Automatic Term Extraction (ATE)

• Automatic Term Extraction (ATE) is the process of identifying domain-specific single-word and multiword terms by assessing termhood and unithood.
• ATE employs methods from statistical ranking and linguistic filtering to supervised neural extraction, aiding tasks like ontology construction and patent analysis.
• Recent advances leverage transformer models and large language model prompting to enhance accuracy, handle nested structures, and mitigate domain shifts.

Searching arXiv for recent and foundational papers on Automatic Term Extraction to ground the article. arxiv_search(query="Automatic Term Extraction survey transformers term extraction arXiv", max_results=5) Automatic Term Extraction (ATE) is the task of automatically identifying domain-specific terms—single-word terms and multiword terms—from corpora. In the ATE literature, terms are linguistic units that denote concepts in a specialized field, and the task is typically framed through the interaction of termhood, which quantifies how representative a candidate is of a domain, and unithood, which quantifies the lexical cohesion of the words forming a multiword unit. Contemporary ATE spans statistical ranking, linguistic filtering, supervised token classification, nested span modeling, contextual-embedding classifiers, and large-language-model prompting or instruction tuning; its outputs support document tagging, ontology construction, terminology management, patent analysis, machine translation, information retrieval, and knowledge base population (Tran et al., 2023, Senger et al., 8 Oct 2025, Zhang et al., 2017).

1. Conceptual foundations

ATE generally targets both single-word terms (SWTs) and multiword terms (MWTs), and many domains also require explicit handling of nested terms. The survey literature uses examples such as “confiscation” and “confiscation of criminal assets” to illustrate SWTs and MWTs, while nested extraction is exemplified by spans such as those contained in “Mouse interleukin-2 receptor alpha gene expression,” where multiple overlapping gold spans encode different conceptual aspects (Tran et al., 2023, Gao et al., 2019). In specialized settings, term extraction is therefore not merely phrase spotting: it must decide whether a span denotes a concept, whether shorter subspans should be retained, and whether inflectional or orthographic variants should be normalized.

A recurrent distinction is between termhood and unithood. The Peacemaker system for Italian ATE defines ATE as identifying domain-specific terms while balancing linguistic unithood—whether tokens belong together as a unit—and termhood—whether a unit is specific and relevant to the domain—while coping with variant forms (Bakhtiyarzadeh et al., 31 May 2026). This distinction underlies both classical ranking methods and neural labeling models. In practice, ATE also has to determine whether adjectives, prepositional structures, abbreviations, and head-modifier combinations should be treated as canonical terms, and these decisions vary by domain and annotation protocol.

ATE is also distinct from both named entity recognition and generic keyword extraction. The domain-independent “wiki-worthy” term study explicitly contrasts ATE with NER and keyword extraction: unlike NER, ATE must capture conceptual, eventive, and compositional noun phrases, and unlike generic keyword extraction, it emphasizes termhood, unithood, and canonicalization (Bilu et al., 2020). Some benchmarks exclude named entities unless they function as scientific or technical terms, whereas others define alternate regimes. In ACTER, for example, ANN evaluates terms only and NES evaluates terms plus named entities, and the difference materially affects system behavior (Tran et al., 2022). The distant-supervision benchmark similarly states that named entities are generally excluded from the target concept set unless they function as scientific or technical terms, as in “heart failure” (Senger et al., 8 Oct 2025).

2. Methodological paradigms

ATE has historically been organized as a pipeline of corpus preparation, candidate generation, candidate scoring or ranking, filtering, and validation. The survey literature groups systems into linguistic or rule-based approaches, statistical or unsupervised approaches, hybrid approaches, traditional supervised machine learning, and neural approaches (Tran et al., 2023). Classical systems frequently rely on noun-phrase patterns, n-grams, dependency templates, and domain-versus-general distributional contrasts. Neural systems increasingly replace explicit candidate generation with sequence labeling or direct generation, but the candidate-generation-plus-ranking paradigm remains central in unsupervised and hybrid ATE.

Family	Characteristic mechanism	Representative systems
Linguistic/rule-based	POS patterns, NP chunking, dependency templates	TermoStat, TTC TermSuite, TExSIS
Statistical/unsupervised	Termhood and unithood ranking	RAKE, TerMine, TermRaider
Hybrid/re-ranking	Linguistic filters plus statistical or semantic re-ranking	RENT, SemRe-Rank, HAMLET
Supervised neural	BIO/BILOU token labeling, candidate classification, seq2seq	BiLSTM-CRF, BERT, RoBERTa, XLM-R, mBART
LLM-oriented	Few-shot prompting, pseudo-label distillation, direct term-list generation	ChatGPT-assisted extraction, DiSTER, syntactic retrieval

Classical formulas remain foundational. The survey lists TF–IDF, C-Value, NC-Value, Weirdness, PMI, Dice, and LLR as standard measures (Tran et al., 2023). In agriculture-oriented tool comparisons, RAKE scores a word by $\mathrm{deg}(w)/\mathrm{freq}(w)$ and a phrase by summing constituent word scores; TerMine uses C-value and extracts only multi-word candidates; TermRaider uses TF-IDF over noun phrases; and RENT combines POS information, regular-expression evidence, and frequency in a domain-specific hybrid weight (Chatterjee et al., 2020). These methods differ primarily in how they operationalize termhood and whether they privilege multiword cohesion, domain salience, or domain-specific seed patterns.

A notable line of work treats ATE enhancement as a separate problem from ATE itself. SemRe-Rank begins from the premise that “there is no existing ATE methods that can consistently outperform others in any domain” and therefore proposes a generic enhancement method rather than another one-size-fit-all extractor (Zhang et al., 2017). It uses word embeddings, document-level graphs of semantically related words, and personalised PageRank to compute word-level semantic importance, then revises the score of a candidate term $t$ as

$$srk(t)=\Big(1.0 + \frac{\sum_{w \in t} nsmi(w)}{|t|}\Big)\, nate(t).$$

This re-ranking strategy improved 13 base ATE methods on four datasets, with reported gains of up to 15 percentage points in average Precision@K and up to 28 percentage points in F1 measured at recoverable true positives (Zhang et al., 2017).

3. Supervised neural extraction and span modeling

The dominant supervised formulation in recent ATE is sequence labeling with BIO/IOB tags. ACTER-based Transformer systems, CoastTerm, and Peacemaker all use token-level labels that are later collapsed into term spans (Tran et al., 2022, Delaunay et al., 2024, Bakhtiyarzadeh et al., 31 May 2026). In CoastTerm, ATE is modeled as sequence labeling that assigns IOB tags to tokens and reconstructs multi-word term spans, while automatic term classification appends an ARDI-aligned role label to each span (Delaunay et al., 2024). Peacemaker similarly converts Italian gold terms to B-TERM, I-TERM, and O, uses the Italian cased tokenizer from dbmdz/bert-base-italian-cased, aligns character spans with tokenizer offsets, and trains with standard token-level cross-entropy (Bakhtiyarzadeh et al., 31 May 2026).

Transformer token classification has become the principal strong baseline. Cross-domain ATE on ACTER and RSDO5 uses fine-tuned token classifiers over monolingual and multilingual pretrained LLMs, followed by exact-match evaluation on aggregated term lists (Tran et al., 2022). In ACTER, monolingual models outperform multilingual ones in some settings and multilingual models outperform monolingual ones in others: for English NES, roberta-base reaches F1 59.14, while for Dutch ANN InfoXLM reaches F1 68.53 (Tran et al., 2022). On CoastTerm, RoBERTa_large reaches F1 81.92 for human-recommended ATE and the best joint extraction-plus-role-labeling score on Fusion is near 69.00 F1 with XLM-R_large, illustrating that span extraction and role classification are separable difficulty levels (Delaunay et al., 2024).

Not all neural ATE is token classification. The nested-term model for GENIA formulates ATE as end-to-end span classification and ranking. It enumerates all token spans up to a fixed maximum length $k$, represents each span by internal, head, boundary, sentence-attention, and length components, and explicitly allows overlapping and nested spans (Gao et al., 2019). The number of candidates is reduced from $\frac{n(n+1)}{2}$ to $n*k - \frac{k(k-1)}{2}$, and the final ranker stage reaches F1 $\approx 0.779$ on GENIA (Gao et al., 2019). This architecture is significant because it supports nested extraction without external linguistic analyzers.

A different supervised direction treats ATE as candidate classification rather than token labeling. The Slovenian contextual-embedding model first derives corpus-informed candidate-selection rules from the term-annotated RSDO5 corpus, then trains a linear SVM over linguistic, statistical, and contextual features derived from Slovene ELMo (Repar et al., 24 Feb 2025). Its results range from F1 0.530 on biomechanics to F1 0.594 on veterinary and yield absolute improvements over KAS-term of +0.191, +0.228, +0.183, and +0.295 across the four held-out domains (Repar et al., 24 Feb 2025). The ablation study shows that contextual signals are the principal driver: S-only and P-only configurations are near zero F1, whereas contextual-only remains usable though recall is lower (Repar et al., 24 Feb 2025).

4. LLMs, retrieval, and distant supervision

ATE with LLMs has developed along at least three lines: LLMs as annotators or candidate generators, LLMs as instruction-tuned student extractors, and LLMs as few-shot extractors guided by retrieval. The mathematical-concept study in category theory is representative of the first line. It uses ChatGPT to extract concepts from 755 abstracts, develops annotation guidelines, and evaluates agreement with human annotators using Jaccard rather than precision and recall (Paiva et al., 2023). In the pilot on 100 sentences, pairwise Jaccard among humans is 0.75–0.79, while ChatGPT versus humans is approximately 0.5; in a 436-sentence comparison, adjudication improves ChatGPT’s Jaccard from 0.531 to 0.631 (Paiva et al., 2023). The same study concludes that LLMs are effective candidate generators but “cannot at this time surpass human performance” in mathematical ATE (Paiva et al., 2023).

The second line is exemplified by DiSTER, which replaces gold supervision with pseudo-labels produced by GPT-4o and then fine-tunes open instruction-tuned LLMs to generate comma-separated term lists (Senger et al., 8 Oct 2025). DiSTER evaluates on seven domains—corruption, equitation, heart failure, wind energy, coastal science, biomedical, and computational linguistics—and reports both corpus-level and document-level evaluation. The Llama-3-8B-Instruct student reaches corpus-level macro F1 51.52 and document-level macro F1 51.34, while lightweight post-hoc document-level consistency (DC) and corpus-level consistency (CC) heuristics raise the mean document-level F1 from approximately 51.91 to 55.92 under DC+CC (Senger et al., 8 Oct 2025). The paper reports that the approach exceeds previous approaches on 5/7 domains with an average improvement of 10 percentage points (Senger et al., 8 Oct 2025).

The third line addresses the demonstration-selection problem in few-shot prompting. Syntactic retrieval for ATE uses unlexicalized constituency parse trees and the FastKASSIM syntactic similarity metric to select demonstrations by structural rather than semantic similarity (Chun et al., 26 Jun 2025). On cross-domain ACTER, Gemma-2-9B-IT with FastKASSIM reaches P=64.3, R=56.6, F1=60.2; on ACLR2 and BCGM, FastKASSIM often improves LLM performance, although tuned PLMs remain stronger in-domain, with RoBERTa-large at F1 87.4 on ACLR2 and F1 88.5 on BCGM (Chun et al., 26 Jun 2025). The lexical-overlap analysis further shows that in cross-domain settings FastKASSIM has TOR=0 and remains competitive, whereas semantic overlap can be negatively correlated with micro F1 (Chun et al., 26 Jun 2025). This suggests that, for ATE boundary control, syntactic alignment can be more stable than lexical similarity when domain overlap is low.

5. Corpora, annotation regimes, and evaluation

ATE research is heavily shaped by corpus design, annotation policy, and matching criteria. Some corpora annotate only spans, some annotate spans plus labels, some distinguish terms from named entities, and some record candidate validity rather than raw spans. The following resources illustrate the range of current practice.

Corpus	Scope and size	Annotation design
ACTER v1.5	EN: 248 docs, 222,281 tokens, 5,534 terms, 1,333 entities; FR: 309 docs, 240,207 tokens, 4,941 terms, 789 entities; NL: 259 docs, 218,989 tokens, 5,454 terms, 884 entities	ANN: terms only; NES: terms + named entities
GENIA 3.02	18,539 sentences, 490,766 words, 99,111 annotated terms	22,675 nested or overlapped terms
CoastTerm	2,491 sentences from 410 scientific abstracts	BIO spans plus Actor, Resource, Process, Quality, Location
RSDO5	12 full texts, about 250,000 words, 38,000 manually annotated term occurrences, 9,657 unique lemma terms	Slovene term-annotated corpus across four domains
TAC category-theory corpus	755 abstracts, approximately 3.2K sentences	Concept-list extraction with adjudication and Jaccard analysis
ATE-IT Task A	Train 2,308; Development/validation 577; Test 1,142 sentences	BIO sequence labeling for Italian waste-management text

Evaluation protocols are correspondingly heterogeneous. Many systems use exact-match string or span comparison, either at corpus level or document level (Tran et al., 2022, Senger et al., 8 Oct 2025). Peacemaker reports both type-level and micro-level precision, recall, and F1, distinguishing unique-term recovery from instance-level extraction (Bakhtiyarzadeh et al., 31 May 2026). CoastTerm compares the set of unique terms predicted on the entire test set against the gold set under strict exact-match criteria and reports both lemmatized and unlemmatized scores (Delaunay et al., 2024). The mathematical-concept study uses Jaccard similarity between sets of concepts instead of F1, reflecting its emphasis on inter-annotator agreement and normalization disputes (Paiva et al., 2023).

Annotation itself is often a major bottleneck. CoastTerm reports mean Krippendorff’s alpha of 43% during initial manual annotation, which the authors explicitly interpret as evidence of the difficulty of consistent term boundary and role labeling in a multidisciplinary setting (Delaunay et al., 2024). Mathematical ATE shows similar boundary and granularity instability, including disagreements about adjectives, sensible subspans, and prepositional spans (Paiva et al., 2023). These cases show that evaluation numbers are inseparable from annotation policy: exact-match scoring penalizes partial-span predictions, while different corpora encode different views of what counts as a term in the first place.

6. Applications, controversies, and research directions

ATE is an enabling layer for several downstream infrastructures. The coastal-science work positions role-aware extraction as “an initial step towards the development of a specialized Knowledge Base dedicated to coastal areas,” where extracted Actors, Resources, Processes, Qualities, and Locations support schema organization, downstream relation extraction, and ontology construction (Delaunay et al., 2024). The agriculture survey frames ATE as essential for populating AGROVOC and the NAL Thesaurus, as well as for ontology creation, recommendation systems, sentiment classification, and query expansion (Chatterjee et al., 2020). DiSTER explicitly identifies document tagging, ontology construction, patent analysis, translation, and IR as downstream settings where document-level term consistency matters (Senger et al., 8 Oct 2025).

Several controversies recur across the literature. First, there is no stable universal winner. SemRe-Rank states that there is no existing ATE method that can consistently outperform others in any domain, and the empirical record across ACTER, RSDO5, CoastTerm, mathematical ATE, and the multi-domain DiSTER benchmark is consistent with that assessment (Zhang et al., 2017). Second, multilingual pretraining is not uniformly superior. In ACTER, monolingual models are often stronger when named entities are included, while multilingual models can dominate in terms-only settings for French and Dutch; in CoastTerm, English-specialized RoBERTa_large is best for English-only ATE, whereas multilingual XLM-R is stronger for role classification (Tran et al., 2022, Delaunay et al., 2024). Third, LLMs do not eliminate the need for curation. ChatGPT remains below human agreement in mathematical ATE, and in-domain PLMs substantially outperform LLM few-shot systems on ACLR2 and BCGM (Paiva et al., 2023, Chun et al., 26 Jun 2025).

The main technical obstacles are also stable across domains. Multi-word boundaries remain difficult: CoastTerm reports under-prediction and fragmentation of longer terms, Peacemaker identifies partial MWEs and truncated spans under domain shift, and syntactic-retrieval experiments highlight over-inclusion and under-inclusion errors such as “blood pressure measurement” versus “blood pressure” (Delaunay et al., 2024, Bakhtiyarzadeh et al., 31 May 2026, Chun et al., 26 Jun 2025). Nested structures remain challenging even when they are explicitly modeled, and exact-match protocols magnify the cost of near misses (Gao et al., 2019). Domain shift is equally persistent: cross-domain IOB baselines are brittle in DiSTER, Italian legal-administrative phrasing shifts across splits in ATE-IT, and domain-general words such as “utente” or “system” can oscillate between term and non-term status depending on context (Senger et al., 8 Oct 2025, Bakhtiyarzadeh et al., 31 May 2026, Repar et al., 24 Feb 2025).

Current research directions therefore converge on richer supervision and richer structure. CoastTerm proposes moving from abstracts to full texts, expanding language coverage, refining hierarchical roles, and exploring CRF layers, span-based decoders, and generative models (Delaunay et al., 2024). The Slovenian contextual-embedding system proposes stronger contextual models such as BERT and extension to ACTER languages (Repar et al., 24 Feb 2025). DiSTER demonstrates that distant supervision with pseudo-labels can make open LLMs competitive across domains without human annotation, while syntactic retrieval shows that structure-sensitive demonstration selection can improve few-shot extraction (Senger et al., 8 Oct 2025, Chun et al., 26 Jun 2025). A plausible implication is that future ATE systems will increasingly combine domain-aware annotation design, exact structural boundary modeling, and downstream-oriented normalization rather than treating extraction as a stand-alone ranking problem.