Keywords

1 Introduction

Open-domain dialogue response generation (RG) models enable machines to converse with humans using natural language and play an important role in human-computer interaction [43]. However, machines lack enough real-world knowledge cognition because they can only access the parametric knowledge of a model besides the dialogue history [45]. Thus, machines struggle to thoroughly understand the semantics of dialogue histories and generate informative responses. Seeking information from external knowledge sources is an effective solution [50], i.e., knowledge-grounded dialogue response generation (KRG) [4, 17].

Compared to RG models, the superiority of KRG models derives from the ability to use external knowledge [42]. The general paradigm of KRG can be summarized as three stages [7, 39]: 1) Knowledge-Retrieval stage: it first employs an efficient Information Retrieval (IR) tool to retrieve a set of knowledge candidates in a coarse-grained way. The retrieved knowledge candidates contain much irrelevant information because IR tools only consider the literal feature; 2) Knowledge-Selection stage: To filter out irrelevant information and select contextually-relevant knowledge, KRG also has a knowledge selection stage using more fine-grained methods; 3) Response Generation stage: it finally generates the target response by accessing the dialogue history and selected knowledge. Among such three stages, the second knowledge selection stage plays the most crucial role in the research of KRG and has received much attention [10, 23, 28].

This paper focuses on commonsense knowledge-grounded dialogue response generation (CKRG). Despite many successes [42, 46], CKRG still suffers from several challenges, especially in the era of pre-trained language models (PLMs) [14, 19]. First, the knowledge candidate space (i.e., the knowledge can be selected and used when generating the response) is fixed and limited. On the one hand, IR tools can only retrieve knowledge candidates already existing in the knowledge base. On the other hand, the model can only use the knowledge candidate already retrieved by IR tools. This may lead to insufficient knowledge coverage [42]. Second, in the knowledge selection stage, previous CKRG works [46, 50] often use deep but complex networks, which lack enough interpretability to explain the knowledge selection procedure. For example, it is hard to determine which knowledge facts have been selected. Finally, although PLMs are powerful, they also bring many thorny restrictions to the downstream applications [15], such as the length (most PLMs can only operate at most 512/1024 tokens), the input format (must be plain text), the network structure, and so on. Consequently, many knowledge selection methods originally proposed for non-PLM-based models have become incapable in the era of PLMs; then, knowledge selection can only rely on the external network or the implicit self-attention mechanism [23, 49].

Fig. 1.
figure 1

An example. SEG-CKRG can use Generative Knowledge Selection to select the existing knowledge and extend the new knowledge, then generates the response.

Full size image

Considering these challenges, we propose SEG-CKRG, a simple but elegant CKRG model. As shown in Fig. 1, SEG-CKRG introduces a novel Generative Knowledge Selection (GenSel) mechanism, which regards knowledge selection as a generative problem. GenSel uses a PLM to explicitly generate contextually-relevant knowledge based on the dialogue history and the knowledge candidate set retrieved by IR tools. By regarding this task as a generative problem, GenSel can not only select knowledge from the candidate set retrieved by IR tools, but can also extend the knowledge by externalizing the inherent knowledge of PLMs. Then, SEG-CKRG generates the target response conditioned on both the generated knowledge and the retrieved knowledge. Considering both the generative knowledge selection procedure and the dialogue generation procedure are generative problems, we can train/infer SEG-CKRG in an end-to-end fashion. We pre-train two GPT-2 models [27] as the backbone PLMs. To boost the knowledge representation density and the infusing of two generative procedures, we propose an Efficient Input Representation technique and a Dual-Head Generator technique, respectively.

We conduct extensive experiments on a Chinese conversational dataset Weibo-ConceptNet [41], whose dialogues have been aligned to a commonsense knowledge base, ConceptNet [32]. Experimental results have verified that SEG-CKRG has significantly outperformed previous state-of-the-art models, and GenSel can not only accurately select the knowledge but also generate new contextually-relevant knowledge. We also bring extensive analyses to investigate our approach further.

2 Methodology

2.1 Preliminary

Response Generation (RG). Suppose \(\mathcal {D}=\{(H_i, R_i)\}^{N}\) is a conversational corpus, where \(H_i=(h_1,\cdots ,h_{|H_i|})\) is the dialogue history, \(R_i=(r_1,\cdots ,r_{|R_i|})\) is the response. Then, RG learns a conditional language model \(P_{RG}(R_i|H_i)\) to generate \(R_i\) conditioned on \(H_i\): \( P_{RG}(R_i|H_i) = \prod P_{RG}(r_t|r_{<t},H_i)\).

Knowledge-Grounded Response Generation (KRG). RG models tend to generate generic responses such as ‘I don’t know.’ [13] because \(P_{RG}\) can only use the insufficient knowledge hidden in the parameters \(\theta _{RG}\) and the dialogue history. To address this issue, KRG methods try to seek more knowledge from the external knowledge base, such as encyclopedic knowledge [7], commonsense knowledge [32], and so on [26].

More specifically, in the commonsense knowledge-grounded dialogue response generation (CKRG) scenario, there is a knowledge base \(\mathcal {K}=\{k_i=(e_i^h, e_i^r,e_i^t)\}^{M}\), where \(k_i\) is a commonsense fact triplet, \(e_i^h\), \(e_i^r\), and \(e_i^t\) are the corresponding head entity, relation, and tail entity, respectively. Then, for each dialogue history \(H_i\), we need to employ an IR tool to retrieve a set of commonsense facts \(K_i=\{k_{i,j}\}^L, L<< M\) form \(\mathcal {K}\). Finally, the problem of CKRG is given by:

$$\begin{aligned} P_{CKRG}(R_i|H_i) = \prod P_{CKRG}(r_t|r_{<t},H_i,K_i) \end{aligned}$$
(1)

where \(P_{CKRG}\) is a conditional language model with the ability to access the knowledge \(K_i\). Although \(K_i\) is the filtered results via IR tools, IR tools can only consider the token-level literal feature. Thus, a more fine-grained context-aware knowledge selection procedure is needed in \(P_{CKRG}\). In non-PLM CKRG works, this procedure can be explicitly modeled and then integrated into \(P_{CKRG}\). For example, [50] employs graph attention network [34]. In the era of PLM, limited by the input format and network structure, this procedure can only be implicitly performed by the integrated self-attention mechanism or external tools, bringing less interpretability but more limitations to the knowledge selection procedure.

Fig. 2.
figure 2

An overview of SEG-CKRG. We show the input/output examples. In this example, SEG-CKRG has selected a knowledge fact ‘(sport, RelatedTo, Hockey)’ and extended a ‘(sport, RelatedTo, badminton)’ in the example.

Full size image

2.2 Problem Definition and Overview

As shown in Fig. 2, unlike previous CKRG works, SEG-CKRG introduces a novel Generative Knowledge Selection (GenSel) mechanism, which regards knowledge selection as a generative problem. The objective of SEG-CKRG is:

$$\begin{aligned} P_{GenSel}(K^G_i|H_i,K_i) \cdot P_{ResGen}(R_i|H_i,K_i,K^G_i) \end{aligned}$$
(2)

where \(P_{GenSel}(K^G_i|H_i,K_i)\) first generates the contextually-relevant knowledge \(K^G_i\) conditioned both the dialogue history \(H_i\) and the retrieved knowledge \(K_i\); subsequently, \(P_{ResGen}(R_i|H_i,K_i,K^G_i)\) generates the target response \(R_i\).

2.3 Generative Knowledge Selection

SEG-CKRG uses a generative method to explicitly select and extend knowledge. Similar to other generation tasks, it is a conditional language modeling problem:

$$\begin{aligned} P_{GenSel}(K^G_i|H_i,K_i) = \prod P_{GenSel}(k^G_t|k^G_{<t},H_i,K_i) \end{aligned}$$
(3)

Efficient Input Representation. Most PLMs can only accept plain texts as input, which means the structural commonsense knowledge must be linearized to plain text. Thus, the input of \( P_{GenSel} \) is given by:

$$\begin{aligned} S_i = [\omega _K(K_i),\omega _H(H_i),Prompting] \end{aligned}$$
(4)

where \(\omega _H(H_i)\) linearizes the dialogue history with role (human/bot) labels and turn identifiers; Prompting is a prompting textFootnote 1 [52] to hint the PLM about the following generation action; \(\omega _K(K_i)\) linearizes the structural \(K_i=\{k_{i,j}=(e_{i,j}^h, e_{i,i}^r,e_{i,i}^t)\}^{M}\) to a sequence. To reduce the loss of structural information and improve the representation density, \(\omega (K_i)\) uses a graph-level pattern:

$$\begin{aligned} \begin{matrix} \omega (K_i) = (\omega _G(g_{i,1});\omega _G(g_{i,2});\cdots ;\omega _G(g_{i,j});\cdots ) \\ \\ \omega _G(g_{i,j}) = ([G], e^{hg}_{i,j}, e^{rg}_{i,j}, e^{tg}_{i,j,1}, [T], e^{tg}_{i,j,2},\cdots ) \end{matrix} \end{aligned}$$
(5)

where \(K_i\) is first compressed as a set of 1-hop graphs \(G_i=\{ g_{i,j} = e^{hg}_{i,j}, e^{rg}_{i,j}, \{e^{tg}_{i,j}\}\}\); namely, \(\forall k \in K_i\) that have the same head entity \(e^{hg}_{i,j}\) and the same relation \(e^{rg}_{i,j}\) are placed to the corresponding 1-hop graph \(g_{i,j}\); then, \(g_{i,j}\) is sequentially linearized and concatenated with a graph separator [G] and a tail entity separator [T]. Compared to previous triplet-level patterns [42, 52], our graph-level pattern can reduce the length of the linearized knowledge and achieve higher representation density. Higher representation density means more knowledge facts can be included under the same length limitation.

Generation. The goal is to generate the linearized contextually-relevant knowledge sequence \(\omega _K(K_i^G)\). We adopt a widely-used auto-regressive GPT-2 [27] to implement \(P_{GenSel}(K^G_i|H_i,K_i)\) and generate the \(\omega _K(K_i^G)\):

$$\begin{aligned} \omega _K(K_i^G) = GPT2(S_i)=GPT2([\omega _K(K_i),\omega _H(H_i),Prompting]) \end{aligned}$$
(6)

In the training stage, we use a weakly-supervised way [51, 52] to construct the generation goal \(K_i^G\). Given a knowledge candidate set \( K_i\) retrieved by IR tools, if there is a knowledge candidate \(k \in K_i\) whose head entity and tail entity appear in the dialogue history \(H_i\) and the dialogue response \(R_i\), respectively; then this k is added to the target \(K^G_i\).

During the generation, \(K^G_i\) is fully generated based on the \(K_i\) and \(H_i\). Intuitively, the generated \(K^G_i\) can select the relevant knowledge from \(K_i\). Besides, as a generative model, GPT2 can also extend to generate the relevant knowledge that is not included in the \(K^G_i\), which is an inherent feature of generative language models [9]. Meanwhile, the generated \(\omega _K(K_i^G)\) is a natural language text, which can explicitly explain the results of knowledge selection and extension.

2.4 Dialogue Response Generation

Finally, we use the same GPT2 to generate the dialogue response \(R_i\) based on the dialogue history \(H_i\), the retrieved knowledge \(K_i\), and the generated contextually-relevant knowledge \(K^G_i\). We feed the input \(S_i^{DG}\) to the GPT2, estimate \(P_{ResGen}(r_t|R_{<t},H_i,K_i,K^{G}_{i})\), and then generate the \(R_i\):

$$\begin{aligned} \begin{matrix} S_i^{DG} = [\omega _K(K_i),\omega _H(H_i),Prompting,\omega _K(K_i^G)] \\ \\ R_i = GPT2(S_i^{DG}) = GPT2([\omega _K(K_i),\omega _H(H_i),Prompting,\omega _K(K_i^G)]) \end{matrix} \end{aligned}$$
(7)

where the generation head \(\mathbf {W^R}\) is newly introduced compared to Eq. 6. This is because two generative procedures have different generation spaces, two separate generation heads help avoid confusion. Such a two-head generation mechanism is called Dual-Head Generator.

2.5 Training

Two generative procedures can be jointly trained in an end-to-end fashion by sharing the same GPT-2. We have pre-trained two different GPT-2 models and our SEG-CKRG on two Nvidia RTX-3090 GPUs:

General GPT2: The general-purpose or dialogue-oriented base sizeFootnote 2 Chinese GPT-2 resources are not very abundant [31]. Consequently, we first pre-train a Chinese GPT2 for our experiments. We implement a base size GPT2 language model network using the Huggingface transformer libraryFootnote 3 and PyTorch. There are 12 layers of 768-dimensional (for both the hidden states and embeddings) and 12-head Transformer layers. The vocabulary includes 30,000 subwords and 200 special symbols (placeholders). For efficiency, the maximum input length is limited to 512 tokens. This GPT-2 is first pre-trained on massive Chinese unsupervised data, including massive open-released news, movie/product comments, and Wikipedia data. In total, there are 18.4M sessions and 5.22B tokens. During the training, the batch size is 512, the number of total training steps is 80,000, and the optimizer is AdamW. After 4,000 warm-up steps, the learning rate will reach 2e−4; then, the learning rate will linearly decay to 0.

Dialogue-Oriented GPT2: We also fine-tune a dialogue-oriented GPT-2. We use the Chinese conversational pre-training corpus LCCC-large released by [37], which includes 7.2M/4.7M sessions of single/multi-turn dialogues and 380M tokens in total. This GPT-2 is initialized from our general GPT-2, the batch size is 512, the number of total training steps is 180,000, and the optimizer is AdamW. It has the same learning rate strategy as GPT-2, except for the highest learning rate is decreased to 1.5e−4.

SEG-CKRG : Finally, SEG-CKRG is fine-tuned on the general GPT2 (by default) or the dialogue-oriented GPT2 (in ablation study). The batch size is set to 32, the maximum training epoch is set to 15. The best epoch on the validation set is adopted in the following test stage.

3 Experiment

3.1 Settings

Dataset. We test models on a Chinese dataset Weibo-ConceptNet [41], which has been aligned to a well-known commonsense knowledge base ConceptNet [32]. The training/validation/test set includes 102K/5.6K/5.6K single-turn dialogues. Each utterance has 10.3 words on average. The commonsense graph has 696,466 facts, 27,189 entities, and 26 relations. On average, each dialogue has 77.7 candidate facts that are retrieved from ConceptNet.

Comparison Models. We first selected several non-PLM baselines: 1) Seq2Seq: an attentive Seq2Seq RG Model [3, 24]; 2) PGN: Seq2Seq + Pointer-Genetor copy network [29]; 3) ConKADI: a KRG model with the felicitous knowledge selection mechanism [41]; 4) GOKC: a KRG model with a novel knowledge copy mechanism [1]. We also selected several fine-tuned base-size PLM methods: 5) BERT2Seq, 6) BERT-PGN: We changed the encoder of Seq2Seq and PGN to the ‘hfl/chinese-bert-wwm-ext’ [5] BERT encoder [6]. 7) CDial-GPT2: An open-released conversational GPT-2 RG models [37]. We select the GPT-2 configuration ‘GPT2LCCC-base’. 8) MHKD-GPT2: A PLM-based KRG models [42], which is based on CDial-GPT2.

Implementation. We use the official codes for ConKADI, GOKC, CDial-GPT2, and MHKD-GPT2, and we re-implement the remaining models using PyTorch. For non-PLM models, we use a 2-layer 768d bi-GRU/LSTMFootnote 4 encoder, 2-layer 768d GRU/LSTM decoder, Adam optimizer, 1e-4 learning rate. For all baselines, we use 32 batch size, up to 20 epochs, and finally select the best model on the validation set. Due to the different requirements, BERT2Seq, BERT-PGN use the corresponding BERT tokenizer and vocab, Seq2Seq, PGN, GOKC, and ConKADI use the original tokenizer and vocab, CDial-GPT2 and MHKD-GPT2 use CDial-GPT2’s tokenizer and vocab. The implementation of our SEG-CKRG will be released at https://github.com/pku-sixing/DASFAA23_GenSel.

Automatic Evaluation Metrics. Different models use different tokenizers; thus, we conduct character-level evaluations to avoid such differences. We use the following automatic metrics: 1) F1: it is the F-measure of character-overlapping relevance [1]; 2) BLEU-4: it is the 4-gram BLEU to evaluate the precision-oriented relevance [25]; 3) ROUGE: we use ROUGE-L to evaluate the recall-oriented relevance [18]; 4) EM-A/G/X: we use embedding evaluate the semantic relevance, the embedding is computed using Average/Greedy/Extrema [21]; 5) DI-1/2: we use Distinct-1/2 to evaluate the diversity [12]; 6) Ent: we use 4-gram entropy to evaluate the informativeness [30]; 7) Mean: following [42], we compute the geometric mean of all previous scores to evaluate the overall performance.

3.2 Automatic Evaluation

Table 1. Automatic Evaluation Results. First/Second denotes the first/second best.
Full size table

As reported in Table 1, SEG-CKRG has achieved tier-1 results (the first and the second best) in all metrics and significantly outperformed previous methods in the Mean score, demonstrating the best overall performance and effectiveness. In addition, rather than pursuing the best score on a single-dimensional metric or only using some handpicked metrics, the philosophy of SEG-CKRG is multi-dimensional because a single automatic metric is not reliable [21].

Relevance: In the three overlapping-based metrics (i.e., F1, BLEU-4, and ROUGE), SEG-CKRG has the best results because our approach can simultaneously seek information from both the pre-trained language model and the external knowledge source to help the dialogue generation. In another three embedding-based relevance metrics (i.e., EMB-A/G/X), SEG-CKRG also has the best overall performance, showing the dialogue responses generated by our approach are more semantically relevant to the ground truth. Besides, we can also find that PLM-based models have better relevance performance than non-PLM-based models in the mass. Indicating the necessity of using PLMs in CKRG.

Diversity and Informativeness: The situation is different in this part. ConKADI and our SEG-CKRG notably surpass other models. Between such two models, SEG-CKRG is slightly lower than ConKADI, and the reason can be summarized as 1) SEG-CKRG does not sacrifice the relevance to improving diversity and informativeness; 2) SEG-CKRG does not use any copy mechanism. Copy mechanism can copy words from the dialogue history or the external knowledge directly, which can significantly boost diversity and informativeness in the automatic evaluation. For example, compared with Seq2Seq/BERT2Seq, the copy variant PGN/BERT-PGN has more notable improvements in such metrics. However, we find previous copy works tend to repeat the given query rather than extend the new information, and then we decide not to equip this mechanism.

Table 2. Human Annotation Results. Scores denotes SEG-CKRG is significantly better (sign-test, p-value < 0.005). The 2/3 agreement ratio (at least 2 judges gave the same) is 95.4%, the 3/3 ratio is 54.8.2%.
Full size table

3.3 Human Evaluation

We employed three well-educated native-speaker to evaluate the practical generation quality of SEG-CKRG. The criteria include three dimensions: 1) Fluency: is this response grammatically correct and fluent? 2) Rationality: does this response logically conform to the current dialogue context? 3) Informativeness: can this response provide enough meaningful information?

As reported in Table 2, we sampled 100 comparison casesFootnote 5 and compared SEG-CKRG with the three best baselines in the automatic evaluation (ConKADI, BERT-PGN, and GOKC) and the naive Seq2Seq. We have several findings: 1) Although Seq2Seq is the naive baseline, the comparison result is not the worst, especially in terms of fluency. This is because the task and the network of Seq2Seq are simple but stable; 2) Compared to GOKC and ConKADI, SEG-CKRG has notable advantages, indicating the importance of introducing the PLMs to CKRG; 3) Compared to BERT-PGN, SEG-CKRG is still better, demonstrating the effectiveness of using external knowledge. Finally, we also compare SEG-CKRG with the human-generated ground-truth response. SEG-CKRG is comparable to the human in terms of fluency. However, SEG-CKRG is still behind the human in terms of rationality and informativeness. This shows we still have a large room to improve CKRG in future works.

Table 3. Generated knowledge types. # is the average counting per response.
Full size table

3.4 More Analyses

Generated Knowledge. As reported in Table 3, we counted the types among the knowledge generated by our Generative Knowledge Selection (GenSel). The original dataset has 77.7 facts per dialogue, but our model could only use 52.5 randomly sampled facts because of the restriction of the backbone GPT-2. We can find GenSel generated 1.282 facts per response on average, and about 9.7% of them are extended knowledge. It indicates 1) GenSel can indeed extend the knowledge and 2) using IR tools to explicitly retrieve knowledge candidates is still necessary.

Table 4. Ablation Study.
Full size table

Backbone GPT2. We have pre-trained two GPT-2 models; the first general GPT-2 is pre-trained on the general unsupervised data, and the second dialogue-oriented GPT-2 is further fine-tuned on a large-scale conversational dataset [37]. As shown in Table 4, if we train a SEG-CKRG based on the second dialogue-oriented GPT-2 (#1), although the perplexity is lower (9.93 vs 11.04), the real performance is even worse than the SEG-CKRG trained on the general GPT-2 (#0). We think our GenSel prefers the general backbone. Meanwhile, we also train a SEG-CKRG without initializing from a pre-trained GPT-2 (#2). The results are worse in every dimension, indicating the necessity of pre-training.

Knowledge Accessing. As shown in Table 4, we first remove the usage of knowledge generated by Generative Knowledge Selection (GenSel) (#3), and the performance is worse. It verifies GenSel can indeed help SEG-CKRG to select the existing knowledge and to extend to generate new relevant knowledge. Subsequently, similar to [52], we remove the usage of retrieved knowledge and try to externalize the implicit knowledge inherent in the backbone PLM (#4). It can be seen that the performance of (#4) is significantly worse than the original model. Thus, we think it is still necessary to explicitly access external knowledge even in the era of PLMs. Finally, we discard all knowledge usages and purely fine-tune two RG models based on our general GPT-2 (#5) and dialogue-oriented GPT-2 (#6). Both (#5) and (#6) are worse than previous model variants, showing the value of modeling dialogue response with external knowledge. We also find if we do not use any external knowledge, the dialogue-oriented GPT-2 is better than the general GPT-2, which differs from the previous (#0 vs #1). It verifies SEG-CKRG prefers to general GPT-2 again. In addition, (#2) and (#5) have similar overall performance, showing that explicit commonsense knowledge and implicit PLM knowledge play similar importance in our SEG-CKRG.

Table 5. Case Study. For SEG-CKRG, we report the generated knowledge (K:) and response (R:). We use color to highlight the knowledge.
Full size table

3.5 Case Study

We report two cases in Table 5 In the first case, SEG-CKRG generated two commonsense knowledge facts. We can find one knowledge fact is selected from the retrieved knowledge candidate set, but the other does not exist in the retrieved knowledge candidate set. This case verifies that our Generative Knowledge Selection mechanism can not only select knowledge but also extend the new knowledge. Subsequently, we can find the response generated by SEG-CKRG is highly relevant to the dialogue history and such generated knowledge facts, demonstrating the important role of the generated knowledge. Besides, compared to baselines, SEG-CKRG is the only one that can keep fluency, rationality, and informativeness at the same time. In the second case, we can find although ConKADI, BERT-PGN, and SEG-CKRG tried to introduce the new knowledge concept phone, only SEG-CKRG answered a proper response.

Table 6. Error Case. Same notations as the Case Study.
Full size table

Limitations. Although SEG-CKRG has surpassed all baselines, we also find a limitation in the current work, i.e., Error Propagation. SEG-CKRG sequentially generates the selected/extended knowledge and the dialogue response. Thus, if irrelevant knowledge has been generated in the first knowledge generation procedure, the next response generation procedure will be impacted. We report two typical error cases in Table 6. In the first case, SEG-CKRG generated a new but incorrect knowledge fact ‘(game, Synonym, play)’ in the knowledge generation procedure. SEG-CKRG wrongly predicted the relation between ‘game’ and ‘play’, where the correct relation should be ‘CapableOf’. But fortunately, this level of error has little impact on the following dialogue response generation. SEG-CKRG still generated a better response than other baselines. In the next case, SEG-CKRG has generated an existing but contextually-irrelevant knowledge fact. This error has significantly impacted the relevance of the generated response. Without considering the dialogue query, the response generated by SEG-CKRG is still fluent.

4 Related Work

Knowledge-Grounded Response Generation (KRG): Due to the inability to access enough knowledge, traditional RG models [33, 35] always generate safe but boring responses in spite of the given query [12, 13]. Consequently, KRG models try to solve this issue by accessing the external knowledge bases [20, 36, 48]. Commonsense knowledge is a popular knowledge type in the current research [32, 51], which helps a model to understand the dialogue, extend the topic, and then generate informative responses [38, 40, 41, 44, 46, 50].

Pretrained Language Models (PLMs): PLMs such as BERT [6], RoBERTa [22], GPTs [2, 27], and BARTs [11] have shown the dominate advantages in many NLP tasks [16]. PLMs can transfer the knowledge learned from massive unsupervised corpus to the open-domain RG models and bring significant improvements [8, 31, 37, 47]. As for KRG models, previous works have shown PLMs can further prompt the text knowledge-grounded dialogue response generation [4, 49] and the commonsense knowledge-grounded dialogue response generation [52].

Knowledge Selection: It is a research focus in KRG [7]. Non-PLM KRG models often adopt specific modules to conduct this job. [50], and [46] adopt graph neural networks, [41] uses the posterior response to help the learning of knowledge selection, [20], and [1] introduces copy networks to select the knowledge, [10] proposes a sequential knowledge selection paradigm, [28] proposes a global-to-local paradigm. In the era of PLMs, most knowledge selection methods that are originally designed for non-PLM KRG models become incompatible due to the restrictions of PLMs. Thus, the knowledge selection can only rely on the self-attention implemented by the Transformers of PLMs [23] or use the external module [49]. Meanwhile, such works can only select knowledge from a fixed and limited knowledge space, and the selection procedure is not very transparent. Different from such works, SEG-CKRG proposes a PLM-friendly Generative Knowledge Selection mechanism, which regards knowledge selection as a generative problem. Thus, our method can not only select the existing knowledge but also extend the new knowledge. Another difference is our work can explain the selection result using the human understandable natural language. In addition, although TBS [52] uses a PLM to generate knowledge, it does not include any knowledge selection procedure. The methodology of TBS is similar to our model (#4) in Table 4. Please refer to the corresponding results.

5 Conclusion

We propose an end-to-end CKRG model SEG-CKRG. Unlike previous works that can only use the limited and fixed knowledge retrieved by IR tools, SEG-CKRG introduces a novel Generative Knowledge Selection (GenSel) mechanism to select existing knowledge and extend new knowledge in a generative way. More importantly, the knowledge selection/extension procedure has higher interpretability than previous works because the output is a natural language text. SEG-CKRG is implemented based on two Chinese GPT-2s pre-trained by ourselves. Finally, experimental results have shown the very competitive performance of SEG-CKRG.

Our future work includes three directions. First, we will continue to address the mentioned limitation; Second, we will explore and verify the effectiveness of GenSel in more different types of knowledge, such as text-based and table-based knowledge; Third, we are considering jointly modeling the CKRG task and the conversational relation extraction simultaneously by extending the potential of GenSel.