CSConv: Cognitive Stimulation Dialogue Dataset

• CSConv is a Chinese dialogue dataset designed for cognitive stimulation in elderly care, integrating 18 CS principles with emotional support strategies.
• The framework employs progressive-mask encoders and utterance-level classification to predict CS principles, emotion, and support strategy for controlled response generation.
• Empirical results show that CSD outperforms baselines in BLEU and BERTScore metrics, demonstrating its effectiveness in generating empathetic and coherent dialogues.

Cognitive Stimulation CSConv denotes a Chinese CS conversation dataset and the associated CSD framework introduced for dialogue with elders with cognitive impairment. The work addresses data sparsity in CS-based dialogue systems, particularly in the Chinese language, by constructing CSConv and by proposing a multi-source knowledge fusion method for CS dialogue that generates open-ended responses guided by the CS principle and emotional support strategy (Jiang et al., 2023). In the reported formulation, cognitive stimulation is coupled with emotional support rather than treated as a purely task-oriented prompting problem, and the system architecture explicitly predicts CS principle, emotion, and support strategy before response generation (Jiang et al., 2023).

1. Dataset origin and annotation protocol

CSConv was constructed from approximately 2,600 one-to-one “elder ↔ therapist” dialogue groups (Jiang et al., 2023). The sources were Cognitive Stimulation Therapy videos in Cantonese translated to Mandarin, with a $2:1$ ratio, and Mandarin dialogues manually authored from the CST handbook, which contributed 18 principles. The corpus was cleaned by truncating overly long turns and removing repeated punctuation, and it was formatted with manual speaker tags distinguishing SPEAKER, representing the elder, from LISTENER, representing the therapist or smart-speaker.

Annotation proceeded by automated initialization of three labels per utterance via BERT classifiers, followed by two rounds of manual review (Jiang et al., 2023). The three label families were CS principle, emotion, and support strategy. This combination reflects the paper’s premise that “making chit chat while providing emotional support is overlooked by the majority of existing cognitive dialogue systems,” and the annotation scheme operationalizes that premise at utterance level.

A plausible implication is that the dataset was designed not only for response generation but also for multi-task supervision, since the same utterances support classification of CS principle, emotion, and support strategy as prerequisites for controlled generation.

2. Label ontology and corpus statistics

The CS principle labels comprise seven types: None, Inquiry, Respect, Reminiscence, Expression, Enjoyment, and Comfort (Jiang et al., 2023). The emotion labels comprise eight types: none, disgust, sadness, fear, surprise, like, happiness, and anger. The support strategy labels also comprise seven types: None, Question, Reflection of Feelings, Self-disclosure, Providing Suggestions, Information, and Others.

The corpus statistics reported for CSConv are as follows:

Statistic	Value
Conversations	2,643
Total utterances	16,845
SPEAKER turns	8,363
LISTENER turns	8,480
Avg. tokens per conv.	60.39
Avg. utterances per conv.	6.37
Avg. tokens per utt.	9.48

These statistics indicate a relatively short-turn conversational regime, with 6.37 utterances per conversation and 9.48 tokens per utterance (Jiang et al., 2023). This suggests that the dataset emphasizes compact, locally coherent exchanges rather than long-horizon dialogue management. Because the labels are attached at utterance level and the conversations are short, the benchmark is especially suited to studying fine-grained conditioning of response style and content under limited context.

3. Progressive-mask encoders and utterance-level classification

The encoder component begins from a standard BERT encoder, specifically bert-base-chinese, and is enriched via two external-knowledge dictionaries: an emotion dictionary from Jiagu sentiment extraction and a keyword dictionary from Jiagu keyword extraction (Jiang et al., 2023). During pre-training, the model randomly masks spans of size 1–4 tokens or full sentences, mixing ratios $\lambda_i$ for 1–4-token masks, a sentence-level mask ratio $\lambda_5$, and 5% standard BERT-style [MASK]/random/unchanged tokens.

At layer $l$, the encoder is defined by the standard residual-attention and residual-FFN updates

$$\begin{aligned} h_i^{(l)} &= \text{LN}\bigl(x_i^{(l-1)} + \mathrm{MHAtt}(x_i^{(l-1)})\bigr)\ x_i^{(l)} &= \text{LN}\bigl(h_i^{(l)} + \mathrm{FFN}(h_i^{(l)})\bigr)\,, \end{aligned}$$

where $\mathrm{MHAtt}$ is multi-head self-attention and $\mathrm{FFN}$ is a two-layer feed-forward network (Jiang et al., 2023). After $L$ layers, the contextualized token embeddings are denoted $\mathbf{C_u}=[\tilde x_0,\ldots,\tilde x_m]$.

The model flattens an $M$-turn context $\lambda_i$0 and feeds it into three parallel classifiers for CS principle, emotion, and strategy (Jiang et al., 2023). For CS principle, the paper gives

$\lambda_i$1

with analogous heads for the 8-way emotion task and the 7-way strategy task. Each classifier is trained with cross-entropy loss; for example,

$\lambda_i$2

The role of these encoders is not merely representational. The paper states that learning encoders as effective classifiers is the prerequisite to predicting the CS principle and emotional support strategy of the target response (Jiang et al., 2023). This places the classification subsystem upstream of generation and makes the predicted labels an explicit control signal.

4. Decoder conditioning and multi-source knowledge fusion

At generation time, the decoder appends three special tokens corresponding to the predicted emotion, CS, and strategy labels, $\lambda_i$3, to the end of the decoder input sequence: $\lambda_i$4 Each token is embedded through word and positional embeddings $\lambda_i$5 and $\lambda_i$6 (Jiang et al., 2023).

An extra encoder processes $\lambda_i$7 to yield hidden states $\lambda_i$8, after which the hidden states at the three label positions are replaced with the corresponding outputs from the three pre-trained classification encoders (Jiang et al., 2023). In the decoder’s cross-attention, these states attend to the conversation context. At decoder layer $\lambda_i$9, the paper writes

$\lambda_5$0

with the label positions swapped as described above.

The model further introduces two auxiliary MSE losses to bias generation toward high-intensity emotion words and keywords (Jiang et al., 2023). The emotion-intensity is derived from Chinese EmoBank, the keyword-intensity from Jiagu scores, and both are matched against self-attention weights $\lambda_5$1 in decoder cross-attention: $\lambda_5$2

Response generation itself is formulated autoregressively: $\lambda_5$3 with the negative log-likelihood

$\lambda_5$4

This architecture is termed a multi-source knowledge fusion method because it fuses conversation context, predicted task labels, and external lexical knowledge sources within a single conditional generation pipeline (Jiang et al., 2023). The design suggests a hybrid control mechanism in which symbolic label information and distributional attention cues jointly shape decoding.

5. Training objective and optimization configuration

The full training objective for generation is

$\lambda_5$5

with $\lambda_5$6 and $\lambda_5$7 (Jiang et al., 2023). The classification losses $\lambda_5$8, $\lambda_5$9, and $l$0 are trained separately in order to pre-train the encoders.

The reported optimization details are specific. Encoder pre-training and classification use bert-base-chinese with progressive mask; decoder initialization uses gpt2-chinese-cluecorpussmall; the optimizer is AdamW with the learning-rate schedule of Vaswani et al. (2017); the batch size is 4; and training is conducted on 4 × A100 GPUs (Jiang et al., 2023). The reported training times are approximately 2 minutes for the encoder and approximately 33 minutes for the decoder. Inference uses temperature $l$1, top-$l$2, and top-$l$3.

These details situate CSD as a comparatively lightweight experimental configuration in hardware time while still employing a staged training regime. A plausible implication is that most of the methodological novelty lies in data construction, controlled conditioning, and attention shaping rather than in scale.

6. Empirical performance, ablations, and interpretation

On CSConv classification, CSD achieves the strongest reported accuracies among the listed models (Jiang et al., 2023).

Model	CS Acc.%	Emo Acc.%	Strat. Acc.%
Transformer	83.67	85.10	91.63
BERT	85.71	87.76	94.49
BERT+CNN	84.90	87.35	94.29
CSD (ours)	87.14	88.37	94.69

For response generation, the reported automatic and human-evaluation results are likewise highest for CSD among the listed baselines (Jiang et al., 2023).

Model	BLEU$l$4	BLEU$l$5	BERTScore
CDialGPT$l$6	17.55	6.22	57.70
GPT2-chitchat	34.61	21.04	66.37
Cluecorpussmall	41.04	26.59	68.65
CSD (ours)	45.53	30.90	74.61

The full generation table additionally reports Dist1, Dist2, Empathy, Support, and Fluency. CSD records Dist1 $l$7, Dist2 $l$8, Empathy $l$9, Support $$\begin{aligned} h_i^{(l)} &= \text{LN}\bigl(x_i^{(l-1)} + \mathrm{MHAtt}(x_i^{(l-1)})\bigr)\ x_i^{(l)} &= \text{LN}\bigl(h_i^{(l)} + \mathrm{FFN}(h_i^{(l)})\bigr)\,, \end{aligned}$$0, and Fluency $$\begin{aligned} h_i^{(l)} &= \text{LN}\bigl(x_i^{(l-1)} + \mathrm{MHAtt}(x_i^{(l-1)})\bigr)\ x_i^{(l)} &= \text{LN}\bigl(h_i^{(l)} + \mathrm{FFN}(h_i^{(l)})\bigr)\,, \end{aligned}$$1, and the paper states that it outperforms all baselines on both automatic and human metrics (Jiang et al., 2023). In direct pairwise comparison, CSD is preferred over the strongest baselines in 41–69% of cases.

The ablation study isolates the contribution of progressive mask, label injection, cross-attention, and attention losses (Jiang et al., 2023).

Model	BLEU$$\begin{aligned} h_i^{(l)} &= \text{LN}\bigl(x_i^{(l-1)} + \mathrm{MHAtt}(x_i^{(l-1)})\bigr)\ x_i^{(l)} &= \text{LN}\bigl(h_i^{(l)} + \mathrm{FFN}(h_i^{(l)})\bigr)\,, \end{aligned}$$2	BLEU$$\begin{aligned} h_i^{(l)} &= \text{LN}\bigl(x_i^{(l-1)} + \mathrm{MHAtt}(x_i^{(l-1)})\bigr)\ x_i^{(l)} &= \text{LN}\bigl(h_i^{(l)} + \mathrm{FFN}(h_i^{(l)})\bigr)\,, \end{aligned}$$3	BERTScore
CSD (full)	45.53	30.90	74.61
w/o NM (no mask)	44.75	30.42	74.27
w/o IL (no labels)	42.88	30.53	73.22
w/o CA (no cross-att)	43.39	28.73	72.79
w/o AL (no attn-loss)	43.66	28.91	70.97

The paper concludes that all three components—progressive mask, label injection, and attention losses—contribute to performance (Jiang et al., 2023). Removing label injection yields the largest drop in BLEU$$\begin{aligned} h_i^{(l)} &= \text{LN}\bigl(x_i^{(l-1)} + \mathrm{MHAtt}(x_i^{(l-1)})\bigr)\ x_i^{(l)} &= \text{LN}\bigl(h_i^{(l)} + \mathrm{FFN}(h_i^{(l)})\bigr)\,, \end{aligned}$$4, while removing cross-attention yields the highest diversity, suggesting that cross-attention tightly focuses the model. This suggests a measurable trade-off between targeted conditioning and lexical diversity.

7. Scope, limitations, and research significance

The work is framed around elders with cognitive impairment and around dialogue grounded in CS principles and emotional support strategies (Jiang et al., 2023). Within that scope, CSConv fills a Chinese-language resource gap and provides a benchmark where cognitive stimulation, emotion perception, and supportive response generation are jointly modeled. The dataset’s construction from CST videos and the CST handbook gives the benchmark an explicit therapeutic lineage rather than a purely open-domain conversational origin.

At the same time, the paper explicitly notes that there is still a large space for improvement compared to human performance (Jiang et al., 2023). This is an important corrective to any misconception that performance gains on BLEU, BERTScore, or human preference imply parity with therapist-level interaction. The reported results establish relative effectiveness over the selected baselines, not closure of the gap between generated and human responses.

A further misconception would be to equate CSConv with a generic empathetic dialogue dataset. The label ontology and decoder conditioning make it specifically a CS dialogue resource in which open-ended responses are guided by CS principle and emotional support strategy (Jiang et al., 2023). A plausible implication is that future work on this benchmark would need to preserve that dual emphasis: cognitive stimulation as the core therapeutic orientation, and emotional support as an explicit conversational requirement rather than an incidental stylistic feature.