, and describe the trade-ofs for model deployment in the context of a production environment.
volves the extraction of relevant information about a candidate from their résumé document into a structured data model. The extracted information, when integrated with downstream recommender systems, can in turn help candidates and recruiters optimize their search.
Résumés1 exhibit significant diversity in their format
and in their use of language due to diferences in the
background, industry, and location of the candidates [
includes a period, a job title, and the employer’s name.
Since résumé parsing occurs early on in the digitalized icant efect on that of the downstream recommender systems. However, the aforementioned diversity in the use of language in résumés makes the parsing problem cial intelligence (AI) approaches. Efective solutions for this task, therefore, require the use of machine learning techniques.
not make public detailed information about their systems.
domain focuses on constrained scenarios that are limited in scope, in the complexity of the target label scheme, or in terms of the size and quality of the annotated datasets.
more stages. In the first stage, they segment the résumé
into sections and groups [
full document as a whole. This is an unusual setting in long text sequences and the set of labels is relatively big.
and convenient from an engineering point of view, but compare it to previous approaches and we also study several design-decisions of our system in terms of their efect on accuracy as well as in time and memory eficiency.
We share experimental observations on résumés in seven tionally, some sections depict a chronological progression previous work experience, and professional skills. Addi- long text documents, this is generally approached as text classification of independent lines without document(e.g. work experience, education), and are naturally fur- level context. The second stage uses a section-specific are typically present in the contact information section. academic literature for sequence labeling, as résumés are recruitment process pipeline, its accuracy has a signif- it is also competitive with the two-stage alternative. We
• Casting the task of résumé parsing as hierarchical Yadav and Bethard (2018) and Li et al. (2022) for a more
sequence labeling, with line-level and token-level comprehensive review of deep neural networks for
seobjectives, and presenting an eficient résumé quence labeling [8, 9].
parsing architecture for simultaneous labeling Prior work on parsing résumés usually divides the
at both levels. We propose two variants of this problem into two tasks, and tackles each separately [1,
model: one optimized for latency and the other 2, 3, 10, 11]. The résumé is first segmented into sections
optimized for performance. and groups, and then section-specific sequence labeling
models are applied to extract target entities. The early
• A comprehensive set of experiments on résumé work by Tosik et al. (2015) focuses on the second task
parsing corpora in English, French, Chinese, only, as they experiment with already-segmented
GerSpanish, German, Portuguese, and Swedish, each man résumés [
Our empirical study suggests that the proposed hi- based on gazetteers and dictionaries.
erarchical sequence labeling model can parse résumés Barducci et al. (2022) work with Italian résumés. They
efectively and outperform previous work, even with a ifrst segment the résumé using a pattern-matching
aptask definition that involves labeling significantly large proach that relies on a language- and country-specific
text sequences and a relatively large number of entity dictionary of keywords [
For the first, we view the résumé as a sequence of lines and infer the per-line labels that belong to diferent section and group types. This is a generalization of the task definition used in previous work, where the label (class) for each line is inferred independent of information about the text or the predicted labels of other lines. We assume that section and group boundaries are always placed at the end of a line, which is the case in all the résumés we came across during this project. The label set for this part of the task includes a total of 18 sections and groups, which are listed in Appendix A.1.
For the second level, we view the résumé as a long sequence of tokens that includes all the tokens from every line concatenated together. We infer the per-token labels that correspond to the diferent entities. The label set for this part of the task includes 17 entities, which are in turn listed in Appendix A.2.
The scope of this paper revolves around the extraction task and therefore we do not focus on the conversion of the original résumé (e.g. a docx or pdf file) into plain text format. Rather, the systems studied in this work assume textual input.
plication, annotation quality is highly important. For that purpose, we implemented a custom web-based annotation tool that allows the user to annotate section and group labels for each line of a résumé, and to annotate entity labels for each arbitrary span of characters.
We developed the annotation guidelines by starting with a rough definition for each label and performing exploratory annotations on a small set of English résumés —a mini-corpus that we later used for onboarding the annotators. The guidelines were then iteratively refined for the whole duration of the project, achieving a stable and rigorous version at the end. In Appendix A we define the section, group, and entity objectives covered in our corpora, and we provide a screenshot of the annotation tool user interface for reference.
Each language corpus was managed as an independent annotation project. We recruited 2 or 3 annotators, who are native speakers of the target language and without specifically seeking domain expertise, through an online freelance marketplace. The annotators did not communicate with each other during the process, maintaining the independence of the multiple annotations. Before starting the annotations on the target corpus, we asked each 4. Corpora annotator to carefully read the guidelines, and annotate the onboarding English mini-corpus. After reviewing We built résumé parsing corpora in English, French, Chi- and providing feedback, the annotator was instructed to nese, Spanish, German, Portuguese, and Swedish. Some annotate all the résumés in the target corpus. statistics on the corpora are reported in Table 1. For each The estimated inter-annotator agreement (IAA) for of these languages, résumés were randomly sampled from the corpus in each language, computed as suggested by public job boards, covering diverse locations and indus- Brandsen et al. (2020) [13] in terms of 1, ranges from tries. For all but Chinese, we controlled the sampling 84.23 to 94.35% and the median is 89.07%. Finally, we adprocess in order to enforce diversity in locations. For ex- judicated the independent annotations in order to obtain ample, although the English corpus is biased toward the the gold standard annotations. This process involved USA, there is a fraction of résumés from other English- resolving any conflicting decisions made by individual speaking countries including the UK, Ireland, Australia, annotators through the majority voting method. In cases New Zealand, South Africa, and India. Although we did where a majority decision was not attainable, the adjunot control for industry variability, we observe a high dicator was instructed to review the decisions of each level of diversity in the selected collections. We then annotator and apply their own criteria to arrive at a final used third-party software to convert into plain text the decision. original files, which came in varied formats such as pdf, doc, and docx.
Text Sequence
Input Text Sequence
Input Text Sequence
Input
Token-level Initial Features Embeddings or Transformer Token-level Initial Features Embeddings or Transformer Token-level Initial Features Embeddings or Transformer
Token-level Representations
each line are concatenated to obtain the input sequence for the BiRNN+CRF architecture. This is visually described in Figure 2. Preliminary experiments, which are The models we use in this work are based on the not presented here because of space constraints, showed BiRNN+CRF architecture. Initial features are first ex- that avoiding the BiRNN component for this last architracted for each token, then combined through bidirec- tecture, i.e. applying CRF directly on the output of the tional recurrent layers, and finally passed through a CRF Transformer-based features, obtains markedly worse relayer to predict the labels. Unless specified otherwise, sults. This is because the two layers capture complementhe input to the model is the entire résumé text after ap- tary aspects of the context: the Transformer encodes plying tokenization. We study two design-decisions: (1) tokens by exclusively considering the context of the curthe choice for initial features, and (2) separate models for rent line, while the BiRNN layer on top contextualizes predicting line and token labels vs. a multi-task model across every line. Because of the typical length of a réthat predicts both jointly. sumé in terms of tokens, we did not explore encoding Initial features. We explore two alternatives: the whole résumé at once with the Transformer encoders used in this work.
Single-task vs. Multi-task. We experiment with: (a) A combination of FastText [14] word embeddings and handcrafted features, which are detailed in
Appendix B. (b) Token representations obtained from the encoder component of a pre-trained T5 [15] model (or an mT5 [16], depending on the language) without ifne-tuning.
(a) Single-task models that perform either line-level sequence labeling (sections and groups) or tokenlevel sequence labeling (entities). (b) Multi-task models that predict labels for both linelevel and token-level tasks simultaneously.
chitecture. For this second case, each line is encoded individually2, and then the token representations for
tokens (see Table 1).
Figure 1 illustrates the model variants. The architecture shown in Figure 1a is a single-task model for linelevel objectives (sections and groups). This architecture takes as input the complete sequence of tokens in the résumé and predicts one label for each line. We train ... ... ta cn oC coCn Transformer
Transformer
Transf. x1 x 2 x 3 x 4 x 5 x 6 x 7 x
8 x9 .. xT-1 xT Line 1
Line 2
Line 3
Line L token: r = ⃖h⃗ ⊕⃖h⃖
. The result is a sequence of line repre- commercial setting. From an operational perspective, the sentations R = (r1, r2, … , r ), which is in turn processed by another BiRNN layer. This aggregation mechanism is training, testing, integration, and maintenance of a single model is simpler and cheaper than for two models.
corpora of Section 4. The main results are summarized in for training and report the micro-average 1 scores (for the positive labels only) on the held-out 10%3. The results compare the two model architectures discussed in
tecture, the two alternatives for initial features: FastText and Transformer-based T5.
The 1 scores for the token sequence labeling task (predicting entities) are reported in Table 2a. Those include the results for the two single-task models that act only on the token-level task, as well as the two multi-task models. The 1 scores for the line sequence labeling task (sections and groups) are shown in Table 2b, again for the two single-task models that act only on the line-level task, and the two multi-task models4.
row 3, and also row 2 with row 4, we see that using
of 2.5% in the goals 1 on English, and a smaller improvement on French, Spanish, Chinese, and Portuguese, but is worse on German and Swedish5. FastText initial features, on the other hand, perform as well or better than Transformer-based features in the line-level task. It is important to consider, though, that the improved error rate of the Transformer-based model comes at a higher computational cost during inference. This consideration is especially important when the model is deployed in a high-load commercial application where latency is a crucial factor.
Section-specific Models
The simplification of model development and
maintenance is even more significant when we contrast the
unified multi-task model described above with the
typical two-stage approach for résumé parsing [
for diferent tasks), and similarly for row 4
markedly less accurate. This might be due to the small size of Swedish data used for pre-training mT5. depicted in Figure 2.
CRF receives as input the concatenation of: (i) the repre- 3Due to the relatively small size of the corpora, we opted against for entity extraction within each specific section type stage, e.g. the error observed for the Single-task models (e.g. one single-task model for the entities related to con- presented in Table 2b. The aim is to provide the practact information, another single-task model for entities titioner with a quantifiable assessment of the trade-of related to work experience, etc). By contrast, the unified between engineering simplicity and task accuracy. multi-task model we proposed is used to label all the entities across the whole résumé at once, regardless of Analysis and Details on Deployment the section type. This simplification, however, comes at a cost of increased error rate since a section-specific model The results already suggest that the Transformer-based has to decide among a much smaller set of labels, and initial features perform generally better for the tokenreceives a shorter text sequence as input. level sequence labeling task. Furthermore, they do not
In this part, we attempt to quantify such degradation. need language-specific handcrafted features, so they can We train section-specific models, i.e. individual models, be readily applied to new languages. On the other hand, for the entities for three of the section types: contact infor- the alternative set of initial features (the combination mation, work experience, and education. Each is trained of word embeddings and handcrafted features) performs and evaluated only on the corresponding segment of better in the line sequence labeling task for detecting the résumés. Segmentation is performed using the gold section and group labels. standard annotations for sections, in order to focus our However, in terms of eficiency, our experiments remeasurements on the token-level task. In Table 3, we re- veal that using word embedding initial features leads port the micro-average 1 scores grouped by the relevant to a considerable improvement in time-eficiency dursections, comparing the performance of each section- ing inference, when compared to the Transformer-based specific model to the proposed unified, multi-task model. features. The inference time for the multi-task model Results are reported for English, French, and Chinese. was measured under both feature sets. On a bare-metal
We show a loss in 1 ranging from 1% to 5% depending server with a single GPU6, we observed a speedup of 7
on the section and language. Since the section-specific of the FastText models compared to Transformer-based
models benefit from the gold standard segmentation of features. Furthermore, when utilizing CPU-only
hardsections, the results should be considered as an upper ware7, the speedup increased substantially to 90. As an
bound of the degradation in error rate. A real-world
system implemented according to the two-stage approach
should expect a compound error carried from the first
6NVIDIA Tesla T4 and Intel® Xeon® Platinum 8259CL CPU @
2.50GHz.
7Intel® Xeon® CPU E5-2630 v2 @ 2.60GHz
example, we note that the multi-task model using Fast- Both result in a significant degradation of performance
Text initial features, deployed on CPU-only servers via with respect to the models including the BiRNN, again
TensorFlow Serving [
The first group involves variants that use, as initial fea- bels. When compared to our proposed model (variant ④),
tures, the combination of FastText word embeddings and we observe a significant drop in performance, suggesting
handcrafted features. Variant ① is the multi-task model that the contextualization across diferent lines in the
presented in Table 2a. The first ablation, variant ②, in- résumé is the critical factor for the performance of the
volves replacing the top-wise CRF layer with a Softmax system. Interestingly, when variant ⑧ is compared to
layer. Both variants have comparable performance, with variant ⑦ —identical, except for fine-tuning— we do see
a small degradation when Softmax is used. The next abla- an improvement in performance, suggesting that without
tion, variant ③, removes the BiRNN layer and thus makes inter-line contextualization, fine-tuning is indeed helpful.
the CRF predict the token labels using the initial features Variant ⑨ is similar to the previous variant but
redirectly. This is a re-implementation of the system pro- places the CRF layer with Softmax. This model is a
posed by Tosik et al. (2015) [
The second group involves variants that apply frozen trol, in which we fine-tune the encoder component of T5 Transformers to each line individually, and then concate- and predict labels using a Softmax on top of the Transnate every line to obtain the initial features (this is visu- former representations for each token. Again, T5 proally described in Figure 2). Variant ④ is the multi-task vides slightly better results with respect to the equivalent model presented in Table 2a. The first ablation, vari- BERT variant. ant ⑤, involves replacing the T5 (or mT5) encoder with In summary, the ablation experiments suggest that the a BERT (or mBERT) encoder [7]. We observe an appre- BiRNN layer, which contextualizes the token representaciable degradation in performance, suggesting that the tions across the entire résumé, has a significant impact pre-trained T5 family of models produces representations that are more useful for our task. Variant ⑥ and ⑦ use T5 and BERT, respectively, but omit the recurrent layer.
in memory as many copies of the Transformer as lines in the target résumé.
Table 4 trade-of analysis of the proposed variants and described
Ablation study. Variants are compared in terms of the micro- challenges for deployment in production environments.
average 1 obtained for the token sequence labeling task. Vari- The ablation experiments suggest that the BiRNN layer
ants ① and ④ represent the models discussed in the previous contextualizing across the résumé is critical for
perforpart of this section. Other model variants depart from either mance, and that the CRF component further provides a
one of these by changing one aspect at a time. In particular, smaller improvement.
vaanrdiavnatr③ianrte-⑨imispleeqmueivnatsletnhtetsoytshteemarocfhTitoescitkueret aplr.o(p20o1s5e)d[b1y], Potential directions for future research include the
folDevlin et al. (2019) [7] for other sequence labeling tasks. IF lowing: using character-based initial features [5, 6] for
denotes initial features. Each result is an average of three in- the FastText variants, as they can complement word
emdependent replications. beddings by incorporating information from the surface
form of the text and may even ofer the opportunity to
Model variant English French Chinese gradually replace handcrafted features; domain-adapting
F①asIFtT+eBxitRinNiNtia+lCfeRaFtures 89.03 86.90 92.66 the Transformer representations with unannotated
ré② IF+BiRNN+Softmax 88.86 86.53 92.67 sumés, considering the reported efectiveness of this
tech③ IF+CRF [
As discussed in Section 4, despite our best eforts to cover as many locations, industries, and seniority levels, it is not feasible for résumé parsing corpora with sizes of up to 1200 résumés to actually contain samples from every subgroup of the population under study. Therefore, we would like to highlight that the findings presented in this work apply specifically to résumés that are similar to those included in the corpora, and may not generalize with the same level of accuracy to other résumés belonging to combinations of location, industry, and work experience that were not seen by the model during training.
Résumé parsing is an important task for digitalized re- Ethics Statement cruitment processes, and the accuracy of the parsing step afects downstream recommender systems significantly. The system described in this work is intended for pars
In this work, we study résumé parsing extensively ing résumés of individuals from diferent backgrounds, in seven languages. We formulated it as a sequence located around the globe. Considering the importance labeling problem in two levels (lines and tokens), and of inclusivity in this context, we made a great efort to studied several variants of a unified model that solves cover the diversity of the use of language in our corboth tasks. We also described the process for develop- pora with the objective in mind. This helps us to provide ing high-quality annotated corpora in seven languages. high-quality résumé parsing for individuals from various We showed through experimental results that the pro- industries and locations. posed models can perform this task efectively despite Furthermore, the data used for training and evaluating the challenges of substantially long input text sequences our models consist of résumés that contain sensitive inforand a large number of labels. We observed that the joint mation from real-world individuals. We have taken the model is more convenient than the typical two-stage so- necessary privacy and security measures for protecting lution in terms of resource eficiency and model life-cycle this information throughout every step of this project. maintainability, and also found that in some cases the joint model yields better performance. We provided a A. Details on the Corpus Besides the categories listed above, lines that belong to sections Work Experience, Education, and Internship Annotations can belong to experience groups. Internally, the tool uses the IOB (inside, outside, beginning) format for the groups Annotators were provided access to a custom annotation within each section type. Note that, for example in the tool that we developed for this task. In Figure 3 we show case of the Education section, the label I-edu denotes a the user interface of this tool, with an example résumé line that is part of the Education section but it’s not part annotated for sections, groups, and entities. of any particular group, whereas the label B-edu_group
The annotators were asked to detect and highlight spe- denotes a line that lies at the beginning of a group in the cific information in résumés written in their language. Education section.
A.2. Labels for Tokens The sequence labeling task at the token-level is intended to extract entities. Entities are annotated on arbitrary spans of characters in the résumé text, in order to generalize the annotations for any possible tokenization. We allow for a total of 17 labels. Although most entities are usually found in specific sections, we allow for annotating any entity in any part of the résumé.
A.2.1. Contact Information entities
information section.
Name Candidate’s name.
Phone number Candidate’s phone number.
Email Candidate’s email address.
St. address Unit information (apartment, floor) and neigh
borhood.
ZipCode An alphanumeric postal code.
City The city the candidate lives in.
State First-level geopolitical subdivision where the candidate
is located.
A.2.2. Work Experience and Internship entities
the work experience or the internship sections. Company Name of a candidate’s employing organization. Job title Title that the employer gave to the candidate while working for the company.
Period Period in which the candidate held the position. A.2.3. Education entities
education section.
School name Name of an institution where the candidate was formally educated.
Degree title he name of the academic program in which a student participates or was awarded a degree.
Degree Period Period of time in which the candidate attended the educational organization in fulfillment of the degree.
Major The core academic discipline the candidate had to focus on while pursuing their degree.
GPA Grade, or any other candidate score, presented as a standardized measurement.
A.2.4. Language entities
Language name The name of a language that the candidate claims to be familiar with.
The non-Transformer models presented in Section 5 employ a combination of FastText word embeddings and handcrafted features. The complete list and details of each handcrafted feature is provided in GitHub because of space considerations9. The final set of features was determined based on the empirical results of preliminary experiments, which are not included in this study due to space constraints. Note that certain features necessitate dictionaries of relevant terms to be computed, thereby requiring separate dictionaries for each language.