RHO ($\rho$): Reducing Hallucination in Open-domain Dialogues with Knowledge Grounding

Researchers have been devoted to reducing hallucination in open-domain dialogue systems incorporating external knowledge. Neural Path Hunter (NPH) (Dziri et al., 2021) leverages a hallucination critic and retrieves faithful entities by a query signal propagated over a sub-graph in the refinement stage. Shuster et al. (2021) explore various neural-retrieval-in-the-loop architectures where a retriever is introduced for knowledge selection. Rashkin et al. (2021) propose an faithfulness control code in decoding using re-sampling techniques. Wu et al. (2021) define a control mechanism with lexical control phrases and inductive attention where potentially uninformative attention links are removed. Our work improves the fusion and interaction between external knowledge and dialogue context via various knowledge groundings and reasoning techniques, further reducing hallucination.

KGs convey large amounts of structured knowledge, which can help to improve dialogue systems’ performance in informativeness Tuan et al. (2019) and empathy (Li et al., 2020). For open-domain dialogue generation, Liu et al. (2019) unify knowledge triples and texts as a graph, and conduct multi-hop reasoning for explainability. Xu et al. (2020a) propose a proactive dialogue generation method based on agnostic meta-learning considering the limited number of KGs. Kumar et al. (2020) learn unified representations by training syntactic graph convolution networks, knowledge, and memory module with triplet loss. Xu et al. (2022); Zhou et al. (2018); Zhang et al. (2020) explore and demonstrate how commonsense KG facilitates language generation in dialogue systems. Besides, Yang et al. (2020); Rony et al. (2022); Chaudhuri et al. (2021) are committed to incorporating KG into task-oriented dialogue models. Different from the above literature, our work employs the factoid knowledge paths from KG to improve the faithfulness of open-domain dialogue systems.

In the response generation task in dialogue systems, each data sample consists of a dialogue history $H$, including a set of utterances $U$ conducted by humans and agents interactively. The goal of the response generation model in the dialogue system is to learn to generate a proper response $R$ based on the dialogue history $H$.

The response generation for KGD task is a special case of the above task that takes a multi-relational KG as an additional input; Multi-relational KG is a directed graph $\mathcal{G}$ formulated by a collection of triples, denoted as $T=\left\langle\texttt{[SBJ],[PRE],[OBJ]}\right\rangle$, as additional input. Here, [SBJ],[OBJ] are subject and object entities, and [PRE] is the relation predicate $r$ indicating the relationship between the subject and object entities $e$. The goal of KGD is to generate a faithful response $R$ based on the history $H$ and a knowledge sub-graph $\mathcal{G}_{H}$, which is the subset of the entire KG $\mathcal{G}$ with triples semantically related to the dialogue history $H$. Our task is also in line with the previous works (Zhou et al., 2021; Dziri et al., 2021). Figure 2 has an illustrated example.

A naive approach of input construction from raw data samples for language models is simply concatenating all triples in $\mathcal{G_{H}}$ with the dialogue history $H$³³3Similar to the baseline approach in Dziri et al. (2021).. However, there is a lack of guidance to specifically excavate the model’s innate ability to handle the KGD task (Brown et al., 2020). Inspired by Raffel et al. (2020), we design a prompt to guide the PLM for KGD and convert the structured $\mathcal{G_{H}}$ into textual information. Here, we linearize the triples in $\mathcal{G_{H}}$ into texts (treated as meta-knowledge) and cooperate the dialogue history utterances $U$ with the following template: “Given the knowledge: [SBJ]₁ <sep> [PRE]₁ <sep> [OBJ]₁ <triple> [SBJ]₂ <sep> [PRE]₂ <sep> [OBJ]₂ <triple> $\cdots$ <user> U₁ <assistant> U₂ $\cdots$”, where <sep>, <triple>, <user>, and <assistant> are special markers.

Although $\mathcal{G_{H}}$ is converted and injected as additional input (§ 3.2), the model using textual information only cannot effectively handle the semantics of KGs which are typically sparse and complex in form Petroni et al. (2019); Logan et al. (2019). Therefore, we ground the language representations with KG to take full advantage of lexical and structured knowledge information simultaneously.

During pre-processing, we obtain the collections of linked mentions of entities ($\mathcal{E}_{H}$) and relations ($\mathcal{R}_{H}$) from the dialogue history $H$ and their KG embeddings as follows:

1. 1.

   Identify entity mention $e_{m}$ that appears in dialogue history $H$ that can be linked to an entity $e$ in the sub-graph $\mathcal{G_{H}}$. We utilise an open-source linking tool named FLAIR Akbik et al. (2019).

2. 2.

   Since relations connecting entities are crucial in knowledge reasoning for PLMs Labutov et al. (2018); Feng et al. (2020), we also link the relation mention $r_{m}$ in $H$ to the relation predicate $r$ in $\mathcal{G}_{H}$.

3. 3.

   We employ TransE (Bordes et al., 2013) to learn the KG embeddings of entities ($\vec{e}_{\mathcal{G}}$) and relation predicates ($\vec{r}_{\mathcal{G}}$) from the entire $\mathcal{G}$ ⁴⁴4We employ TransE via OpenKE (Han et al., 2018) as our underlying KG representation learning algorithm due to its effectiveness and simplicity. We have also experimented with some recent algorithms and observed no improvement in the performance (refer to Appendix C for details)..

Then, we obtain a locally grounded token embedding $\vec{w}_{local}$ for an arbitrary non-special token $w$ in $H$ as follows:

$$\vec{w}_{local}=\left\{\begin{matrix}M(\vec{e}_{\mathcal{G}})&\texttt{substr}(w,\mathcal{E}_{H})\\ M(\vec{r}_{\mathcal{G}})&\texttt{substr}(w,\mathcal{R}_{H})\\ \vec{0}&\text{otherwise}\end{matrix}\right.$$

(1)

where $M(\cdot)$ transforms the space from the KG embeddings to the PLM token embeddings. A typical way of implementing $M(\cdot)$ is through a mapping matrix. Specifically, if $dim(\vec{w})=dim(\vec{e}_{\mathcal{G}})$, $M(\cdot)$ can be further simplified as an identity mapping. $\texttt{substr}(w,\mathcal{E}_{H})$ is a Boolean indicator that returns true if the current token $w$ is a sub-string of any $e_{m}$ in $\mathcal{E}_{H}$. This way, tokens related to specific entities or relation predicates can be grounded by their respective KG embeddings by fusing $\vec{w}_{local}$ into the vanilla token embedding $\vec{w}$. We regard this approach to be local as $\vec{w}_{local}$ is only related to the KG embedding of the corresponding node.

As in Figure 2 and Figure 3, the tokens "X", "-", "Men", "2" in dialogue history are linked to the entity "X-Men 2" in KG. Then, we take the corresponding KG embedding from the database gained by TransE as the local knowledge grounding which will fuse into these tokens’ textual embeddings.

Focusing only on a single token in the context and a single node in the graph is insufficient to enhance the multi-hop reasoning abilities of the dialogue system. In addition to local grounding, we further propose global knowledge grounding which enriches the semantics considering the entire sub-graph $\mathcal{G}_{H}$ and hence offers the model a comprehensive view of the background knowledge.

Following our observation, we adopt the attention mechanism Vaswani et al. (2017) to draw global dependencies between the dialogue history $H$ and a memory bank storing the representations of all knowledge triples in $\mathcal{G}_{H}$. Let $T_{H}$ be the collection of all triples in $\mathcal{G}_{H}$. The memory bank stores $|T_{H}|$ embedding vectors where the $i$-th vector $\vec{v}_{i}$ corresponds to the KG embedding of the $i$-th relation triple $T_{i}=\left\langle\texttt{[SBJ],[PRE],[OBJ]}\right\rangle$:

$$\vec{v}_{i}=M([SBJ])\oplus M([PRE])\oplus M([OBJ])$$

(2)

where $\oplus$ is the concatenation operator for vectors. We gather all vectors and further project them to a global knowledge embedding space by:

$$\begin{split}K_{H}=W_{proj}\cdot[\vec{v}_{1}\oplus\cdots\oplus\vec{v}_{|T_{H}|}]\end{split}$$

(3)

where $W_{proj}$ is a learnable projection matrix.

Based on the formulation of $K_{H}$, we compute how much attention the current token $w$ in $H$ pays to each relation triple according to the semantic relevance and obtain the globally grounded token embedding $\vec{w}_{global}$ as follows:

$\vec{w}_{global}=\left\{\begin{matrix}\text{softmax}(\frac{\vec{w}\cdot K_{H}^{T}}{\sqrt{dim(\vec{w})}})\cdot K_{H}&\texttt{substr}(w,\mathcal{E}_{H}),\\ &\texttt{substr}(w,\mathcal{R}_{H})\\ \vec{0}&otherwise\end{matrix}\right.$

(4)

As in Figure 2 and Figure 3, for ‘‘X-Men 2’’ in the dialogue history, Triple 1 should have more influence on the tokens’ representation than Triple 2 due to its higher relevance.

Finally, the encoder of RHO sums the vanilla token embedding $\vec{w}$, the locally grounded embedding $\vec{w}_{local}$, and the globally grounded embedding $\vec{w}_{global}$ as: $\tilde{w}=\vec{w}+\vec{w}_{local}+\vec{w}_{global}$. During training, while $\vec{w}$ is rapidly updated via back propagation, $\vec{w}_{local}$ and $\vec{w}_{global}$ are relatively fixed with few parameters trainable (e.g., $W_{proj}$).

With the above approaches, our knowledge-grounded model generates $N$ candidate responses by beam search. Yet, the grounding process mainly applies to the embedding level, lacking output constraints. To enhance our RHO’s ability to reduce hallucination, we extend KG-CRUSE Sarkar et al. (2022) and train a conversational reasoning model $\phi$ for response re-ranking, with emphasis on the KG. If the generated response can be reasoned backward to the source, we can assume it is faithful.

In our approach, we obtain the semantic embeddings of the dialogue history $H$ and a possible response $R$ via a contextual sentence encoder, i.e., Sentence-BERT (Reimers and Gurevych, 2019). The model $\phi$ is an LSTM-based decoder that learns the probability $p_{t,\phi}$ of an action $\vec{a}_{t}$ given the state $\vec{s}_{t}$ at step $t$. Here, the action refers to a walking step on the graph $\mathcal{G}_{H}$, represented as $\vec{a}_{t}$, which is the concatenation of the relation and entity embeddings derived from the KG, together with their semantic embeddings based on Sentence-BERT, i.e., ⁵⁵5We have also investigated the impact of the KG embeddings ($\vec{e}_{\mathcal{G}}$ and $\vec{r}_{\mathcal{G}}$) for action modeling in Appendix E.1 by comparing the performance of the re-ranker under two settings: i) $\vec{a}_{t}=(\vec{e}_{\mathcal{G}}+\vec{e}_{\mathcal{S}})\oplus(\vec{r}_{\mathcal{G}}+\vec{r}_{\mathcal{S}})$ and ii) $\vec{a}_{t}=\vec{e}_{\mathcal{S}}\oplus\vec{r}_{\mathcal{S}}$ (the vanilla model in KG-CRUSE).

$$\vec{a}_{t}=(\vec{e}_{\mathcal{G}}+\vec{e}_{\mathcal{S}})\oplus(\vec{r}_{\mathcal{G}}+\vec{r}_{\mathcal{S}})$$

(5)

where $\vec{e}_{\mathcal{S}}$ and $\vec{r}_{\mathcal{S}}$ are the semantic sentence embeddings of an entity $e$ and a relation predicate $r$, respectively. The state $\vec{s}_{t}$ contains the representations of the dialogue history, together with entities and relations already traversed by $\phi$ (action history). It is defined as a tuple $(H,(\vec{a}_{1},\vec{a}_{2},\cdots,\vec{a}_{t\mbox{-}1}))$. Hence, the model $\phi$ explicitly models the process of a traversal upon $\mathcal{G}_{H}$ conditioned on the dialogue history $H$ and a possible response $R$. During training, each action $\vec{a}_{t}$ made by $\phi$ is combined into a path, and the target path is the given context-related sub-graph $\mathcal{G}_{H}$.

After our encoder-decoder model generates $N$ candidate responses $\{R_{1},\cdots,R_{N}\}$, we select the best response $R^{*}$ with the highest probability $\mathbf{p_{\phi}}=\prod_{t}p_{t,\phi}$ over all the generated responses, i.e.,

$$R^{*}=\mathop{\arg\max}\limits_{n\in\{1,\cdots,N\}}\mathbf{p}_{\phi}(\mathbf{A}=\mathcal{G_{H}}|H,R_{n})$$

(6)

where $\mathbf{A}$ is a collection of actions $\vec{a}_{t}$ (i.e. knowledge path) that $\phi$ has already traversed conditioned on the dialogue history $H$ and each response $R_{n}$.

For a more intuitive understanding, refer to the example in Figure 2 where the model generates three candidate responses: A, B, and C. It selects Response C as the final output with the traversal path ‘‘X-Men 2, directed by, Bryan Singer, $\sim$directed by, Superman Returns’’. ⁶⁶6$\sim$directed by refers to the opposite direction of the relation directed by. As seen, there is a higher matching degree between the sub-graph in Figure 2 and the Response C, compared to other candidate responses (i.e., A and B).

OpenDialKG (Moon et al., 2019) contains open-ended dialogues between two speakers, initiated by talking about a given entity and grounded on the relevant facts from a structured KG. Thus, the sequential turn-based dialogues can be regarded as traversing the paths in the KG. To our knowledge, OpenDialKG is currently the only publicly available corpus for English open-ended dialogues with KG path annotations Yu et al. (2022); Ni et al. (2022), and previous works (Dziri et al., 2021; Zhou et al., 2021) evaluate their effectiveness on this corpus. Hence, we also conduct our experiments on OpenDialKG. Consistent with previous works (Dziri et al., 2021; Liu et al., 2019; Zhou et al., 2021), we filter OpenDialKG by keeping only the dialogue samples that are annotated with a KG path. The dataset is divided into training, validation, and testing sets in the ratio of 8:1:1.

The following strong baselines are employed to show the efficiency of our method. We fine-tune pre-trained language models GPT2 Radford et al. (2019) and BART (Lewis et al., 2020) on our task. NPH (Dziri et al., 2021) refines the generated responses by retrieved entities from the KG. To our knowledge, the integration of GPT2 and NPH, called GPT2+NPH, reaches the SOTA performance on OpenDialKG. Since this post-processing technique is agnostic to the generation model, we apply it to BART, named BART+NPH as our baseline. In addition, EARL Zhou et al. (2021) utilizes external KGs for conversation generation without parameterizing specific entity representations. KG-BART (Liu et al., 2021), a KG-augmented pre-trained language generation model based on BART, introduces the information of the relations among concepts for generative commonsense reasoning. We are the first to adapt this model to the KGD generation. Furthermore, we explore ChatGPT on this task in Appendix A. Please refer to it for the details of baseline implementations.

At a more granular level, the extrinsic hallucination problem is more frequent than the intrinsic one in all models. This phenomenon is also observed in other works (Dziri et al., 2021; Nan et al., 2021). Specifically, compared to SOTA (GPT2+NPH), RHO reduces extrinsic hallucination by 42.85%. Compared to BART+NPH, RHO reduces intrinsic hallucination by 13.66% and extrinsic hallucination by 46.74%. According to the A/B test results for fluency in Table 3, RHO is slightly more fluent than SOTA methods. Overall, human evaluation results are in line with automatic evaluation. RHO mitigates both intrinsic and extrinsic hallucination issues without sacrificing fluency.

We conduct an ablation analysis to assess the contribution of each component of our method: Local Knowledge Grounding (LKG), Global Knowledge Grounding (GKG), and Response Re-ranking (RR). As shown in the last four rows of Table 1, fully-implemented RHO performs best in automatic hallucination metrics with a slight sacrifice of classical overlap metrics. Specifically, compared to models equipped with only local/global knowledge grounding, the model equipped with both two (LKG+GKG) gains higher scores in FeQA, QuestEval, and Entity Coverage. The same trend is observed when comparing the fully-implemented model with the model without re-ranking (LKG+GKG) and please refer to Appendix E.2 for an example. The trade-off between hallucination metrics and the others is due to the fact that some reference responses in the dataset diverge from the sources (Dziri et al., 2022). Hence, improving responses’ consistency with the source will inevitably reduce that with the references. Overall, the above three mechanisms synergy to improve the generated responses’ faithfulness without significant reduction in quality.

Although RHO achieves better performance than the baselines in the KDG task, it still generates a few cases of failure. To gain more insight into our model, we present failure examples from RHO in Appendix D and conduct an error analysis.

As shown in Table 2, the extrinsic hallucination problem is more frequent. Because the corpus itself has more extrinsic hallucinations and the model is more likely to produce unverified rather than contradictory information with the limitation of the source. To solve the problem, we suggest that the quality of the corpus used for model training, such as OpenDialKG, can be further improved by filtering out irrelevant and contradictory information. Fact-checking can also be potentially used to find evidence from world knowledge and verify the generated responses Ji et al. (2022a).

In addition to intrinsic and extrinsic hallucination issues, we discover two other types of negative cases, namely No Full Coverage, and Unnatural Connection. No Full Coverage refers to the situation where the output does not cover the full answer as expressed by all triples mentioned in the given sub-graph. We believe that further research on the reasoning ability and interpretability of the model can help address this issue. Unnatural Connection denotes that a response is not connected to the dialogue history especially the last utterance naturally. We find that in a few cases, the dialogue system delivers an irrelevant answer or poorly replies to the dialogue history. This issue occurs because our model focuses more on hallucination reduction and sometimes pays less attention to flexibility and diversity. In the future, we can explore the controllability of the grounding degree on knowledge and strike a dynamic balance between faithfulness and flexibility in response generation.