Efficient Continual Pre-training of LLMs for Low-resource Languages

Open-source Large Language models (OsLLMs) propel the democratization of natural language research by giving the flexibility to augment or update model parameters for performance improvement. Nevertheless, like proprietary LLMs, Os-LLMs offer poorer performance on low-resource languages (LRLs) than high-resource languages (HRLs), owing to smaller amounts of training data and underrepresented vocabulary. On the other hand, continual pre-training (CPT) with large amounts of language-specific data is a costly proposition in terms of data acquisition and computational resources. Our goal is to drastically reduce CPT cost. To that end, we first develop a new algorithm to select a subset of texts from a larger corpus. We show the effectiveness of our technique using very little CPT data. In search of further improvement, we design a new algorithm to select tokens to include in the LLM vocabulary. We experiment with the recent Llama-3 model and nine Indian languages with diverse scripts and extent of resource availability. For evaluation, we use IndicGenBench, a generation task benchmark dataset for Indic languages. We experiment with various CPT corpora and augmented vocabulary size and offer insights across language families.

Graph Edit Distance with General Costs Using Neural Set Divergence

Graph Edit Distance (GED) measures the (dis-)similarity between two given graphs, in terms of the minimum-cost edit sequence that transforms one graph to the other. However, the exact computation of GED is NP-Hard, which has recently motivated the design of neural methods for GED estimation. However, they do not explicitly account for edit operations with different costs. In response, we propose GRAPHEDX, a neural GED estimator that can work with general costs specified for the four edit operations, viz., edge deletion, edge addition, node deletion and node addition. We first present GED as a quadratic assignment problem (QAP) that incorporates these four costs. Then, we represent each graph as a set of node and edge embeddings and use them to design a family of neural set divergence surrogates. We replace the QAP terms corresponding to each operation with their surrogates. Computing such neural set divergence require aligning nodes and edges of the two graphs. We learn these alignments using a Gumbel-Sinkhorn permutation generator, additionally ensuring that the node and edge alignments are consistent with each other. Moreover, these alignments are cognizant of both the presence and absence of edges between node-pairs. Experiments on several datasets, under a variety of edit cost settings, show that GRAPHEDX consistently outperforms state-of-the-art methods and heuristics in terms of prediction error.

Cost-Performance Optimization for Processing Low-Resource Language Tasks Using Commercial LLMs

Large Language Models (LLMs) exhibit impressive zero/few-shot inference and generation quality for high-resource languages(HRLs). A few of them have been trained in low-resource languages (LRLs) and give decent performance. Owing to the prohibitive costs of training LLMs, they are usually used as a network service, with the client charged by the count of input and output tokens. The number of tokens strongly depends on the script and language, as well as the LLM's sub-word vocabulary. We show that LRLs are at a pricing disadvantage, because the well-known LLMs produce more tokens for LRLs than HRLs. This is because most currently popular LLMs are optimized for HRL vocabularies. Our objective is to level the playing field: reduce the cost of processing LRLs in contemporary LLMs while ensuring that predictive and generative qualities are not compromised. As means to reduce the number of tokens processed by the LLM, we consider code-mixing, translation, and transliteration of LRLs to HRLs. We perform an extensive study using the IndicXTREME dataset, covering 15 Indian languages, while using GPT-4 (one of the costliest LLM services released so far) as a commercial LLM. We observe and analyze interesting patterns involving token count, cost,and quality across a multitude of languages and tasks. We show that choosing the best policy to interact with the LLM can reduce cost by 90% while giving better or comparable performance, compared to communicating with the LLM in the original LRL.

Graph Regularized Encoder Training for Extreme Classification

Deep extreme classification (XC) aims to train an encoder architecture and an accompanying classifier architecture to tag a data point with the most relevant subset of labels from a very large universe of labels. XC applications in ranking, recommendation and tagging routinely encounter tail labels for which the amount of training data is exceedingly small. Graph convolutional networks (GCN) present a convenient but computationally expensive way to leverage task metadata and enhance model accuracies in these settings. This paper formally establishes that in several use cases, the steep computational cost of GCNs is entirely avoidable by replacing GCNs with non-GCN architectures. The paper notices that in these settings, it is much more effective to use graph data to regularize encoder training than to implement a GCN. Based on these insights, an alternative paradigm RAMEN is presented to utilize graph metadata in XC settings that offers significant performance boosts with zero increase in inference computational costs. RAMEN scales to datasets with up to 1M labels and offers prediction accuracy up to 15% higher on benchmark datasets than state of the art methods, including those that use graph metadata to train GCNs. RAMEN also offers 10% higher accuracy over the best baseline on a proprietary recommendation dataset sourced from click logs of a popular search engine. Code for RAMEN will be released publicly.

How to think step-by-step: A mechanistic understanding of chain-of-thought reasoning

Despite superior reasoning prowess demonstrated by Large Language Models (LLMs) with Chain-of-Thought (CoT) prompting, a lack of understanding prevails around the internal mechanisms of the models that facilitate CoT generation. This work investigates the neural sub-structures within LLMs that manifest CoT reasoning from a mechanistic point of view. From an analysis of LLaMA-2 7B applied to multistep reasoning over fictional ontologies, we demonstrate that LLMs deploy multiple parallel pathways of answer generation for step-by-step reasoning. These parallel pathways provide sequential answers from the input question context as well as the generated CoT. We observe a striking functional rift in the middle layers of the LLM. Token representations in the initial half remain strongly biased towards the pretraining prior, with the in-context taking over abruptly in the later half. This internal phase shift manifests in different functional components: attention heads that write the answer token predominantly appear in the later half, attention heads that move information along ontological relationships appear exclusively in the initial half, and so on. To the best of our knowledge, this is the first attempt towards mechanistic investigation of CoT reasoning in LLMs.

Frugal LMs Trained to Invoke Symbolic Solvers Achieve Parameter-Efficient Arithmetic Reasoning

Large Language Models (LLM) exhibit zero-shot mathematical reasoning capacity as a behavior emergent with scale, commonly manifesting as chain-of-thoughts (CoT) reasoning. However, multiple empirical findings suggest that this prowess is exclusive to LLMs with exorbitant sizes (beyond 50 billion parameters). Meanwhile, educational neuroscientists suggest that symbolic algebraic manipulation be introduced around the same time as arithmetic word problems to modularize language-to-formulation, symbolic manipulation of the formulation, and endgame arithmetic. In this paper, we start with the hypothesis that much smaller LMs, which are weak at multi-step reasoning, can achieve reasonable arithmetic reasoning if arithmetic word problems are posed as a formalize-then-solve task. In our architecture, which we call SYRELM, the LM serves the role of a translator to map natural language arithmetic questions into a formal language (FL) description. A symbolic solver then evaluates the FL expression to obtain the answer. A small frozen LM, equipped with an efficient low-rank adapter, is capable of generating FL expressions that incorporate natural language descriptions of the arithmetic problem (e.g., variable names and their purposes, formal expressions combining variables, etc.). We adopt policy-gradient reinforcement learning to train the adapted LM, informed by the non-differentiable symbolic solver. This marks a sharp departure from the recent development in tool-augmented LLMs, in which the external tools (e.g., calculator, Web search, etc.) are essentially detached from the learning phase of the LM. SYRELM shows massive improvements (e.g., +30.65 absolute point improvement in accuracy on the SVAMP dataset using GPT-J 6B model) over base LMs, while keeping our testbed easy to diagnose, interpret and within reach of most researchers.

CoRE-CoG: Conversational Recommendation of Entities using Constrained Generation

End-to-end conversational recommendation systems (CRS) generate responses by leveraging both dialog history and a knowledge base (KB). A CRS mainly faces three key challenges: (1) at each turn, it must decide if recommending a KB entity is appropriate; if so, it must identify the most relevant KB entity to recommend; and finally, it must recommend the entity in a fluent utterance that is consistent with the conversation history. Recent CRSs do not pay sufficient attention to these desiderata, often generating unfluent responses or not recommending (relevant) entities at the right turn. We introduce a new CRS we call CoRE-CoG. CoRE-CoG addresses the limitations in prior systems by implementing (1) a recommendation trigger that decides if the system utterance should include an entity, (2) a type pruning module that improves the relevance of recommended entities, and (3) a novel constrained response generator to make recommendations while maintaining fluency. Together, these modules ensure simultaneous accurate recommendation decisions and fluent system utterances. Experiments with recent benchmarks show the superiority particularly on conditional generation sub-tasks with close to 10 F1 and 4 Recall@1 percent points gain over baselines.

CRUSH4SQL: Collective Retrieval Using Schema Hallucination For Text2SQL

Existing Text-to-SQL generators require the entire schema to be encoded with the user text. This is expensive or impractical for large databases with tens of thousands of columns. Standard dense retrieval techniques are inadequate for schema subsetting of a large structured database, where the correct semantics of retrieval demands that we rank sets of schema elements rather than individual elements. In response, we propose a two-stage process for effective coverage during retrieval. First, we instruct an LLM to hallucinate a minimal DB schema deemed adequate to answer the query. We use the hallucinated schema to retrieve a subset of the actual schema, by composing the results from multiple dense retrievals. Remarkably, hallucination $\unicode{x2013}$ generally considered a nuisance $\unicode{x2013}$ turns out to be actually useful as a bridging mechanism. Since no existing benchmarks exist for schema subsetting on large databases, we introduce three benchmarks. Two semi-synthetic datasets are derived from the union of schemas in two well-known datasets, SPIDER and BIRD, resulting in 4502 and 798 schema elements respectively. A real-life benchmark called SocialDB is sourced from an actual large data warehouse comprising 17844 schema elements. We show that our method1 leads to significantly higher recall than SOTA retrieval-based augmentation methods.

Small Language Models Fine-tuned to Coordinate Larger Language Models improve Complex Reasoning

Large Language Models (LLMs) prompted to generate chain-of-thought (CoT) exhibit impressive reasoning capabilities. Recent attempts at prompt decomposition toward solving complex, multi-step reasoning problems depend on the ability of the LLM to simultaneously decompose and solve the problem. A significant disadvantage is that foundational LLMs are typically not available for fine-tuning, making adaptation computationally prohibitive. We believe (and demonstrate) that problem decomposition and solution generation are distinct capabilites, better addressed in separate modules, than by one monolithic LLM. We introduce DaSLaM, which uses a decomposition generator to decompose complex problems into subproblems that require fewer reasoning steps. These subproblems are answered by a solver. We use a relatively small (13B parameters) LM as the decomposition generator, which we train using policy gradient optimization to interact with a solver LM (regarded as black-box) and guide it through subproblems, thereby rendering our method solver-agnostic. Evaluation on multiple different reasoning datasets reveal that with our method, a 175 billion parameter LM (text-davinci-003) can produce competitive or even better performance, compared to its orders-of-magnitude larger successor, GPT-4. Additionally, we show that DaSLaM is not limited by the solver's capabilities as a function of scale; e.g., solver LMs with diverse sizes give significant performance improvement with our solver-agnostic decomposition technique. Exhaustive ablation studies evince the superiority of our modular finetuning technique over exorbitantly large decomposer LLMs, based on prompting alone.

Structured Case-based Reasoning for Inference-time Adaptation of Text-to-SQL parsers

Inference-time adaptation methods for semantic parsing are useful for leveraging examples from newly-observed domains without repeated fine-tuning. Existing approaches typically bias the decoder by simply concatenating input-output example pairs (cases) from the new domain at the encoder's input in a Seq-to-Seq model. Such methods cannot adequately leverage the structure of logical forms in the case examples. We propose StructCBR, a structured case-based reasoning approach, which leverages subtree-level similarity between logical forms of cases and candidate outputs, resulting in better decoder decisions. For the task of adapting Text-to-SQL models to unseen schemas, we show that exploiting case examples in a structured manner via StructCBR offers consistent performance improvements over prior inference-time adaptation methods across five different databases. To the best of our knowledge, we are the first to attempt inference-time adaptation of Text-to-SQL models, and harness trainable structured similarity between subqueries.