Persona-centric Metamorphic Relation guided Robustness Evaluation for Multi-turn Dialogue Modelling

In the case of retrieval dialogues, the requirement for meticulously curated datasets is crucial. These datasets serve as references for evaluating the system’s ability to retrieve relevant responses. The process of creating such a dataset necessitates extensive human effort and the application of relevant knowledge to meticulously choose and annotate suitable responses that align with a given context or query. It involves a significant investment of human labor and requires individuals with expertise in the field to ensure the careful curation of the dataset. The meticulous curation process further adds to the labor-intensive nature of evaluating personalized retrieval-based dialogue systems. Additionally, the constraints of human resources and time can lead to limited assessments, which might not fully capture the complexities and challenges that dialogue systems may encounter in applications.

The examples as illustrated in Figure 1, reveal an important limitation. When we modify the persona to a synonymous sentence with a similar meaning and input it into the model, it becomes confused, resulting in an output that deviates from the expected relationship with the original source output. This indicates a lack of robustness in the model’s performance. Moreover, some researches find that certain models may achieve correct answers through the memorization of specific patterns or keywords rather than a genuine understanding and inference of the answers Gardner et al. (2020).

Traditional reference-based validation methods solely focus on reporting the "consistency" of actual outputs with the true labels, and unfortunately, they do not effectively reveal the true language comprehension capabilities of the models Ribeiro et al. (2021). Additionally, these methods fail to capture the associated shortcomings of model robustness. While traditional reference-based metrics provide a basic understanding of model performance, there is a need for more comprehensive and rigorous evaluation methods, such as metamorphic testing, to assess the consistency and robustness of personalized dialogue retrieval systems under challenging conditions. This can help identify potential weaknesses and limitations of existing methods and guide the development of more reliable and robust personalized dialogue systems.

In the domain of personalized multi-turn dialogue systems, various types of datasets exist, such as personalized datasets Qian et al. (2021), personalized empathy datasets Zhong et al. (2020), and datasets incorporating personalized information with background knowledge Jang et al. (2022). For this paper, our primary focus is on the Persona-Chat dataset Zhang et al. (2018), while acknowledging the presence of other relevant datasets that will be explored and discussed in future studies.

The Persona-Chat dataset, represented as $D$, consisting of n conversation tuples in the format of $(c,p,r,y)$. Specifically, $c=\left\{u_{1},u_{2},...,u_{n_{c}}\right\}$ represents the $n_{c}$ context utterances, $p=\left\{p_{1},p_{2},...,p_{n_{p}}\right\}$ is the $n_{p}$ personas of the speaker, and $r$ is the response candidate for $c$. The label $y\in\left\{0,1\right\}$ indicates whether $r$ is the appropriate response for $(c,p)$, where $y=1$ means it is appropriate, while $y=0$ means it is not. Our objective is to learn a matching function $g$ from $D(c,p,r)$ such that, given any tuple $(c,p,r)$, $g(c,p,r)$ calculates the degree of matching between $(c,p)$ and $r$.

The Persona-Chat dataset Zhang et al. (2018) is currently the most extensive publicly available dataset of multi-turn dialogues conditioned on persona. It consists of 65,719 context-response pairs for the training set, 7,801 for the validation set, and 7,512 for the test set. The dataset includes correct responses from real humans, while incorrect responses are randomly sampled. To increase the challenge of the task, measures are taken to ensure that there is no overlap of contexts and roles between the training, validation, and test sets. This approach guarantees a robust evaluation of models’ generalization capabilities across diverse scenarios. Table 1 presents a selection of illustrative examples from this diverse dataset. Several studies Gu et al. (2019, 2020b, 2021a); Das et al. (2022); Wolf et al. (2019) have explored both non-pre-trained and pre-trained approaches on this dataset to enhance personalized dialogue retrieval and maintain consistent personality in the responses.

In our study of multi-turn personalized dialogue retrieval, we investigate three paradigms and conduct evaluations of the task’s robustness performance. As shown in Figure 2(a), the non-pretraining approach involves training the personalized dialogue retrieval model using dataset $D(c,p,r,y)$, such as DIM Gu et al. (2019) and FIRE Gu et al. (2020b).

Pretraining methods utilize a pretrained model trained on a large-scale corpus. Fine-tuning initializes the personalized dialogue retrieval model with pretrained model parameters and updates its weights using dataset $D$ and matching function $g$ to adapt it to the task, as shown in Figure 2(b). We will test CoBERT Zhong et al. (2020) and BERT-CRA Gu et al. (2021a).

Illustrated in Figure 2(c), the prompt learning includes an appropriate prompt $p$ within dataset $D$, following BERT’s Masked Language Modeling (MLM) and Next Sentence Prediction (NSP) format. These prompts, such as cloze prompts Gao et al. (2021) (i)) and prompts introduced by Sun et al. (2022)(ii)), aid the pre-trained model in assimilating relevant information, improving personalized dialogue retrieval performance by leveraging acquired knowledge. Therefore,we predominantly utilize three models:prompt-MLM(BERT), which is based on BERT-CRA converted to MLM format, prompt-MLM(DialogLM),which is based on the post-train dialogue model DialogLMZhong et al. (2022) and converted to MLM format, and prompt-NSP(BERT), which is converted to NSP format.

The assessment of the software outcomes in study Chen et al. (2021) is quantified as the “violation rate" ($V_{r}$). This metric is employed to gauge the extent to which the models under investigation generate responses that do not satisfy the metamorphic relation. Specifically, we establish the metamorphic relation $MR_{i}\left(t_{i},r\right)$ based on the persona in the personalized dialogue retrieval task, where $t_{i}$ represents a transformation and $r$ denotes the output relation. As previously mentioned, the $r$ signifies consistent output. The test set $\mathbb{S}=\left\{s_{1},s_{2},...,s_{n}\right\}$ is utilized to transform an invariant-example $\mathbb{S^{\mathit{i}}}=\left\{s_{1}^{i},s_{2}^{i},...,s_{n}^{i}\right\}$ based on this relation. Subsequently, we input $\mathbb{S}$ and $\mathbb{S^{\mathit{i}}}$ into the personalized dialogue retrieval model $P$ to obtain results $O_{j}$ and $O_{j}^{i}$, respectively. Based on $r$, we determine that $V_{r}^{i}=1$ if $O_{j}\neq O_{j}^{i}$, and $V_{r}^{i}=0$ otherwise.

For a standardized evaluation of personalized dialogue retrieval models with previous studies, the recall of true positive responses (hits@1) and mean reverse ranking (MRR) are used.

Our study revolves around answering three research questions:

RQ1: To assess the overall validity of metamorphic testing constructed using persona-based metamorphic relations. As previously discussed, persona plays a vital role in maintaining the consistency of the personality of responses in a multi-turn dialogue model. We have developed persona-centric metamorphic relations to assess the model’s robustness based on persona. This experimental question seeks to determine the effectiveness of persona-centric metamorphic testing in identifying flaws in the model under test. It will also evaluate whether our proposed metamorphic relation can achieve the expected advantage.

RQ2: To compare the performance of multi-turn dialogue models trained with different paradigms. By conducting metamorphic testing on models that adhere to the three paradigms of non-pretraining, fine-tuning, and prompt learning, and by presenting the test results using the proposed metamorphic relations, we can effectively demonstrate the performance of these models across various facets of their capabilities.

RQ3: To investigate the relationship between metrics based on reference answers and the violation rate metrics of metamorphic testing. In subsection 5.1, we will present two metrics utilized in our study: the traditional metric hits@1 and the violation rate of the metamorphic testing. These metrics evaluate our test results from distinct perspectives. Specifically, we will compare the extent of the decrease in hits@1 and the percentage of violations reported by the metamorphic testing. We will also assess their error detection effects.

Metamorphic testing was conducted on seven multi-turn dialogue models using three persona-centered metamorphic relations, with test cases drawn from self-persona and partner-persona of the Persona-Chat dataset. Table 2 presents the violation rates observed during testing. The results indicate that all models produced erroneous outputs, as evidenced by violation rates exceeding 0. These findings demonstrate the effectiveness of metamorphic testing in uncovering flaws in the models’ abilities to handle persona-related dialogues, and underscore the importance of designing appropriate metamorphic relations for testing such models.

It is worth noting that we conducted experiments on the last turn of context to compare the impact of noise introduced by persona on personalized dialogue retrieval. Comparing the effects, we observe that the introduction of noise to either persona or context alone has a similar impact on the model, without causing a significant change. However, when noise is simultaneously applied to both persona and context, we observe a sharp increase in the violation rate, nearly twice as much as the effect of applying noise to one aspect alone. Additionally, referencing the significant drop in hits@1 and MRR demonstrated in Figure 3, this phenomenon suggests that even when the persona and context undergo change separately, the model still possesses the ability to identify a response that aligns with the semantics and personality of the other participant. Hence, it indicates that the model has the capacity to learn and incorporate persona information from the dialogue history. This observation underscores the fact that metamorphic testing can reveal gaps in the models’ semantic comprehension abilities.

To summarize, our results indicate that all personalized dialogue retrieval models experienced a decrease in performance when exposed to data perturbation, with the self-persona models showing a great vulnerability. Therefore, our proposed metamorphic testing approach, along with the application of specific metamorphic relations, shows effective in identifying issues within these models.

The research question we investigate is independent of the dataset used. As Table 2 illustrates, the differences in performance between models in the same scenes are comparable. Therefore, our evaluation primarily centers around the scenes of the self-persona’s original.

Figure 4 displays the violation rates of multi-turn dialogue models for the three training paradigms where Seven models are classified into three types. As stated in RQ1, the violation rate for each category exceeds 0, indicating a robustness issue for each model type. Our results reveal that prompt-based models exhibit lower violation rates than those subjected to fine-tune. This aligns with our expectation that prompt learning leverages more latent knowledge acquired during pre-training compared to fine-tune.

Specially, the prompt-MLM of DialogLM, exhibits the highest violation rate in MT1, surpassing DIM by 8.57%, which has the second-highest violation rate. This phenomenon can be attributed to DialogLM’s pre-training using multi-speaker conversations and long dialog understanding, making our simple prompt setup less suited for its pre-training task compared to BERT. Merely employing prompts without careful consideration not only fails to fully leverage the model’s potential knowledge but also hinders its performance, as evidenced by the self-original hits@1 or MRR scores of the prompt-MLM-DialogLM model in Fig 3.

The violation rates observed in the metamorphic testings based on persona constructions indicate that prompt-based models still possess an advantage in discerning between conversational styles of two speakers based on conversation history, even after introducing perturbations to the personas.

The preceding discussion demonstrates that all model categories experience performance degradation when exposed to data perturbation. The robustness of models that incorporate prompts is generally superior to those based on fine-tune, indicating the presence of untapped knowledge in pre-trained models. These findings highlight the need for further research to fully leverage the potential of pre-trained models.

As Figure 5 depicts, the decline in each model’s metamorphic testing is relatively limited when evaluated using the hits@1 metric with labeled answers. This suggests that these models exhibit a certain level of robustness in terms of synonym recognition, resistance to irrelevant sentence influence, and performance affected by noise. However, it is important to note that this approach, which employs standardized answers for testing, can be considered a type of adversarial testing. Nonetheless, it doesn’t provide sufficient insight into the internal deficiencies of the models or their actual understanding of the required knowledge, like the limitations described in Sec. 2.

On the other hand, by utilizing $V_{r}$ to assess the results of the model test, we were able to identify more errors in the models. Moreover, the use of $V_{r}$ highlights the possibility of answerless scenarios in metamorphic testing. By establishing a metamorphic relation based on persona consistency and constructing metamorphic testing, we gain a comparative advantage over adversarial testing. During the analysis of the preceding two experimental questions (RQs), we leveraged the traditional hits@1 metric to support the evaluation of violation rates in personalized dialogue retrieval tasks with labeled answers. This approach enabled us to conduct a more thorough examination of anomalies in the MT results.

Our study introduces three specific metamorphic testings to examine persona-based output consistency in personalized dialogue retrieval. The results effectively demonstrate the utility of our work in identifying potential comprehension issues within models. We observe significant performance degradation across different model types when exposed to intentional noise, emphasizing the critical role of our tests in revealing limitations. The intentionally introduced noise represents an extreme scenario, providing a lower-bound estimate of the model’s performance in real-world scenarios with natural noise. Our analysis reveals the model’s proficiency in grasping context and persona characteristics, successfully matching response to specific persona even with partial information loss.

Our evaluation of three training paradigms reveals that fine-tuning and prompt learning through pre-training exhibit greater stability than non-pretrained models, showing less susceptibility to data perturbations and performance degradation. This enhanced stability can be attributed to the fact that pre-trained models are trained on large-scale corpora, allowing them to better handle noise and improve generalization. Prompt learning effectively leverages the latent knowledge of the pre-trained model, but constructing templates within the pre-trained format does not fully unleash its capability and can lead to performance degradation. In summary, existing personalized dialogue retrieval models have room for improvement in performance.

Hybrid retrieval and generative modules are commonly used in dialogue systems, leading us to investigate the robustness of generative models. Additionally, we find that the persona-centric metamorphic relation in personalization-based dialogue systems is also applicable to generative models. For example, in Sec. 2, we expect a robust generative model to maintain the originality of generated sentences when introducing persona sentences with similar meanings.

The concept of persona was introduced by the researchers to ensure consistent characterization and clear memory for generating rational responses Zhang et al. (2018); Zhong et al. (2020). Evaluation of personalized dialogue systems commonly involves a reference-based approach to ensure persona consistency. Standardized labels, generated through human evaluation, are used alongside traditional metrics like hits@1 to assess system performance Zhang et al. (2018); Gu et al. (2021a); Das et al. (2022); Wolf et al. (2019). This ensures robust and accurate measurement of system effectiveness. Some researchers Li et al. (2019) have explored robustness in retrieval-based dialogue systems by generating adversarial examples in black-box settings, still the continued need for labeled data.

Metamorphic testing complements traditional reference-based metrics, focusing on targeted assessment. An important benefit is its reliance on existing annotated labels, enabling automated generation of additional testing data without further manual annotation.

Metamorphic testing Chen et al. (2020) is an approach used to evaluate the internal consistency of NLP models by examining the preservation of expected relationships between inputs and outputs Ribeiro et al. (2021). A significant focus of existing metamorphic relations in NLP is to assess the robustness of NLP models. These relations specifically evaluate the ability of a model to maintain consistent output Aspillaga et al. (2020); Belinkov and Bisk (2018); Li et al. (2017); Malfa et al. (2020). Demonstrating their effectiveness, robustness relations have been successfully applied in testing various NLP tasks, including Sentiment Analysis Ribeiro et al. (2021); Jia et al. (2019) and NLI Aspillaga et al. (2020); Malfa et al. (2020), among others. Additionally, metamorphic testing has been extended to test other aspects of model performance, including fairness Ma et al. (2020), and moreRibeiro et al. (2019); Manino et al. (2022). This highlights the versatility of metamorphic testing in assessing different dimensions of NLP models.

Metamorphic testing in NLP is being explored in various tasks with mixed success. Our work aims further to investigate persona-based consistent outputs to assess its robustness with the help of MT.