Source-linked AI summary
Language Models that Seek for Knowledge: Modular Search & Generation for Dialogue and Prompt Completion
Kurt Shuster, Mojtaba Komeili, Leonard Adolphs, Stephen Roller, Arthur Szlam, Jason Weston
TL;DR
Language models can be fluent yet factually wrong and lack information newer than their training data. SeeKeR chains internet search, knowledge generation, and response generation in one language model, outperforming BlenderBot 2, GPT2, and GPT3 on the reported dialogue and topical-completion evaluations. Its modular design also permits separate engineering of modules, though it retains inconsistency, factuality, toxicity, bias, and retrieved-content risks.
Problem
Language models can generate factually incorrect statements and cannot learn facts newer than their fixed training data.
Method
SeeKeR applies one transformer successively to search, knowledge generation, and final response generation, passing each module’s output forward.
Results
SeeKeR outperforms BlenderBot 2 in dialogue and GPT2 and GPT3 in factuality and topicality on topical prompt completions.
Takeaways & Limitations
The three-module design supports incorporating current information and allows future engineering of individual modules while improving interpretability.
Takeaways & Limitations
SeeKeR retains occasional inconsistency, contradictions, factual inaccuracies, repetition, shallow reasoning, toxicity, bias, and risks from undesirable retrieved content.
Abstract
from arXiv · showhide
Language models (LMs) have recently been shown to generate more factual responses by employing modularity (Zhou et al., 2021) in combination with retrieval (Adolphs et al., 2021). We extend the recent approach of Adolphs et al. (2021) to include internet search as a module. Our SeeKeR (Search engine->Knowledge->Response) method thus applies a single LM to three modular tasks in succession: search, generating knowledge, and generating a final response. We show that, when using SeeKeR as a dialogue model, it outperforms the state-of-the-art model BlenderBot 2 (Chen et al., 2021) on open-domain knowledge-grounded conversations for the same number of parameters, in terms of consistency, knowledge and per-turn engagingness. SeeKeR applied to topical prompt completions as a standard language model outperforms GPT2 (Radford et al., 2019) and GPT3 (Brown et al., 2020) in terms of factuality and topicality, despite GPT3 being a vastly larger model. Our code and models are made publicly available.
1 Introduction
SeeKeR addresses factuality and freshness limitations in language models by chaining search, knowledge generation, and response generation. It improves human-rated dialogue and topical-completion performance against established baselines.
- Motivation: Standard language models can produce fluent but factually incorrect statements and cannot acquire facts newer than their training data.Increasing model size alone does not solve factual errors, while fixed training data limits knowledge freshness.
- SeeKeR approach: SeeKeR sequentially uses one transformer for search, knowledge generation, and final response, passing each module’s output to the next.The search module generates a query, the knowledge module selects relevant retrieved content, and the response module produces the final answer.
- Results: SeeKeR outperforms BlenderBot 2 in open-domain dialogue on human-rated consistency, knowledge, and per-turn engagingness.
- Results: SeeKeR outperforms GPT2 and GPT3 on factuality and topicality for prompts about January 2022 news unseen during training.
2 Related Work
Prior work combines modular knowledge generation or retrieval with language modeling, but existing systems generally lack both modular knowledge selection and continuously updated internet information.
- Modular retrieval: SeeKeR extends the two-stage knowledge-to-response approach by adding internet search and applying it to open-domain dialogue and general language modeling.
- Dialogue systems: BlenderBot 2 retrieves internet documents but generates responses directly rather than using a separate modular knowledge-generation stage.
- Dialogue systems: Prior retrieval systems can mix facts incorrectly or ignore retrieved knowledge, motivating SeeKeR’s explicit knowledge-generation step.
- Language modeling: Nearest-neighbor and cache-based language models retrieve from fixed databases, leaving them unable to reflect the latest knowledge and current events.
3 SeeKeR Model
SeeKeR reuses a standard transformer across sequential search, knowledge, and response modules, training those modules with dialogue and language-modeling tasks built from web and knowledge-intensive data.
- Architecture: The architecture invokes search, knowledge, and response modules sequentially, feeding each output forward alongside the original context.Special tokens indicate which module is active; the same encoder-decoder or decoder-only transformer is reused.
- Architecture: The search module generates a query and retrieves documents, while the knowledge module extracts relevant phrases or sentences from them.Dialogue experiments use the Bing Web Search API and retain the top five documents after filtering with Common Crawl.
- Architecture: The response module conditions on context plus extracted knowledge to generate a fluent continuation, simplifying the task compared with conventional sequence-to-sequence generation.
- Dialogue tasks: Dialogue training uses supervised search-query data from Wizard of Internet and multi-task knowledge targets from knowledge-grounded dialogue and question-answering datasets.
- Language-modeling tasks: For language modeling, search training predicts simplified document titles, while knowledge training retrieves similar sentences from Common Crawl for use in response prediction.
4 Experiments
Experiments evaluate SeeKeR on knowledge-grounded dialogue and topical prompt completion, including human judgments, automatic comparisons, ablations, and qualitative failure analyses. SeeKeR generally improves factuality, knowledge use, and topicality, while revealing trade-offs involving engagingness, search quality, and conversational behavior.
- 4.1.3 Human Evaluation Results: 94.7% versus 78–79% engagingness for knowledgeable turns, and 44% versus 21–28% for turns both knowledgeable and engaging, with statistically significant improvements over baselines.The reported improvements use an independent two-sample t-test with p < 0.001.
- 4.1.3 Human Evaluation Results: SeeKeR outperforms BlenderBot 1 and 2 on human-rated consistency, knowledge, factual correctness, and per-turn engagingness in open-domain dialogue.The comparison covers a wide variety of metrics in short human-evaluated conversations.
- 4.1.4 Ablations: SeeKeR still compares favorably after controlling BlenderBot pre-training, although language-model training may make knowledge use easier while reducing engagingness in an R2C2 BlenderBot 2 variant.The authors characterize this explanation as a possibility rather than a demonstrated mechanism.
- 4.1.5 Analysis: Qualitative analyses expose repetitive responses, fact recitation, ignoring the conversational partner, non-topical outputs, and hallucinations caused by mixing information from retrieved pages.Appending “January 2022” to search queries raises topicality from 15% to 19% and reduces hallucinations, underscoring the importance of search quality.
- 4.2.1 Automatic Evaluations: A single multi-task SeeKeR model performs similarly to separate search, knowledge, and response models, while improving Gold Docs performance as model size increases.The reported automatic evaluations cover perplexity and F1, with all-in-one models outperforming x3 models in F1 for the largest XL models.
5 Limitations & Discussion
SeeKeR retains limitations common to current language systems, including inconsistency, factual errors, repetition, shallow reasoning, toxic language, and bias. Its reliance on external search also means retrieved undesirable content can influence generations.
- SeeKeR can produce inconsistent, contradictory, factually inaccurate, repetitive, and shallow responses, while also exhibiting toxic language and bias in some contexts.
- External search keeps information current but can also bring undesirable internet content into generated responses.
6 Conclusion
SeeKeR separates search, knowledge selection, and response generation into modular stages, achieving strong results across dialogue and topical prompt completion while supporting future module-specific improvements.
- The three-module design enables future engineering improvements to individual components and makes the system more interpretable.The modules separately handle searching for and choosing knowledge during generation.
- SeeKeR outperforms state-of-the-art dialogue models and GPT-based baselines across knowledge-grounded dialogue and topical prompt completion.The conclusion reports stronger dialogue performance than the state of the art, and better topical prompt performance than GPT2 and GPT3.
- SeeKeR outperforms GPT3 on topical prompts despite using 500 times fewer parameters.
A Appendix: Additional Examples and Evaluations
The appendix provides illustrative prompt and dialogue examples alongside additional human evaluations, ablations, and automated comparisons of SeeKeR variants.
- SeeKeR produces correct topical continuations in examples where GPT3 hallucinates, although some examples show topicality limitations for both systems.
- Additional tables compare SeeKeR variants with GPT3, GPT2-related systems, BlenderBot models, and independently modularized alternatives.
- SeeKeR is rated more knowledgeable than several comparison models, while engagingness differences are not significant in the reported ACUTE-Eval results.
- The appendix includes a multi-tasked-versus-dialogue-tuned SeeKeR comparison on the WizInt validation set.
B.1 SeeKeR 2.7B R2C2 Model Architecture
The 2.7B SeeKeR R2C2 dialogue model uses a deep encoder-decoder architecture, while Figure 4 illustrates its search, knowledge, and response stages from the human user's perspective.
- The dialogue model has 22 encoder and 22 decoder layers, 2048-dimensional embeddings, 8192 hidden units, and 32 attention heads.
- Figure 4 distinguishes user messages, search queries, generated knowledge, and final dialogue responses with separate visual box colors.Human conversationalists saw only the final responses, not the intermediate search queries or generated knowledge.
B.2 SeeKeR 2.7B R2C2 Pre-training Hyperparameters
SeeKeR is pretrained with BART denoising and fine-tuned jointly across search, knowledge, and response tasks, using task-specific decoding procedures for each module.
- B.2 SeeKeR 2.7B R2C2 Pre-training Hyperparameters: The model is pretrained for 500,000 steps with a BART denoising objective on 128 V100 GPUs for approximately 25 days.
- B.2 SeeKeR 2.7B R2C2 Pre-training Hyperparameters: Fine-tuning jointly covers search, knowledge, and response tasks, with early stopping based on task performance.
- B.2 SeeKeR 2.7B R2C2 Pre-training Hyperparameters: Figure 5 presents the same staged interaction format in which human conversationalists see only the model's final responses.
- B.2 SeeKeR 2.7B R2C2 Pre-training Hyperparameters: Search uses greedy decoding, knowledge generation uses beam search with beam size 3, and response generation uses beam search with beam size 10 in evaluation settings.
C.1 Pre-training
SeeKeR’s base model was pretrained on a concatenation of RoBERTa, CC100EN, and Pushshift.io Reddit data, then fine-tuned across task-specific datasets using control tokens to select generation behavior.
- C.1 Pre-training: The base model combines RoBERTa, CC100EN, and Pushshift.io Reddit data for pre-training.The Reddit discussions are flattened into one conversation per tree node before conversation-level denoising.
- C.1 Pre-training: Approximately 100B tokens from RoBERTa and English CC100 corpora are tokenized with the GPT2 dictionary and denoised at the sentence level.
- C.1 Pre-training: Fine-tuning datasets are filtered for high-quality examples, including human-grounded responses and final human-entered search queries.
- C.1 Pre-training: Control tokens identify search, knowledge, and dialogue generation tasks during model training.Search uses __generate-query__; dialogue knowledge is delimited by __knowledge__ and __endknowledge__ tokens.
- C.1 Pre-training: Table 11 summarizes the training datasets and the number of examples used for each fine-tuning task.
D Human Evaluation Details
Human evaluation used crowdworker instructions and an annotation interface to support model-chatting and response annotation, with privacy and policy review conducted before launch.
- D Human Evaluation Details: Crowdworkers received instructions for chatting with models and annotating their responses, supported by a dedicated annotation screen.
- D Human Evaluation Details: The evaluation task was reviewed by privacy and policy experts and did not request workers’ personal information.
- D Human Evaluation Details: Workers were paid well above minimum wage for the crowdsourcing task.