Version classiqueVersion mobile

Proceedings of the Seventh Italian Conference on Computational Linguistics CLiC-it 2020

Felice Dell'Orletta
Johanna Monti
Fabio Tamburini

Contributed Papers

Overprotective Training Environments Fall Short at Testing Time: Let Models Contribute to Their Own Training

Alberto Testoni et Raffaella Bernardi


Despite important progress, conversational systems often generate dialogues that sound unnatural to humans. We conjecture that the reason lies in their different training and testing conditions: agents are trained in a controlled “lab” setting but tested in the “wild”. During training, they learn to generate an utterance given the human dialogue history. On the other hand, during testing, they must interact with each other, and hence deal with noisy data. We propose to fill this gap by training the model with mixed batches containing both samples of human and machine-generated dialogues. We assess the validity of the proposed method on GuessWhat?!, a visual referential game.1

Texte intégral

We kindly acknowledge the support of NVIDIA Corporation with the donation of the GPUs used in our research at the University of Trento. We acknowledge SAP for sponsoring the work.

1. Introduction

1Important progress has been made in the last years on developing conversational agents, thanks to the introduction of the encoder-decoder framework (Sutskever, Vinyals, and Le 2014) that allows learning directly from raw data for both natural language understanding and generation. Promising results were obtained both for chit-chat (Vinyals and Le 2015) and task-oriented dialogues (Lewis et al. 2017). The framework has been further extended to develop agents that can communicate about a visual content using natural language (Vries et al. 2017; Mostafazadeh et al. 2017; Das, Kottur, Gupta, et al. 2017). It is not easy to evaluate the performance of dialogue systems, but one crucial aspect is the quality of the generated dialogue. These systems must in fact produce a dialogue that sounds natural to humans in order to be employed in real-world scenarios. Although there is not a general agreement on what makes a machine-generated text sound natural, some features can be easily identified: for instance, natural language respects syntactic rules and semantic constraints, it is coherent, it contains words with different frequency distribution but that crucially are informative for the conveyed message, and it does not have repetitions, both at a token and a sentence level.

2Unfortunately, even state-of-the-art dialogue systems often generate a language that sounds unnatural to humans, in particular with respect to the large number of repetitions contained in the generated output. We conjecture that part of the problem is due to the training paradigm adopted by most of the systems. In the Supervised Learning training paradigm, the utterances generated by the models during training are used only to compute a Log Likelihood loss function with the gold-standard human dialogues and they are then thrown away. In a multi-turn dialogue setting, for instance, the follow-up utterance is always generated starting from the human dialogue and not from the previously generated output. In this way, conversational agents never really interact one with the other. This procedure resembles a controlled “laboratory setting", where the agents are always exposed to “clean" human data at training time. Crucially, when tested, the agents are instead left alone “in the wild", without any human supervision. They have to “survive" in a new environment by exploiting the skills learned in the controlled lab setting and by interacting one with the other.

3Agents trained in a Reinforcement Learning fashion are instead trained “in the wild" by maximizing a reward function based on the task success of the agent, at the cost of a significant increase of computational complexity. Agents trained according to this paradigm generate many repetitions and the quality of the dialogue degrades. This issue is mildly solved by the Cooperative Learning training, but still, several repetitions occur in the dialogues, making them sound unnatural.

4In this paper, we propose a simple but effective method to alter the training environment so that it becomes more similar to the testing one (see Figure 1). In particular, we propose to replace part of the human training data with dialogues generated by conversational agents talking to each other; these dialogues are “noisy", since they may contain repetitions, a limited vocabulary etc. We then propose to train a new instance of the same conversational agent on this new training set. The model is now trained “out of the lab" since the data it is exposed to are less controlled and they get the model used to live in an environment more similar to the one it will encounter during testing.

Figure 1

Figure 1

Two-steps training method of the C Bot: two Bots, A and B, are trained independently to reproduce human dialogues; then they play together to generate new dialogues (step 1). In step 2 the Bot C is trained on mixed batches of human and machine-generated data (by A and B in step 1).

5We assessed the validity of the proposed method on a referential visual dialogue game, GuessWhat?! (Vries et al. 2017). We found that the model trained according to our method outperforms the one trained only on human data with respect both to the accuracy in the guessing game and to the linguistic quality of the generated dialogues. In particular, the number of games with repeated questions drops significantly.

2. Related Work

6The need of going beyond the task success metric has been highlighted in Shekhar et al. (2019b), where the authors compare the quality of the dialogues generated by their model and other state-of-the-art questioner models according to some linguistic metrics. One striking feature of the dialogues generated by these models is the large number of games containing repeated questions, while the dialogues used to train the model (collected with human annotators) do not contain repetitions. In Shekhar et al. (2019a) the authors enrich the model proposed in Shekhar et al. (2019b) with a module that decides when the agent has gathered enough information and is ready to guess the target object. This approach is effective in reducing repetitions but, crucially, the task accuracy of the game decreases.

7Murahari et al. (2019) propose a Questioner model for the GuessWhich task (Das, Kottur, Moura, et al. 2017) that specifically aims to improve the diversity of generated dialogues by adding a new loss function during training: the authors propose a simple auxiliary loss that penalizes similar dialogue state embeddings in consecutive turns. Although this technique reduces the number of repeated questions compared to the baseline model, there is still a large number of repetitions in the output. Compared to these methods, our method does not require to design ad-hoc loss functions or to plug additional modules in the network.

8The problem of generating repetitions not only affects dialogue systems, but instead it seems to be a general property of current decoding strategies. Holtzman et al. (2020) found that decoding strategies that optimize for an output with high probability, such as the widely used beam/greedy search, lead to a linguistic output that is incredibly degenerate. Although language models generally assign high probabilities to well-formed text, the highest scores for longer texts are often repetitive and incoherent. To address this issue, the authors propose a new decoding strategy (Nucleus Sampling) that shows promising results.

3. Task and Models

3.1 Task

9The GuessWhat?! game (de Vries et al. 2017) is a cooperative two-player game based on a referential communication task where two players collaborate to identify a referent. This setting has been extensively used in human-human collaborative dialogue (Clark 1996; Yule 2013). It is an asymmetric game involving two human participants who see a real-world image. One of the participants (the Oracle) is secretly assigned a target object within the image and the other participant (the Questioner) has to guess it by asking binary (Yes/No) questions to the Oracle.

3.2 Models

10We use the Visually-Grounded State Encoder (GDSE) model of Shekhar et al. , i.e. a Questioner agent for the GuessWhat?! game. We consider the version of GDSE trained in a supervised learning fashion (GDSE-SL). The model uses a visually grounded dialogue state that takes the visual features of the input image and each question-answer pair in the dialogue history to create a shared representation used both for generating a follow-up question (QGen module) and guessing the target object (Guesser module) in a multi-task learning scenario. More specifically, the visual features are extracted with a ResNet-152 network (He et al. 2016) and the dialogue history is encoded with an LSTM network. Since QGen faces a harder task and thus requires more training iterations, the authors made the learning schedule task-dependent. They called this setup modulo-n training, where n specifies after how many epochs of QGen training the Guesser component is updated together with QGen. The QGen component is optimized with the Log Likelihood of the training dialogues, and the Guesser computes a score for each candidate object by performing the dot product between visually grounded dialogue state and each object representation. As standard practice, the dialogues generated by the QGen are used only to compute the loss function, and the Guesser is trained by receiving human dialogues. At test time, instead, the model generates a fixed number of questions (5 in our work) and the answers are obtained with the baseline Oracle agent presented in de Vries et al. (2017). Please refer to Shekhar et al. (2019b) for any additional detail on the model architecture and the training paradigm.

4. Metrics

11The first metric we considered is the simple task accuracy (ACC) of the Questioner agent in guessing the target object among the candidates. We use four metrics to evaluate the quality of the generated dialogues. (1) Games with repeated questions (GRQ), which measures the percentage of games with at least one repeated question verbatim. (2) Mutual Overlap (MO), which represents the average of the BLEU-4 score obtained by comparing each question with the other questions within the same dialogue. (3) Novel questions (NQ), computed as the average number of questions in a generated dialogue that were not seen during training (compared via string matching). (4) Global Recall (GR), which measures the overall percentage of learnable words (i.e. words in the vocabulary) that the models recall (use) while generating new dialogues. MO and NQ metrics are taken from Murahari et al. (2019), while the GR metric is taken from van Miltenburg et al., (2019). We believe that, overall, these metrics represent a good proxy of the quality of the generated dialogues.

5. Datasets

12We are interested in studying how modifying part of the human data in the training set affects the linguistic output and the model’s accuracy on the GuessWhat game. More specifically, we aim at building a training set in which part of the dialogues collected with human annotators are replaced with dialogues generated by the GDSE-SL questioner model while playing with the baseline Oracle model on the same games being replaced. In this way, we build a training set containing dialogues that are more similar to the ones the model will generate at test time while playing with the Oracle.

5.1 Human data

13The training set contains about 108K dialogues and the validation and test sets 23K each. Dialogues contain on average 5.2 turns. The GuessWhat?!dataset was collected via Amazon Mechanical Turk by de Vries et al. (2017). The images used in GuessWhat?!are taken from the MS-COCO dataset (Lin et al. 2014). Each image contains at least three and at most twenty objects. More than ten thousand people in total participated in the dataset collection procedure. Humans could stop asking questions at any time, so the length of the dialogues is not fixed. Humans used a vocabulary of 17657 words to play GuessWhat?!: 10469 of these words appear at least three times, and thus make up the vocabulary given to the models. For our experiments, we considered only those games in which humans succeeded in identifying the target object and that contain less than 20 turns.

5.2 Mixed Batches

14We let the GDSE-SL model play with the baseline Oracle on the same games of the human training dataset. This produces automatically generated data for the whole training set. The model uses less than 3000 words out of a vocabulary of more than 10000 words. We built new training sets according to two criteria: the proportion of human and machine-generated data (50-50 or 75-25) and the length of the generated dialogue. Either we always keep a fixed dialogue length (5 turns, as the average length in the dataset) or we take the same number of turns that the human Questioner used while playing the game we are replacing.

15Table 1 reports some statics of different training sets. Human dialogues have a very low mutual overlap and a much larger vocabulary than both the generated (0-100) and mixed batches datasets (50-50, 75-25). Looking at the number of games with at least one repeated question in the training set (GRQ column in Table 1), it can be observed that human annotators never produce dialogues with repetitions. The 75/25 dataset configuration contains less than 3% of dialogues with repeated questions and this percentage rises to around 5% for the 50/50 configuration and to around 10% for generated dialogues. Looking at the vocabulary size, the human dataset (100-0) contains around ten thousand unique words, the mixed batches datasets (50-50, 75-25) around 4500 words, and the generated dialogues (0-100) approximately 2500 words.

6. Experiment and Results

6.1 Experiment

16As a first step, we trained the GDSE-SL model for 100 epochs as described in Shekhar et al. (2019b). At the end of the training, we used GDSE to play the game with the Oracle on the whole training set, saving all the dialogues. We generate these dialogues with the model trained for all the 100 epochs since it generates fewer repetitions, although it is not the best-performing on the validation set. The dialogues generated by GDSE while playing with the Oracle are noisy: they may contain duplicated questions, wrong answers, etc. See Figure 2 for an example of human and machine-generated dialogues for the same game. We design different training sets as described in Section 5 and train the GDSE-SL model on these datasets. We scrutinize the effect of training on different sets using the metrics described in Section 4 by letting the model generate new dialogues on the test set while playing with the Oracle.

Figure 2

Figure 2

GuessWhat sample dialogues between two human annotators (left) and two conversational agents (right, generated by GDSE-SL as in Shekhar et al. (2019b)). The yellow box highlights the target entity that the Questioner has to guess by asking binary questions to the Oracle. Both humans and conversational agents have to guess the target object only at the end of the dialogue. Note that the machine-generated dialogue on the right contains repetitions from the Questioner and wrong answers from the Oracle (both in italic).

6.2 Results

17Table 2 reports the results of the GDSE model trained on different training sets. To sum up, there are five dataset configurations: apart from the original GuessWhat dataset composed of dialogues produced by human annotators (100% Human Dialogues), there are datasets composed of 75% human dialogues and 25% generated dialogues or 50% human dialogues and 50% generated dialogues. For each dataset configuration, the generated dialogues can be always 5-turns long (“fixed" length) or they can have the same number of turns human annotators used for that game (“variable" length). We do not report the results on the dataset composed of generated dialogues only since it leads to a huge drop in the accuracy of the guessing game.

18By looking at the results on the test set, we can see how even a small number of machine-generated dialogues affects the generation phase at test time, when the model generates 5-turns dialogues and, at the end of the game, it guesses the target object. First of all, it can be noticed that the accuracy of GDSE-SL trained on the new datasets outperforms the one trained on the original training set: in particular, the accuracy of GDSE trained on 50% human dialogues and 50% 5-turns generated dialogues is almost 2% higher (in absolute terms) than the model trained only on human dialogues. The model seems to benefit from being exposed to noisy data at training time to better perform in the guessing game using the dialogues generated by the model itself while playing with the Oracle.

19The linguistic analysis of the dialogues generated on the test set reveals that the models trained on “mixed" batches produce better dialogues according to the metrics described in Section 4. In particular, considering the best-performing model on the test set, the percentage of games with repeated questions drops by 14.3% in absolute terms and the mutual overlap score by 0.09. The percentage of vocabulary used (global recall), on the other hand, remains stable. Interestingly, the only metric that seems to suffer from the model being trained on mixed datasets is the number of novel questions in the generated dialogue: being trained on noisy data does not seem to improve the “creativity" of the model, measured as the ability to generate new questions compared to ones seen at training time.

20Overall, our results show an interesting phenomenon: replacing part of the GuessWhat?!training set with machine-generated noisy dialogues, and training the GDSE-SL questioner model on this new dataset, is found to improve both the accuracy of the guessing game and the linguistic quality of the generated dialogues, in particular with respect to the reduced number of repetitions in the output.


21Despite impressive progress on developing proficient conversational agents, current state-of-the-art systems produce dialogues that do not sound as natural as they should. In particular, they contain a high number of repetitions. To address this issue, methods presented so far in the literature implement new loss functions, or modify the models’ architecture. When applied to referential guessing games, these techniques have the drawback of gaining little improvement, degrading the accuracy of the referential game, or producing incoherent dialogues. Our work presents a simple but effective method to improve the linguistic output of conversational agents playing the GuessWhat?! game. We modify the training set by replacing part of the dialogues produced by human annotators with machine-generated dialogues. We show that a state-of-the-art model benefits from being trained on this new mixed dataset: being exposed to a small number of “imperfect" dialogues at training time improves the quality of the output without deteriorating its accuracy on the task. Our results show an absolute improvement in the accuracy of +1.8% and a drop in the number of dialogues containing duplicated questions of around -14%. Further work is required to check the effectiveness of this approach on other tasks/datasets, and to explore other kinds of perturbations on the input of generative neural dialogue systems.


Herbert H. Clark. 1996. Using Language. Cambridge University Press.

Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, José M. F. Moura, Devi Parikh, and Dhruv Batra. 2017. “Visual Dialog.” In Proceedings of the Ieee Conference on Computer Vision and Pattern Recognition, 326–35.

Abhishek Das, Satwik Kottur, José M. F. Moura, Stefan Lee, and Dhruv Batra. 2017. “Learning Cooperative Visual Dialog Agents with deep Reinforcement Learning.” In 2017 Ieee International Conference on Computer Vision, 2951–60.

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. “Deep Residual Learning for Image Recognition.” In Proceedings of the Ieee Conference on Computer Vision and Pattern Recognition, 770–78.

Mike Lewis, Denis Yarats, Yann N. Dauphin, Devi Parikh, and Dhruv Batra. 2017. “Deal or No Deal? End-to-End Learning for Negotiation Dialogues.” In ArXiv:1706.05125, 2443–53.

Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. “Microsoft COCO: Common Objects in Context.” In Proceedings of Eccv (European Conference on Computer Vision), 740–55.

Nasrin Mostafazadeh, Chris Brockett, Bill Dolan, Michel Galley, Jianfeng Gao, Georgios P. Spithourakis, and Lucy Vanderwende. 2017. “Image-Grounded Conversations: Multimodal Context for Natural Question and Response Generation.” In Proceedings of the the 8th International Joint Conference on Natural Language Processing, 462–72.

Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. 2014. “Sequence to Sequence Learning with Neural Networks.” In Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, 3104–12.

Oriol Vinyals, and Quoc V. Le. 2015. “A Neural Conversational Model.” ICML Deep Learning Workshop.

Harm de Vries, Florian Strub, Sarath Chandar, Olivier Pietquin, Hugo Larochelle, and Aaron C. Courville. 2017. “GuessWhat?! Visual Object Discovery Through Multi-Modal Dialogue.” In 2017 Ieee Conference on Computer Vision and Pattern Recognition, 5503–12.

George Yule. 2013. Referential Communication Tasks. Routledge.


1 Copyright © 2020 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).

Table des illustrations

Titre Figure 1
Légende Two-steps training method of the C Bot: two Bots, A and B, are trained independently to reproduce human dialogues; then they play together to generate new dialogues (step 1). In step 2 the Bot C is trained on mixed batches of human and machine-generated data (by A and B in step 1).
Fichier image/jpeg, 128k
Titre Figure 2
Légende GuessWhat sample dialogues between two human annotators (left) and two conversational agents (right, generated by GDSE-SL as in Shekhar et al. (2019b)). The yellow box highlights the target entity that the Questioner has to guess by asking binary questions to the Oracle. Both humans and conversational agents have to guess the target object only at the end of the dialogue. Note that the machine-generated dialogue on the right contains repetitions from the Questioner and wrong answers from the Oracle (both in italic).
Fichier image/jpeg, 72k

© Accademia University Press, 2020

Conditions d’utilisation :


Open access

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search