Afleveringen
-
In this episode of Neural Search Talks, Andrew Yates (Assistant Prof at the University of Amsterdam) Sergi Castella (Analyst at Zeta Alpha), and Gabriel Bénédict (PhD student at the University of Amsterdam) discuss the prospect of using GPT-like models as a replacement for conventional search engines.Generative Information Retrieval (Gen IR) SIGIR Workshop
Workshop organized by Gabriel Bénédict, Ruqing Zhang, and Donald Metzler https://coda.io/@sigir/gen-ir Resources on Gen IR: https://github.com/gabriben/awesome-generative-information-retrievalReferences
Rethinking Search: https://arxiv.org/abs/2105.02274 Survey on Augmented Language Models: https://arxiv.org/abs/2302.07842 Differentiable Search Index: https://arxiv.org/abs/2202.06991 Recommender Systems with Generative Retrieval: https://shashankrajput.github.io/Generative.pdfTimestamps:00:00 Introduction, ChatGPT Plugins02:01 ChatGPT plugins, LangChain04:37 What is even Information Retrieval?06:14 Index-centric vs. model-centric Retrieval12:22 Generative Information Retrieval (Gen IR)21:34 Gen IR emerging applications24:19 How Retrieval Augmented LMs incorporate external knowledge29:19 What is hallucination?35:04 Factuality and Faithfulness41:04 Evaluating generation of Language Models47:44 Do we even need to "measure" performance?54:07 How would you evaluate Bing's Sydney?57:22 Will language models take over commercial search?1:01:44 NLP academic research in the times of GPT-41:06:59 Outro
-
Andrew Yates (Assistant Prof at University of Amsterdam) and Sergi Castella (Analyst at Zeta Alpha) discuss the paper "Task-aware Retrieval with Instructions" by Akari Asai et al. This paper proposes to augment a conglomerate of existing retrieval and NLP datasets with natural language instructions (BERRI, Bank of Explicit RetRieval Instructions) and use it to train TART (Multi-task Instructed Retriever).
📄 Paper: https://arxiv.org/abs/2211.09260
🍻 BEIR benchmark: https://arxiv.org/abs/2104.08663
📈 LOTTE (Long-Tail Topic-stratified Evaluation, introduced in ColBERT v2): https://arxiv.org/abs/2112.01488
Timestamps:
00:00 Intro: "Task-aware Retrieval with Instructions"
02:20 BERRI, TART, X^2 evaluation
04:00 Background: recent works in domain adaptation
06:50 Instruction Tuning 08:50 Retrieval with descriptions
11:30 Retrieval with instructions
17:28 BERRI, Bank of Explicit RetRieval Instructions
21:48 Repurposing NLP tasks as retrieval tasks
23:53 Negative document selection
27:47 TART, Multi-task Instructed Retriever
31:50 Evaluation: Zero-shot and X^2 evaluation
39:20 Results on Table 3 (BEIR, LOTTE)
50:30 Results on Table 4 (X^2-Retrieval)
55:50 Ablations
57:17 Discussion: user modeling, future work, scale
-
Zijn er afleveringen die ontbreken?
-
Marzieh Fadaee — NLP Research Lead at Zeta Alpha — joins Andrew Yates and Sergi Castella to chat about her work in using large Language Models like GPT-3 to generate domain-specific training data for retrieval models with little-to-no human input. The two papers discussed are "InPars: Data Augmentation for Information Retrieval using Large Language Models" and "Promptagator: Few-shot Dense Retrieval From 8 Examples".
InPars: https://arxiv.org/abs/2202.05144
Promptagator: https://arxiv.org/abs/2209.11755
Timestamps:
00:00 Introduction
02:00 Background and journey of Marzieh Fadaee
03:10 Challenges of leveraging Large LMs in Information Retrieval
05:20 InPars, motivation and method
14:30 Vanilla vs GBQ prompting
24:40 Evaluation and Benchmark
26:30 Baselines
27:40 Main results and takeaways (Table 1, InPars)
35:40 Ablations: prompting, in-domain vs. MSMARCO input documents
40:40 Promptagator overview and main differences with InPars
48:40 Retriever training and filtering in Promptagator
54:37 Main Results (Table 2, Promptagator)
1:02:30 Ablations on consistency filtering (Figure 2, Promptagator)
1:07:39 Is this the magic black-box pipeline for neural retrieval on any documents
1:11:14 Limitations of using LMs for synthetic data
1:13:00 Future directions for this line of research
-
Andrew Yates (Assistant Professor at the University of Amsterdam) and Sergi Castella (Analyst at Zeta Alpha) discus the two influential papers introducing ColBERT (from 2020) and ColBERT v2 (from 2022), which mainly propose a fast late interaction operation to achieve a performance close to full cross-encoders but at a more manageable computational cost at inference; along with many other optimizations.
📄 ColBERT: "ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT" by Omar Khattab and Matei Zaharia. https://arxiv.org/abs/2004.12832
📄 ColBERTv2: "ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction" by Keshav Santhanam, Omar Khattab, Jon Saad-Falcon, Christopher Potts, and Matei Zaharia. https://arxiv.org/abs/2112.01488
📄 PLAID: "An Efficient Engine for Late Interaction Retrieval" by Keshav Santhanam, Omar Khattab, Christopher Potts, and Matei Zaharia. https://arxiv.org/abs/2205.09707
📄 CEDR: "CEDR: Contextualized Embeddings for Document Ranking" by Sean MacAvaney, Andrew Yates, Arman Cohan, and Nazli Goharian. https://arxiv.org/abs/1904.07094
🪃 Feedback form: https://scastella.typeform.com/to/rg7a5GfJ
Timestamps:
00:00 Introduction
00:42 Why ColBERT?
03:34 Retrieval paradigms recap
08:04 ColBERT query formulation and architecture
09:04 Using ColBERT as a reranker or as an end-to-end retriever
11:28 Space Footprint vs. MRR on MS MARCO
12:24 Methodology: datasets and negative sampling
14:37 Terminology for cross encoders, interaction-based models, etc.
16:12 Results (ColBERT v1) on MS MARCO
18:41 Ablations on model components
20:34 Max pooling vs. mean pooling
22:54 Why did ColBERT have a big impact?
26:31 ColBERTv2: knowledge distillation
29:34 ColBERTv2: indexing improvements
33:59 Effects of clustering compression in performance
35:19 Results (ColBERT v2): MS MARCO
38:54 Results (ColBERT v2): BEIR
41:27 Takeaway: strong specially in out-of-domain evaluation
43:59 Qualitatively how do ColBERT scores look like?
46:21 What's the most promising of all current neural IR paradigms
49:34 How come there's still so much interest in Dense retrieval?
51:09 Many to many similarity at different granularities
53:44 What would ColBERT v3 include?
56:39 PLAID: An Efficient Engine for Late Interaction Retrieval
Contact: [email protected]
-
How much of the training and test sets in TREC or MS Marco overlap? Can we evaluate on different splits of the data to isolate the extrapolation performance?
In this episode of Neural Information Retrieval Talks, Andrew Yates and Sergi Castella i Sapé discuss the paper "Evaluating Extrapolation Performance of Dense Retrieval" byJingtao Zhan, Xiaohui Xie, Jiaxin Mao, Yiqun Liu, Min Zhang, and Shaoping Ma.
📄 Paper: https://arxiv.org/abs/2204.11447
❓ About MS Marco: https://microsoft.github.io/msmarco/
❓About TREC: https://trec.nist.gov/
🪃 Feedback form: https://scastella.typeform.com/to/rg7a5GfJ
Timestamps:
00:00 Introduction
01:08 Evaluation in Information Retrieval, why is it exciting
07:40 Extrapolation Performance in Dense Retrieval
10:30 Learning in High Dimension Always Amounts to Extrapolation
11:40 3 Research questions
16:18 Defining Train-Test label overlap: entity and query intent overlap
21:00 Train-test Overlap in existing benchmarks TREC
23:29 Resampling evaluation methods: constructing distinct train-test sets
25:37 Baselines and results: ColBERT, SPLADE
29:36 Table 6: interpolation vs. extrapolation performance in TREC
33:06 Table 7: interplation vs. extrapolation in MS Marco
35:55 Table 8: Comparing different DR training approaches
40:00 Research Question 1 resolved: cross encoders are more robust than dense retrieval in extrapolation
42:00 Extrapolation and Domain Transfer: BEIR benchmark.
44:46 Figure 2: correlation between extrapolation performance and domain transfer performance
48:35 Broad strokes takeaways from this work
52:30 Is there any intuition behind the results where Dense Retrieval generalizes worse than Cross Encoders?
56:14 Will this have an impact on the IR benchmarking culture?
57:40 Outro
Contact: [email protected]
-
Andrew Yates (Assistant Professor at the University of Amsterdam) and Sergi Castella i Sapé discuss the recent "Open Pre-trained Transformer (OPT) Language Models" from Meta AI (formerly Facebook). In this replication work, Meta developed and trained a 175 Billion parameter Transformer very similar to GPT-3 from OpenAI, documenting the process in detail to share their findings with the community. The code, pretrained weights, and logbook are available on their Github repository (links below).
Links
❓Feedback Form: https://scastella.typeform.com/to/rg7a5GfJ
📄 OPT paper: https://arxiv.org/abs/2205.01068
👾 Code: https://github.com/facebookresearch/metaseq
📒 Logbook: https://github.com/facebookresearch/metaseq/blob/main/projects/OPT/chronicles/OPT175B_Logbook.pdf
✍️ OPT Official Blog Post: https://ai.facebook.com/blog/democratizing-access-to-large-scale-language-models-with-opt-175b/
OpenAI Embeddings API: https://openai.com/blog/introducing-text-and-code-embeddings/
Nils Reimers' critique of OpenAI Embeddings API: https://medium.com/@nils_reimers/openai-gpt-3-text-embeddings-really-a-new-state-of-the-art-in-dense-text-embeddings-6571fe3ec9d9
Timestamps:
00:00 Introduction and housekeeping: new feedback form, ACL conference highlights
02:42 The convergence between NLP and Neural IR techniques
06:43 Open Pretrained Transformer motivation and scope, reproducing GPT-3 and open-sourcing
08:16 Basics of OPT: architecture, pre-training objective, teacher forcing, tokenizer, training data
13:40 Preliminary experiments findings: hyperparameters, training stability, spikiness
20:08 Problems that appear at scale when training with 992 GPUs
23:01 Using temperature to check whether GPUs are working
25:00 Training the largest model: what to do when the loss explodes? (which happens quite often)
29:15 When they switched away from AdamW to SGD
32:00 Results: successful but not quite GPT-3 level.
Toxicity? 35:45 Replicability of Large Language Models research. Was GPT-3 replicable? What difference does it make?
37:25 What makes a paper replicable?
40:33 Directions in which large Language Models are applied to Information Retrieval
45:15 Final thoughts and takeaways
-
We discuss Conversational Search with our usual cohosts Andrew Yates and Sergi Castella i Sapé; along with a special guest Antonios Minas Krasakis, PhD candidate at the University of Amsterdam.
We center our discussion around the ConvDR paper: "Few-Shot Conversational Dense Retrieval" by Shi Yu et al. which was the first work to perform Conversational Search without an explicit conversation to query rewriting step.
Timestamps:
00:00 Introduction
00:50 Conversational AI and Conversational Search
05:40 What makes Conversational Search challenging
07:00 ConvDR paper introduction
10:10 Passage representations
11:30 Conversation representations: query rewriting
19:12 ConvDR novel proposed method: teacher-student setup with ANCE
22:50 Datasets and benchmarks: CAsT, CANARD
25:32 Teacher-student advantages and knowledge distillation vs. ranking loss functions
28:09 TREC CAsT and OR-QuAC
35:50 Metrics: MRR, NDCG, holes@10
44:16 Main Results on CAsT and OR-QuAC (Table 2)
57:35 Ablations on combinations of loss functions (Table 4)
1:00:10 How fast is ConvDR? (Table 3)
1:02:40 Qualitative analysis on ConvDR embeddings (Figure 4)
1:04:50 How has this work aged? More recent works in similar directions: Contextualized Quesy Embeddings for Conversational Search.
1:07:02 Is "end-to-end" the silver-bullet for Conversational Search?
1:10:04 Will conversational search become more mainstream?
1:18:44 Latest initiatives for Conversational Search
-
Andrew Yates and Sergi Castella discuss the paper titled "Transformer Memory as a Differentiable Search Index" by Yi Tay et al at Google. This work proposes a new approach to document retrieval in which document ids are memorized by a transformer during training (or "indexing") and for retrieval, a query is fed to the model, which then generates autoregressively relevant doc ids for that query.
Paper: https://arxiv.org/abs/2202.06991
Timestamps:
00:00 Intro: Transformer memory as a Differentiable Search Index (DSI)
01:15 The gist of the paper, motivation
4:20 Related work: Autoregressive Entity Linking
7:38 What is an index? Conventional vs. "differentiable"
10:20 Indexing and Retrieval definitions in the context of the DSI
12:40 Learning representations for documents
17:20 How to represent document ids: atomic, string, semantically relevant
22:00 Zero-shot vs. finetuned settings
24:10 Datasets and baselines
27:08 Dinetuned results
36:40 Zero-shot results
43:50 Ablation results
47:15 Where could this model be useds?
52:00 Is memory efficiency a fundamental problem of this approach?
55:14 What about semantically relevant doc ids?
60:30 Closing remarks
Contact: [email protected]
-
In this third episode of the Neural Information Retrieval Talks podcast, Andrew Yates and Sergi Castella discuss the paper "Learning to Retrieve Passages without Supervision" by Ori Ram et al.
Despite the massive advances in Neural Information Retrieval in the past few years, statistical models still overperform neural models when no annotations are available at all. This paper proposes a new self-supervised pertaining task for Dense Information Retrieval that manages to beat BM25 on some benchmarks without using any label.
Paper: https://arxiv.org/abs/2112.07708
Timestamps:
00:00 Introduction
00:36 "Learning to Retrieve Passages Without Supervision"
02:20 Open Domain Question Answering
05:05 Related work: Families of Retrieval Models
08:30 Contrastive Learning
11:18 Siamese Networks, Bi-Encoders and Dual-Encoders
13:33 Choosing Negative Samples
17:46 Self supervision: how to train IR models without labels.
21:31 The modern recipe for SOTA Retrieval Models
23:50 Methodology: a new proposed self supervision task
26:40 Datasets, metrics and baselines
\33:50 Results: Zero-Shot performance
43:07 Results: Few-shot performance
47:15 Practically, is not using labels relevant after all?
51:37 How would you "break" the Spider model?
53:23 How long until Neural IR models outperform BM25 out-of-the-box robustly?
54:50 Models as a service: OpenAI's text embeddings API
Contact: [email protected]
-
We discuss the Information Retrieval publication "The Curse of Dense Low-Dimensional Information Retrieval for Large Index Sizes" by Nils Reimers and Iryna Gurevych, which explores how Dense Passage Retrieval performance degrades as the index size varies and how it compares to traditional sparse or keyword-based methods.
Timestamps:
00:00 Co-host introduction
00:26 Paper introduction
02:18 Dense vs. Sparse retrieval
05:46 Theoretical analysis of false positives(1)
08:17 What is low vs. high dimensional representations
11:49 Theoretical analysis o false positives (2)
20:10 First results: growing the MS-Marco index
28:35 Adding random strings to the index
39:17 Discussion, takeaways
44:26 Will dense retrieval replace or coexist with sparse methods?
50:50 Sparse, Dense and Attentional Representations for Text Retrieval
Referenced work:
Sparse, Dense and Attentional Representations for Text Retrieval by Yi Luan et al. 2020.
-
In this first episode of Neural Information Retrieval Talks, Andrew Yates and Sergi Castellla discuss the paper "Shallow Pooling for Sparse Labels" by Negar Arabzadeh, Alexandra Vtyurina, Xinyi Yan and Charles L. A. Clarke from the University of Waterloo, Canada.
This paper puts the spotlight on the popular IR benchmark MS MARCO and investigates whether modern neural retrieval models retrieve documents that are even more relevant than the original top relevance annotations. The results have important implications and raise the question of to what degree this benchmark is still an informative north star to follow.
Contact: [email protected]
Timestamps:
00:00 — Introduction.
01:52 — Overview and motivation of the paper.
04:00 — Origins of MS MARCO.
07:30 — Modern approaches to IR: keyword-based, dense retrieval, rerankers and learned sparse representations.
13:40 — What is "better than perfect" performance on MS MARCO?
17:15 — Results and discussion: how often are neural rankers preferred over original annotations on MS MARCO? How should we interpret these results?
26:55 — The authors' proposal to "fix" MS MARCO: shallow pooling
32:40 — How does TREC Deep Learning compare?
38:30 — How do models compare after re-annotating MS MARCO passages?
45:00 — Figure 5 audio description.
47:00 — Discussion on models' performance after re-annotations.
51:50 — Exciting directions in the space of IR benchmarking.
1:06:20 — Outro.
Related material:
- Leo Boystov paper critique blog post: http://searchivarius.org/blog/ir-leaderboards-never-tell-full-story-they-are-still-useful-and-what-can-be-done-make-them-even
- "MS MARCO Chameleons: Challenging the MS MARCO Leaderboard with Extremely Obstinate Queries" https://dl.acm.org/doi/abs/10.1145/3459637.3482011
