Large Language Models as Source Planner for Personalized Knowledge-grounded Dialogues

To build a personalized dialog agent, Zhang et al. (2018) extensively investigated this task with a new dataset Persona-Chat, where a pre-defined persona set is a form of multiple sentences of textual description. Lots of works follow this setting and have taken mutual persona perception (Liu et al., 2020; Kim et al., 2020; Xu et al., 2022a), persona-sparse scenario (Song et al., 2021; Welch et al., 2022), long-term persona memory (Xu et al., 2022c), persona extending (Liu et al., 2022b) and persona order bias (Chen et al., 2023) into consideration. Although some of them complement the insufficient semantics in short persona descriptions by further utilizing an external commonsense knowledge base to extend existing persona sets (Majumder et al., 2020; Liu et al., 2022b), they still fall into the conventional framework coupling the knowledge selection with the response generation Wu et al. (2022b), rendering it infeasible to handle various sources of knowledge. There have also been works showing that the combination of different knowledge sources such as persona descriptions and Wikipedia can further improve the overall performance Jang et al. (2022); Wu et al. (2021, 2022a). However, they still fail to capture possible dependency between knowledge sources. In their framework, knowledge is not used as the role to assist persona-consistent response generation, but as an additional resource to generate a more informative response or select a suitable persona Jang et al. (2022); Fu et al. (2022).

Large Language Models (LLMs) show remarkable capabilities in planning the use of various external resources, such as tools Schick et al. (2023), models Shen et al. (2023), and APIs Li et al. (2023), to solve various NLP tasks and suit different applications in practice. Alternatively, different types of knowledge can be retrieved from external sources, as illustrated in WebGPT Nakano et al. (2022) and ReAct Yao et al. (2023). Integrating various knowledge sources to improve the quality of LLM generation becomes increasingly challenging due to the need for strategic planning, sequential decision-making, and complex reasoning. Previous research primarily focuses on either earlier decision-making stages Nakano et al. (2022); Shen et al. (2023) or the subsequent response generation Sun et al. (2023); Schick et al. (2023); Deng et al. (2023a), instead of establishing a complete framework for planning the use of multiple knowledge sources to generate appropriate responses. There is a latest work named TPE which regards different knowledge sources as conceptual tools and proposes a multi-persona collaboration framework to model the decision-making process of the call order for multiple knowledge sources (Wang et al., 2023a). We differ in exploring the planning capability of LLMs to decide whether or not require knowledge, which source to call, and when to call in both the supervised and unsupervised manner.

Seeds Preparation. To reduce annotation cost, we take advantage of the currently available persona dialogue dataset: DuLemon (Xu et al., 2022c) and two widely-adopted Chinese knowledge bases: Baike³³3http://www.openkg.cn and Ownthink⁴⁴4http://github.com/ownthink/KnowledgeGraphData to produce seed data. Specifically, we first cluster all persona sentences from DuLeMon into 10 topics. After removing duplicate, rare, and similar personas, we carefully choose around 20 personas for each left topic as seed personas⁵⁵5Some topics are removed if they contain less than 20 personas. We don’t pick hundreds of personas because one persona sentence has a vast amount of knowledge behind it.. In addition, we manually add some personas for the existing topics and new topics. The final personas consist of age, nation, personality, career, movie, music, sport, book, constellation, locality, gender, others. For retrieving persona-related knowledge, we simply combine two aforementioned knowledge bases with similar filtering operations and store them as (head entity, attribute, tail entity) tuples.

Collection Setting. Following the setting of Jang et al. (2022), annotators are instructed to make a dialogue by considering persona and corresponding knowledge under the single-person setup. In this way, one person can better understand what persona to ask as the human and what knowledge to use as the system, in comparison with two independent persons playing two roles separately. During the collection, each annotator first selects a suitable persona and then optionally identifies relevant knowledge, giving a knowledge-enhanced and persona-consistent response at last.

We organize the collected personas (PERSONA source), persona-related knowledge ⁹⁹9The knowledge related to the same persona forms a document, and different documents form DOCUMENTS source. (DOCUMENTS source), and dialogues in the form shown in Figure 1(c). We finally collect 2,477 dialogues and 24,554 utterances with 5 turns per dialogue on average. We then split the collected data into train, validation, and test sets using the 8:1:1 ratio. The dataset statistics are summarized in Table 1, including the number of dialogues, utterances, average length, as well as data sources used. The average length per utterance reaches 17.6, hinting at the informativity and depth of the conversation. It is shown that over 86% of responses used persona (i.e., resp w/ persona) and 76% used both persona and knowledge (i.e., resp w/ p_and_k), which shows that KBP is capable as a benchmark to evaluate the different grounding abilities of models.

We first provide a general definition of a dialogue system that requires multiple sources and then we instantiate the definition in the context of personalized knowledge-grounded dialogue. For each dialogue, the dialogue context $\boldsymbol{c}=\{u_{1},s_{1},u_{2},s_{2},...,u_{t}\}$ and different knowledge sources $\boldsymbol{K}=\{K_{1},K_{2},...,K_{i}\}$ are provided, where $K_{i}=\{k_{i}^{1},k_{i}^{2},...,k_{i}^{j}\}$ indicates the $i_{th}$ source’s name of $\boldsymbol{K}$. $k_{i}^{j}$ denotes the $j_{th}$ knowledge in natural language from $K_{i}$. If $K_{2}$ is reliant on $K_{1}$, knowledge should be retrieved from $K_{2}$ based on the selected knowledge in $K_{1}$. Such reliance should also be embodied in the construction of $K_{2}$, in a way such as $K_{2}$ = $\{k_{1}^{1}:\{k_{2}^{1},k_{2}^{2}\},k_{1}^{2}:\{k_{2}^{3},k_{2}^{4}\},...\}$. The goal of the system is to generate a response $s_{t}$ conditioned on $\boldsymbol{c}$ and a set of knowledge $\{k_{i}^{j},...,k_{n}^{m}\}$ retrieved from $\boldsymbol{K}$ if required¹⁰¹⁰10Unlike most of the previous works, we also consider the response that does not need any knowledge in our setting.. Specifically in the context of personalized knowledge-grounded dialogue, we regard $\{\texttt{PERSONA},\texttt{DOCUMENTS}\}$ as $\{K_{1},K_{2}\}$ respectively. There is a potential dependency between these two sources, and the goal is to generate a response $s_{t}$ conditioned on a set of knowledge $\{p_{i}^{j},...k_{m}^{n}\}$, where $p_{i}^{j}$ is retrieved from PERSONA and $k_{m}^{n}$ is retrieved from DOCUMENTS. The response can also be generated conditioned on a set of knowledge from a single source PERSONA or without any sources.

There are three different steps in our proposed SAFARI framework: Planning, Retrieval, and Assembling as shown in Figure 1(b). We will introduce them step-by-step under both supervised and unsupervised settings.

Inspired by the recent progress using LLMs as a controller to plan a call order of different models Shen et al. (2023), we adopt a similar way here by providing detailed prompts to leverage the LLMs’ capability to address the dependency between different knowledge sources. We consider two settings: zero-shot and in-context learning for planning and assembling steps here since the retrieval step is the same as above.

The full prompts of the unsupervised planning step can be found in Appendix D. For few-shot in-context learning, we prepend three corresponding demonstrations from the train set to the zero-shot prompts during evaluation.

Implementation Details. We mainly choose BELLE-LLAMA-7B-2M Ji et al. (2023) and ChatGLM-6B Du et al. (2022); Zeng et al. (2023) as two backbone models for supervised setting since they are two popular open-source Chinese models. And we additionally add ChatGPT (gpt-3.5-turbo-0301)¹¹¹¹11https://openai.com/blog/chatgpt for the unsupervised setting. For training, we set the batch size as 8, train models with 3 epochs and save the checkpoint with the lowest validation loss. For other hyper-parameter settings, we mainly follow the corresponding official code¹²¹²12https://github.com/THUDM/ChatGLM-6B and https://github.com/LianjiaTech/BELLE. Due to the computation limit, we conduct training with LoRA Hu et al. (2021) at one single 3090 GPU, and it cost about 4-6 hours. For the unsupervised setting, we set both the temperature and top p as 0.1 to reduce the randomness of LLMs. Besides that, we use three types of retrievers including both sparse and dense retrieval: BM25 Robertson and Zaragoza (2009), DPR Karpukhin et al. (2020), and RocketQAv2 Ren et al. (2021). We only retrieve the top-ranked result from each source in the experiments¹³¹³13The effects of different choices is analyzed at Section 6..

There are three types of decisions representing different sources required in the next step: NULL, PERSONA, and Both (selecting both PERSONA and DOCUMENTS). Table 4 demonstrates the F1 of planning under different settings. Under supervised settings, despite LLMs achieving high F1 scores at Both, the performance at NULL and Persona is still unsatisfactory, since there are fewer training samples in these two cases. On the other hand, under unsupervised settings, the LLMs are over-confident in their decisions to use NULL, and they misunderstand the dependency between different sources (sometimes deciding to only use DOCUMENTS without PERSONA)¹⁴¹⁴14We assign the case that LLMs predict DOCUMENTS only as NULL since this case does not exist in KBP.. This result reveals the LLMs’ low accuracy in expressing uncertainty and fetching unknown knowledge. Furthermore, in-context learning cannot improve this situation, which is similar to the observation in Amayuelas et al. (2023).

With the ground-truth planning labels (except NULL), we examine three types of retrievers, including BM25, RocketQAv2, and DPR, to evaluate the retrieval performance. Table 5 presents the Recall@1 (R@1) of the different retrievers. We found that the DPR and RocketQAv2 can achieve over 80% R@1 when retrieving from PERSONA source while only about 50% from DOCUMENTS and the R@1 at DOCUMENTS^† further decreases after removing the dependency. First, the semantics between different knowledge from DOCUMENTS with the dependency are similar to the same underlying persona $p^{*}$, making them more difficult to be distinguished. In addition, noisy knowledge sentences are introduced since there exists no dependency. Moreover, we observe that DPR performs the best out of these three retrievers in all sources of knowledge while BM25 performs worst¹⁵¹⁵15RocketQAv2 is generally not competitive with DPR because of the pre-trained weights in the RocketQAv2, since it is pre-trained using QA datasets and the length of the question is much shorter than dialogue context., revealing the importance of dense retrieval models in this task. Therefore, we set DPR as the retriever in our experiments afterward.

Table 6 demonstrates the performance of response generation under both supervised and unsupervised settings. Referring to Table 4, the performance of the planning step largely affects the results in the assembling step, when the retriever is the same. Mostly, better planning leads to better responses in all metrics. The supervised models are much better than unsupervised models since their planning results are much better, while ChatGPT performs best under unsupervised settings due to a similar reason. We found that BELLE achieves higher BLEU1 and Rouge-L, K.C but lower P.C than ChatGLM since the planning gap between them mainly comes from PERSONA. In addition, due to poor retrieval performance at DOCUMENTS (Table 5), the consistency score K.C is also much lower than P.C. With demonstrations in the prompt, we observe generally better performance on most metrics, since LLMs tend to accept the personalized role, rather than generating responses like “As an AI language model, I do not have persona ….”. Overall, we conclude that the grounding ability of supervised models is much better than unsupervised ones, and ChatGPT performs best under the unsupervised setting.

We investigate the effects of individual steps by providing the model ground-truth labels from each step to generate the response, enabling us to analyze and understand the specific effects of each step in a clear and systematic way. Table 7 presents the results. First, we note that the inclusion of ground-truth planning labels or knowledge casts a positive impact on performance. Planning primarily enhances P.C and K.C, while grounding knowledge contributes to BLEU1 and Rouge-L scores. The best results are obtained when both signals are combined. Secondly, we also conduct an ablation study by removing some modules: 1) removing dependency information (-Dependency); 2) removing DOCUMENTS and only using PERSONA (-Documents); and 3) removing the planning step by always selecting PERSONA (-Planning^∗) or always selecting PERSONA and DOCUMENTS (-Planning^∗∗) for each turn. It can be found at the bottom of Table 7 that all metrics are dropped differently after removing different components except for Rouge-L when always selecting two knowledge sources. To conclude, SAFARI can effectively incorporate multiple sources (compared with -Documents) and further address dependency issues (compared with -Dependency). Moreover, SAFARI demonstrates its versatility by effectively handling multiple sources and efficiently selecting relevant ground knowledge. Notably, SAFARI outperforms existing methods that indiscriminately utilize all available sources (compared with -Planning^∗∗).

The number of retrieved results plays a key role in the response generation. There is a trade-off between accuracy and recall, while a small number of retrieved results may not cover enough semantics but a large number may introduce additional noises. Table 8 presents the results of the different numbers of retrieved results. We observe that the performance of response generation decreases with the number, which indicates that noisy knowledge will harm the quality of the generated responses.

Human evaluation is conducted to evaluate the quality of generated response in terms of three metrics: coherence score (Coh.), persona consistency score (Per.Cs), and knowledge consistency score (Know.Cs). We randomly sample 100 responses with grounding information for each model and ask three annotators to indicate its coherence score (1-5) and whether the response is consistent with the given persona (1/0), and knowledge (1/0). Table 9 shows the result. We observe that supervised methods achieve higher performance than unsupervised ones, which corroborates the findings of the automatic evaluation results presented in Table 6. Besides that, we found BELLE achieves the highest performance across all metrics and outperforms ChatGLM since the effects of planning are not considered during human evaluation. Moreover, we also found that in-context learning brings a lower rejection rate and more human-like responses. Specifically, the rejection rate of BELLE under the setting of zero-shot learning is about 32%, while the number is reduced to 12% under in-context learning¹⁶¹⁶16More analysis can be found in Appendix E.