---
kind: concept
slug: conversational-recommendation
title: Conversational recommendation
summary: >-
  Dialog systems that recommend items during a chat, such as movies. One paper with a card labels each recommender turn with a social strategy (opinion, similarity, encouragement, credibility) and conditions a GPT-2 model on it. Author-reported gains on perplexity, BLEU and human ratings, on a small dataset; the link to success is correlational. Trade-off: needs strategy-annotated dialogues.
tags: [recommendation, dialogue, sociable-agents, dataset]
papers: ["2009.14306"]
related: [dialogue-generation, dialogue-evaluation, decoding-strategies, human-evaluation, collaborative-filtering]
status: draft
updated_at: 2026-10-19
---

## Overview

A conversational recommender talks with a seeker, learns what they want, and
proposes items during the chat. The question the INSPIRED paper asks is whether
the social moves a human recommender makes, such as sharing an opinion or
encouraging the seeker, make a suggestion more likely to be accepted, and whether
a dialog model can learn to make those moves [arxiv:2009.14306 §1].

Success is defined from the seeker's behaviour after the suggestion. A
recommendation counts as successful when the seeker watched more than half of
the trailer and rated it 4 or 5 [arxiv:2009.14306 §5.2]. The same paper also
treats a reply classified as a recommendation as the trigger for a movie
lookup, so the dialog model and the recommender are linked [arxiv:2009.14306
§6.2].

Only one paper with a card covers this concept. Its results are the authors' own,
on a dataset they collected, and they are not compared with other conversational
recommenders. Treat the numbers below as one data point, not a benchmark.

## Variants

### Sociable-strategy labels on a GPT-2 dialog model

The data is human-human movie dialogues, with each utterance labelled by the
strategy it uses. The scheme has eight sociable strategies (personal opinion,
personal experience, similarity, encouragement, offering help, preference
confirmation, self-modelling and credibility), two preference-elicitation
inquiries, and non-strategy classes such as acknowledgement and recommendation
[arxiv:2009.14306 §4.1]. The paper reports 1,001 dialogues and 35,811 utterances,
split 801 / 100 / 100 for training, validation and test [arxiv:2009.14306 §6]
[arxiv:2009.14306 tl;dr].

The model has two GPT-2 small language models, one per speaker, and both
condition on the shared history [arxiv:2009.14306 §6.1]. The strategy-labelled
version writes the strategy as a special token before each recommender turn.
The baseline uses the same architecture without the token and is trained on the
same dialogues [arxiv:2009.14306 §6.1] [arxiv:2009.14306 §6.2].

Reported results, author-reported on 100 test dialogues [arxiv:2009.14306 tab. 7]:

| Metric | Baseline | Strategy-labelled |
|---|---|---|
| Perplexity (lower is better) | 9.28 | 8.93 |
| BLEU-4 (higher is better) | 5.11 | 6.63 |

In a human comparison with 28 participants, the labelled model was preferred on
all five aspects, with the largest gap on engagingness [arxiv:2009.14306 §6.3].
Five-point ratings from 50 participants give the same ordering, with the
engagingness scores at 3.56 for the baseline and 3.72 for the labelled model
[arxiv:2009.14306 tab. 10]. The text gives no counts or significance test for the
pairwise comparison [arxiv:2009.14306 §6.3].

The link to success is correlational. In a logistic regression over 667 dialogues
with a clear outcome, personal opinion (0.12), similarity (0.23), encouragement
(0.20) and credibility (0.09) have positive coefficients, and similarity and
encouragement are significant at the stated levels [arxiv:2009.14306 §5.2]
[arxiv:2009.14306 tab. 6]. The authors do not run a controlled test, so these
coefficients show association only [arxiv:2009.14306 §5.2].

## When to use which

There is one variant, so the choice is between using its method or not. The
constraints below are the ones the paper's evidence bears on.

| Situation | Use | Why |
|---|---|---|
| You can collect or buy human-human dialogues and label utterances with a strategy scheme | This variant | It is the only one with evidence here. The label is one special token, so the architecture cost does not change [arxiv:2009.14306 §6.2] |
| You have no labelled dialogues | Not covered | The baseline also needs the same dialogues, just without labels [arxiv:2009.14306 §6.1]. The paper reports no method that avoids labelling |
| Limited GPU budget | Budget for two GPT-2 small models and a 1024-token history | Training is the main cost. Reported times: 1.37 hours per pretraining epoch and 16 minutes per fine-tuning epoch on one GTX 1080 Ti [arxiv:2009.14306 §6.1] [arxiv:2009.14306 app. D] |
| You need a claim that a strategy increases acceptance | Run your own controlled test | The paper's acceptance evidence is correlational [arxiv:2009.14306 §5.2] |
| You need the recommendation trigger | Plan for a classifier | The paper's BERT classifier reports 95.4% accuracy and 91.2% F1 for recommendations, on the authors' data [arxiv:2009.14306 §6.2] |

## Implementation sketch

The core loop is: prepend the strategy label, sample several candidate replies,
drop degenerate ones, and pick one. The paper's selection rule differs from the
released code, so the sketch shows the paper's steps and the notes below list
the differences.

```pseudocode
# Fine-tune one GPT-2 model per speaker on labelled turns  [§6.1]
# Each recommender turn starts with its strategy as a special token  [§6.2]
turn_tokens = [STRATEGY_TOKEN[turn.label]] + tokenize(turn.text)

# Sample candidate replies from the same history  [§6.2]
# The paper samples five candidates per turn
candidates = [sample(model_a, history, temperature=0.8,
                     top_k=400, top_p=0.9)
              for _ in range(5)]  # [app. D]

# Drop candidates with repeated bigrams  [code note below]
candidates = [c for c in candidates if not has_repeated_bigram(c)]

# Paper: prefer candidates with encouragement or the longest length
# [§6.2]
chosen = pick_candidate(candidates)
```

Notes for implementers:

- The paper's selection rule (encouragement or greatest length) is not what
  its released code runs, which chooses at random among the filtered candidates.
  The code also samples three candidates, not five. The
  [paper card](/papers/2009.14306) lists these differences in full.
- The paper gives a temperature of 0.82 for the baseline and 0.8 for the labelled
  model [arxiv:2009.14306 app. D]. The released code has only 0.8.
- The paper says training runs "until convergence", about two epochs for the
  baseline and three for the labelled model [arxiv:2009.14306 app. D]. The
  released code runs ten epochs.
- The labelled model does not enforce its strategy. The label is generated as
  part of the reply, and nothing checks that the chosen reply carries it.

## Contradictions

Only one paper with a card covers this concept, so there are no disagreements
between papers. Within the paper, the text reports 79.7% of cases as successful
[arxiv:2009.14306 §3.3], but the case counts in its table do not produce that
figure under any simple grouping the paper card checked [arxiv:2009.14306 tab. 3].
Treat that percentage as unreconciled.

## Papers

- [INSPIRED: Toward Sociable Recommendation Dialog Systems](/papers/2009.14306):
  adds labelled social strategies to a GPT-2 recommender and reports author-run
  gains on one small dataset.