L’apprentissage des représentations des personnages améliore la compréhension du dialogue

Dans cette étude, nous présentons une nouvelle approche pour améliorer la génération de dialogues en tirant parti de l'apprentissage de la représentation des personnages. Notre modèle est formé sur un ensemble de données à grande échelle de romans chinois et modélise explicitement la relation entre les personnages de l'histoire. Cela permet à notre modèle de générer un dialogue plus cohérent et raisonnable en considérant les motivations et les interactions des différents personnages. Nous évaluons notre modèle sur deux tâches de génération de dialogue - la génération de dialogue masqué et la reconnaissance du locuteur du dialogue - et démontrons sa supériorité sur des bases de référence solides. Nos travaux mettent en évidence l’importance de l’apprentissage de la représentation des personnages pour la génération de dialogues et ouvrent de nouvelles voies pour de futures recherches.

Authors:

Auteurs:

(1) Jianzhu Yao, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(1) Jianzhu Yao, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(2) Ziqi Liu, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(2) Ziqi Liu, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(3) Jian Guan, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(3) Jian Guan, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(4) Minlie Huang, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology. Table of Links Abstract and Intro Related Works DIALSTORY Dataset Proposed Tasks Methodology Experiments Discussion Future Work Conclusion Limitations and References 6 Experiments 6.1 Masked Dialogue Generation Implementation Details To conduct experiments on the masked dialogue generation task, we decide the hyper-parameters based on the performance of the validation set. We train Shao et al. (2021)’s BART model for 4.6 epochs with a 1e-4 learning rate for 1 day, and for our model, we train it for 5.6 epochs with a 1e-4 learning rate for 1 day. All baselines and our model are trained using the Adam optimizer. During the training process for our method, we computed the selection coverage of characters within a single story. And it showed that in every 1000 training steps, the coverage of different characters ranged from 98.64% to 99.00%, which meant nearly all the characters are selected during training, and all the characters contributed to the generated dialogue. It further proved that the argmax in Eq. 3 operation doesn’t break the gradient progress when training for this task. Automatic Evaluation Following previous works, we use several standard, widely used automatic evaluation metrics. We use BLEUn (Papineni et al., 2002) to measure the average word overlap between each generated and groundtruth dialogue turn (n=1,2), and Distinct-n (Li et al., 2015) to evaluate n-gram diversity of generated dialogue turns (n=2,3,4). To be more specific, for the coherence classifier, we construct the training and validation sets by randomly shuffling the order of dialogue turns and keeping other content in the correct order. We regard the perturbed story as a negative example and the original story as a positive example. We sample another 195k stories (except those in DIALSTORY) from the novels of Guan et al. (2022) to construct the training set (190k examples) and the validation set (5k examples). We train the model for 4 epochs with a 2e-5 learning rate and a 16- batch size, using the Adam optimizer. During the evaluation, we consider an example coherent when the probability of being coherent predicted by the classifier is greater than 0.5. We use the ratio of outputs (along with the input) that are classified as coherent by the classifier to all generated outputs as the coherence score. The result of the automatic evaluation is presented in Table 3. According to the table, compared to the BART baseline, our model consistently generates more word overlaps with ground truth and achieves better diversity under the guidance of character representations, which means our model can generate more diverse but not commonplace responses. Figure 3 plots the coherence score varying with the number of masked dialogue turns. The result shows that our model gets a higher coherence score than BART when required to generate more than seven turns of dialogue in one story. Manual Evaluation We conduct a pairwise comparison between our model and the BART baseline. We randomly select 100 examples from the test set. For each pair of outputs along with the input, we ask three annotators to give a preference (win, lose and tie) in terms of fluency, coherence, and informativeness. All the annotations are native Chinese speakers. We adopt majority voting to make final decisions among the annotators. The three aspects of manual evaluation are as follows: Fluency: Grammatical correctness and intra-sentence linguistic quality.

(4) Minlie Huang, The CoAI group, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information. Tableau des liens Résumé et introduction Travaux connexes Ensemble de données DIALSTORY Tâches proposées Méthodologie Expériences Discussion Travaux futurs Conclusion Limites et références 6 Expériences 6.1 Détails de mise en œuvre de la génération de dialogue masqué Pour mener des expériences sur la tâche de génération de dialogue masqué, nous décidons des hyper-paramètres en fonction des performances de l’ensemble de validation. Nous formons Shao et al. (2021) pour 4,6 époques avec un taux d'apprentissage de 1e-4 pendant 1 jour, et pour notre modèle, nous l'entraînons pendant 5,6 époques avec un taux d'apprentissage de 1e-4 pendant 1 jour. Toutes les lignes de base et notre modèle sont formés à l'aide de l'optimiseur Adam. Au cours du processus de formation à notre méthode, nous avons calculé la couverture de sélection des personnages au sein d'une seule histoire. Et cela a montré que toutes les 1000 étapes de formation, la couverture des différents personnages variait de 98,64 % à 99,00 %, ce qui signifie que presque tous les personnages sont sélectionnés pendant la formation et que tous les personnages ont contribué au dialogue généré. Il a en outre prouvé que l'argmax dans l'équation. L’opération 3 n’interrompt pas la progression du gradient lors de l’entraînement à cette tâche. Évaluation automatique Suite à des travaux antérieurs, nous utilisons plusieurs métriques d'évaluation automatique standards et largement utilisées. Nous utilisons BLEUn (Papineni et al., 2002) pour mesurer le chevauchement moyen de mots entre chaque tour de dialogue généré et vérité fondamentale (n = 1,2), et Distinct-n (Li et al., 2015) pour évaluer la diversité des n-grammes. de tours de dialogue générés (n=2,3,4). Pour être plus précis, pour le classificateur de cohérence, nous construisons les ensembles de formation et de validation en mélangeant aléatoirement l'ordre des tours de dialogue et en gardant les autres contenus dans le bon ordre. Nous considérons l’histoire perturbée comme un exemple négatif et l’histoire originale comme un exemple positif. Nous échantillonnons 195 000 autres histoires (sauf celles de DIALSTORY) tirées des romans de Guan et al. (2022) pour construire l'ensemble de formation (190 000 exemples) et l'ensemble de validation (5 000 exemples). Nous entraînons le modèle pendant 4 époques avec un taux d'apprentissage de 2e-5 et une taille de 16 lots, à l'aide de l'optimiseur Adam. Lors de l'évaluation, nous considérons un exemple cohérent lorsque la probabilité d'être cohérent prédite par le classifieur est supérieure à 0,5. Nous utilisons le rapport entre les sorties (ainsi que les entrées) classées comme cohérentes par le classificateur et toutes les sorties générées comme score de cohérence. Le résultat de l'évaluation automatique est présenté dans le tableau 3. Selon le tableau, par rapport à la ligne de base BART, notre modèle génère systématiquement plus de chevauchements de mots avec la vérité terrain et atteint une meilleure diversité sous la direction des représentations de caractères, ce qui signifie que notre modèle peut générer des réponses plus diverses mais pas courantes. La figure 3 représente le score de cohérence variant en fonction du nombre de tours de dialogue masqués. Le résultat montre que notre modèle obtient un score de cohérence plus élevé que BART lorsqu'il est nécessaire de générer plus de sept tours de dialogue dans une histoire. Évaluation manuelle Nous effectuons une comparaison par paires entre notre modèle et la référence BART. Nous sélectionnons au hasard 100 exemples dans l’ensemble de tests. Pour chaque paire de sorties ainsi que l'entrée, nous demandons à trois annotateurs de donner une préférence (gagner, perdre et égalité) en termes de fluidité, de cohérence et d'information. Toutes les annotations sont de langue maternelle chinoise. Nous adoptons le vote majoritaire pour prendre les décisions finales parmi les annotateurs. Les trois aspects de l'évaluation manuelle sont les suivants : Maîtrise : exactitude grammaticale et qualité linguistique intra-phrase.

Coherence: Inter-sentence relatedness, causal and temporal dependencies. We judge the coherence between the story and a dialogue turn by following the criterion in Table 5. We add the scores of all the generated dialogue turns in a story to get the overall coherence score of the story, which is then used to compare with each other.

Cohérence : relation entre les phrases, dépendances causales et temporelles. Nous jugeons la cohérence entre l'histoire et un tour de dialogue en suivant le critère du tableau 5. Nous additionnons les scores de tous les tours de dialogue générés dans une histoire pour obtenir le score de cohérence global de l'histoire, qui est ensuite utilisé pour comparer avec chacun. autre.

Informativeness: Interesting, diverse and rich details. As shown in Table 4, all the results show moderate (κ > 0.4) agreement, which shows our model outperforms the BART baseline significantly in dialogue informativeness and coherence. Case Study Figure 5 showed two examples to investigate how learning character representations can help our model generate more coherent dialogue. We found that our model can better model the relationship between different characters and the direction of the storyline. For example, in the first case, we can see that the BART’s generation confuses different characters’ fathers, while our model captures the relationship between different characters, and generates proper responses for the corresponding characters, which also moves the plot forward. And in the second case, we can see that BART’s generation is commonplace and contradicts the plot development. In contrast, our model captures the intentions of the speaker and the development trend of the plot, generating an appropriate and coherent response. Since these two models use the same pretrained weight, we can infer that the character modeling module leverages the coherent and reasonable generation. We also summarize four error types of the generated dialogue turn for the DialGen task: (1) Intersentence Contradiction; (2) Inter-sentence Repetition; (3) Intra-sentence Contradiction; (4) Intrasentence Repetition. We show the typical corresponding cases in Figure 4. We conducted a quantitative analysis of those 4 error types on our model’s generation. We analyzed 20 stories with 103 dialog turns and the results are shown in Figure 6. We found that both our model and BART suffer from these errors, suggesting that there is still space for model improvement, especially in the inter-sentence repetition. 6.2 Dialogue Speaker Recognition Implementation Details To conduct experiments on the speaker recognition task, we decide the hyper-parameters based on the performance of the validation set. For the BART baseline and our approach, we insert a mask token before each dialogue needed to be predicted, and a person id token before and after each character name span. Then, we insert all the unique person id tokens before the input stories as different options, and make predictions based on the cos similarity of option tokens and mask tokens. We train Shao et al. (2021)’s BART model for 30 epochs with a 5e-5 learning rate for 3 days, For encoder-only baselines, we implemented BERT, RoBERTa, and MacBERT and trained them for 15 epochs wity a le-5 learning rate for 2 days. For our model, we train it for 22 epochs with a le-6 learning rate for 2 days. All baselines and our model are trained using the Adam optimizer. Metrics We evaluate the DialSpk task using two automatic metrics including dialogue-level accuracy (DAC) and story-level accuracy (SAC). DAC is calculated as the ratio of the correct predictions to the total number of specifies dialogue turns, while SAC is the ratio of the number of stories where all dialogue turns are correctly predicted to the number of all test examples. These two metrics provide the evaluation for dialogue understanding with different granularities. Results As shown in Table 7, our model outperforms all the baselines significantly (p< 0.01, Wilcoxon signed-rank test) on both DAS and SAC scores, suggesting the benefit of learning character representations. We tested the accuracy of automatic training set annotations, and the DAC/SAC scores are 86.78%/67.80%. Together with the model’s performance on the test set, we can see the automatic annotation for the training set is of good quality. We also conducted the human prediction experiment, and the DAC/SAC scores are 97.90%/90.70%, which are much higher than the best model. So there is much room for further improvement for machine-based approaches. This paper is available on arxiv under CC 4.0 DEED license. [2] https://huggingface.co/fnlp/ bart-base-chinese [3] https://huggingface.co/bert-base-chinese [4] https://huggingface.co/hfl/ chinese-roberta-wwm-ext [5] https://huggingface.co/hfl/ chinese-macbert-base Authors:

Auteurs:

(1) Jianzhu Yao, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(1) Jianzhu Yao, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(2) Ziqi Liu, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(2) Ziqi Liu, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(3) Jian Guan, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(3) Jian Guan, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(4) Minlie Huang, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology. Authors:

Auteurs:

Authors:

Auteurs:

(1) Jianzhu Yao, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(1) Jianzhu Yao, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(2) Ziqi Liu, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(2) Ziqi Liu, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(3) Jian Guan, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology;

(3) Jian Guan, groupe CoAI, Université Tsinghua, Pékin, Département chinois d'informatique et de technologie, Université Tsinghua, Pékin, Chine Centre national de recherche de Pékin pour les sciences et technologies de l'information ;

(4) Minlie Huang, The CoAI group, Tsinghua University, Beijing, China Department of Computer Science and Technology, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology. Table of Links Abstract and Intro Abstract and Intro Related Works Related Works DIALSTORY Dataset DIALSTORY Dataset Proposed Tasks Proposed Tasks Methodology Methodology Experiments Experiments Discussion Discussion Future Work Future Work Conclusion Conclusion Limitations and References Limitations and References 6 Experiments 6.1 Masked Dialogue Generation Implementation Details To conduct experiments on the masked dialogue generation task, we decide the hyper-parameters based on the performance of the validation set. We train Shao et al. (2021)’s BART model for 4.6 epochs with a 1e-4 learning rate for 1 day, and for our model, we train it for 5.6 epochs with a 1e-4 learning rate for 1 day. All baselines and our model are trained using the Adam optimizer. Implementation Details During the training process for our method, we computed the selection coverage of characters within a single story. And it showed that in every 1000 training steps, the coverage of different characters ranged from 98.64% to 99.00%, which meant nearly all the characters are selected during training, and all the characters contributed to the generated dialogue. It further proved that the argmax in Eq. 3 operation doesn’t break the gradient progress when training for this task. Automatic Evaluation Following previous works, we use several standard, widely used automatic evaluation metrics. We use BLEUn (Papineni et al., 2002) to measure the average word overlap between each generated and groundtruth dialogue turn (n=1,2), and Distinct-n (Li et al., 2015) to evaluate n-gram diversity of generated dialogue turns (n=2,3,4). Automatic Evaluation To be more specific, for the coherence classifier, we construct the training and validation sets by randomly shuffling the order of dialogue turns and keeping other content in the correct order. We regard the perturbed story as a negative example and the original story as a positive example. We sample another 195k stories (except those in DIALSTORY) from the novels of Guan et al. (2022) to construct the training set (190k examples) and the validation set (5k examples). We train the model for 4 epochs with a 2e-5 learning rate and a 16- batch size, using the Adam optimizer. During the evaluation, we consider an example coherent when the probability of being coherent predicted by the classifier is greater than 0.5. We use the ratio of outputs (along with the input) that are classified as coherent by the classifier to all generated outputs as the coherence score. The result of the automatic evaluation is presented in Table 3. According to the table, compared to the BART baseline, our model consistently generates more word overlaps with ground truth and achieves better diversity under the guidance of character representations, which means our model can generate more diverse but not commonplace responses. Figure 3 plots the coherence score varying with the number of masked dialogue turns. The result shows that our model gets a higher coherence score than BART when required to generate more than seven turns of dialogue in one story. Manual Evaluation We conduct a pairwise comparison between our model and the BART baseline. We randomly select 100 examples from the test set. For each pair of outputs along with the input, we ask three annotators to give a preference (win, lose and tie) in terms of fluency, coherence, and informativeness. All the annotations are native Chinese speakers. We adopt majority voting to make final decisions among the annotators. The three aspects of manual evaluation are as follows: Manual Evaluation Fluency: Grammatical correctness and intra-sentence linguistic quality. Coherence: Inter-sentence relatedness, causal and temporal dependencies. We judge the coherence between the story and a dialogue turn by following the criterion in Table 5. We add the scores of all the generated dialogue turns in a story to get the overall coherence score of the story, which is then used to compare with each other.

Cohérence : relation entre les phrases, dépendances causales et temporelles. Nous jugeons la cohérence entre l'histoire et un tour de dialogue en suivant le critère du tableau 5. Nous additionnons les scores de tous les tours de dialogue générés dans une histoire pour obtenir le score de cohérence global de l'histoire, qui est ensuite utilisé pour comparer avec chacun. autre.

Informativeness: Interesting, diverse and rich details. Fluency: Grammatical correctness and intra-sentence linguistic quality. Fluency : Grammatical correctness and intra-sentence linguistic quality. Fluency Coherence: Inter-sentence relatedness, causal and temporal dependencies. We judge the coherence between the story and a dialogue turn by following the criterion in Table 5. We add the scores of all the generated dialogue turns in a story to get the overall coherence score of the story, which is then used to compare with each other.

Cohérence : relation entre les phrases, dépendances causales et temporelles. Nous jugeons la cohérence entre l'histoire et un tour de dialogue en suivant le critère du tableau 5. Nous additionnons les scores de tous les tours de dialogue générés dans une histoire pour obtenir le score de cohérence global de l'histoire, qui est ensuite utilisé pour comparer avec chacun. autre.

Coherence : Inter-sentence relatedness, causal and temporal dependencies. We judge the coherence between the story and a dialogue turn by following the criterion in Table 5. We add the scores of all the generated dialogue turns in a story to get the overall coherence score of the story, which is then used to compare with each other. Coherence Informativeness: Interesting, diverse and rich details. Informativeness : Interesting, diverse and rich details. Informativeness As shown in Table 4, all the results show moderate (κ > 0.4) agreement, which shows our model outperforms the BART baseline significantly in dialogue informativeness and coherence. Case Study Figure 5 showed two examples to investigate how learning character representations can help our model generate more coherent dialogue. We found that our model can better model the relationship between different characters and the direction of the storyline. For example, in the first case, we can see that the BART’s generation confuses different characters’ fathers, while our model captures the relationship between different characters, and generates proper responses for the corresponding characters, which also moves the plot forward. And in the second case, we can see that BART’s generation is commonplace and contradicts the plot development. In contrast, our model captures the intentions of the speaker and the development trend of the plot, generating an appropriate and coherent response. Since these two models use the same pretrained weight, we can infer that the character modeling module leverages the coherent and reasonable generation. Case Study We also summarize four error types of the generated dialogue turn for the DialGen task: (1) Intersentence Contradiction; (2) Inter-sentence Repetition; (3) Intra-sentence Contradiction; (4) Intrasentence Repetition. We show the typical corresponding cases in Figure 4. We conducted a quantitative analysis of those 4 error types on our model’s generation. We analyzed 20 stories with 103 dialog turns and the results are shown in Figure 6. We found that both our model and BART suffer from these errors, suggesting that there is still space for model improvement, especially in the inter-sentence repetition. 6.2 Dialogue Speaker Recognition Implementation Details To conduct experiments on the speaker recognition task, we decide the hyper-parameters based on the performance of the validation set. For the BART baseline and our approach, we insert a mask token before each dialogue needed to be predicted, and a person id token before and after each character name span. Then, we insert all the unique person id tokens before the input stories as different options, and make predictions based on the cos similarity of option tokens and mask tokens. We train Shao et al. (2021)’s BART model for 30 epochs with a 5e-5 learning rate for 3 days, For encoder-only baselines, we implemented BERT, RoBERTa, and MacBERT and trained them for 15 epochs wity a le-5 learning rate for 2 days. For our model, we train it for 22 epochs with a le-6 learning rate for 2 days. All baselines and our model are trained using the Adam optimizer. Metrics We evaluate the DialSpk task using two automatic metrics including dialogue-level accuracy (DAC) and story-level accuracy (SAC). DAC is calculated as the ratio of the correct predictions to the total number of specifies dialogue turns, while SAC is the ratio of the number of stories where all dialogue turns are correctly predicted to the number of all test examples. These two metrics provide the evaluation for dialogue understanding with different granularities. Metrics Results As shown in Table 7, our model outperforms all the baselines significantly (p< 0.01, Wilcoxon signed-rank test) on both DAS and SAC scores, suggesting the benefit of learning character representations. We tested the accuracy of automatic training set annotations, and the DAC/SAC scores are 86.78%/67.80%. Together with the model’s performance on the test set, we can see the automatic annotation for the training set is of good quality. We also conducted the human prediction experiment, and the DAC/SAC scores are 97.90%/90.70%, which are much higher than the best model. So there is much room for further improvement for machine-based approaches. Results This paper is available on arxiv under CC 4.0 DEED license. This paper is available on arxiv under CC 4.0 DEED license. available on arxiv [2] https://huggingface.co/fnlp/ bart-base-chinese [3] https://huggingface.co/bert-base-chinese [4] https://huggingface.co/hfl/ chinese-roberta-wwm-ext [5] https://huggingface.co/hfl/ chinese-macbert-base