“Will It Ever Get Better?”: Understanding Psychological Distress Among Indian Students Through Reddit

An NLP pipeline for characterising how psychological distress is expressed — and how it changes — across the Indian student lifecycle, from school and competitive-exam preparation through college, career, and higher education. It models 70,425 code-mixed (Hinglish) Reddit posts and 164,367 linked comments from 11 student communities using multilingual Transformer embeddings and BERTopic.

Paper: submitted to the IEEE International Conference on Big Data (IEEE BigData 2026), Phoenix, Arizona, USA — under review. Research carried out with the Data Science for Healthcare Lab, Wright State University.

---

Highlights

• Five embedding models benchmarked (MuRIL, MentalBERT, HingBERT-Mixed-v2, HingRoBERTa-Mixed, BGE-m3) on topic coherence, diversity, stability, and retrieval. BGE-m3 was selected — best coherence (c_v 0.566) and retrieval (MAP 0.933, MRR 1.000, nDCG@10 0.437).
• Topic-model refinement: a 30-configuration search turned an over-fragmented 98-topic model into a stable 41-topic model (c_v 0.652, c_npmi 0.142, diversity 0.720, 0.82% outliers, seed-to-seed ARI 0.97).
• Taxonomy: 41 topics → 8 higher-level themes → 6 distress categories. Distress is mostly anticipatory — Academic (34.4%) and Career (26.8%) — rather than clinical (Emotional 6.9%).
• Lifecycle shift: academic distress dominates at the school stage (65.6%), while career distress dominates at the career stage (92.0%); code-mixed coping/meme expression peaks during competitive-exam preparation (21.3%).
• Community response (comments): explicit emotional support is scarce, and on suicidal-ideation posts crisis-escalating replies receive more up-votes than supportive ones (3.53 vs 2.79, one-sided Mann–Whitney p = 0.0015).

---

Pipeline

flowchart LR
    A["Reddit collection<br/>207,546 posts"] --> B["Two-stage distress filtering<br/>lexicon + classifier"]
    B --> C["De-duplication<br/>70,425 posts"]
    C --> D["Preprocessing<br/>artifact removal · Hinglish<br/>normalisation · phrase preservation"]
    D --> E["Embeddings<br/>5 Transformer models<br/>256 tokens · L2-normalised"]
    E --> F["Benchmark<br/>coherence · diversity<br/>stability · retrieval"]
    F -->|BGE-m3| G["BERTopic<br/>UMAP → HDBSCAN<br/>→ c-TF-IDF + KeyBERT + MMR"]
    G --> H["Refinement<br/>30-config search<br/>98 → 41 topics"]
    H --> I["8 themes →<br/>6 distress categories"]
    I --> J["Lifecycle analysis"]
    K["164,367 comments"] --> L["Same preprocessing<br/>+ embeddings"]
    L --> M["Interaction taxonomy +<br/>cross-corpus analysis"]
    I --> M
    N["Human annotation<br/>(Cohen's κ)"] -.-> I
    N -.-> M

Loading

---

System design

Resumable chunked inference

scripts/posts/generate_posts_embeddings.py · scripts/comments/generate_comments_embeddings.py

flowchart TD
    A["Load corpus<br/>70,425 preprocessed texts"] --> B["Select device<br/>MPS → CUDA → CPU fallback"]
    B --> V["Probe model with one input<br/>verify embedding dimension"]
    V --> C{"Next 10K chunk"}

    subgraph W["Per-chunk worker"]
        direction TB
        D{"Checkpoint file<br/>already on disk?"}
        E["Sort chunk by token length<br/>minimises padding waste"]
        F["Batched inference<br/>batch 64 · 256 tokens · CLS vector"]
        G["Scatter vectors back to<br/>original document order"]
        P["Save chunk checkpoint"]
        D -->|"miss — compute"| E
        E --> F
        F --> G
        G --> P
    end

    C -->|"chunk pending"| D
    D -->|"hit — skip, already done"| C
    P -->|"chunk complete"| C

    C -->|"all chunks issued"| M{"Every chunk<br/>present on disk?"}
    M -->|"no — missing chunk"| X["Abort, non-zero exit"]
    M -->|"yes"| H["Assemble chunks<br/>into one matrix"]
    H --> S["Save embeddings .npy"]
    S --> R["Release device memory"]

Loading

• A killed run resumes at chunk granularity — completed chunks are skipped, nothing is recomputed.
• Documents are reordered by length for throughput, then scattered back to their original indices, so output row order still matches the input corpus.
• Assembly refuses to write a partial matrix: a missing chunk aborts the run with a non-zero exit code.
• Vectors are written unnormalised; consumers L2-normalise on load.

Cached configuration sweep

src/pipeline/bertopic_search.py

sequenceDiagram
    participant S as Sweep driver
    participant J as search_results.jsonl
    participant C as UMAP cache on disk
    participant U as UMAP fit — expensive
    participant H as HDBSCAN — cheap

    S->>J: read completed config_ids
    J-->>S: resume set

    loop each UMAP config
        S->>C: reduction for config + seed 42
        alt cache hit
            C-->>S: load reduced matrix
        else cache miss
            S->>U: fit_transform L2-normalised embeddings
            U-->>S: reduced matrix
            S->>C: persist reduction
        end

        loop each HDBSCAN config
            alt already in resume set
                S->>S: skip, already scored
            else not yet scored
                S->>H: cluster seed-42 reduction
                H-->>S: labels
                alt fewer than 6 topics or over 70% outliers
                    S->>J: append invalid record + reason
                else passes sanity gate
                    S->>S: c-TF-IDF keywords, coherence, diversity, separation
                    opt second seed not yet reduced
                        S->>C: reduction for config + seed 7
                        C-->>S: reduced matrix
                    end
                    S->>H: cluster seed-7 reduction
                    H-->>S: labels
                    S->>S: stability = ARI over both seeds
                    S->>J: append metrics record
                end
            end
        end
    end

    S->>J: read all records
    J-->>S: rank configs and write report

Loading

• The sweep is itself resumable: completed config_ids are read back from the append-only JSONL log and skipped, so an interrupted search continues where it stopped.
• The cache key is the UMAP configuration plus the seed, so one expensive reduction serves every clustering configuration that shares it — this is what makes a 30-configuration search affordable. The second seed is reduced lazily, only once a configuration actually reaches the stability check.
• Collapsed configurations (fewer than 6 topics, or over 70% outliers) are rejected before any scoring runs.
• Surviving configurations are scored on count-independent quality (coherence, stability, keyword distinctiveness, near-duplicate pairs), because centroid-separation metrics were confounded with topic count (r = −0.88).
• Every stage writes frozen intermediate artifacts, so any stage can be re-run in isolation.

---

Results

Embedding benchmark (70,425 posts)

Model	c_v	c_npmi	Diversity	P@10	MAP	MRR	nDCG@10
BGE-m3	0.566	0.099	0.487	0.844	0.933	1.000	0.437
HingRoBERTa-Mixed	0.535	0.079	0.444	0.844	0.928	0.970	0.342
MuRIL	0.524	0.080	0.413	0.612	0.732	0.791	0.218
MentalBERT	0.498	0.050	0.365	0.444	0.543	0.601	0.149
HingBERT-Mixed-v2	0.487	0.044	0.334	0.544	0.721	0.816	0.189

On the comment corpus the ranking changes: MuRIL has the best weighted score (0.742; best coherence and stability), while BGE-m3 keeps the best retrieval and diversity — the best embedding model depends on the corpus.

Topic-model refinement (BGE-m3)

Metric	98-topic baseline	Final 41-topic model
c_v coherence	0.626	0.652
c_npmi	0.136	0.142
Topic diversity	0.641	0.720
Near-duplicate topic pairs (cos > 0.90)	280	128
Tiny topics (< 200 posts)	4	0
Smallest topic	149	498
Outlier rate	0.03%	0.82%

Final configuration: UMAP (n_neighbors=50, n_components=10, min_dist=0.0, cosine) · HDBSCAN (min_cluster_size=300, min_samples=10, leaf) · c-TF-IDF + KeyBERTInspired + MMR (0.3) · c-TF-IDF outlier reduction (threshold 0.05). Two-seed stability: ARI 0.97.

Themes and distress taxonomy

Theme	Posts	%		Distress category	Posts	%
Entrance Exam Preparation	19,136	27.4		Academic	24,045	34.4
Careers, Placements & Jobs	11,277	16.2		Career	18,722	26.8
College & Academic Decisions	10,376	14.9		Informational / Low-Distress	10,689	15.3
Coaching & Study Resources	7,937	11.4		Coping Expression	8,850	12.7
Higher Studies & Global Mobility	7,445	10.7		Emotional	4,824	6.9
Mental Health & Wellbeing	4,824	6.9		Institutional	2,715	3.9
General Venting	4,770	6.8
Coping & Meme Culture	4,080	5.8

Distress across the lifecycle

Stage	Dominant distress
School	Academic 65.6%
Competitive examination	Academic 45.0% · Coping/meme 21.3%
Undergraduate	Informational 40.9% · Career 29.1%
Career	Career 92.0%
Higher education	Career 88.6%

Community response (138,989 cleaned comments)

• Replies are mostly neutral discussion (43.7%), advice (17.5%) and humour/meme-coping (17.3%); the emotional-support-to-harm ratio is 0.67 : 1.
• Suicidal-ideation posts attract the most engagement (22.8 comments per post), yet 50.1% of them receive no explicitly supportive reply and crisis-escalating replies are 3.83× more frequent than the corpus baseline.
• Parent-post distress and response type are associated (χ² = 9,281, df = 66, p < 0.001; Cramér's V = 0.105).

Validation

Task	Samples	Agreement (95% bootstrap CI)	Cohen's κ
Post distress labels	300	70.0% (0.643–0.747)	≈ 0.00
Comment interaction labels	300	17.7% (0.133–0.217)	0.10

The distress sample is dominated by one class (210/300), so raw agreement is high but κ is near zero; the interaction labels come from a keyword-based classifier. Both are therefore reported as descriptive labels rather than validated classifiers.

---

Repository structure

.
├── README.md
├── requirements.txt
├── paper/                    # manuscript (added after review)
├── data/                     # NOT tracked — see data/README.md for the expected layout
├── src/
│   ├── preprocessing/        # artifact removal, Hinglish normalisation, phrase preservation
│   ├── pipeline/             # BERTopic search / finalisation, paper figures & tables
│   └── analysis/             # comment-corpus BERTopic, coherence, retrieval, stability, taxonomy
├── scripts/
│   ├── posts/                # post embeddings, benchmark, retrieval, benchmark figures
│   ├── comments/             # comment preprocessing, embeddings, analysis, statistics
│   ├── annotation/           # annotation sampling and agreement
│   └── figures/              # revised paper figures
└── final_outputs/            # metrics, reports, tables and figures (see final_outputs/README.md)

---

Setup

git clone <repo-url>
cd <repo-folder>
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Tested with Python 3.13 on Apple Silicon (MPS); CUDA and CPU are selected automatically. Models are downloaded from Hugging Face on first use into model_cache/.

Reproducing the results

All commands run from the repository root and expect the data layout in data/README.md. Embedding generation takes several hours per corpus; the runners checkpoint and can simply be re-run after an interruption.

Posts

bash scripts/posts/run_all_posts_embeddings_256.sh            # 5 models, 256 tokens

for m in MuRIL:muril MentalBERT:mentalbert HingBERT-Mixed-v2:hingbert_mixed_v2 \
         HingRoBERTa-Mixed:hingroberta_mixed BGE-m3:bge_m3; do
  python scripts/posts/run_posts_bertopic.py --model-name "${m%%:*}" \
         --embedding-path "data/embeddings/posts/${m##*:}_embeddings.npy"
  python scripts/posts/run_posts_coherence.py --model "${m%%:*}"
done
python scripts/posts/run_posts_retrieval.py                    # 25-query retrieval benchmark
python scripts/posts/generate_posts_figures.py

python src/pipeline/bertopic_search.py                         # 30-config search (BGE-m3)
python src/pipeline/bertopic_finalize.py                       # final 41-topic model
python src/pipeline/paper_assets.py                            # figures, taxonomy, metadata
python src/pipeline/paper_extras.py
python src/pipeline/paper_tables.py

src/pipeline/bertopic_refinement.py is an optional stop-word / artifact audit (--phase audit) used while refining the topic model.

Comments

python scripts/comments/preprocess_comments.py                 # 164,367 -> 138,989 comments
bash scripts/comments/run_comments_embeddings_256.sh           # 5 models, 256 tokens
bash scripts/comments/run_comments_analysis.sh                 # BERTopic, coherence, retrieval, stability, aggregate
python src/analysis/comments_taxonomy_v2.py                    # interaction taxonomy
python scripts/comments/cross_corpus_analysis.py               # response vs. parent-post distress
python scripts/comments/suicidal_interaction_stats.py          # chi-square, Cramér's V, Mann-Whitney

Figures and validation

python scripts/figures/make_revised_figures.py
python scripts/figures/make_theme_fig.py
python scripts/figures/make_cross_corpus_fig.py
python scripts/annotation/annotation_agreement.py

---

Data and ethics

The corpus consists of public Reddit content that includes usernames and sensitive first-person disclosures. To protect the people who wrote it, no raw text, post/comment IDs, or per-document labels are published — only aggregate results. The work is a descriptive, cross-sectional study; it is not a diagnostic tool and does not attempt to identify individuals.

Limitations

• Reddit users are not representative of all Indian students.
• Mostly English and Hinglish; other Indian languages are not modelled.
• Cross-sectional: no causal or longitudinal claims.
• Human validation is limited (see Validation); topic, theme and retrieval validation were not human-annotated.

Authors

Author	Affiliation
Chinmaya Krishna B	School of Computer Engineering, Manipal Institute of Technology, Manipal, India
Animesh Kumar	School of Electrical Engineering, Manipal Institute of Technology, Manipal, India
Piyush Kumar	School of Basic Sciences, Humanities and Management, Manipal Institute of Technology, Manipal, India
Misty M	School of Electrical Engineering, Manipal Institute of Technology, Manipal, India
Tanvi Banerjee	Data Science for Healthcare Lab, Wright State University, Dayton, USA
William Romine	Data Science for Healthcare Lab, Wright State University, Dayton, USA
Ramakrishna Mundugar	School of Computer Engineering, Manipal Institute of Technology, Manipal, India

Citation

The paper is under review; please cite it as below until a published version is available.

@unpublished{krishnab2026willitevergetbetter,
  title  = {``Will It Ever Get Better?'': Understanding Psychological Distress Among Indian Students Through Reddit},
  author = {{Chinmaya Krishna B} and {Animesh Kumar} and {Piyush Kumar} and {Misty M} and
            {Tanvi Banerjee} and {William Romine} and {Ramakrishna Mundugar}},
  note   = {Submitted to the IEEE International Conference on Big Data (IEEE BigData 2026)},
  year   = {2026}
}