In long context scenarios, large language models (LLMs) face three main challenges: higher computational cost, performance reduction, and position bias. Research indicates that LLM performance hinges on the density and position of key information in the input prompt. Inspired by these findings, we propose LongLLMLingua for prompt compression towards improving LLMs’ perception of the key information to simultaneously address the three challenges. Our extensive evaluation across various long context scenarios demonstrates that LongLLMLingua not only enhances performance but also significantly reduces costs and latency. For instance, in the NaturalQuestions benchmark, LongLLMLingua boosts performance by up to 21.4% with around 4x fewer tokens in GPT-3.5-Turbo, leading to substantial cost savings. It achieves a 94.0% cost reduction in the LooGLE benchmark. Moreover, when compressing prompts of about 10k tokens at ratios of 2x-6x, LongLLMLingua can accelerate end-to-end latency by 1.4x-2.6x.

This paper focuses on task-agnostic prompt compression for better generalizability and efficiency. Considering the redundancy in natural language, existing approaches compress prompts by removing tokens or lexical units according to their information entropy obtained from a causal language model such as LLaMa-7B. The challenge is that information entropy may be a suboptimal compression metric: (i) it only leverages unidirectional context and may fail to capture all essential information needed for prompt compression; (ii) it is not aligned with the prompt compression objective.To address these issues, we propose a data distillation procedure to derive knowledge from an LLM to compress prompts without losing crucial information, and meantime, introduce an extractive text compression dataset. We formulate prompt compression as a token classification problem to guarantee the faithfulness of the compressed prompt to the original one, and use a Transformer encoder as the base architecture to capture all essential information for prompt compression from the full bidirectional context. Our approach leads to lower latency by explicitly learning the compression objective with smaller models such as XLM-RoBERTa-large and mBERT.We evaluate our method on both in-domain and out-of-domain datasets, including MeetingBank, LongBench, ZeroScrolls, GSM8K, and BBH. Despite its small size, our model shows significant performance gains over strong baselines and demonstrates robust generalization ability across different LLMs. Additionally, our model is 3x-6x faster than existing prompt compression methods, while accelerating the end-to-end latency by 1.6x-2.9x with compression ratios of 2x-5x.

Cross-lingual named entity recognition (NER) aims to train an NER system that generalizes well to a target language by leveraging labeled data in a given source language. Previous work alleviates the data scarcity problem by translating source-language labeled data or performing knowledge distillation on target-language unlabeled data. However, these methods may suffer from label noise due to the automatic labeling process. In this paper, we propose CoLaDa, a Collaborative Label Denoising Framework, to address this problem. Specifically, we first explore a model-collaboration-based denoising scheme that enables models trained on different data sources to collaboratively denoise pseudo labels used by each other. We then present an instance-collaboration-based strategy that considers the label consistency of each token’s neighborhood in the representation space for denoising. Experiments on different benchmark datasets show that the proposed CoLaDa achieves superior results compared to previous methods, especially when generalizing to distant languages.

Self-supervised representation learning has proved to be a valuable component for out-of-distribution (OoD) detection with only the texts of in-distribution (ID) examples. These approaches either train a language model from scratch or fine-tune a pre-trained language model using ID examples, and then take the perplexity output by the language model as OoD scores. In this paper, we analyze the complementary characteristic of both methods and propose a multi-level knowledge distillation approach that integrates their strengths while mitigating their limitations. Specifically, we use a fine-tuned model as the teacher to teach a randomly initialized student model on the ID examples. Besides the prediction layer distillation, we present a similarity-based intermediate layer distillation method to thoroughly explore the representation space of the teacher model. In this way, the learned student can better represent the ID data manifold while gaining a stronger ability to map OoD examples outside the ID data manifold with the regularization inherited from pre-training. Besides, the student model sees only ID examples during parameter learning, further promoting more distinguishable features for OoD detection. We conduct extensive experiments over multiple benchmark datasets, i.e., CLINC150, SST, ROSTD, 20 NewsGroups, and AG News; showing that the proposed method yields new state-of-the-art performance. We also explore its application as an AIGC detector to distinguish answers generated by ChatGPT and human experts. It is observed that our model exceeds human evaluators in the pair-expert task on the Human ChatGPT Comparison Corpus.

Large language models (LLMs) have been applied in various applications due to their astonishing capabilities. With advancements in technologies such as chain-of-thought (CoT) prompting and in-context learning (ICL), the prompts fed to LLMs are becoming increasingly lengthy, even exceeding tens of thousands of tokens. To accelerate model inference and reduce cost, this paper presents LLMLingua, a coarse-to-fine prompt compression method that involves a budget controller to maintain semantic integrity under high compression ratios, a token-level iterative compression algorithm to better model the interdependence between compressed contents, and an instruction tuning based method for distribution alignment between language models. We conduct experiments and analysis over four datasets from different scenarios, i.e., GSM8K, BBH, ShareGPT, and Arxiv-March23; showing that the proposed approach yields state-of-the-art performance and allows for up to 20x compression with little performance loss.

This paper presents ReasonFormer, a unified reasoning framework for mirroring the modular and compositional reasoning process of humans in complex decision-making. Inspired by dual-process theory in cognitive science, the representation module (automatic thinking) and reasoning modules (controlled thinking) are decoupled to capture different levels of cognition. Upon the top of the representation module, the pre-trained reasoning modules are modular and professional in specific and fundamental reasoning skills (e.g., logic, simple QA, etc). To mimic the controlled compositional thinking process, different reasoning modules are dynamically activated and composed in both parallel and cascaded manners to control what reasoning skills are activated and how deep the reasoning process will be reached to solve the current problems. The unified reasoning framework solves multiple tasks with a single model, and is trained and inferred in an end-to-end manner. Evaluated on 11 datasets requiring different reasoning skills and complexity, ReasonFormer demonstrates substantial performance boosts, revealing the compositional reasoning ability. Few-shot experiments exhibit better generalization ability by learning to compose pre-trained skills for new tasks with limited data, and decoupling the representation module and the reasoning modules. Further analysis shows the modularity of reasoning modules as different tasks activate distinct reasoning skills at different reasoning depths.

Pre-trained language models (PLMs) have shown their effectiveness in multiple scenarios. However, KBQA remains challenging, especially regarding coverage and generalization settings. This is due to two main factors: i) understanding the semantics of both questions and relevant knowledge from the KB; ii) generating executable logical forms with both semantic and syntactic correctness. In this paper, we present a new KBQA model, TIARA, which addresses those issues by applying multi-grained retrieval to help the PLM focus on the most relevant KB context, viz., entities, exemplary logical forms, and schema items. Moreover, constrained decoding is used to control the output space and reduce generation errors. Experiments over important benchmarks demonstrate the effectiveness of our approach. TIARA outperforms previous SOTA, including those using PLMs or oracle entity annotations, by at least 4.1 and 1.1 F1 points on GrailQA and WebQuestionsSP, respectively. Specifically on GrailQA, TIARA outperforms previous models in all categories, with an improvement of 4.7 F1 points in zero-shot generalization.

Few-shot named entity recognition (NER) systems aim at recognizing novel-class named entities based on only a few labeled examples. In this paper, we present a decomposed meta-learning approach which addresses the problem of few-shot NER by sequentially tackling few-shot span detection and few-shot entity typing using meta-learning. In particular, we take the few-shot span detection as a sequence labeling problem and train the span detector by introducing the model-agnostic meta-learning (MAML) algorithm to find a good model parameter initialization that could fast adapt to new entity classes. For few-shot entity typing, we propose MAML-ProtoNet, i.e., MAML-enhanced prototypical networks to find a good embedding space that can better distinguish text span representations from different entity classes. Extensive experiments on various benchmarks show that our approach achieves superior performance over prior methods.

Recent studies on few-shot intent detection have attempted to formulate the task as a meta-learning problem, where a meta-learning model is trained with a certain capability to quickly adapt to newly specified few-shot tasks with potentially unseen intent categories. Prototypical networks have been commonly used in this setting, with the hope that good prototypical representations could be learned to capture the semantic similarity between the query and a few labeled instances. This intuition naturally leaves a question of whether or not a good sentence representation scheme could suffice for the task without further domain-specific adaptation. In this paper, we conduct empirical studies on a number of general-purpose sentence embedding schemes, showing that good sentence embeddings without any fine-tuning on intent detection data could produce a non-trivially strong performance. Inspired by the results from our qualitative analysis, we propose a frustratingly easy modification, which leads to consistent improvements over all sentence encoding schemes, including those from the state-of-the-art prototypical network variants with task-specific fine-tuning.

The general format of natural language inference (NLI) makes it tempting to be used for zero-shot text classification by casting any target label into a sentence of hypothesis and verifying whether or not it could be entailed by the input, aiming at generic classification applicable on any specified label space. In this opinion piece, we point out a few overlooked issues that are yet to be discussed in this line of work. We observe huge variance across different classification datasets amongst standard BERT-based NLI models and surprisingly find that pre-trained BERT without any fine-tuning can yield competitive performance against BERT fine-tuned for NLI. With the concern that these models heavily rely on spurious lexical patterns for prediction, we also experiment with preliminary approaches for more robust NLI, but the results are in general negative. Our observations reveal implicit but challenging difficulties in entailment-based zero-shot text classification.

We present Retriever-Transducer-Checker (ReTraCk), a neural semantic parsing framework for large scale knowledge base question answering (KBQA). ReTraCk is designed as a modular framework to maintain high flexibility. It includes a retriever to retrieve relevant KB items efficiently, a transducer to generate logical form with syntax correctness guarantees and a checker to improve transduction procedure. ReTraCk is ranked at top1 overall performance on the GrailQA leaderboard and obtains highly competitive performance on the typical WebQuestionsSP benchmark. Our system can interact with users timely, demonstrating the efficiency of the proposed framework.

Learning semantic correspondences between structured input data (e.g., slot-value pairs) and associated texts is a core problem for many downstream NLP applications, e.g., data-to-text generation. Large-scale datasets recently proposed for generation contain loosely corresponding data text pairs, where part of spans in text cannot be aligned to its incomplete paired input. To learn semantic correspondences from such datasets, we propose a two-stage local-to-global alignment (L2GA) framework. First, a local model based on multi-instance learning is applied to build alignments for texts spans that can be directly grounded to its paired structured input. Then, a novel global model built upon a memory-guided conditional random field (CRF) layer aims to infer missing alignments for text spans which not supported by paired incomplete inputs, where the memory is designed to leverage alignment clues provided by the local model to strengthen the global model. In this way, the local model and global model can work jointly to learn semantic correspondences in the same framework. Experimental results show that our proposed method can be generalized to both restaurant and computer domains and improve the alignment accuracy.

Recent neural models for data-to-text generation rely on massive parallel pairs of data and text to learn the writing knowledge. They often assume that writing knowledge can be acquired from the training data alone. However, when people are writing, they not only rely on the data but also consider related knowledge. In this paper, we enhance neural data-to-text models with external knowledge in a simple but effective way to improve the fidelity of generated text. Besides relying on parallel data and text as in previous work, our model attends to relevant external knowledge, encoded as a temporary memory, and combines this knowledge with the context representation of data before generating words. This allows the model to infer relevant facts which are not explicitly stated in the data table from an external knowledge source. Experimental results on twenty-one Wikipedia infobox-to-text datasets show our model, KBAtt, consistently improves a state-of-the-art model on most of the datasets. In addition, to quantify when and why external knowledge is effective, we design a metric, KBGain, which shows a strong correlation with the observed performance boost. This result demonstrates the relevance of external knowledge and sparseness of original data are the main factors affecting system performance.

Recent neural language generation systems often hallucinate contents (i.e., producing irrelevant or contradicted facts), especially when trained on loosely corresponding pairs of the input structure and text. To mitigate this issue, we propose to integrate a language understanding module for data refinement with self-training iterations to effectively induce strong equivalence between the input data and the paired text. Experiments on the E2E challenge dataset show that our proposed framework can reduce more than 50% relative unaligned noise from the original data-text pairs. A vanilla sequence-to-sequence neural NLG model trained on the refined data has improved on content correctness compared with the current state-of-the-art ensemble generator.

Most of the recently proposed neural models for named entity recognition have been purely data-driven, with a strong emphasis on getting rid of the efforts for collecting external resources or designing hand-crafted features. This could increase the chance of overfitting since the models cannot access any supervision signal beyond the small amount of annotated data, limiting their power to generalize beyond the annotated entities. In this work, we show that properly utilizing external gazetteers could benefit segmental neural NER models. We add a simple module on the recently proposed hybrid semi-Markov CRF architecture and observe some promising results.

Data-to-text generation aims to generate descriptions given a structured input data (i.e., a table with multiple records). Existing neural methods for encoding input data can be divided into two categories: a) pooling based encoders which ignore dependencies between input records or b) recurrent encoders which model only sequential dependencies between input records. In our investigation, although the recurrent encoder generally outperforms the pooling based encoder by learning the sequential dependencies, it is sensitive to the order of the input records (i.e., performance decreases when injecting the random shuffling noise over input data). To overcome this problem, we propose to adopt the self-attention mechanism to learn dependencies between arbitrary input records. Experimental results show the proposed method achieves comparable results and remains stable under random shuffling over input data.

Sequence-to-sequence model has been applied to solve math word problems. The model takes math problem descriptions as input and generates equations as output. The advantage of sequence-to-sequence model requires no feature engineering and can generate equations that do not exist in training data. However, our experimental analysis reveals that this model suffers from two shortcomings: (1) generate spurious numbers; (2) generate numbers at wrong positions. In this paper, we propose incorporating copy and alignment mechanism to the sequence-to-sequence model (namely CASS) to address these shortcomings. To train our model, we apply reinforcement learning to directly optimize the solution accuracy. It overcomes the “train-test discrepancy” issue of maximum likelihood estimation, which uses the surrogate objective of maximizing equation likelihood during training while the evaluation metric is solution accuracy (non-differentiable) at test time. Furthermore, to explore the effectiveness of our neural model, we use our model output as a feature and incorporate it into the feature-based model. Experimental results show that (1) The copy and alignment mechanism is effective to address the two issues; (2) Reinforcement learning leads to better performance than maximum likelihood on this task; (3) Our neural model is complementary to the feature-based model and their combination significantly outperforms the state-of-the-art results.

Previous work on grounded language learning did not fully capture the semantics underlying the correspondences between structured world state representations and texts, especially those between numerical values and lexical terms. In this paper, we attempt at learning explicit latent semantic annotations from paired structured tables and texts, establishing correspondences between various types of values and texts. We model the joint probability of data fields, texts, phrasal spans, and latent annotations with an adapted semi-hidden Markov model, and impose a soft statistical constraint to further improve the performance. As a by-product, we leverage the induced annotations to extract templates for language generation. Experimental results suggest the feasibility of the setting in this study, as well as the effectiveness of our proposed framework.

Recent neural models for data-to-text generation are mostly based on data-driven end-to-end training over encoder-decoder networks. Even though the generated texts are mostly fluent and informative, they often generate descriptions that are not consistent with the input structured data. This is a critical issue especially in domains that require inference or calculations over raw data. In this paper, we attempt to improve the fidelity of neural data-to-text generation by utilizing pre-executed symbolic operations. We propose a framework called Operation-guided Attention-based sequence-to-sequence network (OpAtt), with a specifically designed gating mechanism as well as a quantization module for operation results to utilize information from pre-executed operations. Experiments on two sports datasets show our proposed method clearly improves the fidelity of the generated texts to the input structured data.

Data2Text Studio is a platform for automated text generation from structured data. It is equipped with a Semi-HMMs model to extract high-quality templates and corresponding trigger conditions from parallel data automatically, which improves the interactivity and interpretability of the generated text. In addition, several easy-to-use tools are provided for developers to edit templates of pre-trained models, and APIs are released for developers to call the pre-trained model to generate texts in third-party applications. We conduct experiments on RotoWire datasets for template extraction and text generation. The results show that our model achieves improvements on both tasks.

The task of entity linking aims to identify concepts mentioned in a text fragments and link them to a reference knowledge base. Entity linking in long text has been well studied in previous work. However, short text entity linking is more challenging since the text are noisy and less coherent. To better utilize the local information provided in short texts, we propose a novel neural network framework, Aggregated Semantic Matching (ASM), in which two different aspects of semantic information between the local context and the candidate entity are captured via representation-based and interaction-based neural semantic matching models, and then two matching signals work jointly for disambiguation with a rank aggregation mechanism. Our evaluation shows that the proposed model outperforms the state-of-the-arts on public tweet datasets.

To solve math word problems, previous statistical approaches attempt at learning a direct mapping from a problem description to its corresponding equation system. However, such mappings do not include the information of a few higher-order operations that cannot be explicitly represented in equations but are required to solve the problem. The gap between natural language and equations makes it difficult for a learned model to generalize from limited data. In this work we present an intermediate meaning representation scheme that tries to reduce this gap. We use a sequence-to-sequence model with a novel attention regularization term to generate the intermediate forms, then execute them to obtain the final answers. Since the intermediate forms are latent, we propose an iterative labeling framework for learning by leveraging supervision signals from both equations and answers. Our experiments show using intermediate forms outperforms directly predicting equations.

This paper presents a novel template-based method to solve math word problems. This method learns the mappings between math concept phrases in math word problems and their math expressions from training data. For each equation template, we automatically construct a rich template sketch by aggregating information from various problems with the same template. Our approach is implemented in a two-stage system. It first retrieves a few relevant equation system templates and aligns numbers in math word problems to those templates for candidate equation generation. It then does a fine-grained inference to obtain the final answer. Experiment results show that our method achieves an accuracy of 28.4% on the linear Dolphin18K benchmark, which is 10% (54% relative) higher than previous state-of-the-art systems while achieving an accuracy increase of 12% (59% relative) on the TS6 benchmark subset.

We introduce automatic verification as a post-processing step for entity linking (EL). The proposed method trusts EL system results collectively, by assuming entity mentions are mostly linked correctly, in order to create a semantic profile of the given text using geospatial and temporal information, as well as fine-grained entity types. This profile is then used to automatically verify each linked mention individually, i.e., to predict whether it has been linked correctly or not. Verification allows leveraging a rich set of global and pairwise features that would be prohibitively expensive for EL systems employing global inference. Evaluation shows consistent improvements across datasets and systems. In particular, when applied to state-of-the-art systems, our method yields an absolute improvement in linking performance of up to 1.7 F1 on AIDA/CoNLL’03 and up to 2.4 F1 on the English TAC KBP 2015 TEDL dataset.

We present in this paper a statistical framework that generates accurate and fluent product description from product attributes. Specifically, after extracting templates and learning writing knowledge from attribute-description parallel data, we use the learned knowledge to decide what to say and how to say for product description generation. To evaluate accuracy and fluency for the generated descriptions, in addition to BLEU and Recall, we propose to measure what to say (in terms of attribute coverage) and to measure how to say (by attribute-specified generation) separately. Experimental results show that our framework is effective.

Traditional Entity Linking (EL) technologies rely on rich structures and properties in the target knowledge base (KB). However, in many applications, the KB may be as simple and sparse as lists of names of the same type (e.g., lists of products). We call it as List-only Entity Linking problem. Fortunately, some mentions may have more cues for linking, which can be used as seed mentions to bridge other mentions and the uninformative entities. In this work, we select most linkable mentions as seed mentions and disambiguate other mentions by comparing them with the seed mentions rather than directly with the entities. Our experiments on linking mentions to seven automatically mined lists show promising results and demonstrate the effectiveness of our approach.

As part of evaluating a summary automati-cally, it is usual to determine how much of the contents of one or more human-produced “ideal” summaries it contains. Past automated methods such as ROUGE compare using fixed word ngrams, which are not ideal for a variety of reasons. In this paper we describe a framework in which summary evaluation measures can be instantiated and compared, and we implement a specific evaluation method using very small units of content, called Basic Elements that address some of the shortcomings of ngrams. This method is tested on DUC 2003, 2004, and 2005 systems and produces very good correlations with human judgments.

Recent work in several computational linguistics (CL) applications (especially question answering) has shown the value of semantics (in fact, many people argue that the current performance ceiling experienced by so many CL applications derives from their inability to perform any kind of semantic processing). But the absence of a large semantic information repository that provides representations for sentences prevents the training of statistical CL engines and thus hampers the development of such semantics-enabled applications. This talk refers to recent work in several projects that seek to annotate large volumes of text with shallower or deeper representations of some semantic phenomena. It describes one of the essential problems—creating, managing, and annotating (at large scale) the meanings of words, and outlines the Omega ontology, being built at ISI, that acts as term repository. The talk illustrates how one can proceed from words via senses to concepts, and how the annotation process can help verify good concept decisions and expose bad ones. Much of this work is performed in the context of the OntoNotes project, joint with BBN, the Universities of Colorado and Pennsylvania, and ISI, that is working to build a corpus of about 1M words (English, Chinese, and Arabic), annotated for shallow semantics, over the next few years.

In this paper we describe the design and implementation of MuST, a multilingual information retrieval, summarization, and translation system. MuST integrates machine translation and other text processing services to enable users to perform cross-language information retrieval using available search services such as commercial Internet search engines. To handle non-standard languages, a new Internet indexing agent can be deployed, specialized local search services can be built, and shallow MT can be added to provide useful functionality. A case study of augmenting MuST with Indonesian is included. MuST adopts ubiquitous web browsers as its primary user interface, and provides tightly integrated automated shallow translation and user biased summarization to help users quickly judge the relevance of documents.