Bilingual is At Least Monolingual (BALM): A Novel Translation Algorithm that Encodes Monolingual Priors

Chris Callison-Burch; Jeffrey Cheng

REVIEW 4 major objections 6 minor 14 references

Bilingual is At Least Monolingual (BALM): A Novel Translation Algorithm that Encodes Monolingual Priors

T0 review · 4 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read BALM claims that machine translation can be reduced to a fixed-length mapping between BERT thought-spaces, letting a shallow feedforward network translate with less parallel data.

desk verdict A plausible but unproven idea: the paper's real result is a feedforward translator on BERT mean-pooled embeddings that gets 24.8 BLEU, but the abstract's near-SOTA claim is false on its own numbers and the autoencoder validation is confounded by fine-tuning. read the letter →

arxiv 1909.01146 v1 pith:3F3SQFXH submitted 2019-08-30 cs.CL cs.LG

Jeffrey Cheng , Chris Callison-Burch This is my paper · ORCID

classification cs.CLcs.LG

keywords machinetranslationBERTsentenceembeddingsthought-spacefeedforwardnetworklow-resourceautoencoderMulti30k

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 tries to establish that machine translation need not be learned as an end-to-end sequence-to-sequence problem. It proposes BALM, which uses pretrained BERT encoders to project each language's sentences into fixed-length thought-spaces, trains a shallow feedforward network to map one language's thought-space to the other, and decodes with a GRU pretrained to invert the target BERT embedding. The claim is that this makes translation a fixed-length regression problem, so a simple model and little parallel data suffice. The paper reports an English autoencoder BLEU of 0.605 and a German-to-English translation BLEU of 0.248, below the 0.35 state of the art it cites, and frames the result as evidence that monolingual priors can carry most of the translation burden.

What carries the argument

BALM's load-bearing object is the composition $T = B_{\mathrm{German}} \circ F \circ B_{\mathrm{English}}^{-1}$. Here $B_L$ is a pretrained BERT encoder that maps sentences of language $L$ to a fixed-length thought-space $\mathcal{S}_{L,768} \subseteq \mathbb{R}^{768}$ by mean-pooling the final hidden layer; $F$ is a shallow feedforward network with a 768-768-768 architecture and ReLU activations, trained to align German and English thought-vectors; and $B_{\mathrm{English}}^{-1}$ is a single-layer GRU decoder trained to reconstruct English from the thought-vector with teacher forcing. This composition converts variable-length sequence-to-sequence translation into a fixed-length regression problem, which is why a feedforward network suffices for the bilingual mapping. In the experiments, the BERT encoders are fine-tuned with gradient updates during both autoencoder and translator training, and the GRU decoder is initialized from the autoencoder and then fine-tuned.

What would settle it

Freeze a pretrained English BERT encoder, train only the GRU decoder to reconstruct Multi30k captions, and compare reconstruction BLEU with the reported 0.605; a large drop would show the autoencoder's success depended on fine-tuning rather than the pretrained thought-space. Separately, train a standard attention-based sequence-to-sequence system and BALM on the same small parallel subsets of Multi30k and compare BLEU as a function of the number of parallel sentences; if BALM does not reach comparable quality with markedly fewer parallel pairs, the low-resource data-efficiency claim is unsupported.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a mean-pooled BERT vector is a rich enough thought-space to be inverted into natural language, and that translating between two such spaces is a fixed-length regression. The BALM model composes a German BERT encoder, a feedforward thought-translator, and an English BERT-inverting GRU decoder; the paper reports an English autoencoder with BLEU 0.605 and a German-to-English translator at BLEU 0.248, with cross-entropy losses near zero on a 28,996-token vocabulary. The paper takes the autoencoder's success as evidence that BERT creates sentence embeddings, and the translator's success as evidence that sequence-to-sequence problems like MT can be solved by simple fixed-length models.

Load-bearing premise

The whole framework assumes that a single fixed-length summary of a sentence, computed by averaging BERT's word vectors, contains enough information to reconstruct the entire sentence, and that this invertibility holds for the pretrained embeddings used for translation rather than emerging only after task-specific fine-tuning.

Editorial extensions

If this is right

Under BALM, parallel data are needed only to train the fixed-length thought-translator, so translation between language pairs with scarce bilingual corpora becomes a small supervised learning problem on top of monolingual pretraining.
Since the only model trained from scratch in translation is a shallow feedforward network, the framework should converge in far fewer epochs than recurrent or transformer baselines.
If the autoencoder result transfers, the pretrained English thought-decoder can be reused for any source language that maps into the same English thought-space.
The separation of encoder, translator, and decoder gives a diagnostic trail: bad translations can be attributed to one of the three submodules by re-encoding outputs.
Because the fixed-length thought-space can be fed to any classifier, replacing the feedforward translator with even simpler models or regularized variants is a direct next test.

Reading between the lines

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

Editorial inference: Because the autoencoder fine-tunes BERT during training, the reported reconstruction quality does not by itself prove that the published pretrained mean-pooled embeddings are invertible; a frozen-encoder reconstruction experiment would settle which component carries the capacity.
Editorial inference: The measured 0.248 BLEU against the paper's own stated 0.35 SOTA is not near-SOTA in the usual sense; the defensible claim is that a dramatically simpler model learns nontrivial translation, and the more interesting comparison is data efficiency, not final BLEU.
Editorial inference: The same three-module scheme could be applied to any language pair with two monolingual BERT-style encoders and no parallel data except the thought-translator; a natural test is a low-resource pair without image-captioning bias.
Editorial inference: The thought-space view suggests testable diagnostics: if the translator produces a bad sentence, re-encode that output with the English BERT and compare the recovered thought-vector to the source German thought-vector to see whether the error lies in the thought-translator or in the decoder.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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

Colleague,

The one thing to know: this is a plausible but unproven idea, and the abstract overstates it. The actual result is a feedforward mapping between mean-pooled BERT sentence vectors plus a GRU decoder that gets 24.8 BLEU on German-English Multi30k. That is nontrivial – BLEU above 15 usually means real learning – and the paper is honest that it trails the 35 BLEU state of the art. But calling 24.8 "near-SOTA" is wrong, and the conclusion that BERT lets simple models solve MT is not supported by the experiments as run.

What's new: composing a pretrained BERT encoder, a 768x768x768 feedforward translator, and a transfer-learned GRU decoder. Each piece exists, but this specific composition is not in the cited literature. The paper also ships code and pretrained models, which is useful for anyone wanting to reproduce or build on it.

Where it falls down: the central premise is that the fixed-length mean-pooled BERT vector is a thought-space rich enough to reconstruct a sentence. The only evidence is the autoencoder in §4.1, but §3.2 says gradient updates are allowed to fine-tune the BERT encoder. So the 0.605 reconstruction BLEU shows a fine-tuned encoder is invertible, not that the off-the-shelf pretrained embedding is. Then the same decoder is fine-tuned again in the translation stage, so the monolingual prior contribution is never isolated. There are no baselines: no randomly-initialized BERT, no fixed BERT without fine-tuning, no standard seq2seq under identical conditions, and no repeated runs or error bars. The low-resource advantage is asserted, not demonstrated. On the plus side, the paper explicitly notes the translation is sub-SOTA and overfits the training set, so it is not hiding its weaknesses.

Who is this for: someone working on low-resource MT or on whether sentence embeddings can be lossy thought-spaces. It's a preprint from 2019, so much of this has been overtaken by later work, but as a statement of a framework it could be a useful discussion piece.

Should a serious referee spend time on it? Yes, I think so. The core question – can monolingual priors turn MT into a fixed-length mapping – is worth testing rigorously, and this paper provides a concrete architecture and public artifacts for that test. It just needs a corrected abstract, controlled baselines, and separation of the pretrained-embedding claim from the fine-tuning confound.

My vote: send to peer review with a request for major revision, and flag the abstract and the autoencoder confound as the two things that must be fixed.

Referee Report

4 major / 6 minor

Summary. The paper proposes BALM, a translation framework that uses pretrained BERT sentence embeddings as fixed-length 'thought-spaces' for source and target languages, so that translation reduces to a fixed-length feedforward mapping between thought-spaces plus a learned recurrent decoder. The authors first train an English autoencoder (BERT encoder + GRU decoder) and report a reconstruction BLEU of 0.605; they then train a German-to-English translator using a multilingual BERT encoder, a 768x768x768 feedforward network, and the transferred GRU decoder, reporting a BLEU of 0.248 on Multi30k against a stated SOTA of 0.35. They conclude that BERT embeddings allow complex sequence-to-sequence tasks to be solved with simple feedforward networks and that the approach is promising for low-resource translation.

Significance. If the central claims held, the paper would make a useful conceptual contribution: casting MT as a fixed-length mapping, potentially reducing data hunger, and providing evidence that mean-pooled BERT vectors are invertible sentence representations. The paper ships code and model links, and the reported 24.8 BLEU with a feedforward translator is a real empirical data point. However, the strength of the conclusions far exceeds what the experiments establish: the key validation is confounded by fine-tuning, the 'near-SOTA' characterization is contradicted by the paper's own numbers, and no comparison is made with the baselines used to support the convergence and low-resource claims.

major comments (4)

[Abstract and §4.2.2] The abstract calls the English-German result 'near-SOTA', but §4.2.2 reports BLEU 0.248 against a stated SOTA of 0.35, a relative gap of roughly 29% (0.102 BLEU). This is not near-SOTA by any standard, and the same overstatement appears in the conclusion (§5). The claim should be corrected to 'sub-SOTA but above random chance' or supported by substantially better results.
[§3.2 and §4.1] The autoencoder used to validate BERT mean-pooling as a sentence embedding allows gradient updates to backpropagate through the pretrained BERT encoder and fine-tune it (§3.2). Therefore the reported 0.605 reconstruction BLEU demonstrates invertibility of a task-fine-tuned encoder, not of the published pretrained mean-pooled representation that the method claims to exploit. Since the translator also fine-tunes both the BERT encoder and the decoder (§3.2), the experiments never isolate the contribution of the monolingual thought-space. The authors should either freeze BERT during autoencoder training or compare against a randomly initialized encoder to support the claim that pretrained BERT embeddings are the source of the capability.
[§4.2.1 and §5] The paper claims that the BALM translator 'does seem to converge faster than both seq2seq and transformer-based MT systems' (§5), but no seq2seq or transformer baseline is trained or plotted in §4.2.1. The learning curve alone cannot support this claim; a direct comparison of training curves on the same data and hardware is required, or the claim should be removed.
[§1, §6, and §7] The motivation and impact sections emphasize the value for low-resource language pairs, but no experiment varies the amount of parallel data or tests a genuinely low-resource pair. The Multi30k setup uses 29,000 training captions and 200 epochs, which does not demonstrate data efficiency. This claim should be labeled as speculation or supported by a data-ablation experiment.

minor comments (6)

[§2.2] The composition order is written incorrectly: the translator should be T = B^{-1}_English ∘ F ∘ BGerman, but the paper writes T = BGerman ∘ F ∘ B^{-1}_English, which would apply the English decoder first. The same issue appears in §2.1 for the autoencoder.
[§2.1] The equation 'BGerman : LEnglish → SEnglish,k' uses the German encoder symbol for an English encoding; this should be BEnglish.
[Throughout] There are numerous typos and grammatical errors, including 'A observation' (§1), 'langauge' (§1.1), 'disamguation' (§1.2), 'BLAM' (§7), 'taken' for 'token' (§3.2), and the missing 'that' in the abstract ('BALM is a framework incorporates').
[§4.1.2 and §4.2.2] The BLEU numbers are reported without evaluation details (tokenization, case sensitivity, which reference split, whether sacrebleu or another BLEU implementation was used) or variance across runs; this information should be provided for reproducibility.
[§4.2.2] The SOTA value of 0.35 for Multi30k is cited without a reference; please provide the source.
[§1.3] Reference [14] is a software repository rather than a peer-reviewed study establishing mean-pooling as a sentence embedding; a citable publication would strengthen the motivation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the translation result is measured against an external benchmark, and the autoencoder fine-tuning issue is a confound rather than a circular reduction.

full rationale

The paper's central empirical claims are not circular. The translation BLEU of 0.248 is computed against the held-out Multi30k test set and compared with an external SOTA figure of 0.35, so the result does not reduce to the training objective. The BALM architecture is defined as T = B_German composed with F composed with B^{-1}_English, and F is trained as a fixed-length mapping; no equation in the paper is shown to be equivalent to its own input by construction. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in through the authors' prior work. The main evidentiary weakness is in Section 4.1: the autoencoder offered as proof that 'mean-pooling over BERT's word embedding does create sentence embeddings' allows gradient updates to fine-tune the BERT encoder during training (Section 3.2), so the 0.605 reconstruction BLEU partly reflects the task-fitted encoder rather than the frozen pretrained representation. That is a confound and an overclaim, but it is not circularity: the BLEU score is a measured outcome of a trained system, not a quantity defined as its own prediction. The abstract's 'near-SOTA' characterization is also contradicted by the paper's own Section 4.2.2, which states that the score 'falls short of the state-of-the-art performance on Multi30K of 0.35'; that tension is a correctness or calibration issue, not a circular derivation. The derivation chain is therefore self-contained with respect to circularity, even though the evidence for one of the paper's premises is weaker than the conclusion claims.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The central claim rests on assumptions about BERT embeddings, transferability of the decoder, and representativeness of Multi30k. The main uncharged input is that BERT's mean-pooled representation is an invertible sentence embedding; this is tested in a way that confounds fine-tuning. No new entities beyond a relabeled thought-space concept are introduced.

free parameters (4)

Hyperparameters (values in Figure 5) = not stated in text
Tuned to minimize validation loss; exact values only in a figure not reproduced in the text.
Feedforward thought-translator weights (768x768x768) = learned on Multi30k parallel captions
Core mapping F_German_to_English; all translation behavior is fit here.
Fine-tuned BERT encoder weights = fine-tuned on autoencoder and translation losses
Both BERT encoders are fine-tuned, so the pretrained embeddings are not used as fixed priors in the final model; this weakens the claim that observed success comes from pretrained monolingual priors.
Mean-pooling choice (last hidden layer, 768 dims) = 768
Manual choice of which BERT layer and pooling strategy to use; not derived from theory.

assumptions (5)

domain assumption Mean-pooled BERT word embeddings form a sentence embedding (a thought-space) that contains enough information for reconstruction.
Assumed in §1.3 and §2.1; the autoencoder test in §4.1 is meant to verify it, but BERT is fine-tuned during that test.
domain assumption A fixed-length vector can capture the semantic content of a sentence for translation.
Core premise of BALM (Definition 2.1), but the paper's own section 5 quotes Ray Mooney's contrary claim and does not fully resolve it.
domain assumption Pretrained BERT models provide strong monolingual language priors that transfer to translation.
The method downloads pretrained BERT (§3.2) and relies on its embeddings; no non-BERT control is reported.
domain assumption The English GRU decoder trained in the autoencoder can be transfer-learned to decode English sentences from German-derived thought vectors.
Used in §2.2; the decoder is fine-tuned on parallel data, so the transfer assumption is not isolated.
domain assumption Multi30k captions are an adequate testbed for general MT claims.
The paper itself notes the dataset contains only descriptive image captions (§3.1), limiting generalization.

invented entities (2)

Thought-space S_L,k
purpose: A formal named embedding space for sentences of language L, used to frame translation as a fixed-length mapping.
This is a relabeling of the latent space that any encoder-decoder already produces; no independent falsifiable prediction is attached to the term.
BALM framework
purpose: Pipeline combining BERT encoders, a feedforward thought translator, and a GRU decoder.
A system composition rather than a new physical or mathematical entity; its value is entirely empirical.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Bilingual is At Least Monolingual (BALM): A Novel Translation Algorithm that Encodes Monolingual Priors." pith.science (2026). https://pith.science/paper/3F3SQFXH

@misc{pith2026190901146,
  author       = {Pith},
  title        = {Pith review of: Bilingual is At Least Monolingual (BALM): A Novel Translation Algorithm that Encodes Monolingual Priors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3F3SQFXH}},
  note         = {Machine review of arXiv:1909.01146}
}

read the original abstract

State-of-the-art machine translation (MT) models do not use knowledge of any single language's structure; this is the equivalent of asking someone to translate from English to German while knowing neither language. BALM is a framework incorporates monolingual priors into an MT pipeline; by casting input and output languages into embedded space using BERT, we can solve machine translation with much simpler models. We find that English-to-German translation on the Multi30k dataset can be solved with a simple feedforward network under the BALM framework with near-SOTA BLEU scores.

Figures

Figures reproduced from arXiv: 1909.01146 by the authors.

**Figure 1.** We see two images here. (Right) Each image has an independently written caption in both English and German. (Left) Each independent caption has a corresponding translation in the dataset. Image courtesy of Elliott et al. By the nature of image captioning, this dataset is biased towards static structures (e.g. "This is a white house.") and animals doing things ("The brown dog is drinking from the bowl."). Notably, th… view at source ↗

**Figure 2.** Three libraries transform natural language data into model-ready batch-wise tensors. Original flow chart diagram and original descriptions. We now expand out the model and explain the forward passes of the autoencoder and the translator. 3.2 Model implementation Autoencoder Implementation We implement the BALM autoencoder model in native PyTorch with the following modules: • The pre-trained English BERT encoder. Th… view at source ↗

**Figure 3.** Original flow chart diagram and original descriptions. Translator Implementation We implement the translator with the same framework as the autoencoder, substituting the English BERT encoder for a German BERT encoder. We also add in an intermediate module to learn the desired function FGerman→English mapping between the thought-spaces of the two languages [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Original flow chart diagram and original descriptions [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: The final tuned hyperparameters of both the autoencoder setup and the translator setup. They are the same for both pipelines for consistency. Given the above hyperparameters, we run the autoencoder model and the translator model in sequence. Each model’s training run o…

**Figure 6.** Figure 6: Selected examples of the BALM autoencoder’s reconstructions. We note that on examples from the Multi30K dataset (train and test), the autoencoder has perfect reconstructions. For manually-written examples that are somewhat similar [PITH_FULL_IMAGE:figures/full_fig_p01…

**Figure 7.** Figure 7: Selected examples of the BALM translator’s reconstructions. We see that the translation model does relatively well on the training examples. In fact, it produces the strongest signal of natural language understanding: a correct translation that is synonymous but not id…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 6 canonical work pages

[1]

Neural machine translation by jointly learning to align and translate, 2014

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate, 2014

2014
[2]

Using monolingual data in neural machine trans- lation: a systematic study

Franck Burlot and François Yvon. Using monolingual data in neural machine trans- lation: a systematic study. Proceedings of the Third Conference on Machine Translation , 2018

work page 2018
[3]

Bert: Pre-training of deep bidirectional transformers for language understanding, 2018

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2018

2018
[4]

Multi30k: Multilin- gual english-german image descriptions

Desmond Elliott, Stella Frank, Khalil Sima’an, and Lucia Specia. Multi30k: Multilin- gual english-german image descriptions. CoRR, abs/1605.00459, 2016

arXiv 2016
[5]

Long short-term memory

Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computa- tion, 9(8):1735–1780, 1997

work page 1997
[6]

Ryan Kiros, Ruslan Salakhutdinov, and Richard S. Zemel. Unifying visual-semantic embeddings with multimodal neural language models. CoRR, abs/1411.2539, 2014

arXiv 2014
[7]

Six challenges for neural machine translation

Philipp Koehn and Rebecca Knowles. Six challenges for neural machine translation. CoRR, abs/1706.03872, 2017

arXiv 2017
[8]

Bleu: A method for automatic evaluation of machine translation

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, pages 311–318, Stroudsburg, PA, USA, 2002. Association for Computational Linguistics

work page 2002

Show all 14 references

[9]

Romanian text corpora

Michael Rundell. Romanian text corpora. https://www.sketchengine.eu/user-guide/user-manual/corpora/by- language/romanian-text-corpora/, 2018

2018
[10]

Improving neural machine translation models with monolingual data

Rico Sennrich, Barry Haddow, and Alexandra Birch. Improving neural machine translation models with monolingual data. CoRR, abs/1511.06709, 2015

2015 arXiv
[11]

Cold fusion: Training seq2seq models together with language models

Anuroop Sriram, Heewoo Jun, Sanjeev Satheesh, and Adam Coates. Cold fusion: Training seq2seq models together with language models. CoRR, abs/1708.06426, 2017

2017 arXiv
[12]

Parallel corpora in the clarin infrastructure

Clarin Eric: Utrecht University. Parallel corpora in the clarin infrastructure. https://www.clarin.eu/resource-families/parallel-corpora, 2018

2018
[13]

Gomez, Lukasz Kaiser, and Illia Polosukhin

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2017

2017
[14]

Bert-as-a-service

Han Xiao. Bert-as-a-service. https://github.com/hanxiao/bert-as-service, 2018

2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Neural machine translation by jointly learning to align and translate, 2014

Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate, 2014

2014

[2] [2]

Using monolingual data in neural machine trans- lation: a systematic study

Franck Burlot and François Yvon. Using monolingual data in neural machine trans- lation: a systematic study. Proceedings of the Third Conference on Machine Translation , 2018

work page 2018

[3] [3]

Bert: Pre-training of deep bidirectional transformers for language understanding, 2018

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2018

2018

[4] [4]

Multi30k: Multilin- gual english-german image descriptions

Desmond Elliott, Stella Frank, Khalil Sima’an, and Lucia Specia. Multi30k: Multilin- gual english-german image descriptions. CoRR, abs/1605.00459, 2016

arXiv 2016

[5] [5]

Long short-term memory

Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computa- tion, 9(8):1735–1780, 1997

work page 1997

[6] [6]

Ryan Kiros, Ruslan Salakhutdinov, and Richard S. Zemel. Unifying visual-semantic embeddings with multimodal neural language models. CoRR, abs/1411.2539, 2014

arXiv 2014

[7] [7]

Six challenges for neural machine translation

Philipp Koehn and Rebecca Knowles. Six challenges for neural machine translation. CoRR, abs/1706.03872, 2017

arXiv 2017

[8] [8]

Bleu: A method for automatic evaluation of machine translation

Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: A method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association for Computational Linguistics, ACL ’02, pages 311–318, Stroudsburg, PA, USA, 2002. Association for Computational Linguistics

work page 2002