{"id":"a262adfa-051b-4221-8338-512b1834b70d","arxiv_id":"2411.14463","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A systematic review and UMAP-based gap analysis showing that NLP in bank marketing is under-researched, with opportunities in customer acquisition, retention, and personalized engagement.","lead":"This paper systematically reviews research on AI and NLP in bank marketing and maps where the literature is thin. It finds that NLP is heavily studied in general marketing but rarely applied to bank marketing, and sketches where banks could use it next.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sparse NLP-bank-marketing finding may be an artifact of the Scopus-only, journal-only, ABDC-filtered search; Q3's count is not reported, so the paper does not yet establish that the field is truly underexplored.","rationale":"The paper is a competent systematic review with a plausible descriptive claim, and the thematic synthesis, tables, and PRISMA flowcharts show genuine effort. The reader's CONDITIONAL verdict is appropriate. My read converges on the same weakest assumption: the search regime determines the central finding. Section 2.1 restricts the review to Scopus, journal articles, and ABDC C-or-higher venues; Section 3.1 reports that the AllIntersect query 'yielded a limited number of studies' without a numeric count; and the limitations section explicitly concedes that conference papers and industry reports were excluded. Since the headline contribution is that the area is underexplored, the missing evidence is whether that finding survives a broader search. I do not see an internal inconsistency or evidence of fabrication; the issue is evidentiary completeness. If an expanded search confirms sparsity, the conclusion stands; if not, the gap map describes a corpus artifact rather than the research landscape. The proposed test is cheap and decisive, so the verdict should remain CONDITIONAL from the reader, hence UNCHANGED.","tokens_in":31189,"tokens_out":5115,"duration_ms":59805,"concrete_test":"Run the AllIntersect expression (MarketingQuery AND BankingQuery AND NLPQuery) in Scopus with no source-type or ABDC restriction, and in Web of Science and IEEE Xplore including conference proceedings; screen the added records using the same title/abstract criteria and report the unique relevant count. Also build a gold set of 20-30 known relevant papers from NLP/IS conferences and non-ABDC venues and measure how many the original strategy retrieves. If the expanded count or recall substantially exceeds the original, the 'limited research' conclusion is a search artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that research at the intersection of NLP and bank marketing is 'limited' rests on the AllIntersect query in Section 2.1, yet that query is never reported as an independent search with a numeric count: Section 3.1 says only that it 'yielded a limited number of studies.' More importantly, the search corpus is restricted to Scopus, peer-reviewed journals with ABDC rank C or higher, and excludes conference papers (Section 2.1). NLP and applied-AI results are heavily published at ACL/EMNLP, KDD, ICIS/ECIS/HICSS, and in venues not on the ABDC list; omitting them can turn a search-coverage gap into an apparent research gap. The paper's own limitation paragraph acknowledges that conference and industry sources are excluded, but it does not test whether including them changes the conclusion. As reported, the evidence supports 'limited research in this selected corpus,' not 'limited research in the field.' The underlying reality may still be sparse, but the current data do not demonstrate it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a PRISMA-based systematic review of two literatures—AI/analytics in bank marketing (Query 1) and NLP in general marketing (Query 2)—drawn from Scopus-indexed peer-reviewed journals ranked C or higher on the ABDC list, published 2014–2024. It then combines the included studies into a semantic map using Sentence Transformers and UMAP, manually labels clusters, and interprets sparse regions as gaps for NLP in bank marketing. On this basis it claims that research at the intersection of NLP and bank marketing is limited, and it proposes a customer-journey framework plus specific recommendations for NLP-driven engagement and operational excellence.","tokens_in":31360,"tokens_out":5395,"duration_ms":55045,"significance":"If the scarcity claim is correct, the paper fills a useful niche: it is one of the few systematic reviews to connect the NLP-in-marketing literature with the bank-marketing literature, and the semantic-mapping approach is a creative way to visualize a research landscape. The paper has genuine strengths: the PRISMA flowchart, the explicit Boolean queries in Table 1, the structured thematic tables (Tables 3–7), and a conceptual framework linking NLP to the marketing mix and customer journey. However, the central claim is not yet fully evidenced: the intersection query (AllIntersect) is not reported with numeric counts, and the search corpus excludes conference papers and non-ABDC venues where much NLP research appears. The gap analysis is also not fully reproducible as reported. These issues are load-bearing for the paper's main conclusions and need to be addressed before publication.","major_comments":[{"comment":"The central claim that research at the intersection of NLP and bank marketing is 'limited' is not quantitatively supported. Table 1 states that AllIntersect is 'derived from combining individual query outcomes, eliminating a separate search,' and Section 3.1 reports only that this query 'yielded a limited number of studies' without giving a number. Because the corpus is restricted to Scopus, peer-reviewed journals, and ABDC rank C or higher, and explicitly excludes conference papers, sparse results may reflect search coverage rather than a genuine research gap. Please report the full AllIntersect numbers (records identified, screened, eligible, included), and test robustness by repeating the search in at least one additional database or by including major NLP/AI conferences (ACL, EMNLP, KDD, ICIS/ECIS/HICSS). If the conclusion changes, the abstract and Section 3.1 should be softened to 'limited in this selected corpus.'","section":"§3.1 and Table 1"},{"comment":"The gap analysis relies on a UMAP projection and 'scatter variance' statistics, but the paper does not report the UMAP hyperparameters (n_neighbors, min_dist, metric), the random seed, or the definition of scatter variance used in Figure 6. UMAP is stochastic and hyperparameter-sensitive, so the sparse regions that motivate the recommendations may not be stable. The manual labeling of clusters is also described only as 'manual labeling' with no detail on procedure or inter-rater agreement. Please provide the hyperparameters and a definition of scatter variance, and ideally a robustness check (varying n_neighbors/min_dist or using a different embedding model) to show that the identified gaps persist.","section":"§4.1 and Figure 6"},{"comment":"The two query corpora are not mutually exclusive: for instance, De Caigny et al. (2020), Shumanov et al. (2022), and Afolabi et al. (2017) appear in both the MarketingBanking table (Table 4) and the NLP-marketing sector table (Table 7). If such studies are counted in both Query 1 and Query 2, the UMAP plot in Figure 5 may double-count studies, distorting cluster densities and the inferred gaps. Please report the degree of overlap between the two query result sets and clarify how overlapping studies were treated in the semantic map and in the thematic counts in Table 3.","section":"§3.3, §3.4, and Figure 5"}],"minor_comments":[{"comment":"The 'Additional Sources' step is under-specified: the flowcharts in Figure 2 add 3 and 2 studies respectively, but the text does not state how these were identified, whether they satisfied the same eligibility criteria, or whether they are included in the reported totals.","section":"§2.1.1"},{"comment":"The PRISMA flowchart covers only Queries 1 and 2; since AllIntersect is the basis for the paper's main conclusion, consider adding a third flowchart or a separate counts table for this query so readers can see the screening stages.","section":"§3.1 and Figure 2"},{"comment":"The row counts in Table 3 sum to 41 for MarketingBanking and 112 for MarketingNLP, exceeding the numbers of included studies (35+3 and 82+2). Please clarify that categories are not mutually exclusive and report the number of unique studies per dimension if relevant.","section":"Table 3"},{"comment":"The sentence transformer model is named as 'all-mpnet-base-v2' but no citation is given; please cite the Sentence-BERT source (Reimers and Gurevych, 2019) and the UMAP implementation (McInnes et al., 2018).","section":"§4.1"},{"comment":"There are minor typographical issues, including 'A S YSTEMATIC REVIEW' in the title, 'anomni-channel' in Section 3.4.3, and irregular spacing in name formatting such as 'V o' and 'Co¸ ser'; these should be cleaned up during copyediting.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"One included study (De Caigny et al., 2020) is co-authored by co-author Stefan Lessmann. The analysis does not appear to depend on this study, but given the overlap, a conflict-of-interest or author-contribution disclosure would be prudent. The manuscript's scope (an IS/marketing review) is reasonable, though it is submitted to cs.CL; the NLP content is mostly a survey of NLP applications, not an NLP methods contribution, which may matter for venue fit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent, well-scoped systematic review of AI/NLP in bank marketing, and the main claim—that the intersection is under-researched—is plausible but only demonstrated for a deliberately narrow slice of the literature. The paper does what prior reviews (Berger et al., Mustak et al., Kumar & Ravi) did not: it focuses specifically on banking, and it adds a UMAP-based gap map plus a customer-journey framework for where NLP could be applied. The two-query PRISMA design (MarketingBanking and MarketingNLP) is sound, the thematic classification into CRM, marketing mix, and strategic insights is useful, and the recommendations are concrete and tied to the reviewed studies. The authors also state their main limitation openly: conference papers and industry reports are excluded.\n\nThe soft spots are in the gap analysis and the central scarcity claim. AllIntersect—the query that establishes that NLP+bank+marketing is sparse—is never given a numeric count; the paper just says it \"yielded a limited number of studies.\" That makes the headline finding hard to evaluate. The search corpus is Scopus-only, journal-only, ABDC C-or-higher, with no conferences. Applied NLP is heavily published at ACL/EMNLP, KDD, ICIS/ECIS/HICSS, and in non-ABDC venues; leaving those out can convert a search-coverage gap into an apparent research gap. The paper's own limitation paragraph acknowledges this, but it never tests whether including those sources would change the conclusion. As reported, the evidence supports \"limited research in this selected corpus,\" not \"limited research in the field.\"\n\nThe UMAP-based gap map is also less reproducible than it should be. The authors do not provide the UMAP hyperparameters, the code, or the cluster labeling procedure; they report scatter variance figures, but without parameters those numbers are not checkable. The map is suggestive, not a measurement. This is a moderate issue, not a fatal one—the visual is an aid to the review, not the whole contribution.\n\nNo red flags: the literature table looks consistent with the search criteria, and the one included study co-authored by Lessmann (De Caigny et al. 2020) does not affect the central claim.\n\nMy recommendation: send it to a serious referee. It is a useful synthesis for anyone scoping research in bank marketing and NLP, and the weaknesses are addressable: put a number to AllIntersect, broaden or justify the corpus, and release the mapping code and screening data. I would cite it if I needed a landscape map, but not as a cornerstone reference. Reading group: maybe, mostly to discuss how search design can create apparent gaps.","headline":"A solid, well-scoped review whose central scarcity claim is only proven for a narrow journal-only corpus; the gap map is suggestive but not reproducible as reported.","tokens_in":31912,"tokens_out":2745,"would_cite":true,"duration_ms":76997,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review claims that NLP applications in bank marketing are scarce, and it maps where text analytics could create the most value for banks.","keywords":["bank marketing","natural language processing","systematic review","gap analysis","PRISMA","semantic mapping","UMAP","customer journey"],"falsifier":"Run the same three Boolean queries against a second database such as Web of Science, include conference proceedings, working papers, and industry reports, and extend the window to 2026. If the triple-intersection query returns a substantial body of NLP-in-bank-marketing studies with practical applications, the paper's 'limited research' claim fails; if the added venues still yield only a handful of studies, the gap claim survives.","tokens_in":30948,"feed_emoji":"🏦","tokens_out":7890,"duration_ms":65209,"temperature":0.7,"pith_summary":"This paper is a systematic review and gap analysis of artificial intelligence and natural language processing (NLP) in bank marketing. Using the PRISMA methodology, it reviews two bodies of work separately—analytical marketing in banking and NLP in general marketing—and then examines their intersection. The paper's core finding is that research specifically connecting marketing, banking, and NLP is limited, and it uses semantic mapping to visualize where the gaps lie. The authors embed abstracts with a sentence transformer and project them with UMAP, then label the clusters to reveal underexplored areas such as customer acquisition, retention, personalized engagement, and pricing. The intended contribution is a roadmap for academics and banking practitioners to target future NLP work in bank marketing.","feed_headline":"Bank marketing largely ignores NLP; map shows where it could help","feed_subtitle":"Review and semantic map pinpoint NLP opportunities in acquisition, retention, and pricing.","key_machinery":"Three pieces of machinery carry the argument. The first is the PRISMA search design, which runs two parallel queries—MarketingBanking and MarketingNLP—and derives a third intersection, AllIntersect, from their combination; the scarcity of the third query is the paper's core evidence. The second is semantic mapping: abstracts are embedded with the Sentence Transformer all-mpnet-base-v2 and projected into two dimensions with UMAP, producing a landscape where dense clusters indicate well-studied topics and sparse regions indicate gaps, interpreted along the axes Customer Focus vs. Strategic Focus and Internal vs. External Focus. The third is a conceptual framework that merges Kannan and Li's digital-marketing value creation with a customer-journey adaptation, which turns empty regions on the map into concrete recommendations for NLP deployment. The named technique that carries the gap claim is the UMAP projection combined with manual cluster labeling; without it, the paper would only have a bibliometric count rather than a spatial argument about where research is missing.","core_discovery":"The paper's central claim is that NLP is underused in bank marketing even though it is widely applied in general marketing and elsewhere in banking. The PRISMA review's third query, which searches for studies combining marketing, banking, and NLP, 'yielded a limited number of studies,' and the authors treat this scarcity as the motivating fact. The gap analysis positions the existing literature in a two-dimensional semantic space whose axes are Customer Focus versus Strategic Focus and Internal Focus versus External Focus. Sparse regions around customer-centric applications—acquisition, retention, personalization, pricing—are interpreted as genuine research and practice gaps. The paper also contributes a conceptual framework connecting NLP to the banking customer journey (awareness, consideration, conversion, retention) and to value creation for customers, the bank, and customer equity, with regulatory compliance as a precondition.","pith_inferences":["The authors do not claim, but their search design implies a testable hypothesis: the scarcity at the triple intersection may be partly a product of excluding conference papers and industry reports; a broader search could reveal more practice than the journal-only pool captures.","The gap map's axes suggest a natural research agenda: the largest empty regions are customer-facing applications, so a next study could pick one, such as NLP for pricing perception, and benchmark it against the e-commerce results the review summarizes.","Because the review window starts with Word2Vec (2014), transformer-era tools such as large language models are only just entering the captured literature; if the current trajectory of LLM-based marketing continues, the 'limited research' claim may age quickly and the map should be re-run on a 2024-2026 corpus.","The paper's four-stage customer journey is borrowed from general marketing; an untested extension would be a bank-specific lifecycle that includes onboarding and cross-selling within the retention stage, where the map already shows dense existing work."],"forward_implications":["Banks can improve customer acquisition by using sentiment analysis and topic modeling on social media data to identify prospects in real time, a direct recommendation from the gap analysis.","Adding textual data such as call logs, chat transcripts, and customer messages to churn models can improve retention predictions beyond structured data alone, as the aggregated churn studies show.","NLP analysis of customer feedback on pricing can inform dynamic and personalized pricing strategies, an application the map shows as a sparse region.","The proposed framework maps specific NLP tasks—sentiment analysis, topic modeling, text generation, chatbots—onto each stage of the banking customer journey, giving practitioners a checklist for where text analytics fits.","The value-creation framework ties NLP to customer equity, bank revenue, and customer value, with regulatory compliance and ethics as preconditions that any implementation must satisfy."],"supporting_citations":[{"why":"Supplies the PRISMA methodology that structures the systematic review and the search flow.","marker":"[Moher et al., 2009]"},{"why":"Provides the digital-marketing framework, marketing mix, and customer-journey/value-creation scaffolding used in the gap analysis.","marker":"[Kannan and Li, 2017]"},{"why":"Supplies the 'textual universe' framework and the view of text as a marketing data source that organizes the NLP-in-marketing review.","marker":"[Berger et al., 2020]"},{"why":"Provides the finance-domain NLP survey that the paper contrasts with and draws on for the 'excellence' dimension.","marker":"[Kumar and Ravi, 2016]"},{"why":"Grounds the claim that AI's potential for unstructured-data marketing insight remains underexplored.","marker":"[Wedel and Kannan, 2016]"},{"why":"Word2Vec sets the 2014 start date of the review window, bounding the literature considered.","marker":"[Mikolov et al., 2013]"}],"fun_headline_variants":["NLP barely used in bank marketing; map reveals untapped niches","Bank marketing: NLP gap mapped to acquisition, retention, pricing","NLP in bank marketing: a systematic review finds big gaps","Mapping NLP's missing role in bank marketing strategies","Bank marketing skips NLP; semantic map shows where to apply"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The review assumes that searching only Scopus, only peer-reviewed journals ranked C or higher on the ABDC list, and only the years 2014-2024 captures the full research landscape, so the sparse regions on the map are real research gaps rather than gaps left by the search itself.","fun_headline_variants_meta":{"raw":{"variants":["NLP barely used in bank marketing; map reveals untapped niches","Bank marketing: NLP gap mapped to acquisition, retention, pricing","NLP in bank marketing: a systematic review finds big gaps","Mapping NLP's missing role in bank marketing strategies","Bank marketing skips NLP; semantic map shows where to apply"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000529,"raw_usage":{"total_tokens":2555,"prompt_tokens":958,"completion_tokens":1597,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":1513}},"tokens_in":574,"tokens_out":1597,"duration_ms":10108,"temperature":1.0,"reasoning_tokens":1513,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:54:28.185784+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same three Boolean queries against a second database such as Web of Science, include conference proceedings, working papers, and industry reports, and extend the window to 2026. If the triple-intersection query returns a substantial body of NLP-in-bank-marketing studies with practical applications, the paper's 'limited research' claim fails; if the added venues still yield only a handful of studies, the gap claim survives.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Grounds the claim that AI's potential for unstructured-data marketing insight remains underexplored."}],"review_version":1}