Attention with Dependency Parsing Augmentation for Fine-Grained Attribution

Lvzhou Luo; Ping Luo; Qiang Ding; Yixuan Cao

REVIEW 3 major objections 5 minor 2 cited by

Attention with Dependency Parsing Augmentation for Fine-Grained Attribution

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that fine-grained attribution for RAG answers can be made state-of-the-art by unioning token-level attention evidence and expanding target spans via dependency parsing.

desk verdict Attention + dependency-parsing set-union attribution is a plausible advance, but the SOTA claim needs a lexical-overlap baseline and error bars before it fully lands. read the letter →

arxiv 2412.11404 v1 pith:IWE72NPD submitted 2024-12-16 cs.CL cs.AI

Qiang Ding , Lvzhou Luo , Yixuan Cao , Ping Luo This is my paper · ORCID

classification cs.CLcs.AI

keywords fine-grainedattributionretrieval-augmentedgenerationattentionweightsdependencyparsingmodelinternalsevidencecitationfaithfulness

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proposes that fine-grained attribution—pinpointing, for each span of a retrieval-augmented answer, the supporting text in the retrieved documents—can be made substantially more accurate by two model-internal techniques. The first attributes each response token separately and then takes the union of its evidence sets, preserving token-level granularity instead of averaging it away. The second uses dependency parsing to expand the target span with the tokens that form its atomic fact, such as the subject, verb, and object, so the attributor can use context that a causal decoder cannot see from the target position. With averaged attention weights from a selected middle layer as the similarity score, the method reports a new state of the art on the QuoteSum and VERI-GRAN benchmarks.

What carries the argument

The machinery rests on a token-wise attribution set e(ri)={dj | w(ri,dj)>0}, built from a similarity matrix S in which each response token's score against each prompt token is thresholded at the top-k value (k=2). For a target span t, the evidence is the union ⋃_{ri∈t} e(ri), with scores summed and isolated evidence tokens removed if no other evidence token lies within τ=2 positions. The second mechanism DEP applies a dependency parse (via LAL-Parser) to the answer sentence, finds the closest verb ancestor of a target token, collects that verb's successors except punctuation and irrelevant coordinating constituents, and then augments the token's attribution with the union of those atomic-fact tokens' attributions. The similarity score S is the mean over heads of the attention weights at the layer ⌊L/2⌋+1 of the decoder, which the paper argues outperforms hidden-state cosine similarity and is much cheaper than gradient-based saliency.

What would settle it

Rerun ATTN UNION DEP on the same benchmarks after paraphrasing the retrieved documents so that no target span appears verbatim; if accuracy collapses, the reported gains are explained by near-verbatim span overlap instead of by the attention-plus-dependency mechanism.

Watch

Extended reading notes

Core claim

The central claim is that the two mechanisms are complementary and each produces a large gain. Token-wise union alone performs on par with the existing average-hidden-state method (HSSAVG); dependency augmentation alone also helps; but the combination lifts fine-grained attribution accuracy to 93.3 and 84.6 for Qwen2, and 94.0 and 78.2 for Llama2, on QuoteSum and VERI-GRAN respectively, outperforming the two published fine-grained attributors (CCI and HSSAVG) and a reported 2-shot GPT-4 baseline. The paper additionally claims the dependency-augmented union is faithful to the generator as measured by log-probability drops when the attributed evidence is removed, and that union aggregation makes the method faster because token-wise evidence is computed once and reused for every target span.

Load-bearing premise

The method assumes that a response token's mean attention to prompt tokens in a chosen middle layer encodes genuine semantic support, so that the top-k attended tokens are the right evidence; if attention is determined more by syntax, position, or attention sinks, the reported gains could be coincidental.

Editorial extensions

If this is right

The method establishes a new accuracy benchmark for fine-grained attribution on QuoteSum and VERI-GRAN, beating the previous best fine-grained attributors by roughly 13–17 percentage points.
Because token-wise evidence is computed once and reused across all spans of a response, fine-grained attribution becomes fast enough for interactive use, with per-span latency of 22.7 ms on QuoteSum versus 84.5 ms for HSSAVG and 2921.8 ms for CCI.
Dependency augmentation improves faithfulness of attribution to the generation process, approaching the oracle upper bound in the paper's log-probability-drop measurement.
The two techniques transfer to sentence-level attribution and to other similarity metrics, improving the citation quality of self-citation and attribute-then-generate pipelines.
Approximating attention with an open-source LLM preserves accuracy, so the method can attribute answers from black-box generators.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The dependency-expansion idea might be portable to other model-internal probes, such as hallucination detection, where a false claim usually hinges on one atomic fact rather than a whole sentence.
The paper's rule-based coordinating-constituent pruning is English-specific, but the attention-union core is language-neutral, so a learned or multilingual dependency parser could extend the same recipe to other languages with less hand-tuning.
Because evidence sets are computed once per response, the same machinery could enable interactive post hoc attribution, letting a user click any span of an already-generated answer without re-running the model.
The benchmarks contain mostly verbatim spans; if the method is applied to abstractive answers, the dependency enrichment may need to be paired with a paraphrase-aware similarity metric to keep accuracy.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First, the useful part: this is a solid, clearly written empirical paper with a genuinely new combination—token-wise set-union aggregation and dependency-parsing augmentation—applied to fine-grained attribution in RAG. The results are consistent across two backbones and two datasets, and the ablation showing that DEP helps with union but not with hidden-state averaging is a good piece of evidence that the mechanism is more than just adding context tokens. The latency numbers also give practitioners a reason to try it over gradient-based CCI.

The soft spot is real but not fatal. The fine-grained benchmarks use near-verbatim spans, and the paper provides no lexical-overlap baseline. That is a genuine gap, as the stress-test note says. If a simple Jaccard-overlap scorer gets 93% on QuoteSum, then the attention weights may be riding on surface matching, and the DEP gain could be just an artifact of expanding spans to include more words that appear verbatim in the passage. The paper concedes the verbatim nature of the benchmarks in Limitations, which is honest, but it does not resolve the confound. The faithfulness experiment is run on freshly generated answers, not the benchmark spans, so it does not certify the Table 1 results. The absence of error bars or significance tests on the main table is also a problem: differences of 5–10 points are large, but we have no sense of variance.

That said, the paper is not hiding anything. The limitations section is straightforward, the hyperparameter choices are validated, and the comparison to HSSAVG and CCI is fair and internally consistent. The concern that attention weights are biased by syntax or position is real, but the paper's own ablation—where DEP hurts HSSAVG—suggests the method is not simply adding overlapping tokens; if it were, hidden-state averaging should also benefit from an expanded span.

Bottom line: this paper deserves a serious referee. The method is new, the experiments are well-designed, and the claim is within reason, but the SOTA claim is conditional. I would ask a referee to require a lexical-overlap baseline (e.g., Jaccard, TF-IDF, or BM25 on the expanded span against passages) and uncertainty estimates before publication. The paper is a meaningful contribution to fine-grained attribution, and with that additional control, it would be a solid accept.

Referee Report

3 major / 5 minor

Summary. This paper proposes ATTN UNION DEP, a model-internals method for fine-grained attribution in RAG. For each response token, the top-k prompt tokens by averaged attention from a selected middle layer are collected as evidence; evidence sets are combined by union across target-span tokens, isolated tokens are removed, and dependency parsing is used to expand each target token to the clause containing its governing verb and related constituents. The method is evaluated on QuoteSum and VERI-GRAN against CCI and HSSAVG, on sentence-level attribution against self-citation and ATTR FIRST, on a faithfulness test using log-probability drops, and in a latency comparison. The main empirical claim is that ATTN UNION DEP sets a new state of the art in fine-grained attribution (Table 1).

Significance. The two proposed techniques are simple and generally applicable: set-union aggregation avoids recomputation over target spans, and dependency parsing supplies a linguistically motivated way to give decoder representations access to later context. The paper includes careful ablations, hyperparameter sensitivity plots, a cross-lingual transfer experiment, and an efficiency study, all of which are strengths. If the fine-grained results survive a control for lexical overlap, the method would be a practically useful and substantially faster attributor than existing gradient- or hidden-state based approaches.

major comments (3)

[§4.1, Table 1; §6 (Limitations)] The paper's central SOTA claim is not yet secured because the fine-grained benchmarks use near-verbatim target spans, as the authors concede in §6, and no lexical-overlap baseline is reported. ATTN UNION DEP expands each target token into a dependency clause (§3.3), so the passage-level score is accumulated over many tokens of the expanded span; for spans that are copied from a source passage, this score may largely reflect token overlap between the expanded span and each candidate passage. The reported gains over HSSAVG (e.g., 80.4→93.3 on Qwen2/QuoteSum, 77.1→94.0 on Llama2/QuoteSum) could therefore be explained by surface matching rather than by attention weights encoding semantic support. Please add baselines such as BM25, TF-IDF cosine, or exact n-gram overlap between the DEP-expanded span and each passage, and an ablation that replaces attention scores with uniform or random weights while keeping the DEP expansion and union aggregation identical.
[§3.3, §4.2.1] The DEP ablation does not isolate the contribution of the dependency expansion from the contribution of the underlying similarity metric. The comparisons against SENT COMP and against JinaBERT encoder features all use representational or lexical similarity signals; none controls for the expanded span itself. A direct control is to apply the same expansion and evidence aggregation while replacing the attention-based similarity S with a bag-of-words overlap score. Without such a control, the conclusion that 'recognizing atomic facts via DEP is more effective' (§4.2.1) conflates the dependency expansion with the chosen similarity metric and does not establish that attention weights provide the semantic support claimed for the method.
[§4.3, Table 3] The faithfulness experiment does not evaluate the benchmark spans used in Table 1. The protocol generates new answers with greedy decoding, applies CTI to identify context-sensitive tokens, and then measures log-probability drops after deleting the attributed passage. This is a reasonable causal check, but it does not show that the high Table 1 accuracy on QuoteSum/VERI-GRAN spans reflects causal support for those particular spans. Please either run the deletion test on the benchmark spans with the original generator or explicitly restrict the faithfulness claim to newly generated answers.

minor comments (5)

[§4.2.3, Table 2] The text states that UNION underperforms AVG without DEP, but Table 2 shows the opposite in three of the four comparisons: HSSUNION vs. HSSAVG are 80.0 vs. 80.4 (Qwen2/QuoteSum), 80.1 vs. 77.1 (Llama2/QuoteSum), 73.1 vs. 67.1 (Qwen2/VERI-GRAN), and 65.3 vs. 64.5 (Llama2/VERI-GRAN). Only the first comparison favors AVG, and by only 0.4 points. This sentence should be corrected and the surrounding interpretation revised.
[§4.1, Table 1] The fine-grained results are reported without error bars or significance tests. Because VERI-GRAN contains only 197 instances, a bootstrap confidence interval would help the reader judge whether differences of a few points (e.g., ATTN UNION vs. HSSAVG on Llama2/VERI-GRAN, 66.7 vs. 64.5) are meaningful.
[§3.4] The claim that attention-based similarity is 'superior' to hidden-state similarity is supported only by selected comparisons in Fig. 3; the adjacent-bar comparisons are not quantified or tested. Please report the relevant numbers and, if possible, a paired comparison.
[§4.1, Baselines] The GPT-4 baseline is cited from Phukan et al. (2024) and may use a different prompt template or evaluation protocol; this should be stated explicitly so that the reader does not treat the comparison as controlled.
[Appendix H] The Chinese synthetic dataset construction translates only the answers while keeping questions and passages in English. This is a reasonable first test, but the paper should note that it does not evaluate full Chinese RAG pipelines with Chinese documents.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is defined independently of the benchmark labels and its gains are measured against external human-annotated data.

full rationale

The paper's derivation chain is self-contained and does not reduce to its own inputs. ATTN UNION is defined by Eq. (1) as a top-k thresholded attention similarity with set-union aggregation over target tokens; DEP expands a target token to dependency-parse atomic-fact elements and sums their attention scores. Neither definition uses the human-annotated evidence labels, so the reported accuracy on QuoteSum and VERI-GRAN is a genuine external evaluation rather than a fitted quantity renamed as a prediction. Hyperparameters (k, tau, L*, W, and the HSS/attention layer choices) are selected on validation sets, which is standard practice and not circular. The faithfulness experiment in Sec. 4.3 is also externally grounded: it removes the attributed passage, regenerates, and measures log-probability drop against generated answers, with RANDOM and ORACLE controls. The paper cites prior methods by other author groups (Phukan et al., Qi et al., Cohen-Wang et al.), but these citations are not self-citations and their results are used as baselines, not as justification for the proposed mechanism. The limitation stated in Sec. 6 — that QuoteSum and VERI-GRAN attribute verbatim spans — is a real external-validity concern about lexical overlap confounding the SOTA claim, but it is not a circularity defect: the model's output is still computed from attention weights and dependency parses, not from the benchmark labels. No equation is equivalent to its input by construction, and no load-bearing premise is imported from self-citation. Therefore the appropriate finding is no significant circularity, score 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The method has three tuned hyperparameters (k, tau, L*) and relies on three unproved domain assumptions about attention informativeness, dependency-parse approximation of atomic facts, and benchmark label validity. No new theoretical entities are introduced.

free parameters (3)

k (top-k prompt tokens per response token) = 2
Selected by human evaluation on validation sets; controls how many prompt tokens count as evidence per response token (Sec 3.2, Appendix F).
tau (isolated-token removal threshold) = 2
Chosen on validation sets; removes evidence tokens isolated by at least tau from other evidence tokens (Sec 3.2, Appendix F).
L* (attention extraction layer) = floor(L/2)+1 (15 for Qwen2-7B, 17 for Llama2-7B)
Layer for attention weights selected for strong validation performance (Sec 3.4, Sec 4.1).

assumptions (3)

domain assumption Mean attention weights from one selected layer quantify semantic relevance between response tokens and prompt tokens.
Core similarity metric for evidence selection (Sec 3.4); faithfulness test in Sec 4.3 provides partial support but does not cover benchmark spans.
domain assumption Dependency parse structure approximates atomic facts: closest verb ancestor and its successors capture the supporting clause for a target token.
Applies LAL-Parser to approximate atomic fact elements (Sec 3.3); rule-based coordination handling in Appendix C.
domain assumption Human-annotated evidence in QuoteSum and VERI-GRAN is a valid ground truth for fine-grained attribution.
Evaluation depends on these labels (Sec 4.1); paper notes target spans are model-selected, not human-selected (Sec 6).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Attention with Dependency Parsing Augmentation for Fine-Grained Attribution." pith.science (2026). https://pith.science/paper/IWE72NPD

@misc{pith2026241211404,
  author       = {Pith},
  title        = {Pith review of: Attention with Dependency Parsing Augmentation for Fine-Grained Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IWE72NPD}},
  note         = {Machine review of arXiv:2412.11404}
}

read the original abstract

To assist humans in efficiently validating RAG-generated content, developing a fine-grained attribution mechanism that provides supporting evidence from retrieved documents for every answer span is essential. Existing fine-grained attribution methods rely on model-internal similarity metrics between responses and documents, such as saliency scores and hidden state similarity. However, these approaches suffer from either high computational complexity or coarse-grained representations. Additionally, a common problem shared by the previous works is their reliance on decoder-only Transformers, limiting their ability to incorporate contextual information after the target span. To address the above problems, we propose two techniques applicable to all model-internals-based methods. First, we aggregate token-wise evidence through set union operations, preserving the granularity of representations. Second, we enhance the attributor by integrating dependency parsing to enrich the semantic completeness of target spans. For practical implementation, our approach employs attention weights as the similarity metric. Experimental results demonstrate that the proposed method consistently outperforms all prior works.

Figures

Figures reproduced from arXiv: 2412.11404 by the authors.

**Figure 2.** An illustration of dependency parsing augmentation. Suppose the target span is the token “one”. The [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Results of ablating DEP. Here AU and HU represent ATTNUNION and HSSUNION, respectively. Model QuoteSum VERI-GRAN Baselines 2-SHOT GPT-4† 90.6 62.1 HSSAVG Qwen2 80.4 67.1 Llama2 77.1 64.5 Llama2† 87.5 77.3 CCI Qwen2 71.3 64.5 Llama2 72.2 59.0 Our Methods ATTNUNION Qwen2 79.4 70.9 Llama2 81.3 66.7 ATTNUNIONDEP Qwen2 93.3 84.6 Llama2 94.0 78.2 [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: An illustration of reforming the coordinate [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

**Figure 5.** Figure 5: The validation accuracy of ATTNUNIONDEP against the layer from which the attention weights are extracted (fixing k = 2, τ = 2). For Qwen2 7B, L = 28, and ⌊L/2⌋+ 1 = 15. For Llama2 7B, L = 32, , and ⌊L/2⌋ + 1 = 17. 1 2 3 4 5 10 k 80 90 100 Accuracy Qwen on QuoteSum Qwen…

**Figure 6.** Figure 6: The validation accuracy of ATTNUNIONDEP against k (fixing τ = 2, L ∗ = 15 and 17 for Qwen2 7B and Llama2 7B, respectively). 1 2 3 4 5 80 90 100 Accuracy Qwen on QuoteSum Qwen on VERI-GRAN Llama on QuoteSum Llama on VERI-GRAN [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]

**Figure 7.** Figure 7: The validation accuracy of ATTNUNIONDEP against τ (fixing k = 2, L ∗ = 15 and 17 for Qwen2 7B and Llama2 7B, respectively). Here, τ = ∞ means no filtering out isolated evidence tokens. H Evaluating Attribution on Chinese Synthetic Datasets We conducted experiments on C…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

LAQuer: Localized Attribution Queries in Content-grounded Generation
cs.CL 2025-06 conditional novelty 6.0 of 10

LAQuer defines user-initiated, span-level attribution for grounded generation and shows it can cut the text users must read to verify a claim by about two orders of magnitude, at the cost of lower attribution accuracy.
TokenShapley: Token Level Context Attribution with Shapley Value
cs.CL 2025-06 conditional novelty 5.0 of 10

TokenShapley computes token-level Shapley attributions from context to response by treating context tokens as (prefix, token) data points in a KNN datastore.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages · cited by 2 Pith papers

[1]

The coordi- nating structures are searched by enumerating the potential leaders (shown as follows)

Identify coordinating structures. The coordi- nating structures are searched by enumerating the potential leaders (shown as follows). def find_coordinations ( dep_head, dep_label ): coordinations = [] in_coordination_words = set() for j in range (len(dep_head)-1): if j in in_coordination_words: continue new_coordination = [j] for k in range (j+1, len(dep_...

work page
[2]

For the convenience of the following process, we temporarily reform the local structures for all coordinate structures, as Fig

Reform the tree. For the convenience of the following process, we temporarily reform the local structures for all coordinate structures, as Fig. 4 shows. We replace the heads of non-leader com- ponents with the head of the leader, ending the asymmetric relationship between the leader and the other components. For other children of the leader, we retain it...

work page
[3]

Identify the path from v to ri

work page
[4]

In Findings of the Association for Compu- tational Linguistics ACL 2024, pages 11481–11495

Peering into the mind of language models: An approach for attribution in contextual question an- swering. In Findings of the Association for Compu- tational Linguistics ACL 2024, pages 11481–11495. Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun, Sean Welleck, Yejin Choi, and Zaid Harchaoui. 2021. Mauve: Measuring the gap be- tween neu...

arXiv 2024
[5]

one million dollars,

Process Coordinating Structures that do not intersect with v → ri. For non-intersecting coor- dinating structures G, the algorithm first determines whether there is a parallel coordinating structure of it, where parallel coordinating structures are those that have the same number of constituents, e.g., (“one million dollars,” “two million dollars”) and (“...

work page 2012
[6]

""stage 1: do not output attention, output kv cache on prompt_ids

Recollect. The algorithm recollects all v’s suc- cessors (except punctuation marks) with the new tree, yielding the final A(ri). D Details of Huggingface Implementation of ATTN UNION def forward_stage1 (self, batch): """stage 1: do not output attention, output kv cache on prompt_ids """ outputs = self.model( input_ids=batch[’prompt_ids’][:,:-1].to( self.d...

work page
[8]

For coordinating structures that intersect with v → ri, the tree retains the intersec- tion and removes all other components

Process Coordinating Structures that inter- 13 sect with v → ri. For coordinating structures that intersect with v → ri, the tree retains the intersec- tion and removes all other components

work page
[9]

Here we name the alternative AUGMENT BY- ATTN. We also evaluate a variant of AUGMENT - BYATTN that limits the augmentation tokens in the target span’s local sentence, to ensure the augmen- tation tokens are more relevant to the target span. As the results show, AUGMENT BYATTN and its variant improve ATTN UNION but are not as good as dependency parsing aug...

work page

Show all 17 references

[11]

L∗: the layer to extract attention weights
[12]

k: how many top-scored prompt tokens are selected as evidence for each response token
[13]

τ: the threshold for recognizing isolated to- kens. The experiments are conducted on validation sets of QuoteSum and VERI-GRAN, with the met- ric of accuracy (%), the attributor of ATTN UNION - DEP, and the models of Qwen2 7B and Llama 7B. The results are shown in Fig. 5, 6, a...
[14]

segmenting the answer by the boundaries of target spans
[15]

separately translating each segment into Chi- nese
[16]

2012” might be attributed to “one mil- lion dollars

concatenating all translated segments to build the translated answer. We adapt the dependency parsing augmentation to Chinese without modifying the rules (except for mapping Chinese dependency parsing labels to the English counterpart). We compare HSSAVG, AT- TNUNION , and ATT...

2012
[2020]

In Findings of the Asso- ciation for Computational Linguistics: EMNLP 2020, pages 731–742

Rethinking self-attention: Towards inter- pretability in neural parsing. In Findings of the Asso- ciation for Computational Linguistics: EMNLP 2020, pages 731–742. Dor Muhlgay, Ori Ram, Inbal Magar, Yoav Levine, Nir Ratner, Yonatan Belinkov, Omri Abend, Kevin Leyton-Brown, Amn...

2020
[2022]

arXiv preprint arXiv:2203.11147

Teaching language models to support answers with verified quotes. arXiv preprint arXiv:2203.11147. Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettle- moyer, and Hannaneh Hajishirzi. 2023. FActScore: Fine-grained atomic evaluatio...

2023 arXiv
[2024]

In Proceedings of the 18th Conference of the European Chapter of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 49–66

Generating benchmarks for factuality evalua- tion of language models. In Proceedings of the 18th Conference of the European Chapter of the Associa- tion for Computational Linguistics (Volume 1: Long Papers), pages 49–66. Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, ...

2021 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

The coordi- nating structures are searched by enumerating the potential leaders (shown as follows)

Identify coordinating structures. The coordi- nating structures are searched by enumerating the potential leaders (shown as follows). def find_coordinations ( dep_head, dep_label ): coordinations = [] in_coordination_words = set() for j in range (len(dep_head)-1): if j in in_coordination_words: continue new_coordination = [j] for k in range (j+1, len(dep_...

work page

[2] [2]

For the convenience of the following process, we temporarily reform the local structures for all coordinate structures, as Fig

Reform the tree. For the convenience of the following process, we temporarily reform the local structures for all coordinate structures, as Fig. 4 shows. We replace the heads of non-leader com- ponents with the head of the leader, ending the asymmetric relationship between the leader and the other components. For other children of the leader, we retain it...

work page

[3] [3]

Identify the path from v to ri

work page

[4] [4]

In Findings of the Association for Compu- tational Linguistics ACL 2024, pages 11481–11495

Peering into the mind of language models: An approach for attribution in contextual question an- swering. In Findings of the Association for Compu- tational Linguistics ACL 2024, pages 11481–11495. Krishna Pillutla, Swabha Swayamdipta, Rowan Zellers, John Thickstun, Sean Welleck, Yejin Choi, and Zaid Harchaoui. 2021. Mauve: Measuring the gap be- tween neu...

arXiv 2024

[5] [5]

one million dollars,

Process Coordinating Structures that do not intersect with v → ri. For non-intersecting coor- dinating structures G, the algorithm first determines whether there is a parallel coordinating structure of it, where parallel coordinating structures are those that have the same number of constituents, e.g., (“one million dollars,” “two million dollars”) and (“...

work page 2012

[6] [6]

""stage 1: do not output attention, output kv cache on prompt_ids

Recollect. The algorithm recollects all v’s suc- cessors (except punctuation marks) with the new tree, yielding the final A(ri). D Details of Huggingface Implementation of ATTN UNION def forward_stage1 (self, batch): """stage 1: do not output attention, output kv cache on prompt_ids """ outputs = self.model( input_ids=batch[’prompt_ids’][:,:-1].to( self.d...

work page

[7] [8]

For coordinating structures that intersect with v → ri, the tree retains the intersec- tion and removes all other components

Process Coordinating Structures that inter- 13 sect with v → ri. For coordinating structures that intersect with v → ri, the tree retains the intersec- tion and removes all other components

work page

[8] [9]

Here we name the alternative AUGMENT BY- ATTN. We also evaluate a variant of AUGMENT - BYATTN that limits the augmentation tokens in the target span’s local sentence, to ensure the augmen- tation tokens are more relevant to the target span. As the results show, AUGMENT BYATTN and its variant improve ATTN UNION but are not as good as dependency parsing aug...

work page